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 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node
{
110 struct rb_root rb_root
;
111 struct rb_node
*rb_rightmost
;
115 struct mem_cgroup_tree
{
116 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
122 struct mem_cgroup_eventfd_list
{
123 struct list_head list
;
124 struct eventfd_ctx
*eventfd
;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event
{
132 * memcg which the event belongs to.
134 struct mem_cgroup
*memcg
;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx
*eventfd
;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list
;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event
)(struct mem_cgroup
*memcg
,
149 struct eventfd_ctx
*eventfd
, const char *args
);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event
)(struct mem_cgroup
*memcg
,
156 struct eventfd_ctx
*eventfd
);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t
*wqh
;
163 wait_queue_entry_t wait
;
164 struct work_struct remove
;
167 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
168 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct
{
180 spinlock_t lock
; /* for from, to */
181 struct mm_struct
*mm
;
182 struct mem_cgroup
*from
;
183 struct mem_cgroup
*to
;
185 unsigned long precharge
;
186 unsigned long moved_charge
;
187 unsigned long moved_swap
;
188 struct task_struct
*moving_task
; /* a task moving charges */
189 wait_queue_head_t waitq
; /* a waitq for other context */
191 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
192 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON
,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
243 (current
->flags
& PF_EXITING
);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
250 memcg
= root_mem_cgroup
;
251 return &memcg
->vmpressure
;
254 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
256 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
259 #ifdef CONFIG_MEMCG_KMEM
261 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
262 * The main reason for not using cgroup id for this:
263 * this works better in sparse environments, where we have a lot of memcgs,
264 * but only a few kmem-limited. Or also, if we have, for instance, 200
265 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
266 * 200 entry array for that.
268 * The current size of the caches array is stored in memcg_nr_cache_ids. It
269 * will double each time we have to increase it.
271 static DEFINE_IDA(memcg_cache_ida
);
272 int memcg_nr_cache_ids
;
274 /* Protects memcg_nr_cache_ids */
275 static DECLARE_RWSEM(memcg_cache_ids_sem
);
277 void memcg_get_cache_ids(void)
279 down_read(&memcg_cache_ids_sem
);
282 void memcg_put_cache_ids(void)
284 up_read(&memcg_cache_ids_sem
);
288 * MIN_SIZE is different than 1, because we would like to avoid going through
289 * the alloc/free process all the time. In a small machine, 4 kmem-limited
290 * cgroups is a reasonable guess. In the future, it could be a parameter or
291 * tunable, but that is strictly not necessary.
293 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
294 * this constant directly from cgroup, but it is understandable that this is
295 * better kept as an internal representation in cgroup.c. In any case, the
296 * cgrp_id space is not getting any smaller, and we don't have to necessarily
297 * increase ours as well if it increases.
299 #define MEMCG_CACHES_MIN_SIZE 4
300 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
303 * A lot of the calls to the cache allocation functions are expected to be
304 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
305 * conditional to this static branch, we'll have to allow modules that does
306 * kmem_cache_alloc and the such to see this symbol as well
308 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
309 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
311 struct workqueue_struct
*memcg_kmem_cache_wq
;
314 static int memcg_shrinker_map_size
;
315 static DEFINE_MUTEX(memcg_shrinker_map_mutex
);
317 static void memcg_free_shrinker_map_rcu(struct rcu_head
*head
)
319 kvfree(container_of(head
, struct memcg_shrinker_map
, rcu
));
322 static int memcg_expand_one_shrinker_map(struct mem_cgroup
*memcg
,
323 int size
, int old_size
)
325 struct memcg_shrinker_map
*new, *old
;
328 lockdep_assert_held(&memcg_shrinker_map_mutex
);
331 old
= rcu_dereference_protected(
332 mem_cgroup_nodeinfo(memcg
, nid
)->shrinker_map
, true);
333 /* Not yet online memcg */
337 new = kvmalloc_node(sizeof(*new) + size
, GFP_KERNEL
, nid
);
341 /* Set all old bits, clear all new bits */
342 memset(new->map
, (int)0xff, old_size
);
343 memset((void *)new->map
+ old_size
, 0, size
- old_size
);
345 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, new);
346 call_rcu(&old
->rcu
, memcg_free_shrinker_map_rcu
);
352 static void memcg_free_shrinker_maps(struct mem_cgroup
*memcg
)
354 struct mem_cgroup_per_node
*pn
;
355 struct memcg_shrinker_map
*map
;
358 if (mem_cgroup_is_root(memcg
))
362 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
363 map
= rcu_dereference_protected(pn
->shrinker_map
, true);
366 rcu_assign_pointer(pn
->shrinker_map
, NULL
);
370 static int memcg_alloc_shrinker_maps(struct mem_cgroup
*memcg
)
372 struct memcg_shrinker_map
*map
;
373 int nid
, size
, ret
= 0;
375 if (mem_cgroup_is_root(memcg
))
378 mutex_lock(&memcg_shrinker_map_mutex
);
379 size
= memcg_shrinker_map_size
;
381 map
= kvzalloc_node(sizeof(*map
) + size
, GFP_KERNEL
, nid
);
383 memcg_free_shrinker_maps(memcg
);
387 rcu_assign_pointer(memcg
->nodeinfo
[nid
]->shrinker_map
, map
);
389 mutex_unlock(&memcg_shrinker_map_mutex
);
394 int memcg_expand_shrinker_maps(int new_id
)
396 int size
, old_size
, ret
= 0;
397 struct mem_cgroup
*memcg
;
399 size
= DIV_ROUND_UP(new_id
+ 1, BITS_PER_LONG
) * sizeof(unsigned long);
400 old_size
= memcg_shrinker_map_size
;
401 if (size
<= old_size
)
404 mutex_lock(&memcg_shrinker_map_mutex
);
405 if (!root_mem_cgroup
)
408 for_each_mem_cgroup(memcg
) {
409 if (mem_cgroup_is_root(memcg
))
411 ret
= memcg_expand_one_shrinker_map(memcg
, size
, old_size
);
413 mem_cgroup_iter_break(NULL
, memcg
);
419 memcg_shrinker_map_size
= size
;
420 mutex_unlock(&memcg_shrinker_map_mutex
);
424 void memcg_set_shrinker_bit(struct mem_cgroup
*memcg
, int nid
, int shrinker_id
)
426 if (shrinker_id
>= 0 && memcg
&& !mem_cgroup_is_root(memcg
)) {
427 struct memcg_shrinker_map
*map
;
430 map
= rcu_dereference(memcg
->nodeinfo
[nid
]->shrinker_map
);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id
, map
->map
);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
451 struct mem_cgroup
*memcg
;
453 memcg
= page
->mem_cgroup
;
455 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
456 memcg
= root_mem_cgroup
;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t
page_cgroup_ino(struct page
*page
)
476 struct mem_cgroup
*memcg
;
477 unsigned long ino
= 0;
480 if (PageSlab(page
) && !PageTail(page
))
481 memcg
= memcg_from_slab_page(page
);
483 memcg
= READ_ONCE(page
->mem_cgroup
);
484 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
485 memcg
= parent_mem_cgroup(memcg
);
487 ino
= cgroup_ino(memcg
->css
.cgroup
);
492 static struct mem_cgroup_per_node
*
493 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
495 int nid
= page_to_nid(page
);
497 return memcg
->nodeinfo
[nid
];
500 static struct mem_cgroup_tree_per_node
*
501 soft_limit_tree_node(int nid
)
503 return soft_limit_tree
.rb_tree_per_node
[nid
];
506 static struct mem_cgroup_tree_per_node
*
507 soft_limit_tree_from_page(struct page
*page
)
509 int nid
= page_to_nid(page
);
511 return soft_limit_tree
.rb_tree_per_node
[nid
];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
515 struct mem_cgroup_tree_per_node
*mctz
,
516 unsigned long new_usage_in_excess
)
518 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
519 struct rb_node
*parent
= NULL
;
520 struct mem_cgroup_per_node
*mz_node
;
521 bool rightmost
= true;
526 mz
->usage_in_excess
= new_usage_in_excess
;
527 if (!mz
->usage_in_excess
)
531 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
533 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
547 mctz
->rb_rightmost
= &mz
->tree_node
;
549 rb_link_node(&mz
->tree_node
, parent
, p
);
550 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
555 struct mem_cgroup_tree_per_node
*mctz
)
560 if (&mz
->tree_node
== mctz
->rb_rightmost
)
561 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
563 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
568 struct mem_cgroup_tree_per_node
*mctz
)
572 spin_lock_irqsave(&mctz
->lock
, flags
);
573 __mem_cgroup_remove_exceeded(mz
, mctz
);
574 spin_unlock_irqrestore(&mctz
->lock
, flags
);
577 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
579 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
580 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
581 unsigned long excess
= 0;
583 if (nr_pages
> soft_limit
)
584 excess
= nr_pages
- soft_limit
;
589 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
591 unsigned long excess
;
592 struct mem_cgroup_per_node
*mz
;
593 struct mem_cgroup_tree_per_node
*mctz
;
595 mctz
= soft_limit_tree_from_page(page
);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
603 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
604 excess
= soft_limit_excess(memcg
);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess
|| mz
->on_tree
) {
612 spin_lock_irqsave(&mctz
->lock
, flags
);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz
, mctz
);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
621 spin_unlock_irqrestore(&mctz
->lock
, flags
);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
628 struct mem_cgroup_tree_per_node
*mctz
;
629 struct mem_cgroup_per_node
*mz
;
633 mz
= mem_cgroup_nodeinfo(memcg
, nid
);
634 mctz
= soft_limit_tree_node(nid
);
636 mem_cgroup_remove_exceeded(mz
, mctz
);
640 static struct mem_cgroup_per_node
*
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
643 struct mem_cgroup_per_node
*mz
;
647 if (!mctz
->rb_rightmost
)
648 goto done
; /* Nothing to reclaim from */
650 mz
= rb_entry(mctz
->rb_rightmost
,
651 struct mem_cgroup_per_node
, tree_node
);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz
, mctz
);
658 if (!soft_limit_excess(mz
->memcg
) ||
659 !css_tryget(&mz
->memcg
->css
))
665 static struct mem_cgroup_per_node
*
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
668 struct mem_cgroup_per_node
*mz
;
670 spin_lock_irq(&mctz
->lock
);
671 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
672 spin_unlock_irq(&mctz
->lock
);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
686 if (mem_cgroup_disabled())
689 x
= val
+ __this_cpu_read(memcg
->vmstats_percpu
->stat
[idx
]);
690 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
691 struct mem_cgroup
*mi
;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg
->vmstats_local
->stat
[idx
], x
);
698 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
699 atomic_long_add(x
, &mi
->vmstats
[idx
]);
702 __this_cpu_write(memcg
->vmstats_percpu
->stat
[idx
], x
);
705 static struct mem_cgroup_per_node
*
706 parent_nodeinfo(struct mem_cgroup_per_node
*pn
, int nid
)
708 struct mem_cgroup
*parent
;
710 parent
= parent_mem_cgroup(pn
->memcg
);
713 return mem_cgroup_nodeinfo(parent
, nid
);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
729 pg_data_t
*pgdat
= lruvec_pgdat(lruvec
);
730 struct mem_cgroup_per_node
*pn
;
731 struct mem_cgroup
*memcg
;
735 __mod_node_page_state(pgdat
, idx
, val
);
737 if (mem_cgroup_disabled())
740 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
744 __mod_memcg_state(memcg
, idx
, val
);
747 __this_cpu_add(pn
->lruvec_stat_local
->count
[idx
], val
);
749 x
= val
+ __this_cpu_read(pn
->lruvec_stat_cpu
->count
[idx
]);
750 if (unlikely(abs(x
) > MEMCG_CHARGE_BATCH
)) {
751 struct mem_cgroup_per_node
*pi
;
753 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, pgdat
->node_id
))
754 atomic_long_add(x
, &pi
->lruvec_stat
[idx
]);
757 __this_cpu_write(pn
->lruvec_stat_cpu
->count
[idx
], x
);
760 void __mod_lruvec_slab_state(void *p
, enum node_stat_item idx
, int val
)
762 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
763 struct mem_cgroup
*memcg
;
764 struct lruvec
*lruvec
;
767 memcg
= mem_cgroup_from_obj(p
);
769 /* Untracked pages have no memcg, no lruvec. Update only the node */
770 if (!memcg
|| memcg
== root_mem_cgroup
) {
771 __mod_node_page_state(pgdat
, idx
, val
);
773 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
774 __mod_lruvec_state(lruvec
, idx
, val
);
779 void mod_memcg_obj_state(void *p
, int idx
, int val
)
781 struct mem_cgroup
*memcg
;
784 memcg
= mem_cgroup_from_obj(p
);
786 mod_memcg_state(memcg
, idx
, val
);
791 * __count_memcg_events - account VM events in a cgroup
792 * @memcg: the memory cgroup
793 * @idx: the event item
794 * @count: the number of events that occured
796 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
801 if (mem_cgroup_disabled())
804 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
805 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
806 struct mem_cgroup
*mi
;
809 * Batch local counters to keep them in sync with
810 * the hierarchical ones.
812 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
813 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
814 atomic_long_add(x
, &mi
->vmevents
[idx
]);
817 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
820 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
822 return atomic_long_read(&memcg
->vmevents
[event
]);
825 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
830 for_each_possible_cpu(cpu
)
831 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
835 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
837 bool compound
, int nr_pages
)
840 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
841 * counted as CACHE even if it's on ANON LRU.
844 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
846 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
847 if (PageSwapBacked(page
))
848 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
852 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
853 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
856 /* pagein of a big page is an event. So, ignore page size */
858 __count_memcg_events(memcg
, PGPGIN
, 1);
860 __count_memcg_events(memcg
, PGPGOUT
, 1);
861 nr_pages
= -nr_pages
; /* for event */
864 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
867 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
868 enum mem_cgroup_events_target target
)
870 unsigned long val
, next
;
872 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
873 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
874 /* from time_after() in jiffies.h */
875 if ((long)(next
- val
) < 0) {
877 case MEM_CGROUP_TARGET_THRESH
:
878 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
880 case MEM_CGROUP_TARGET_SOFTLIMIT
:
881 next
= val
+ SOFTLIMIT_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
))) {
903 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
904 MEM_CGROUP_TARGET_SOFTLIMIT
);
905 mem_cgroup_threshold(memcg
);
906 if (unlikely(do_softlimit
))
907 mem_cgroup_update_tree(memcg
, page
);
911 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
914 * mm_update_next_owner() may clear mm->owner to NULL
915 * if it races with swapoff, page migration, etc.
916 * So this can be called with p == NULL.
921 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
923 EXPORT_SYMBOL(mem_cgroup_from_task
);
926 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
927 * @mm: mm from which memcg should be extracted. It can be NULL.
929 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
930 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
933 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
935 struct mem_cgroup
*memcg
;
937 if (mem_cgroup_disabled())
943 * Page cache insertions can happen withou an
944 * actual mm context, e.g. during disk probing
945 * on boot, loopback IO, acct() writes etc.
948 memcg
= root_mem_cgroup
;
950 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
951 if (unlikely(!memcg
))
952 memcg
= root_mem_cgroup
;
954 } while (!css_tryget(&memcg
->css
));
958 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
961 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
962 * @page: page from which memcg should be extracted.
964 * Obtain a reference on page->memcg and returns it if successful. Otherwise
965 * root_mem_cgroup is returned.
967 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
969 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
971 if (mem_cgroup_disabled())
975 /* Page should not get uncharged and freed memcg under us. */
976 if (!memcg
|| WARN_ON_ONCE(!css_tryget(&memcg
->css
)))
977 memcg
= root_mem_cgroup
;
981 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
984 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
986 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
988 if (unlikely(current
->active_memcg
)) {
989 struct mem_cgroup
*memcg
;
992 /* current->active_memcg must hold a ref. */
993 if (WARN_ON_ONCE(!css_tryget(¤t
->active_memcg
->css
)))
994 memcg
= root_mem_cgroup
;
996 memcg
= current
->active_memcg
;
1000 return get_mem_cgroup_from_mm(current
->mm
);
1004 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1005 * @root: hierarchy root
1006 * @prev: previously returned memcg, NULL on first invocation
1007 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1009 * Returns references to children of the hierarchy below @root, or
1010 * @root itself, or %NULL after a full round-trip.
1012 * Caller must pass the return value in @prev on subsequent
1013 * invocations for reference counting, or use mem_cgroup_iter_break()
1014 * to cancel a hierarchy walk before the round-trip is complete.
1016 * Reclaimers can specify a node and a priority level in @reclaim to
1017 * divide up the memcgs in the hierarchy among all concurrent
1018 * reclaimers operating on the same node and priority.
1020 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1021 struct mem_cgroup
*prev
,
1022 struct mem_cgroup_reclaim_cookie
*reclaim
)
1024 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1025 struct cgroup_subsys_state
*css
= NULL
;
1026 struct mem_cgroup
*memcg
= NULL
;
1027 struct mem_cgroup
*pos
= NULL
;
1029 if (mem_cgroup_disabled())
1033 root
= root_mem_cgroup
;
1035 if (prev
&& !reclaim
)
1038 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1047 struct mem_cgroup_per_node
*mz
;
1049 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1052 if (prev
&& reclaim
->generation
!= iter
->generation
)
1056 pos
= READ_ONCE(iter
->position
);
1057 if (!pos
|| css_tryget(&pos
->css
))
1060 * css reference reached zero, so iter->position will
1061 * be cleared by ->css_released. However, we should not
1062 * rely on this happening soon, because ->css_released
1063 * is called from a work queue, and by busy-waiting we
1064 * might block it. So we clear iter->position right
1067 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1075 css
= css_next_descendant_pre(css
, &root
->css
);
1078 * Reclaimers share the hierarchy walk, and a
1079 * new one might jump in right at the end of
1080 * the hierarchy - make sure they see at least
1081 * one group and restart from the beginning.
1089 * Verify the css and acquire a reference. The root
1090 * is provided by the caller, so we know it's alive
1091 * and kicking, and don't take an extra reference.
