1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
72 EXPORT_SYMBOL(memory_cgrp_subsys
);
74 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
76 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket
;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem
;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly
;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char *const mem_cgroup_lru_names
[] = {
109 #define THRESHOLDS_EVENTS_TARGET 128
110 #define SOFTLIMIT_EVENTS_TARGET 1024
111 #define NUMAINFO_EVENTS_TARGET 1024
114 * Cgroups above their limits are maintained in a RB-Tree, independent of
115 * their hierarchy representation
118 struct mem_cgroup_tree_per_node
{
119 struct rb_root rb_root
;
120 struct rb_node
*rb_rightmost
;
124 struct mem_cgroup_tree
{
125 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
128 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
131 struct mem_cgroup_eventfd_list
{
132 struct list_head list
;
133 struct eventfd_ctx
*eventfd
;
137 * cgroup_event represents events which userspace want to receive.
139 struct mem_cgroup_event
{
141 * memcg which the event belongs to.
143 struct mem_cgroup
*memcg
;
145 * eventfd to signal userspace about the event.
147 struct eventfd_ctx
*eventfd
;
149 * Each of these stored in a list by the cgroup.
151 struct list_head list
;
153 * register_event() callback will be used to add new userspace
154 * waiter for changes related to this event. Use eventfd_signal()
155 * on eventfd to send notification to userspace.
157 int (*register_event
)(struct mem_cgroup
*memcg
,
158 struct eventfd_ctx
*eventfd
, const char *args
);
160 * unregister_event() callback will be called when userspace closes
161 * the eventfd or on cgroup removing. This callback must be set,
162 * if you want provide notification functionality.
164 void (*unregister_event
)(struct mem_cgroup
*memcg
,
165 struct eventfd_ctx
*eventfd
);
167 * All fields below needed to unregister event when
168 * userspace closes eventfd.
171 wait_queue_head_t
*wqh
;
172 wait_queue_entry_t wait
;
173 struct work_struct remove
;
176 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
177 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
179 /* Stuffs for move charges at task migration. */
181 * Types of charges to be moved.
183 #define MOVE_ANON 0x1U
184 #define MOVE_FILE 0x2U
185 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
187 /* "mc" and its members are protected by cgroup_mutex */
188 static struct move_charge_struct
{
189 spinlock_t lock
; /* for from, to */
190 struct mm_struct
*mm
;
191 struct mem_cgroup
*from
;
192 struct mem_cgroup
*to
;
194 unsigned long precharge
;
195 unsigned long moved_charge
;
196 unsigned long moved_swap
;
197 struct task_struct
*moving_task
; /* a task moving charges */
198 wait_queue_head_t waitq
; /* a waitq for other context */
200 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
201 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
205 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
206 * limit reclaim to prevent infinite loops, if they ever occur.
208 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
209 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
212 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
213 MEM_CGROUP_CHARGE_TYPE_ANON
,
214 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
215 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
219 /* for encoding cft->private value on file */
228 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
229 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
230 #define MEMFILE_ATTR(val) ((val) & 0xffff)
231 /* Used for OOM nofiier */
232 #define OOM_CONTROL (0)
235 * Iteration constructs for visiting all cgroups (under a tree). If
236 * loops are exited prematurely (break), mem_cgroup_iter_break() must
237 * be used for reference counting.
239 #define for_each_mem_cgroup_tree(iter, root) \
240 for (iter = mem_cgroup_iter(root, NULL, NULL); \
242 iter = mem_cgroup_iter(root, iter, NULL))
244 #define for_each_mem_cgroup(iter) \
245 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
247 iter = mem_cgroup_iter(NULL, iter, NULL))
249 static inline bool should_force_charge(void)
251 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
252 (current
->flags
& PF_EXITING
);
255 /* Some nice accessors for the vmpressure. */
256 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
259 memcg
= root_mem_cgroup
;
260 return &memcg
->vmpressure
;
263 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
265 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
268 #ifdef CONFIG_MEMCG_KMEM
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida
);
281 int memcg_nr_cache_ids
;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem
);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem
);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem
);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
320 struct workqueue_struct
*memcg_kmem_cache_wq
;
323 static int memcg_shrinker_map_size
;
324 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
326 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
328 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
331 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
332 int size
, int old_size
)
334 struct memcg_shrinker_map
*new, *old
;
337 lockdep_assert_held(&memcg_shrinker_map_mutex
);
340 old
= rcu_dereference_protected(
341 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
342 /* Not yet online memcg */
346 new = kvmalloc(sizeof(*new) + size
, GFP_KERNEL
);
350 /* Set all old bits, clear all new bits */
351 memset(new->map
, (int)0xff, old_size
);
352 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
354 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
355 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
361 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
363 struct mem_cgroup_per_node
*pn
;
364 struct memcg_shrinker_map
*map
;
367 if (mem_cgroup_is_root(memcg
))
371 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
372 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
375 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
379 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
381 struct memcg_shrinker_map
*map
;
382 int nid
, size
, ret
= 0;
384 if (mem_cgroup_is_root(memcg
))
387 mutex_lock(&memcg_shrinker_map_mutex
);
388 size
= memcg_shrinker_map_size
;
390 map
= kvzalloc(sizeof(*map
) + size
, GFP_KERNEL
);
392 memcg_free_shrinker_maps(memcg
);
396 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
398 mutex_unlock(&memcg_shrinker_map_mutex
);
403 int memcg_expand_shrinker_maps(int new_id
)
405 int size
, old_size
, ret
= 0;
406 struct mem_cgroup
*memcg
;
408 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
409 old_size
= memcg_shrinker_map_size
;
410 if (size
<= old_size
)
413 mutex_lock(&memcg_shrinker_map_mutex
);
414 if (!root_mem_cgroup
)
417 for_each_mem_cgroup(memcg
) {
418 if (mem_cgroup_is_root(memcg
))
420 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
426 memcg_shrinker_map_size
= size
;
427 mutex_unlock(&memcg_shrinker_map_mutex
);
431 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
433 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
434 struct memcg_shrinker_map
*map
;
437 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
438 /* Pairs with smp mb in shrink_slab() */
439 smp_mb__before_atomic();
440 set_bit(shrinker_id
, map
->map
);
446 * mem_cgroup_css_from_page - css of the memcg associated with a page
447 * @page: page of interest
449 * If memcg is bound to the default hierarchy, css of the memcg associated
450 * with @page is returned. The returned css remains associated with @page
451 * until it is released.
453 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
456 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
458 struct mem_cgroup
*memcg
;
460 memcg
= page
->mem_cgroup
;
462 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
463 memcg
= root_mem_cgroup
;
469 * page_cgroup_ino - return inode number of the memcg a page is charged to
472 * Look up the closest online ancestor of the memory cgroup @page is charged to
473 * and return its inode number or 0 if @page is not charged to any cgroup. It
474 * is safe to call this function without holding a reference to @page.
476 * Note, this function is inherently racy, because there is nothing to prevent
477 * the cgroup inode from getting torn down and potentially reallocated a moment
478 * after page_cgroup_ino() returns, so it only should be used by callers that
479 * do not care (such as procfs interfaces).
481 ino_t
page_cgroup_ino(struct page
*page
)
483 struct mem_cgroup
*memcg
;
484 unsigned long ino
= 0;
487 if (PageHead(page
) && PageSlab(page
))
488 memcg
= memcg_from_slab_page(page
);
490 memcg
= READ_ONCE(page
->mem_cgroup
);
491 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
492 memcg
= parent_mem_cgroup(memcg
);
494 ino
= cgroup_ino(memcg
->css
.cgroup
);
499 static struct mem_cgroup_per_node
*
500 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
502 int nid
= page_to_nid(page
);
504 return memcg
->nodeinfo
[nid
];
507 static struct mem_cgroup_tree_per_node
*
508 soft_limit_tree_node(int nid
)
510 return soft_limit_tree
.rb_tree_per_node
[nid
];
513 static struct mem_cgroup_tree_per_node
*
514 soft_limit_tree_from_page(struct page
*page
)
516 int nid
= page_to_nid(page
);
518 return soft_limit_tree
.rb_tree_per_node
[nid
];
521 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
522 struct mem_cgroup_tree_per_node
*mctz
,
523 unsigned long new_usage_in_excess
)
525 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
526 struct rb_node
*parent
= NULL
;
527 struct mem_cgroup_per_node
*mz_node
;
528 bool rightmost
= true;
533 mz
->usage_in_excess
= new_usage_in_excess
;
534 if (!mz
->usage_in_excess
)
538 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
540 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
546 * We can't avoid mem cgroups that are over their soft
547 * limit by the same amount
549 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
554 mctz
->rb_rightmost
= &mz
->tree_node
;
556 rb_link_node(&mz
->tree_node
, parent
, p
);
557 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
561 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
562 struct mem_cgroup_tree_per_node
*mctz
)
567 if (&mz
->tree_node
== mctz
->rb_rightmost
)
568 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
570 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
574 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
575 struct mem_cgroup_tree_per_node
*mctz
)
579 spin_lock_irqsave(&mctz
->lock
, flags
);
580 __mem_cgroup_remove_exceeded(mz
, mctz
);
581 spin_unlock_irqrestore(&mctz
->lock
, flags
);
584 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
586 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
587 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
588 unsigned long excess
= 0;
590 if (nr_pages
> soft_limit
)
591 excess
= nr_pages
- soft_limit
;
596 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
598 unsigned long excess
;
599 struct mem_cgroup_per_node
*mz
;
600 struct mem_cgroup_tree_per_node
*mctz
;
602 mctz
= soft_limit_tree_from_page(page
);
606 * Necessary to update all ancestors when hierarchy is used.
607 * because their event counter is not touched.
609 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
610 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
611 excess
= soft_limit_excess(memcg
);
613 * We have to update the tree if mz is on RB-tree or
614 * mem is over its softlimit.
616 if (excess
|| mz
->on_tree
) {
619 spin_lock_irqsave(&mctz
->lock
, flags
);
620 /* if on-tree, remove it */
622 __mem_cgroup_remove_exceeded(mz
, mctz
);
624 * Insert again. mz->usage_in_excess will be updated.
625 * If excess is 0, no tree ops.
627 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
628 spin_unlock_irqrestore(&mctz
->lock
, flags
);
633 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
635 struct mem_cgroup_tree_per_node
*mctz
;
636 struct mem_cgroup_per_node
*mz
;
640 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
641 mctz
= soft_limit_tree_node(nid
);
643 mem_cgroup_remove_exceeded(mz
, mctz
);
647 static struct mem_cgroup_per_node
*
648 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
650 struct mem_cgroup_per_node
*mz
;
654 if (!mctz
->rb_rightmost
)
655 goto done
; /* Nothing to reclaim from */
657 mz
= rb_entry(mctz
->rb_rightmost
,
658 struct mem_cgroup_per_node
, tree_node
);
660 * Remove the node now but someone else can add it back,
661 * we will to add it back at the end of reclaim to its correct
662 * position in the tree.
664 __mem_cgroup_remove_exceeded(mz
, mctz
);
665 if (!soft_limit_excess(mz
->memcg
) ||
666 !css_tryget_online(&mz
->memcg
->css
))
672 static struct mem_cgroup_per_node
*
673 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
675 struct mem_cgroup_per_node
*mz
;
677 spin_lock_irq(&mctz
->lock
);
678 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
679 spin_unlock_irq(&mctz
->lock
);
684 * __mod_memcg_state - update cgroup memory statistics
685 * @memcg: the memory cgroup
686 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
687 * @val: delta to add to the counter, can be negative
689 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
693 if (mem_cgroup_disabled())
696 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
697 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
698 struct mem_cgroup
*mi
;
701 * Batch local counters to keep them in sync with
702 * the hierarchical ones.
704 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
705 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
706 atomic_long_add(x
, &mi
->vmstats
[idx
]);
709 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
712 static struct mem_cgroup_per_node
*
713 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
715 struct mem_cgroup
*parent
;
717 parent
= parent_mem_cgroup(pn
->memcg
);
720 return mem_cgroup_nodeinfo(parent
, nid
);
724 * __mod_lruvec_state - update lruvec memory statistics
725 * @lruvec: the lruvec
726 * @idx: the stat item
727 * @val: delta to add to the counter, can be negative
729 * The lruvec is the intersection of the NUMA node and a cgroup. This
730 * function updates the all three counters that are affected by a
731 * change of state at this level: per-node, per-cgroup, per-lruvec.
733 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
736 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
737 struct mem_cgroup_per_node
*pn
;
738 struct mem_cgroup
*memcg
;
742 __mod_node_page_state(pgdat
, idx
, val
);
744 if (mem_cgroup_disabled())
747 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
751 __mod_memcg_state(memcg
, idx
, val
);
754 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
756 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
757 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
758 struct mem_cgroup_per_node
*pi
;
760 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
761 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
764 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
767 void __mod_lruvec_slab_state(void *p
, enum node_stat_item idx
, int val
)
769 struct page
*page
= virt_to_head_page(p
);
770 pg_data_t
*pgdat
= page_pgdat(page
);
771 struct mem_cgroup
*memcg
;
772 struct lruvec
*lruvec
;
775 memcg
= memcg_from_slab_page(page
);
777 /* Untracked pages have no memcg, no lruvec. Update only the node */
778 if (!memcg
|| memcg
== root_mem_cgroup
) {
779 __mod_node_page_state(pgdat
, idx
, val
);
781 lruvec
= mem_cgroup_lruvec(pgdat
, memcg
);
782 __mod_lruvec_state(lruvec
, idx
, val
);
788 * __count_memcg_events - account VM events in a cgroup
789 * @memcg: the memory cgroup
790 * @idx: the event item
791 * @count: the number of events that occured
793 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
798 if (mem_cgroup_disabled())
801 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
802 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
803 struct mem_cgroup
*mi
;
806 * Batch local counters to keep them in sync with
807 * the hierarchical ones.
809 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
810 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
811 atomic_long_add(x
, &mi
->vmevents
[idx
]);
814 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
817 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
819 return atomic_long_read(&memcg
->vmevents
[event
]);
822 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
827 for_each_possible_cpu(cpu
)
828 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
832 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
834 bool compound
, int nr_pages
)
837 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
838 * counted as CACHE even if it's on ANON LRU.
841 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
843 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
844 if (PageSwapBacked(page
))
845 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
849 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
850 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
853 /* pagein of a big page is an event. So, ignore page size */
855 __count_memcg_events(memcg
, PGPGIN
, 1);
857 __count_memcg_events(memcg
, PGPGOUT
, 1);
858 nr_pages
= -nr_pages
; /* for event */
861 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
864 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
865 enum mem_cgroup_events_target target
)
867 unsigned long val
, next
;
869 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
870 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
871 /* from time_after() in jiffies.h */
872 if ((long)(next
- val
) < 0) {
874 case MEM_CGROUP_TARGET_THRESH
:
875 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
877 case MEM_CGROUP_TARGET_SOFTLIMIT
:
878 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
880 case MEM_CGROUP_TARGET_NUMAINFO
:
881 next
= val
+ NUMAINFO_EVENTS_TARGET
;
886 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
893 * Check events in order.
896 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
898 /* threshold event is triggered in finer grain than soft limit */
899 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
900 MEM_CGROUP_TARGET_THRESH
))) {
902 bool do_numainfo __maybe_unused
;
904 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
905 MEM_CGROUP_TARGET_SOFTLIMIT
);
907 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
908 MEM_CGROUP_TARGET_NUMAINFO
);
910 mem_cgroup_threshold(memcg
);
911 if (unlikely(do_softlimit
))
912 mem_cgroup_update_tree(memcg
, page
);
914 if (unlikely(do_numainfo
))
915 atomic_inc(&memcg
->numainfo_events
);
920 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
923 * mm_update_next_owner() may clear mm->owner to NULL
924 * if it races with swapoff, page migration, etc.
925 * So this can be called with p == NULL.
930 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
932 EXPORT_SYMBOL(mem_cgroup_from_task
);
935 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
936 * @mm: mm from which memcg should be extracted. It can be NULL.
938 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
939 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
942 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
944 struct mem_cgroup
*memcg
;
946 if (mem_cgroup_disabled())
952 * Page cache insertions can happen withou an
953 * actual mm context, e.g. during disk probing
954 * on boot, loopback IO, acct() writes etc.
957 memcg
= root_mem_cgroup
;
959 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
960 if (unlikely(!memcg
))
961 memcg
= root_mem_cgroup
;
963 } while (!css_tryget_online(&memcg
->css
));
967 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
970 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
971 * @page: page from which memcg should be extracted.
973 * Obtain a reference on page->memcg and returns it if successful. Otherwise
974 * root_mem_cgroup is returned.
976 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
978 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
980 if (mem_cgroup_disabled())
984 if (!memcg
|| !css_tryget_online(&memcg
->css
))
985 memcg
= root_mem_cgroup
;
989 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
992 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
994 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
996 if (unlikely(current
->active_memcg
)) {
997 struct mem_cgroup
*memcg
= root_mem_cgroup
;
1000 if (css_tryget_online(¤t
->active_memcg
->css
))
1001 memcg
= current
->active_memcg
;
1005 return get_mem_cgroup_from_mm(current
->mm
);
1009 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1010 * @root: hierarchy root
1011 * @prev: previously returned memcg, NULL on first invocation
1012 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1014 * Returns references to children of the hierarchy below @root, or
1015 * @root itself, or %NULL after a full round-trip.
1017 * Caller must pass the return value in @prev on subsequent
1018 * invocations for reference counting, or use mem_cgroup_iter_break()
1019 * to cancel a hierarchy walk before the round-trip is complete.
1021 * Reclaimers can specify a node and a priority level in @reclaim to
1022 * divide up the memcgs in the hierarchy among all concurrent
1023 * reclaimers operating on the same node and priority.
1025 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1026 struct mem_cgroup
*prev
,
1027 struct mem_cgroup_reclaim_cookie
*reclaim
)
1029 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1030 struct cgroup_subsys_state
*css
= NULL
;
1031 struct mem_cgroup
*memcg
= NULL
;
1032 struct mem_cgroup
*pos
= NULL
;
1034 if (mem_cgroup_disabled())
1038 root
= root_mem_cgroup
;
1040 if (prev
&& !reclaim
)
1043 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1052 struct mem_cgroup_per_node
*mz
;
1054 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1055 iter
= &mz
->iter
[reclaim
->priority
];
1057 if (prev
&& reclaim
->generation
!= iter
->generation
)
1061 pos
= READ_ONCE(iter
->position
);
1062 if (!pos
|| css_tryget(&pos
->css
))
1065 * css reference reached zero, so iter->position will
1066 * be cleared by ->css_released. However, we should not
1067 * rely on this happening soon, because ->css_released
1068 * is called from a work queue, and by busy-waiting we
1069 * might block it. So we clear iter->position right
1072 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1080 css
= css_next_descendant_pre(css
, &root
->css
);
1083 * Reclaimers share the hierarchy walk, and a
1084 * new one might jump in right at the end of
1085 * the hierarchy - make sure they see at least
1086 * one group and restart from the beginning.
1094 * Verify the css and acquire a reference. The root
1095 * is provided by the caller, so we know it's alive
1096 * and kicking, and don't take an extra reference.
