1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
65 #include <linux/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
74 #define MEM_CGROUP_RECLAIM_RETRIES 5
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket
;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem
;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly
;
86 #define do_swap_account 0
89 /* Whether legacy memory+swap accounting is active */
90 static bool do_memsw_account(void)
92 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
95 static const char *const mem_cgroup_lru_names
[] = {
103 #define THRESHOLDS_EVENTS_TARGET 128
104 #define SOFTLIMIT_EVENTS_TARGET 1024
105 #define NUMAINFO_EVENTS_TARGET 1024
108 * Cgroups above their limits are maintained in a RB-Tree, independent of
109 * their hierarchy representation
112 struct mem_cgroup_tree_per_node
{
113 struct rb_root rb_root
;
114 struct rb_node
*rb_rightmost
;
118 struct mem_cgroup_tree
{
119 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
122 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
125 struct mem_cgroup_eventfd_list
{
126 struct list_head list
;
127 struct eventfd_ctx
*eventfd
;
131 * cgroup_event represents events which userspace want to receive.
133 struct mem_cgroup_event
{
135 * memcg which the event belongs to.
137 struct mem_cgroup
*memcg
;
139 * eventfd to signal userspace about the event.
141 struct eventfd_ctx
*eventfd
;
143 * Each of these stored in a list by the cgroup.
145 struct list_head list
;
147 * register_event() callback will be used to add new userspace
148 * waiter for changes related to this event. Use eventfd_signal()
149 * on eventfd to send notification to userspace.
151 int (*register_event
)(struct mem_cgroup
*memcg
,
152 struct eventfd_ctx
*eventfd
, const char *args
);
154 * unregister_event() callback will be called when userspace closes
155 * the eventfd or on cgroup removing. This callback must be set,
156 * if you want provide notification functionality.
158 void (*unregister_event
)(struct mem_cgroup
*memcg
,
159 struct eventfd_ctx
*eventfd
);
161 * All fields below needed to unregister event when
162 * userspace closes eventfd.
165 wait_queue_head_t
*wqh
;
166 wait_queue_entry_t wait
;
167 struct work_struct remove
;
170 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
171 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
173 /* Stuffs for move charges at task migration. */
175 * Types of charges to be moved.
177 #define MOVE_ANON 0x1U
178 #define MOVE_FILE 0x2U
179 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
181 /* "mc" and its members are protected by cgroup_mutex */
182 static struct move_charge_struct
{
183 spinlock_t lock
; /* for from, to */
184 struct mm_struct
*mm
;
185 struct mem_cgroup
*from
;
186 struct mem_cgroup
*to
;
188 unsigned long precharge
;
189 unsigned long moved_charge
;
190 unsigned long moved_swap
;
191 struct task_struct
*moving_task
; /* a task moving charges */
192 wait_queue_head_t waitq
; /* a waitq for other context */
194 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
195 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
199 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
200 * limit reclaim to prevent infinite loops, if they ever occur.
202 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
203 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
207 MEM_CGROUP_CHARGE_TYPE_ANON
,
208 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
209 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
213 /* for encoding cft->private value on file */
222 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
223 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
224 #define MEMFILE_ATTR(val) ((val) & 0xffff)
225 /* Used for OOM nofiier */
226 #define OOM_CONTROL (0)
229 * Iteration constructs for visiting all cgroups (under a tree). If
230 * loops are exited prematurely (break), mem_cgroup_iter_break() must
231 * be used for reference counting.
233 #define for_each_mem_cgroup_tree(iter, root) \
234 for (iter = mem_cgroup_iter(root, NULL, NULL); \
236 iter = mem_cgroup_iter(root, iter, NULL))
238 #define for_each_mem_cgroup(iter) \
239 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
241 iter = mem_cgroup_iter(NULL, iter, NULL))
243 static inline bool should_force_charge(void)
245 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
246 (current
->flags
& PF_EXITING
);
249 /* Some nice accessors for the vmpressure. */
250 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
253 memcg
= root_mem_cgroup
;
254 return &memcg
->vmpressure
;
257 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
259 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
262 #ifdef CONFIG_MEMCG_KMEM
264 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
265 * The main reason for not using cgroup id for this:
266 * this works better in sparse environments, where we have a lot of memcgs,
267 * but only a few kmem-limited. Or also, if we have, for instance, 200
268 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
269 * 200 entry array for that.
271 * The current size of the caches array is stored in memcg_nr_cache_ids. It
272 * will double each time we have to increase it.
274 static DEFINE_IDA(memcg_cache_ida
);
275 int memcg_nr_cache_ids
;
277 /* Protects memcg_nr_cache_ids */
278 static DECLARE_RWSEM(memcg_cache_ids_sem
);
280 void memcg_get_cache_ids(void)
282 down_read(&memcg_cache_ids_sem
);
285 void memcg_put_cache_ids(void)
287 up_read(&memcg_cache_ids_sem
);
291 * MIN_SIZE is different than 1, because we would like to avoid going through
292 * the alloc/free process all the time. In a small machine, 4 kmem-limited
293 * cgroups is a reasonable guess. In the future, it could be a parameter or
294 * tunable, but that is strictly not necessary.
296 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
297 * this constant directly from cgroup, but it is understandable that this is
298 * better kept as an internal representation in cgroup.c. In any case, the
299 * cgrp_id space is not getting any smaller, and we don't have to necessarily
300 * increase ours as well if it increases.
302 #define MEMCG_CACHES_MIN_SIZE 4
303 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
306 * A lot of the calls to the cache allocation functions are expected to be
307 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
308 * conditional to this static branch, we'll have to allow modules that does
309 * kmem_cache_alloc and the such to see this symbol as well
311 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
312 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
314 struct workqueue_struct
*memcg_kmem_cache_wq
;
316 static int memcg_shrinker_map_size
;
317 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
319 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
321 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
324 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
325 int size
, int old_size
)
327 struct memcg_shrinker_map
*new, *old
;
330 lockdep_assert_held(&memcg_shrinker_map_mutex
);
333 old
= rcu_dereference_protected(
334 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
335 /* Not yet online memcg */
339 new = kvmalloc(sizeof(*new) + size
, GFP_KERNEL
);
343 /* Set all old bits, clear all new bits */
344 memset(new->map
, (int)0xff, old_size
);
345 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
347 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
348 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
354 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
356 struct mem_cgroup_per_node
*pn
;
357 struct memcg_shrinker_map
*map
;
360 if (mem_cgroup_is_root(memcg
))
364 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
365 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
368 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
372 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
374 struct memcg_shrinker_map
*map
;
375 int nid
, size
, ret
= 0;
377 if (mem_cgroup_is_root(memcg
))
380 mutex_lock(&memcg_shrinker_map_mutex
);
381 size
= memcg_shrinker_map_size
;
383 map
= kvzalloc(sizeof(*map
) + size
, GFP_KERNEL
);
385 memcg_free_shrinker_maps(memcg
);
389 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
391 mutex_unlock(&memcg_shrinker_map_mutex
);
396 int memcg_expand_shrinker_maps(int new_id
)
398 int size
, old_size
, ret
= 0;
399 struct mem_cgroup
*memcg
;
401 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
402 old_size
= memcg_shrinker_map_size
;
403 if (size
<= old_size
)
406 mutex_lock(&memcg_shrinker_map_mutex
);
407 if (!root_mem_cgroup
)
410 for_each_mem_cgroup(memcg
) {
411 if (mem_cgroup_is_root(memcg
))
413 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
419 memcg_shrinker_map_size
= size
;
420 mutex_unlock(&memcg_shrinker_map_mutex
);
424 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
426 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
427 struct memcg_shrinker_map
*map
;
430 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id
, map
->map
);
438 #else /* CONFIG_MEMCG_KMEM */
439 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
443 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
) { }
444 #endif /* CONFIG_MEMCG_KMEM */
447 * mem_cgroup_css_from_page - css of the memcg associated with a page
448 * @page: page of interest
450 * If memcg is bound to the default hierarchy, css of the memcg associated
451 * with @page is returned. The returned css remains associated with @page
452 * until it is released.
454 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
457 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
459 struct mem_cgroup
*memcg
;
461 memcg
= page
->mem_cgroup
;
463 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
464 memcg
= root_mem_cgroup
;
470 * page_cgroup_ino - return inode number of the memcg a page is charged to
473 * Look up the closest online ancestor of the memory cgroup @page is charged to
474 * and return its inode number or 0 if @page is not charged to any cgroup. It
475 * is safe to call this function without holding a reference to @page.
477 * Note, this function is inherently racy, because there is nothing to prevent
478 * the cgroup inode from getting torn down and potentially reallocated a moment
479 * after page_cgroup_ino() returns, so it only should be used by callers that
480 * do not care (such as procfs interfaces).
482 ino_t
page_cgroup_ino(struct page
*page
)
484 struct mem_cgroup
*memcg
;
485 unsigned long ino
= 0;
488 memcg
= READ_ONCE(page
->mem_cgroup
);
489 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
490 memcg
= parent_mem_cgroup(memcg
);
492 ino
= cgroup_ino(memcg
->css
.cgroup
);
497 static struct mem_cgroup_per_node
*
498 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
500 int nid
= page_to_nid(page
);
502 return memcg
->nodeinfo
[nid
];
505 static struct mem_cgroup_tree_per_node
*
506 soft_limit_tree_node(int nid
)
508 return soft_limit_tree
.rb_tree_per_node
[nid
];
511 static struct mem_cgroup_tree_per_node
*
512 soft_limit_tree_from_page(struct page
*page
)
514 int nid
= page_to_nid(page
);
516 return soft_limit_tree
.rb_tree_per_node
[nid
];
519 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
520 struct mem_cgroup_tree_per_node
*mctz
,
521 unsigned long new_usage_in_excess
)
523 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
524 struct rb_node
*parent
= NULL
;
525 struct mem_cgroup_per_node
*mz_node
;
526 bool rightmost
= true;
531 mz
->usage_in_excess
= new_usage_in_excess
;
532 if (!mz
->usage_in_excess
)
536 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
538 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
544 * We can't avoid mem cgroups that are over their soft
545 * limit by the same amount
547 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
552 mctz
->rb_rightmost
= &mz
->tree_node
;
554 rb_link_node(&mz
->tree_node
, parent
, p
);
555 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
559 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
560 struct mem_cgroup_tree_per_node
*mctz
)
565 if (&mz
->tree_node
== mctz
->rb_rightmost
)
566 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
568 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
572 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
573 struct mem_cgroup_tree_per_node
*mctz
)
577 spin_lock_irqsave(&mctz
->lock
, flags
);
578 __mem_cgroup_remove_exceeded(mz
, mctz
);
579 spin_unlock_irqrestore(&mctz
->lock
, flags
);
582 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
584 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
585 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
586 unsigned long excess
= 0;
588 if (nr_pages
> soft_limit
)
589 excess
= nr_pages
- soft_limit
;
594 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
596 unsigned long excess
;
597 struct mem_cgroup_per_node
*mz
;
598 struct mem_cgroup_tree_per_node
*mctz
;
600 mctz
= soft_limit_tree_from_page(page
);
604 * Necessary to update all ancestors when hierarchy is used.
605 * because their event counter is not touched.
607 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
608 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
609 excess
= soft_limit_excess(memcg
);
611 * We have to update the tree if mz is on RB-tree or
612 * mem is over its softlimit.
614 if (excess
|| mz
->on_tree
) {
617 spin_lock_irqsave(&mctz
->lock
, flags
);
618 /* if on-tree, remove it */
620 __mem_cgroup_remove_exceeded(mz
, mctz
);
622 * Insert again. mz->usage_in_excess will be updated.
623 * If excess is 0, no tree ops.
625 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
626 spin_unlock_irqrestore(&mctz
->lock
, flags
);
631 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
633 struct mem_cgroup_tree_per_node
*mctz
;
634 struct mem_cgroup_per_node
*mz
;
638 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
639 mctz
= soft_limit_tree_node(nid
);
641 mem_cgroup_remove_exceeded(mz
, mctz
);
645 static struct mem_cgroup_per_node
*
646 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
648 struct mem_cgroup_per_node
*mz
;
652 if (!mctz
->rb_rightmost
)
653 goto done
; /* Nothing to reclaim from */
655 mz
= rb_entry(mctz
->rb_rightmost
,
656 struct mem_cgroup_per_node
, tree_node
);
658 * Remove the node now but someone else can add it back,
659 * we will to add it back at the end of reclaim to its correct
660 * position in the tree.
662 __mem_cgroup_remove_exceeded(mz
, mctz
);
663 if (!soft_limit_excess(mz
->memcg
) ||
664 !css_tryget_online(&mz
->memcg
->css
))
670 static struct mem_cgroup_per_node
*
671 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
673 struct mem_cgroup_per_node
*mz
;
675 spin_lock_irq(&mctz
->lock
);
676 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
677 spin_unlock_irq(&mctz
->lock
);
682 * __mod_memcg_state - update cgroup memory statistics
683 * @memcg: the memory cgroup
684 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
685 * @val: delta to add to the counter, can be negative
687 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
691 if (mem_cgroup_disabled())
694 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
695 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
696 struct mem_cgroup
*mi
;
698 atomic_long_add(x
, &memcg
->vmstats_local
[idx
]);
699 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
700 atomic_long_add(x
, &mi
->vmstats
[idx
]);
703 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
706 static struct mem_cgroup_per_node
*
707 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
709 struct mem_cgroup
*parent
;
711 parent
= parent_mem_cgroup(pn
->memcg
);
714 return mem_cgroup_nodeinfo(parent
, nid
);
718 * __mod_lruvec_state - update lruvec memory statistics
719 * @lruvec: the lruvec
720 * @idx: the stat item
721 * @val: delta to add to the counter, can be negative
723 * The lruvec is the intersection of the NUMA node and a cgroup. This
724 * function updates the all three counters that are affected by a
725 * change of state at this level: per-node, per-cgroup, per-lruvec.
727 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
730 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
731 struct mem_cgroup_per_node
*pn
;
732 struct mem_cgroup
*memcg
;
736 __mod_node_page_state(pgdat
, idx
, val
);
738 if (mem_cgroup_disabled())
741 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
745 __mod_memcg_state(memcg
, idx
, val
);
748 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
749 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
750 struct mem_cgroup_per_node
*pi
;
752 atomic_long_add(x
, &pn
->lruvec_stat_local
[idx
]);
753 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
754 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
757 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
761 * __count_memcg_events - account VM events in a cgroup
762 * @memcg: the memory cgroup
763 * @idx: the event item
764 * @count: the number of events that occured
766 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
771 if (mem_cgroup_disabled())
774 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
775 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
776 struct mem_cgroup
*mi
;
778 atomic_long_add(x
, &memcg
->vmevents_local
[idx
]);
779 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
780 atomic_long_add(x
, &mi
->vmevents
[idx
]);
783 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
786 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
788 return atomic_long_read(&memcg
->vmevents
[event
]);
791 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
793 return atomic_long_read(&memcg
->vmevents_local
[event
]);
796 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
798 bool compound
, int nr_pages
)
801 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
802 * counted as CACHE even if it's on ANON LRU.
805 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
807 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
808 if (PageSwapBacked(page
))
809 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
813 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
814 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
817 /* pagein of a big page is an event. So, ignore page size */
819 __count_memcg_events(memcg
, PGPGIN
, 1);
821 __count_memcg_events(memcg
, PGPGOUT
, 1);
822 nr_pages
= -nr_pages
; /* for event */
825 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
828 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
829 enum mem_cgroup_events_target target
)
831 unsigned long val
, next
;
833 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
834 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
835 /* from time_after() in jiffies.h */
836 if ((long)(next
- val
) < 0) {
838 case MEM_CGROUP_TARGET_THRESH
:
839 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
841 case MEM_CGROUP_TARGET_SOFTLIMIT
:
842 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
844 case MEM_CGROUP_TARGET_NUMAINFO
:
845 next
= val
+ NUMAINFO_EVENTS_TARGET
;
850 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
857 * Check events in order.
860 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
862 /* threshold event is triggered in finer grain than soft limit */
863 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
864 MEM_CGROUP_TARGET_THRESH
))) {
866 bool do_numainfo __maybe_unused
;
868 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
869 MEM_CGROUP_TARGET_SOFTLIMIT
);
871 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
872 MEM_CGROUP_TARGET_NUMAINFO
);
874 mem_cgroup_threshold(memcg
);
875 if (unlikely(do_softlimit
))
876 mem_cgroup_update_tree(memcg
, page
);
878 if (unlikely(do_numainfo
))
879 atomic_inc(&memcg
->numainfo_events
);
884 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
887 * mm_update_next_owner() may clear mm->owner to NULL
888 * if it races with swapoff, page migration, etc.
889 * So this can be called with p == NULL.
894 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
896 EXPORT_SYMBOL(mem_cgroup_from_task
);
899 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
900 * @mm: mm from which memcg should be extracted. It can be NULL.
902 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
903 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
906 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
908 struct mem_cgroup
*memcg
;
910 if (mem_cgroup_disabled())
916 * Page cache insertions can happen withou an
917 * actual mm context, e.g. during disk probing
918 * on boot, loopback IO, acct() writes etc.
921 memcg
= root_mem_cgroup
;
923 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
924 if (unlikely(!memcg
))
925 memcg
= root_mem_cgroup
;
927 } while (!css_tryget_online(&memcg
->css
));
931 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
934 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
935 * @page: page from which memcg should be extracted.
937 * Obtain a reference on page->memcg and returns it if successful. Otherwise
938 * root_mem_cgroup is returned.
940 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
942 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
944 if (mem_cgroup_disabled())
948 if (!memcg
|| !css_tryget_online(&memcg
->css
))
949 memcg
= root_mem_cgroup
;
953 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
956 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
958 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
960 if (unlikely(current
->active_memcg
)) {
961 struct mem_cgroup
*memcg
= root_mem_cgroup
;
964 if (css_tryget_online(¤t
->active_memcg
->css
))
965 memcg
= current
->active_memcg
;
969 return get_mem_cgroup_from_mm(current
->mm
);
973 * mem_cgroup_iter - iterate over memory cgroup hierarchy
974 * @root: hierarchy root
975 * @prev: previously returned memcg, NULL on first invocation
976 * @reclaim: cookie for shared reclaim walks, NULL for full walks
978 * Returns references to children of the hierarchy below @root, or
979 * @root itself, or %NULL after a full round-trip.