1093 memcg
= mem_cgroup_from_css(css
);
1095 if (css
== &root
->css
)
1098 if (css_tryget(css
))
1106 * The position could have already been updated by a competing
1107 * thread, so check that the value hasn't changed since we read
1108 * it to avoid reclaiming from the same cgroup twice.
1110 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1118 reclaim
->generation
= iter
->generation
;
1124 if (prev
&& prev
!= root
)
1125 css_put(&prev
->css
);
1131 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1132 * @root: hierarchy root
1133 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1135 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1136 struct mem_cgroup
*prev
)
1139 root
= root_mem_cgroup
;
1140 if (prev
&& prev
!= root
)
1141 css_put(&prev
->css
);
1144 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1145 struct mem_cgroup
*dead_memcg
)
1147 struct mem_cgroup_reclaim_iter
*iter
;
1148 struct mem_cgroup_per_node
*mz
;
1151 for_each_node(nid
) {
1152 mz
= mem_cgroup_nodeinfo(from
, nid
);
1154 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1158 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1160 struct mem_cgroup
*memcg
= dead_memcg
;
1161 struct mem_cgroup
*last
;
1164 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1166 } while ((memcg
= parent_mem_cgroup(memcg
)));
1169 * When cgruop1 non-hierarchy mode is used,
1170 * parent_mem_cgroup() does not walk all the way up to the
1171 * cgroup root (root_mem_cgroup). So we have to handle
1172 * dead_memcg from cgroup root separately.
1174 if (last
!= root_mem_cgroup
)
1175 __invalidate_reclaim_iterators(root_mem_cgroup
,
1180 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1181 * @memcg: hierarchy root
1182 * @fn: function to call for each task
1183 * @arg: argument passed to @fn
1185 * This function iterates over tasks attached to @memcg or to any of its
1186 * descendants and calls @fn for each task. If @fn returns a non-zero
1187 * value, the function breaks the iteration loop and returns the value.
1188 * Otherwise, it will iterate over all tasks and return 0.
1190 * This function must not be called for the root memory cgroup.
1192 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1193 int (*fn
)(struct task_struct
*, void *), void *arg
)
1195 struct mem_cgroup
*iter
;
1198 BUG_ON(memcg
== root_mem_cgroup
);
1200 for_each_mem_cgroup_tree(iter
, memcg
) {
1201 struct css_task_iter it
;
1202 struct task_struct
*task
;
1204 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1205 while (!ret
&& (task
= css_task_iter_next(&it
)))
1206 ret
= fn(task
, arg
);
1207 css_task_iter_end(&it
);
1209 mem_cgroup_iter_break(memcg
, iter
);
1217 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1219 * @pgdat: pgdat of the page
1221 * This function is only safe when following the LRU page isolation
1222 * and putback protocol: the LRU lock must be held, and the page must
1223 * either be PageLRU() or the caller must have isolated/allocated it.
1225 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1227 struct mem_cgroup_per_node
*mz
;
1228 struct mem_cgroup
*memcg
;
1229 struct lruvec
*lruvec
;
1231 if (mem_cgroup_disabled()) {
1232 lruvec
= &pgdat
->__lruvec
;
1236 memcg
= page
->mem_cgroup
;
1238 * Swapcache readahead pages are added to the LRU - and
1239 * possibly migrated - before they are charged.
1242 memcg
= root_mem_cgroup
;
1244 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1245 lruvec
= &mz
->lruvec
;
1248 * Since a node can be onlined after the mem_cgroup was created,
1249 * we have to be prepared to initialize lruvec->zone here;
1250 * and if offlined then reonlined, we need to reinitialize it.
1252 if (unlikely(lruvec
->pgdat
!= pgdat
))
1253 lruvec
->pgdat
= pgdat
;
1258 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1259 * @lruvec: mem_cgroup per zone lru vector
1260 * @lru: index of lru list the page is sitting on
1261 * @zid: zone id of the accounted pages
1262 * @nr_pages: positive when adding or negative when removing
1264 * This function must be called under lru_lock, just before a page is added
1265 * to or just after a page is removed from an lru list (that ordering being
1266 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1268 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1269 int zid
, int nr_pages
)
1271 struct mem_cgroup_per_node
*mz
;
1272 unsigned long *lru_size
;
1275 if (mem_cgroup_disabled())
1278 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1279 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1282 *lru_size
+= nr_pages
;
1285 if (WARN_ONCE(size
< 0,
1286 "%s(%p, %d, %d): lru_size %ld\n",
1287 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1293 *lru_size
+= nr_pages
;
1297 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1298 * @memcg: the memory cgroup
1300 * Returns the maximum amount of memory @mem can be charged with, in
1303 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1305 unsigned long margin
= 0;
1306 unsigned long count
;
1307 unsigned long limit
;
1309 count
= page_counter_read(&memcg
->memory
);
1310 limit
= READ_ONCE(memcg
->memory
.max
);
1312 margin
= limit
- count
;
1314 if (do_memsw_account()) {
1315 count
= page_counter_read(&memcg
->memsw
);
1316 limit
= READ_ONCE(memcg
->memsw
.max
);
1318 margin
= min(margin
, limit
- count
);
1327 * A routine for checking "mem" is under move_account() or not.
1329 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1330 * moving cgroups. This is for waiting at high-memory pressure
1333 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1335 struct mem_cgroup
*from
;
1336 struct mem_cgroup
*to
;
1339 * Unlike task_move routines, we access mc.to, mc.from not under
1340 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1342 spin_lock(&mc
.lock
);
1348 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1349 mem_cgroup_is_descendant(to
, memcg
);
1351 spin_unlock(&mc
.lock
);
1355 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1357 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1358 if (mem_cgroup_under_move(memcg
)) {
1360 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1361 /* moving charge context might have finished. */
1364 finish_wait(&mc
.waitq
, &wait
);
1371 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1376 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1381 * Provide statistics on the state of the memory subsystem as
1382 * well as cumulative event counters that show past behavior.
1384 * This list is ordered following a combination of these gradients:
1385 * 1) generic big picture -> specifics and details
1386 * 2) reflecting userspace activity -> reflecting kernel heuristics
1388 * Current memory state:
1391 seq_buf_printf(&s
, "anon %llu\n",
1392 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) *
1394 seq_buf_printf(&s
, "file %llu\n",
1395 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) *
1397 seq_buf_printf(&s
, "kernel_stack %llu\n",
1398 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) *
1400 seq_buf_printf(&s
, "slab %llu\n",
1401 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
1402 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
1404 seq_buf_printf(&s
, "sock %llu\n",
1405 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) *
1408 seq_buf_printf(&s
, "shmem %llu\n",
1409 (u64
)memcg_page_state(memcg
, NR_SHMEM
) *
1411 seq_buf_printf(&s
, "file_mapped %llu\n",
1412 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) *
1414 seq_buf_printf(&s
, "file_dirty %llu\n",
1415 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) *
1417 seq_buf_printf(&s
, "file_writeback %llu\n",
1418 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) *
1422 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1423 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1424 * arse because it requires migrating the work out of rmap to a place
1425 * where the page->mem_cgroup is set up and stable.
1427 seq_buf_printf(&s
, "anon_thp %llu\n",
1428 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) *
1431 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1432 seq_buf_printf(&s
, "%s %llu\n", lru_list_name(i
),
1433 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
1436 seq_buf_printf(&s
, "slab_reclaimable %llu\n",
1437 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
1439 seq_buf_printf(&s
, "slab_unreclaimable %llu\n",
1440 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
1443 /* Accumulated memory events */
1445 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1446 memcg_events(memcg
, PGFAULT
));
1447 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1448 memcg_events(memcg
, PGMAJFAULT
));
1450 seq_buf_printf(&s
, "workingset_refault %lu\n",
1451 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
1452 seq_buf_printf(&s
, "workingset_activate %lu\n",
1453 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
1454 seq_buf_printf(&s
, "workingset_nodereclaim %lu\n",
1455 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
1457 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1458 memcg_events(memcg
, PGREFILL
));
1459 seq_buf_printf(&s
, "pgscan %lu\n",
1460 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1461 memcg_events(memcg
, PGSCAN_DIRECT
));
1462 seq_buf_printf(&s
, "pgsteal %lu\n",
1463 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1464 memcg_events(memcg
, PGSTEAL_DIRECT
));
1465 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1466 memcg_events(memcg
, PGACTIVATE
));
1467 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1468 memcg_events(memcg
, PGDEACTIVATE
));
1469 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1470 memcg_events(memcg
, PGLAZYFREE
));
1471 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1472 memcg_events(memcg
, PGLAZYFREED
));
1474 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1475 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1476 memcg_events(memcg
, THP_FAULT_ALLOC
));
1477 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1478 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1479 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1481 /* The above should easily fit into one page */
1482 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1487 #define K(x) ((x) << (PAGE_SHIFT-10))
1489 * mem_cgroup_print_oom_context: Print OOM information relevant to
1490 * memory controller.
1491 * @memcg: The memory cgroup that went over limit
1492 * @p: Task that is going to be killed
1494 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1497 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1502 pr_cont(",oom_memcg=");
1503 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1505 pr_cont(",global_oom");
1507 pr_cont(",task_memcg=");
1508 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1514 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1515 * memory controller.
1516 * @memcg: The memory cgroup that went over limit
1518 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1522 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1523 K((u64
)page_counter_read(&memcg
->memory
)),
1524 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1525 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1526 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1527 K((u64
)page_counter_read(&memcg
->swap
)),
1528 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1530 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64
)page_counter_read(&memcg
->memsw
)),
1532 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1533 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1534 K((u64
)page_counter_read(&memcg
->kmem
)),
1535 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1538 pr_info("Memory cgroup stats for ");
1539 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1541 buf
= memory_stat_format(memcg
);
1549 * Return the memory (and swap, if configured) limit for a memcg.
1551 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1555 max
= READ_ONCE(memcg
->memory
.max
);
1556 if (mem_cgroup_swappiness(memcg
)) {
1557 unsigned long memsw_max
;
1558 unsigned long swap_max
;
1560 memsw_max
= memcg
->memsw
.max
;
1561 swap_max
= READ_ONCE(memcg
->swap
.max
);
1562 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1563 max
= min(max
+ swap_max
, memsw_max
);
1568 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1570 return page_counter_read(&memcg
->memory
);
1573 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1576 struct oom_control oc
= {
1580 .gfp_mask
= gfp_mask
,
1585 if (mutex_lock_killable(&oom_lock
))
1588 * A few threads which were not waiting at mutex_lock_killable() can
1589 * fail to bail out. Therefore, check again after holding oom_lock.
1591 ret
= should_force_charge() || out_of_memory(&oc
);
1592 mutex_unlock(&oom_lock
);
1596 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1599 unsigned long *total_scanned
)
1601 struct mem_cgroup
*victim
= NULL
;
1604 unsigned long excess
;
1605 unsigned long nr_scanned
;
1606 struct mem_cgroup_reclaim_cookie reclaim
= {
1610 excess
= soft_limit_excess(root_memcg
);
1613 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1618 * If we have not been able to reclaim
1619 * anything, it might because there are
1620 * no reclaimable pages under this hierarchy
1625 * We want to do more targeted reclaim.
1626 * excess >> 2 is not to excessive so as to
1627 * reclaim too much, nor too less that we keep
1628 * coming back to reclaim from this cgroup
1630 if (total
>= (excess
>> 2) ||
1631 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1636 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1637 pgdat
, &nr_scanned
);
1638 *total_scanned
+= nr_scanned
;
1639 if (!soft_limit_excess(root_memcg
))
1642 mem_cgroup_iter_break(root_memcg
, victim
);
1646 #ifdef CONFIG_LOCKDEP
1647 static struct lockdep_map memcg_oom_lock_dep_map
= {
1648 .name
= "memcg_oom_lock",
1652 static DEFINE_SPINLOCK(memcg_oom_lock
);
1655 * Check OOM-Killer is already running under our hierarchy.
1656 * If someone is running, return false.
1658 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1660 struct mem_cgroup
*iter
, *failed
= NULL
;
1662 spin_lock(&memcg_oom_lock
);
1664 for_each_mem_cgroup_tree(iter
, memcg
) {
1665 if (iter
->oom_lock
) {
1667 * this subtree of our hierarchy is already locked
1668 * so we cannot give a lock.
1671 mem_cgroup_iter_break(memcg
, iter
);
1674 iter
->oom_lock
= true;
1679 * OK, we failed to lock the whole subtree so we have
1680 * to clean up what we set up to the failing subtree
1682 for_each_mem_cgroup_tree(iter
, memcg
) {
1683 if (iter
== failed
) {
1684 mem_cgroup_iter_break(memcg
, iter
);
1687 iter
->oom_lock
= false;
1690 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1692 spin_unlock(&memcg_oom_lock
);
1697 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1699 struct mem_cgroup
*iter
;
1701 spin_lock(&memcg_oom_lock
);
1702 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1703 for_each_mem_cgroup_tree(iter
, memcg
)
1704 iter
->oom_lock
= false;
1705 spin_unlock(&memcg_oom_lock
);
1708 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1710 struct mem_cgroup
*iter
;
1712 spin_lock(&memcg_oom_lock
);
1713 for_each_mem_cgroup_tree(iter
, memcg
)
1715 spin_unlock(&memcg_oom_lock
);
1718 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1720 struct mem_cgroup
*iter
;
1723 * When a new child is created while the hierarchy is under oom,
1724 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1726 spin_lock(&memcg_oom_lock
);
1727 for_each_mem_cgroup_tree(iter
, memcg
)
1728 if (iter
->under_oom
> 0)
1730 spin_unlock(&memcg_oom_lock
);
1733 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1735 struct oom_wait_info
{
1736 struct mem_cgroup
*memcg
;
1737 wait_queue_entry_t wait
;
1740 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1741 unsigned mode
, int sync
, void *arg
)
1743 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1744 struct mem_cgroup
*oom_wait_memcg
;
1745 struct oom_wait_info
*oom_wait_info
;
1747 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1748 oom_wait_memcg
= oom_wait_info
->memcg
;
1750 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1751 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1753 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1756 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1759 * For the following lockless ->under_oom test, the only required
1760 * guarantee is that it must see the state asserted by an OOM when
1761 * this function is called as a result of userland actions
1762 * triggered by the notification of the OOM. This is trivially
1763 * achieved by invoking mem_cgroup_mark_under_oom() before
1764 * triggering notification.
1766 if (memcg
&& memcg
->under_oom
)
1767 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1777 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1779 enum oom_status ret
;
1782 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1785 memcg_memory_event(memcg
, MEMCG_OOM
);
1788 * We are in the middle of the charge context here, so we
1789 * don't want to block when potentially sitting on a callstack
1790 * that holds all kinds of filesystem and mm locks.
1792 * cgroup1 allows disabling the OOM killer and waiting for outside
1793 * handling until the charge can succeed; remember the context and put
1794 * the task to sleep at the end of the page fault when all locks are
1797 * On the other hand, in-kernel OOM killer allows for an async victim
1798 * memory reclaim (oom_reaper) and that means that we are not solely
1799 * relying on the oom victim to make a forward progress and we can
1800 * invoke the oom killer here.
1802 * Please note that mem_cgroup_out_of_memory might fail to find a
1803 * victim and then we have to bail out from the charge path.
1805 if (memcg
->oom_kill_disable
) {
1806 if (!current
->in_user_fault
)
1808 css_get(&memcg
->css
);
1809 current
->memcg_in_oom
= memcg
;
1810 current
->memcg_oom_gfp_mask
= mask
;
1811 current
->memcg_oom_order
= order
;
1816 mem_cgroup_mark_under_oom(memcg
);
1818 locked
= mem_cgroup_oom_trylock(memcg
);
1821 mem_cgroup_oom_notify(memcg
);
1823 mem_cgroup_unmark_under_oom(memcg
);
1824 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1830 mem_cgroup_oom_unlock(memcg
);
1836 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1837 * @handle: actually kill/wait or just clean up the OOM state
1839 * This has to be called at the end of a page fault if the memcg OOM
1840 * handler was enabled.
1842 * Memcg supports userspace OOM handling where failed allocations must
1843 * sleep on a waitqueue until the userspace task resolves the
1844 * situation. Sleeping directly in the charge context with all kinds
1845 * of locks held is not a good idea, instead we remember an OOM state
1846 * in the task and mem_cgroup_oom_synchronize() has to be called at
1847 * the end of the page fault to complete the OOM handling.
1849 * Returns %true if an ongoing memcg OOM situation was detected and
1850 * completed, %false otherwise.
1852 bool mem_cgroup_oom_synchronize(bool handle
)
1854 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1855 struct oom_wait_info owait
;
1858 /* OOM is global, do not handle */
1865 owait
.memcg
= memcg
;
1866 owait
.wait
.flags
= 0;
1867 owait
.wait
.func
= memcg_oom_wake_function
;
1868 owait
.wait
.private = current
;
1869 INIT_LIST_HEAD(&owait
.wait
.entry
);
1871 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1872 mem_cgroup_mark_under_oom(memcg
);
1874 locked
= mem_cgroup_oom_trylock(memcg
);
1877 mem_cgroup_oom_notify(memcg
);
1879 if (locked
&& !memcg
->oom_kill_disable
) {
1880 mem_cgroup_unmark_under_oom(memcg
);
1881 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1882 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1883 current
->memcg_oom_order
);
1886 mem_cgroup_unmark_under_oom(memcg
);
1887 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1891 mem_cgroup_oom_unlock(memcg
);
1893 * There is no guarantee that an OOM-lock contender
1894 * sees the wakeups triggered by the OOM kill
1895 * uncharges. Wake any sleepers explicitely.
1897 memcg_oom_recover(memcg
);
1900 current
->memcg_in_oom
= NULL
;
1901 css_put(&memcg
->css
);
1906 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1907 * @victim: task to be killed by the OOM killer
1908 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1910 * Returns a pointer to a memory cgroup, which has to be cleaned up
1911 * by killing all belonging OOM-killable tasks.
1913 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1915 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1916 struct mem_cgroup
*oom_domain
)
1918 struct mem_cgroup
*oom_group
= NULL
;
1919 struct mem_cgroup
*memcg
;
1921 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1925 oom_domain
= root_mem_cgroup
;
1929 memcg
= mem_cgroup_from_task(victim
);
1930 if (memcg
== root_mem_cgroup
)
1934 * If the victim task has been asynchronously moved to a different
1935 * memory cgroup, we might end up killing tasks outside oom_domain.