1098 memcg
= mem_cgroup_from_css(css
);
1100 if (css
== &root
->css
)
1103 if (css_tryget(css
))
1111 * The position could have already been updated by a competing
1112 * thread, so check that the value hasn't changed since we read
1113 * it to avoid reclaiming from the same cgroup twice.
1115 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1123 reclaim
->generation
= iter
->generation
;
1129 if (prev
&& prev
!= root
)
1130 css_put(&prev
->css
);
1136 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1137 * @root: hierarchy root
1138 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1140 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1141 struct mem_cgroup
*prev
)
1144 root
= root_mem_cgroup
;
1145 if (prev
&& prev
!= root
)
1146 css_put(&prev
->css
);
1149 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1150 struct mem_cgroup
*dead_memcg
)
1152 struct mem_cgroup_reclaim_iter
*iter
;
1153 struct mem_cgroup_per_node
*mz
;
1157 for_each_node(nid
) {
1158 mz
= mem_cgroup_nodeinfo(from
, nid
);
1159 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
1160 iter
= &mz
->iter
[i
];
1161 cmpxchg(&iter
->position
,
1167 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1169 struct mem_cgroup
*memcg
= dead_memcg
;
1170 struct mem_cgroup
*last
;
1173 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1175 } while ((memcg
= parent_mem_cgroup(memcg
)));
1178 * When cgruop1 non-hierarchy mode is used,
1179 * parent_mem_cgroup() does not walk all the way up to the
1180 * cgroup root (root_mem_cgroup). So we have to handle
1181 * dead_memcg from cgroup root separately.
1183 if (last
!= root_mem_cgroup
)
1184 __invalidate_reclaim_iterators(root_mem_cgroup
,
1189 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1190 * @memcg: hierarchy root
1191 * @fn: function to call for each task
1192 * @arg: argument passed to @fn
1194 * This function iterates over tasks attached to @memcg or to any of its
1195 * descendants and calls @fn for each task. If @fn returns a non-zero
1196 * value, the function breaks the iteration loop and returns the value.
1197 * Otherwise, it will iterate over all tasks and return 0.
1199 * This function must not be called for the root memory cgroup.
1201 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1202 int (*fn
)(struct task_struct
*, void *), void *arg
)
1204 struct mem_cgroup
*iter
;
1207 BUG_ON(memcg
== root_mem_cgroup
);
1209 for_each_mem_cgroup_tree(iter
, memcg
) {
1210 struct css_task_iter it
;
1211 struct task_struct
*task
;
1213 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1214 while (!ret
&& (task
= css_task_iter_next(&it
)))
1215 ret
= fn(task
, arg
);
1216 css_task_iter_end(&it
);
1218 mem_cgroup_iter_break(memcg
, iter
);
1226 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1228 * @pgdat: pgdat of the page
1230 * This function is only safe when following the LRU page isolation
1231 * and putback protocol: the LRU lock must be held, and the page must
1232 * either be PageLRU() or the caller must have isolated/allocated it.
1234 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1236 struct mem_cgroup_per_node
*mz
;
1237 struct mem_cgroup
*memcg
;
1238 struct lruvec
*lruvec
;
1240 if (mem_cgroup_disabled()) {
1241 lruvec
= &pgdat
->lruvec
;
1245 memcg
= page
->mem_cgroup
;
1247 * Swapcache readahead pages are added to the LRU - and
1248 * possibly migrated - before they are charged.
1251 memcg
= root_mem_cgroup
;
1253 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1254 lruvec
= &mz
->lruvec
;
1257 * Since a node can be onlined after the mem_cgroup was created,
1258 * we have to be prepared to initialize lruvec->zone here;
1259 * and if offlined then reonlined, we need to reinitialize it.
1261 if (unlikely(lruvec
->pgdat
!= pgdat
))
1262 lruvec
->pgdat
= pgdat
;
1267 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1268 * @lruvec: mem_cgroup per zone lru vector
1269 * @lru: index of lru list the page is sitting on
1270 * @zid: zone id of the accounted pages
1271 * @nr_pages: positive when adding or negative when removing
1273 * This function must be called under lru_lock, just before a page is added
1274 * to or just after a page is removed from an lru list (that ordering being
1275 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1277 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1278 int zid
, int nr_pages
)
1280 struct mem_cgroup_per_node
*mz
;
1281 unsigned long *lru_size
;
1284 if (mem_cgroup_disabled())
1287 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1288 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1291 *lru_size
+= nr_pages
;
1294 if (WARN_ONCE(size
< 0,
1295 "%s(%p, %d, %d): lru_size %ld\n",
1296 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1302 *lru_size
+= nr_pages
;
1306 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1307 * @memcg: the memory cgroup
1309 * Returns the maximum amount of memory @mem can be charged with, in
1312 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1314 unsigned long margin
= 0;
1315 unsigned long count
;
1316 unsigned long limit
;
1318 count
= page_counter_read(&memcg
->memory
);
1319 limit
= READ_ONCE(memcg
->memory
.max
);
1321 margin
= limit
- count
;
1323 if (do_memsw_account()) {
1324 count
= page_counter_read(&memcg
->memsw
);
1325 limit
= READ_ONCE(memcg
->memsw
.max
);
1327 margin
= min(margin
, limit
- count
);
1336 * A routine for checking "mem" is under move_account() or not.
1338 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1339 * moving cgroups. This is for waiting at high-memory pressure
1342 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1344 struct mem_cgroup
*from
;
1345 struct mem_cgroup
*to
;
1348 * Unlike task_move routines, we access mc.to, mc.from not under
1349 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1351 spin_lock(&mc
.lock
);
1357 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1358 mem_cgroup_is_descendant(to
, memcg
);
1360 spin_unlock(&mc
.lock
);
1364 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1366 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1367 if (mem_cgroup_under_move(memcg
)) {
1369 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1370 /* moving charge context might have finished. */
1373 finish_wait(&mc
.waitq
, &wait
);
1380 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1385 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1390 * Provide statistics on the state of the memory subsystem as
1391 * well as cumulative event counters that show past behavior.
1393 * This list is ordered following a combination of these gradients:
1394 * 1) generic big picture -> specifics and details
1395 * 2) reflecting userspace activity -> reflecting kernel heuristics
1397 * Current memory state:
1400 seq_buf_printf(&s
, "anon %llu\n",
1401 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) *
1403 seq_buf_printf(&s
, "file %llu\n",
1404 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) *
1406 seq_buf_printf(&s
, "kernel_stack %llu\n",
1407 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) *
1409 seq_buf_printf(&s
, "slab %llu\n",
1410 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
1411 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
1413 seq_buf_printf(&s
, "sock %llu\n",
1414 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) *
1417 seq_buf_printf(&s
, "shmem %llu\n",
1418 (u64
)memcg_page_state(memcg
, NR_SHMEM
) *
1420 seq_buf_printf(&s
, "file_mapped %llu\n",
1421 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) *
1423 seq_buf_printf(&s
, "file_dirty %llu\n",
1424 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) *
1426 seq_buf_printf(&s
, "file_writeback %llu\n",
1427 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) *
1431 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1432 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1433 * arse because it requires migrating the work out of rmap to a place
1434 * where the page->mem_cgroup is set up and stable.
1436 seq_buf_printf(&s
, "anon_thp %llu\n",
1437 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) *
1440 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1441 seq_buf_printf(&s
, "%s %llu\n", mem_cgroup_lru_names
[i
],
1442 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
1445 seq_buf_printf(&s
, "slab_reclaimable %llu\n",
1446 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
1448 seq_buf_printf(&s
, "slab_unreclaimable %llu\n",
1449 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
1452 /* Accumulated memory events */
1454 seq_buf_printf(&s
, "pgfault %lu\n", memcg_events(memcg
, PGFAULT
));
1455 seq_buf_printf(&s
, "pgmajfault %lu\n", memcg_events(memcg
, PGMAJFAULT
));
1457 seq_buf_printf(&s
, "workingset_refault %lu\n",
1458 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
1459 seq_buf_printf(&s
, "workingset_activate %lu\n",
1460 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
1461 seq_buf_printf(&s
, "workingset_nodereclaim %lu\n",
1462 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
1464 seq_buf_printf(&s
, "pgrefill %lu\n", memcg_events(memcg
, PGREFILL
));
1465 seq_buf_printf(&s
, "pgscan %lu\n",
1466 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1467 memcg_events(memcg
, PGSCAN_DIRECT
));
1468 seq_buf_printf(&s
, "pgsteal %lu\n",
1469 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1470 memcg_events(memcg
, PGSTEAL_DIRECT
));
1471 seq_buf_printf(&s
, "pgactivate %lu\n", memcg_events(memcg
, PGACTIVATE
));
1472 seq_buf_printf(&s
, "pgdeactivate %lu\n", memcg_events(memcg
, PGDEACTIVATE
));
1473 seq_buf_printf(&s
, "pglazyfree %lu\n", memcg_events(memcg
, PGLAZYFREE
));
1474 seq_buf_printf(&s
, "pglazyfreed %lu\n", memcg_events(memcg
, PGLAZYFREED
));
1476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1477 seq_buf_printf(&s
, "thp_fault_alloc %lu\n",
1478 memcg_events(memcg
, THP_FAULT_ALLOC
));
1479 seq_buf_printf(&s
, "thp_collapse_alloc %lu\n",
1480 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1483 /* The above should easily fit into one page */
1484 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1489 #define K(x) ((x) << (PAGE_SHIFT-10))
1491 * mem_cgroup_print_oom_context: Print OOM information relevant to
1492 * memory controller.
1493 * @memcg: The memory cgroup that went over limit
1494 * @p: Task that is going to be killed
1496 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1499 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1504 pr_cont(",oom_memcg=");
1505 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1507 pr_cont(",global_oom");
1509 pr_cont(",task_memcg=");
1510 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1516 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1517 * memory controller.
1518 * @memcg: The memory cgroup that went over limit
1520 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1524 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1525 K((u64
)page_counter_read(&memcg
->memory
)),
1526 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1527 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1528 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1529 K((u64
)page_counter_read(&memcg
->swap
)),
1530 K((u64
)memcg
->swap
.max
), memcg
->swap
.failcnt
);
1532 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1533 K((u64
)page_counter_read(&memcg
->memsw
)),
1534 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1535 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1536 K((u64
)page_counter_read(&memcg
->kmem
)),
1537 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1540 pr_info("Memory cgroup stats for ");
1541 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1543 buf
= memory_stat_format(memcg
);
1551 * Return the memory (and swap, if configured) limit for a memcg.
1553 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1557 max
= memcg
->memory
.max
;
1558 if (mem_cgroup_swappiness(memcg
)) {
1559 unsigned long memsw_max
;
1560 unsigned long swap_max
;
1562 memsw_max
= memcg
->memsw
.max
;
1563 swap_max
= memcg
->swap
.max
;
1564 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1565 max
= min(max
+ swap_max
, memsw_max
);
1570 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1573 struct oom_control oc
= {
1577 .gfp_mask
= gfp_mask
,
1582 if (mutex_lock_killable(&oom_lock
))
1585 * A few threads which were not waiting at mutex_lock_killable() can
1586 * fail to bail out. Therefore, check again after holding oom_lock.
1588 ret
= should_force_charge() || out_of_memory(&oc
);
1589 mutex_unlock(&oom_lock
);
1593 #if MAX_NUMNODES > 1
1596 * test_mem_cgroup_node_reclaimable
1597 * @memcg: the target memcg
1598 * @nid: the node ID to be checked.
1599 * @noswap : specify true here if the user wants flle only information.
1601 * This function returns whether the specified memcg contains any
1602 * reclaimable pages on a node. Returns true if there are any reclaimable
1603 * pages in the node.
1605 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1606 int nid
, bool noswap
)
1608 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
1610 if (lruvec_page_state(lruvec
, NR_INACTIVE_FILE
) ||
1611 lruvec_page_state(lruvec
, NR_ACTIVE_FILE
))
1613 if (noswap
|| !total_swap_pages
)
1615 if (lruvec_page_state(lruvec
, NR_INACTIVE_ANON
) ||
1616 lruvec_page_state(lruvec
, NR_ACTIVE_ANON
))
1623 * Always updating the nodemask is not very good - even if we have an empty
1624 * list or the wrong list here, we can start from some node and traverse all
1625 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1628 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1632 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1633 * pagein/pageout changes since the last update.
1635 if (!atomic_read(&memcg
->numainfo_events
))
1637 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1640 /* make a nodemask where this memcg uses memory from */
1641 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1643 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1645 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1646 node_clear(nid
, memcg
->scan_nodes
);
1649 atomic_set(&memcg
->numainfo_events
, 0);
1650 atomic_set(&memcg
->numainfo_updating
, 0);
1654 * Selecting a node where we start reclaim from. Because what we need is just
1655 * reducing usage counter, start from anywhere is O,K. Considering
1656 * memory reclaim from current node, there are pros. and cons.
1658 * Freeing memory from current node means freeing memory from a node which
1659 * we'll use or we've used. So, it may make LRU bad. And if several threads
1660 * hit limits, it will see a contention on a node. But freeing from remote
1661 * node means more costs for memory reclaim because of memory latency.
1663 * Now, we use round-robin. Better algorithm is welcomed.
1665 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1669 mem_cgroup_may_update_nodemask(memcg
);
1670 node
= memcg
->last_scanned_node
;
1672 node
= next_node_in(node
, memcg
->scan_nodes
);
1674 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1675 * last time it really checked all the LRUs due to rate limiting.
1676 * Fallback to the current node in that case for simplicity.
1678 if (unlikely(node
== MAX_NUMNODES
))
1679 node
= numa_node_id();
1681 memcg
->last_scanned_node
= node
;
1685 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1691 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1694 unsigned long *total_scanned
)
1696 struct mem_cgroup
*victim
= NULL
;
1699 unsigned long excess
;
1700 unsigned long nr_scanned
;
1701 struct mem_cgroup_reclaim_cookie reclaim
= {
1706 excess
= soft_limit_excess(root_memcg
);
1709 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1714 * If we have not been able to reclaim
1715 * anything, it might because there are
1716 * no reclaimable pages under this hierarchy
1721 * We want to do more targeted reclaim.
1722 * excess >> 2 is not to excessive so as to
1723 * reclaim too much, nor too less that we keep
1724 * coming back to reclaim from this cgroup
1726 if (total
>= (excess
>> 2) ||
1727 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1732 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1733 pgdat
, &nr_scanned
);
1734 *total_scanned
+= nr_scanned
;
1735 if (!soft_limit_excess(root_memcg
))
1738 mem_cgroup_iter_break(root_memcg
, victim
);
1742 #ifdef CONFIG_LOCKDEP
1743 static struct lockdep_map memcg_oom_lock_dep_map
= {
1744 .name
= "memcg_oom_lock",
1748 static DEFINE_SPINLOCK(memcg_oom_lock
);
1751 * Check OOM-Killer is already running under our hierarchy.
1752 * If someone is running, return false.
1754 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1756 struct mem_cgroup
*iter
, *failed
= NULL
;
1758 spin_lock(&memcg_oom_lock
);
1760 for_each_mem_cgroup_tree(iter
, memcg
) {
1761 if (iter
->oom_lock
) {
1763 * this subtree of our hierarchy is already locked
1764 * so we cannot give a lock.
1767 mem_cgroup_iter_break(memcg
, iter
);
1770 iter
->oom_lock
= true;
1775 * OK, we failed to lock the whole subtree so we have
1776 * to clean up what we set up to the failing subtree
1778 for_each_mem_cgroup_tree(iter
, memcg
) {
1779 if (iter
== failed
) {
1780 mem_cgroup_iter_break(memcg
, iter
);
1783 iter
->oom_lock
= false;
1786 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1788 spin_unlock(&memcg_oom_lock
);
1793 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1795 struct mem_cgroup
*iter
;
1797 spin_lock(&memcg_oom_lock
);
1798 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1799 for_each_mem_cgroup_tree(iter
, memcg
)
1800 iter
->oom_lock
= false;
1801 spin_unlock(&memcg_oom_lock
);
1804 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1806 struct mem_cgroup
*iter
;
1808 spin_lock(&memcg_oom_lock
);
1809 for_each_mem_cgroup_tree(iter
, memcg
)
1811 spin_unlock(&memcg_oom_lock
);
1814 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1816 struct mem_cgroup
*iter
;
1819 * When a new child is created while the hierarchy is under oom,
1820 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1822 spin_lock(&memcg_oom_lock
);
1823 for_each_mem_cgroup_tree(iter
, memcg
)
1824 if (iter
->under_oom
> 0)
1826 spin_unlock(&memcg_oom_lock
);
1829 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1831 struct oom_wait_info
{
1832 struct mem_cgroup
*memcg
;
1833 wait_queue_entry_t wait
;
1836 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1837 unsigned mode
, int sync
, void *arg
)
1839 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1840 struct mem_cgroup
*oom_wait_memcg
;
1841 struct oom_wait_info
*oom_wait_info
;
1843 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1844 oom_wait_memcg
= oom_wait_info
->memcg
;
1846 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1847 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1849 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1852 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1855 * For the following lockless ->under_oom test, the only required
1856 * guarantee is that it must see the state asserted by an OOM when
1857 * this function is called as a result of userland actions
1858 * triggered by the notification of the OOM. This is trivially
1859 * achieved by invoking mem_cgroup_mark_under_oom() before
1860 * triggering notification.
1862 if (memcg
&& memcg
->under_oom
)
1863 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1873 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1875 enum oom_status ret
;
1878 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1881 memcg_memory_event(memcg
, MEMCG_OOM
);
1884 * We are in the middle of the charge context here, so we
1885 * don't want to block when potentially sitting on a callstack
1886 * that holds all kinds of filesystem and mm locks.
1888 * cgroup1 allows disabling the OOM killer and waiting for outside
1889 * handling until the charge can succeed; remember the context and put
1890 * the task to sleep at the end of the page fault when all locks are
1893 * On the other hand, in-kernel OOM killer allows for an async victim
1894 * memory reclaim (oom_reaper) and that means that we are not solely
1895 * relying on the oom victim to make a forward progress and we can
1896 * invoke the oom killer here.
1898 * Please note that mem_cgroup_out_of_memory might fail to find a
1899 * victim and then we have to bail out from the charge path.
1901 if (memcg
->oom_kill_disable
) {
1902 if (!current
->in_user_fault
)
1904 css_get(&memcg
->css
);
1905 current
->memcg_in_oom
= memcg
;
1906 current
->memcg_oom_gfp_mask
= mask
;
1907 current
->memcg_oom_order
= order
;
1912 mem_cgroup_mark_under_oom(memcg
);
1914 locked
= mem_cgroup_oom_trylock(memcg
);
1917 mem_cgroup_oom_notify(memcg
);
1919 mem_cgroup_unmark_under_oom(memcg
);
1920 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1926 mem_cgroup_oom_unlock(memcg
);
1932 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1933 * @handle: actually kill/wait or just clean up the OOM state
1935 * This has to be called at the end of a page fault if the memcg OOM
1936 * handler was enabled.