981 * Caller must pass the return value in @prev on subsequent
982 * invocations for reference counting, or use mem_cgroup_iter_break()
983 * to cancel a hierarchy walk before the round-trip is complete.
985 * Reclaimers can specify a node and a priority level in @reclaim to
986 * divide up the memcgs in the hierarchy among all concurrent
987 * reclaimers operating on the same node and priority.
989 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
990 struct mem_cgroup
*prev
,
991 struct mem_cgroup_reclaim_cookie
*reclaim
)
993 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
994 struct cgroup_subsys_state
*css
= NULL
;
995 struct mem_cgroup
*memcg
= NULL
;
996 struct mem_cgroup
*pos
= NULL
;
998 if (mem_cgroup_disabled())
1002 root
= root_mem_cgroup
;
1004 if (prev
&& !reclaim
)
1007 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1016 struct mem_cgroup_per_node
*mz
;
1018 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1019 iter
= &mz
->iter
[reclaim
->priority
];
1021 if (prev
&& reclaim
->generation
!= iter
->generation
)
1025 pos
= READ_ONCE(iter
->position
);
1026 if (!pos
|| css_tryget(&pos
->css
))
1029 * css reference reached zero, so iter->position will
1030 * be cleared by ->css_released. However, we should not
1031 * rely on this happening soon, because ->css_released
1032 * is called from a work queue, and by busy-waiting we
1033 * might block it. So we clear iter->position right
1036 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1044 css
= css_next_descendant_pre(css
, &root
->css
);
1047 * Reclaimers share the hierarchy walk, and a
1048 * new one might jump in right at the end of
1049 * the hierarchy - make sure they see at least
1050 * one group and restart from the beginning.
1058 * Verify the css and acquire a reference. The root
1059 * is provided by the caller, so we know it's alive
1060 * and kicking, and don't take an extra reference.
1062 memcg
= mem_cgroup_from_css(css
);
1064 if (css
== &root
->css
)
1067 if (css_tryget(css
))
1075 * The position could have already been updated by a competing
1076 * thread, so check that the value hasn't changed since we read
1077 * it to avoid reclaiming from the same cgroup twice.
1079 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1087 reclaim
->generation
= iter
->generation
;
1093 if (prev
&& prev
!= root
)
1094 css_put(&prev
->css
);
1100 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1101 * @root: hierarchy root
1102 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1104 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1105 struct mem_cgroup
*prev
)
1108 root
= root_mem_cgroup
;
1109 if (prev
&& prev
!= root
)
1110 css_put(&prev
->css
);
1113 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1115 struct mem_cgroup
*memcg
= dead_memcg
;
1116 struct mem_cgroup_reclaim_iter
*iter
;
1117 struct mem_cgroup_per_node
*mz
;
1121 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1122 for_each_node(nid
) {
1123 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
1124 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
1125 iter
= &mz
->iter
[i
];
1126 cmpxchg(&iter
->position
,
1134 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1135 * @memcg: hierarchy root
1136 * @fn: function to call for each task
1137 * @arg: argument passed to @fn
1139 * This function iterates over tasks attached to @memcg or to any of its
1140 * descendants and calls @fn for each task. If @fn returns a non-zero
1141 * value, the function breaks the iteration loop and returns the value.
1142 * Otherwise, it will iterate over all tasks and return 0.
1144 * This function must not be called for the root memory cgroup.
1146 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1147 int (*fn
)(struct task_struct
*, void *), void *arg
)
1149 struct mem_cgroup
*iter
;
1152 BUG_ON(memcg
== root_mem_cgroup
);
1154 for_each_mem_cgroup_tree(iter
, memcg
) {
1155 struct css_task_iter it
;
1156 struct task_struct
*task
;
1158 css_task_iter_start(&iter
->css
, 0, &it
);
1159 while (!ret
&& (task
= css_task_iter_next(&it
)))
1160 ret
= fn(task
, arg
);
1161 css_task_iter_end(&it
);
1163 mem_cgroup_iter_break(memcg
, iter
);
1171 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1173 * @pgdat: pgdat of the page
1175 * This function is only safe when following the LRU page isolation
1176 * and putback protocol: the LRU lock must be held, and the page must
1177 * either be PageLRU() or the caller must have isolated/allocated it.
1179 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1181 struct mem_cgroup_per_node
*mz
;
1182 struct mem_cgroup
*memcg
;
1183 struct lruvec
*lruvec
;
1185 if (mem_cgroup_disabled()) {
1186 lruvec
= &pgdat
->lruvec
;
1190 memcg
= page
->mem_cgroup
;
1192 * Swapcache readahead pages are added to the LRU - and
1193 * possibly migrated - before they are charged.
1196 memcg
= root_mem_cgroup
;
1198 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1199 lruvec
= &mz
->lruvec
;
1202 * Since a node can be onlined after the mem_cgroup was created,
1203 * we have to be prepared to initialize lruvec->zone here;
1204 * and if offlined then reonlined, we need to reinitialize it.
1206 if (unlikely(lruvec
->pgdat
!= pgdat
))
1207 lruvec
->pgdat
= pgdat
;
1212 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1213 * @lruvec: mem_cgroup per zone lru vector
1214 * @lru: index of lru list the page is sitting on
1215 * @zid: zone id of the accounted pages
1216 * @nr_pages: positive when adding or negative when removing
1218 * This function must be called under lru_lock, just before a page is added
1219 * to or just after a page is removed from an lru list (that ordering being
1220 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1222 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1223 int zid
, int nr_pages
)
1225 struct mem_cgroup_per_node
*mz
;
1226 unsigned long *lru_size
;
1229 if (mem_cgroup_disabled())
1232 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1233 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1236 *lru_size
+= nr_pages
;
1239 if (WARN_ONCE(size
< 0,
1240 "%s(%p, %d, %d): lru_size %ld\n",
1241 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1247 *lru_size
+= nr_pages
;
1250 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1252 struct mem_cgroup
*task_memcg
;
1253 struct task_struct
*p
;
1256 p
= find_lock_task_mm(task
);
1258 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1262 * All threads may have already detached their mm's, but the oom
1263 * killer still needs to detect if they have already been oom
1264 * killed to prevent needlessly killing additional tasks.
1267 task_memcg
= mem_cgroup_from_task(task
);
1268 css_get(&task_memcg
->css
);
1271 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1272 css_put(&task_memcg
->css
);
1277 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1278 * @memcg: the memory cgroup
1280 * Returns the maximum amount of memory @mem can be charged with, in
1283 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1285 unsigned long margin
= 0;
1286 unsigned long count
;
1287 unsigned long limit
;
1289 count
= page_counter_read(&memcg
->memory
);
1290 limit
= READ_ONCE(memcg
->memory
.max
);
1292 margin
= limit
- count
;
1294 if (do_memsw_account()) {
1295 count
= page_counter_read(&memcg
->memsw
);
1296 limit
= READ_ONCE(memcg
->memsw
.max
);
1298 margin
= min(margin
, limit
- count
);
1307 * A routine for checking "mem" is under move_account() or not.
1309 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1310 * moving cgroups. This is for waiting at high-memory pressure
1313 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1315 struct mem_cgroup
*from
;
1316 struct mem_cgroup
*to
;
1319 * Unlike task_move routines, we access mc.to, mc.from not under
1320 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1322 spin_lock(&mc
.lock
);
1328 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1329 mem_cgroup_is_descendant(to
, memcg
);
1331 spin_unlock(&mc
.lock
);
1335 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1337 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1338 if (mem_cgroup_under_move(memcg
)) {
1340 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1341 /* moving charge context might have finished. */
1344 finish_wait(&mc
.waitq
, &wait
);
1351 static const unsigned int memcg1_stats
[] = {
1362 static const char *const memcg1_stat_names
[] = {
1373 #define K(x) ((x) << (PAGE_SHIFT-10))
1375 * mem_cgroup_print_oom_context: Print OOM information relevant to
1376 * memory controller.
1377 * @memcg: The memory cgroup that went over limit
1378 * @p: Task that is going to be killed
1380 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1383 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1388 pr_cont(",oom_memcg=");
1389 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1391 pr_cont(",global_oom");
1393 pr_cont(",task_memcg=");
1394 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1400 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1401 * memory controller.
1402 * @memcg: The memory cgroup that went over limit
1404 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1406 struct mem_cgroup
*iter
;
1409 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1410 K((u64
)page_counter_read(&memcg
->memory
)),
1411 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1412 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1413 K((u64
)page_counter_read(&memcg
->memsw
)),
1414 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1415 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1416 K((u64
)page_counter_read(&memcg
->kmem
)),
1417 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1419 for_each_mem_cgroup_tree(iter
, memcg
) {
1420 pr_info("Memory cgroup stats for ");
1421 pr_cont_cgroup_path(iter
->css
.cgroup
);
1424 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
1425 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_swap_account
)
1427 pr_cont(" %s:%luKB", memcg1_stat_names
[i
],
1428 K(memcg_page_state_local(iter
,
1432 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1433 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1434 K(memcg_page_state_local(iter
,
1442 * Return the memory (and swap, if configured) limit for a memcg.
1444 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1448 max
= memcg
->memory
.max
;
1449 if (mem_cgroup_swappiness(memcg
)) {
1450 unsigned long memsw_max
;
1451 unsigned long swap_max
;
1453 memsw_max
= memcg
->memsw
.max
;
1454 swap_max
= memcg
->swap
.max
;
1455 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1456 max
= min(max
+ swap_max
, memsw_max
);
1461 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1464 struct oom_control oc
= {
1468 .gfp_mask
= gfp_mask
,
1473 if (mutex_lock_killable(&oom_lock
))
1476 * A few threads which were not waiting at mutex_lock_killable() can
1477 * fail to bail out. Therefore, check again after holding oom_lock.
1479 ret
= should_force_charge() || out_of_memory(&oc
);
1480 mutex_unlock(&oom_lock
);
1484 #if MAX_NUMNODES > 1
1487 * test_mem_cgroup_node_reclaimable
1488 * @memcg: the target memcg
1489 * @nid: the node ID to be checked.
1490 * @noswap : specify true here if the user wants flle only information.
1492 * This function returns whether the specified memcg contains any
1493 * reclaimable pages on a node. Returns true if there are any reclaimable
1494 * pages in the node.
1496 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1497 int nid
, bool noswap
)
1499 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
1501 if (lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) ||
1502 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
))
1504 if (noswap
|| !total_swap_pages
)
1506 if (lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) ||
1507 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
))
1514 * Always updating the nodemask is not very good - even if we have an empty
1515 * list or the wrong list here, we can start from some node and traverse all
1516 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1519 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1523 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1524 * pagein/pageout changes since the last update.
1526 if (!atomic_read(&memcg
->numainfo_events
))
1528 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1531 /* make a nodemask where this memcg uses memory from */
1532 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1534 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1536 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1537 node_clear(nid
, memcg
->scan_nodes
);
1540 atomic_set(&memcg
->numainfo_events
, 0);
1541 atomic_set(&memcg
->numainfo_updating
, 0);
1545 * Selecting a node where we start reclaim from. Because what we need is just
1546 * reducing usage counter, start from anywhere is O,K. Considering
1547 * memory reclaim from current node, there are pros. and cons.
1549 * Freeing memory from current node means freeing memory from a node which
1550 * we'll use or we've used. So, it may make LRU bad. And if several threads
1551 * hit limits, it will see a contention on a node. But freeing from remote
1552 * node means more costs for memory reclaim because of memory latency.
1554 * Now, we use round-robin. Better algorithm is welcomed.
1556 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1560 mem_cgroup_may_update_nodemask(memcg
);
1561 node
= memcg
->last_scanned_node
;
1563 node
= next_node_in(node
, memcg
->scan_nodes
);
1565 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1566 * last time it really checked all the LRUs due to rate limiting.
1567 * Fallback to the current node in that case for simplicity.
1569 if (unlikely(node
== MAX_NUMNODES
))
1570 node
= numa_node_id();
1572 memcg
->last_scanned_node
= node
;
1576 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1582 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1585 unsigned long *total_scanned
)
1587 struct mem_cgroup
*victim
= NULL
;
1590 unsigned long excess
;
1591 unsigned long nr_scanned
;
1592 struct mem_cgroup_reclaim_cookie reclaim
= {
1597 excess
= soft_limit_excess(root_memcg
);
1600 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1605 * If we have not been able to reclaim
1606 * anything, it might because there are
1607 * no reclaimable pages under this hierarchy
1612 * We want to do more targeted reclaim.
1613 * excess >> 2 is not to excessive so as to
1614 * reclaim too much, nor too less that we keep
1615 * coming back to reclaim from this cgroup
1617 if (total
>= (excess
>> 2) ||
1618 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1623 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1624 pgdat
, &nr_scanned
);
1625 *total_scanned
+= nr_scanned
;
1626 if (!soft_limit_excess(root_memcg
))
1629 mem_cgroup_iter_break(root_memcg
, victim
);
1633 #ifdef CONFIG_LOCKDEP
1634 static struct lockdep_map memcg_oom_lock_dep_map
= {
1635 .name
= "memcg_oom_lock",
1639 static DEFINE_SPINLOCK(memcg_oom_lock
);
1642 * Check OOM-Killer is already running under our hierarchy.
1643 * If someone is running, return false.
1645 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1647 struct mem_cgroup
*iter
, *failed
= NULL
;
1649 spin_lock(&memcg_oom_lock
);
1651 for_each_mem_cgroup_tree(iter
, memcg
) {
1652 if (iter
->oom_lock
) {
1654 * this subtree of our hierarchy is already locked
1655 * so we cannot give a lock.
1658 mem_cgroup_iter_break(memcg
, iter
);
1661 iter
->oom_lock
= true;
1666 * OK, we failed to lock the whole subtree so we have
1667 * to clean up what we set up to the failing subtree
1669 for_each_mem_cgroup_tree(iter
, memcg
) {
1670 if (iter
== failed
) {
1671 mem_cgroup_iter_break(memcg
, iter
);
1674 iter
->oom_lock
= false;
1677 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1679 spin_unlock(&memcg_oom_lock
);
1684 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1686 struct mem_cgroup
*iter
;
1688 spin_lock(&memcg_oom_lock
);
1689 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1690 for_each_mem_cgroup_tree(iter
, memcg
)
1691 iter
->oom_lock
= false;
1692 spin_unlock(&memcg_oom_lock
);
1695 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1697 struct mem_cgroup
*iter
;
1699 spin_lock(&memcg_oom_lock
);
1700 for_each_mem_cgroup_tree(iter
, memcg
)
1702 spin_unlock(&memcg_oom_lock
);
1705 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1707 struct mem_cgroup
*iter
;
1710 * When a new child is created while the hierarchy is under oom,
1711 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1713 spin_lock(&memcg_oom_lock
);
1714 for_each_mem_cgroup_tree(iter
, memcg
)
1715 if (iter
->under_oom
> 0)
1717 spin_unlock(&memcg_oom_lock
);
1720 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1722 struct oom_wait_info
{
1723 struct mem_cgroup
*memcg
;
1724 wait_queue_entry_t wait
;
1727 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1728 unsigned mode
, int sync
, void *arg
)
1730 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1731 struct mem_cgroup
*oom_wait_memcg
;
1732 struct oom_wait_info
*oom_wait_info
;
1734 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1735 oom_wait_memcg
= oom_wait_info
->memcg
;
1737 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1738 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1740 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1743 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1746 * For the following lockless ->under_oom test, the only required
1747 * guarantee is that it must see the state asserted by an OOM when
1748 * this function is called as a result of userland actions
1749 * triggered by the notification of the OOM. This is trivially
1750 * achieved by invoking mem_cgroup_mark_under_oom() before
1751 * triggering notification.
1753 if (memcg
&& memcg
->under_oom
)
1754 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1764 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1766 enum oom_status ret
;
1769 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1772 memcg_memory_event(memcg
, MEMCG_OOM
);
1775 * We are in the middle of the charge context here, so we
1776 * don't want to block when potentially sitting on a callstack
1777 * that holds all kinds of filesystem and mm locks.
1779 * cgroup1 allows disabling the OOM killer and waiting for outside
1780 * handling until the charge can succeed; remember the context and put
1781 * the task to sleep at the end of the page fault when all locks are
1784 * On the other hand, in-kernel OOM killer allows for an async victim
1785 * memory reclaim (oom_reaper) and that means that we are not solely
1786 * relying on the oom victim to make a forward progress and we can
1787 * invoke the oom killer here.
1789 * Please note that mem_cgroup_out_of_memory might fail to find a
1790 * victim and then we have to bail out from the charge path.
1792 if (memcg
->oom_kill_disable
) {
1793 if (!current
->in_user_fault
)
1795 css_get(&memcg
->css
);
1796 current
->memcg_in_oom
= memcg
;
1797 current
->memcg_oom_gfp_mask
= mask
;
1798 current
->memcg_oom_order
= order
;
1803 mem_cgroup_mark_under_oom(memcg
);
1805 locked
= mem_cgroup_oom_trylock(memcg
);
1808 mem_cgroup_oom_notify(memcg
);
1810 mem_cgroup_unmark_under_oom(memcg
);
1811 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1817 mem_cgroup_oom_unlock(memcg
);
1823 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1824 * @handle: actually kill/wait or just clean up the OOM state
1826 * This has to be called at the end of a page fault if the memcg OOM
1827 * handler was enabled.
1829 * Memcg supports userspace OOM handling where failed allocations must
1830 * sleep on a waitqueue until the userspace task resolves the
1831 * situation. Sleeping directly in the charge context with all kinds
1832 * of locks held is not a good idea, instead we remember an OOM state
1833 * in the task and mem_cgroup_oom_synchronize() has to be called at
1834 * the end of the page fault to complete the OOM handling.