1936 * In this case it's better to ignore memory.group.oom.
1938 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
1942 * Traverse the memory cgroup hierarchy from the victim task's
1943 * cgroup up to the OOMing cgroup (or root) to find the
1944 * highest-level memory cgroup with oom.group set.
1946 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1947 if (memcg
->oom_group
)
1950 if (memcg
== oom_domain
)
1955 css_get(&oom_group
->css
);
1962 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1964 pr_info("Tasks in ");
1965 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1966 pr_cont(" are going to be killed due to memory.oom.group set\n");
1970 * lock_page_memcg - lock a page->mem_cgroup binding
1973 * This function protects unlocked LRU pages from being moved to
1976 * It ensures lifetime of the returned memcg. Caller is responsible
1977 * for the lifetime of the page; __unlock_page_memcg() is available
1978 * when @page might get freed inside the locked section.
1980 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1982 struct mem_cgroup
*memcg
;
1983 unsigned long flags
;
1986 * The RCU lock is held throughout the transaction. The fast
1987 * path can get away without acquiring the memcg->move_lock
1988 * because page moving starts with an RCU grace period.
1990 * The RCU lock also protects the memcg from being freed when
1991 * the page state that is going to change is the only thing
1992 * preventing the page itself from being freed. E.g. writeback
1993 * doesn't hold a page reference and relies on PG_writeback to
1994 * keep off truncation, migration and so forth.
1998 if (mem_cgroup_disabled())
2001 memcg
= page
->mem_cgroup
;
2002 if (unlikely(!memcg
))
2005 if (atomic_read(&memcg
->moving_account
) <= 0)
2008 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2009 if (memcg
!= page
->mem_cgroup
) {
2010 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2015 * When charge migration first begins, we can have locked and
2016 * unlocked page stat updates happening concurrently. Track
2017 * the task who has the lock for unlock_page_memcg().
2019 memcg
->move_lock_task
= current
;
2020 memcg
->move_lock_flags
= flags
;
2024 EXPORT_SYMBOL(lock_page_memcg
);
2027 * __unlock_page_memcg - unlock and unpin a memcg
2030 * Unlock and unpin a memcg returned by lock_page_memcg().
2032 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2034 if (memcg
&& memcg
->move_lock_task
== current
) {
2035 unsigned long flags
= memcg
->move_lock_flags
;
2037 memcg
->move_lock_task
= NULL
;
2038 memcg
->move_lock_flags
= 0;
2040 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2047 * unlock_page_memcg - unlock a page->mem_cgroup binding
2050 void unlock_page_memcg(struct page
*page
)
2052 __unlock_page_memcg(page
->mem_cgroup
);
2054 EXPORT_SYMBOL(unlock_page_memcg
);
2056 struct memcg_stock_pcp
{
2057 struct mem_cgroup
*cached
; /* this never be root cgroup */
2058 unsigned int nr_pages
;
2059 struct work_struct work
;
2060 unsigned long flags
;
2061 #define FLUSHING_CACHED_CHARGE 0
2063 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2064 static DEFINE_MUTEX(percpu_charge_mutex
);
2067 * consume_stock: Try to consume stocked charge on this cpu.
2068 * @memcg: memcg to consume from.
2069 * @nr_pages: how many pages to charge.
2071 * The charges will only happen if @memcg matches the current cpu's memcg
2072 * stock, and at least @nr_pages are available in that stock. Failure to
2073 * service an allocation will refill the stock.
2075 * returns true if successful, false otherwise.
2077 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2079 struct memcg_stock_pcp
*stock
;
2080 unsigned long flags
;
2083 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2086 local_irq_save(flags
);
2088 stock
= this_cpu_ptr(&memcg_stock
);
2089 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2090 stock
->nr_pages
-= nr_pages
;
2094 local_irq_restore(flags
);
2100 * Returns stocks cached in percpu and reset cached information.
2102 static void drain_stock(struct memcg_stock_pcp
*stock
)
2104 struct mem_cgroup
*old
= stock
->cached
;
2106 if (stock
->nr_pages
) {
2107 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2108 if (do_memsw_account())
2109 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2110 css_put_many(&old
->css
, stock
->nr_pages
);
2111 stock
->nr_pages
= 0;
2113 stock
->cached
= NULL
;
2116 static void drain_local_stock(struct work_struct
*dummy
)
2118 struct memcg_stock_pcp
*stock
;
2119 unsigned long flags
;
2122 * The only protection from memory hotplug vs. drain_stock races is
2123 * that we always operate on local CPU stock here with IRQ disabled
2125 local_irq_save(flags
);
2127 stock
= this_cpu_ptr(&memcg_stock
);
2129 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2131 local_irq_restore(flags
);
2135 * Cache charges(val) to local per_cpu area.
2136 * This will be consumed by consume_stock() function, later.
2138 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2140 struct memcg_stock_pcp
*stock
;
2141 unsigned long flags
;
2143 local_irq_save(flags
);
2145 stock
= this_cpu_ptr(&memcg_stock
);
2146 if (stock
->cached
!= memcg
) { /* reset if necessary */
2148 stock
->cached
= memcg
;
2150 stock
->nr_pages
+= nr_pages
;
2152 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2155 local_irq_restore(flags
);
2159 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2160 * of the hierarchy under it.
2162 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2166 /* If someone's already draining, avoid adding running more workers. */
2167 if (!mutex_trylock(&percpu_charge_mutex
))
2170 * Notify other cpus that system-wide "drain" is running
2171 * We do not care about races with the cpu hotplug because cpu down
2172 * as well as workers from this path always operate on the local
2173 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2176 for_each_online_cpu(cpu
) {
2177 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2178 struct mem_cgroup
*memcg
;
2182 memcg
= stock
->cached
;
2183 if (memcg
&& stock
->nr_pages
&&
2184 mem_cgroup_is_descendant(memcg
, root_memcg
))
2189 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2191 drain_local_stock(&stock
->work
);
2193 schedule_work_on(cpu
, &stock
->work
);
2197 mutex_unlock(&percpu_charge_mutex
);
2200 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2202 struct memcg_stock_pcp
*stock
;
2203 struct mem_cgroup
*memcg
, *mi
;
2205 stock
= &per_cpu(memcg_stock
, cpu
);
2208 for_each_mem_cgroup(memcg
) {
2211 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2215 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2217 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2218 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2220 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2223 for_each_node(nid
) {
2224 struct mem_cgroup_per_node
*pn
;
2226 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2227 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2230 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2231 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2235 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2238 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2240 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2241 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2248 static void reclaim_high(struct mem_cgroup
*memcg
,
2249 unsigned int nr_pages
,
2253 if (page_counter_read(&memcg
->memory
) <= READ_ONCE(memcg
->high
))
2255 memcg_memory_event(memcg
, MEMCG_HIGH
);
2256 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2257 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2258 !mem_cgroup_is_root(memcg
));
2261 static void high_work_func(struct work_struct
*work
)
2263 struct mem_cgroup
*memcg
;
2265 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2266 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2270 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2271 * enough to still cause a significant slowdown in most cases, while still
2272 * allowing diagnostics and tracing to proceed without becoming stuck.
2274 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2277 * When calculating the delay, we use these either side of the exponentiation to
2278 * maintain precision and scale to a reasonable number of jiffies (see the table
2281 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2282 * overage ratio to a delay.
2283 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2284 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2285 * to produce a reasonable delay curve.
2287 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2288 * reasonable delay curve compared to precision-adjusted overage, not
2289 * penalising heavily at first, but still making sure that growth beyond the
2290 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2291 * example, with a high of 100 megabytes:
2293 * +-------+------------------------+
2294 * | usage | time to allocate in ms |
2295 * +-------+------------------------+
2317 * +-------+------------------------+
2319 #define MEMCG_DELAY_PRECISION_SHIFT 20
2320 #define MEMCG_DELAY_SCALING_SHIFT 14
2323 * Get the number of jiffies that we should penalise a mischievous cgroup which
2324 * is exceeding its memory.high by checking both it and its ancestors.
2326 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2327 unsigned int nr_pages
)
2329 unsigned long penalty_jiffies
;
2330 u64 max_overage
= 0;
2333 unsigned long usage
, high
;
2336 usage
= page_counter_read(&memcg
->memory
);
2337 high
= READ_ONCE(memcg
->high
);
2340 * Prevent division by 0 in overage calculation by acting as if
2341 * it was a threshold of 1 page
2343 high
= max(high
, 1UL);
2345 overage
= usage
- high
;
2346 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2347 overage
= div64_u64(overage
, high
);
2349 if (overage
> max_overage
)
2350 max_overage
= overage
;
2351 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2352 !mem_cgroup_is_root(memcg
));
2358 * We use overage compared to memory.high to calculate the number of
2359 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2360 * fairly lenient on small overages, and increasingly harsh when the
2361 * memcg in question makes it clear that it has no intention of stopping
2362 * its crazy behaviour, so we exponentially increase the delay based on
2365 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2366 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2367 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2370 * Factor in the task's own contribution to the overage, such that four
2371 * N-sized allocations are throttled approximately the same as one
2372 * 4N-sized allocation.
2374 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2375 * larger the current charge patch is than that.
2377 penalty_jiffies
= penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2380 * Clamp the max delay per usermode return so as to still keep the
2381 * application moving forwards and also permit diagnostics, albeit
2384 return min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2388 * Scheduled by try_charge() to be executed from the userland return path
2389 * and reclaims memory over the high limit.
2391 void mem_cgroup_handle_over_high(void)
2393 unsigned long penalty_jiffies
;
2394 unsigned long pflags
;
2395 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2396 struct mem_cgroup
*memcg
;
2398 if (likely(!nr_pages
))
2401 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2402 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2403 current
->memcg_nr_pages_over_high
= 0;
2406 * memory.high is breached and reclaim is unable to keep up. Throttle
2407 * allocators proactively to slow down excessive growth.
2409 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
);
2412 * Don't sleep if the amount of jiffies this memcg owes us is so low
2413 * that it's not even worth doing, in an attempt to be nice to those who
2414 * go only a small amount over their memory.high value and maybe haven't
2415 * been aggressively reclaimed enough yet.
2417 if (penalty_jiffies
<= HZ
/ 100)
2421 * If we exit early, we're guaranteed to die (since
2422 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2423 * need to account for any ill-begotten jiffies to pay them off later.
2425 psi_memstall_enter(&pflags
);
2426 schedule_timeout_killable(penalty_jiffies
);
2427 psi_memstall_leave(&pflags
);
2430 css_put(&memcg
->css
);
2433 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2434 unsigned int nr_pages
)
2436 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2437 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2438 struct mem_cgroup
*mem_over_limit
;
2439 struct page_counter
*counter
;
2440 unsigned long nr_reclaimed
;
2441 bool may_swap
= true;
2442 bool drained
= false;
2443 enum oom_status oom_status
;
2445 if (mem_cgroup_is_root(memcg
))
2448 if (consume_stock(memcg
, nr_pages
))
2451 if (!do_memsw_account() ||
2452 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2453 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2455 if (do_memsw_account())
2456 page_counter_uncharge(&memcg
->memsw
, batch
);
2457 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2459 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2463 if (batch
> nr_pages
) {
2469 * Memcg doesn't have a dedicated reserve for atomic
2470 * allocations. But like the global atomic pool, we need to
2471 * put the burden of reclaim on regular allocation requests
2472 * and let these go through as privileged allocations.
2474 if (gfp_mask
& __GFP_ATOMIC
)
2478 * Unlike in global OOM situations, memcg is not in a physical
2479 * memory shortage. Allow dying and OOM-killed tasks to
2480 * bypass the last charges so that they can exit quickly and
2481 * free their memory.
2483 if (unlikely(should_force_charge()))
2487 * Prevent unbounded recursion when reclaim operations need to
2488 * allocate memory. This might exceed the limits temporarily,
2489 * but we prefer facilitating memory reclaim and getting back
2490 * under the limit over triggering OOM kills in these cases.
2492 if (unlikely(current
->flags
& PF_MEMALLOC
))
2495 if (unlikely(task_in_memcg_oom(current
)))
2498 if (!gfpflags_allow_blocking(gfp_mask
))
2501 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2503 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2504 gfp_mask
, may_swap
);
2506 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2510 drain_all_stock(mem_over_limit
);
2515 if (gfp_mask
& __GFP_NORETRY
)
2518 * Even though the limit is exceeded at this point, reclaim
2519 * may have been able to free some pages. Retry the charge
2520 * before killing the task.
2522 * Only for regular pages, though: huge pages are rather
2523 * unlikely to succeed so close to the limit, and we fall back
2524 * to regular pages anyway in case of failure.
2526 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2529 * At task move, charge accounts can be doubly counted. So, it's
2530 * better to wait until the end of task_move if something is going on.
2532 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2538 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2541 if (gfp_mask
& __GFP_NOFAIL
)
2544 if (fatal_signal_pending(current
))
2548 * keep retrying as long as the memcg oom killer is able to make
2549 * a forward progress or bypass the charge if the oom killer
2550 * couldn't make any progress.
2552 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2553 get_order(nr_pages
* PAGE_SIZE
));
2554 switch (oom_status
) {
2556 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2564 if (!(gfp_mask
& __GFP_NOFAIL
))
2568 * The allocation either can't fail or will lead to more memory
2569 * being freed very soon. Allow memory usage go over the limit
2570 * temporarily by force charging it.
2572 page_counter_charge(&memcg
->memory
, nr_pages
);
2573 if (do_memsw_account())
2574 page_counter_charge(&memcg
->memsw
, nr_pages
);
2575 css_get_many(&memcg
->css
, nr_pages
);
2580 css_get_many(&memcg
->css
, batch
);
2581 if (batch
> nr_pages
)
2582 refill_stock(memcg
, batch
- nr_pages
);
2585 * If the hierarchy is above the normal consumption range, schedule
2586 * reclaim on returning to userland. We can perform reclaim here
2587 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2588 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2589 * not recorded as it most likely matches current's and won't
2590 * change in the meantime. As high limit is checked again before
2591 * reclaim, the cost of mismatch is negligible.
2594 if (page_counter_read(&memcg
->memory
) > READ_ONCE(memcg
->high
)) {
2595 /* Don't bother a random interrupted task */
2596 if (in_interrupt()) {
2597 schedule_work(&memcg
->high_work
);
2600 current
->memcg_nr_pages_over_high
+= batch
;
2601 set_notify_resume(current
);
2604 } while ((memcg
= parent_mem_cgroup(memcg
)));
2609 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2611 if (mem_cgroup_is_root(memcg
))
2614 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2615 if (do_memsw_account())
2616 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2618 css_put_many(&memcg
->css
, nr_pages
);
2621 static void lock_page_lru(struct page
*page
, int *isolated
)
2623 pg_data_t
*pgdat
= page_pgdat(page
);
2625 spin_lock_irq(&pgdat
->lru_lock
);
2626 if (PageLRU(page
)) {
2627 struct lruvec
*lruvec
;
2629 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2631 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2637 static void unlock_page_lru(struct page
*page
, int isolated
)
2639 pg_data_t
*pgdat
= page_pgdat(page
);
2642 struct lruvec
*lruvec
;
2644 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2645 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2647 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2649 spin_unlock_irq(&pgdat
->lru_lock
);
2652 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2657 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2660 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2661 * may already be on some other mem_cgroup's LRU. Take care of it.
2664 lock_page_lru(page
, &isolated
);
2667 * Nobody should be changing or seriously looking at
2668 * page->mem_cgroup at this point:
2670 * - the page is uncharged
2672 * - the page is off-LRU
2674 * - an anonymous fault has exclusive page access, except for
2675 * a locked page table
2677 * - a page cache insertion, a swapin fault, or a migration
2678 * have the page locked
2680 page
->mem_cgroup
= memcg
;
2683 unlock_page_lru(page
, isolated
);
2686 #ifdef CONFIG_MEMCG_KMEM
2688 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2690 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2691 * cgroup_mutex, etc.
2693 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2697 if (mem_cgroup_disabled())
2700 page
= virt_to_head_page(p
);
2703 * Slab pages don't have page->mem_cgroup set because corresponding
2704 * kmem caches can be reparented during the lifetime. That's why
2705 * memcg_from_slab_page() should be used instead.
2708 return memcg_from_slab_page(page
);
2710 /* All other pages use page->mem_cgroup */
2711 return page
->mem_cgroup
;
2714 static int memcg_alloc_cache_id(void)
2719 id
= ida_simple_get(&memcg_cache_ida
,
2720 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2724 if (id
< memcg_nr_cache_ids
)
2728 * There's no space for the new id in memcg_caches arrays,
2729 * so we have to grow them.
2731 down_write(&memcg_cache_ids_sem
);
2733 size
= 2 * (id
+ 1);
2734 if (size
< MEMCG_CACHES_MIN_SIZE
)
2735 size
= MEMCG_CACHES_MIN_SIZE
;
2736 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2737 size
= MEMCG_CACHES_MAX_SIZE
;
2739 err
= memcg_update_all_caches(size
);
2741 err
= memcg_update_all_list_lrus(size
);
2743 memcg_nr_cache_ids
= size
;
2745 up_write(&memcg_cache_ids_sem
);
2748 ida_simple_remove(&memcg_cache_ida
, id
);
2754 static void memcg_free_cache_id(int id
)
2756 ida_simple_remove(&memcg_cache_ida
, id
);
2759 struct memcg_kmem_cache_create_work
{
2760 struct mem_cgroup
*memcg
;
2761 struct kmem_cache
*cachep
;
2762 struct work_struct work
;
2765 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2767 struct memcg_kmem_cache_create_work
*cw
=
2768 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2769 struct mem_cgroup
*memcg
= cw
->memcg
;
2770 struct kmem_cache
*cachep
= cw
->cachep
;
2772 memcg_create_kmem_cache(memcg
, cachep
);
2774 css_put(&memcg
->css
);
2779 * Enqueue the creation of a per-memcg kmem_cache.