1938 * Memcg supports userspace OOM handling where failed allocations must
1939 * sleep on a waitqueue until the userspace task resolves the
1940 * situation. Sleeping directly in the charge context with all kinds
1941 * of locks held is not a good idea, instead we remember an OOM state
1942 * in the task and mem_cgroup_oom_synchronize() has to be called at
1943 * the end of the page fault to complete the OOM handling.
1945 * Returns %true if an ongoing memcg OOM situation was detected and
1946 * completed, %false otherwise.
1948 bool mem_cgroup_oom_synchronize(bool handle
)
1950 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1951 struct oom_wait_info owait
;
1954 /* OOM is global, do not handle */
1961 owait
.memcg
= memcg
;
1962 owait
.wait
.flags
= 0;
1963 owait
.wait
.func
= memcg_oom_wake_function
;
1964 owait
.wait
.private = current
;
1965 INIT_LIST_HEAD(&owait
.wait
.entry
);
1967 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1968 mem_cgroup_mark_under_oom(memcg
);
1970 locked
= mem_cgroup_oom_trylock(memcg
);
1973 mem_cgroup_oom_notify(memcg
);
1975 if (locked
&& !memcg
->oom_kill_disable
) {
1976 mem_cgroup_unmark_under_oom(memcg
);
1977 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1978 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1979 current
->memcg_oom_order
);
1982 mem_cgroup_unmark_under_oom(memcg
);
1983 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1987 mem_cgroup_oom_unlock(memcg
);
1989 * There is no guarantee that an OOM-lock contender
1990 * sees the wakeups triggered by the OOM kill
1991 * uncharges. Wake any sleepers explicitely.
1993 memcg_oom_recover(memcg
);
1996 current
->memcg_in_oom
= NULL
;
1997 css_put(&memcg
->css
);
2002 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2003 * @victim: task to be killed by the OOM killer
2004 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2006 * Returns a pointer to a memory cgroup, which has to be cleaned up
2007 * by killing all belonging OOM-killable tasks.
2009 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2011 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2012 struct mem_cgroup
*oom_domain
)
2014 struct mem_cgroup
*oom_group
= NULL
;
2015 struct mem_cgroup
*memcg
;
2017 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2021 oom_domain
= root_mem_cgroup
;
2025 memcg
= mem_cgroup_from_task(victim
);
2026 if (memcg
== root_mem_cgroup
)
2030 * Traverse the memory cgroup hierarchy from the victim task's
2031 * cgroup up to the OOMing cgroup (or root) to find the
2032 * highest-level memory cgroup with oom.group set.
2034 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2035 if (memcg
->oom_group
)
2038 if (memcg
== oom_domain
)
2043 css_get(&oom_group
->css
);
2050 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2052 pr_info("Tasks in ");
2053 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2054 pr_cont(" are going to be killed due to memory.oom.group set\n");
2058 * lock_page_memcg - lock a page->mem_cgroup binding
2061 * This function protects unlocked LRU pages from being moved to
2064 * It ensures lifetime of the returned memcg. Caller is responsible
2065 * for the lifetime of the page; __unlock_page_memcg() is available
2066 * when @page might get freed inside the locked section.
2068 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
2070 struct mem_cgroup
*memcg
;
2071 unsigned long flags
;
2074 * The RCU lock is held throughout the transaction. The fast
2075 * path can get away without acquiring the memcg->move_lock
2076 * because page moving starts with an RCU grace period.
2078 * The RCU lock also protects the memcg from being freed when
2079 * the page state that is going to change is the only thing
2080 * preventing the page itself from being freed. E.g. writeback
2081 * doesn't hold a page reference and relies on PG_writeback to
2082 * keep off truncation, migration and so forth.
2086 if (mem_cgroup_disabled())
2089 memcg
= page
->mem_cgroup
;
2090 if (unlikely(!memcg
))
2093 if (atomic_read(&memcg
->moving_account
) <= 0)
2096 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2097 if (memcg
!= page
->mem_cgroup
) {
2098 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2103 * When charge migration first begins, we can have locked and
2104 * unlocked page stat updates happening concurrently. Track
2105 * the task who has the lock for unlock_page_memcg().
2107 memcg
->move_lock_task
= current
;
2108 memcg
->move_lock_flags
= flags
;
2112 EXPORT_SYMBOL(lock_page_memcg
);
2115 * __unlock_page_memcg - unlock and unpin a memcg
2118 * Unlock and unpin a memcg returned by lock_page_memcg().
2120 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2122 if (memcg
&& memcg
->move_lock_task
== current
) {
2123 unsigned long flags
= memcg
->move_lock_flags
;
2125 memcg
->move_lock_task
= NULL
;
2126 memcg
->move_lock_flags
= 0;
2128 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2135 * unlock_page_memcg - unlock a page->mem_cgroup binding
2138 void unlock_page_memcg(struct page
*page
)
2140 __unlock_page_memcg(page
->mem_cgroup
);
2142 EXPORT_SYMBOL(unlock_page_memcg
);
2144 struct memcg_stock_pcp
{
2145 struct mem_cgroup
*cached
; /* this never be root cgroup */
2146 unsigned int nr_pages
;
2147 struct work_struct work
;
2148 unsigned long flags
;
2149 #define FLUSHING_CACHED_CHARGE 0
2151 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2152 static DEFINE_MUTEX(percpu_charge_mutex
);
2155 * consume_stock: Try to consume stocked charge on this cpu.
2156 * @memcg: memcg to consume from.
2157 * @nr_pages: how many pages to charge.
2159 * The charges will only happen if @memcg matches the current cpu's memcg
2160 * stock, and at least @nr_pages are available in that stock. Failure to
2161 * service an allocation will refill the stock.
2163 * returns true if successful, false otherwise.
2165 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2167 struct memcg_stock_pcp
*stock
;
2168 unsigned long flags
;
2171 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2174 local_irq_save(flags
);
2176 stock
= this_cpu_ptr(&memcg_stock
);
2177 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2178 stock
->nr_pages
-= nr_pages
;
2182 local_irq_restore(flags
);
2188 * Returns stocks cached in percpu and reset cached information.
2190 static void drain_stock(struct memcg_stock_pcp
*stock
)
2192 struct mem_cgroup
*old
= stock
->cached
;
2194 if (stock
->nr_pages
) {
2195 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2196 if (do_memsw_account())
2197 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2198 css_put_many(&old
->css
, stock
->nr_pages
);
2199 stock
->nr_pages
= 0;
2201 stock
->cached
= NULL
;
2204 static void drain_local_stock(struct work_struct
*dummy
)
2206 struct memcg_stock_pcp
*stock
;
2207 unsigned long flags
;
2210 * The only protection from memory hotplug vs. drain_stock races is
2211 * that we always operate on local CPU stock here with IRQ disabled
2213 local_irq_save(flags
);
2215 stock
= this_cpu_ptr(&memcg_stock
);
2217 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2219 local_irq_restore(flags
);
2223 * Cache charges(val) to local per_cpu area.
2224 * This will be consumed by consume_stock() function, later.
2226 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2228 struct memcg_stock_pcp
*stock
;
2229 unsigned long flags
;
2231 local_irq_save(flags
);
2233 stock
= this_cpu_ptr(&memcg_stock
);
2234 if (stock
->cached
!= memcg
) { /* reset if necessary */
2236 stock
->cached
= memcg
;
2238 stock
->nr_pages
+= nr_pages
;
2240 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2243 local_irq_restore(flags
);
2247 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2248 * of the hierarchy under it.
2250 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2254 /* If someone's already draining, avoid adding running more workers. */
2255 if (!mutex_trylock(&percpu_charge_mutex
))
2258 * Notify other cpus that system-wide "drain" is running
2259 * We do not care about races with the cpu hotplug because cpu down
2260 * as well as workers from this path always operate on the local
2261 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2264 for_each_online_cpu(cpu
) {
2265 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2266 struct mem_cgroup
*memcg
;
2270 memcg
= stock
->cached
;
2271 if (memcg
&& stock
->nr_pages
&&
2272 mem_cgroup_is_descendant(memcg
, root_memcg
))
2277 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2279 drain_local_stock(&stock
->work
);
2281 schedule_work_on(cpu
, &stock
->work
);
2285 mutex_unlock(&percpu_charge_mutex
);
2288 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2290 struct memcg_stock_pcp
*stock
;
2291 struct mem_cgroup
*memcg
, *mi
;
2293 stock
= &per_cpu(memcg_stock
, cpu
);
2296 for_each_mem_cgroup(memcg
) {
2299 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2303 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2305 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2306 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2308 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2311 for_each_node(nid
) {
2312 struct mem_cgroup_per_node
*pn
;
2314 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2315 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2318 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2319 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2323 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2326 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2328 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2329 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2336 static void reclaim_high(struct mem_cgroup
*memcg
,
2337 unsigned int nr_pages
,
2341 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2343 memcg_memory_event(memcg
, MEMCG_HIGH
);
2344 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2345 } while ((memcg
= parent_mem_cgroup(memcg
)));
2348 static void high_work_func(struct work_struct
*work
)
2350 struct mem_cgroup
*memcg
;
2352 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2353 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2357 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2358 * enough to still cause a significant slowdown in most cases, while still
2359 * allowing diagnostics and tracing to proceed without becoming stuck.
2361 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2364 * When calculating the delay, we use these either side of the exponentiation to
2365 * maintain precision and scale to a reasonable number of jiffies (see the table
2368 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2369 * overage ratio to a delay.
2370 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2371 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2372 * to produce a reasonable delay curve.
2374 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2375 * reasonable delay curve compared to precision-adjusted overage, not
2376 * penalising heavily at first, but still making sure that growth beyond the
2377 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2378 * example, with a high of 100 megabytes:
2380 * +-------+------------------------+
2381 * | usage | time to allocate in ms |
2382 * +-------+------------------------+
2404 * +-------+------------------------+
2406 #define MEMCG_DELAY_PRECISION_SHIFT 20
2407 #define MEMCG_DELAY_SCALING_SHIFT 14
2410 * Scheduled by try_charge() to be executed from the userland return path
2411 * and reclaims memory over the high limit.
2413 void mem_cgroup_handle_over_high(void)
2415 unsigned long usage
, high
, clamped_high
;
2416 unsigned long pflags
;
2417 unsigned long penalty_jiffies
, overage
;
2418 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2419 struct mem_cgroup
*memcg
;
2421 if (likely(!nr_pages
))
2424 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2425 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2426 current
->memcg_nr_pages_over_high
= 0;
2429 * memory.high is breached and reclaim is unable to keep up. Throttle
2430 * allocators proactively to slow down excessive growth.
2432 * We use overage compared to memory.high to calculate the number of
2433 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2434 * fairly lenient on small overages, and increasingly harsh when the
2435 * memcg in question makes it clear that it has no intention of stopping
2436 * its crazy behaviour, so we exponentially increase the delay based on
2440 usage
= page_counter_read(&memcg
->memory
);
2441 high
= READ_ONCE(memcg
->high
);
2447 * Prevent division by 0 in overage calculation by acting as if it was a
2448 * threshold of 1 page
2450 clamped_high
= max(high
, 1UL);
2452 overage
= div_u64((u64
)(usage
- high
) << MEMCG_DELAY_PRECISION_SHIFT
,
2455 penalty_jiffies
= ((u64
)overage
* overage
* HZ
)
2456 >> (MEMCG_DELAY_PRECISION_SHIFT
+ MEMCG_DELAY_SCALING_SHIFT
);
2459 * Factor in the task's own contribution to the overage, such that four
2460 * N-sized allocations are throttled approximately the same as one
2461 * 4N-sized allocation.
2463 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2464 * larger the current charge patch is than that.
2466 penalty_jiffies
= penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2469 * Clamp the max delay per usermode return so as to still keep the
2470 * application moving forwards and also permit diagnostics, albeit
2473 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2476 * Don't sleep if the amount of jiffies this memcg owes us is so low
2477 * that it's not even worth doing, in an attempt to be nice to those who
2478 * go only a small amount over their memory.high value and maybe haven't
2479 * been aggressively reclaimed enough yet.
2481 if (penalty_jiffies
<= HZ
/ 100)
2485 * If we exit early, we're guaranteed to die (since
2486 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2487 * need to account for any ill-begotten jiffies to pay them off later.
2489 psi_memstall_enter(&pflags
);
2490 schedule_timeout_killable(penalty_jiffies
);
2491 psi_memstall_leave(&pflags
);
2494 css_put(&memcg
->css
);
2497 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2498 unsigned int nr_pages
)
2500 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2501 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2502 struct mem_cgroup
*mem_over_limit
;
2503 struct page_counter
*counter
;
2504 unsigned long nr_reclaimed
;
2505 bool may_swap
= true;
2506 bool drained
= false;
2507 enum oom_status oom_status
;
2509 if (mem_cgroup_is_root(memcg
))
2512 if (consume_stock(memcg
, nr_pages
))
2515 if (!do_memsw_account() ||
2516 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2517 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2519 if (do_memsw_account())
2520 page_counter_uncharge(&memcg
->memsw
, batch
);
2521 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2523 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2527 if (batch
> nr_pages
) {
2533 * Unlike in global OOM situations, memcg is not in a physical
2534 * memory shortage. Allow dying and OOM-killed tasks to
2535 * bypass the last charges so that they can exit quickly and
2536 * free their memory.
2538 if (unlikely(should_force_charge()))
2542 * Prevent unbounded recursion when reclaim operations need to
2543 * allocate memory. This might exceed the limits temporarily,
2544 * but we prefer facilitating memory reclaim and getting back
2545 * under the limit over triggering OOM kills in these cases.
2547 if (unlikely(current
->flags
& PF_MEMALLOC
))
2550 if (unlikely(task_in_memcg_oom(current
)))
2553 if (!gfpflags_allow_blocking(gfp_mask
))
2556 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2558 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2559 gfp_mask
, may_swap
);
2561 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2565 drain_all_stock(mem_over_limit
);
2570 if (gfp_mask
& __GFP_NORETRY
)
2573 * Even though the limit is exceeded at this point, reclaim
2574 * may have been able to free some pages. Retry the charge
2575 * before killing the task.
2577 * Only for regular pages, though: huge pages are rather
2578 * unlikely to succeed so close to the limit, and we fall back
2579 * to regular pages anyway in case of failure.
2581 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2584 * At task move, charge accounts can be doubly counted. So, it's
2585 * better to wait until the end of task_move if something is going on.
2587 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2593 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2596 if (gfp_mask
& __GFP_NOFAIL
)
2599 if (fatal_signal_pending(current
))
2603 * keep retrying as long as the memcg oom killer is able to make
2604 * a forward progress or bypass the charge if the oom killer
2605 * couldn't make any progress.
2607 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2608 get_order(nr_pages
* PAGE_SIZE
));
2609 switch (oom_status
) {
2611 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2619 if (!(gfp_mask
& __GFP_NOFAIL
))
2623 * The allocation either can't fail or will lead to more memory
2624 * being freed very soon. Allow memory usage go over the limit
2625 * temporarily by force charging it.
2627 page_counter_charge(&memcg
->memory
, nr_pages
);
2628 if (do_memsw_account())
2629 page_counter_charge(&memcg
->memsw
, nr_pages
);
2630 css_get_many(&memcg
->css
, nr_pages
);
2635 css_get_many(&memcg
->css
, batch
);
2636 if (batch
> nr_pages
)
2637 refill_stock(memcg
, batch
- nr_pages
);
2640 * If the hierarchy is above the normal consumption range, schedule
2641 * reclaim on returning to userland. We can perform reclaim here
2642 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2643 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2644 * not recorded as it most likely matches current's and won't
2645 * change in the meantime. As high limit is checked again before
2646 * reclaim, the cost of mismatch is negligible.
2649 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2650 /* Don't bother a random interrupted task */
2651 if (in_interrupt()) {
2652 schedule_work(&memcg
->high_work
);
2655 current
->memcg_nr_pages_over_high
+= batch
;
2656 set_notify_resume(current
);
2659 } while ((memcg
= parent_mem_cgroup(memcg
)));
2664 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2666 if (mem_cgroup_is_root(memcg
))
2669 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2670 if (do_memsw_account())
2671 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2673 css_put_many(&memcg
->css
, nr_pages
);
2676 static void lock_page_lru(struct page
*page
, int *isolated
)
2678 pg_data_t
*pgdat
= page_pgdat(page
);
2680 spin_lock_irq(&pgdat
->lru_lock
);
2681 if (PageLRU(page
)) {
2682 struct lruvec
*lruvec
;
2684 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2686 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2692 static void unlock_page_lru(struct page
*page
, int isolated
)
2694 pg_data_t
*pgdat
= page_pgdat(page
);
2697 struct lruvec
*lruvec
;
2699 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2700 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2702 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2704 spin_unlock_irq(&pgdat
->lru_lock
);
2707 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2712 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2715 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2716 * may already be on some other mem_cgroup's LRU. Take care of it.
2719 lock_page_lru(page
, &isolated
);
2722 * Nobody should be changing or seriously looking at
2723 * page->mem_cgroup at this point:
2725 * - the page is uncharged
2727 * - the page is off-LRU
2729 * - an anonymous fault has exclusive page access, except for
2730 * a locked page table
2732 * - a page cache insertion, a swapin fault, or a migration
2733 * have the page locked
2735 page
->mem_cgroup
= memcg
;
2738 unlock_page_lru(page
, isolated
);
2741 #ifdef CONFIG_MEMCG_KMEM
2742 static int memcg_alloc_cache_id(void)
2747 id
= ida_simple_get(&memcg_cache_ida
,
2748 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2752 if (id
< memcg_nr_cache_ids
)
2756 * There's no space for the new id in memcg_caches arrays,
2757 * so we have to grow them.
2759 down_write(&memcg_cache_ids_sem
);
2761 size
= 2 * (id
+ 1);
2762 if (size
< MEMCG_CACHES_MIN_SIZE
)
2763 size
= MEMCG_CACHES_MIN_SIZE
;
2764 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2765 size
= MEMCG_CACHES_MAX_SIZE
;
2767 err
= memcg_update_all_caches(size
);
2769 err
= memcg_update_all_list_lrus(size
);
2771 memcg_nr_cache_ids
= size
;
2773 up_write(&memcg_cache_ids_sem
);
2776 ida_simple_remove(&memcg_cache_ida
, id
);
2782 static void memcg_free_cache_id(int id
)
2784 ida_simple_remove(&memcg_cache_ida
, id
);
2787 struct memcg_kmem_cache_create_work
{
2788 struct mem_cgroup
*memcg
;
2789 struct kmem_cache
*cachep
;
2790 struct work_struct work
;
2793 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2795 struct memcg_kmem_cache_create_work
*cw
=
2796 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2797 struct mem_cgroup
*memcg
= cw
->memcg
;
2798 struct kmem_cache
*cachep
= cw
->cachep
;
2800 memcg_create_kmem_cache(memcg
, cachep
);
2802 css_put(&memcg
->css
);
2807 * Enqueue the creation of a per-memcg kmem_cache.
2809 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2810 struct kmem_cache
*cachep
)
2812 struct memcg_kmem_cache_create_work
*cw
;
2814 if (!css_tryget_online(&memcg
->css
))
2817 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2822 cw
->cachep
= cachep
;
2823 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2825 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2828 static inline bool memcg_kmem_bypass(void)
2830 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2836 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2837 * @cachep: the original global kmem cache
2839 * Return the kmem_cache we're supposed to use for a slab allocation.