1836 * Returns %true if an ongoing memcg OOM situation was detected and
1837 * completed, %false otherwise.
1839 bool mem_cgroup_oom_synchronize(bool handle
)
1841 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1842 struct oom_wait_info owait
;
1845 /* OOM is global, do not handle */
1852 owait
.memcg
= memcg
;
1853 owait
.wait
.flags
= 0;
1854 owait
.wait
.func
= memcg_oom_wake_function
;
1855 owait
.wait
.private = current
;
1856 INIT_LIST_HEAD(&owait
.wait
.entry
);
1858 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1859 mem_cgroup_mark_under_oom(memcg
);
1861 locked
= mem_cgroup_oom_trylock(memcg
);
1864 mem_cgroup_oom_notify(memcg
);
1866 if (locked
&& !memcg
->oom_kill_disable
) {
1867 mem_cgroup_unmark_under_oom(memcg
);
1868 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1869 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1870 current
->memcg_oom_order
);
1873 mem_cgroup_unmark_under_oom(memcg
);
1874 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1878 mem_cgroup_oom_unlock(memcg
);
1880 * There is no guarantee that an OOM-lock contender
1881 * sees the wakeups triggered by the OOM kill
1882 * uncharges. Wake any sleepers explicitely.
1884 memcg_oom_recover(memcg
);
1887 current
->memcg_in_oom
= NULL
;
1888 css_put(&memcg
->css
);
1893 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1894 * @victim: task to be killed by the OOM killer
1895 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1897 * Returns a pointer to a memory cgroup, which has to be cleaned up
1898 * by killing all belonging OOM-killable tasks.
1900 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1902 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1903 struct mem_cgroup
*oom_domain
)
1905 struct mem_cgroup
*oom_group
= NULL
;
1906 struct mem_cgroup
*memcg
;
1908 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1912 oom_domain
= root_mem_cgroup
;
1916 memcg
= mem_cgroup_from_task(victim
);
1917 if (memcg
== root_mem_cgroup
)
1921 * Traverse the memory cgroup hierarchy from the victim task's
1922 * cgroup up to the OOMing cgroup (or root) to find the
1923 * highest-level memory cgroup with oom.group set.
1925 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1926 if (memcg
->oom_group
)
1929 if (memcg
== oom_domain
)
1934 css_get(&oom_group
->css
);
1941 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1943 pr_info("Tasks in ");
1944 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1945 pr_cont(" are going to be killed due to memory.oom.group set\n");
1949 * lock_page_memcg - lock a page->mem_cgroup binding
1952 * This function protects unlocked LRU pages from being moved to
1955 * It ensures lifetime of the returned memcg. Caller is responsible
1956 * for the lifetime of the page; __unlock_page_memcg() is available
1957 * when @page might get freed inside the locked section.
1959 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1961 struct mem_cgroup
*memcg
;
1962 unsigned long flags
;
1965 * The RCU lock is held throughout the transaction. The fast
1966 * path can get away without acquiring the memcg->move_lock
1967 * because page moving starts with an RCU grace period.
1969 * The RCU lock also protects the memcg from being freed when
1970 * the page state that is going to change is the only thing
1971 * preventing the page itself from being freed. E.g. writeback
1972 * doesn't hold a page reference and relies on PG_writeback to
1973 * keep off truncation, migration and so forth.
1977 if (mem_cgroup_disabled())
1980 memcg
= page
->mem_cgroup
;
1981 if (unlikely(!memcg
))
1984 if (atomic_read(&memcg
->moving_account
) <= 0)
1987 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1988 if (memcg
!= page
->mem_cgroup
) {
1989 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1994 * When charge migration first begins, we can have locked and
1995 * unlocked page stat updates happening concurrently. Track
1996 * the task who has the lock for unlock_page_memcg().
1998 memcg
->move_lock_task
= current
;
1999 memcg
->move_lock_flags
= flags
;
2003 EXPORT_SYMBOL(lock_page_memcg
);
2006 * __unlock_page_memcg - unlock and unpin a memcg
2009 * Unlock and unpin a memcg returned by lock_page_memcg().
2011 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2013 if (memcg
&& memcg
->move_lock_task
== current
) {
2014 unsigned long flags
= memcg
->move_lock_flags
;
2016 memcg
->move_lock_task
= NULL
;
2017 memcg
->move_lock_flags
= 0;
2019 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2026 * unlock_page_memcg - unlock a page->mem_cgroup binding
2029 void unlock_page_memcg(struct page
*page
)
2031 __unlock_page_memcg(page
->mem_cgroup
);
2033 EXPORT_SYMBOL(unlock_page_memcg
);
2035 struct memcg_stock_pcp
{
2036 struct mem_cgroup
*cached
; /* this never be root cgroup */
2037 unsigned int nr_pages
;
2038 struct work_struct work
;
2039 unsigned long flags
;
2040 #define FLUSHING_CACHED_CHARGE 0
2042 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2043 static DEFINE_MUTEX(percpu_charge_mutex
);
2046 * consume_stock: Try to consume stocked charge on this cpu.
2047 * @memcg: memcg to consume from.
2048 * @nr_pages: how many pages to charge.
2050 * The charges will only happen if @memcg matches the current cpu's memcg
2051 * stock, and at least @nr_pages are available in that stock. Failure to
2052 * service an allocation will refill the stock.
2054 * returns true if successful, false otherwise.
2056 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2058 struct memcg_stock_pcp
*stock
;
2059 unsigned long flags
;
2062 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2065 local_irq_save(flags
);
2067 stock
= this_cpu_ptr(&memcg_stock
);
2068 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2069 stock
->nr_pages
-= nr_pages
;
2073 local_irq_restore(flags
);
2079 * Returns stocks cached in percpu and reset cached information.
2081 static void drain_stock(struct memcg_stock_pcp
*stock
)
2083 struct mem_cgroup
*old
= stock
->cached
;
2085 if (stock
->nr_pages
) {
2086 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2087 if (do_memsw_account())
2088 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2089 css_put_many(&old
->css
, stock
->nr_pages
);
2090 stock
->nr_pages
= 0;
2092 stock
->cached
= NULL
;
2095 static void drain_local_stock(struct work_struct
*dummy
)
2097 struct memcg_stock_pcp
*stock
;
2098 unsigned long flags
;
2101 * The only protection from memory hotplug vs. drain_stock races is
2102 * that we always operate on local CPU stock here with IRQ disabled
2104 local_irq_save(flags
);
2106 stock
= this_cpu_ptr(&memcg_stock
);
2108 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2110 local_irq_restore(flags
);
2114 * Cache charges(val) to local per_cpu area.
2115 * This will be consumed by consume_stock() function, later.
2117 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2119 struct memcg_stock_pcp
*stock
;
2120 unsigned long flags
;
2122 local_irq_save(flags
);
2124 stock
= this_cpu_ptr(&memcg_stock
);
2125 if (stock
->cached
!= memcg
) { /* reset if necessary */
2127 stock
->cached
= memcg
;
2129 stock
->nr_pages
+= nr_pages
;
2131 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2134 local_irq_restore(flags
);
2138 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2139 * of the hierarchy under it.
2141 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2145 /* If someone's already draining, avoid adding running more workers. */
2146 if (!mutex_trylock(&percpu_charge_mutex
))
2149 * Notify other cpus that system-wide "drain" is running
2150 * We do not care about races with the cpu hotplug because cpu down
2151 * as well as workers from this path always operate on the local
2152 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2155 for_each_online_cpu(cpu
) {
2156 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2157 struct mem_cgroup
*memcg
;
2159 memcg
= stock
->cached
;
2160 if (!memcg
|| !stock
->nr_pages
|| !css_tryget(&memcg
->css
))
2162 if (!mem_cgroup_is_descendant(memcg
, root_memcg
)) {
2163 css_put(&memcg
->css
);
2166 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2168 drain_local_stock(&stock
->work
);
2170 schedule_work_on(cpu
, &stock
->work
);
2172 css_put(&memcg
->css
);
2175 mutex_unlock(&percpu_charge_mutex
);
2178 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2180 struct memcg_stock_pcp
*stock
;
2181 struct mem_cgroup
*memcg
, *mi
;
2183 stock
= &per_cpu(memcg_stock
, cpu
);
2186 for_each_mem_cgroup(memcg
) {
2189 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2193 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2195 atomic_long_add(x
, &memcg
->vmstats_local
[i
]);
2196 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2197 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2200 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2203 for_each_node(nid
) {
2204 struct mem_cgroup_per_node
*pn
;
2206 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2207 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2209 atomic_long_add(x
, &pn
->lruvec_stat_local
[i
]);
2211 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2212 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2217 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2220 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2222 atomic_long_add(x
, &memcg
->vmevents_local
[i
]);
2223 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2224 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2232 static void reclaim_high(struct mem_cgroup
*memcg
,
2233 unsigned int nr_pages
,
2237 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2239 memcg_memory_event(memcg
, MEMCG_HIGH
);
2240 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2241 } while ((memcg
= parent_mem_cgroup(memcg
)));
2244 static void high_work_func(struct work_struct
*work
)
2246 struct mem_cgroup
*memcg
;
2248 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2249 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2253 * Scheduled by try_charge() to be executed from the userland return path
2254 * and reclaims memory over the high limit.
2256 void mem_cgroup_handle_over_high(void)
2258 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2259 struct mem_cgroup
*memcg
;
2261 if (likely(!nr_pages
))
2264 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2265 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2266 css_put(&memcg
->css
);
2267 current
->memcg_nr_pages_over_high
= 0;
2270 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2271 unsigned int nr_pages
)
2273 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2274 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2275 struct mem_cgroup
*mem_over_limit
;
2276 struct page_counter
*counter
;
2277 unsigned long nr_reclaimed
;
2278 bool may_swap
= true;
2279 bool drained
= false;
2281 enum oom_status oom_status
;
2283 if (mem_cgroup_is_root(memcg
))
2286 if (consume_stock(memcg
, nr_pages
))
2289 if (!do_memsw_account() ||
2290 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2291 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2293 if (do_memsw_account())
2294 page_counter_uncharge(&memcg
->memsw
, batch
);
2295 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2297 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2301 if (batch
> nr_pages
) {
2307 * Unlike in global OOM situations, memcg is not in a physical
2308 * memory shortage. Allow dying and OOM-killed tasks to
2309 * bypass the last charges so that they can exit quickly and
2310 * free their memory.
2312 if (unlikely(should_force_charge()))
2316 * Prevent unbounded recursion when reclaim operations need to
2317 * allocate memory. This might exceed the limits temporarily,
2318 * but we prefer facilitating memory reclaim and getting back
2319 * under the limit over triggering OOM kills in these cases.
2321 if (unlikely(current
->flags
& PF_MEMALLOC
))
2324 if (unlikely(task_in_memcg_oom(current
)))
2327 if (!gfpflags_allow_blocking(gfp_mask
))
2330 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2332 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2333 gfp_mask
, may_swap
);
2335 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2339 drain_all_stock(mem_over_limit
);
2344 if (gfp_mask
& __GFP_NORETRY
)
2347 * Even though the limit is exceeded at this point, reclaim
2348 * may have been able to free some pages. Retry the charge
2349 * before killing the task.
2351 * Only for regular pages, though: huge pages are rather
2352 * unlikely to succeed so close to the limit, and we fall back
2353 * to regular pages anyway in case of failure.
2355 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2358 * At task move, charge accounts can be doubly counted. So, it's
2359 * better to wait until the end of task_move if something is going on.
2361 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2367 if (gfp_mask
& __GFP_RETRY_MAYFAIL
&& oomed
)
2370 if (gfp_mask
& __GFP_NOFAIL
)
2373 if (fatal_signal_pending(current
))
2377 * keep retrying as long as the memcg oom killer is able to make
2378 * a forward progress or bypass the charge if the oom killer
2379 * couldn't make any progress.
2381 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2382 get_order(nr_pages
* PAGE_SIZE
));
2383 switch (oom_status
) {
2385 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2394 if (!(gfp_mask
& __GFP_NOFAIL
))
2398 * The allocation either can't fail or will lead to more memory
2399 * being freed very soon. Allow memory usage go over the limit
2400 * temporarily by force charging it.
2402 page_counter_charge(&memcg
->memory
, nr_pages
);
2403 if (do_memsw_account())
2404 page_counter_charge(&memcg
->memsw
, nr_pages
);
2405 css_get_many(&memcg
->css
, nr_pages
);
2410 css_get_many(&memcg
->css
, batch
);
2411 if (batch
> nr_pages
)
2412 refill_stock(memcg
, batch
- nr_pages
);
2415 * If the hierarchy is above the normal consumption range, schedule
2416 * reclaim on returning to userland. We can perform reclaim here
2417 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2418 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2419 * not recorded as it most likely matches current's and won't
2420 * change in the meantime. As high limit is checked again before
2421 * reclaim, the cost of mismatch is negligible.
2424 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2425 /* Don't bother a random interrupted task */
2426 if (in_interrupt()) {
2427 schedule_work(&memcg
->high_work
);
2430 current
->memcg_nr_pages_over_high
+= batch
;
2431 set_notify_resume(current
);
2434 } while ((memcg
= parent_mem_cgroup(memcg
)));
2439 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2441 if (mem_cgroup_is_root(memcg
))
2444 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2445 if (do_memsw_account())
2446 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2448 css_put_many(&memcg
->css
, nr_pages
);
2451 static void lock_page_lru(struct page
*page
, int *isolated
)
2453 pg_data_t
*pgdat
= page_pgdat(page
);
2455 spin_lock_irq(&pgdat
->lru_lock
);
2456 if (PageLRU(page
)) {
2457 struct lruvec
*lruvec
;
2459 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2461 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2467 static void unlock_page_lru(struct page
*page
, int isolated
)
2469 pg_data_t
*pgdat
= page_pgdat(page
);
2472 struct lruvec
*lruvec
;
2474 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2475 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2477 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2479 spin_unlock_irq(&pgdat
->lru_lock
);
2482 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2487 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2490 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2491 * may already be on some other mem_cgroup's LRU. Take care of it.
2494 lock_page_lru(page
, &isolated
);
2497 * Nobody should be changing or seriously looking at
2498 * page->mem_cgroup at this point:
2500 * - the page is uncharged
2502 * - the page is off-LRU
2504 * - an anonymous fault has exclusive page access, except for
2505 * a locked page table
2507 * - a page cache insertion, a swapin fault, or a migration
2508 * have the page locked
2510 page
->mem_cgroup
= memcg
;
2513 unlock_page_lru(page
, isolated
);
2516 #ifdef CONFIG_MEMCG_KMEM
2517 static int memcg_alloc_cache_id(void)
2522 id
= ida_simple_get(&memcg_cache_ida
,
2523 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2527 if (id
< memcg_nr_cache_ids
)
2531 * There's no space for the new id in memcg_caches arrays,
2532 * so we have to grow them.
2534 down_write(&memcg_cache_ids_sem
);
2536 size
= 2 * (id
+ 1);
2537 if (size
< MEMCG_CACHES_MIN_SIZE
)
2538 size
= MEMCG_CACHES_MIN_SIZE
;
2539 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2540 size
= MEMCG_CACHES_MAX_SIZE
;
2542 err
= memcg_update_all_caches(size
);
2544 err
= memcg_update_all_list_lrus(size
);
2546 memcg_nr_cache_ids
= size
;
2548 up_write(&memcg_cache_ids_sem
);
2551 ida_simple_remove(&memcg_cache_ida
, id
);
2557 static void memcg_free_cache_id(int id
)
2559 ida_simple_remove(&memcg_cache_ida
, id
);
2562 struct memcg_kmem_cache_create_work
{
2563 struct mem_cgroup
*memcg
;
2564 struct kmem_cache
*cachep
;
2565 struct work_struct work
;
2568 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2570 struct memcg_kmem_cache_create_work
*cw
=
2571 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2572 struct mem_cgroup
*memcg
= cw
->memcg
;
2573 struct kmem_cache
*cachep
= cw
->cachep
;
2575 memcg_create_kmem_cache(memcg
, cachep
);
2577 css_put(&memcg
->css
);
2582 * Enqueue the creation of a per-memcg kmem_cache.
2584 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2585 struct kmem_cache
*cachep
)
2587 struct memcg_kmem_cache_create_work
*cw
;
2589 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2593 css_get(&memcg
->css
);
2596 cw
->cachep
= cachep
;
2597 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2599 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2602 static inline bool memcg_kmem_bypass(void)
2604 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2610 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2611 * @cachep: the original global kmem cache
2613 * Return the kmem_cache we're supposed to use for a slab allocation.
2614 * We try to use the current memcg's version of the cache.
2616 * If the cache does not exist yet, if we are the first user of it, we
2617 * create it asynchronously in a workqueue and let the current allocation
2618 * go through with the original cache.
2620 * This function takes a reference to the cache it returns to assure it
2621 * won't get destroyed while we are working with it. Once the caller is
2622 * done with it, memcg_kmem_put_cache() must be called to release the
2625 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2627 struct mem_cgroup
*memcg
;
2628 struct kmem_cache
*memcg_cachep
;
2631 VM_BUG_ON(!is_root_cache(cachep
));
2633 if (memcg_kmem_bypass())
2636 memcg
= get_mem_cgroup_from_current();
2637 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2641 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2642 if (likely(memcg_cachep
))
2643 return memcg_cachep
;
2646 * If we are in a safe context (can wait, and not in interrupt
2647 * context), we could be be predictable and return right away.
2648 * This would guarantee that the allocation being performed
2649 * already belongs in the new cache.
2651 * However, there are some clashes that can arrive from locking.