2781 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2782 struct kmem_cache
*cachep
)
2784 struct memcg_kmem_cache_create_work
*cw
;
2786 if (!css_tryget_online(&memcg
->css
))
2789 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2794 cw
->cachep
= cachep
;
2795 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2797 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2800 static inline bool memcg_kmem_bypass(void)
2802 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2808 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2809 * @cachep: the original global kmem cache
2811 * Return the kmem_cache we're supposed to use for a slab allocation.
2812 * We try to use the current memcg's version of the cache.
2814 * If the cache does not exist yet, if we are the first user of it, we
2815 * create it asynchronously in a workqueue and let the current allocation
2816 * go through with the original cache.
2818 * This function takes a reference to the cache it returns to assure it
2819 * won't get destroyed while we are working with it. Once the caller is
2820 * done with it, memcg_kmem_put_cache() must be called to release the
2823 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2825 struct mem_cgroup
*memcg
;
2826 struct kmem_cache
*memcg_cachep
;
2827 struct memcg_cache_array
*arr
;
2830 VM_BUG_ON(!is_root_cache(cachep
));
2832 if (memcg_kmem_bypass())
2837 if (unlikely(current
->active_memcg
))
2838 memcg
= current
->active_memcg
;
2840 memcg
= mem_cgroup_from_task(current
);
2842 if (!memcg
|| memcg
== root_mem_cgroup
)
2845 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2849 arr
= rcu_dereference(cachep
->memcg_params
.memcg_caches
);
2852 * Make sure we will access the up-to-date value. The code updating
2853 * memcg_caches issues a write barrier to match the data dependency
2854 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2856 memcg_cachep
= READ_ONCE(arr
->entries
[kmemcg_id
]);
2859 * If we are in a safe context (can wait, and not in interrupt
2860 * context), we could be be predictable and return right away.
2861 * This would guarantee that the allocation being performed
2862 * already belongs in the new cache.
2864 * However, there are some clashes that can arrive from locking.
2865 * For instance, because we acquire the slab_mutex while doing
2866 * memcg_create_kmem_cache, this means no further allocation
2867 * could happen with the slab_mutex held. So it's better to
2870 * If the memcg is dying or memcg_cache is about to be released,
2871 * don't bother creating new kmem_caches. Because memcg_cachep
2872 * is ZEROed as the fist step of kmem offlining, we don't need
2873 * percpu_ref_tryget_live() here. css_tryget_online() check in
2874 * memcg_schedule_kmem_cache_create() will prevent us from
2875 * creation of a new kmem_cache.
2877 if (unlikely(!memcg_cachep
))
2878 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2879 else if (percpu_ref_tryget(&memcg_cachep
->memcg_params
.refcnt
))
2880 cachep
= memcg_cachep
;
2887 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2888 * @cachep: the cache returned by memcg_kmem_get_cache
2890 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2892 if (!is_root_cache(cachep
))
2893 percpu_ref_put(&cachep
->memcg_params
.refcnt
);
2897 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2898 * @memcg: memory cgroup to charge
2899 * @gfp: reclaim mode
2900 * @nr_pages: number of pages to charge
2902 * Returns 0 on success, an error code on failure.
2904 int __memcg_kmem_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
2905 unsigned int nr_pages
)
2907 struct page_counter
*counter
;
2910 ret
= try_charge(memcg
, gfp
, nr_pages
);
2914 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2915 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2918 * Enforce __GFP_NOFAIL allocation because callers are not
2919 * prepared to see failures and likely do not have any failure
2922 if (gfp
& __GFP_NOFAIL
) {
2923 page_counter_charge(&memcg
->kmem
, nr_pages
);
2926 cancel_charge(memcg
, nr_pages
);
2933 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2934 * @memcg: memcg to uncharge
2935 * @nr_pages: number of pages to uncharge
2937 void __memcg_kmem_uncharge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2939 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2940 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2942 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2943 if (do_memsw_account())
2944 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2948 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2949 * @page: page to charge
2950 * @gfp: reclaim mode
2951 * @order: allocation order
2953 * Returns 0 on success, an error code on failure.
2955 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
2957 struct mem_cgroup
*memcg
;
2960 if (memcg_kmem_bypass())
2963 memcg
= get_mem_cgroup_from_current();
2964 if (!mem_cgroup_is_root(memcg
)) {
2965 ret
= __memcg_kmem_charge(memcg
, gfp
, 1 << order
);
2967 page
->mem_cgroup
= memcg
;
2968 __SetPageKmemcg(page
);
2971 css_put(&memcg
->css
);
2976 * __memcg_kmem_uncharge_page: uncharge a kmem page
2977 * @page: page to uncharge
2978 * @order: allocation order
2980 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
2982 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2983 unsigned int nr_pages
= 1 << order
;
2988 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2989 __memcg_kmem_uncharge(memcg
, nr_pages
);
2990 page
->mem_cgroup
= NULL
;
2992 /* slab pages do not have PageKmemcg flag set */
2993 if (PageKmemcg(page
))
2994 __ClearPageKmemcg(page
);
2996 css_put_many(&memcg
->css
, nr_pages
);
2998 #endif /* CONFIG_MEMCG_KMEM */
3000 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3003 * Because tail pages are not marked as "used", set it. We're under
3004 * pgdat->lru_lock and migration entries setup in all page mappings.
3006 void mem_cgroup_split_huge_fixup(struct page
*head
)
3010 if (mem_cgroup_disabled())
3013 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
3014 head
[i
].mem_cgroup
= head
->mem_cgroup
;
3016 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
3018 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3020 #ifdef CONFIG_MEMCG_SWAP
3022 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3023 * @entry: swap entry to be moved
3024 * @from: mem_cgroup which the entry is moved from
3025 * @to: mem_cgroup which the entry is moved to
3027 * It succeeds only when the swap_cgroup's record for this entry is the same
3028 * as the mem_cgroup's id of @from.
3030 * Returns 0 on success, -EINVAL on failure.
3032 * The caller must have charged to @to, IOW, called page_counter_charge() about
3033 * both res and memsw, and called css_get().
3035 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3036 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3038 unsigned short old_id
, new_id
;
3040 old_id
= mem_cgroup_id(from
);
3041 new_id
= mem_cgroup_id(to
);
3043 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3044 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3045 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3051 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3052 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3058 static DEFINE_MUTEX(memcg_max_mutex
);
3060 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3061 unsigned long max
, bool memsw
)
3063 bool enlarge
= false;
3064 bool drained
= false;
3066 bool limits_invariant
;
3067 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3070 if (signal_pending(current
)) {
3075 mutex_lock(&memcg_max_mutex
);
3077 * Make sure that the new limit (memsw or memory limit) doesn't
3078 * break our basic invariant rule memory.max <= memsw.max.
3080 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3081 max
<= memcg
->memsw
.max
;
3082 if (!limits_invariant
) {
3083 mutex_unlock(&memcg_max_mutex
);
3087 if (max
> counter
->max
)
3089 ret
= page_counter_set_max(counter
, max
);
3090 mutex_unlock(&memcg_max_mutex
);
3096 drain_all_stock(memcg
);
3101 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3102 GFP_KERNEL
, !memsw
)) {
3108 if (!ret
&& enlarge
)
3109 memcg_oom_recover(memcg
);
3114 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3116 unsigned long *total_scanned
)
3118 unsigned long nr_reclaimed
= 0;
3119 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3120 unsigned long reclaimed
;
3122 struct mem_cgroup_tree_per_node
*mctz
;
3123 unsigned long excess
;
3124 unsigned long nr_scanned
;
3129 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3132 * Do not even bother to check the largest node if the root
3133 * is empty. Do it lockless to prevent lock bouncing. Races
3134 * are acceptable as soft limit is best effort anyway.
3136 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3140 * This loop can run a while, specially if mem_cgroup's continuously
3141 * keep exceeding their soft limit and putting the system under
3148 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3153 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3154 gfp_mask
, &nr_scanned
);
3155 nr_reclaimed
+= reclaimed
;
3156 *total_scanned
+= nr_scanned
;
3157 spin_lock_irq(&mctz
->lock
);
3158 __mem_cgroup_remove_exceeded(mz
, mctz
);
3161 * If we failed to reclaim anything from this memory cgroup
3162 * it is time to move on to the next cgroup
3166 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3168 excess
= soft_limit_excess(mz
->memcg
);
3170 * One school of thought says that we should not add
3171 * back the node to the tree if reclaim returns 0.
3172 * But our reclaim could return 0, simply because due
3173 * to priority we are exposing a smaller subset of
3174 * memory to reclaim from. Consider this as a longer
3177 /* If excess == 0, no tree ops */
3178 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3179 spin_unlock_irq(&mctz
->lock
);
3180 css_put(&mz
->memcg
->css
);
3183 * Could not reclaim anything and there are no more
3184 * mem cgroups to try or we seem to be looping without
3185 * reclaiming anything.
3187 if (!nr_reclaimed
&&
3189 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3191 } while (!nr_reclaimed
);
3193 css_put(&next_mz
->memcg
->css
);
3194 return nr_reclaimed
;
3198 * Test whether @memcg has children, dead or alive. Note that this
3199 * function doesn't care whether @memcg has use_hierarchy enabled and
3200 * returns %true if there are child csses according to the cgroup
3201 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3203 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3208 ret
= css_next_child(NULL
, &memcg
->css
);
3214 * Reclaims as many pages from the given memcg as possible.
3216 * Caller is responsible for holding css reference for memcg.
3218 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3220 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3222 /* we call try-to-free pages for make this cgroup empty */
3223 lru_add_drain_all();
3225 drain_all_stock(memcg
);
3227 /* try to free all pages in this cgroup */
3228 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3231 if (signal_pending(current
))
3234 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3238 /* maybe some writeback is necessary */
3239 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3247 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3248 char *buf
, size_t nbytes
,
3251 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3253 if (mem_cgroup_is_root(memcg
))
3255 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3258 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3261 return mem_cgroup_from_css(css
)->use_hierarchy
;
3264 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3265 struct cftype
*cft
, u64 val
)
3268 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3269 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3271 if (memcg
->use_hierarchy
== val
)
3275 * If parent's use_hierarchy is set, we can't make any modifications
3276 * in the child subtrees. If it is unset, then the change can
3277 * occur, provided the current cgroup has no children.
3279 * For the root cgroup, parent_mem is NULL, we allow value to be
3280 * set if there are no children.
3282 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3283 (val
== 1 || val
== 0)) {
3284 if (!memcg_has_children(memcg
))
3285 memcg
->use_hierarchy
= val
;
3294 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3298 if (mem_cgroup_is_root(memcg
)) {
3299 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3300 memcg_page_state(memcg
, MEMCG_RSS
);
3302 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3305 val
= page_counter_read(&memcg
->memory
);
3307 val
= page_counter_read(&memcg
->memsw
);
3320 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3323 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3324 struct page_counter
*counter
;
3326 switch (MEMFILE_TYPE(cft
->private)) {
3328 counter
= &memcg
->memory
;
3331 counter
= &memcg
->memsw
;
3334 counter
= &memcg
->kmem
;
3337 counter
= &memcg
->tcpmem
;
3343 switch (MEMFILE_ATTR(cft
->private)) {
3345 if (counter
== &memcg
->memory
)
3346 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3347 if (counter
== &memcg
->memsw
)
3348 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3349 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3351 return (u64
)counter
->max
* PAGE_SIZE
;
3353 return (u64
)counter
->watermark
* PAGE_SIZE
;
3355 return counter
->failcnt
;
3356 case RES_SOFT_LIMIT
:
3357 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3363 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3365 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3366 struct mem_cgroup
*mi
;
3369 for_each_online_cpu(cpu
)
3370 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3371 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3373 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3374 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3375 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3377 for_each_node(node
) {
3378 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3379 struct mem_cgroup_per_node
*pi
;
3381 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3384 for_each_online_cpu(cpu
)
3385 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3387 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3389 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3390 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3391 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3395 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3397 unsigned long events
[NR_VM_EVENT_ITEMS
];
3398 struct mem_cgroup
*mi
;
3401 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3404 for_each_online_cpu(cpu
)
3405 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3406 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3409 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3410 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3411 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3414 #ifdef CONFIG_MEMCG_KMEM
3415 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3419 if (cgroup_memory_nokmem
)
3422 BUG_ON(memcg
->kmemcg_id
>= 0);
3423 BUG_ON(memcg
->kmem_state
);
3425 memcg_id
= memcg_alloc_cache_id();
3429 static_branch_inc(&memcg_kmem_enabled_key
);
3431 * A memory cgroup is considered kmem-online as soon as it gets
3432 * kmemcg_id. Setting the id after enabling static branching will
3433 * guarantee no one starts accounting before all call sites are
3436 memcg
->kmemcg_id
= memcg_id
;
3437 memcg
->kmem_state
= KMEM_ONLINE
;
3438 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3443 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3445 struct cgroup_subsys_state
*css
;
3446 struct mem_cgroup
*parent
, *child
;
3449 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3452 * Clear the online state before clearing memcg_caches array
3453 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3454 * guarantees that no cache will be created for this cgroup
3455 * after we are done (see memcg_create_kmem_cache()).
3457 memcg
->kmem_state
= KMEM_ALLOCATED
;
3459 parent
= parent_mem_cgroup(memcg
);
3461 parent
= root_mem_cgroup
;
3464 * Deactivate and reparent kmem_caches.
3466 memcg_deactivate_kmem_caches(memcg
, parent
);
3468 kmemcg_id
= memcg
->kmemcg_id
;
3469 BUG_ON(kmemcg_id
< 0);
3472 * Change kmemcg_id of this cgroup and all its descendants to the
3473 * parent's id, and then move all entries from this cgroup's list_lrus
3474 * to ones of the parent. After we have finished, all list_lrus
3475 * corresponding to this cgroup are guaranteed to remain empty. The
3476 * ordering is imposed by list_lru_node->lock taken by
3477 * memcg_drain_all_list_lrus().
3479 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3480 css_for_each_descendant_pre(css
, &memcg
->css
) {
3481 child
= mem_cgroup_from_css(css
);
3482 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3483 child
->kmemcg_id
= parent
->kmemcg_id
;
3484 if (!memcg
->use_hierarchy
)
3489 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3491 memcg_free_cache_id(kmemcg_id
);
3494 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3496 /* css_alloc() failed, offlining didn't happen */
3497 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3498 memcg_offline_kmem(memcg
);
3500 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3501 WARN_ON(!list_empty(&memcg
->kmem_caches
));
3502 static_branch_dec(&memcg_kmem_enabled_key
);
3506 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3510 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3513 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3516 #endif /* CONFIG_MEMCG_KMEM */
3518 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3523 mutex_lock(&memcg_max_mutex
);
3524 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3525 mutex_unlock(&memcg_max_mutex
);
3529 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3533 mutex_lock(&memcg_max_mutex
);
3535 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3539 if (!memcg
->tcpmem_active
) {
3541 * The active flag needs to be written after the static_key
3542 * update. This is what guarantees that the socket activation
3543 * function is the last one to run. See mem_cgroup_sk_alloc()
3544 * for details, and note that we don't mark any socket as
3545 * belonging to this memcg until that flag is up.
3547 * We need to do this, because static_keys will span multiple
3548 * sites, but we can't control their order. If we mark a socket
3549 * as accounted, but the accounting functions are not patched in
3550 * yet, we'll lose accounting.
3552 * We never race with the readers in mem_cgroup_sk_alloc(),
3553 * because when this value change, the code to process it is not
3556 static_branch_inc(&memcg_sockets_enabled_key
);
3557 memcg
->tcpmem_active
= true;
3560 mutex_unlock(&memcg_max_mutex
);
3565 * The user of this function is...
3568 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3569 char *buf
, size_t nbytes
, loff_t off
)
3571 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3572 unsigned long nr_pages
;
3575 buf
= strstrip(buf
);
3576 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3580 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3582 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3586 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3588 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3591 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3594 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3595 "Please report your usecase to linux-mm@kvack.org if you "
3596 "depend on this functionality.\n");
3597 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3600 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3604 case RES_SOFT_LIMIT
:
3605 memcg
->soft_limit
= nr_pages
;
3609 return ret
?: nbytes
;
3612 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3613 size_t nbytes
, loff_t off
)
3615 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3616 struct page_counter
*counter
;
3618 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3620 counter
= &memcg
->memory
;
3623 counter
= &memcg
->memsw
;
3626 counter
= &memcg
->kmem
;
3629 counter
= &memcg
->tcpmem
;
3635 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3637 page_counter_reset_watermark(counter
);
3640 counter
->failcnt
= 0;
3649 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3652 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3656 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3657 struct cftype
*cft
, u64 val
)
3659 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3661 if (val
& ~MOVE_MASK
)
3665 * No kind of locking is needed in here, because ->can_attach() will
3666 * check this value once in the beginning of the process, and then carry
3667 * on with stale data. This means that changes to this value will only
3668 * affect task migrations starting after the change.