2840 * We try to use the current memcg's version of the cache.
2842 * If the cache does not exist yet, if we are the first user of it, we
2843 * create it asynchronously in a workqueue and let the current allocation
2844 * go through with the original cache.
2846 * This function takes a reference to the cache it returns to assure it
2847 * won't get destroyed while we are working with it. Once the caller is
2848 * done with it, memcg_kmem_put_cache() must be called to release the
2851 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2853 struct mem_cgroup
*memcg
;
2854 struct kmem_cache
*memcg_cachep
;
2855 struct memcg_cache_array
*arr
;
2858 VM_BUG_ON(!is_root_cache(cachep
));
2860 if (memcg_kmem_bypass())
2865 if (unlikely(current
->active_memcg
))
2866 memcg
= current
->active_memcg
;
2868 memcg
= mem_cgroup_from_task(current
);
2870 if (!memcg
|| memcg
== root_mem_cgroup
)
2873 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2877 arr
= rcu_dereference(cachep
->memcg_params
.memcg_caches
);
2880 * Make sure we will access the up-to-date value. The code updating
2881 * memcg_caches issues a write barrier to match the data dependency
2882 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2884 memcg_cachep
= READ_ONCE(arr
->entries
[kmemcg_id
]);
2887 * If we are in a safe context (can wait, and not in interrupt
2888 * context), we could be be predictable and return right away.
2889 * This would guarantee that the allocation being performed
2890 * already belongs in the new cache.
2892 * However, there are some clashes that can arrive from locking.
2893 * For instance, because we acquire the slab_mutex while doing
2894 * memcg_create_kmem_cache, this means no further allocation
2895 * could happen with the slab_mutex held. So it's better to
2898 * If the memcg is dying or memcg_cache is about to be released,
2899 * don't bother creating new kmem_caches. Because memcg_cachep
2900 * is ZEROed as the fist step of kmem offlining, we don't need
2901 * percpu_ref_tryget_live() here. css_tryget_online() check in
2902 * memcg_schedule_kmem_cache_create() will prevent us from
2903 * creation of a new kmem_cache.
2905 if (unlikely(!memcg_cachep
))
2906 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2907 else if (percpu_ref_tryget(&memcg_cachep
->memcg_params
.refcnt
))
2908 cachep
= memcg_cachep
;
2915 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2916 * @cachep: the cache returned by memcg_kmem_get_cache
2918 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2920 if (!is_root_cache(cachep
))
2921 percpu_ref_put(&cachep
->memcg_params
.refcnt
);
2925 * __memcg_kmem_charge_memcg: charge a kmem page
2926 * @page: page to charge
2927 * @gfp: reclaim mode
2928 * @order: allocation order
2929 * @memcg: memory cgroup to charge
2931 * Returns 0 on success, an error code on failure.
2933 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2934 struct mem_cgroup
*memcg
)
2936 unsigned int nr_pages
= 1 << order
;
2937 struct page_counter
*counter
;
2940 ret
= try_charge(memcg
, gfp
, nr_pages
);
2944 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2945 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2948 * Enforce __GFP_NOFAIL allocation because callers are not
2949 * prepared to see failures and likely do not have any failure
2952 if (gfp
& __GFP_NOFAIL
) {
2953 page_counter_charge(&memcg
->kmem
, nr_pages
);
2956 cancel_charge(memcg
, nr_pages
);
2963 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2964 * @page: page to charge
2965 * @gfp: reclaim mode
2966 * @order: allocation order
2968 * Returns 0 on success, an error code on failure.
2970 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2972 struct mem_cgroup
*memcg
;
2975 if (memcg_kmem_bypass())
2978 memcg
= get_mem_cgroup_from_current();
2979 if (!mem_cgroup_is_root(memcg
)) {
2980 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2982 page
->mem_cgroup
= memcg
;
2983 __SetPageKmemcg(page
);
2986 css_put(&memcg
->css
);
2991 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2992 * @memcg: memcg to uncharge
2993 * @nr_pages: number of pages to uncharge
2995 void __memcg_kmem_uncharge_memcg(struct mem_cgroup
*memcg
,
2996 unsigned int nr_pages
)
2998 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2999 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
3001 page_counter_uncharge(&memcg
->memory
, nr_pages
);
3002 if (do_memsw_account())
3003 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
3006 * __memcg_kmem_uncharge: uncharge a kmem page
3007 * @page: page to uncharge
3008 * @order: allocation order
3010 void __memcg_kmem_uncharge(struct page
*page
, int order
)
3012 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
3013 unsigned int nr_pages
= 1 << order
;
3018 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3019 __memcg_kmem_uncharge_memcg(memcg
, nr_pages
);
3020 page
->mem_cgroup
= NULL
;
3022 /* slab pages do not have PageKmemcg flag set */
3023 if (PageKmemcg(page
))
3024 __ClearPageKmemcg(page
);
3026 css_put_many(&memcg
->css
, nr_pages
);
3028 #endif /* CONFIG_MEMCG_KMEM */
3030 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3033 * Because tail pages are not marked as "used", set it. We're under
3034 * pgdat->lru_lock and migration entries setup in all page mappings.
3036 void mem_cgroup_split_huge_fixup(struct page
*head
)
3040 if (mem_cgroup_disabled())
3043 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
3044 head
[i
].mem_cgroup
= head
->mem_cgroup
;
3046 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
3048 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3050 #ifdef CONFIG_MEMCG_SWAP
3052 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3053 * @entry: swap entry to be moved
3054 * @from: mem_cgroup which the entry is moved from
3055 * @to: mem_cgroup which the entry is moved to
3057 * It succeeds only when the swap_cgroup's record for this entry is the same
3058 * as the mem_cgroup's id of @from.
3060 * Returns 0 on success, -EINVAL on failure.
3062 * The caller must have charged to @to, IOW, called page_counter_charge() about
3063 * both res and memsw, and called css_get().
3065 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3066 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3068 unsigned short old_id
, new_id
;
3070 old_id
= mem_cgroup_id(from
);
3071 new_id
= mem_cgroup_id(to
);
3073 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3074 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3075 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3081 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3082 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3088 static DEFINE_MUTEX(memcg_max_mutex
);
3090 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3091 unsigned long max
, bool memsw
)
3093 bool enlarge
= false;
3094 bool drained
= false;
3096 bool limits_invariant
;
3097 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3100 if (signal_pending(current
)) {
3105 mutex_lock(&memcg_max_mutex
);
3107 * Make sure that the new limit (memsw or memory limit) doesn't
3108 * break our basic invariant rule memory.max <= memsw.max.
3110 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
3111 max
<= memcg
->memsw
.max
;
3112 if (!limits_invariant
) {
3113 mutex_unlock(&memcg_max_mutex
);
3117 if (max
> counter
->max
)
3119 ret
= page_counter_set_max(counter
, max
);
3120 mutex_unlock(&memcg_max_mutex
);
3126 drain_all_stock(memcg
);
3131 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3132 GFP_KERNEL
, !memsw
)) {
3138 if (!ret
&& enlarge
)
3139 memcg_oom_recover(memcg
);
3144 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3146 unsigned long *total_scanned
)
3148 unsigned long nr_reclaimed
= 0;
3149 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3150 unsigned long reclaimed
;
3152 struct mem_cgroup_tree_per_node
*mctz
;
3153 unsigned long excess
;
3154 unsigned long nr_scanned
;
3159 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3162 * Do not even bother to check the largest node if the root
3163 * is empty. Do it lockless to prevent lock bouncing. Races
3164 * are acceptable as soft limit is best effort anyway.
3166 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3170 * This loop can run a while, specially if mem_cgroup's continuously
3171 * keep exceeding their soft limit and putting the system under
3178 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3183 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3184 gfp_mask
, &nr_scanned
);
3185 nr_reclaimed
+= reclaimed
;
3186 *total_scanned
+= nr_scanned
;
3187 spin_lock_irq(&mctz
->lock
);
3188 __mem_cgroup_remove_exceeded(mz
, mctz
);
3191 * If we failed to reclaim anything from this memory cgroup
3192 * it is time to move on to the next cgroup
3196 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3198 excess
= soft_limit_excess(mz
->memcg
);
3200 * One school of thought says that we should not add
3201 * back the node to the tree if reclaim returns 0.
3202 * But our reclaim could return 0, simply because due
3203 * to priority we are exposing a smaller subset of
3204 * memory to reclaim from. Consider this as a longer
3207 /* If excess == 0, no tree ops */
3208 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3209 spin_unlock_irq(&mctz
->lock
);
3210 css_put(&mz
->memcg
->css
);
3213 * Could not reclaim anything and there are no more
3214 * mem cgroups to try or we seem to be looping without
3215 * reclaiming anything.
3217 if (!nr_reclaimed
&&
3219 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3221 } while (!nr_reclaimed
);
3223 css_put(&next_mz
->memcg
->css
);
3224 return nr_reclaimed
;
3228 * Test whether @memcg has children, dead or alive. Note that this
3229 * function doesn't care whether @memcg has use_hierarchy enabled and
3230 * returns %true if there are child csses according to the cgroup
3231 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3233 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3238 ret
= css_next_child(NULL
, &memcg
->css
);
3244 * Reclaims as many pages from the given memcg as possible.
3246 * Caller is responsible for holding css reference for memcg.
3248 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3250 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3252 /* we call try-to-free pages for make this cgroup empty */
3253 lru_add_drain_all();
3255 drain_all_stock(memcg
);
3257 /* try to free all pages in this cgroup */
3258 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3261 if (signal_pending(current
))
3264 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3268 /* maybe some writeback is necessary */
3269 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3277 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3278 char *buf
, size_t nbytes
,
3281 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3283 if (mem_cgroup_is_root(memcg
))
3285 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3288 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3291 return mem_cgroup_from_css(css
)->use_hierarchy
;
3294 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3295 struct cftype
*cft
, u64 val
)
3298 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3299 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3301 if (memcg
->use_hierarchy
== val
)
3305 * If parent's use_hierarchy is set, we can't make any modifications
3306 * in the child subtrees. If it is unset, then the change can
3307 * occur, provided the current cgroup has no children.
3309 * For the root cgroup, parent_mem is NULL, we allow value to be
3310 * set if there are no children.
3312 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3313 (val
== 1 || val
== 0)) {
3314 if (!memcg_has_children(memcg
))
3315 memcg
->use_hierarchy
= val
;
3324 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3328 if (mem_cgroup_is_root(memcg
)) {
3329 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3330 memcg_page_state(memcg
, MEMCG_RSS
);
3332 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3335 val
= page_counter_read(&memcg
->memory
);
3337 val
= page_counter_read(&memcg
->memsw
);
3350 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3353 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3354 struct page_counter
*counter
;
3356 switch (MEMFILE_TYPE(cft
->private)) {
3358 counter
= &memcg
->memory
;
3361 counter
= &memcg
->memsw
;
3364 counter
= &memcg
->kmem
;
3367 counter
= &memcg
->tcpmem
;
3373 switch (MEMFILE_ATTR(cft
->private)) {
3375 if (counter
== &memcg
->memory
)
3376 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3377 if (counter
== &memcg
->memsw
)
3378 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3379 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3381 return (u64
)counter
->max
* PAGE_SIZE
;
3383 return (u64
)counter
->watermark
* PAGE_SIZE
;
3385 return counter
->failcnt
;
3386 case RES_SOFT_LIMIT
:
3387 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3393 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
, bool slab_only
)
3395 unsigned long stat
[MEMCG_NR_STAT
];
3396 struct mem_cgroup
*mi
;
3398 int min_idx
, max_idx
;
3401 min_idx
= NR_SLAB_RECLAIMABLE
;
3402 max_idx
= NR_SLAB_UNRECLAIMABLE
;
3405 max_idx
= MEMCG_NR_STAT
;
3408 for (i
= min_idx
; i
< max_idx
; i
++)
3411 for_each_online_cpu(cpu
)
3412 for (i
= min_idx
; i
< max_idx
; i
++)
3413 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3415 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3416 for (i
= min_idx
; i
< max_idx
; i
++)
3417 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3420 max_idx
= NR_VM_NODE_STAT_ITEMS
;
3422 for_each_node(node
) {
3423 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3424 struct mem_cgroup_per_node
*pi
;
3426 for (i
= min_idx
; i
< max_idx
; i
++)
3429 for_each_online_cpu(cpu
)
3430 for (i
= min_idx
; i
< max_idx
; i
++)
3432 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3434 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3435 for (i
= min_idx
; i
< max_idx
; i
++)
3436 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3440 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3442 unsigned long events
[NR_VM_EVENT_ITEMS
];
3443 struct mem_cgroup
*mi
;
3446 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3449 for_each_online_cpu(cpu
)
3450 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3451 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3454 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3455 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3456 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3459 #ifdef CONFIG_MEMCG_KMEM
3460 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3464 if (cgroup_memory_nokmem
)
3467 BUG_ON(memcg
->kmemcg_id
>= 0);
3468 BUG_ON(memcg
->kmem_state
);
3470 memcg_id
= memcg_alloc_cache_id();
3474 static_branch_inc(&memcg_kmem_enabled_key
);
3476 * A memory cgroup is considered kmem-online as soon as it gets
3477 * kmemcg_id. Setting the id after enabling static branching will
3478 * guarantee no one starts accounting before all call sites are
3481 memcg
->kmemcg_id
= memcg_id
;
3482 memcg
->kmem_state
= KMEM_ONLINE
;
3483 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3488 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3490 struct cgroup_subsys_state
*css
;
3491 struct mem_cgroup
*parent
, *child
;
3494 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3497 * Clear the online state before clearing memcg_caches array
3498 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3499 * guarantees that no cache will be created for this cgroup
3500 * after we are done (see memcg_create_kmem_cache()).
3502 memcg
->kmem_state
= KMEM_ALLOCATED
;
3504 parent
= parent_mem_cgroup(memcg
);
3506 parent
= root_mem_cgroup
;
3509 * Deactivate and reparent kmem_caches. Then flush percpu
3510 * slab statistics to have precise values at the parent and
3511 * all ancestor levels. It's required to keep slab stats
3512 * accurate after the reparenting of kmem_caches.
3514 memcg_deactivate_kmem_caches(memcg
, parent
);
3515 memcg_flush_percpu_vmstats(memcg
, true);
3517 kmemcg_id
= memcg
->kmemcg_id
;
3518 BUG_ON(kmemcg_id
< 0);
3521 * Change kmemcg_id of this cgroup and all its descendants to the
3522 * parent's id, and then move all entries from this cgroup's list_lrus
3523 * to ones of the parent. After we have finished, all list_lrus
3524 * corresponding to this cgroup are guaranteed to remain empty. The
3525 * ordering is imposed by list_lru_node->lock taken by
3526 * memcg_drain_all_list_lrus().
3528 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3529 css_for_each_descendant_pre(css
, &memcg
->css
) {
3530 child
= mem_cgroup_from_css(css
);
3531 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3532 child
->kmemcg_id
= parent
->kmemcg_id
;
3533 if (!memcg
->use_hierarchy
)
3538 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3540 memcg_free_cache_id(kmemcg_id
);
3543 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3545 /* css_alloc() failed, offlining didn't happen */
3546 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3547 memcg_offline_kmem(memcg
);
3549 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3550 WARN_ON(!list_empty(&memcg
->kmem_caches
));
3551 static_branch_dec(&memcg_kmem_enabled_key
);
3555 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3559 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3562 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3565 #endif /* CONFIG_MEMCG_KMEM */
3567 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3572 mutex_lock(&memcg_max_mutex
);
3573 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3574 mutex_unlock(&memcg_max_mutex
);
3578 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3582 mutex_lock(&memcg_max_mutex
);
3584 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3588 if (!memcg
->tcpmem_active
) {
3590 * The active flag needs to be written after the static_key
3591 * update. This is what guarantees that the socket activation
3592 * function is the last one to run. See mem_cgroup_sk_alloc()
3593 * for details, and note that we don't mark any socket as
3594 * belonging to this memcg until that flag is up.
3596 * We need to do this, because static_keys will span multiple
3597 * sites, but we can't control their order. If we mark a socket
3598 * as accounted, but the accounting functions are not patched in
3599 * yet, we'll lose accounting.
3601 * We never race with the readers in mem_cgroup_sk_alloc(),
3602 * because when this value change, the code to process it is not
3605 static_branch_inc(&memcg_sockets_enabled_key
);
3606 memcg
->tcpmem_active
= true;
3609 mutex_unlock(&memcg_max_mutex
);
3614 * The user of this function is...
3617 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3618 char *buf
, size_t nbytes
, loff_t off
)
3620 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3621 unsigned long nr_pages
;
3624 buf
= strstrip(buf
);
3625 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3629 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3631 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3635 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3637 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3640 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3643 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3644 "Please report your usecase to linux-mm@kvack.org if you "
3645 "depend on this functionality.\n");
3646 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3649 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3653 case RES_SOFT_LIMIT
:
3654 memcg
->soft_limit
= nr_pages
;
3658 return ret
?: nbytes
;
3661 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3662 size_t nbytes
, loff_t off
)
3664 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3665 struct page_counter
*counter
;
3667 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3669 counter
= &memcg
->memory
;
3672 counter
= &memcg
->memsw
;
3675 counter
= &memcg
->kmem
;
3678 counter
= &memcg
->tcpmem
;
3684 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3686 page_counter_reset_watermark(counter
);
3689 counter
->failcnt
= 0;
3698 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3701 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3705 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3706 struct cftype
*cft
, u64 val
)
3708 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3710 if (val
& ~MOVE_MASK
)
3714 * No kind of locking is needed in here, because ->can_attach() will
3715 * check this value once in the beginning of the process, and then carry
3716 * on with stale data. This means that changes to this value will only
3717 * affect task migrations starting after the change.