2652 * For instance, because we acquire the slab_mutex while doing
2653 * memcg_create_kmem_cache, this means no further allocation
2654 * could happen with the slab_mutex held. So it's better to
2657 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2659 css_put(&memcg
->css
);
2664 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2665 * @cachep: the cache returned by memcg_kmem_get_cache
2667 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2669 if (!is_root_cache(cachep
))
2670 css_put(&cachep
->memcg_params
.memcg
->css
);
2674 * __memcg_kmem_charge_memcg: charge a kmem page
2675 * @page: page to charge
2676 * @gfp: reclaim mode
2677 * @order: allocation order
2678 * @memcg: memory cgroup to charge
2680 * Returns 0 on success, an error code on failure.
2682 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2683 struct mem_cgroup
*memcg
)
2685 unsigned int nr_pages
= 1 << order
;
2686 struct page_counter
*counter
;
2689 ret
= try_charge(memcg
, gfp
, nr_pages
);
2693 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2694 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2695 cancel_charge(memcg
, nr_pages
);
2699 page
->mem_cgroup
= memcg
;
2705 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2706 * @page: page to charge
2707 * @gfp: reclaim mode
2708 * @order: allocation order
2710 * Returns 0 on success, an error code on failure.
2712 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2714 struct mem_cgroup
*memcg
;
2717 if (memcg_kmem_bypass())
2720 memcg
= get_mem_cgroup_from_current();
2721 if (!mem_cgroup_is_root(memcg
)) {
2722 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2724 __SetPageKmemcg(page
);
2726 css_put(&memcg
->css
);
2730 * __memcg_kmem_uncharge: uncharge a kmem page
2731 * @page: page to uncharge
2732 * @order: allocation order
2734 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2736 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2737 unsigned int nr_pages
= 1 << order
;
2742 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2744 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2745 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2747 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2748 if (do_memsw_account())
2749 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2751 page
->mem_cgroup
= NULL
;
2753 /* slab pages do not have PageKmemcg flag set */
2754 if (PageKmemcg(page
))
2755 __ClearPageKmemcg(page
);
2757 css_put_many(&memcg
->css
, nr_pages
);
2759 #endif /* CONFIG_MEMCG_KMEM */
2761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2764 * Because tail pages are not marked as "used", set it. We're under
2765 * pgdat->lru_lock and migration entries setup in all page mappings.
2767 void mem_cgroup_split_huge_fixup(struct page
*head
)
2771 if (mem_cgroup_disabled())
2774 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2775 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2777 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
2779 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2781 #ifdef CONFIG_MEMCG_SWAP
2783 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2784 * @entry: swap entry to be moved
2785 * @from: mem_cgroup which the entry is moved from
2786 * @to: mem_cgroup which the entry is moved to
2788 * It succeeds only when the swap_cgroup's record for this entry is the same
2789 * as the mem_cgroup's id of @from.
2791 * Returns 0 on success, -EINVAL on failure.
2793 * The caller must have charged to @to, IOW, called page_counter_charge() about
2794 * both res and memsw, and called css_get().
2796 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2797 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2799 unsigned short old_id
, new_id
;
2801 old_id
= mem_cgroup_id(from
);
2802 new_id
= mem_cgroup_id(to
);
2804 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2805 mod_memcg_state(from
, MEMCG_SWAP
, -1);
2806 mod_memcg_state(to
, MEMCG_SWAP
, 1);
2812 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2813 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2819 static DEFINE_MUTEX(memcg_max_mutex
);
2821 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
2822 unsigned long max
, bool memsw
)
2824 bool enlarge
= false;
2825 bool drained
= false;
2827 bool limits_invariant
;
2828 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
2831 if (signal_pending(current
)) {
2836 mutex_lock(&memcg_max_mutex
);
2838 * Make sure that the new limit (memsw or memory limit) doesn't
2839 * break our basic invariant rule memory.max <= memsw.max.
2841 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
2842 max
<= memcg
->memsw
.max
;
2843 if (!limits_invariant
) {
2844 mutex_unlock(&memcg_max_mutex
);
2848 if (max
> counter
->max
)
2850 ret
= page_counter_set_max(counter
, max
);
2851 mutex_unlock(&memcg_max_mutex
);
2857 drain_all_stock(memcg
);
2862 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
2863 GFP_KERNEL
, !memsw
)) {
2869 if (!ret
&& enlarge
)
2870 memcg_oom_recover(memcg
);
2875 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
2877 unsigned long *total_scanned
)
2879 unsigned long nr_reclaimed
= 0;
2880 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
2881 unsigned long reclaimed
;
2883 struct mem_cgroup_tree_per_node
*mctz
;
2884 unsigned long excess
;
2885 unsigned long nr_scanned
;
2890 mctz
= soft_limit_tree_node(pgdat
->node_id
);
2893 * Do not even bother to check the largest node if the root
2894 * is empty. Do it lockless to prevent lock bouncing. Races
2895 * are acceptable as soft limit is best effort anyway.
2897 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
2901 * This loop can run a while, specially if mem_cgroup's continuously
2902 * keep exceeding their soft limit and putting the system under
2909 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2914 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
2915 gfp_mask
, &nr_scanned
);
2916 nr_reclaimed
+= reclaimed
;
2917 *total_scanned
+= nr_scanned
;
2918 spin_lock_irq(&mctz
->lock
);
2919 __mem_cgroup_remove_exceeded(mz
, mctz
);
2922 * If we failed to reclaim anything from this memory cgroup
2923 * it is time to move on to the next cgroup
2927 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2929 excess
= soft_limit_excess(mz
->memcg
);
2931 * One school of thought says that we should not add
2932 * back the node to the tree if reclaim returns 0.
2933 * But our reclaim could return 0, simply because due
2934 * to priority we are exposing a smaller subset of
2935 * memory to reclaim from. Consider this as a longer
2938 /* If excess == 0, no tree ops */
2939 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2940 spin_unlock_irq(&mctz
->lock
);
2941 css_put(&mz
->memcg
->css
);
2944 * Could not reclaim anything and there are no more
2945 * mem cgroups to try or we seem to be looping without
2946 * reclaiming anything.
2948 if (!nr_reclaimed
&&
2950 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2952 } while (!nr_reclaimed
);
2954 css_put(&next_mz
->memcg
->css
);
2955 return nr_reclaimed
;
2959 * Test whether @memcg has children, dead or alive. Note that this
2960 * function doesn't care whether @memcg has use_hierarchy enabled and
2961 * returns %true if there are child csses according to the cgroup
2962 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2964 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2969 ret
= css_next_child(NULL
, &memcg
->css
);
2975 * Reclaims as many pages from the given memcg as possible.
2977 * Caller is responsible for holding css reference for memcg.
2979 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2981 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2983 /* we call try-to-free pages for make this cgroup empty */
2984 lru_add_drain_all();
2986 drain_all_stock(memcg
);
2988 /* try to free all pages in this cgroup */
2989 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2992 if (signal_pending(current
))
2995 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2999 /* maybe some writeback is necessary */
3000 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3008 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3009 char *buf
, size_t nbytes
,
3012 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3014 if (mem_cgroup_is_root(memcg
))
3016 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3019 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3022 return mem_cgroup_from_css(css
)->use_hierarchy
;
3025 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3026 struct cftype
*cft
, u64 val
)
3029 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3030 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3032 if (memcg
->use_hierarchy
== val
)
3036 * If parent's use_hierarchy is set, we can't make any modifications
3037 * in the child subtrees. If it is unset, then the change can
3038 * occur, provided the current cgroup has no children.
3040 * For the root cgroup, parent_mem is NULL, we allow value to be
3041 * set if there are no children.
3043 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3044 (val
== 1 || val
== 0)) {
3045 if (!memcg_has_children(memcg
))
3046 memcg
->use_hierarchy
= val
;
3055 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3059 if (mem_cgroup_is_root(memcg
)) {
3060 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3061 memcg_page_state(memcg
, MEMCG_RSS
);
3063 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3066 val
= page_counter_read(&memcg
->memory
);
3068 val
= page_counter_read(&memcg
->memsw
);
3081 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3084 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3085 struct page_counter
*counter
;
3087 switch (MEMFILE_TYPE(cft
->private)) {
3089 counter
= &memcg
->memory
;
3092 counter
= &memcg
->memsw
;
3095 counter
= &memcg
->kmem
;
3098 counter
= &memcg
->tcpmem
;
3104 switch (MEMFILE_ATTR(cft
->private)) {
3106 if (counter
== &memcg
->memory
)
3107 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3108 if (counter
== &memcg
->memsw
)
3109 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3110 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3112 return (u64
)counter
->max
* PAGE_SIZE
;
3114 return (u64
)counter
->watermark
* PAGE_SIZE
;
3116 return counter
->failcnt
;
3117 case RES_SOFT_LIMIT
:
3118 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3124 #ifdef CONFIG_MEMCG_KMEM
3125 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3129 if (cgroup_memory_nokmem
)
3132 BUG_ON(memcg
->kmemcg_id
>= 0);
3133 BUG_ON(memcg
->kmem_state
);
3135 memcg_id
= memcg_alloc_cache_id();
3139 static_branch_inc(&memcg_kmem_enabled_key
);
3141 * A memory cgroup is considered kmem-online as soon as it gets
3142 * kmemcg_id. Setting the id after enabling static branching will
3143 * guarantee no one starts accounting before all call sites are
3146 memcg
->kmemcg_id
= memcg_id
;
3147 memcg
->kmem_state
= KMEM_ONLINE
;
3148 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3153 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3155 struct cgroup_subsys_state
*css
;
3156 struct mem_cgroup
*parent
, *child
;
3159 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3162 * Clear the online state before clearing memcg_caches array
3163 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3164 * guarantees that no cache will be created for this cgroup
3165 * after we are done (see memcg_create_kmem_cache()).
3167 memcg
->kmem_state
= KMEM_ALLOCATED
;
3169 memcg_deactivate_kmem_caches(memcg
);
3171 kmemcg_id
= memcg
->kmemcg_id
;
3172 BUG_ON(kmemcg_id
< 0);
3174 parent
= parent_mem_cgroup(memcg
);
3176 parent
= root_mem_cgroup
;
3179 * Change kmemcg_id of this cgroup and all its descendants to the
3180 * parent's id, and then move all entries from this cgroup's list_lrus
3181 * to ones of the parent. After we have finished, all list_lrus
3182 * corresponding to this cgroup are guaranteed to remain empty. The
3183 * ordering is imposed by list_lru_node->lock taken by
3184 * memcg_drain_all_list_lrus().
3186 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3187 css_for_each_descendant_pre(css
, &memcg
->css
) {
3188 child
= mem_cgroup_from_css(css
);
3189 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3190 child
->kmemcg_id
= parent
->kmemcg_id
;
3191 if (!memcg
->use_hierarchy
)
3196 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3198 memcg_free_cache_id(kmemcg_id
);
3201 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3203 /* css_alloc() failed, offlining didn't happen */
3204 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3205 memcg_offline_kmem(memcg
);
3207 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3208 memcg_destroy_kmem_caches(memcg
);
3209 static_branch_dec(&memcg_kmem_enabled_key
);
3210 WARN_ON(page_counter_read(&memcg
->kmem
));
3214 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3218 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3221 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3224 #endif /* CONFIG_MEMCG_KMEM */
3226 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3231 mutex_lock(&memcg_max_mutex
);
3232 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3233 mutex_unlock(&memcg_max_mutex
);
3237 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3241 mutex_lock(&memcg_max_mutex
);
3243 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3247 if (!memcg
->tcpmem_active
) {
3249 * The active flag needs to be written after the static_key
3250 * update. This is what guarantees that the socket activation
3251 * function is the last one to run. See mem_cgroup_sk_alloc()
3252 * for details, and note that we don't mark any socket as
3253 * belonging to this memcg until that flag is up.
3255 * We need to do this, because static_keys will span multiple
3256 * sites, but we can't control their order. If we mark a socket
3257 * as accounted, but the accounting functions are not patched in
3258 * yet, we'll lose accounting.
3260 * We never race with the readers in mem_cgroup_sk_alloc(),
3261 * because when this value change, the code to process it is not
3264 static_branch_inc(&memcg_sockets_enabled_key
);
3265 memcg
->tcpmem_active
= true;
3268 mutex_unlock(&memcg_max_mutex
);
3273 * The user of this function is...
3276 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3277 char *buf
, size_t nbytes
, loff_t off
)
3279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3280 unsigned long nr_pages
;
3283 buf
= strstrip(buf
);
3284 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3288 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3290 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3294 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3296 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3299 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3302 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3305 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3309 case RES_SOFT_LIMIT
:
3310 memcg
->soft_limit
= nr_pages
;
3314 return ret
?: nbytes
;
3317 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3318 size_t nbytes
, loff_t off
)
3320 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3321 struct page_counter
*counter
;
3323 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3325 counter
= &memcg
->memory
;
3328 counter
= &memcg
->memsw
;
3331 counter
= &memcg
->kmem
;
3334 counter
= &memcg
->tcpmem
;
3340 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3342 page_counter_reset_watermark(counter
);
3345 counter
->failcnt
= 0;
3354 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3357 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3361 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3362 struct cftype
*cft
, u64 val
)
3364 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3366 if (val
& ~MOVE_MASK
)
3370 * No kind of locking is needed in here, because ->can_attach() will
3371 * check this value once in the beginning of the process, and then carry
3372 * on with stale data. This means that changes to this value will only
3373 * affect task migrations starting after the change.
3375 memcg
->move_charge_at_immigrate
= val
;
3379 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3380 struct cftype
*cft
, u64 val
)
3388 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3389 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3390 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3392 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3393 int nid
, unsigned int lru_mask
)
3395 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
3396 unsigned long nr
= 0;
3399 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3402 if (!(BIT(lru
) & lru_mask
))
3404 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3409 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3410 unsigned int lru_mask
)
3412 unsigned long nr
= 0;
3416 if (!(BIT(lru
) & lru_mask
))
3418 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3423 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3427 unsigned int lru_mask
;
3430 static const struct numa_stat stats
[] = {
3431 { "total", LRU_ALL
},
3432 { "file", LRU_ALL_FILE
},
3433 { "anon", LRU_ALL_ANON
},
3434 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3436 const struct numa_stat
*stat
;
3439 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3441 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3442 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3443 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3444 for_each_node_state(nid
, N_MEMORY
) {
3445 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3447 seq_printf(m
, " N%d=%lu", nid
, nr
);
3452 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3453 struct mem_cgroup
*iter
;
3456 for_each_mem_cgroup_tree(iter
, memcg
)
3457 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3458 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3459 for_each_node_state(nid
, N_MEMORY
) {
3461 for_each_mem_cgroup_tree(iter
, memcg
)
3462 nr
+= mem_cgroup_node_nr_lru_pages(
3463 iter
, nid
, stat
->lru_mask
);
3464 seq_printf(m
, " N%d=%lu", nid
, nr
);
3471 #endif /* CONFIG_NUMA */
3473 /* Universal VM events cgroup1 shows, original sort order */
3474 static const unsigned int memcg1_events
[] = {
3481 static const char *const memcg1_event_names
[] = {
3488 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3490 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3491 unsigned long memory
, memsw
;
3492 struct mem_cgroup
*mi
;
3495 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3496 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3498 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3499 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3501 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3502 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3506 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3507 seq_printf(m
, "%s %lu\n", memcg1_event_names
[i
],
3508 memcg_events_local(memcg
, memcg1_events
[i
]));
3510 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3511 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3512 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3515 /* Hierarchical information */
3516 memory
= memsw
= PAGE_COUNTER_MAX
;
3517 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3518 memory
= min(memory
, mi
->memory
.max
);
3519 memsw
= min(memsw
, mi
->memsw
.max
);
3521 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3522 (u64
)memory
* PAGE_SIZE
);
3523 if (do_memsw_account())
3524 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3525 (u64
)memsw
* PAGE_SIZE
);
3527 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3528 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3530 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3531 (u64
)memcg_page_state(memcg
, i
) * PAGE_SIZE
);
3534 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3535 seq_printf(m
, "total_%s %llu\n", memcg1_event_names
[i
],
3536 (u64
)memcg_events(memcg
, i
));
3538 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3539 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
],
3540 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3543 #ifdef CONFIG_DEBUG_VM
3546 struct mem_cgroup_per_node
*mz
;
3547 struct zone_reclaim_stat
*rstat
;
3548 unsigned long recent_rotated
[2] = {0, 0};
3549 unsigned long recent_scanned
[2] = {0, 0};
3551 for_each_online_pgdat(pgdat
) {
3552 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3553 rstat
= &mz
->lruvec
.reclaim_stat
;
3555 recent_rotated
[0] += rstat
->recent_rotated
[0];
3556 recent_rotated
[1] += rstat
->recent_rotated
[1];
3557 recent_scanned
[0] += rstat
->recent_scanned
[0];
3558 recent_scanned
[1] += rstat
->recent_scanned
[1];
3560 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3561 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3562 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3563 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3570 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3573 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3575 return mem_cgroup_swappiness(memcg
);
3578 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3579 struct cftype
*cft
, u64 val
)
3581 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3587 memcg
->swappiness
= val
;
3589 vm_swappiness
= val
;
3594 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3596 struct mem_cgroup_threshold_ary
*t
;
3597 unsigned long usage
;
3602 t
= rcu_dereference(memcg
->thresholds
.primary
);
3604 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3609 usage
= mem_cgroup_usage(memcg
, swap
);
3612 * current_threshold points to threshold just below or equal to usage.
3613 * If it's not true, a threshold was crossed after last
3614 * call of __mem_cgroup_threshold().
3616 i
= t
->current_threshold
;
3619 * Iterate backward over array of thresholds starting from
3620 * current_threshold and check if a threshold is crossed.
3621 * If none of thresholds below usage is crossed, we read
3622 * only one element of the array here.
3624 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3625 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3627 /* i = current_threshold + 1 */
3631 * Iterate forward over array of thresholds starting from
3632 * current_threshold+1 and check if a threshold is crossed.
3633 * If none of thresholds above usage is crossed, we read
3634 * only one element of the array here.