3670 memcg
->move_charge_at_immigrate
= val
;
3674 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3675 struct cftype
*cft
, u64 val
)
3683 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3684 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3685 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3687 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3688 int nid
, unsigned int lru_mask
)
3690 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3691 unsigned long nr
= 0;
3694 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3697 if (!(BIT(lru
) & lru_mask
))
3699 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3704 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3705 unsigned int lru_mask
)
3707 unsigned long nr
= 0;
3711 if (!(BIT(lru
) & lru_mask
))
3713 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3718 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3722 unsigned int lru_mask
;
3725 static const struct numa_stat stats
[] = {
3726 { "total", LRU_ALL
},
3727 { "file", LRU_ALL_FILE
},
3728 { "anon", LRU_ALL_ANON
},
3729 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3731 const struct numa_stat
*stat
;
3734 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3736 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3737 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3738 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3739 for_each_node_state(nid
, N_MEMORY
) {
3740 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3742 seq_printf(m
, " N%d=%lu", nid
, nr
);
3747 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3748 struct mem_cgroup
*iter
;
3751 for_each_mem_cgroup_tree(iter
, memcg
)
3752 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3753 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3754 for_each_node_state(nid
, N_MEMORY
) {
3756 for_each_mem_cgroup_tree(iter
, memcg
)
3757 nr
+= mem_cgroup_node_nr_lru_pages(
3758 iter
, nid
, stat
->lru_mask
);
3759 seq_printf(m
, " N%d=%lu", nid
, nr
);
3766 #endif /* CONFIG_NUMA */
3768 static const unsigned int memcg1_stats
[] = {
3779 static const char *const memcg1_stat_names
[] = {
3790 /* Universal VM events cgroup1 shows, original sort order */
3791 static const unsigned int memcg1_events
[] = {
3798 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3800 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3801 unsigned long memory
, memsw
;
3802 struct mem_cgroup
*mi
;
3805 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3807 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3808 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3810 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3811 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3815 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3816 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
3817 memcg_events_local(memcg
, memcg1_events
[i
]));
3819 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3820 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
3821 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3824 /* Hierarchical information */
3825 memory
= memsw
= PAGE_COUNTER_MAX
;
3826 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3827 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
3828 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
3830 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3831 (u64
)memory
* PAGE_SIZE
);
3832 if (do_memsw_account())
3833 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3834 (u64
)memsw
* PAGE_SIZE
);
3836 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3837 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3839 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3840 (u64
)memcg_page_state(memcg
, memcg1_stats
[i
]) *
3844 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3845 seq_printf(m
, "total_%s %llu\n",
3846 vm_event_name(memcg1_events
[i
]),
3847 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
3849 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3850 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
3851 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3854 #ifdef CONFIG_DEBUG_VM
3857 struct mem_cgroup_per_node
*mz
;
3858 struct zone_reclaim_stat
*rstat
;
3859 unsigned long recent_rotated
[2] = {0, 0};
3860 unsigned long recent_scanned
[2] = {0, 0};
3862 for_each_online_pgdat(pgdat
) {
3863 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3864 rstat
= &mz
->lruvec
.reclaim_stat
;
3866 recent_rotated
[0] += rstat
->recent_rotated
[0];
3867 recent_rotated
[1] += rstat
->recent_rotated
[1];
3868 recent_scanned
[0] += rstat
->recent_scanned
[0];
3869 recent_scanned
[1] += rstat
->recent_scanned
[1];
3871 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3872 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3873 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3874 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3881 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3884 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3886 return mem_cgroup_swappiness(memcg
);
3889 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3890 struct cftype
*cft
, u64 val
)
3892 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3898 memcg
->swappiness
= val
;
3900 vm_swappiness
= val
;
3905 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3907 struct mem_cgroup_threshold_ary
*t
;
3908 unsigned long usage
;
3913 t
= rcu_dereference(memcg
->thresholds
.primary
);
3915 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3920 usage
= mem_cgroup_usage(memcg
, swap
);
3923 * current_threshold points to threshold just below or equal to usage.
3924 * If it's not true, a threshold was crossed after last
3925 * call of __mem_cgroup_threshold().
3927 i
= t
->current_threshold
;
3930 * Iterate backward over array of thresholds starting from
3931 * current_threshold and check if a threshold is crossed.
3932 * If none of thresholds below usage is crossed, we read
3933 * only one element of the array here.
3935 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3936 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3938 /* i = current_threshold + 1 */
3942 * Iterate forward over array of thresholds starting from
3943 * current_threshold+1 and check if a threshold is crossed.
3944 * If none of thresholds above usage is crossed, we read
3945 * only one element of the array here.
3947 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3948 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3950 /* Update current_threshold */
3951 t
->current_threshold
= i
- 1;
3956 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3959 __mem_cgroup_threshold(memcg
, false);
3960 if (do_memsw_account())
3961 __mem_cgroup_threshold(memcg
, true);
3963 memcg
= parent_mem_cgroup(memcg
);
3967 static int compare_thresholds(const void *a
, const void *b
)
3969 const struct mem_cgroup_threshold
*_a
= a
;
3970 const struct mem_cgroup_threshold
*_b
= b
;
3972 if (_a
->threshold
> _b
->threshold
)
3975 if (_a
->threshold
< _b
->threshold
)
3981 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3983 struct mem_cgroup_eventfd_list
*ev
;
3985 spin_lock(&memcg_oom_lock
);
3987 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3988 eventfd_signal(ev
->eventfd
, 1);
3990 spin_unlock(&memcg_oom_lock
);
3994 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3996 struct mem_cgroup
*iter
;
3998 for_each_mem_cgroup_tree(iter
, memcg
)
3999 mem_cgroup_oom_notify_cb(iter
);
4002 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4003 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4005 struct mem_cgroup_thresholds
*thresholds
;
4006 struct mem_cgroup_threshold_ary
*new;
4007 unsigned long threshold
;
4008 unsigned long usage
;
4011 ret
= page_counter_memparse(args
, "-1", &threshold
);
4015 mutex_lock(&memcg
->thresholds_lock
);
4018 thresholds
= &memcg
->thresholds
;
4019 usage
= mem_cgroup_usage(memcg
, false);
4020 } else if (type
== _MEMSWAP
) {
4021 thresholds
= &memcg
->memsw_thresholds
;
4022 usage
= mem_cgroup_usage(memcg
, true);
4026 /* Check if a threshold crossed before adding a new one */
4027 if (thresholds
->primary
)
4028 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4030 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4032 /* Allocate memory for new array of thresholds */
4033 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4040 /* Copy thresholds (if any) to new array */
4041 if (thresholds
->primary
) {
4042 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4043 sizeof(struct mem_cgroup_threshold
));
4046 /* Add new threshold */
4047 new->entries
[size
- 1].eventfd
= eventfd
;
4048 new->entries
[size
- 1].threshold
= threshold
;
4050 /* Sort thresholds. Registering of new threshold isn't time-critical */
4051 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4052 compare_thresholds
, NULL
);
4054 /* Find current threshold */
4055 new->current_threshold
= -1;
4056 for (i
= 0; i
< size
; i
++) {
4057 if (new->entries
[i
].threshold
<= usage
) {
4059 * new->current_threshold will not be used until
4060 * rcu_assign_pointer(), so it's safe to increment
4063 ++new->current_threshold
;
4068 /* Free old spare buffer and save old primary buffer as spare */
4069 kfree(thresholds
->spare
);
4070 thresholds
->spare
= thresholds
->primary
;
4072 rcu_assign_pointer(thresholds
->primary
, new);
4074 /* To be sure that nobody uses thresholds */
4078 mutex_unlock(&memcg
->thresholds_lock
);
4083 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4084 struct eventfd_ctx
*eventfd
, const char *args
)
4086 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4089 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4090 struct eventfd_ctx
*eventfd
, const char *args
)
4092 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4095 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4096 struct eventfd_ctx
*eventfd
, enum res_type type
)
4098 struct mem_cgroup_thresholds
*thresholds
;
4099 struct mem_cgroup_threshold_ary
*new;
4100 unsigned long usage
;
4101 int i
, j
, size
, entries
;
4103 mutex_lock(&memcg
->thresholds_lock
);
4106 thresholds
= &memcg
->thresholds
;
4107 usage
= mem_cgroup_usage(memcg
, false);
4108 } else if (type
== _MEMSWAP
) {
4109 thresholds
= &memcg
->memsw_thresholds
;
4110 usage
= mem_cgroup_usage(memcg
, true);
4114 if (!thresholds
->primary
)
4117 /* Check if a threshold crossed before removing */
4118 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4120 /* Calculate new number of threshold */
4122 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4123 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4129 new = thresholds
->spare
;
4131 /* If no items related to eventfd have been cleared, nothing to do */
4135 /* Set thresholds array to NULL if we don't have thresholds */
4144 /* Copy thresholds and find current threshold */
4145 new->current_threshold
= -1;
4146 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4147 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4150 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4151 if (new->entries
[j
].threshold
<= usage
) {
4153 * new->current_threshold will not be used
4154 * until rcu_assign_pointer(), so it's safe to increment
4157 ++new->current_threshold
;
4163 /* Swap primary and spare array */
4164 thresholds
->spare
= thresholds
->primary
;
4166 rcu_assign_pointer(thresholds
->primary
, new);
4168 /* To be sure that nobody uses thresholds */
4171 /* If all events are unregistered, free the spare array */
4173 kfree(thresholds
->spare
);
4174 thresholds
->spare
= NULL
;
4177 mutex_unlock(&memcg
->thresholds_lock
);
4180 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4181 struct eventfd_ctx
*eventfd
)
4183 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4186 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4187 struct eventfd_ctx
*eventfd
)
4189 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4192 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4193 struct eventfd_ctx
*eventfd
, const char *args
)
4195 struct mem_cgroup_eventfd_list
*event
;
4197 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4201 spin_lock(&memcg_oom_lock
);
4203 event
->eventfd
= eventfd
;
4204 list_add(&event
->list
, &memcg
->oom_notify
);
4206 /* already in OOM ? */
4207 if (memcg
->under_oom
)
4208 eventfd_signal(eventfd
, 1);
4209 spin_unlock(&memcg_oom_lock
);
4214 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4215 struct eventfd_ctx
*eventfd
)
4217 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4219 spin_lock(&memcg_oom_lock
);
4221 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4222 if (ev
->eventfd
== eventfd
) {
4223 list_del(&ev
->list
);
4228 spin_unlock(&memcg_oom_lock
);
4231 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4233 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4235 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4236 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4237 seq_printf(sf
, "oom_kill %lu\n",
4238 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4242 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4243 struct cftype
*cft
, u64 val
)
4245 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4247 /* cannot set to root cgroup and only 0 and 1 are allowed */
4248 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4251 memcg
->oom_kill_disable
= val
;
4253 memcg_oom_recover(memcg
);
4258 #ifdef CONFIG_CGROUP_WRITEBACK
4260 #include <trace/events/writeback.h>
4262 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4264 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4267 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4269 wb_domain_exit(&memcg
->cgwb_domain
);
4272 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4274 wb_domain_size_changed(&memcg
->cgwb_domain
);
4277 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4281 if (!memcg
->css
.parent
)
4284 return &memcg
->cgwb_domain
;
4288 * idx can be of type enum memcg_stat_item or node_stat_item.
4289 * Keep in sync with memcg_exact_page().
4291 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4293 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4296 for_each_online_cpu(cpu
)
4297 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4304 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4305 * @wb: bdi_writeback in question
4306 * @pfilepages: out parameter for number of file pages
4307 * @pheadroom: out parameter for number of allocatable pages according to memcg
4308 * @pdirty: out parameter for number of dirty pages
4309 * @pwriteback: out parameter for number of pages under writeback
4311 * Determine the numbers of file, headroom, dirty, and writeback pages in
4312 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4313 * is a bit more involved.
4315 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4316 * headroom is calculated as the lowest headroom of itself and the
4317 * ancestors. Note that this doesn't consider the actual amount of
4318 * available memory in the system. The caller should further cap
4319 * *@pheadroom accordingly.
4321 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4322 unsigned long *pheadroom
, unsigned long *pdirty
,
4323 unsigned long *pwriteback
)
4325 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4326 struct mem_cgroup
*parent
;
4328 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4330 /* this should eventually include NR_UNSTABLE_NFS */
4331 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4332 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4333 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4334 *pheadroom
= PAGE_COUNTER_MAX
;
4336 while ((parent
= parent_mem_cgroup(memcg
))) {
4337 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4338 READ_ONCE(memcg
->high
));
4339 unsigned long used
= page_counter_read(&memcg
->memory
);
4341 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4347 * Foreign dirty flushing
4349 * There's an inherent mismatch between memcg and writeback. The former
4350 * trackes ownership per-page while the latter per-inode. This was a
4351 * deliberate design decision because honoring per-page ownership in the
4352 * writeback path is complicated, may lead to higher CPU and IO overheads
4353 * and deemed unnecessary given that write-sharing an inode across
4354 * different cgroups isn't a common use-case.
4356 * Combined with inode majority-writer ownership switching, this works well
4357 * enough in most cases but there are some pathological cases. For
4358 * example, let's say there are two cgroups A and B which keep writing to
4359 * different but confined parts of the same inode. B owns the inode and
4360 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4361 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4362 * triggering background writeback. A will be slowed down without a way to
4363 * make writeback of the dirty pages happen.
4365 * Conditions like the above can lead to a cgroup getting repatedly and
4366 * severely throttled after making some progress after each
4367 * dirty_expire_interval while the underyling IO device is almost
4370 * Solving this problem completely requires matching the ownership tracking
4371 * granularities between memcg and writeback in either direction. However,
4372 * the more egregious behaviors can be avoided by simply remembering the
4373 * most recent foreign dirtying events and initiating remote flushes on
4374 * them when local writeback isn't enough to keep the memory clean enough.
4376 * The following two functions implement such mechanism. When a foreign
4377 * page - a page whose memcg and writeback ownerships don't match - is
4378 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4379 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4380 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4381 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4382 * foreign bdi_writebacks which haven't expired. Both the numbers of
4383 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4384 * limited to MEMCG_CGWB_FRN_CNT.
4386 * The mechanism only remembers IDs and doesn't hold any object references.
4387 * As being wrong occasionally doesn't matter, updates and accesses to the
4388 * records are lockless and racy.
4390 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4391 struct bdi_writeback
*wb
)
4393 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4394 struct memcg_cgwb_frn
*frn
;
4395 u64 now
= get_jiffies_64();
4396 u64 oldest_at
= now
;
4400 trace_track_foreign_dirty(page
, wb
);
4403 * Pick the slot to use. If there is already a slot for @wb, keep
4404 * using it. If not replace the oldest one which isn't being
4407 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4408 frn
= &memcg
->cgwb_frn
[i
];
4409 if (frn
->bdi_id
== wb
->bdi
->id
&&
4410 frn
->memcg_id
== wb
->memcg_css
->id
)
4412 if (time_before64(frn
->at
, oldest_at
) &&
4413 atomic_read(&frn
->done
.cnt
) == 1) {
4415 oldest_at
= frn
->at
;
4419 if (i
< MEMCG_CGWB_FRN_CNT
) {
4421 * Re-using an existing one. Update timestamp lazily to
4422 * avoid making the cacheline hot. We want them to be
4423 * reasonably up-to-date and significantly shorter than
4424 * dirty_expire_interval as that's what expires the record.
4425 * Use the shorter of 1s and dirty_expire_interval / 8.
4427 unsigned long update_intv
=
4428 min_t(unsigned long, HZ
,
4429 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4431 if (time_before64(frn
->at
, now
- update_intv
))
4433 } else if (oldest
>= 0) {
4434 /* replace the oldest free one */
4435 frn
= &memcg
->cgwb_frn
[oldest
];
4436 frn
->bdi_id
= wb
->bdi
->id
;
4437 frn
->memcg_id
= wb
->memcg_css
->id
;
4442 /* issue foreign writeback flushes for recorded foreign dirtying events */
4443 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4445 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4446 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4447 u64 now
= jiffies_64
;
4450 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4451 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4454 * If the record is older than dirty_expire_interval,
4455 * writeback on it has already started. No need to kick it
4456 * off again. Also, don't start a new one if there's
4457 * already one in flight.
4459 if (time_after64(frn
->at
, now
- intv
) &&
4460 atomic_read(&frn
->done
.cnt
) == 1) {
4462 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4463 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4464 WB_REASON_FOREIGN_FLUSH
,
4470 #else /* CONFIG_CGROUP_WRITEBACK */
4472 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4477 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4481 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4485 #endif /* CONFIG_CGROUP_WRITEBACK */
4488 * DO NOT USE IN NEW FILES.
4490 * "cgroup.event_control" implementation.
4492 * This is way over-engineered. It tries to support fully configurable
4493 * events for each user. Such level of flexibility is completely
4494 * unnecessary especially in the light of the planned unified hierarchy.
4496 * Please deprecate this and replace with something simpler if at all
4501 * Unregister event and free resources.
4503 * Gets called from workqueue.
4505 static void memcg_event_remove(struct work_struct
*work
)
4507 struct mem_cgroup_event
*event
=
4508 container_of(work
, struct mem_cgroup_event
, remove
);
4509 struct mem_cgroup
*memcg
= event
->memcg
;
4511 remove_wait_queue(event
->wqh
, &event
->wait
);
4513 event
->unregister_event(memcg
, event
->eventfd
);
4515 /* Notify userspace the event is going away. */
4516 eventfd_signal(event
->eventfd
, 1);
4518 eventfd_ctx_put(event
->eventfd
);
4520 css_put(&memcg
->css
);
4524 * Gets called on EPOLLHUP on eventfd when user closes it.
4526 * Called with wqh->lock held and interrupts disabled.
4528 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4529 int sync
, void *key
)
4531 struct mem_cgroup_event
*event
=
4532 container_of(wait
, struct mem_cgroup_event
, wait
);
4533 struct mem_cgroup
*memcg
= event
->memcg
;
4534 __poll_t flags
= key_to_poll(key
);
4536 if (flags
& EPOLLHUP
) {
4538 * If the event has been detached at cgroup removal, we
4539 * can simply return knowing the other side will cleanup
4542 * We can't race against event freeing since the other
4543 * side will require wqh->lock via remove_wait_queue(),
4546 spin_lock(&memcg
->event_list_lock
);
4547 if (!list_empty(&event
->list
)) {
4548 list_del_init(&event
->list
);
4550 * We are in atomic context, but cgroup_event_remove()
4551 * may sleep, so we have to call it in workqueue.
4553 schedule_work(&event
->remove
);
4555 spin_unlock(&memcg
->event_list_lock
);
4561 static void memcg_event_ptable_queue_proc(struct file
*file
,
4562 wait_queue_head_t
*wqh
, poll_table
*pt
)
4564 struct mem_cgroup_event
*event
=
4565 container_of(pt
, struct mem_cgroup_event
, pt
);
4568 add_wait_queue(wqh
, &event
->wait
);
4572 * DO NOT USE IN NEW FILES.
4574 * Parse input and register new cgroup event handler.
4576 * Input must be in format '<event_fd> <control_fd> <args>'.
4577 * Interpretation of args is defined by control file implementation.