3719 memcg
->move_charge_at_immigrate
= val
;
3723 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3724 struct cftype
*cft
, u64 val
)
3732 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3733 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3734 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3736 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3737 int nid
, unsigned int lru_mask
)
3739 struct lruvec
*lruvec
= mem_cgroup_lruvec(NODE_DATA(nid
), memcg
);
3740 unsigned long nr
= 0;
3743 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3746 if (!(BIT(lru
) & lru_mask
))
3748 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3753 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3754 unsigned int lru_mask
)
3756 unsigned long nr
= 0;
3760 if (!(BIT(lru
) & lru_mask
))
3762 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3767 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3771 unsigned int lru_mask
;
3774 static const struct numa_stat stats
[] = {
3775 { "total", LRU_ALL
},
3776 { "file", LRU_ALL_FILE
},
3777 { "anon", LRU_ALL_ANON
},
3778 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3780 const struct numa_stat
*stat
;
3783 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3785 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3786 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3787 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3788 for_each_node_state(nid
, N_MEMORY
) {
3789 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3791 seq_printf(m
, " N%d=%lu", nid
, nr
);
3796 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3797 struct mem_cgroup
*iter
;
3800 for_each_mem_cgroup_tree(iter
, memcg
)
3801 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3802 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3803 for_each_node_state(nid
, N_MEMORY
) {
3805 for_each_mem_cgroup_tree(iter
, memcg
)
3806 nr
+= mem_cgroup_node_nr_lru_pages(
3807 iter
, nid
, stat
->lru_mask
);
3808 seq_printf(m
, " N%d=%lu", nid
, nr
);
3815 #endif /* CONFIG_NUMA */
3817 static const unsigned int memcg1_stats
[] = {
3828 static const char *const memcg1_stat_names
[] = {
3839 /* Universal VM events cgroup1 shows, original sort order */
3840 static const unsigned int memcg1_events
[] = {
3847 static const char *const memcg1_event_names
[] = {
3854 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3856 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3857 unsigned long memory
, memsw
;
3858 struct mem_cgroup
*mi
;
3861 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3862 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3864 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3865 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3867 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3868 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3872 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3873 seq_printf(m
, "%s %lu\n", memcg1_event_names
[i
],
3874 memcg_events_local(memcg
, memcg1_events
[i
]));
3876 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3877 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3878 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3881 /* Hierarchical information */
3882 memory
= memsw
= PAGE_COUNTER_MAX
;
3883 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3884 memory
= min(memory
, mi
->memory
.max
);
3885 memsw
= min(memsw
, mi
->memsw
.max
);
3887 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3888 (u64
)memory
* PAGE_SIZE
);
3889 if (do_memsw_account())
3890 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3891 (u64
)memsw
* PAGE_SIZE
);
3893 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3894 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3896 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3897 (u64
)memcg_page_state(memcg
, memcg1_stats
[i
]) *
3901 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3902 seq_printf(m
, "total_%s %llu\n", memcg1_event_names
[i
],
3903 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
3905 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3906 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
],
3907 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3910 #ifdef CONFIG_DEBUG_VM
3913 struct mem_cgroup_per_node
*mz
;
3914 struct zone_reclaim_stat
*rstat
;
3915 unsigned long recent_rotated
[2] = {0, 0};
3916 unsigned long recent_scanned
[2] = {0, 0};
3918 for_each_online_pgdat(pgdat
) {
3919 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3920 rstat
= &mz
->lruvec
.reclaim_stat
;
3922 recent_rotated
[0] += rstat
->recent_rotated
[0];
3923 recent_rotated
[1] += rstat
->recent_rotated
[1];
3924 recent_scanned
[0] += rstat
->recent_scanned
[0];
3925 recent_scanned
[1] += rstat
->recent_scanned
[1];
3927 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3928 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3929 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3930 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3937 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3940 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3942 return mem_cgroup_swappiness(memcg
);
3945 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3946 struct cftype
*cft
, u64 val
)
3948 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3954 memcg
->swappiness
= val
;
3956 vm_swappiness
= val
;
3961 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3963 struct mem_cgroup_threshold_ary
*t
;
3964 unsigned long usage
;
3969 t
= rcu_dereference(memcg
->thresholds
.primary
);
3971 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3976 usage
= mem_cgroup_usage(memcg
, swap
);
3979 * current_threshold points to threshold just below or equal to usage.
3980 * If it's not true, a threshold was crossed after last
3981 * call of __mem_cgroup_threshold().
3983 i
= t
->current_threshold
;
3986 * Iterate backward over array of thresholds starting from
3987 * current_threshold and check if a threshold is crossed.
3988 * If none of thresholds below usage is crossed, we read
3989 * only one element of the array here.
3991 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3992 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3994 /* i = current_threshold + 1 */
3998 * Iterate forward over array of thresholds starting from
3999 * current_threshold+1 and check if a threshold is crossed.
4000 * If none of thresholds above usage is crossed, we read
4001 * only one element of the array here.
4003 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4004 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4006 /* Update current_threshold */
4007 t
->current_threshold
= i
- 1;
4012 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4015 __mem_cgroup_threshold(memcg
, false);
4016 if (do_memsw_account())
4017 __mem_cgroup_threshold(memcg
, true);
4019 memcg
= parent_mem_cgroup(memcg
);
4023 static int compare_thresholds(const void *a
, const void *b
)
4025 const struct mem_cgroup_threshold
*_a
= a
;
4026 const struct mem_cgroup_threshold
*_b
= b
;
4028 if (_a
->threshold
> _b
->threshold
)
4031 if (_a
->threshold
< _b
->threshold
)
4037 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4039 struct mem_cgroup_eventfd_list
*ev
;
4041 spin_lock(&memcg_oom_lock
);
4043 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4044 eventfd_signal(ev
->eventfd
, 1);
4046 spin_unlock(&memcg_oom_lock
);
4050 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4052 struct mem_cgroup
*iter
;
4054 for_each_mem_cgroup_tree(iter
, memcg
)
4055 mem_cgroup_oom_notify_cb(iter
);
4058 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4059 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4061 struct mem_cgroup_thresholds
*thresholds
;
4062 struct mem_cgroup_threshold_ary
*new;
4063 unsigned long threshold
;
4064 unsigned long usage
;
4067 ret
= page_counter_memparse(args
, "-1", &threshold
);
4071 mutex_lock(&memcg
->thresholds_lock
);
4074 thresholds
= &memcg
->thresholds
;
4075 usage
= mem_cgroup_usage(memcg
, false);
4076 } else if (type
== _MEMSWAP
) {
4077 thresholds
= &memcg
->memsw_thresholds
;
4078 usage
= mem_cgroup_usage(memcg
, true);
4082 /* Check if a threshold crossed before adding a new one */
4083 if (thresholds
->primary
)
4084 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4086 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4088 /* Allocate memory for new array of thresholds */
4089 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4096 /* Copy thresholds (if any) to new array */
4097 if (thresholds
->primary
) {
4098 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4099 sizeof(struct mem_cgroup_threshold
));
4102 /* Add new threshold */
4103 new->entries
[size
- 1].eventfd
= eventfd
;
4104 new->entries
[size
- 1].threshold
= threshold
;
4106 /* Sort thresholds. Registering of new threshold isn't time-critical */
4107 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4108 compare_thresholds
, NULL
);
4110 /* Find current threshold */
4111 new->current_threshold
= -1;
4112 for (i
= 0; i
< size
; i
++) {
4113 if (new->entries
[i
].threshold
<= usage
) {
4115 * new->current_threshold will not be used until
4116 * rcu_assign_pointer(), so it's safe to increment
4119 ++new->current_threshold
;
4124 /* Free old spare buffer and save old primary buffer as spare */
4125 kfree(thresholds
->spare
);
4126 thresholds
->spare
= thresholds
->primary
;
4128 rcu_assign_pointer(thresholds
->primary
, new);
4130 /* To be sure that nobody uses thresholds */
4134 mutex_unlock(&memcg
->thresholds_lock
);
4139 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4140 struct eventfd_ctx
*eventfd
, const char *args
)
4142 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4145 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4146 struct eventfd_ctx
*eventfd
, const char *args
)
4148 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4151 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4152 struct eventfd_ctx
*eventfd
, enum res_type type
)
4154 struct mem_cgroup_thresholds
*thresholds
;
4155 struct mem_cgroup_threshold_ary
*new;
4156 unsigned long usage
;
4159 mutex_lock(&memcg
->thresholds_lock
);
4162 thresholds
= &memcg
->thresholds
;
4163 usage
= mem_cgroup_usage(memcg
, false);
4164 } else if (type
== _MEMSWAP
) {
4165 thresholds
= &memcg
->memsw_thresholds
;
4166 usage
= mem_cgroup_usage(memcg
, true);
4170 if (!thresholds
->primary
)
4173 /* Check if a threshold crossed before removing */
4174 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4176 /* Calculate new number of threshold */
4178 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4179 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4183 new = thresholds
->spare
;
4185 /* Set thresholds array to NULL if we don't have thresholds */
4194 /* Copy thresholds and find current threshold */
4195 new->current_threshold
= -1;
4196 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4197 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4200 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4201 if (new->entries
[j
].threshold
<= usage
) {
4203 * new->current_threshold will not be used
4204 * until rcu_assign_pointer(), so it's safe to increment
4207 ++new->current_threshold
;
4213 /* Swap primary and spare array */
4214 thresholds
->spare
= thresholds
->primary
;
4216 rcu_assign_pointer(thresholds
->primary
, new);
4218 /* To be sure that nobody uses thresholds */
4221 /* If all events are unregistered, free the spare array */
4223 kfree(thresholds
->spare
);
4224 thresholds
->spare
= NULL
;
4227 mutex_unlock(&memcg
->thresholds_lock
);
4230 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4231 struct eventfd_ctx
*eventfd
)
4233 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4236 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4237 struct eventfd_ctx
*eventfd
)
4239 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4242 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4243 struct eventfd_ctx
*eventfd
, const char *args
)
4245 struct mem_cgroup_eventfd_list
*event
;
4247 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4251 spin_lock(&memcg_oom_lock
);
4253 event
->eventfd
= eventfd
;
4254 list_add(&event
->list
, &memcg
->oom_notify
);
4256 /* already in OOM ? */
4257 if (memcg
->under_oom
)
4258 eventfd_signal(eventfd
, 1);
4259 spin_unlock(&memcg_oom_lock
);
4264 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4265 struct eventfd_ctx
*eventfd
)
4267 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4269 spin_lock(&memcg_oom_lock
);
4271 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4272 if (ev
->eventfd
== eventfd
) {
4273 list_del(&ev
->list
);
4278 spin_unlock(&memcg_oom_lock
);
4281 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4283 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4285 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4286 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4287 seq_printf(sf
, "oom_kill %lu\n",
4288 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4292 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4293 struct cftype
*cft
, u64 val
)
4295 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4297 /* cannot set to root cgroup and only 0 and 1 are allowed */
4298 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4301 memcg
->oom_kill_disable
= val
;
4303 memcg_oom_recover(memcg
);
4308 #ifdef CONFIG_CGROUP_WRITEBACK
4310 #include <trace/events/writeback.h>
4312 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4314 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4317 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4319 wb_domain_exit(&memcg
->cgwb_domain
);
4322 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4324 wb_domain_size_changed(&memcg
->cgwb_domain
);
4327 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4329 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4331 if (!memcg
->css
.parent
)
4334 return &memcg
->cgwb_domain
;
4338 * idx can be of type enum memcg_stat_item or node_stat_item.
4339 * Keep in sync with memcg_exact_page().
4341 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4343 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4346 for_each_online_cpu(cpu
)
4347 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4354 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4355 * @wb: bdi_writeback in question
4356 * @pfilepages: out parameter for number of file pages
4357 * @pheadroom: out parameter for number of allocatable pages according to memcg
4358 * @pdirty: out parameter for number of dirty pages
4359 * @pwriteback: out parameter for number of pages under writeback
4361 * Determine the numbers of file, headroom, dirty, and writeback pages in
4362 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4363 * is a bit more involved.
4365 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4366 * headroom is calculated as the lowest headroom of itself and the
4367 * ancestors. Note that this doesn't consider the actual amount of
4368 * available memory in the system. The caller should further cap
4369 * *@pheadroom accordingly.
4371 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4372 unsigned long *pheadroom
, unsigned long *pdirty
,
4373 unsigned long *pwriteback
)
4375 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4376 struct mem_cgroup
*parent
;
4378 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4380 /* this should eventually include NR_UNSTABLE_NFS */
4381 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4382 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4383 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4384 *pheadroom
= PAGE_COUNTER_MAX
;
4386 while ((parent
= parent_mem_cgroup(memcg
))) {
4387 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
4388 unsigned long used
= page_counter_read(&memcg
->memory
);
4390 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4396 * Foreign dirty flushing
4398 * There's an inherent mismatch between memcg and writeback. The former
4399 * trackes ownership per-page while the latter per-inode. This was a
4400 * deliberate design decision because honoring per-page ownership in the
4401 * writeback path is complicated, may lead to higher CPU and IO overheads
4402 * and deemed unnecessary given that write-sharing an inode across
4403 * different cgroups isn't a common use-case.
4405 * Combined with inode majority-writer ownership switching, this works well
4406 * enough in most cases but there are some pathological cases. For
4407 * example, let's say there are two cgroups A and B which keep writing to
4408 * different but confined parts of the same inode. B owns the inode and
4409 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4410 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4411 * triggering background writeback. A will be slowed down without a way to
4412 * make writeback of the dirty pages happen.
4414 * Conditions like the above can lead to a cgroup getting repatedly and
4415 * severely throttled after making some progress after each
4416 * dirty_expire_interval while the underyling IO device is almost
4419 * Solving this problem completely requires matching the ownership tracking
4420 * granularities between memcg and writeback in either direction. However,
4421 * the more egregious behaviors can be avoided by simply remembering the
4422 * most recent foreign dirtying events and initiating remote flushes on
4423 * them when local writeback isn't enough to keep the memory clean enough.
4425 * The following two functions implement such mechanism. When a foreign
4426 * page - a page whose memcg and writeback ownerships don't match - is
4427 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4428 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4429 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4430 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4431 * foreign bdi_writebacks which haven't expired. Both the numbers of
4432 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4433 * limited to MEMCG_CGWB_FRN_CNT.
4435 * The mechanism only remembers IDs and doesn't hold any object references.
4436 * As being wrong occasionally doesn't matter, updates and accesses to the
4437 * records are lockless and racy.
4439 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4440 struct bdi_writeback
*wb
)
4442 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4443 struct memcg_cgwb_frn
*frn
;
4444 u64 now
= get_jiffies_64();
4445 u64 oldest_at
= now
;
4449 trace_track_foreign_dirty(page
, wb
);
4452 * Pick the slot to use. If there is already a slot for @wb, keep
4453 * using it. If not replace the oldest one which isn't being
4456 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4457 frn
= &memcg
->cgwb_frn
[i
];
4458 if (frn
->bdi_id
== wb
->bdi
->id
&&
4459 frn
->memcg_id
== wb
->memcg_css
->id
)
4461 if (time_before64(frn
->at
, oldest_at
) &&
4462 atomic_read(&frn
->done
.cnt
) == 1) {
4464 oldest_at
= frn
->at
;
4468 if (i
< MEMCG_CGWB_FRN_CNT
) {
4470 * Re-using an existing one. Update timestamp lazily to
4471 * avoid making the cacheline hot. We want them to be
4472 * reasonably up-to-date and significantly shorter than
4473 * dirty_expire_interval as that's what expires the record.
4474 * Use the shorter of 1s and dirty_expire_interval / 8.
4476 unsigned long update_intv
=
4477 min_t(unsigned long, HZ
,
4478 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4480 if (time_before64(frn
->at
, now
- update_intv
))
4482 } else if (oldest
>= 0) {
4483 /* replace the oldest free one */
4484 frn
= &memcg
->cgwb_frn
[oldest
];
4485 frn
->bdi_id
= wb
->bdi
->id
;
4486 frn
->memcg_id
= wb
->memcg_css
->id
;
4491 /* issue foreign writeback flushes for recorded foreign dirtying events */
4492 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4494 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4495 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4496 u64 now
= jiffies_64
;
4499 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4500 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4503 * If the record is older than dirty_expire_interval,
4504 * writeback on it has already started. No need to kick it
4505 * off again. Also, don't start a new one if there's
4506 * already one in flight.
4508 if (time_after64(frn
->at
, now
- intv
) &&
4509 atomic_read(&frn
->done
.cnt
) == 1) {
4511 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4512 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4513 WB_REASON_FOREIGN_FLUSH
,
4519 #else /* CONFIG_CGROUP_WRITEBACK */
4521 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4526 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4530 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4534 #endif /* CONFIG_CGROUP_WRITEBACK */
4537 * DO NOT USE IN NEW FILES.
4539 * "cgroup.event_control" implementation.
4541 * This is way over-engineered. It tries to support fully configurable
4542 * events for each user. Such level of flexibility is completely
4543 * unnecessary especially in the light of the planned unified hierarchy.
4545 * Please deprecate this and replace with something simpler if at all
4550 * Unregister event and free resources.
4552 * Gets called from workqueue.
4554 static void memcg_event_remove(struct work_struct
*work
)
4556 struct mem_cgroup_event
*event
=
4557 container_of(work
, struct mem_cgroup_event
, remove
);
4558 struct mem_cgroup
*memcg
= event
->memcg
;
4560 remove_wait_queue(event
->wqh
, &event
->wait
);
4562 event
->unregister_event(memcg
, event
->eventfd
);
4564 /* Notify userspace the event is going away. */
4565 eventfd_signal(event
->eventfd
, 1);
4567 eventfd_ctx_put(event
->eventfd
);
4569 css_put(&memcg
->css
);
4573 * Gets called on EPOLLHUP on eventfd when user closes it.
4575 * Called with wqh->lock held and interrupts disabled.
4577 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4578 int sync
, void *key
)
4580 struct mem_cgroup_event
*event
=
4581 container_of(wait
, struct mem_cgroup_event
, wait
);
4582 struct mem_cgroup
*memcg
= event
->memcg
;
4583 __poll_t flags
= key_to_poll(key
);
4585 if (flags
& EPOLLHUP
) {
4587 * If the event has been detached at cgroup removal, we
4588 * can simply return knowing the other side will cleanup
4591 * We can't race against event freeing since the other
4592 * side will require wqh->lock via remove_wait_queue(),
4595 spin_lock(&memcg
->event_list_lock
);
4596 if (!list_empty(&event
->list
)) {
4597 list_del_init(&event
->list
);
4599 * We are in atomic context, but cgroup_event_remove()
4600 * may sleep, so we have to call it in workqueue.
4602 schedule_work(&event
->remove
);
4604 spin_unlock(&memcg
->event_list_lock
);
4610 static void memcg_event_ptable_queue_proc(struct file
*file
,
4611 wait_queue_head_t
*wqh
, poll_table
*pt
)
4613 struct mem_cgroup_event
*event
=
4614 container_of(pt
, struct mem_cgroup_event
, pt
);
4617 add_wait_queue(wqh
, &event
->wait
);
4621 * DO NOT USE IN NEW FILES.
4623 * Parse input and register new cgroup event handler.
4625 * Input must be in format '<event_fd> <control_fd> <args>'.
4626 * Interpretation of args is defined by control file implementation.
4628 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4629 char *buf
, size_t nbytes
, loff_t off
)
4631 struct cgroup_subsys_state
*css
= of_css(of
);
4632 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4633 struct mem_cgroup_event
*event
;
4634 struct cgroup_subsys_state
*cfile_css
;
4635 unsigned int efd
, cfd
;
4642 buf
= strstrip(buf
);
4644 efd
= simple_strtoul(buf
, &endp
, 10);
4649 cfd
= simple_strtoul(buf
, &endp
, 10);
4650 if ((*endp
!= ' ') && (*endp
!= '\0'))
4654 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4658 event
->memcg
= memcg
;
4659 INIT_LIST_HEAD(&event
->list
);
4660 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4661 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4662 INIT_WORK(&event
->remove
, memcg_event_remove
);
4670 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4671 if (IS_ERR(event
->eventfd
)) {
4672 ret
= PTR_ERR(event
->eventfd
);
4679 goto out_put_eventfd
;
4682 /* the process need read permission on control file */
4683 /* AV: shouldn't we check that it's been opened for read instead? */
4684 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4689 * Determine the event callbacks and set them in @event. This used
4690 * to be done via struct cftype but cgroup core no longer knows
4691 * about these events. The following is crude but the whole thing
4692 * is for compatibility anyway.