3636 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3637 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3639 /* Update current_threshold */
3640 t
->current_threshold
= i
- 1;
3645 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3648 __mem_cgroup_threshold(memcg
, false);
3649 if (do_memsw_account())
3650 __mem_cgroup_threshold(memcg
, true);
3652 memcg
= parent_mem_cgroup(memcg
);
3656 static int compare_thresholds(const void *a
, const void *b
)
3658 const struct mem_cgroup_threshold
*_a
= a
;
3659 const struct mem_cgroup_threshold
*_b
= b
;
3661 if (_a
->threshold
> _b
->threshold
)
3664 if (_a
->threshold
< _b
->threshold
)
3670 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3672 struct mem_cgroup_eventfd_list
*ev
;
3674 spin_lock(&memcg_oom_lock
);
3676 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3677 eventfd_signal(ev
->eventfd
, 1);
3679 spin_unlock(&memcg_oom_lock
);
3683 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3685 struct mem_cgroup
*iter
;
3687 for_each_mem_cgroup_tree(iter
, memcg
)
3688 mem_cgroup_oom_notify_cb(iter
);
3691 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3692 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3694 struct mem_cgroup_thresholds
*thresholds
;
3695 struct mem_cgroup_threshold_ary
*new;
3696 unsigned long threshold
;
3697 unsigned long usage
;
3700 ret
= page_counter_memparse(args
, "-1", &threshold
);
3704 mutex_lock(&memcg
->thresholds_lock
);
3707 thresholds
= &memcg
->thresholds
;
3708 usage
= mem_cgroup_usage(memcg
, false);
3709 } else if (type
== _MEMSWAP
) {
3710 thresholds
= &memcg
->memsw_thresholds
;
3711 usage
= mem_cgroup_usage(memcg
, true);
3715 /* Check if a threshold crossed before adding a new one */
3716 if (thresholds
->primary
)
3717 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3719 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3721 /* Allocate memory for new array of thresholds */
3722 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
3729 /* Copy thresholds (if any) to new array */
3730 if (thresholds
->primary
) {
3731 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3732 sizeof(struct mem_cgroup_threshold
));
3735 /* Add new threshold */
3736 new->entries
[size
- 1].eventfd
= eventfd
;
3737 new->entries
[size
- 1].threshold
= threshold
;
3739 /* Sort thresholds. Registering of new threshold isn't time-critical */
3740 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3741 compare_thresholds
, NULL
);
3743 /* Find current threshold */
3744 new->current_threshold
= -1;
3745 for (i
= 0; i
< size
; i
++) {
3746 if (new->entries
[i
].threshold
<= usage
) {
3748 * new->current_threshold will not be used until
3749 * rcu_assign_pointer(), so it's safe to increment
3752 ++new->current_threshold
;
3757 /* Free old spare buffer and save old primary buffer as spare */
3758 kfree(thresholds
->spare
);
3759 thresholds
->spare
= thresholds
->primary
;
3761 rcu_assign_pointer(thresholds
->primary
, new);
3763 /* To be sure that nobody uses thresholds */
3767 mutex_unlock(&memcg
->thresholds_lock
);
3772 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3773 struct eventfd_ctx
*eventfd
, const char *args
)
3775 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3778 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3779 struct eventfd_ctx
*eventfd
, const char *args
)
3781 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3784 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3785 struct eventfd_ctx
*eventfd
, enum res_type type
)
3787 struct mem_cgroup_thresholds
*thresholds
;
3788 struct mem_cgroup_threshold_ary
*new;
3789 unsigned long usage
;
3792 mutex_lock(&memcg
->thresholds_lock
);
3795 thresholds
= &memcg
->thresholds
;
3796 usage
= mem_cgroup_usage(memcg
, false);
3797 } else if (type
== _MEMSWAP
) {
3798 thresholds
= &memcg
->memsw_thresholds
;
3799 usage
= mem_cgroup_usage(memcg
, true);
3803 if (!thresholds
->primary
)
3806 /* Check if a threshold crossed before removing */
3807 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3809 /* Calculate new number of threshold */
3811 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3812 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3816 new = thresholds
->spare
;
3818 /* Set thresholds array to NULL if we don't have thresholds */
3827 /* Copy thresholds and find current threshold */
3828 new->current_threshold
= -1;
3829 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3830 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3833 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3834 if (new->entries
[j
].threshold
<= usage
) {
3836 * new->current_threshold will not be used
3837 * until rcu_assign_pointer(), so it's safe to increment
3840 ++new->current_threshold
;
3846 /* Swap primary and spare array */
3847 thresholds
->spare
= thresholds
->primary
;
3849 rcu_assign_pointer(thresholds
->primary
, new);
3851 /* To be sure that nobody uses thresholds */
3854 /* If all events are unregistered, free the spare array */
3856 kfree(thresholds
->spare
);
3857 thresholds
->spare
= NULL
;
3860 mutex_unlock(&memcg
->thresholds_lock
);
3863 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3864 struct eventfd_ctx
*eventfd
)
3866 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3869 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3870 struct eventfd_ctx
*eventfd
)
3872 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3875 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3876 struct eventfd_ctx
*eventfd
, const char *args
)
3878 struct mem_cgroup_eventfd_list
*event
;
3880 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3884 spin_lock(&memcg_oom_lock
);
3886 event
->eventfd
= eventfd
;
3887 list_add(&event
->list
, &memcg
->oom_notify
);
3889 /* already in OOM ? */
3890 if (memcg
->under_oom
)
3891 eventfd_signal(eventfd
, 1);
3892 spin_unlock(&memcg_oom_lock
);
3897 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3898 struct eventfd_ctx
*eventfd
)
3900 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3902 spin_lock(&memcg_oom_lock
);
3904 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3905 if (ev
->eventfd
== eventfd
) {
3906 list_del(&ev
->list
);
3911 spin_unlock(&memcg_oom_lock
);
3914 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3916 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
3918 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3919 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3920 seq_printf(sf
, "oom_kill %lu\n",
3921 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
3925 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3926 struct cftype
*cft
, u64 val
)
3928 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3930 /* cannot set to root cgroup and only 0 and 1 are allowed */
3931 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3934 memcg
->oom_kill_disable
= val
;
3936 memcg_oom_recover(memcg
);
3941 #ifdef CONFIG_CGROUP_WRITEBACK
3943 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3945 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3948 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3950 wb_domain_exit(&memcg
->cgwb_domain
);
3953 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3955 wb_domain_size_changed(&memcg
->cgwb_domain
);
3958 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3960 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3962 if (!memcg
->css
.parent
)
3965 return &memcg
->cgwb_domain
;
3969 * idx can be of type enum memcg_stat_item or node_stat_item.
3970 * Keep in sync with memcg_exact_page().
3972 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
3974 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
3977 for_each_online_cpu(cpu
)
3978 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
3985 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3986 * @wb: bdi_writeback in question
3987 * @pfilepages: out parameter for number of file pages
3988 * @pheadroom: out parameter for number of allocatable pages according to memcg
3989 * @pdirty: out parameter for number of dirty pages
3990 * @pwriteback: out parameter for number of pages under writeback
3992 * Determine the numbers of file, headroom, dirty, and writeback pages in
3993 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3994 * is a bit more involved.
3996 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3997 * headroom is calculated as the lowest headroom of itself and the
3998 * ancestors. Note that this doesn't consider the actual amount of
3999 * available memory in the system. The caller should further cap
4000 * *@pheadroom accordingly.
4002 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4003 unsigned long *pheadroom
, unsigned long *pdirty
,
4004 unsigned long *pwriteback
)
4006 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4007 struct mem_cgroup
*parent
;
4009 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4011 /* this should eventually include NR_UNSTABLE_NFS */
4012 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4013 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4014 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4015 *pheadroom
= PAGE_COUNTER_MAX
;
4017 while ((parent
= parent_mem_cgroup(memcg
))) {
4018 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
4019 unsigned long used
= page_counter_read(&memcg
->memory
);
4021 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4026 #else /* CONFIG_CGROUP_WRITEBACK */
4028 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4033 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4037 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4041 #endif /* CONFIG_CGROUP_WRITEBACK */
4044 * DO NOT USE IN NEW FILES.
4046 * "cgroup.event_control" implementation.
4048 * This is way over-engineered. It tries to support fully configurable
4049 * events for each user. Such level of flexibility is completely
4050 * unnecessary especially in the light of the planned unified hierarchy.
4052 * Please deprecate this and replace with something simpler if at all
4057 * Unregister event and free resources.
4059 * Gets called from workqueue.
4061 static void memcg_event_remove(struct work_struct
*work
)
4063 struct mem_cgroup_event
*event
=
4064 container_of(work
, struct mem_cgroup_event
, remove
);
4065 struct mem_cgroup
*memcg
= event
->memcg
;
4067 remove_wait_queue(event
->wqh
, &event
->wait
);
4069 event
->unregister_event(memcg
, event
->eventfd
);
4071 /* Notify userspace the event is going away. */
4072 eventfd_signal(event
->eventfd
, 1);
4074 eventfd_ctx_put(event
->eventfd
);
4076 css_put(&memcg
->css
);
4080 * Gets called on EPOLLHUP on eventfd when user closes it.
4082 * Called with wqh->lock held and interrupts disabled.
4084 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4085 int sync
, void *key
)
4087 struct mem_cgroup_event
*event
=
4088 container_of(wait
, struct mem_cgroup_event
, wait
);
4089 struct mem_cgroup
*memcg
= event
->memcg
;
4090 __poll_t flags
= key_to_poll(key
);
4092 if (flags
& EPOLLHUP
) {
4094 * If the event has been detached at cgroup removal, we
4095 * can simply return knowing the other side will cleanup
4098 * We can't race against event freeing since the other
4099 * side will require wqh->lock via remove_wait_queue(),
4102 spin_lock(&memcg
->event_list_lock
);
4103 if (!list_empty(&event
->list
)) {
4104 list_del_init(&event
->list
);
4106 * We are in atomic context, but cgroup_event_remove()
4107 * may sleep, so we have to call it in workqueue.
4109 schedule_work(&event
->remove
);
4111 spin_unlock(&memcg
->event_list_lock
);
4117 static void memcg_event_ptable_queue_proc(struct file
*file
,
4118 wait_queue_head_t
*wqh
, poll_table
*pt
)
4120 struct mem_cgroup_event
*event
=
4121 container_of(pt
, struct mem_cgroup_event
, pt
);
4124 add_wait_queue(wqh
, &event
->wait
);
4128 * DO NOT USE IN NEW FILES.
4130 * Parse input and register new cgroup event handler.
4132 * Input must be in format '<event_fd> <control_fd> <args>'.
4133 * Interpretation of args is defined by control file implementation.
4135 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4136 char *buf
, size_t nbytes
, loff_t off
)
4138 struct cgroup_subsys_state
*css
= of_css(of
);
4139 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4140 struct mem_cgroup_event
*event
;
4141 struct cgroup_subsys_state
*cfile_css
;
4142 unsigned int efd
, cfd
;
4149 buf
= strstrip(buf
);
4151 efd
= simple_strtoul(buf
, &endp
, 10);
4156 cfd
= simple_strtoul(buf
, &endp
, 10);
4157 if ((*endp
!= ' ') && (*endp
!= '\0'))
4161 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4165 event
->memcg
= memcg
;
4166 INIT_LIST_HEAD(&event
->list
);
4167 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4168 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4169 INIT_WORK(&event
->remove
, memcg_event_remove
);
4177 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4178 if (IS_ERR(event
->eventfd
)) {
4179 ret
= PTR_ERR(event
->eventfd
);
4186 goto out_put_eventfd
;
4189 /* the process need read permission on control file */
4190 /* AV: shouldn't we check that it's been opened for read instead? */
4191 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4196 * Determine the event callbacks and set them in @event. This used
4197 * to be done via struct cftype but cgroup core no longer knows
4198 * about these events. The following is crude but the whole thing
4199 * is for compatibility anyway.
4201 * DO NOT ADD NEW FILES.
4203 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4205 if (!strcmp(name
, "memory.usage_in_bytes")) {
4206 event
->register_event
= mem_cgroup_usage_register_event
;
4207 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4208 } else if (!strcmp(name
, "memory.oom_control")) {
4209 event
->register_event
= mem_cgroup_oom_register_event
;
4210 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4211 } else if (!strcmp(name
, "memory.pressure_level")) {
4212 event
->register_event
= vmpressure_register_event
;
4213 event
->unregister_event
= vmpressure_unregister_event
;
4214 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4215 event
->register_event
= memsw_cgroup_usage_register_event
;
4216 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4223 * Verify @cfile should belong to @css. Also, remaining events are
4224 * automatically removed on cgroup destruction but the removal is
4225 * asynchronous, so take an extra ref on @css.
4227 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4228 &memory_cgrp_subsys
);
4230 if (IS_ERR(cfile_css
))
4232 if (cfile_css
!= css
) {
4237 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4241 vfs_poll(efile
.file
, &event
->pt
);
4243 spin_lock(&memcg
->event_list_lock
);
4244 list_add(&event
->list
, &memcg
->event_list
);
4245 spin_unlock(&memcg
->event_list_lock
);
4257 eventfd_ctx_put(event
->eventfd
);
4266 static struct cftype mem_cgroup_legacy_files
[] = {
4268 .name
= "usage_in_bytes",
4269 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4270 .read_u64
= mem_cgroup_read_u64
,
4273 .name
= "max_usage_in_bytes",
4274 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4275 .write
= mem_cgroup_reset
,
4276 .read_u64
= mem_cgroup_read_u64
,
4279 .name
= "limit_in_bytes",
4280 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4281 .write
= mem_cgroup_write
,
4282 .read_u64
= mem_cgroup_read_u64
,
4285 .name
= "soft_limit_in_bytes",
4286 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4287 .write
= mem_cgroup_write
,
4288 .read_u64
= mem_cgroup_read_u64
,
4292 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4293 .write
= mem_cgroup_reset
,
4294 .read_u64
= mem_cgroup_read_u64
,
4298 .seq_show
= memcg_stat_show
,
4301 .name
= "force_empty",
4302 .write
= mem_cgroup_force_empty_write
,
4305 .name
= "use_hierarchy",
4306 .write_u64
= mem_cgroup_hierarchy_write
,
4307 .read_u64
= mem_cgroup_hierarchy_read
,
4310 .name
= "cgroup.event_control", /* XXX: for compat */
4311 .write
= memcg_write_event_control
,
4312 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4315 .name
= "swappiness",
4316 .read_u64
= mem_cgroup_swappiness_read
,
4317 .write_u64
= mem_cgroup_swappiness_write
,
4320 .name
= "move_charge_at_immigrate",
4321 .read_u64
= mem_cgroup_move_charge_read
,
4322 .write_u64
= mem_cgroup_move_charge_write
,
4325 .name
= "oom_control",
4326 .seq_show
= mem_cgroup_oom_control_read
,
4327 .write_u64
= mem_cgroup_oom_control_write
,
4328 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4331 .name
= "pressure_level",
4335 .name
= "numa_stat",
4336 .seq_show
= memcg_numa_stat_show
,
4340 .name
= "kmem.limit_in_bytes",
4341 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4342 .write
= mem_cgroup_write
,
4343 .read_u64
= mem_cgroup_read_u64
,
4346 .name
= "kmem.usage_in_bytes",
4347 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4348 .read_u64
= mem_cgroup_read_u64
,
4351 .name
= "kmem.failcnt",
4352 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4353 .write
= mem_cgroup_reset
,
4354 .read_u64
= mem_cgroup_read_u64
,
4357 .name
= "kmem.max_usage_in_bytes",
4358 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4359 .write
= mem_cgroup_reset
,
4360 .read_u64
= mem_cgroup_read_u64
,
4362 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4364 .name
= "kmem.slabinfo",
4365 .seq_start
= memcg_slab_start
,
4366 .seq_next
= memcg_slab_next
,
4367 .seq_stop
= memcg_slab_stop
,
4368 .seq_show
= memcg_slab_show
,
4372 .name
= "kmem.tcp.limit_in_bytes",
4373 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4374 .write
= mem_cgroup_write
,
4375 .read_u64
= mem_cgroup_read_u64
,
4378 .name
= "kmem.tcp.usage_in_bytes",
4379 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4380 .read_u64
= mem_cgroup_read_u64
,
4383 .name
= "kmem.tcp.failcnt",
4384 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4385 .write
= mem_cgroup_reset
,
4386 .read_u64
= mem_cgroup_read_u64
,
4389 .name
= "kmem.tcp.max_usage_in_bytes",
4390 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4391 .write
= mem_cgroup_reset
,
4392 .read_u64
= mem_cgroup_read_u64
,
4394 { }, /* terminate */
4398 * Private memory cgroup IDR
4400 * Swap-out records and page cache shadow entries need to store memcg
4401 * references in constrained space, so we maintain an ID space that is
4402 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4403 * memory-controlled cgroups to 64k.
4405 * However, there usually are many references to the oflline CSS after
4406 * the cgroup has been destroyed, such as page cache or reclaimable
4407 * slab objects, that don't need to hang on to the ID. We want to keep
4408 * those dead CSS from occupying IDs, or we might quickly exhaust the
4409 * relatively small ID space and prevent the creation of new cgroups
4410 * even when there are much fewer than 64k cgroups - possibly none.
4412 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4413 * be freed and recycled when it's no longer needed, which is usually
4414 * when the CSS is offlined.
4416 * The only exception to that are records of swapped out tmpfs/shmem
4417 * pages that need to be attributed to live ancestors on swapin. But
4418 * those references are manageable from userspace.
4421 static DEFINE_IDR(mem_cgroup_idr
);
4423 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4425 if (memcg
->id
.id
> 0) {
4426 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4431 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4433 refcount_add(n
, &memcg
->id
.ref
);
4436 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4438 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4439 mem_cgroup_id_remove(memcg
);
4441 /* Memcg ID pins CSS */
4442 css_put(&memcg
->css
);
4446 static inline void mem_cgroup_id_get(struct mem_cgroup
*memcg
)
4448 mem_cgroup_id_get_many(memcg
, 1);
4451 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4453 mem_cgroup_id_put_many(memcg
, 1);
4457 * mem_cgroup_from_id - look up a memcg from a memcg id
4458 * @id: the memcg id to look up
4460 * Caller must hold rcu_read_lock().