4579 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4580 char *buf
, size_t nbytes
, loff_t off
)
4582 struct cgroup_subsys_state
*css
= of_css(of
);
4583 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4584 struct mem_cgroup_event
*event
;
4585 struct cgroup_subsys_state
*cfile_css
;
4586 unsigned int efd
, cfd
;
4593 buf
= strstrip(buf
);
4595 efd
= simple_strtoul(buf
, &endp
, 10);
4600 cfd
= simple_strtoul(buf
, &endp
, 10);
4601 if ((*endp
!= ' ') && (*endp
!= '\0'))
4605 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4609 event
->memcg
= memcg
;
4610 INIT_LIST_HEAD(&event
->list
);
4611 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4612 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4613 INIT_WORK(&event
->remove
, memcg_event_remove
);
4621 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4622 if (IS_ERR(event
->eventfd
)) {
4623 ret
= PTR_ERR(event
->eventfd
);
4630 goto out_put_eventfd
;
4633 /* the process need read permission on control file */
4634 /* AV: shouldn't we check that it's been opened for read instead? */
4635 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4640 * Determine the event callbacks and set them in @event. This used
4641 * to be done via struct cftype but cgroup core no longer knows
4642 * about these events. The following is crude but the whole thing
4643 * is for compatibility anyway.
4645 * DO NOT ADD NEW FILES.
4647 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4649 if (!strcmp(name
, "memory.usage_in_bytes")) {
4650 event
->register_event
= mem_cgroup_usage_register_event
;
4651 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4652 } else if (!strcmp(name
, "memory.oom_control")) {
4653 event
->register_event
= mem_cgroup_oom_register_event
;
4654 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4655 } else if (!strcmp(name
, "memory.pressure_level")) {
4656 event
->register_event
= vmpressure_register_event
;
4657 event
->unregister_event
= vmpressure_unregister_event
;
4658 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4659 event
->register_event
= memsw_cgroup_usage_register_event
;
4660 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4667 * Verify @cfile should belong to @css. Also, remaining events are
4668 * automatically removed on cgroup destruction but the removal is
4669 * asynchronous, so take an extra ref on @css.
4671 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4672 &memory_cgrp_subsys
);
4674 if (IS_ERR(cfile_css
))
4676 if (cfile_css
!= css
) {
4681 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4685 vfs_poll(efile
.file
, &event
->pt
);
4687 spin_lock(&memcg
->event_list_lock
);
4688 list_add(&event
->list
, &memcg
->event_list
);
4689 spin_unlock(&memcg
->event_list_lock
);
4701 eventfd_ctx_put(event
->eventfd
);
4710 static struct cftype mem_cgroup_legacy_files
[] = {
4712 .name
= "usage_in_bytes",
4713 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4714 .read_u64
= mem_cgroup_read_u64
,
4717 .name
= "max_usage_in_bytes",
4718 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4719 .write
= mem_cgroup_reset
,
4720 .read_u64
= mem_cgroup_read_u64
,
4723 .name
= "limit_in_bytes",
4724 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4725 .write
= mem_cgroup_write
,
4726 .read_u64
= mem_cgroup_read_u64
,
4729 .name
= "soft_limit_in_bytes",
4730 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4731 .write
= mem_cgroup_write
,
4732 .read_u64
= mem_cgroup_read_u64
,
4736 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4737 .write
= mem_cgroup_reset
,
4738 .read_u64
= mem_cgroup_read_u64
,
4742 .seq_show
= memcg_stat_show
,
4745 .name
= "force_empty",
4746 .write
= mem_cgroup_force_empty_write
,
4749 .name
= "use_hierarchy",
4750 .write_u64
= mem_cgroup_hierarchy_write
,
4751 .read_u64
= mem_cgroup_hierarchy_read
,
4754 .name
= "cgroup.event_control", /* XXX: for compat */
4755 .write
= memcg_write_event_control
,
4756 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4759 .name
= "swappiness",
4760 .read_u64
= mem_cgroup_swappiness_read
,
4761 .write_u64
= mem_cgroup_swappiness_write
,
4764 .name
= "move_charge_at_immigrate",
4765 .read_u64
= mem_cgroup_move_charge_read
,
4766 .write_u64
= mem_cgroup_move_charge_write
,
4769 .name
= "oom_control",
4770 .seq_show
= mem_cgroup_oom_control_read
,
4771 .write_u64
= mem_cgroup_oom_control_write
,
4772 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4775 .name
= "pressure_level",
4779 .name
= "numa_stat",
4780 .seq_show
= memcg_numa_stat_show
,
4784 .name
= "kmem.limit_in_bytes",
4785 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4786 .write
= mem_cgroup_write
,
4787 .read_u64
= mem_cgroup_read_u64
,
4790 .name
= "kmem.usage_in_bytes",
4791 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4792 .read_u64
= mem_cgroup_read_u64
,
4795 .name
= "kmem.failcnt",
4796 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4797 .write
= mem_cgroup_reset
,
4798 .read_u64
= mem_cgroup_read_u64
,
4801 .name
= "kmem.max_usage_in_bytes",
4802 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4803 .write
= mem_cgroup_reset
,
4804 .read_u64
= mem_cgroup_read_u64
,
4806 #if defined(CONFIG_MEMCG_KMEM) && \
4807 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4809 .name
= "kmem.slabinfo",
4810 .seq_start
= memcg_slab_start
,
4811 .seq_next
= memcg_slab_next
,
4812 .seq_stop
= memcg_slab_stop
,
4813 .seq_show
= memcg_slab_show
,
4817 .name
= "kmem.tcp.limit_in_bytes",
4818 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4819 .write
= mem_cgroup_write
,
4820 .read_u64
= mem_cgroup_read_u64
,
4823 .name
= "kmem.tcp.usage_in_bytes",
4824 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4825 .read_u64
= mem_cgroup_read_u64
,
4828 .name
= "kmem.tcp.failcnt",
4829 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4830 .write
= mem_cgroup_reset
,
4831 .read_u64
= mem_cgroup_read_u64
,
4834 .name
= "kmem.tcp.max_usage_in_bytes",
4835 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4836 .write
= mem_cgroup_reset
,
4837 .read_u64
= mem_cgroup_read_u64
,
4839 { }, /* terminate */
4843 * Private memory cgroup IDR
4845 * Swap-out records and page cache shadow entries need to store memcg
4846 * references in constrained space, so we maintain an ID space that is
4847 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4848 * memory-controlled cgroups to 64k.
4850 * However, there usually are many references to the oflline CSS after
4851 * the cgroup has been destroyed, such as page cache or reclaimable
4852 * slab objects, that don't need to hang on to the ID. We want to keep
4853 * those dead CSS from occupying IDs, or we might quickly exhaust the
4854 * relatively small ID space and prevent the creation of new cgroups
4855 * even when there are much fewer than 64k cgroups - possibly none.
4857 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4858 * be freed and recycled when it's no longer needed, which is usually
4859 * when the CSS is offlined.
4861 * The only exception to that are records of swapped out tmpfs/shmem
4862 * pages that need to be attributed to live ancestors on swapin. But
4863 * those references are manageable from userspace.
4866 static DEFINE_IDR(mem_cgroup_idr
);
4868 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4870 if (memcg
->id
.id
> 0) {
4871 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4876 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
4879 refcount_add(n
, &memcg
->id
.ref
);
4882 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4884 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4885 mem_cgroup_id_remove(memcg
);
4887 /* Memcg ID pins CSS */
4888 css_put(&memcg
->css
);
4892 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4894 mem_cgroup_id_put_many(memcg
, 1);
4898 * mem_cgroup_from_id - look up a memcg from a memcg id
4899 * @id: the memcg id to look up
4901 * Caller must hold rcu_read_lock().
4903 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4905 WARN_ON_ONCE(!rcu_read_lock_held());
4906 return idr_find(&mem_cgroup_idr
, id
);
4909 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4911 struct mem_cgroup_per_node
*pn
;
4914 * This routine is called against possible nodes.
4915 * But it's BUG to call kmalloc() against offline node.
4917 * TODO: this routine can waste much memory for nodes which will
4918 * never be onlined. It's better to use memory hotplug callback
4921 if (!node_state(node
, N_NORMAL_MEMORY
))
4923 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4927 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
4928 if (!pn
->lruvec_stat_local
) {
4933 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4934 if (!pn
->lruvec_stat_cpu
) {
4935 free_percpu(pn
->lruvec_stat_local
);
4940 lruvec_init(&pn
->lruvec
);
4941 pn
->usage_in_excess
= 0;
4942 pn
->on_tree
= false;
4945 memcg
->nodeinfo
[node
] = pn
;
4949 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4951 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4956 free_percpu(pn
->lruvec_stat_cpu
);
4957 free_percpu(pn
->lruvec_stat_local
);
4961 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4966 free_mem_cgroup_per_node_info(memcg
, node
);
4967 free_percpu(memcg
->vmstats_percpu
);
4968 free_percpu(memcg
->vmstats_local
);
4972 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4974 memcg_wb_domain_exit(memcg
);
4976 * Flush percpu vmstats and vmevents to guarantee the value correctness
4977 * on parent's and all ancestor levels.
4979 memcg_flush_percpu_vmstats(memcg
);
4980 memcg_flush_percpu_vmevents(memcg
);
4981 __mem_cgroup_free(memcg
);
4984 static struct mem_cgroup
*mem_cgroup_alloc(void)
4986 struct mem_cgroup
*memcg
;
4989 int __maybe_unused i
;
4991 size
= sizeof(struct mem_cgroup
);
4992 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4994 memcg
= kzalloc(size
, GFP_KERNEL
);
4998 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4999 1, MEM_CGROUP_ID_MAX
,
5001 if (memcg
->id
.id
< 0)
5004 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
5005 if (!memcg
->vmstats_local
)
5008 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
5009 if (!memcg
->vmstats_percpu
)
5013 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5016 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5019 INIT_WORK(&memcg
->high_work
, high_work_func
);
5020 INIT_LIST_HEAD(&memcg
->oom_notify
);
5021 mutex_init(&memcg
->thresholds_lock
);
5022 spin_lock_init(&memcg
->move_lock
);
5023 vmpressure_init(&memcg
->vmpressure
);
5024 INIT_LIST_HEAD(&memcg
->event_list
);
5025 spin_lock_init(&memcg
->event_list_lock
);
5026 memcg
->socket_pressure
= jiffies
;
5027 #ifdef CONFIG_MEMCG_KMEM
5028 memcg
->kmemcg_id
= -1;
5030 #ifdef CONFIG_CGROUP_WRITEBACK
5031 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5032 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5033 memcg
->cgwb_frn
[i
].done
=
5034 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5036 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5037 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5038 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5039 memcg
->deferred_split_queue
.split_queue_len
= 0;
5041 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5044 mem_cgroup_id_remove(memcg
);
5045 __mem_cgroup_free(memcg
);
5049 static struct cgroup_subsys_state
* __ref
5050 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5052 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5053 struct mem_cgroup
*memcg
;
5054 long error
= -ENOMEM
;
5056 memcg
= mem_cgroup_alloc();
5058 return ERR_PTR(error
);
5060 WRITE_ONCE(memcg
->high
, PAGE_COUNTER_MAX
);
5061 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5063 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5064 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5066 if (parent
&& parent
->use_hierarchy
) {
5067 memcg
->use_hierarchy
= true;
5068 page_counter_init(&memcg
->memory
, &parent
->memory
);
5069 page_counter_init(&memcg
->swap
, &parent
->swap
);
5070 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
5071 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5072 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5074 page_counter_init(&memcg
->memory
, NULL
);
5075 page_counter_init(&memcg
->swap
, NULL
);
5076 page_counter_init(&memcg
->memsw
, NULL
);
5077 page_counter_init(&memcg
->kmem
, NULL
);
5078 page_counter_init(&memcg
->tcpmem
, NULL
);
5080 * Deeper hierachy with use_hierarchy == false doesn't make
5081 * much sense so let cgroup subsystem know about this
5082 * unfortunate state in our controller.
5084 if (parent
!= root_mem_cgroup
)
5085 memory_cgrp_subsys
.broken_hierarchy
= true;
5088 /* The following stuff does not apply to the root */
5090 #ifdef CONFIG_MEMCG_KMEM
5091 INIT_LIST_HEAD(&memcg
->kmem_caches
);
5093 root_mem_cgroup
= memcg
;
5097 error
= memcg_online_kmem(memcg
);
5101 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5102 static_branch_inc(&memcg_sockets_enabled_key
);
5106 mem_cgroup_id_remove(memcg
);
5107 mem_cgroup_free(memcg
);
5108 return ERR_PTR(-ENOMEM
);
5111 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5113 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5116 * A memcg must be visible for memcg_expand_shrinker_maps()
5117 * by the time the maps are allocated. So, we allocate maps
5118 * here, when for_each_mem_cgroup() can't skip it.
5120 if (memcg_alloc_shrinker_maps(memcg
)) {
5121 mem_cgroup_id_remove(memcg
);
5125 /* Online state pins memcg ID, memcg ID pins CSS */
5126 refcount_set(&memcg
->id
.ref
, 1);
5131 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5133 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5134 struct mem_cgroup_event
*event
, *tmp
;
5137 * Unregister events and notify userspace.
5138 * Notify userspace about cgroup removing only after rmdir of cgroup
5139 * directory to avoid race between userspace and kernelspace.
5141 spin_lock(&memcg
->event_list_lock
);
5142 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5143 list_del_init(&event
->list
);
5144 schedule_work(&event
->remove
);
5146 spin_unlock(&memcg
->event_list_lock
);
5148 page_counter_set_min(&memcg
->memory
, 0);
5149 page_counter_set_low(&memcg
->memory
, 0);
5151 memcg_offline_kmem(memcg
);
5152 wb_memcg_offline(memcg
);
5154 drain_all_stock(memcg
);
5156 mem_cgroup_id_put(memcg
);
5159 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5161 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5163 invalidate_reclaim_iterators(memcg
);
5166 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5168 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5169 int __maybe_unused i
;
5171 #ifdef CONFIG_CGROUP_WRITEBACK
5172 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5173 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5175 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5176 static_branch_dec(&memcg_sockets_enabled_key
);
5178 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5179 static_branch_dec(&memcg_sockets_enabled_key
);
5181 vmpressure_cleanup(&memcg
->vmpressure
);
5182 cancel_work_sync(&memcg
->high_work
);
5183 mem_cgroup_remove_from_trees(memcg
);
5184 memcg_free_shrinker_maps(memcg
);
5185 memcg_free_kmem(memcg
);
5186 mem_cgroup_free(memcg
);
5190 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5191 * @css: the target css
5193 * Reset the states of the mem_cgroup associated with @css. This is
5194 * invoked when the userland requests disabling on the default hierarchy
5195 * but the memcg is pinned through dependency. The memcg should stop
5196 * applying policies and should revert to the vanilla state as it may be
5197 * made visible again.
5199 * The current implementation only resets the essential configurations.
5200 * This needs to be expanded to cover all the visible parts.
5202 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5204 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5206 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5207 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5208 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
5209 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5210 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5211 page_counter_set_min(&memcg
->memory
, 0);
5212 page_counter_set_low(&memcg
->memory
, 0);
5213 WRITE_ONCE(memcg
->high
, PAGE_COUNTER_MAX
);
5214 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5215 memcg_wb_domain_size_changed(memcg
);
5219 /* Handlers for move charge at task migration. */
5220 static int mem_cgroup_do_precharge(unsigned long count
)
5224 /* Try a single bulk charge without reclaim first, kswapd may wake */
5225 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5227 mc
.precharge
+= count
;
5231 /* Try charges one by one with reclaim, but do not retry */
5233 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5247 enum mc_target_type
{
5254 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5255 unsigned long addr
, pte_t ptent
)
5257 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5259 if (!page
|| !page_mapped(page
))
5261 if (PageAnon(page
)) {
5262 if (!(mc
.flags
& MOVE_ANON
))
5265 if (!(mc
.flags
& MOVE_FILE
))
5268 if (!get_page_unless_zero(page
))
5274 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5275 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5276 pte_t ptent
, swp_entry_t
*entry
)
5278 struct page
*page
= NULL
;
5279 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5281 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
5285 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5286 * a device and because they are not accessible by CPU they are store
5287 * as special swap entry in the CPU page table.
5289 if (is_device_private_entry(ent
)) {
5290 page
= device_private_entry_to_page(ent
);
5292 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5293 * a refcount of 1 when free (unlike normal page)
5295 if (!page_ref_add_unless(page
, 1, 1))
5301 * Because lookup_swap_cache() updates some statistics counter,
5302 * we call find_get_page() with swapper_space directly.
5304 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5305 if (do_memsw_account())
5306 entry
->val
= ent
.val
;
5311 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5312 pte_t ptent
, swp_entry_t
*entry
)
5318 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5319 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5321 struct page
*page
= NULL
;
5322 struct address_space
*mapping
;
5325 if (!vma
->vm_file
) /* anonymous vma */
5327 if (!(mc
.flags
& MOVE_FILE
))
5330 mapping
= vma
->vm_file
->f_mapping
;
5331 pgoff
= linear_page_index(vma
, addr
);
5333 /* page is moved even if it's not RSS of this task(page-faulted). */
5335 /* shmem/tmpfs may report page out on swap: account for that too. */
5336 if (shmem_mapping(mapping
)) {
5337 page
= find_get_entry(mapping
, pgoff
);
5338 if (xa_is_value(page
)) {
5339 swp_entry_t swp
= radix_to_swp_entry(page
);
5340 if (do_memsw_account())
5342 page
= find_get_page(swap_address_space(swp
),
5346 page
= find_get_page(mapping
, pgoff
);
5348 page
= find_get_page(mapping
, pgoff
);
5354 * mem_cgroup_move_account - move account of the page
5356 * @compound: charge the page as compound or small page
5357 * @from: mem_cgroup which the page is moved from.
5358 * @to: mem_cgroup which the page is moved to. @from != @to.
5360 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5362 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5365 static int mem_cgroup_move_account(struct page
*page
,
5367 struct mem_cgroup
*from
,
5368 struct mem_cgroup
*to
)
5370 struct lruvec
*from_vec
, *to_vec
;
5371 struct pglist_data
*pgdat
;
5372 unsigned long flags
;
5373 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5377 VM_BUG_ON(from
== to
);
5378 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5379 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5382 * Prevent mem_cgroup_migrate() from looking at
5383 * page->mem_cgroup of its source page while we change it.