4694 * DO NOT ADD NEW FILES.
4696 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4698 if (!strcmp(name
, "memory.usage_in_bytes")) {
4699 event
->register_event
= mem_cgroup_usage_register_event
;
4700 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4701 } else if (!strcmp(name
, "memory.oom_control")) {
4702 event
->register_event
= mem_cgroup_oom_register_event
;
4703 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4704 } else if (!strcmp(name
, "memory.pressure_level")) {
4705 event
->register_event
= vmpressure_register_event
;
4706 event
->unregister_event
= vmpressure_unregister_event
;
4707 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4708 event
->register_event
= memsw_cgroup_usage_register_event
;
4709 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4716 * Verify @cfile should belong to @css. Also, remaining events are
4717 * automatically removed on cgroup destruction but the removal is
4718 * asynchronous, so take an extra ref on @css.
4720 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4721 &memory_cgrp_subsys
);
4723 if (IS_ERR(cfile_css
))
4725 if (cfile_css
!= css
) {
4730 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4734 vfs_poll(efile
.file
, &event
->pt
);
4736 spin_lock(&memcg
->event_list_lock
);
4737 list_add(&event
->list
, &memcg
->event_list
);
4738 spin_unlock(&memcg
->event_list_lock
);
4750 eventfd_ctx_put(event
->eventfd
);
4759 static struct cftype mem_cgroup_legacy_files
[] = {
4761 .name
= "usage_in_bytes",
4762 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4763 .read_u64
= mem_cgroup_read_u64
,
4766 .name
= "max_usage_in_bytes",
4767 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4768 .write
= mem_cgroup_reset
,
4769 .read_u64
= mem_cgroup_read_u64
,
4772 .name
= "limit_in_bytes",
4773 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4774 .write
= mem_cgroup_write
,
4775 .read_u64
= mem_cgroup_read_u64
,
4778 .name
= "soft_limit_in_bytes",
4779 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4780 .write
= mem_cgroup_write
,
4781 .read_u64
= mem_cgroup_read_u64
,
4785 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4786 .write
= mem_cgroup_reset
,
4787 .read_u64
= mem_cgroup_read_u64
,
4791 .seq_show
= memcg_stat_show
,
4794 .name
= "force_empty",
4795 .write
= mem_cgroup_force_empty_write
,
4798 .name
= "use_hierarchy",
4799 .write_u64
= mem_cgroup_hierarchy_write
,
4800 .read_u64
= mem_cgroup_hierarchy_read
,
4803 .name
= "cgroup.event_control", /* XXX: for compat */
4804 .write
= memcg_write_event_control
,
4805 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4808 .name
= "swappiness",
4809 .read_u64
= mem_cgroup_swappiness_read
,
4810 .write_u64
= mem_cgroup_swappiness_write
,
4813 .name
= "move_charge_at_immigrate",
4814 .read_u64
= mem_cgroup_move_charge_read
,
4815 .write_u64
= mem_cgroup_move_charge_write
,
4818 .name
= "oom_control",
4819 .seq_show
= mem_cgroup_oom_control_read
,
4820 .write_u64
= mem_cgroup_oom_control_write
,
4821 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4824 .name
= "pressure_level",
4828 .name
= "numa_stat",
4829 .seq_show
= memcg_numa_stat_show
,
4833 .name
= "kmem.limit_in_bytes",
4834 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4835 .write
= mem_cgroup_write
,
4836 .read_u64
= mem_cgroup_read_u64
,
4839 .name
= "kmem.usage_in_bytes",
4840 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4841 .read_u64
= mem_cgroup_read_u64
,
4844 .name
= "kmem.failcnt",
4845 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4846 .write
= mem_cgroup_reset
,
4847 .read_u64
= mem_cgroup_read_u64
,
4850 .name
= "kmem.max_usage_in_bytes",
4851 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4852 .write
= mem_cgroup_reset
,
4853 .read_u64
= mem_cgroup_read_u64
,
4855 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4857 .name
= "kmem.slabinfo",
4858 .seq_start
= memcg_slab_start
,
4859 .seq_next
= memcg_slab_next
,
4860 .seq_stop
= memcg_slab_stop
,
4861 .seq_show
= memcg_slab_show
,
4865 .name
= "kmem.tcp.limit_in_bytes",
4866 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4867 .write
= mem_cgroup_write
,
4868 .read_u64
= mem_cgroup_read_u64
,
4871 .name
= "kmem.tcp.usage_in_bytes",
4872 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4873 .read_u64
= mem_cgroup_read_u64
,
4876 .name
= "kmem.tcp.failcnt",
4877 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4878 .write
= mem_cgroup_reset
,
4879 .read_u64
= mem_cgroup_read_u64
,
4882 .name
= "kmem.tcp.max_usage_in_bytes",
4883 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4884 .write
= mem_cgroup_reset
,
4885 .read_u64
= mem_cgroup_read_u64
,
4887 { }, /* terminate */
4891 * Private memory cgroup IDR
4893 * Swap-out records and page cache shadow entries need to store memcg
4894 * references in constrained space, so we maintain an ID space that is
4895 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4896 * memory-controlled cgroups to 64k.
4898 * However, there usually are many references to the oflline CSS after
4899 * the cgroup has been destroyed, such as page cache or reclaimable
4900 * slab objects, that don't need to hang on to the ID. We want to keep
4901 * those dead CSS from occupying IDs, or we might quickly exhaust the
4902 * relatively small ID space and prevent the creation of new cgroups
4903 * even when there are much fewer than 64k cgroups - possibly none.
4905 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4906 * be freed and recycled when it's no longer needed, which is usually
4907 * when the CSS is offlined.
4909 * The only exception to that are records of swapped out tmpfs/shmem
4910 * pages that need to be attributed to live ancestors on swapin. But
4911 * those references are manageable from userspace.
4914 static DEFINE_IDR(mem_cgroup_idr
);
4916 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4918 if (memcg
->id
.id
> 0) {
4919 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4924 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4926 refcount_add(n
, &memcg
->id
.ref
);
4929 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4931 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4932 mem_cgroup_id_remove(memcg
);
4934 /* Memcg ID pins CSS */
4935 css_put(&memcg
->css
);
4939 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4941 mem_cgroup_id_put_many(memcg
, 1);
4945 * mem_cgroup_from_id - look up a memcg from a memcg id
4946 * @id: the memcg id to look up
4948 * Caller must hold rcu_read_lock().
4950 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4952 WARN_ON_ONCE(!rcu_read_lock_held());
4953 return idr_find(&mem_cgroup_idr
, id
);
4956 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4958 struct mem_cgroup_per_node
*pn
;
4961 * This routine is called against possible nodes.
4962 * But it's BUG to call kmalloc() against offline node.
4964 * TODO: this routine can waste much memory for nodes which will
4965 * never be onlined. It's better to use memory hotplug callback
4968 if (!node_state(node
, N_NORMAL_MEMORY
))
4970 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4974 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
4975 if (!pn
->lruvec_stat_local
) {
4980 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4981 if (!pn
->lruvec_stat_cpu
) {
4982 free_percpu(pn
->lruvec_stat_local
);
4987 lruvec_init(&pn
->lruvec
);
4988 pn
->usage_in_excess
= 0;
4989 pn
->on_tree
= false;
4992 memcg
->nodeinfo
[node
] = pn
;
4996 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4998 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5003 free_percpu(pn
->lruvec_stat_cpu
);
5004 free_percpu(pn
->lruvec_stat_local
);
5008 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5013 * Flush percpu vmstats and vmevents to guarantee the value correctness
5014 * on parent's and all ancestor levels.
5016 memcg_flush_percpu_vmstats(memcg
, false);
5017 memcg_flush_percpu_vmevents(memcg
);
5019 free_mem_cgroup_per_node_info(memcg
, node
);
5020 free_percpu(memcg
->vmstats_percpu
);
5021 free_percpu(memcg
->vmstats_local
);
5025 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5027 memcg_wb_domain_exit(memcg
);
5028 __mem_cgroup_free(memcg
);
5031 static struct mem_cgroup
*mem_cgroup_alloc(void)
5033 struct mem_cgroup
*memcg
;
5036 int __maybe_unused i
;
5038 size
= sizeof(struct mem_cgroup
);
5039 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5041 memcg
= kzalloc(size
, GFP_KERNEL
);
5045 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5046 1, MEM_CGROUP_ID_MAX
,
5048 if (memcg
->id
.id
< 0)
5051 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
5052 if (!memcg
->vmstats_local
)
5055 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
5056 if (!memcg
->vmstats_percpu
)
5060 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5063 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5066 INIT_WORK(&memcg
->high_work
, high_work_func
);
5067 memcg
->last_scanned_node
= MAX_NUMNODES
;
5068 INIT_LIST_HEAD(&memcg
->oom_notify
);
5069 mutex_init(&memcg
->thresholds_lock
);
5070 spin_lock_init(&memcg
->move_lock
);
5071 vmpressure_init(&memcg
->vmpressure
);
5072 INIT_LIST_HEAD(&memcg
->event_list
);
5073 spin_lock_init(&memcg
->event_list_lock
);
5074 memcg
->socket_pressure
= jiffies
;
5075 #ifdef CONFIG_MEMCG_KMEM
5076 memcg
->kmemcg_id
= -1;
5078 #ifdef CONFIG_CGROUP_WRITEBACK
5079 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5080 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5081 memcg
->cgwb_frn
[i
].done
=
5082 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5084 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5085 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5086 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5087 memcg
->deferred_split_queue
.split_queue_len
= 0;
5089 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5092 mem_cgroup_id_remove(memcg
);
5093 __mem_cgroup_free(memcg
);
5097 static struct cgroup_subsys_state
* __ref
5098 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5100 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5101 struct mem_cgroup
*memcg
;
5102 long error
= -ENOMEM
;
5104 memcg
= mem_cgroup_alloc();
5106 return ERR_PTR(error
);
5108 memcg
->high
= PAGE_COUNTER_MAX
;
5109 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5111 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5112 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5114 if (parent
&& parent
->use_hierarchy
) {
5115 memcg
->use_hierarchy
= true;
5116 page_counter_init(&memcg
->memory
, &parent
->memory
);
5117 page_counter_init(&memcg
->swap
, &parent
->swap
);
5118 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
5119 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5120 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5122 page_counter_init(&memcg
->memory
, NULL
);
5123 page_counter_init(&memcg
->swap
, NULL
);
5124 page_counter_init(&memcg
->memsw
, NULL
);
5125 page_counter_init(&memcg
->kmem
, NULL
);
5126 page_counter_init(&memcg
->tcpmem
, NULL
);
5128 * Deeper hierachy with use_hierarchy == false doesn't make
5129 * much sense so let cgroup subsystem know about this
5130 * unfortunate state in our controller.
5132 if (parent
!= root_mem_cgroup
)
5133 memory_cgrp_subsys
.broken_hierarchy
= true;
5136 /* The following stuff does not apply to the root */
5138 #ifdef CONFIG_MEMCG_KMEM
5139 INIT_LIST_HEAD(&memcg
->kmem_caches
);
5141 root_mem_cgroup
= memcg
;
5145 error
= memcg_online_kmem(memcg
);
5149 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5150 static_branch_inc(&memcg_sockets_enabled_key
);
5154 mem_cgroup_id_remove(memcg
);
5155 mem_cgroup_free(memcg
);
5156 return ERR_PTR(-ENOMEM
);
5159 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5161 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5164 * A memcg must be visible for memcg_expand_shrinker_maps()
5165 * by the time the maps are allocated. So, we allocate maps
5166 * here, when for_each_mem_cgroup() can't skip it.
5168 if (memcg_alloc_shrinker_maps(memcg
)) {
5169 mem_cgroup_id_remove(memcg
);
5173 /* Online state pins memcg ID, memcg ID pins CSS */
5174 refcount_set(&memcg
->id
.ref
, 1);
5179 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5181 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5182 struct mem_cgroup_event
*event
, *tmp
;
5185 * Unregister events and notify userspace.
5186 * Notify userspace about cgroup removing only after rmdir of cgroup
5187 * directory to avoid race between userspace and kernelspace.
5189 spin_lock(&memcg
->event_list_lock
);
5190 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5191 list_del_init(&event
->list
);
5192 schedule_work(&event
->remove
);
5194 spin_unlock(&memcg
->event_list_lock
);
5196 page_counter_set_min(&memcg
->memory
, 0);
5197 page_counter_set_low(&memcg
->memory
, 0);
5199 memcg_offline_kmem(memcg
);
5200 wb_memcg_offline(memcg
);
5202 drain_all_stock(memcg
);
5204 mem_cgroup_id_put(memcg
);
5207 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5209 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5211 invalidate_reclaim_iterators(memcg
);
5214 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5216 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5217 int __maybe_unused i
;
5219 #ifdef CONFIG_CGROUP_WRITEBACK
5220 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5221 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5223 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5224 static_branch_dec(&memcg_sockets_enabled_key
);
5226 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5227 static_branch_dec(&memcg_sockets_enabled_key
);
5229 vmpressure_cleanup(&memcg
->vmpressure
);
5230 cancel_work_sync(&memcg
->high_work
);
5231 mem_cgroup_remove_from_trees(memcg
);
5232 memcg_free_shrinker_maps(memcg
);
5233 memcg_free_kmem(memcg
);
5234 mem_cgroup_free(memcg
);
5238 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5239 * @css: the target css
5241 * Reset the states of the mem_cgroup associated with @css. This is
5242 * invoked when the userland requests disabling on the default hierarchy
5243 * but the memcg is pinned through dependency. The memcg should stop
5244 * applying policies and should revert to the vanilla state as it may be
5245 * made visible again.
5247 * The current implementation only resets the essential configurations.
5248 * This needs to be expanded to cover all the visible parts.
5250 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5252 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5254 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5255 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5256 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
5257 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5258 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5259 page_counter_set_min(&memcg
->memory
, 0);
5260 page_counter_set_low(&memcg
->memory
, 0);
5261 memcg
->high
= PAGE_COUNTER_MAX
;
5262 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5263 memcg_wb_domain_size_changed(memcg
);
5267 /* Handlers for move charge at task migration. */
5268 static int mem_cgroup_do_precharge(unsigned long count
)
5272 /* Try a single bulk charge without reclaim first, kswapd may wake */
5273 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5275 mc
.precharge
+= count
;
5279 /* Try charges one by one with reclaim, but do not retry */
5281 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5295 enum mc_target_type
{
5302 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5303 unsigned long addr
, pte_t ptent
)
5305 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5307 if (!page
|| !page_mapped(page
))
5309 if (PageAnon(page
)) {
5310 if (!(mc
.flags
& MOVE_ANON
))
5313 if (!(mc
.flags
& MOVE_FILE
))
5316 if (!get_page_unless_zero(page
))
5322 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5323 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5324 pte_t ptent
, swp_entry_t
*entry
)
5326 struct page
*page
= NULL
;
5327 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5329 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
5333 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5334 * a device and because they are not accessible by CPU they are store
5335 * as special swap entry in the CPU page table.
5337 if (is_device_private_entry(ent
)) {
5338 page
= device_private_entry_to_page(ent
);
5340 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5341 * a refcount of 1 when free (unlike normal page)
5343 if (!page_ref_add_unless(page
, 1, 1))
5349 * Because lookup_swap_cache() updates some statistics counter,
5350 * we call find_get_page() with swapper_space directly.
5352 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5353 if (do_memsw_account())
5354 entry
->val
= ent
.val
;
5359 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5360 pte_t ptent
, swp_entry_t
*entry
)
5366 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5367 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5369 struct page
*page
= NULL
;
5370 struct address_space
*mapping
;
5373 if (!vma
->vm_file
) /* anonymous vma */
5375 if (!(mc
.flags
& MOVE_FILE
))
5378 mapping
= vma
->vm_file
->f_mapping
;
5379 pgoff
= linear_page_index(vma
, addr
);
5381 /* page is moved even if it's not RSS of this task(page-faulted). */
5383 /* shmem/tmpfs may report page out on swap: account for that too. */
5384 if (shmem_mapping(mapping
)) {
5385 page
= find_get_entry(mapping
, pgoff
);
5386 if (xa_is_value(page
)) {
5387 swp_entry_t swp
= radix_to_swp_entry(page
);
5388 if (do_memsw_account())
5390 page
= find_get_page(swap_address_space(swp
),
5394 page
= find_get_page(mapping
, pgoff
);
5396 page
= find_get_page(mapping
, pgoff
);
5402 * mem_cgroup_move_account - move account of the page
5404 * @compound: charge the page as compound or small page
5405 * @from: mem_cgroup which the page is moved from.
5406 * @to: mem_cgroup which the page is moved to. @from != @to.
5408 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5410 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5413 static int mem_cgroup_move_account(struct page
*page
,
5415 struct mem_cgroup
*from
,
5416 struct mem_cgroup
*to
)
5418 unsigned long flags
;
5419 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5423 VM_BUG_ON(from
== to
);
5424 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5425 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5428 * Prevent mem_cgroup_migrate() from looking at
5429 * page->mem_cgroup of its source page while we change it.
5432 if (!trylock_page(page
))
5436 if (page
->mem_cgroup
!= from
)
5439 anon
= PageAnon(page
);
5441 spin_lock_irqsave(&from
->move_lock
, flags
);
5443 if (!anon
&& page_mapped(page
)) {
5444 __mod_memcg_state(from
, NR_FILE_MAPPED
, -nr_pages
);
5445 __mod_memcg_state(to
, NR_FILE_MAPPED
, nr_pages
);
5449 * move_lock grabbed above and caller set from->moving_account, so
5450 * mod_memcg_page_state will serialize updates to PageDirty.
5451 * So mapping should be stable for dirty pages.
5453 if (!anon
&& PageDirty(page
)) {
5454 struct address_space
*mapping
= page_mapping(page
);
5456 if (mapping_cap_account_dirty(mapping
)) {
5457 __mod_memcg_state(from
, NR_FILE_DIRTY
, -nr_pages
);
5458 __mod_memcg_state(to
, NR_FILE_DIRTY
, nr_pages
);
5462 if (PageWriteback(page
)) {
5463 __mod_memcg_state(from
, NR_WRITEBACK
, -nr_pages
);
5464 __mod_memcg_state(to
, NR_WRITEBACK
, nr_pages
);
5467 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5468 if (compound
&& !list_empty(page_deferred_list(page
))) {
5469 spin_lock(&from
->deferred_split_queue
.split_queue_lock
);
5470 list_del_init(page_deferred_list(page
));
5471 from
->deferred_split_queue
.split_queue_len
--;
5472 spin_unlock(&from
->deferred_split_queue
.split_queue_lock
);
5476 * It is safe to change page->mem_cgroup here because the page
5477 * is referenced, charged, and isolated - we can't race with
5478 * uncharging, charging, migration, or LRU putback.