4462 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4464 WARN_ON_ONCE(!rcu_read_lock_held());
4465 return idr_find(&mem_cgroup_idr
, id
);
4468 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4470 struct mem_cgroup_per_node
*pn
;
4473 * This routine is called against possible nodes.
4474 * But it's BUG to call kmalloc() against offline node.
4476 * TODO: this routine can waste much memory for nodes which will
4477 * never be onlined. It's better to use memory hotplug callback
4480 if (!node_state(node
, N_NORMAL_MEMORY
))
4482 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4486 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4487 if (!pn
->lruvec_stat_cpu
) {
4492 lruvec_init(&pn
->lruvec
);
4493 pn
->usage_in_excess
= 0;
4494 pn
->on_tree
= false;
4497 memcg
->nodeinfo
[node
] = pn
;
4501 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4503 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4508 free_percpu(pn
->lruvec_stat_cpu
);
4512 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4517 free_mem_cgroup_per_node_info(memcg
, node
);
4518 free_percpu(memcg
->vmstats_percpu
);
4522 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4524 memcg_wb_domain_exit(memcg
);
4525 __mem_cgroup_free(memcg
);
4528 static struct mem_cgroup
*mem_cgroup_alloc(void)
4530 struct mem_cgroup
*memcg
;
4534 size
= sizeof(struct mem_cgroup
);
4535 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4537 memcg
= kzalloc(size
, GFP_KERNEL
);
4541 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4542 1, MEM_CGROUP_ID_MAX
,
4544 if (memcg
->id
.id
< 0)
4547 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
4548 if (!memcg
->vmstats_percpu
)
4552 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4555 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4558 INIT_WORK(&memcg
->high_work
, high_work_func
);
4559 memcg
->last_scanned_node
= MAX_NUMNODES
;
4560 INIT_LIST_HEAD(&memcg
->oom_notify
);
4561 mutex_init(&memcg
->thresholds_lock
);
4562 spin_lock_init(&memcg
->move_lock
);
4563 vmpressure_init(&memcg
->vmpressure
);
4564 INIT_LIST_HEAD(&memcg
->event_list
);
4565 spin_lock_init(&memcg
->event_list_lock
);
4566 memcg
->socket_pressure
= jiffies
;
4567 #ifdef CONFIG_MEMCG_KMEM
4568 memcg
->kmemcg_id
= -1;
4570 #ifdef CONFIG_CGROUP_WRITEBACK
4571 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4573 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4576 mem_cgroup_id_remove(memcg
);
4577 __mem_cgroup_free(memcg
);
4581 static struct cgroup_subsys_state
* __ref
4582 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4584 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4585 struct mem_cgroup
*memcg
;
4586 long error
= -ENOMEM
;
4588 memcg
= mem_cgroup_alloc();
4590 return ERR_PTR(error
);
4592 memcg
->high
= PAGE_COUNTER_MAX
;
4593 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4595 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4596 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4598 if (parent
&& parent
->use_hierarchy
) {
4599 memcg
->use_hierarchy
= true;
4600 page_counter_init(&memcg
->memory
, &parent
->memory
);
4601 page_counter_init(&memcg
->swap
, &parent
->swap
);
4602 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4603 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4604 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4606 page_counter_init(&memcg
->memory
, NULL
);
4607 page_counter_init(&memcg
->swap
, NULL
);
4608 page_counter_init(&memcg
->memsw
, NULL
);
4609 page_counter_init(&memcg
->kmem
, NULL
);
4610 page_counter_init(&memcg
->tcpmem
, NULL
);
4612 * Deeper hierachy with use_hierarchy == false doesn't make
4613 * much sense so let cgroup subsystem know about this
4614 * unfortunate state in our controller.
4616 if (parent
!= root_mem_cgroup
)
4617 memory_cgrp_subsys
.broken_hierarchy
= true;
4620 /* The following stuff does not apply to the root */
4622 root_mem_cgroup
= memcg
;
4626 error
= memcg_online_kmem(memcg
);
4630 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4631 static_branch_inc(&memcg_sockets_enabled_key
);
4635 mem_cgroup_id_remove(memcg
);
4636 mem_cgroup_free(memcg
);
4637 return ERR_PTR(-ENOMEM
);
4640 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4642 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4645 * A memcg must be visible for memcg_expand_shrinker_maps()
4646 * by the time the maps are allocated. So, we allocate maps
4647 * here, when for_each_mem_cgroup() can't skip it.
4649 if (memcg_alloc_shrinker_maps(memcg
)) {
4650 mem_cgroup_id_remove(memcg
);
4654 /* Online state pins memcg ID, memcg ID pins CSS */
4655 refcount_set(&memcg
->id
.ref
, 1);
4660 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4662 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4663 struct mem_cgroup_event
*event
, *tmp
;
4666 * Unregister events and notify userspace.
4667 * Notify userspace about cgroup removing only after rmdir of cgroup
4668 * directory to avoid race between userspace and kernelspace.
4670 spin_lock(&memcg
->event_list_lock
);
4671 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4672 list_del_init(&event
->list
);
4673 schedule_work(&event
->remove
);
4675 spin_unlock(&memcg
->event_list_lock
);
4677 page_counter_set_min(&memcg
->memory
, 0);
4678 page_counter_set_low(&memcg
->memory
, 0);
4680 memcg_offline_kmem(memcg
);
4681 wb_memcg_offline(memcg
);
4683 drain_all_stock(memcg
);
4685 mem_cgroup_id_put(memcg
);
4688 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4690 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4692 invalidate_reclaim_iterators(memcg
);
4695 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4697 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4699 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4700 static_branch_dec(&memcg_sockets_enabled_key
);
4702 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4703 static_branch_dec(&memcg_sockets_enabled_key
);
4705 vmpressure_cleanup(&memcg
->vmpressure
);
4706 cancel_work_sync(&memcg
->high_work
);
4707 mem_cgroup_remove_from_trees(memcg
);
4708 memcg_free_shrinker_maps(memcg
);
4709 memcg_free_kmem(memcg
);
4710 mem_cgroup_free(memcg
);
4714 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4715 * @css: the target css
4717 * Reset the states of the mem_cgroup associated with @css. This is
4718 * invoked when the userland requests disabling on the default hierarchy
4719 * but the memcg is pinned through dependency. The memcg should stop
4720 * applying policies and should revert to the vanilla state as it may be
4721 * made visible again.
4723 * The current implementation only resets the essential configurations.
4724 * This needs to be expanded to cover all the visible parts.
4726 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4728 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4730 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
4731 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
4732 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
4733 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
4734 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
4735 page_counter_set_min(&memcg
->memory
, 0);
4736 page_counter_set_low(&memcg
->memory
, 0);
4737 memcg
->high
= PAGE_COUNTER_MAX
;
4738 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4739 memcg_wb_domain_size_changed(memcg
);
4743 /* Handlers for move charge at task migration. */
4744 static int mem_cgroup_do_precharge(unsigned long count
)
4748 /* Try a single bulk charge without reclaim first, kswapd may wake */
4749 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4751 mc
.precharge
+= count
;
4755 /* Try charges one by one with reclaim, but do not retry */
4757 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
4771 enum mc_target_type
{
4778 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4779 unsigned long addr
, pte_t ptent
)
4781 struct page
*page
= _vm_normal_page(vma
, addr
, ptent
, true);
4783 if (!page
|| !page_mapped(page
))
4785 if (PageAnon(page
)) {
4786 if (!(mc
.flags
& MOVE_ANON
))
4789 if (!(mc
.flags
& MOVE_FILE
))
4792 if (!get_page_unless_zero(page
))
4798 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4799 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4800 pte_t ptent
, swp_entry_t
*entry
)
4802 struct page
*page
= NULL
;
4803 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4805 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4809 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4810 * a device and because they are not accessible by CPU they are store
4811 * as special swap entry in the CPU page table.
4813 if (is_device_private_entry(ent
)) {
4814 page
= device_private_entry_to_page(ent
);
4816 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4817 * a refcount of 1 when free (unlike normal page)
4819 if (!page_ref_add_unless(page
, 1, 1))
4825 * Because lookup_swap_cache() updates some statistics counter,
4826 * we call find_get_page() with swapper_space directly.
4828 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
4829 if (do_memsw_account())
4830 entry
->val
= ent
.val
;
4835 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4836 pte_t ptent
, swp_entry_t
*entry
)
4842 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4843 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4845 struct page
*page
= NULL
;
4846 struct address_space
*mapping
;
4849 if (!vma
->vm_file
) /* anonymous vma */
4851 if (!(mc
.flags
& MOVE_FILE
))
4854 mapping
= vma
->vm_file
->f_mapping
;
4855 pgoff
= linear_page_index(vma
, addr
);
4857 /* page is moved even if it's not RSS of this task(page-faulted). */
4859 /* shmem/tmpfs may report page out on swap: account for that too. */
4860 if (shmem_mapping(mapping
)) {
4861 page
= find_get_entry(mapping
, pgoff
);
4862 if (xa_is_value(page
)) {
4863 swp_entry_t swp
= radix_to_swp_entry(page
);
4864 if (do_memsw_account())
4866 page
= find_get_page(swap_address_space(swp
),
4870 page
= find_get_page(mapping
, pgoff
);
4872 page
= find_get_page(mapping
, pgoff
);
4878 * mem_cgroup_move_account - move account of the page
4880 * @compound: charge the page as compound or small page
4881 * @from: mem_cgroup which the page is moved from.
4882 * @to: mem_cgroup which the page is moved to. @from != @to.
4884 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4886 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4889 static int mem_cgroup_move_account(struct page
*page
,
4891 struct mem_cgroup
*from
,
4892 struct mem_cgroup
*to
)
4894 unsigned long flags
;
4895 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4899 VM_BUG_ON(from
== to
);
4900 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4901 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4904 * Prevent mem_cgroup_migrate() from looking at
4905 * page->mem_cgroup of its source page while we change it.
4908 if (!trylock_page(page
))
4912 if (page
->mem_cgroup
!= from
)
4915 anon
= PageAnon(page
);
4917 spin_lock_irqsave(&from
->move_lock
, flags
);
4919 if (!anon
&& page_mapped(page
)) {
4920 __mod_memcg_state(from
, NR_FILE_MAPPED
, -nr_pages
);
4921 __mod_memcg_state(to
, NR_FILE_MAPPED
, nr_pages
);
4925 * move_lock grabbed above and caller set from->moving_account, so
4926 * mod_memcg_page_state will serialize updates to PageDirty.
4927 * So mapping should be stable for dirty pages.
4929 if (!anon
&& PageDirty(page
)) {
4930 struct address_space
*mapping
= page_mapping(page
);
4932 if (mapping_cap_account_dirty(mapping
)) {
4933 __mod_memcg_state(from
, NR_FILE_DIRTY
, -nr_pages
);
4934 __mod_memcg_state(to
, NR_FILE_DIRTY
, nr_pages
);
4938 if (PageWriteback(page
)) {
4939 __mod_memcg_state(from
, NR_WRITEBACK
, -nr_pages
);
4940 __mod_memcg_state(to
, NR_WRITEBACK
, nr_pages
);
4944 * It is safe to change page->mem_cgroup here because the page
4945 * is referenced, charged, and isolated - we can't race with
4946 * uncharging, charging, migration, or LRU putback.
4949 /* caller should have done css_get */
4950 page
->mem_cgroup
= to
;
4951 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4955 local_irq_disable();
4956 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4957 memcg_check_events(to
, page
);
4958 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4959 memcg_check_events(from
, page
);
4968 * get_mctgt_type - get target type of moving charge
4969 * @vma: the vma the pte to be checked belongs
4970 * @addr: the address corresponding to the pte to be checked
4971 * @ptent: the pte to be checked
4972 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4975 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4976 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4977 * move charge. if @target is not NULL, the page is stored in target->page
4978 * with extra refcnt got(Callers should handle it).
4979 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4980 * target for charge migration. if @target is not NULL, the entry is stored
4982 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4983 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4984 * For now we such page is charge like a regular page would be as for all
4985 * intent and purposes it is just special memory taking the place of a
4988 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4990 * Called with pte lock held.
4993 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4994 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4996 struct page
*page
= NULL
;
4997 enum mc_target_type ret
= MC_TARGET_NONE
;
4998 swp_entry_t ent
= { .val
= 0 };
5000 if (pte_present(ptent
))
5001 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5002 else if (is_swap_pte(ptent
))
5003 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5004 else if (pte_none(ptent
))
5005 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5007 if (!page
&& !ent
.val
)
5011 * Do only loose check w/o serialization.
5012 * mem_cgroup_move_account() checks the page is valid or
5013 * not under LRU exclusion.
5015 if (page
->mem_cgroup
== mc
.from
) {
5016 ret
= MC_TARGET_PAGE
;
5017 if (is_device_private_page(page
) ||
5018 is_device_public_page(page
))
5019 ret
= MC_TARGET_DEVICE
;
5021 target
->page
= page
;
5023 if (!ret
|| !target
)
5027 * There is a swap entry and a page doesn't exist or isn't charged.
5028 * But we cannot move a tail-page in a THP.
5030 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5031 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5032 ret
= MC_TARGET_SWAP
;
5039 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5041 * We don't consider PMD mapped swapping or file mapped pages because THP does
5042 * not support them for now.
5043 * Caller should make sure that pmd_trans_huge(pmd) is true.
5045 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5046 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5048 struct page
*page
= NULL
;
5049 enum mc_target_type ret
= MC_TARGET_NONE
;
5051 if (unlikely(is_swap_pmd(pmd
))) {
5052 VM_BUG_ON(thp_migration_supported() &&
5053 !is_pmd_migration_entry(pmd
));
5056 page
= pmd_page(pmd
);
5057 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5058 if (!(mc
.flags
& MOVE_ANON
))
5060 if (page
->mem_cgroup
== mc
.from
) {
5061 ret
= MC_TARGET_PAGE
;
5064 target
->page
= page
;
5070 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5071 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5073 return MC_TARGET_NONE
;
5077 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5078 unsigned long addr
, unsigned long end
,
5079 struct mm_walk
*walk
)
5081 struct vm_area_struct
*vma
= walk
->vma
;
5085 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5088 * Note their can not be MC_TARGET_DEVICE for now as we do not
5089 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
5090 * MEMORY_DEVICE_PRIVATE but this might change.
5092 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5093 mc
.precharge
+= HPAGE_PMD_NR
;
5098 if (pmd_trans_unstable(pmd
))
5100 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5101 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5102 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5103 mc
.precharge
++; /* increment precharge temporarily */
5104 pte_unmap_unlock(pte
- 1, ptl
);
5110 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5112 unsigned long precharge
;
5114 struct mm_walk mem_cgroup_count_precharge_walk
= {
5115 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5118 down_read(&mm
->mmap_sem
);
5119 walk_page_range(0, mm
->highest_vm_end
,
5120 &mem_cgroup_count_precharge_walk
);
5121 up_read(&mm
->mmap_sem
);
5123 precharge
= mc
.precharge
;
5129 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5131 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5133 VM_BUG_ON(mc
.moving_task
);
5134 mc
.moving_task
= current
;
5135 return mem_cgroup_do_precharge(precharge
);
5138 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5139 static void __mem_cgroup_clear_mc(void)
5141 struct mem_cgroup
*from
= mc
.from
;
5142 struct mem_cgroup
*to
= mc
.to
;
5144 /* we must uncharge all the leftover precharges from mc.to */
5146 cancel_charge(mc
.to
, mc
.precharge
);
5150 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5151 * we must uncharge here.
5153 if (mc
.moved_charge
) {
5154 cancel_charge(mc
.from
, mc
.moved_charge
);
5155 mc
.moved_charge
= 0;
5157 /* we must fixup refcnts and charges */
5158 if (mc
.moved_swap
) {
5159 /* uncharge swap account from the old cgroup */
5160 if (!mem_cgroup_is_root(mc
.from
))
5161 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5163 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5166 * we charged both to->memory and to->memsw, so we
5167 * should uncharge to->memory.
5169 if (!mem_cgroup_is_root(mc
.to
))
5170 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5172 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5173 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5177 memcg_oom_recover(from
);
5178 memcg_oom_recover(to
);
5179 wake_up_all(&mc
.waitq
);
5182 static void mem_cgroup_clear_mc(void)
5184 struct mm_struct
*mm
= mc
.mm
;
5187 * we must clear moving_task before waking up waiters at the end of
5190 mc
.moving_task
= NULL
;
5191 __mem_cgroup_clear_mc();
5192 spin_lock(&mc
.lock
);
5196 spin_unlock(&mc
.lock
);
5201 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5203 struct cgroup_subsys_state
*css
;
5204 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5205 struct mem_cgroup
*from
;
5206 struct task_struct
*leader
, *p
;
5207 struct mm_struct
*mm
;
5208 unsigned long move_flags
;
5211 /* charge immigration isn't supported on the default hierarchy */
5212 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5216 * Multi-process migrations only happen on the default hierarchy
5217 * where charge immigration is not used. Perform charge
5218 * immigration if @tset contains a leader and whine if there are
5222 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5225 memcg
= mem_cgroup_from_css(css
);
5231 * We are now commited to this value whatever it is. Changes in this
5232 * tunable will only affect upcoming migrations, not the current one.
5233 * So we need to save it, and keep it going.