5386 if (!trylock_page(page
))
5390 if (page
->mem_cgroup
!= from
)
5393 anon
= PageAnon(page
);
5395 pgdat
= page_pgdat(page
);
5396 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5397 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5399 spin_lock_irqsave(&from
->move_lock
, flags
);
5401 if (!anon
&& page_mapped(page
)) {
5402 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5403 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5407 * move_lock grabbed above and caller set from->moving_account, so
5408 * mod_memcg_page_state will serialize updates to PageDirty.
5409 * So mapping should be stable for dirty pages.
5411 if (!anon
&& PageDirty(page
)) {
5412 struct address_space
*mapping
= page_mapping(page
);
5414 if (mapping_cap_account_dirty(mapping
)) {
5415 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
, -nr_pages
);
5416 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
, nr_pages
);
5420 if (PageWriteback(page
)) {
5421 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5422 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5426 * It is safe to change page->mem_cgroup here because the page
5427 * is referenced, charged, and isolated - we can't race with
5428 * uncharging, charging, migration, or LRU putback.
5431 /* caller should have done css_get */
5432 page
->mem_cgroup
= to
;
5434 spin_unlock_irqrestore(&from
->move_lock
, flags
);
5438 local_irq_disable();
5439 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
5440 memcg_check_events(to
, page
);
5441 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
5442 memcg_check_events(from
, page
);
5451 * get_mctgt_type - get target type of moving charge
5452 * @vma: the vma the pte to be checked belongs
5453 * @addr: the address corresponding to the pte to be checked
5454 * @ptent: the pte to be checked
5455 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5458 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5459 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5460 * move charge. if @target is not NULL, the page is stored in target->page
5461 * with extra refcnt got(Callers should handle it).
5462 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5463 * target for charge migration. if @target is not NULL, the entry is stored
5465 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5466 * (so ZONE_DEVICE page and thus not on the lru).
5467 * For now we such page is charge like a regular page would be as for all
5468 * intent and purposes it is just special memory taking the place of a
5471 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5473 * Called with pte lock held.
5476 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5477 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5479 struct page
*page
= NULL
;
5480 enum mc_target_type ret
= MC_TARGET_NONE
;
5481 swp_entry_t ent
= { .val
= 0 };
5483 if (pte_present(ptent
))
5484 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5485 else if (is_swap_pte(ptent
))
5486 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5487 else if (pte_none(ptent
))
5488 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5490 if (!page
&& !ent
.val
)
5494 * Do only loose check w/o serialization.
5495 * mem_cgroup_move_account() checks the page is valid or
5496 * not under LRU exclusion.
5498 if (page
->mem_cgroup
== mc
.from
) {
5499 ret
= MC_TARGET_PAGE
;
5500 if (is_device_private_page(page
))
5501 ret
= MC_TARGET_DEVICE
;
5503 target
->page
= page
;
5505 if (!ret
|| !target
)
5509 * There is a swap entry and a page doesn't exist or isn't charged.
5510 * But we cannot move a tail-page in a THP.
5512 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5513 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5514 ret
= MC_TARGET_SWAP
;
5521 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5523 * We don't consider PMD mapped swapping or file mapped pages because THP does
5524 * not support them for now.
5525 * Caller should make sure that pmd_trans_huge(pmd) is true.
5527 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5528 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5530 struct page
*page
= NULL
;
5531 enum mc_target_type ret
= MC_TARGET_NONE
;
5533 if (unlikely(is_swap_pmd(pmd
))) {
5534 VM_BUG_ON(thp_migration_supported() &&
5535 !is_pmd_migration_entry(pmd
));
5538 page
= pmd_page(pmd
);
5539 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5540 if (!(mc
.flags
& MOVE_ANON
))
5542 if (page
->mem_cgroup
== mc
.from
) {
5543 ret
= MC_TARGET_PAGE
;
5546 target
->page
= page
;
5552 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5553 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5555 return MC_TARGET_NONE
;
5559 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5560 unsigned long addr
, unsigned long end
,
5561 struct mm_walk
*walk
)
5563 struct vm_area_struct
*vma
= walk
->vma
;
5567 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5570 * Note their can not be MC_TARGET_DEVICE for now as we do not
5571 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5572 * this might change.
5574 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5575 mc
.precharge
+= HPAGE_PMD_NR
;
5580 if (pmd_trans_unstable(pmd
))
5582 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5583 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5584 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5585 mc
.precharge
++; /* increment precharge temporarily */
5586 pte_unmap_unlock(pte
- 1, ptl
);
5592 static const struct mm_walk_ops precharge_walk_ops
= {
5593 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5596 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5598 unsigned long precharge
;
5600 down_read(&mm
->mmap_sem
);
5601 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5602 up_read(&mm
->mmap_sem
);
5604 precharge
= mc
.precharge
;
5610 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5612 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5614 VM_BUG_ON(mc
.moving_task
);
5615 mc
.moving_task
= current
;
5616 return mem_cgroup_do_precharge(precharge
);
5619 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5620 static void __mem_cgroup_clear_mc(void)
5622 struct mem_cgroup
*from
= mc
.from
;
5623 struct mem_cgroup
*to
= mc
.to
;
5625 /* we must uncharge all the leftover precharges from mc.to */
5627 cancel_charge(mc
.to
, mc
.precharge
);
5631 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5632 * we must uncharge here.
5634 if (mc
.moved_charge
) {
5635 cancel_charge(mc
.from
, mc
.moved_charge
);
5636 mc
.moved_charge
= 0;
5638 /* we must fixup refcnts and charges */
5639 if (mc
.moved_swap
) {
5640 /* uncharge swap account from the old cgroup */
5641 if (!mem_cgroup_is_root(mc
.from
))
5642 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5644 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5647 * we charged both to->memory and to->memsw, so we
5648 * should uncharge to->memory.
5650 if (!mem_cgroup_is_root(mc
.to
))
5651 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5653 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5654 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5658 memcg_oom_recover(from
);
5659 memcg_oom_recover(to
);
5660 wake_up_all(&mc
.waitq
);
5663 static void mem_cgroup_clear_mc(void)
5665 struct mm_struct
*mm
= mc
.mm
;
5668 * we must clear moving_task before waking up waiters at the end of
5671 mc
.moving_task
= NULL
;
5672 __mem_cgroup_clear_mc();
5673 spin_lock(&mc
.lock
);
5677 spin_unlock(&mc
.lock
);
5682 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5684 struct cgroup_subsys_state
*css
;
5685 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5686 struct mem_cgroup
*from
;
5687 struct task_struct
*leader
, *p
;
5688 struct mm_struct
*mm
;
5689 unsigned long move_flags
;
5692 /* charge immigration isn't supported on the default hierarchy */
5693 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5697 * Multi-process migrations only happen on the default hierarchy
5698 * where charge immigration is not used. Perform charge
5699 * immigration if @tset contains a leader and whine if there are
5703 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5706 memcg
= mem_cgroup_from_css(css
);
5712 * We are now commited to this value whatever it is. Changes in this
5713 * tunable will only affect upcoming migrations, not the current one.
5714 * So we need to save it, and keep it going.
5716 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5720 from
= mem_cgroup_from_task(p
);
5722 VM_BUG_ON(from
== memcg
);
5724 mm
= get_task_mm(p
);
5727 /* We move charges only when we move a owner of the mm */
5728 if (mm
->owner
== p
) {
5731 VM_BUG_ON(mc
.precharge
);
5732 VM_BUG_ON(mc
.moved_charge
);
5733 VM_BUG_ON(mc
.moved_swap
);
5735 spin_lock(&mc
.lock
);
5739 mc
.flags
= move_flags
;
5740 spin_unlock(&mc
.lock
);
5741 /* We set mc.moving_task later */
5743 ret
= mem_cgroup_precharge_mc(mm
);
5745 mem_cgroup_clear_mc();
5752 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5755 mem_cgroup_clear_mc();
5758 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5759 unsigned long addr
, unsigned long end
,
5760 struct mm_walk
*walk
)
5763 struct vm_area_struct
*vma
= walk
->vma
;
5766 enum mc_target_type target_type
;
5767 union mc_target target
;
5770 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5772 if (mc
.precharge
< HPAGE_PMD_NR
) {
5776 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5777 if (target_type
== MC_TARGET_PAGE
) {
5779 if (!isolate_lru_page(page
)) {
5780 if (!mem_cgroup_move_account(page
, true,
5782 mc
.precharge
-= HPAGE_PMD_NR
;
5783 mc
.moved_charge
+= HPAGE_PMD_NR
;
5785 putback_lru_page(page
);
5788 } else if (target_type
== MC_TARGET_DEVICE
) {
5790 if (!mem_cgroup_move_account(page
, true,
5792 mc
.precharge
-= HPAGE_PMD_NR
;
5793 mc
.moved_charge
+= HPAGE_PMD_NR
;
5801 if (pmd_trans_unstable(pmd
))
5804 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5805 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5806 pte_t ptent
= *(pte
++);
5807 bool device
= false;
5813 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5814 case MC_TARGET_DEVICE
:
5817 case MC_TARGET_PAGE
:
5820 * We can have a part of the split pmd here. Moving it
5821 * can be done but it would be too convoluted so simply
5822 * ignore such a partial THP and keep it in original
5823 * memcg. There should be somebody mapping the head.
5825 if (PageTransCompound(page
))
5827 if (!device
&& isolate_lru_page(page
))
5829 if (!mem_cgroup_move_account(page
, false,
5832 /* we uncharge from mc.from later. */
5836 putback_lru_page(page
);
5837 put
: /* get_mctgt_type() gets the page */
5840 case MC_TARGET_SWAP
:
5842 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5844 /* we fixup refcnts and charges later. */
5852 pte_unmap_unlock(pte
- 1, ptl
);
5857 * We have consumed all precharges we got in can_attach().
5858 * We try charge one by one, but don't do any additional
5859 * charges to mc.to if we have failed in charge once in attach()
5862 ret
= mem_cgroup_do_precharge(1);
5870 static const struct mm_walk_ops charge_walk_ops
= {
5871 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5874 static void mem_cgroup_move_charge(void)
5876 lru_add_drain_all();
5878 * Signal lock_page_memcg() to take the memcg's move_lock
5879 * while we're moving its pages to another memcg. Then wait
5880 * for already started RCU-only updates to finish.
5882 atomic_inc(&mc
.from
->moving_account
);
5885 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5887 * Someone who are holding the mmap_sem might be waiting in
5888 * waitq. So we cancel all extra charges, wake up all waiters,
5889 * and retry. Because we cancel precharges, we might not be able
5890 * to move enough charges, but moving charge is a best-effort
5891 * feature anyway, so it wouldn't be a big problem.
5893 __mem_cgroup_clear_mc();
5898 * When we have consumed all precharges and failed in doing
5899 * additional charge, the page walk just aborts.
5901 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
5904 up_read(&mc
.mm
->mmap_sem
);
5905 atomic_dec(&mc
.from
->moving_account
);
5908 static void mem_cgroup_move_task(void)
5911 mem_cgroup_move_charge();
5912 mem_cgroup_clear_mc();
5915 #else /* !CONFIG_MMU */
5916 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5920 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5923 static void mem_cgroup_move_task(void)
5929 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5930 * to verify whether we're attached to the default hierarchy on each mount
5933 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5936 * use_hierarchy is forced on the default hierarchy. cgroup core
5937 * guarantees that @root doesn't have any children, so turning it
5938 * on for the root memcg is enough.
5940 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5941 root_mem_cgroup
->use_hierarchy
= true;
5943 root_mem_cgroup
->use_hierarchy
= false;
5946 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
5948 if (value
== PAGE_COUNTER_MAX
)
5949 seq_puts(m
, "max\n");
5951 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
5956 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5959 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5961 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5964 static int memory_min_show(struct seq_file
*m
, void *v
)
5966 return seq_puts_memcg_tunable(m
,
5967 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
5970 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
5971 char *buf
, size_t nbytes
, loff_t off
)
5973 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5977 buf
= strstrip(buf
);
5978 err
= page_counter_memparse(buf
, "max", &min
);
5982 page_counter_set_min(&memcg
->memory
, min
);
5987 static int memory_low_show(struct seq_file
*m
, void *v
)
5989 return seq_puts_memcg_tunable(m
,
5990 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
5993 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5994 char *buf
, size_t nbytes
, loff_t off
)
5996 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6000 buf
= strstrip(buf
);
6001 err
= page_counter_memparse(buf
, "max", &low
);
6005 page_counter_set_low(&memcg
->memory
, low
);
6010 static int memory_high_show(struct seq_file
*m
, void *v
)
6012 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
6015 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6016 char *buf
, size_t nbytes
, loff_t off
)
6018 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6019 unsigned int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
6020 bool drained
= false;
6024 buf
= strstrip(buf
);
6025 err
= page_counter_memparse(buf
, "max", &high
);
6029 WRITE_ONCE(memcg
->high
, high
);
6032 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6033 unsigned long reclaimed
;
6035 if (nr_pages
<= high
)
6038 if (signal_pending(current
))
6042 drain_all_stock(memcg
);
6047 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6050 if (!reclaimed
&& !nr_retries
--)
6057 static int memory_max_show(struct seq_file
*m
, void *v
)
6059 return seq_puts_memcg_tunable(m
,
6060 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6063 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6064 char *buf
, size_t nbytes
, loff_t off
)
6066 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6067 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
6068 bool drained
= false;
6072 buf
= strstrip(buf
);
6073 err
= page_counter_memparse(buf
, "max", &max
);
6077 xchg(&memcg
->memory
.max
, max
);
6080 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6082 if (nr_pages
<= max
)
6085 if (signal_pending(current
))
6089 drain_all_stock(memcg
);
6095 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6101 memcg_memory_event(memcg
, MEMCG_OOM
);
6102 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6106 memcg_wb_domain_size_changed(memcg
);
6110 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6112 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6113 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6114 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6115 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6116 seq_printf(m
, "oom_kill %lu\n",
6117 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6120 static int memory_events_show(struct seq_file
*m
, void *v
)
6122 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6124 __memory_events_show(m
, memcg
->memory_events
);
6128 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6130 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6132 __memory_events_show(m
, memcg
->memory_events_local
);
6136 static int memory_stat_show(struct seq_file
*m
, void *v
)
6138 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6141 buf
= memory_stat_format(memcg
);
6149 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6151 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6153 seq_printf(m
, "%d\n", memcg
->oom_group
);
6158 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6159 char *buf
, size_t nbytes
, loff_t off
)
6161 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6164 buf
= strstrip(buf
);
6168 ret
= kstrtoint(buf
, 0, &oom_group
);
6172 if (oom_group
!= 0 && oom_group
!= 1)
6175 memcg
->oom_group
= oom_group
;
6180 static struct cftype memory_files
[] = {
6183 .flags
= CFTYPE_NOT_ON_ROOT
,
6184 .read_u64
= memory_current_read
,
6188 .flags
= CFTYPE_NOT_ON_ROOT
,
6189 .seq_show
= memory_min_show
,
6190 .write
= memory_min_write
,
6194 .flags
= CFTYPE_NOT_ON_ROOT
,
6195 .seq_show
= memory_low_show
,
6196 .write
= memory_low_write
,
6200 .flags
= CFTYPE_NOT_ON_ROOT
,
6201 .seq_show
= memory_high_show
,
6202 .write
= memory_high_write
,
6206 .flags
= CFTYPE_NOT_ON_ROOT
,
6207 .seq_show
= memory_max_show
,
6208 .write
= memory_max_write
,
6212 .flags
= CFTYPE_NOT_ON_ROOT
,
6213 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6214 .seq_show
= memory_events_show
,
6217 .name
= "events.local",
6218 .flags
= CFTYPE_NOT_ON_ROOT
,
6219 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6220 .seq_show
= memory_events_local_show
,
6224 .flags
= CFTYPE_NOT_ON_ROOT
,
6225 .seq_show
= memory_stat_show
,
6228 .name
= "oom.group",
6229 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6230 .seq_show
= memory_oom_group_show
,
6231 .write
= memory_oom_group_write
,
6236 struct cgroup_subsys memory_cgrp_subsys
= {
6237 .css_alloc
= mem_cgroup_css_alloc
,
6238 .css_online
= mem_cgroup_css_online
,
6239 .css_offline
= mem_cgroup_css_offline
,
6240 .css_released
= mem_cgroup_css_released
,
6241 .css_free
= mem_cgroup_css_free
,
6242 .css_reset
= mem_cgroup_css_reset
,
6243 .can_attach
= mem_cgroup_can_attach
,
6244 .cancel_attach
= mem_cgroup_cancel_attach
,
6245 .post_attach
= mem_cgroup_move_task
,
6246 .bind
= mem_cgroup_bind
,
6247 .dfl_cftypes
= memory_files
,
6248 .legacy_cftypes
= mem_cgroup_legacy_files
,
6253 * This function calculates an individual cgroup's effective
6254 * protection which is derived from its own memory.min/low, its
6255 * parent's and siblings' settings, as well as the actual memory
6256 * distribution in the tree.
6258 * The following rules apply to the effective protection values:
6260 * 1. At the first level of reclaim, effective protection is equal to
6261 * the declared protection in memory.min and memory.low.
6263 * 2. To enable safe delegation of the protection configuration, at
6264 * subsequent levels the effective protection is capped to the
6265 * parent's effective protection.
6267 * 3. To make complex and dynamic subtrees easier to configure, the
6268 * user is allowed to overcommit the declared protection at a given
6269 * level. If that is the case, the parent's effective protection is
6270 * distributed to the children in proportion to how much protection
6271 * they have declared and how much of it they are utilizing.
6273 * This makes distribution proportional, but also work-conserving:
6274 * if one cgroup claims much more protection than it uses memory,
6275 * the unused remainder is available to its siblings.
6277 * 4. Conversely, when the declared protection is undercommitted at a
6278 * given level, the distribution of the larger parental protection
6279 * budget is NOT proportional. A cgroup's protection from a sibling
6280 * is capped to its own memory.min/low setting.