5481 /* caller should have done css_get */
5482 page
->mem_cgroup
= to
;
5484 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5485 if (compound
&& list_empty(page_deferred_list(page
))) {
5486 spin_lock(&to
->deferred_split_queue
.split_queue_lock
);
5487 list_add_tail(page_deferred_list(page
),
5488 &to
->deferred_split_queue
.split_queue
);
5489 to
->deferred_split_queue
.split_queue_len
++;
5490 spin_unlock(&to
->deferred_split_queue
.split_queue_lock
);
5494 spin_unlock_irqrestore(&from
->move_lock
, flags
);
5498 local_irq_disable();
5499 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
5500 memcg_check_events(to
, page
);
5501 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
5502 memcg_check_events(from
, page
);
5511 * get_mctgt_type - get target type of moving charge
5512 * @vma: the vma the pte to be checked belongs
5513 * @addr: the address corresponding to the pte to be checked
5514 * @ptent: the pte to be checked
5515 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5518 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5519 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5520 * move charge. if @target is not NULL, the page is stored in target->page
5521 * with extra refcnt got(Callers should handle it).
5522 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5523 * target for charge migration. if @target is not NULL, the entry is stored
5525 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5526 * (so ZONE_DEVICE page and thus not on the lru).
5527 * For now we such page is charge like a regular page would be as for all
5528 * intent and purposes it is just special memory taking the place of a
5531 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5533 * Called with pte lock held.
5536 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5537 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5539 struct page
*page
= NULL
;
5540 enum mc_target_type ret
= MC_TARGET_NONE
;
5541 swp_entry_t ent
= { .val
= 0 };
5543 if (pte_present(ptent
))
5544 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5545 else if (is_swap_pte(ptent
))
5546 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5547 else if (pte_none(ptent
))
5548 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5550 if (!page
&& !ent
.val
)
5554 * Do only loose check w/o serialization.
5555 * mem_cgroup_move_account() checks the page is valid or
5556 * not under LRU exclusion.
5558 if (page
->mem_cgroup
== mc
.from
) {
5559 ret
= MC_TARGET_PAGE
;
5560 if (is_device_private_page(page
))
5561 ret
= MC_TARGET_DEVICE
;
5563 target
->page
= page
;
5565 if (!ret
|| !target
)
5569 * There is a swap entry and a page doesn't exist or isn't charged.
5570 * But we cannot move a tail-page in a THP.
5572 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5573 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5574 ret
= MC_TARGET_SWAP
;
5581 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5583 * We don't consider PMD mapped swapping or file mapped pages because THP does
5584 * not support them for now.
5585 * Caller should make sure that pmd_trans_huge(pmd) is true.
5587 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5588 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5590 struct page
*page
= NULL
;
5591 enum mc_target_type ret
= MC_TARGET_NONE
;
5593 if (unlikely(is_swap_pmd(pmd
))) {
5594 VM_BUG_ON(thp_migration_supported() &&
5595 !is_pmd_migration_entry(pmd
));
5598 page
= pmd_page(pmd
);
5599 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5600 if (!(mc
.flags
& MOVE_ANON
))
5602 if (page
->mem_cgroup
== mc
.from
) {
5603 ret
= MC_TARGET_PAGE
;
5606 target
->page
= page
;
5612 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5613 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5615 return MC_TARGET_NONE
;
5619 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5620 unsigned long addr
, unsigned long end
,
5621 struct mm_walk
*walk
)
5623 struct vm_area_struct
*vma
= walk
->vma
;
5627 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5630 * Note their can not be MC_TARGET_DEVICE for now as we do not
5631 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5632 * this might change.
5634 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5635 mc
.precharge
+= HPAGE_PMD_NR
;
5640 if (pmd_trans_unstable(pmd
))
5642 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5643 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5644 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5645 mc
.precharge
++; /* increment precharge temporarily */
5646 pte_unmap_unlock(pte
- 1, ptl
);
5652 static const struct mm_walk_ops precharge_walk_ops
= {
5653 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5656 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5658 unsigned long precharge
;
5660 down_read(&mm
->mmap_sem
);
5661 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5662 up_read(&mm
->mmap_sem
);
5664 precharge
= mc
.precharge
;
5670 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5672 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5674 VM_BUG_ON(mc
.moving_task
);
5675 mc
.moving_task
= current
;
5676 return mem_cgroup_do_precharge(precharge
);
5679 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5680 static void __mem_cgroup_clear_mc(void)
5682 struct mem_cgroup
*from
= mc
.from
;
5683 struct mem_cgroup
*to
= mc
.to
;
5685 /* we must uncharge all the leftover precharges from mc.to */
5687 cancel_charge(mc
.to
, mc
.precharge
);
5691 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5692 * we must uncharge here.
5694 if (mc
.moved_charge
) {
5695 cancel_charge(mc
.from
, mc
.moved_charge
);
5696 mc
.moved_charge
= 0;
5698 /* we must fixup refcnts and charges */
5699 if (mc
.moved_swap
) {
5700 /* uncharge swap account from the old cgroup */
5701 if (!mem_cgroup_is_root(mc
.from
))
5702 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5704 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5707 * we charged both to->memory and to->memsw, so we
5708 * should uncharge to->memory.
5710 if (!mem_cgroup_is_root(mc
.to
))
5711 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5713 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5714 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5718 memcg_oom_recover(from
);
5719 memcg_oom_recover(to
);
5720 wake_up_all(&mc
.waitq
);
5723 static void mem_cgroup_clear_mc(void)
5725 struct mm_struct
*mm
= mc
.mm
;
5728 * we must clear moving_task before waking up waiters at the end of
5731 mc
.moving_task
= NULL
;
5732 __mem_cgroup_clear_mc();
5733 spin_lock(&mc
.lock
);
5737 spin_unlock(&mc
.lock
);
5742 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5744 struct cgroup_subsys_state
*css
;
5745 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5746 struct mem_cgroup
*from
;
5747 struct task_struct
*leader
, *p
;
5748 struct mm_struct
*mm
;
5749 unsigned long move_flags
;
5752 /* charge immigration isn't supported on the default hierarchy */
5753 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5757 * Multi-process migrations only happen on the default hierarchy
5758 * where charge immigration is not used. Perform charge
5759 * immigration if @tset contains a leader and whine if there are
5763 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5766 memcg
= mem_cgroup_from_css(css
);
5772 * We are now commited to this value whatever it is. Changes in this
5773 * tunable will only affect upcoming migrations, not the current one.
5774 * So we need to save it, and keep it going.
5776 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5780 from
= mem_cgroup_from_task(p
);
5782 VM_BUG_ON(from
== memcg
);
5784 mm
= get_task_mm(p
);
5787 /* We move charges only when we move a owner of the mm */
5788 if (mm
->owner
== p
) {
5791 VM_BUG_ON(mc
.precharge
);
5792 VM_BUG_ON(mc
.moved_charge
);
5793 VM_BUG_ON(mc
.moved_swap
);
5795 spin_lock(&mc
.lock
);
5799 mc
.flags
= move_flags
;
5800 spin_unlock(&mc
.lock
);
5801 /* We set mc.moving_task later */
5803 ret
= mem_cgroup_precharge_mc(mm
);
5805 mem_cgroup_clear_mc();
5812 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5815 mem_cgroup_clear_mc();
5818 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5819 unsigned long addr
, unsigned long end
,
5820 struct mm_walk
*walk
)
5823 struct vm_area_struct
*vma
= walk
->vma
;
5826 enum mc_target_type target_type
;
5827 union mc_target target
;
5830 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5832 if (mc
.precharge
< HPAGE_PMD_NR
) {
5836 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5837 if (target_type
== MC_TARGET_PAGE
) {
5839 if (!isolate_lru_page(page
)) {
5840 if (!mem_cgroup_move_account(page
, true,
5842 mc
.precharge
-= HPAGE_PMD_NR
;
5843 mc
.moved_charge
+= HPAGE_PMD_NR
;
5845 putback_lru_page(page
);
5848 } else if (target_type
== MC_TARGET_DEVICE
) {
5850 if (!mem_cgroup_move_account(page
, true,
5852 mc
.precharge
-= HPAGE_PMD_NR
;
5853 mc
.moved_charge
+= HPAGE_PMD_NR
;
5861 if (pmd_trans_unstable(pmd
))
5864 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5865 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5866 pte_t ptent
= *(pte
++);
5867 bool device
= false;
5873 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5874 case MC_TARGET_DEVICE
:
5877 case MC_TARGET_PAGE
:
5880 * We can have a part of the split pmd here. Moving it
5881 * can be done but it would be too convoluted so simply
5882 * ignore such a partial THP and keep it in original
5883 * memcg. There should be somebody mapping the head.
5885 if (PageTransCompound(page
))
5887 if (!device
&& isolate_lru_page(page
))
5889 if (!mem_cgroup_move_account(page
, false,
5892 /* we uncharge from mc.from later. */
5896 putback_lru_page(page
);
5897 put
: /* get_mctgt_type() gets the page */
5900 case MC_TARGET_SWAP
:
5902 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5904 /* we fixup refcnts and charges later. */
5912 pte_unmap_unlock(pte
- 1, ptl
);
5917 * We have consumed all precharges we got in can_attach().
5918 * We try charge one by one, but don't do any additional
5919 * charges to mc.to if we have failed in charge once in attach()
5922 ret
= mem_cgroup_do_precharge(1);
5930 static const struct mm_walk_ops charge_walk_ops
= {
5931 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5934 static void mem_cgroup_move_charge(void)
5936 lru_add_drain_all();
5938 * Signal lock_page_memcg() to take the memcg's move_lock
5939 * while we're moving its pages to another memcg. Then wait
5940 * for already started RCU-only updates to finish.
5942 atomic_inc(&mc
.from
->moving_account
);
5945 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5947 * Someone who are holding the mmap_sem might be waiting in
5948 * waitq. So we cancel all extra charges, wake up all waiters,
5949 * and retry. Because we cancel precharges, we might not be able
5950 * to move enough charges, but moving charge is a best-effort
5951 * feature anyway, so it wouldn't be a big problem.
5953 __mem_cgroup_clear_mc();
5958 * When we have consumed all precharges and failed in doing
5959 * additional charge, the page walk just aborts.
5961 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
5964 up_read(&mc
.mm
->mmap_sem
);
5965 atomic_dec(&mc
.from
->moving_account
);
5968 static void mem_cgroup_move_task(void)
5971 mem_cgroup_move_charge();
5972 mem_cgroup_clear_mc();
5975 #else /* !CONFIG_MMU */
5976 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5980 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5983 static void mem_cgroup_move_task(void)
5989 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5990 * to verify whether we're attached to the default hierarchy on each mount
5993 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5996 * use_hierarchy is forced on the default hierarchy. cgroup core
5997 * guarantees that @root doesn't have any children, so turning it
5998 * on for the root memcg is enough.
6000 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6001 root_mem_cgroup
->use_hierarchy
= true;
6003 root_mem_cgroup
->use_hierarchy
= false;
6006 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6008 if (value
== PAGE_COUNTER_MAX
)
6009 seq_puts(m
, "max\n");
6011 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6016 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6019 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6021 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6024 static int memory_min_show(struct seq_file
*m
, void *v
)
6026 return seq_puts_memcg_tunable(m
,
6027 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6030 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6031 char *buf
, size_t nbytes
, loff_t off
)
6033 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6037 buf
= strstrip(buf
);
6038 err
= page_counter_memparse(buf
, "max", &min
);
6042 page_counter_set_min(&memcg
->memory
, min
);
6047 static int memory_low_show(struct seq_file
*m
, void *v
)
6049 return seq_puts_memcg_tunable(m
,
6050 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6053 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6054 char *buf
, size_t nbytes
, loff_t off
)
6056 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6060 buf
= strstrip(buf
);
6061 err
= page_counter_memparse(buf
, "max", &low
);
6065 page_counter_set_low(&memcg
->memory
, low
);
6070 static int memory_high_show(struct seq_file
*m
, void *v
)
6072 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
6075 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6076 char *buf
, size_t nbytes
, loff_t off
)
6078 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6079 unsigned long nr_pages
;
6083 buf
= strstrip(buf
);
6084 err
= page_counter_memparse(buf
, "max", &high
);
6090 nr_pages
= page_counter_read(&memcg
->memory
);
6091 if (nr_pages
> high
)
6092 try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6095 memcg_wb_domain_size_changed(memcg
);
6099 static int memory_max_show(struct seq_file
*m
, void *v
)
6101 return seq_puts_memcg_tunable(m
,
6102 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6105 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6106 char *buf
, size_t nbytes
, loff_t off
)
6108 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6109 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
6110 bool drained
= false;
6114 buf
= strstrip(buf
);
6115 err
= page_counter_memparse(buf
, "max", &max
);
6119 xchg(&memcg
->memory
.max
, max
);
6122 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6124 if (nr_pages
<= max
)
6127 if (signal_pending(current
)) {
6133 drain_all_stock(memcg
);
6139 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6145 memcg_memory_event(memcg
, MEMCG_OOM
);
6146 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6150 memcg_wb_domain_size_changed(memcg
);
6154 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6156 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6157 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6158 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6159 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6160 seq_printf(m
, "oom_kill %lu\n",
6161 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6164 static int memory_events_show(struct seq_file
*m
, void *v
)
6166 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6168 __memory_events_show(m
, memcg
->memory_events
);
6172 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6174 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6176 __memory_events_show(m
, memcg
->memory_events_local
);
6180 static int memory_stat_show(struct seq_file
*m
, void *v
)
6182 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6185 buf
= memory_stat_format(memcg
);
6193 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6195 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6197 seq_printf(m
, "%d\n", memcg
->oom_group
);
6202 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6203 char *buf
, size_t nbytes
, loff_t off
)
6205 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6208 buf
= strstrip(buf
);
6212 ret
= kstrtoint(buf
, 0, &oom_group
);
6216 if (oom_group
!= 0 && oom_group
!= 1)
6219 memcg
->oom_group
= oom_group
;
6224 static struct cftype memory_files
[] = {
6227 .flags
= CFTYPE_NOT_ON_ROOT
,
6228 .read_u64
= memory_current_read
,
6232 .flags
= CFTYPE_NOT_ON_ROOT
,
6233 .seq_show
= memory_min_show
,
6234 .write
= memory_min_write
,
6238 .flags
= CFTYPE_NOT_ON_ROOT
,
6239 .seq_show
= memory_low_show
,
6240 .write
= memory_low_write
,
6244 .flags
= CFTYPE_NOT_ON_ROOT
,
6245 .seq_show
= memory_high_show
,
6246 .write
= memory_high_write
,
6250 .flags
= CFTYPE_NOT_ON_ROOT
,
6251 .seq_show
= memory_max_show
,
6252 .write
= memory_max_write
,
6256 .flags
= CFTYPE_NOT_ON_ROOT
,
6257 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6258 .seq_show
= memory_events_show
,
6261 .name
= "events.local",
6262 .flags
= CFTYPE_NOT_ON_ROOT
,
6263 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6264 .seq_show
= memory_events_local_show
,
6268 .flags
= CFTYPE_NOT_ON_ROOT
,
6269 .seq_show
= memory_stat_show
,
6272 .name
= "oom.group",
6273 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6274 .seq_show
= memory_oom_group_show
,
6275 .write
= memory_oom_group_write
,
6280 struct cgroup_subsys memory_cgrp_subsys
= {
6281 .css_alloc
= mem_cgroup_css_alloc
,
6282 .css_online
= mem_cgroup_css_online
,
6283 .css_offline
= mem_cgroup_css_offline
,
6284 .css_released
= mem_cgroup_css_released
,
6285 .css_free
= mem_cgroup_css_free
,
6286 .css_reset
= mem_cgroup_css_reset
,
6287 .can_attach
= mem_cgroup_can_attach
,
6288 .cancel_attach
= mem_cgroup_cancel_attach
,
6289 .post_attach
= mem_cgroup_move_task
,
6290 .bind
= mem_cgroup_bind
,
6291 .dfl_cftypes
= memory_files
,
6292 .legacy_cftypes
= mem_cgroup_legacy_files
,
6297 * mem_cgroup_protected - check if memory consumption is in the normal range
6298 * @root: the top ancestor of the sub-tree being checked
6299 * @memcg: the memory cgroup to check
6301 * WARNING: This function is not stateless! It can only be used as part
6302 * of a top-down tree iteration, not for isolated queries.
6304 * Returns one of the following:
6305 * MEMCG_PROT_NONE: cgroup memory is not protected
6306 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6307 * an unprotected supply of reclaimable memory from other cgroups.
6308 * MEMCG_PROT_MIN: cgroup memory is protected
6310 * @root is exclusive; it is never protected when looked at directly
6312 * To provide a proper hierarchical behavior, effective memory.min/low values
6313 * are used. Below is the description of how effective memory.low is calculated.
6314 * Effective memory.min values is calculated in the same way.
6316 * Effective memory.low is always equal or less than the original memory.low.
6317 * If there is no memory.low overcommittment (which is always true for
6318 * top-level memory cgroups), these two values are equal.
6319 * Otherwise, it's a part of parent's effective memory.low,
6320 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6321 * memory.low usages, where memory.low usage is the size of actually
6325 * elow = min( memory.low, parent->elow * ------------------ ),
6326 * siblings_low_usage
6328 * | memory.current, if memory.current < memory.low
6333 * Such definition of the effective memory.low provides the expected
6334 * hierarchical behavior: parent's memory.low value is limiting
6335 * children, unprotected memory is reclaimed first and cgroups,
6336 * which are not using their guarantee do not affect actual memory
6339 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6341 * A A/memory.low = 2G, A/memory.current = 6G
6343 * BC DE B/memory.low = 3G B/memory.current = 2G
6344 * C/memory.low = 1G C/memory.current = 2G
6345 * D/memory.low = 0 D/memory.current = 2G
6346 * E/memory.low = 10G E/memory.current = 0
6348 * and the memory pressure is applied, the following memory distribution
6349 * is expected (approximately):
6351 * A/memory.current = 2G
6353 * B/memory.current = 1.3G
6354 * C/memory.current = 0.6G
6355 * D/memory.current = 0
6356 * E/memory.current = 0
6358 * These calculations require constant tracking of the actual low usages
6359 * (see propagate_protected_usage()), as well as recursive calculation of
6360 * effective memory.low values. But as we do call mem_cgroup_protected()
6361 * path for each memory cgroup top-down from the reclaim,
6362 * it's possible to optimize this part, and save calculated elow
6363 * for next usage. This part is intentionally racy, but it's ok,
6364 * as memory.low is a best-effort mechanism.