5235 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5239 from
= mem_cgroup_from_task(p
);
5241 VM_BUG_ON(from
== memcg
);
5243 mm
= get_task_mm(p
);
5246 /* We move charges only when we move a owner of the mm */
5247 if (mm
->owner
== p
) {
5250 VM_BUG_ON(mc
.precharge
);
5251 VM_BUG_ON(mc
.moved_charge
);
5252 VM_BUG_ON(mc
.moved_swap
);
5254 spin_lock(&mc
.lock
);
5258 mc
.flags
= move_flags
;
5259 spin_unlock(&mc
.lock
);
5260 /* We set mc.moving_task later */
5262 ret
= mem_cgroup_precharge_mc(mm
);
5264 mem_cgroup_clear_mc();
5271 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5274 mem_cgroup_clear_mc();
5277 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5278 unsigned long addr
, unsigned long end
,
5279 struct mm_walk
*walk
)
5282 struct vm_area_struct
*vma
= walk
->vma
;
5285 enum mc_target_type target_type
;
5286 union mc_target target
;
5289 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5291 if (mc
.precharge
< HPAGE_PMD_NR
) {
5295 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5296 if (target_type
== MC_TARGET_PAGE
) {
5298 if (!isolate_lru_page(page
)) {
5299 if (!mem_cgroup_move_account(page
, true,
5301 mc
.precharge
-= HPAGE_PMD_NR
;
5302 mc
.moved_charge
+= HPAGE_PMD_NR
;
5304 putback_lru_page(page
);
5307 } else if (target_type
== MC_TARGET_DEVICE
) {
5309 if (!mem_cgroup_move_account(page
, true,
5311 mc
.precharge
-= HPAGE_PMD_NR
;
5312 mc
.moved_charge
+= HPAGE_PMD_NR
;
5320 if (pmd_trans_unstable(pmd
))
5323 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5324 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5325 pte_t ptent
= *(pte
++);
5326 bool device
= false;
5332 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5333 case MC_TARGET_DEVICE
:
5336 case MC_TARGET_PAGE
:
5339 * We can have a part of the split pmd here. Moving it
5340 * can be done but it would be too convoluted so simply
5341 * ignore such a partial THP and keep it in original
5342 * memcg. There should be somebody mapping the head.
5344 if (PageTransCompound(page
))
5346 if (!device
&& isolate_lru_page(page
))
5348 if (!mem_cgroup_move_account(page
, false,
5351 /* we uncharge from mc.from later. */
5355 putback_lru_page(page
);
5356 put
: /* get_mctgt_type() gets the page */
5359 case MC_TARGET_SWAP
:
5361 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5363 /* we fixup refcnts and charges later. */
5371 pte_unmap_unlock(pte
- 1, ptl
);
5376 * We have consumed all precharges we got in can_attach().
5377 * We try charge one by one, but don't do any additional
5378 * charges to mc.to if we have failed in charge once in attach()
5381 ret
= mem_cgroup_do_precharge(1);
5389 static void mem_cgroup_move_charge(void)
5391 struct mm_walk mem_cgroup_move_charge_walk
= {
5392 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5396 lru_add_drain_all();
5398 * Signal lock_page_memcg() to take the memcg's move_lock
5399 * while we're moving its pages to another memcg. Then wait
5400 * for already started RCU-only updates to finish.
5402 atomic_inc(&mc
.from
->moving_account
);
5405 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5407 * Someone who are holding the mmap_sem might be waiting in
5408 * waitq. So we cancel all extra charges, wake up all waiters,
5409 * and retry. Because we cancel precharges, we might not be able
5410 * to move enough charges, but moving charge is a best-effort
5411 * feature anyway, so it wouldn't be a big problem.
5413 __mem_cgroup_clear_mc();
5418 * When we have consumed all precharges and failed in doing
5419 * additional charge, the page walk just aborts.
5421 walk_page_range(0, mc
.mm
->highest_vm_end
, &mem_cgroup_move_charge_walk
);
5423 up_read(&mc
.mm
->mmap_sem
);
5424 atomic_dec(&mc
.from
->moving_account
);
5427 static void mem_cgroup_move_task(void)
5430 mem_cgroup_move_charge();
5431 mem_cgroup_clear_mc();
5434 #else /* !CONFIG_MMU */
5435 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5439 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5442 static void mem_cgroup_move_task(void)
5448 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5449 * to verify whether we're attached to the default hierarchy on each mount
5452 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5455 * use_hierarchy is forced on the default hierarchy. cgroup core
5456 * guarantees that @root doesn't have any children, so turning it
5457 * on for the root memcg is enough.
5459 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5460 root_mem_cgroup
->use_hierarchy
= true;
5462 root_mem_cgroup
->use_hierarchy
= false;
5465 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
5467 if (value
== PAGE_COUNTER_MAX
)
5468 seq_puts(m
, "max\n");
5470 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
5475 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5478 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5480 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5483 static int memory_min_show(struct seq_file
*m
, void *v
)
5485 return seq_puts_memcg_tunable(m
,
5486 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
5489 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
5490 char *buf
, size_t nbytes
, loff_t off
)
5492 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5496 buf
= strstrip(buf
);
5497 err
= page_counter_memparse(buf
, "max", &min
);
5501 page_counter_set_min(&memcg
->memory
, min
);
5506 static int memory_low_show(struct seq_file
*m
, void *v
)
5508 return seq_puts_memcg_tunable(m
,
5509 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
5512 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5513 char *buf
, size_t nbytes
, loff_t off
)
5515 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5519 buf
= strstrip(buf
);
5520 err
= page_counter_memparse(buf
, "max", &low
);
5524 page_counter_set_low(&memcg
->memory
, low
);
5529 static int memory_high_show(struct seq_file
*m
, void *v
)
5531 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
5534 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5535 char *buf
, size_t nbytes
, loff_t off
)
5537 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5538 unsigned long nr_pages
;
5542 buf
= strstrip(buf
);
5543 err
= page_counter_memparse(buf
, "max", &high
);
5549 nr_pages
= page_counter_read(&memcg
->memory
);
5550 if (nr_pages
> high
)
5551 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5554 memcg_wb_domain_size_changed(memcg
);
5558 static int memory_max_show(struct seq_file
*m
, void *v
)
5560 return seq_puts_memcg_tunable(m
,
5561 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
5564 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5565 char *buf
, size_t nbytes
, loff_t off
)
5567 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5568 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5569 bool drained
= false;
5573 buf
= strstrip(buf
);
5574 err
= page_counter_memparse(buf
, "max", &max
);
5578 xchg(&memcg
->memory
.max
, max
);
5581 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5583 if (nr_pages
<= max
)
5586 if (signal_pending(current
)) {
5592 drain_all_stock(memcg
);
5598 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
5604 memcg_memory_event(memcg
, MEMCG_OOM
);
5605 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
5609 memcg_wb_domain_size_changed(memcg
);
5613 static int memory_events_show(struct seq_file
*m
, void *v
)
5615 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5617 seq_printf(m
, "low %lu\n",
5618 atomic_long_read(&memcg
->memory_events
[MEMCG_LOW
]));
5619 seq_printf(m
, "high %lu\n",
5620 atomic_long_read(&memcg
->memory_events
[MEMCG_HIGH
]));
5621 seq_printf(m
, "max %lu\n",
5622 atomic_long_read(&memcg
->memory_events
[MEMCG_MAX
]));
5623 seq_printf(m
, "oom %lu\n",
5624 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM
]));
5625 seq_printf(m
, "oom_kill %lu\n",
5626 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
5631 static int memory_stat_show(struct seq_file
*m
, void *v
)
5633 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5637 * Provide statistics on the state of the memory subsystem as
5638 * well as cumulative event counters that show past behavior.
5640 * This list is ordered following a combination of these gradients:
5641 * 1) generic big picture -> specifics and details
5642 * 2) reflecting userspace activity -> reflecting kernel heuristics
5644 * Current memory state:
5647 seq_printf(m
, "anon %llu\n",
5648 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) * PAGE_SIZE
);
5649 seq_printf(m
, "file %llu\n",
5650 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) * PAGE_SIZE
);
5651 seq_printf(m
, "kernel_stack %llu\n",
5652 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) * 1024);
5653 seq_printf(m
, "slab %llu\n",
5654 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
5655 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
5657 seq_printf(m
, "sock %llu\n",
5658 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) * PAGE_SIZE
);
5660 seq_printf(m
, "shmem %llu\n",
5661 (u64
)memcg_page_state(memcg
, NR_SHMEM
) * PAGE_SIZE
);
5662 seq_printf(m
, "file_mapped %llu\n",
5663 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) * PAGE_SIZE
);
5664 seq_printf(m
, "file_dirty %llu\n",
5665 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) * PAGE_SIZE
);
5666 seq_printf(m
, "file_writeback %llu\n",
5667 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) * PAGE_SIZE
);
5670 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
5671 * with the NR_ANON_THP vm counter, but right now it's a pain in the
5672 * arse because it requires migrating the work out of rmap to a place
5673 * where the page->mem_cgroup is set up and stable.
5675 seq_printf(m
, "anon_thp %llu\n",
5676 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) * PAGE_SIZE
);
5678 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5679 seq_printf(m
, "%s %llu\n", mem_cgroup_lru_names
[i
],
5680 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
5683 seq_printf(m
, "slab_reclaimable %llu\n",
5684 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
5686 seq_printf(m
, "slab_unreclaimable %llu\n",
5687 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
5690 /* Accumulated memory events */
5692 seq_printf(m
, "pgfault %lu\n", memcg_events(memcg
, PGFAULT
));
5693 seq_printf(m
, "pgmajfault %lu\n", memcg_events(memcg
, PGMAJFAULT
));
5695 seq_printf(m
, "workingset_refault %lu\n",
5696 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
5697 seq_printf(m
, "workingset_activate %lu\n",
5698 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
5699 seq_printf(m
, "workingset_nodereclaim %lu\n",
5700 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
5702 seq_printf(m
, "pgrefill %lu\n", memcg_events(memcg
, PGREFILL
));
5703 seq_printf(m
, "pgscan %lu\n", memcg_events(memcg
, PGSCAN_KSWAPD
) +
5704 memcg_events(memcg
, PGSCAN_DIRECT
));
5705 seq_printf(m
, "pgsteal %lu\n", memcg_events(memcg
, PGSTEAL_KSWAPD
) +
5706 memcg_events(memcg
, PGSTEAL_DIRECT
));
5707 seq_printf(m
, "pgactivate %lu\n", memcg_events(memcg
, PGACTIVATE
));
5708 seq_printf(m
, "pgdeactivate %lu\n", memcg_events(memcg
, PGDEACTIVATE
));
5709 seq_printf(m
, "pglazyfree %lu\n", memcg_events(memcg
, PGLAZYFREE
));
5710 seq_printf(m
, "pglazyfreed %lu\n", memcg_events(memcg
, PGLAZYFREED
));
5712 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5713 seq_printf(m
, "thp_fault_alloc %lu\n",
5714 memcg_events(memcg
, THP_FAULT_ALLOC
));
5715 seq_printf(m
, "thp_collapse_alloc %lu\n",
5716 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
5717 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
5722 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
5724 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
5726 seq_printf(m
, "%d\n", memcg
->oom_group
);
5731 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
5732 char *buf
, size_t nbytes
, loff_t off
)
5734 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5737 buf
= strstrip(buf
);
5741 ret
= kstrtoint(buf
, 0, &oom_group
);
5745 if (oom_group
!= 0 && oom_group
!= 1)
5748 memcg
->oom_group
= oom_group
;
5753 static struct cftype memory_files
[] = {
5756 .flags
= CFTYPE_NOT_ON_ROOT
,
5757 .read_u64
= memory_current_read
,
5761 .flags
= CFTYPE_NOT_ON_ROOT
,
5762 .seq_show
= memory_min_show
,
5763 .write
= memory_min_write
,
5767 .flags
= CFTYPE_NOT_ON_ROOT
,
5768 .seq_show
= memory_low_show
,
5769 .write
= memory_low_write
,
5773 .flags
= CFTYPE_NOT_ON_ROOT
,
5774 .seq_show
= memory_high_show
,
5775 .write
= memory_high_write
,
5779 .flags
= CFTYPE_NOT_ON_ROOT
,
5780 .seq_show
= memory_max_show
,
5781 .write
= memory_max_write
,
5785 .flags
= CFTYPE_NOT_ON_ROOT
,
5786 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5787 .seq_show
= memory_events_show
,
5791 .flags
= CFTYPE_NOT_ON_ROOT
,
5792 .seq_show
= memory_stat_show
,
5795 .name
= "oom.group",
5796 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
5797 .seq_show
= memory_oom_group_show
,
5798 .write
= memory_oom_group_write
,
5803 struct cgroup_subsys memory_cgrp_subsys
= {
5804 .css_alloc
= mem_cgroup_css_alloc
,
5805 .css_online
= mem_cgroup_css_online
,
5806 .css_offline
= mem_cgroup_css_offline
,
5807 .css_released
= mem_cgroup_css_released
,
5808 .css_free
= mem_cgroup_css_free
,
5809 .css_reset
= mem_cgroup_css_reset
,
5810 .can_attach
= mem_cgroup_can_attach
,
5811 .cancel_attach
= mem_cgroup_cancel_attach
,
5812 .post_attach
= mem_cgroup_move_task
,
5813 .bind
= mem_cgroup_bind
,
5814 .dfl_cftypes
= memory_files
,
5815 .legacy_cftypes
= mem_cgroup_legacy_files
,
5820 * mem_cgroup_protected - check if memory consumption is in the normal range
5821 * @root: the top ancestor of the sub-tree being checked
5822 * @memcg: the memory cgroup to check
5824 * WARNING: This function is not stateless! It can only be used as part
5825 * of a top-down tree iteration, not for isolated queries.
5827 * Returns one of the following:
5828 * MEMCG_PROT_NONE: cgroup memory is not protected
5829 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5830 * an unprotected supply of reclaimable memory from other cgroups.
5831 * MEMCG_PROT_MIN: cgroup memory is protected
5833 * @root is exclusive; it is never protected when looked at directly
5835 * To provide a proper hierarchical behavior, effective memory.min/low values
5836 * are used. Below is the description of how effective memory.low is calculated.
5837 * Effective memory.min values is calculated in the same way.
5839 * Effective memory.low is always equal or less than the original memory.low.
5840 * If there is no memory.low overcommittment (which is always true for
5841 * top-level memory cgroups), these two values are equal.
5842 * Otherwise, it's a part of parent's effective memory.low,
5843 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5844 * memory.low usages, where memory.low usage is the size of actually
5848 * elow = min( memory.low, parent->elow * ------------------ ),
5849 * siblings_low_usage
5851 * | memory.current, if memory.current < memory.low
5856 * Such definition of the effective memory.low provides the expected
5857 * hierarchical behavior: parent's memory.low value is limiting
5858 * children, unprotected memory is reclaimed first and cgroups,
5859 * which are not using their guarantee do not affect actual memory
5862 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5864 * A A/memory.low = 2G, A/memory.current = 6G
5866 * BC DE B/memory.low = 3G B/memory.current = 2G
5867 * C/memory.low = 1G C/memory.current = 2G
5868 * D/memory.low = 0 D/memory.current = 2G
5869 * E/memory.low = 10G E/memory.current = 0
5871 * and the memory pressure is applied, the following memory distribution
5872 * is expected (approximately):
5874 * A/memory.current = 2G
5876 * B/memory.current = 1.3G
5877 * C/memory.current = 0.6G
5878 * D/memory.current = 0
5879 * E/memory.current = 0
5881 * These calculations require constant tracking of the actual low usages
5882 * (see propagate_protected_usage()), as well as recursive calculation of
5883 * effective memory.low values. But as we do call mem_cgroup_protected()
5884 * path for each memory cgroup top-down from the reclaim,
5885 * it's possible to optimize this part, and save calculated elow
5886 * for next usage. This part is intentionally racy, but it's ok,
5887 * as memory.low is a best-effort mechanism.
5889 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
5890 struct mem_cgroup
*memcg
)
5892 struct mem_cgroup
*parent
;
5893 unsigned long emin
, parent_emin
;
5894 unsigned long elow
, parent_elow
;
5895 unsigned long usage
;
5897 if (mem_cgroup_disabled())
5898 return MEMCG_PROT_NONE
;
5901 root
= root_mem_cgroup
;
5903 return MEMCG_PROT_NONE
;
5905 usage
= page_counter_read(&memcg
->memory
);
5907 return MEMCG_PROT_NONE
;
5909 emin
= memcg
->memory
.min
;
5910 elow
= memcg
->memory
.low
;
5912 parent
= parent_mem_cgroup(memcg
);
5913 /* No parent means a non-hierarchical mode on v1 memcg */
5915 return MEMCG_PROT_NONE
;
5920 parent_emin
= READ_ONCE(parent
->memory
.emin
);
5921 emin
= min(emin
, parent_emin
);
5922 if (emin
&& parent_emin
) {
5923 unsigned long min_usage
, siblings_min_usage
;
5925 min_usage
= min(usage
, memcg
->memory
.min
);
5926 siblings_min_usage
= atomic_long_read(
5927 &parent
->memory
.children_min_usage
);
5929 if (min_usage
&& siblings_min_usage
)
5930 emin
= min(emin
, parent_emin
* min_usage
/
5931 siblings_min_usage
);
5934 parent_elow
= READ_ONCE(parent
->memory
.elow
);
5935 elow
= min(elow
, parent_elow
);
5936 if (elow
&& parent_elow
) {
5937 unsigned long low_usage
, siblings_low_usage
;
5939 low_usage
= min(usage
, memcg
->memory
.low
);
5940 siblings_low_usage
= atomic_long_read(
5941 &parent
->memory
.children_low_usage
);
5943 if (low_usage
&& siblings_low_usage
)
5944 elow
= min(elow
, parent_elow
* low_usage
/
5945 siblings_low_usage
);
5949 memcg
->memory
.emin
= emin
;
5950 memcg
->memory
.elow
= elow
;
5953 return MEMCG_PROT_MIN
;
5954 else if (usage
<= elow
)
5955 return MEMCG_PROT_LOW
;
5957 return MEMCG_PROT_NONE
;
5961 * mem_cgroup_try_charge - try charging a page
5962 * @page: page to charge
5963 * @mm: mm context of the victim
5964 * @gfp_mask: reclaim mode
5965 * @memcgp: charged memcg return
5966 * @compound: charge the page as compound or small page
5968 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5969 * pages according to @gfp_mask if necessary.