6366 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
6367 struct mem_cgroup
*memcg
)
6369 struct mem_cgroup
*parent
;
6370 unsigned long emin
, parent_emin
;
6371 unsigned long elow
, parent_elow
;
6372 unsigned long usage
;
6374 if (mem_cgroup_disabled())
6375 return MEMCG_PROT_NONE
;
6378 root
= root_mem_cgroup
;
6380 return MEMCG_PROT_NONE
;
6382 usage
= page_counter_read(&memcg
->memory
);
6384 return MEMCG_PROT_NONE
;
6386 emin
= memcg
->memory
.min
;
6387 elow
= memcg
->memory
.low
;
6389 parent
= parent_mem_cgroup(memcg
);
6390 /* No parent means a non-hierarchical mode on v1 memcg */
6392 return MEMCG_PROT_NONE
;
6397 parent_emin
= READ_ONCE(parent
->memory
.emin
);
6398 emin
= min(emin
, parent_emin
);
6399 if (emin
&& parent_emin
) {
6400 unsigned long min_usage
, siblings_min_usage
;
6402 min_usage
= min(usage
, memcg
->memory
.min
);
6403 siblings_min_usage
= atomic_long_read(
6404 &parent
->memory
.children_min_usage
);
6406 if (min_usage
&& siblings_min_usage
)
6407 emin
= min(emin
, parent_emin
* min_usage
/
6408 siblings_min_usage
);
6411 parent_elow
= READ_ONCE(parent
->memory
.elow
);
6412 elow
= min(elow
, parent_elow
);
6413 if (elow
&& parent_elow
) {
6414 unsigned long low_usage
, siblings_low_usage
;
6416 low_usage
= min(usage
, memcg
->memory
.low
);
6417 siblings_low_usage
= atomic_long_read(
6418 &parent
->memory
.children_low_usage
);
6420 if (low_usage
&& siblings_low_usage
)
6421 elow
= min(elow
, parent_elow
* low_usage
/
6422 siblings_low_usage
);
6426 memcg
->memory
.emin
= emin
;
6427 memcg
->memory
.elow
= elow
;
6430 return MEMCG_PROT_MIN
;
6431 else if (usage
<= elow
)
6432 return MEMCG_PROT_LOW
;
6434 return MEMCG_PROT_NONE
;
6438 * mem_cgroup_try_charge - try charging a page
6439 * @page: page to charge
6440 * @mm: mm context of the victim
6441 * @gfp_mask: reclaim mode
6442 * @memcgp: charged memcg return
6443 * @compound: charge the page as compound or small page
6445 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6446 * pages according to @gfp_mask if necessary.
6448 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6449 * Otherwise, an error code is returned.
6451 * After page->mapping has been set up, the caller must finalize the
6452 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6453 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6455 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6456 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6459 struct mem_cgroup
*memcg
= NULL
;
6460 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6463 if (mem_cgroup_disabled())
6466 if (PageSwapCache(page
)) {
6468 * Every swap fault against a single page tries to charge the
6469 * page, bail as early as possible. shmem_unuse() encounters
6470 * already charged pages, too. The USED bit is protected by
6471 * the page lock, which serializes swap cache removal, which
6472 * in turn serializes uncharging.
6474 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6475 if (compound_head(page
)->mem_cgroup
)
6478 if (do_swap_account
) {
6479 swp_entry_t ent
= { .val
= page_private(page
), };
6480 unsigned short id
= lookup_swap_cgroup_id(ent
);
6483 memcg
= mem_cgroup_from_id(id
);
6484 if (memcg
&& !css_tryget_online(&memcg
->css
))
6491 memcg
= get_mem_cgroup_from_mm(mm
);
6493 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6495 css_put(&memcg
->css
);
6501 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6502 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6505 struct mem_cgroup
*memcg
;
6508 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6510 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6515 * mem_cgroup_commit_charge - commit a page charge
6516 * @page: page to charge
6517 * @memcg: memcg to charge the page to
6518 * @lrucare: page might be on LRU already
6519 * @compound: charge the page as compound or small page
6521 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6522 * after page->mapping has been set up. This must happen atomically
6523 * as part of the page instantiation, i.e. under the page table lock
6524 * for anonymous pages, under the page lock for page and swap cache.
6526 * In addition, the page must not be on the LRU during the commit, to
6527 * prevent racing with task migration. If it might be, use @lrucare.
6529 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6531 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6532 bool lrucare
, bool compound
)
6534 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6536 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6537 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6539 if (mem_cgroup_disabled())
6542 * Swap faults will attempt to charge the same page multiple
6543 * times. But reuse_swap_page() might have removed the page
6544 * from swapcache already, so we can't check PageSwapCache().
6549 commit_charge(page
, memcg
, lrucare
);
6551 local_irq_disable();
6552 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6553 memcg_check_events(memcg
, page
);
6556 if (do_memsw_account() && PageSwapCache(page
)) {
6557 swp_entry_t entry
= { .val
= page_private(page
) };
6559 * The swap entry might not get freed for a long time,
6560 * let's not wait for it. The page already received a
6561 * memory+swap charge, drop the swap entry duplicate.
6563 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6568 * mem_cgroup_cancel_charge - cancel a page charge
6569 * @page: page to charge
6570 * @memcg: memcg to charge the page to
6571 * @compound: charge the page as compound or small page
6573 * Cancel a charge transaction started by mem_cgroup_try_charge().
6575 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6578 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6580 if (mem_cgroup_disabled())
6583 * Swap faults will attempt to charge the same page multiple
6584 * times. But reuse_swap_page() might have removed the page
6585 * from swapcache already, so we can't check PageSwapCache().
6590 cancel_charge(memcg
, nr_pages
);
6593 struct uncharge_gather
{
6594 struct mem_cgroup
*memcg
;
6595 unsigned long pgpgout
;
6596 unsigned long nr_anon
;
6597 unsigned long nr_file
;
6598 unsigned long nr_kmem
;
6599 unsigned long nr_huge
;
6600 unsigned long nr_shmem
;
6601 struct page
*dummy_page
;
6604 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6606 memset(ug
, 0, sizeof(*ug
));
6609 static void uncharge_batch(const struct uncharge_gather
*ug
)
6611 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6612 unsigned long flags
;
6614 if (!mem_cgroup_is_root(ug
->memcg
)) {
6615 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6616 if (do_memsw_account())
6617 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6618 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6619 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6620 memcg_oom_recover(ug
->memcg
);
6623 local_irq_save(flags
);
6624 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6625 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6626 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6627 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6628 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6629 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6630 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6631 local_irq_restore(flags
);
6633 if (!mem_cgroup_is_root(ug
->memcg
))
6634 css_put_many(&ug
->memcg
->css
, nr_pages
);
6637 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6639 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6640 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6641 !PageHWPoison(page
) , page
);
6643 if (!page
->mem_cgroup
)
6647 * Nobody should be changing or seriously looking at
6648 * page->mem_cgroup at this point, we have fully
6649 * exclusive access to the page.
6652 if (ug
->memcg
!= page
->mem_cgroup
) {
6655 uncharge_gather_clear(ug
);
6657 ug
->memcg
= page
->mem_cgroup
;
6660 if (!PageKmemcg(page
)) {
6661 unsigned int nr_pages
= 1;
6663 if (PageTransHuge(page
)) {
6664 nr_pages
= compound_nr(page
);
6665 ug
->nr_huge
+= nr_pages
;
6668 ug
->nr_anon
+= nr_pages
;
6670 ug
->nr_file
+= nr_pages
;
6671 if (PageSwapBacked(page
))
6672 ug
->nr_shmem
+= nr_pages
;
6676 ug
->nr_kmem
+= compound_nr(page
);
6677 __ClearPageKmemcg(page
);
6680 ug
->dummy_page
= page
;
6681 page
->mem_cgroup
= NULL
;
6684 static void uncharge_list(struct list_head
*page_list
)
6686 struct uncharge_gather ug
;
6687 struct list_head
*next
;
6689 uncharge_gather_clear(&ug
);
6692 * Note that the list can be a single page->lru; hence the
6693 * do-while loop instead of a simple list_for_each_entry().
6695 next
= page_list
->next
;
6699 page
= list_entry(next
, struct page
, lru
);
6700 next
= page
->lru
.next
;
6702 uncharge_page(page
, &ug
);
6703 } while (next
!= page_list
);
6706 uncharge_batch(&ug
);
6710 * mem_cgroup_uncharge - uncharge a page
6711 * @page: page to uncharge
6713 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6714 * mem_cgroup_commit_charge().
6716 void mem_cgroup_uncharge(struct page
*page
)
6718 struct uncharge_gather ug
;
6720 if (mem_cgroup_disabled())
6723 /* Don't touch page->lru of any random page, pre-check: */
6724 if (!page
->mem_cgroup
)
6727 uncharge_gather_clear(&ug
);
6728 uncharge_page(page
, &ug
);
6729 uncharge_batch(&ug
);
6733 * mem_cgroup_uncharge_list - uncharge a list of page
6734 * @page_list: list of pages to uncharge
6736 * Uncharge a list of pages previously charged with
6737 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6739 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6741 if (mem_cgroup_disabled())
6744 if (!list_empty(page_list
))
6745 uncharge_list(page_list
);
6749 * mem_cgroup_migrate - charge a page's replacement
6750 * @oldpage: currently circulating page
6751 * @newpage: replacement page
6753 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6754 * be uncharged upon free.
6756 * Both pages must be locked, @newpage->mapping must be set up.
6758 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6760 struct mem_cgroup
*memcg
;
6761 unsigned int nr_pages
;
6763 unsigned long flags
;
6765 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6766 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6767 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6768 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6771 if (mem_cgroup_disabled())
6774 /* Page cache replacement: new page already charged? */
6775 if (newpage
->mem_cgroup
)
6778 /* Swapcache readahead pages can get replaced before being charged */
6779 memcg
= oldpage
->mem_cgroup
;
6783 /* Force-charge the new page. The old one will be freed soon */
6784 compound
= PageTransHuge(newpage
);
6785 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
6787 page_counter_charge(&memcg
->memory
, nr_pages
);
6788 if (do_memsw_account())
6789 page_counter_charge(&memcg
->memsw
, nr_pages
);
6790 css_get_many(&memcg
->css
, nr_pages
);
6792 commit_charge(newpage
, memcg
, false);
6794 local_irq_save(flags
);
6795 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
6796 memcg_check_events(memcg
, newpage
);
6797 local_irq_restore(flags
);
6800 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6801 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6803 void mem_cgroup_sk_alloc(struct sock
*sk
)
6805 struct mem_cgroup
*memcg
;
6807 if (!mem_cgroup_sockets_enabled
)
6811 * Socket cloning can throw us here with sk_memcg already
6812 * filled. It won't however, necessarily happen from
6813 * process context. So the test for root memcg given
6814 * the current task's memcg won't help us in this case.
6816 * Respecting the original socket's memcg is a better
6817 * decision in this case.
6820 css_get(&sk
->sk_memcg
->css
);
6825 memcg
= mem_cgroup_from_task(current
);
6826 if (memcg
== root_mem_cgroup
)
6828 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6830 if (css_tryget_online(&memcg
->css
))
6831 sk
->sk_memcg
= memcg
;
6836 void mem_cgroup_sk_free(struct sock
*sk
)
6839 css_put(&sk
->sk_memcg
->css
);
6843 * mem_cgroup_charge_skmem - charge socket memory
6844 * @memcg: memcg to charge
6845 * @nr_pages: number of pages to charge
6847 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6848 * @memcg's configured limit, %false if the charge had to be forced.
6850 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6852 gfp_t gfp_mask
= GFP_KERNEL
;
6854 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6855 struct page_counter
*fail
;
6857 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6858 memcg
->tcpmem_pressure
= 0;
6861 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6862 memcg
->tcpmem_pressure
= 1;
6866 /* Don't block in the packet receive path */
6868 gfp_mask
= GFP_NOWAIT
;
6870 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6872 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6875 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6880 * mem_cgroup_uncharge_skmem - uncharge socket memory
6881 * @memcg: memcg to uncharge
6882 * @nr_pages: number of pages to uncharge
6884 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6886 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6887 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6891 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6893 refill_stock(memcg
, nr_pages
);
6896 static int __init
cgroup_memory(char *s
)
6900 while ((token
= strsep(&s
, ",")) != NULL
) {
6903 if (!strcmp(token
, "nosocket"))
6904 cgroup_memory_nosocket
= true;
6905 if (!strcmp(token
, "nokmem"))
6906 cgroup_memory_nokmem
= true;
6910 __setup("cgroup.memory=", cgroup_memory
);
6913 * subsys_initcall() for memory controller.
6915 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6916 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6917 * basically everything that doesn't depend on a specific mem_cgroup structure
6918 * should be initialized from here.
6920 static int __init
mem_cgroup_init(void)
6924 #ifdef CONFIG_MEMCG_KMEM
6926 * Kmem cache creation is mostly done with the slab_mutex held,
6927 * so use a workqueue with limited concurrency to avoid stalling
6928 * all worker threads in case lots of cgroups are created and
6929 * destroyed simultaneously.
6931 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6932 BUG_ON(!memcg_kmem_cache_wq
);
6935 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6936 memcg_hotplug_cpu_dead
);
6938 for_each_possible_cpu(cpu
)
6939 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6942 for_each_node(node
) {
6943 struct mem_cgroup_tree_per_node
*rtpn
;
6945 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6946 node_online(node
) ? node
: NUMA_NO_NODE
);
6948 rtpn
->rb_root
= RB_ROOT
;
6949 rtpn
->rb_rightmost
= NULL
;
6950 spin_lock_init(&rtpn
->lock
);
6951 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6956 subsys_initcall(mem_cgroup_init
);
6958 #ifdef CONFIG_MEMCG_SWAP
6959 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6961 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6963 * The root cgroup cannot be destroyed, so it's refcount must
6966 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6970 memcg
= parent_mem_cgroup(memcg
);
6972 memcg
= root_mem_cgroup
;
6978 * mem_cgroup_swapout - transfer a memsw charge to swap
6979 * @page: page whose memsw charge to transfer
6980 * @entry: swap entry to move the charge to
6982 * Transfer the memsw charge of @page to @entry.
6984 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6986 struct mem_cgroup
*memcg
, *swap_memcg
;
6987 unsigned int nr_entries
;
6988 unsigned short oldid
;
6990 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6991 VM_BUG_ON_PAGE(page_count(page
), page
);
6993 if (!do_memsw_account())
6996 memcg
= page
->mem_cgroup
;
6998 /* Readahead page, never charged */
7003 * In case the memcg owning these pages has been offlined and doesn't
7004 * have an ID allocated to it anymore, charge the closest online
7005 * ancestor for the swap instead and transfer the memory+swap charge.
7007 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7008 nr_entries
= hpage_nr_pages(page
);
7009 /* Get references for the tail pages, too */
7011 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7012 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7014 VM_BUG_ON_PAGE(oldid
, page
);
7015 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7017 page
->mem_cgroup
= NULL
;
7019 if (!mem_cgroup_is_root(memcg
))
7020 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7022 if (memcg
!= swap_memcg
) {
7023 if (!mem_cgroup_is_root(swap_memcg
))
7024 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7025 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7029 * Interrupts should be disabled here because the caller holds the
7030 * i_pages lock which is taken with interrupts-off. It is
7031 * important here to have the interrupts disabled because it is the
7032 * only synchronisation we have for updating the per-CPU variables.
7034 VM_BUG_ON(!irqs_disabled());
7035 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
7037 memcg_check_events(memcg
, page
);
7039 if (!mem_cgroup_is_root(memcg
))
7040 css_put_many(&memcg
->css
, nr_entries
);
7044 * mem_cgroup_try_charge_swap - try charging swap space for a page
7045 * @page: page being added to swap
7046 * @entry: swap entry to charge
7048 * Try to charge @page's memcg for the swap space at @entry.
7050 * Returns 0 on success, -ENOMEM on failure.
7052 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7054 unsigned int nr_pages
= hpage_nr_pages(page
);
7055 struct page_counter
*counter
;
7056 struct mem_cgroup
*memcg
;
7057 unsigned short oldid
;
7059 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
7062 memcg
= page
->mem_cgroup
;
7064 /* Readahead page, never charged */
7069 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7073 memcg
= mem_cgroup_id_get_online(memcg
);
7075 if (!mem_cgroup_is_root(memcg
) &&
7076 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7077 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7078 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7079 mem_cgroup_id_put(memcg
);
7083 /* Get references for the tail pages, too */
7085 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7086 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7087 VM_BUG_ON_PAGE(oldid
, page
);
7088 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7094 * mem_cgroup_uncharge_swap - uncharge swap space
7095 * @entry: swap entry to uncharge
7096 * @nr_pages: the amount of swap space to uncharge
7098 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7100 struct mem_cgroup
*memcg
;
7103 if (!do_swap_account
)
7106 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7108 memcg
= mem_cgroup_from_id(id
);
7110 if (!mem_cgroup_is_root(memcg
)) {
7111 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7112 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7114 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7116 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7117 mem_cgroup_id_put_many(memcg
, nr_pages
);
7122 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7124 long nr_swap_pages
= get_nr_swap_pages();
7126 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7127 return nr_swap_pages
;
7128 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7129 nr_swap_pages
= min_t(long, nr_swap_pages
,
7130 READ_ONCE(memcg
->swap
.max
) -
7131 page_counter_read(&memcg
->swap
));
7132 return nr_swap_pages
;
7135 bool mem_cgroup_swap_full(struct page
*page
)
7137 struct mem_cgroup
*memcg
;
7139 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7143 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7146 memcg
= page
->mem_cgroup
;
7150 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7151 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
7157 /* for remember boot option*/
7158 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7159 static int really_do_swap_account __initdata
= 1;
7161 static int really_do_swap_account __initdata
;
7164 static int __init
enable_swap_account(char *s
)
7166 if (!strcmp(s
, "1"))
7167 really_do_swap_account
= 1;
7168 else if (!strcmp(s
, "0"))
7169 really_do_swap_account
= 0;
7172 __setup("swapaccount=", enable_swap_account
);
7174 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7177 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7179 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7182 static int swap_max_show(struct seq_file
*m
, void *v
)
7184 return seq_puts_memcg_tunable(m
,
7185 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7188 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7189 char *buf
, size_t nbytes
, loff_t off
)
7191 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7195 buf
= strstrip(buf
);
7196 err
= page_counter_memparse(buf
, "max", &max
);
7200 xchg(&memcg
->swap
.max
, max
);
7205 static int swap_events_show(struct seq_file
*m
, void *v
)
7207 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7209 seq_printf(m
, "max %lu\n",
7210 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7211 seq_printf(m
, "fail %lu\n",
7212 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7217 static struct cftype swap_files
[] = {
7219 .name
= "swap.current",
7220 .flags
= CFTYPE_NOT_ON_ROOT
,
7221 .read_u64
= swap_current_read
,
7225 .flags
= CFTYPE_NOT_ON_ROOT
,
7226 .seq_show
= swap_max_show
,
7227 .write
= swap_max_write
,
7230 .name
= "swap.events",
7231 .flags
= CFTYPE_NOT_ON_ROOT
,
7232 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7233 .seq_show
= swap_events_show
,
7238 static struct cftype memsw_cgroup_files
[] = {
7240 .name
= "memsw.usage_in_bytes",
7241 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7242 .read_u64
= mem_cgroup_read_u64
,
7245 .name
= "memsw.max_usage_in_bytes",
7246 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7247 .write
= mem_cgroup_reset
,
7248 .read_u64
= mem_cgroup_read_u64
,
7251 .name
= "memsw.limit_in_bytes",
7252 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7253 .write
= mem_cgroup_write
,
7254 .read_u64
= mem_cgroup_read_u64
,
7257 .name
= "memsw.failcnt",
7258 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7259 .write
= mem_cgroup_reset
,
7260 .read_u64
= mem_cgroup_read_u64
,
7262 { }, /* terminate */
7265 static int __init
mem_cgroup_swap_init(void)
7267 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7268 do_swap_account
= 1;
7269 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
7271 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
7272 memsw_cgroup_files
));
7276 subsys_initcall(mem_cgroup_swap_init
);
7278 #endif /* CONFIG_MEMCG_SWAP */