5971 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5972 * Otherwise, an error code is returned.
5974 * After page->mapping has been set up, the caller must finalize the
5975 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5976 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5978 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5979 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5982 struct mem_cgroup
*memcg
= NULL
;
5983 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5986 if (mem_cgroup_disabled())
5989 if (PageSwapCache(page
)) {
5991 * Every swap fault against a single page tries to charge the
5992 * page, bail as early as possible. shmem_unuse() encounters
5993 * already charged pages, too. The USED bit is protected by
5994 * the page lock, which serializes swap cache removal, which
5995 * in turn serializes uncharging.
5997 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5998 if (compound_head(page
)->mem_cgroup
)
6001 if (do_swap_account
) {
6002 swp_entry_t ent
= { .val
= page_private(page
), };
6003 unsigned short id
= lookup_swap_cgroup_id(ent
);
6006 memcg
= mem_cgroup_from_id(id
);
6007 if (memcg
&& !css_tryget_online(&memcg
->css
))
6014 memcg
= get_mem_cgroup_from_mm(mm
);
6016 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6018 css_put(&memcg
->css
);
6024 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6025 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6028 struct mem_cgroup
*memcg
;
6031 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6033 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6038 * mem_cgroup_commit_charge - commit a page charge
6039 * @page: page to charge
6040 * @memcg: memcg to charge the page to
6041 * @lrucare: page might be on LRU already
6042 * @compound: charge the page as compound or small page
6044 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6045 * after page->mapping has been set up. This must happen atomically
6046 * as part of the page instantiation, i.e. under the page table lock
6047 * for anonymous pages, under the page lock for page and swap cache.
6049 * In addition, the page must not be on the LRU during the commit, to
6050 * prevent racing with task migration. If it might be, use @lrucare.
6052 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6054 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6055 bool lrucare
, bool compound
)
6057 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6059 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6060 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6062 if (mem_cgroup_disabled())
6065 * Swap faults will attempt to charge the same page multiple
6066 * times. But reuse_swap_page() might have removed the page
6067 * from swapcache already, so we can't check PageSwapCache().
6072 commit_charge(page
, memcg
, lrucare
);
6074 local_irq_disable();
6075 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6076 memcg_check_events(memcg
, page
);
6079 if (do_memsw_account() && PageSwapCache(page
)) {
6080 swp_entry_t entry
= { .val
= page_private(page
) };
6082 * The swap entry might not get freed for a long time,
6083 * let's not wait for it. The page already received a
6084 * memory+swap charge, drop the swap entry duplicate.
6086 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6091 * mem_cgroup_cancel_charge - cancel a page charge
6092 * @page: page to charge
6093 * @memcg: memcg to charge the page to
6094 * @compound: charge the page as compound or small page
6096 * Cancel a charge transaction started by mem_cgroup_try_charge().
6098 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6101 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6103 if (mem_cgroup_disabled())
6106 * Swap faults will attempt to charge the same page multiple
6107 * times. But reuse_swap_page() might have removed the page
6108 * from swapcache already, so we can't check PageSwapCache().
6113 cancel_charge(memcg
, nr_pages
);
6116 struct uncharge_gather
{
6117 struct mem_cgroup
*memcg
;
6118 unsigned long pgpgout
;
6119 unsigned long nr_anon
;
6120 unsigned long nr_file
;
6121 unsigned long nr_kmem
;
6122 unsigned long nr_huge
;
6123 unsigned long nr_shmem
;
6124 struct page
*dummy_page
;
6127 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6129 memset(ug
, 0, sizeof(*ug
));
6132 static void uncharge_batch(const struct uncharge_gather
*ug
)
6134 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6135 unsigned long flags
;
6137 if (!mem_cgroup_is_root(ug
->memcg
)) {
6138 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6139 if (do_memsw_account())
6140 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6141 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6142 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6143 memcg_oom_recover(ug
->memcg
);
6146 local_irq_save(flags
);
6147 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6148 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6149 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6150 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6151 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6152 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6153 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6154 local_irq_restore(flags
);
6156 if (!mem_cgroup_is_root(ug
->memcg
))
6157 css_put_many(&ug
->memcg
->css
, nr_pages
);
6160 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6162 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6163 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6164 !PageHWPoison(page
) , page
);
6166 if (!page
->mem_cgroup
)
6170 * Nobody should be changing or seriously looking at
6171 * page->mem_cgroup at this point, we have fully
6172 * exclusive access to the page.
6175 if (ug
->memcg
!= page
->mem_cgroup
) {
6178 uncharge_gather_clear(ug
);
6180 ug
->memcg
= page
->mem_cgroup
;
6183 if (!PageKmemcg(page
)) {
6184 unsigned int nr_pages
= 1;
6186 if (PageTransHuge(page
)) {
6187 nr_pages
<<= compound_order(page
);
6188 ug
->nr_huge
+= nr_pages
;
6191 ug
->nr_anon
+= nr_pages
;
6193 ug
->nr_file
+= nr_pages
;
6194 if (PageSwapBacked(page
))
6195 ug
->nr_shmem
+= nr_pages
;
6199 ug
->nr_kmem
+= 1 << compound_order(page
);
6200 __ClearPageKmemcg(page
);
6203 ug
->dummy_page
= page
;
6204 page
->mem_cgroup
= NULL
;
6207 static void uncharge_list(struct list_head
*page_list
)
6209 struct uncharge_gather ug
;
6210 struct list_head
*next
;
6212 uncharge_gather_clear(&ug
);
6215 * Note that the list can be a single page->lru; hence the
6216 * do-while loop instead of a simple list_for_each_entry().
6218 next
= page_list
->next
;
6222 page
= list_entry(next
, struct page
, lru
);
6223 next
= page
->lru
.next
;
6225 uncharge_page(page
, &ug
);
6226 } while (next
!= page_list
);
6229 uncharge_batch(&ug
);
6233 * mem_cgroup_uncharge - uncharge a page
6234 * @page: page to uncharge
6236 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6237 * mem_cgroup_commit_charge().
6239 void mem_cgroup_uncharge(struct page
*page
)
6241 struct uncharge_gather ug
;
6243 if (mem_cgroup_disabled())
6246 /* Don't touch page->lru of any random page, pre-check: */
6247 if (!page
->mem_cgroup
)
6250 uncharge_gather_clear(&ug
);
6251 uncharge_page(page
, &ug
);
6252 uncharge_batch(&ug
);
6256 * mem_cgroup_uncharge_list - uncharge a list of page
6257 * @page_list: list of pages to uncharge
6259 * Uncharge a list of pages previously charged with
6260 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6262 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6264 if (mem_cgroup_disabled())
6267 if (!list_empty(page_list
))
6268 uncharge_list(page_list
);
6272 * mem_cgroup_migrate - charge a page's replacement
6273 * @oldpage: currently circulating page
6274 * @newpage: replacement page
6276 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6277 * be uncharged upon free.
6279 * Both pages must be locked, @newpage->mapping must be set up.
6281 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6283 struct mem_cgroup
*memcg
;
6284 unsigned int nr_pages
;
6286 unsigned long flags
;
6288 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6289 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6290 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6291 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6294 if (mem_cgroup_disabled())
6297 /* Page cache replacement: new page already charged? */
6298 if (newpage
->mem_cgroup
)
6301 /* Swapcache readahead pages can get replaced before being charged */
6302 memcg
= oldpage
->mem_cgroup
;
6306 /* Force-charge the new page. The old one will be freed soon */
6307 compound
= PageTransHuge(newpage
);
6308 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
6310 page_counter_charge(&memcg
->memory
, nr_pages
);
6311 if (do_memsw_account())
6312 page_counter_charge(&memcg
->memsw
, nr_pages
);
6313 css_get_many(&memcg
->css
, nr_pages
);
6315 commit_charge(newpage
, memcg
, false);
6317 local_irq_save(flags
);
6318 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
6319 memcg_check_events(memcg
, newpage
);
6320 local_irq_restore(flags
);
6323 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6324 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6326 void mem_cgroup_sk_alloc(struct sock
*sk
)
6328 struct mem_cgroup
*memcg
;
6330 if (!mem_cgroup_sockets_enabled
)
6334 * Socket cloning can throw us here with sk_memcg already
6335 * filled. It won't however, necessarily happen from
6336 * process context. So the test for root memcg given
6337 * the current task's memcg won't help us in this case.
6339 * Respecting the original socket's memcg is a better
6340 * decision in this case.
6343 css_get(&sk
->sk_memcg
->css
);
6348 memcg
= mem_cgroup_from_task(current
);
6349 if (memcg
== root_mem_cgroup
)
6351 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6353 if (css_tryget_online(&memcg
->css
))
6354 sk
->sk_memcg
= memcg
;
6359 void mem_cgroup_sk_free(struct sock
*sk
)
6362 css_put(&sk
->sk_memcg
->css
);
6366 * mem_cgroup_charge_skmem - charge socket memory
6367 * @memcg: memcg to charge
6368 * @nr_pages: number of pages to charge
6370 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6371 * @memcg's configured limit, %false if the charge had to be forced.
6373 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6375 gfp_t gfp_mask
= GFP_KERNEL
;
6377 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6378 struct page_counter
*fail
;
6380 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6381 memcg
->tcpmem_pressure
= 0;
6384 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6385 memcg
->tcpmem_pressure
= 1;
6389 /* Don't block in the packet receive path */
6391 gfp_mask
= GFP_NOWAIT
;
6393 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6395 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6398 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6403 * mem_cgroup_uncharge_skmem - uncharge socket memory
6404 * @memcg: memcg to uncharge
6405 * @nr_pages: number of pages to uncharge
6407 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6409 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6410 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6414 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6416 refill_stock(memcg
, nr_pages
);
6419 static int __init
cgroup_memory(char *s
)
6423 while ((token
= strsep(&s
, ",")) != NULL
) {
6426 if (!strcmp(token
, "nosocket"))
6427 cgroup_memory_nosocket
= true;
6428 if (!strcmp(token
, "nokmem"))
6429 cgroup_memory_nokmem
= true;
6433 __setup("cgroup.memory=", cgroup_memory
);
6436 * subsys_initcall() for memory controller.
6438 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6439 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6440 * basically everything that doesn't depend on a specific mem_cgroup structure
6441 * should be initialized from here.
6443 static int __init
mem_cgroup_init(void)
6447 #ifdef CONFIG_MEMCG_KMEM
6449 * Kmem cache creation is mostly done with the slab_mutex held,
6450 * so use a workqueue with limited concurrency to avoid stalling
6451 * all worker threads in case lots of cgroups are created and
6452 * destroyed simultaneously.
6454 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6455 BUG_ON(!memcg_kmem_cache_wq
);
6458 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6459 memcg_hotplug_cpu_dead
);
6461 for_each_possible_cpu(cpu
)
6462 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6465 for_each_node(node
) {
6466 struct mem_cgroup_tree_per_node
*rtpn
;
6468 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6469 node_online(node
) ? node
: NUMA_NO_NODE
);
6471 rtpn
->rb_root
= RB_ROOT
;
6472 rtpn
->rb_rightmost
= NULL
;
6473 spin_lock_init(&rtpn
->lock
);
6474 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6479 subsys_initcall(mem_cgroup_init
);
6481 #ifdef CONFIG_MEMCG_SWAP
6482 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6484 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6486 * The root cgroup cannot be destroyed, so it's refcount must
6489 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6493 memcg
= parent_mem_cgroup(memcg
);
6495 memcg
= root_mem_cgroup
;
6501 * mem_cgroup_swapout - transfer a memsw charge to swap
6502 * @page: page whose memsw charge to transfer
6503 * @entry: swap entry to move the charge to
6505 * Transfer the memsw charge of @page to @entry.
6507 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6509 struct mem_cgroup
*memcg
, *swap_memcg
;
6510 unsigned int nr_entries
;
6511 unsigned short oldid
;
6513 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6514 VM_BUG_ON_PAGE(page_count(page
), page
);
6516 if (!do_memsw_account())
6519 memcg
= page
->mem_cgroup
;
6521 /* Readahead page, never charged */
6526 * In case the memcg owning these pages has been offlined and doesn't
6527 * have an ID allocated to it anymore, charge the closest online
6528 * ancestor for the swap instead and transfer the memory+swap charge.
6530 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6531 nr_entries
= hpage_nr_pages(page
);
6532 /* Get references for the tail pages, too */
6534 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6535 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6537 VM_BUG_ON_PAGE(oldid
, page
);
6538 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6540 page
->mem_cgroup
= NULL
;
6542 if (!mem_cgroup_is_root(memcg
))
6543 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6545 if (memcg
!= swap_memcg
) {
6546 if (!mem_cgroup_is_root(swap_memcg
))
6547 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6548 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6552 * Interrupts should be disabled here because the caller holds the
6553 * i_pages lock which is taken with interrupts-off. It is
6554 * important here to have the interrupts disabled because it is the
6555 * only synchronisation we have for updating the per-CPU variables.
6557 VM_BUG_ON(!irqs_disabled());
6558 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6560 memcg_check_events(memcg
, page
);
6562 if (!mem_cgroup_is_root(memcg
))
6563 css_put_many(&memcg
->css
, nr_entries
);
6567 * mem_cgroup_try_charge_swap - try charging swap space for a page
6568 * @page: page being added to swap
6569 * @entry: swap entry to charge
6571 * Try to charge @page's memcg for the swap space at @entry.
6573 * Returns 0 on success, -ENOMEM on failure.
6575 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6577 unsigned int nr_pages
= hpage_nr_pages(page
);
6578 struct page_counter
*counter
;
6579 struct mem_cgroup
*memcg
;
6580 unsigned short oldid
;
6582 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6585 memcg
= page
->mem_cgroup
;
6587 /* Readahead page, never charged */
6592 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6596 memcg
= mem_cgroup_id_get_online(memcg
);
6598 if (!mem_cgroup_is_root(memcg
) &&
6599 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
6600 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
6601 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6602 mem_cgroup_id_put(memcg
);
6606 /* Get references for the tail pages, too */
6608 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
6609 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
6610 VM_BUG_ON_PAGE(oldid
, page
);
6611 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
6617 * mem_cgroup_uncharge_swap - uncharge swap space
6618 * @entry: swap entry to uncharge
6619 * @nr_pages: the amount of swap space to uncharge
6621 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
6623 struct mem_cgroup
*memcg
;
6626 if (!do_swap_account
)
6629 id
= swap_cgroup_record(entry
, 0, nr_pages
);
6631 memcg
= mem_cgroup_from_id(id
);
6633 if (!mem_cgroup_is_root(memcg
)) {
6634 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6635 page_counter_uncharge(&memcg
->swap
, nr_pages
);
6637 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
6639 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
6640 mem_cgroup_id_put_many(memcg
, nr_pages
);
6645 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
6647 long nr_swap_pages
= get_nr_swap_pages();
6649 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6650 return nr_swap_pages
;
6651 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6652 nr_swap_pages
= min_t(long, nr_swap_pages
,
6653 READ_ONCE(memcg
->swap
.max
) -
6654 page_counter_read(&memcg
->swap
));
6655 return nr_swap_pages
;
6658 bool mem_cgroup_swap_full(struct page
*page
)
6660 struct mem_cgroup
*memcg
;
6662 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6666 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6669 memcg
= page
->mem_cgroup
;
6673 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6674 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
6680 /* for remember boot option*/
6681 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6682 static int really_do_swap_account __initdata
= 1;
6684 static int really_do_swap_account __initdata
;
6687 static int __init
enable_swap_account(char *s
)
6689 if (!strcmp(s
, "1"))
6690 really_do_swap_account
= 1;
6691 else if (!strcmp(s
, "0"))
6692 really_do_swap_account
= 0;
6695 __setup("swapaccount=", enable_swap_account
);
6697 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
6700 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6702 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
6705 static int swap_max_show(struct seq_file
*m
, void *v
)
6707 return seq_puts_memcg_tunable(m
,
6708 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
6711 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
6712 char *buf
, size_t nbytes
, loff_t off
)
6714 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6718 buf
= strstrip(buf
);
6719 err
= page_counter_memparse(buf
, "max", &max
);
6723 xchg(&memcg
->swap
.max
, max
);
6728 static int swap_events_show(struct seq_file
*m
, void *v
)
6730 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6732 seq_printf(m
, "max %lu\n",
6733 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
6734 seq_printf(m
, "fail %lu\n",
6735 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
6740 static struct cftype swap_files
[] = {
6742 .name
= "swap.current",
6743 .flags
= CFTYPE_NOT_ON_ROOT
,
6744 .read_u64
= swap_current_read
,
6748 .flags
= CFTYPE_NOT_ON_ROOT
,
6749 .seq_show
= swap_max_show
,
6750 .write
= swap_max_write
,
6753 .name
= "swap.events",
6754 .flags
= CFTYPE_NOT_ON_ROOT
,
6755 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
6756 .seq_show
= swap_events_show
,
6761 static struct cftype memsw_cgroup_files
[] = {
6763 .name
= "memsw.usage_in_bytes",
6764 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6765 .read_u64
= mem_cgroup_read_u64
,
6768 .name
= "memsw.max_usage_in_bytes",
6769 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6770 .write
= mem_cgroup_reset
,
6771 .read_u64
= mem_cgroup_read_u64
,
6774 .name
= "memsw.limit_in_bytes",
6775 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6776 .write
= mem_cgroup_write
,
6777 .read_u64
= mem_cgroup_read_u64
,
6780 .name
= "memsw.failcnt",
6781 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6782 .write
= mem_cgroup_reset
,
6783 .read_u64
= mem_cgroup_read_u64
,
6785 { }, /* terminate */
6788 static int __init
mem_cgroup_swap_init(void)
6790 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6791 do_swap_account
= 1;
6792 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
6794 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6795 memsw_cgroup_files
));
6799 subsys_initcall(mem_cgroup_swap_init
);
6801 #endif /* CONFIG_MEMCG_SWAP */