6282 * 5. However, to allow protecting recursive subtrees from each other
6283 * without having to declare each individual cgroup's fixed share
6284 * of the ancestor's claim to protection, any unutilized -
6285 * "floating" - protection from up the tree is distributed in
6286 * proportion to each cgroup's *usage*. This makes the protection
6287 * neutral wrt sibling cgroups and lets them compete freely over
6288 * the shared parental protection budget, but it protects the
6289 * subtree as a whole from neighboring subtrees.
6291 * Note that 4. and 5. are not in conflict: 4. is about protecting
6292 * against immediate siblings whereas 5. is about protecting against
6293 * neighboring subtrees.
6295 static unsigned long effective_protection(unsigned long usage
,
6296 unsigned long parent_usage
,
6297 unsigned long setting
,
6298 unsigned long parent_effective
,
6299 unsigned long siblings_protected
)
6301 unsigned long protected;
6304 protected = min(usage
, setting
);
6306 * If all cgroups at this level combined claim and use more
6307 * protection then what the parent affords them, distribute
6308 * shares in proportion to utilization.
6310 * We are using actual utilization rather than the statically
6311 * claimed protection in order to be work-conserving: claimed
6312 * but unused protection is available to siblings that would
6313 * otherwise get a smaller chunk than what they claimed.
6315 if (siblings_protected
> parent_effective
)
6316 return protected * parent_effective
/ siblings_protected
;
6319 * Ok, utilized protection of all children is within what the
6320 * parent affords them, so we know whatever this child claims
6321 * and utilizes is effectively protected.
6323 * If there is unprotected usage beyond this value, reclaim
6324 * will apply pressure in proportion to that amount.
6326 * If there is unutilized protection, the cgroup will be fully
6327 * shielded from reclaim, but we do return a smaller value for
6328 * protection than what the group could enjoy in theory. This
6329 * is okay. With the overcommit distribution above, effective
6330 * protection is always dependent on how memory is actually
6331 * consumed among the siblings anyway.
6336 * If the children aren't claiming (all of) the protection
6337 * afforded to them by the parent, distribute the remainder in
6338 * proportion to the (unprotected) memory of each cgroup. That
6339 * way, cgroups that aren't explicitly prioritized wrt each
6340 * other compete freely over the allowance, but they are
6341 * collectively protected from neighboring trees.
6343 * We're using unprotected memory for the weight so that if
6344 * some cgroups DO claim explicit protection, we don't protect
6345 * the same bytes twice.
6347 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6350 if (parent_effective
> siblings_protected
&& usage
> protected) {
6351 unsigned long unclaimed
;
6353 unclaimed
= parent_effective
- siblings_protected
;
6354 unclaimed
*= usage
- protected;
6355 unclaimed
/= parent_usage
- siblings_protected
;
6364 * mem_cgroup_protected - check if memory consumption is in the normal range
6365 * @root: the top ancestor of the sub-tree being checked
6366 * @memcg: the memory cgroup to check
6368 * WARNING: This function is not stateless! It can only be used as part
6369 * of a top-down tree iteration, not for isolated queries.
6371 * Returns one of the following:
6372 * MEMCG_PROT_NONE: cgroup memory is not protected
6373 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6374 * an unprotected supply of reclaimable memory from other cgroups.
6375 * MEMCG_PROT_MIN: cgroup memory is protected
6377 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
6378 struct mem_cgroup
*memcg
)
6380 unsigned long usage
, parent_usage
;
6381 struct mem_cgroup
*parent
;
6383 if (mem_cgroup_disabled())
6384 return MEMCG_PROT_NONE
;
6387 root
= root_mem_cgroup
;
6389 return MEMCG_PROT_NONE
;
6391 usage
= page_counter_read(&memcg
->memory
);
6393 return MEMCG_PROT_NONE
;
6395 parent
= parent_mem_cgroup(memcg
);
6396 /* No parent means a non-hierarchical mode on v1 memcg */
6398 return MEMCG_PROT_NONE
;
6400 if (parent
== root
) {
6401 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6402 memcg
->memory
.elow
= memcg
->memory
.low
;
6406 parent_usage
= page_counter_read(&parent
->memory
);
6408 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6409 READ_ONCE(memcg
->memory
.min
),
6410 READ_ONCE(parent
->memory
.emin
),
6411 atomic_long_read(&parent
->memory
.children_min_usage
)));
6413 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6414 memcg
->memory
.low
, READ_ONCE(parent
->memory
.elow
),
6415 atomic_long_read(&parent
->memory
.children_low_usage
)));
6418 if (usage
<= memcg
->memory
.emin
)
6419 return MEMCG_PROT_MIN
;
6420 else if (usage
<= memcg
->memory
.elow
)
6421 return MEMCG_PROT_LOW
;
6423 return MEMCG_PROT_NONE
;
6427 * mem_cgroup_try_charge - try charging a page
6428 * @page: page to charge
6429 * @mm: mm context of the victim
6430 * @gfp_mask: reclaim mode
6431 * @memcgp: charged memcg return
6432 * @compound: charge the page as compound or small page
6434 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6435 * pages according to @gfp_mask if necessary.
6437 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6438 * Otherwise, an error code is returned.
6440 * After page->mapping has been set up, the caller must finalize the
6441 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6442 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6444 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6445 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6448 struct mem_cgroup
*memcg
= NULL
;
6449 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6452 if (mem_cgroup_disabled())
6455 if (PageSwapCache(page
)) {
6457 * Every swap fault against a single page tries to charge the
6458 * page, bail as early as possible. shmem_unuse() encounters
6459 * already charged pages, too. The USED bit is protected by
6460 * the page lock, which serializes swap cache removal, which
6461 * in turn serializes uncharging.
6463 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6464 if (compound_head(page
)->mem_cgroup
)
6467 if (do_swap_account
) {
6468 swp_entry_t ent
= { .val
= page_private(page
), };
6469 unsigned short id
= lookup_swap_cgroup_id(ent
);
6472 memcg
= mem_cgroup_from_id(id
);
6473 if (memcg
&& !css_tryget_online(&memcg
->css
))
6480 memcg
= get_mem_cgroup_from_mm(mm
);
6482 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6484 css_put(&memcg
->css
);
6490 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6491 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6494 struct mem_cgroup
*memcg
;
6497 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6499 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6504 * mem_cgroup_commit_charge - commit a page charge
6505 * @page: page to charge
6506 * @memcg: memcg to charge the page to
6507 * @lrucare: page might be on LRU already
6508 * @compound: charge the page as compound or small page
6510 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6511 * after page->mapping has been set up. This must happen atomically
6512 * as part of the page instantiation, i.e. under the page table lock
6513 * for anonymous pages, under the page lock for page and swap cache.
6515 * In addition, the page must not be on the LRU during the commit, to
6516 * prevent racing with task migration. If it might be, use @lrucare.
6518 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6520 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6521 bool lrucare
, bool compound
)
6523 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6525 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6526 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6528 if (mem_cgroup_disabled())
6531 * Swap faults will attempt to charge the same page multiple
6532 * times. But reuse_swap_page() might have removed the page
6533 * from swapcache already, so we can't check PageSwapCache().
6538 commit_charge(page
, memcg
, lrucare
);
6540 local_irq_disable();
6541 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6542 memcg_check_events(memcg
, page
);
6545 if (do_memsw_account() && PageSwapCache(page
)) {
6546 swp_entry_t entry
= { .val
= page_private(page
) };
6548 * The swap entry might not get freed for a long time,
6549 * let's not wait for it. The page already received a
6550 * memory+swap charge, drop the swap entry duplicate.
6552 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6557 * mem_cgroup_cancel_charge - cancel a page charge
6558 * @page: page to charge
6559 * @memcg: memcg to charge the page to
6560 * @compound: charge the page as compound or small page
6562 * Cancel a charge transaction started by mem_cgroup_try_charge().
6564 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6567 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6569 if (mem_cgroup_disabled())
6572 * Swap faults will attempt to charge the same page multiple
6573 * times. But reuse_swap_page() might have removed the page
6574 * from swapcache already, so we can't check PageSwapCache().
6579 cancel_charge(memcg
, nr_pages
);
6582 struct uncharge_gather
{
6583 struct mem_cgroup
*memcg
;
6584 unsigned long pgpgout
;
6585 unsigned long nr_anon
;
6586 unsigned long nr_file
;
6587 unsigned long nr_kmem
;
6588 unsigned long nr_huge
;
6589 unsigned long nr_shmem
;
6590 struct page
*dummy_page
;
6593 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6595 memset(ug
, 0, sizeof(*ug
));
6598 static void uncharge_batch(const struct uncharge_gather
*ug
)
6600 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6601 unsigned long flags
;
6603 if (!mem_cgroup_is_root(ug
->memcg
)) {
6604 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6605 if (do_memsw_account())
6606 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6607 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6608 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6609 memcg_oom_recover(ug
->memcg
);
6612 local_irq_save(flags
);
6613 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6614 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6615 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6616 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6617 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6618 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6619 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6620 local_irq_restore(flags
);
6622 if (!mem_cgroup_is_root(ug
->memcg
))
6623 css_put_many(&ug
->memcg
->css
, nr_pages
);
6626 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6628 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6629 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6630 !PageHWPoison(page
) , page
);
6632 if (!page
->mem_cgroup
)
6636 * Nobody should be changing or seriously looking at
6637 * page->mem_cgroup at this point, we have fully
6638 * exclusive access to the page.
6641 if (ug
->memcg
!= page
->mem_cgroup
) {
6644 uncharge_gather_clear(ug
);
6646 ug
->memcg
= page
->mem_cgroup
;
6649 if (!PageKmemcg(page
)) {
6650 unsigned int nr_pages
= 1;
6652 if (PageTransHuge(page
)) {
6653 nr_pages
= compound_nr(page
);
6654 ug
->nr_huge
+= nr_pages
;
6657 ug
->nr_anon
+= nr_pages
;
6659 ug
->nr_file
+= nr_pages
;
6660 if (PageSwapBacked(page
))
6661 ug
->nr_shmem
+= nr_pages
;
6665 ug
->nr_kmem
+= compound_nr(page
);
6666 __ClearPageKmemcg(page
);
6669 ug
->dummy_page
= page
;
6670 page
->mem_cgroup
= NULL
;
6673 static void uncharge_list(struct list_head
*page_list
)
6675 struct uncharge_gather ug
;
6676 struct list_head
*next
;
6678 uncharge_gather_clear(&ug
);
6681 * Note that the list can be a single page->lru; hence the
6682 * do-while loop instead of a simple list_for_each_entry().
6684 next
= page_list
->next
;
6688 page
= list_entry(next
, struct page
, lru
);
6689 next
= page
->lru
.next
;
6691 uncharge_page(page
, &ug
);
6692 } while (next
!= page_list
);
6695 uncharge_batch(&ug
);
6699 * mem_cgroup_uncharge - uncharge a page
6700 * @page: page to uncharge
6702 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6703 * mem_cgroup_commit_charge().
6705 void mem_cgroup_uncharge(struct page
*page
)
6707 struct uncharge_gather ug
;
6709 if (mem_cgroup_disabled())
6712 /* Don't touch page->lru of any random page, pre-check: */
6713 if (!page
->mem_cgroup
)
6716 uncharge_gather_clear(&ug
);
6717 uncharge_page(page
, &ug
);
6718 uncharge_batch(&ug
);
6722 * mem_cgroup_uncharge_list - uncharge a list of page
6723 * @page_list: list of pages to uncharge
6725 * Uncharge a list of pages previously charged with
6726 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6728 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6730 if (mem_cgroup_disabled())
6733 if (!list_empty(page_list
))
6734 uncharge_list(page_list
);
6738 * mem_cgroup_migrate - charge a page's replacement
6739 * @oldpage: currently circulating page
6740 * @newpage: replacement page
6742 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6743 * be uncharged upon free.
6745 * Both pages must be locked, @newpage->mapping must be set up.
6747 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6749 struct mem_cgroup
*memcg
;
6750 unsigned int nr_pages
;
6751 unsigned long flags
;
6753 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6754 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6755 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6756 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6759 if (mem_cgroup_disabled())
6762 /* Page cache replacement: new page already charged? */
6763 if (newpage
->mem_cgroup
)
6766 /* Swapcache readahead pages can get replaced before being charged */
6767 memcg
= oldpage
->mem_cgroup
;
6771 /* Force-charge the new page. The old one will be freed soon */
6772 nr_pages
= hpage_nr_pages(newpage
);
6774 page_counter_charge(&memcg
->memory
, nr_pages
);
6775 if (do_memsw_account())
6776 page_counter_charge(&memcg
->memsw
, nr_pages
);
6777 css_get_many(&memcg
->css
, nr_pages
);
6779 commit_charge(newpage
, memcg
, false);
6781 local_irq_save(flags
);
6782 mem_cgroup_charge_statistics(memcg
, newpage
, PageTransHuge(newpage
),
6784 memcg_check_events(memcg
, newpage
);
6785 local_irq_restore(flags
);
6788 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6789 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6791 void mem_cgroup_sk_alloc(struct sock
*sk
)
6793 struct mem_cgroup
*memcg
;
6795 if (!mem_cgroup_sockets_enabled
)
6798 /* Do not associate the sock with unrelated interrupted task's memcg. */
6803 memcg
= mem_cgroup_from_task(current
);
6804 if (memcg
== root_mem_cgroup
)
6806 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6808 if (css_tryget(&memcg
->css
))
6809 sk
->sk_memcg
= memcg
;
6814 void mem_cgroup_sk_free(struct sock
*sk
)
6817 css_put(&sk
->sk_memcg
->css
);
6821 * mem_cgroup_charge_skmem - charge socket memory
6822 * @memcg: memcg to charge
6823 * @nr_pages: number of pages to charge
6825 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6826 * @memcg's configured limit, %false if the charge had to be forced.
6828 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6830 gfp_t gfp_mask
= GFP_KERNEL
;
6832 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6833 struct page_counter
*fail
;
6835 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6836 memcg
->tcpmem_pressure
= 0;
6839 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6840 memcg
->tcpmem_pressure
= 1;
6844 /* Don't block in the packet receive path */
6846 gfp_mask
= GFP_NOWAIT
;
6848 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6850 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6853 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6858 * mem_cgroup_uncharge_skmem - uncharge socket memory
6859 * @memcg: memcg to uncharge
6860 * @nr_pages: number of pages to uncharge
6862 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6864 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6865 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6869 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6871 refill_stock(memcg
, nr_pages
);
6874 static int __init
cgroup_memory(char *s
)
6878 while ((token
= strsep(&s
, ",")) != NULL
) {
6881 if (!strcmp(token
, "nosocket"))
6882 cgroup_memory_nosocket
= true;
6883 if (!strcmp(token
, "nokmem"))
6884 cgroup_memory_nokmem
= true;
6888 __setup("cgroup.memory=", cgroup_memory
);
6891 * subsys_initcall() for memory controller.
6893 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6894 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6895 * basically everything that doesn't depend on a specific mem_cgroup structure
6896 * should be initialized from here.
6898 static int __init
mem_cgroup_init(void)
6902 #ifdef CONFIG_MEMCG_KMEM
6904 * Kmem cache creation is mostly done with the slab_mutex held,
6905 * so use a workqueue with limited concurrency to avoid stalling
6906 * all worker threads in case lots of cgroups are created and
6907 * destroyed simultaneously.
6909 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6910 BUG_ON(!memcg_kmem_cache_wq
);
6913 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6914 memcg_hotplug_cpu_dead
);
6916 for_each_possible_cpu(cpu
)
6917 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6920 for_each_node(node
) {
6921 struct mem_cgroup_tree_per_node
*rtpn
;
6923 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6924 node_online(node
) ? node
: NUMA_NO_NODE
);
6926 rtpn
->rb_root
= RB_ROOT
;
6927 rtpn
->rb_rightmost
= NULL
;
6928 spin_lock_init(&rtpn
->lock
);
6929 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6934 subsys_initcall(mem_cgroup_init
);
6936 #ifdef CONFIG_MEMCG_SWAP
6937 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6939 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6941 * The root cgroup cannot be destroyed, so it's refcount must
6944 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6948 memcg
= parent_mem_cgroup(memcg
);
6950 memcg
= root_mem_cgroup
;
6956 * mem_cgroup_swapout - transfer a memsw charge to swap
6957 * @page: page whose memsw charge to transfer
6958 * @entry: swap entry to move the charge to
6960 * Transfer the memsw charge of @page to @entry.
6962 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6964 struct mem_cgroup
*memcg
, *swap_memcg
;
6965 unsigned int nr_entries
;
6966 unsigned short oldid
;
6968 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6969 VM_BUG_ON_PAGE(page_count(page
), page
);
6971 if (!do_memsw_account())
6974 memcg
= page
->mem_cgroup
;
6976 /* Readahead page, never charged */
6981 * In case the memcg owning these pages has been offlined and doesn't
6982 * have an ID allocated to it anymore, charge the closest online
6983 * ancestor for the swap instead and transfer the memory+swap charge.
6985 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6986 nr_entries
= hpage_nr_pages(page
);
6987 /* Get references for the tail pages, too */
6989 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6990 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6992 VM_BUG_ON_PAGE(oldid
, page
);
6993 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6995 page
->mem_cgroup
= NULL
;
6997 if (!mem_cgroup_is_root(memcg
))
6998 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7000 if (memcg
!= swap_memcg
) {
7001 if (!mem_cgroup_is_root(swap_memcg
))
7002 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7003 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7007 * Interrupts should be disabled here because the caller holds the
7008 * i_pages lock which is taken with interrupts-off. It is
7009 * important here to have the interrupts disabled because it is the
7010 * only synchronisation we have for updating the per-CPU variables.
7012 VM_BUG_ON(!irqs_disabled());
7013 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
7015 memcg_check_events(memcg
, page
);
7017 if (!mem_cgroup_is_root(memcg
))
7018 css_put_many(&memcg
->css
, nr_entries
);
7022 * mem_cgroup_try_charge_swap - try charging swap space for a page
7023 * @page: page being added to swap
7024 * @entry: swap entry to charge
7026 * Try to charge @page's memcg for the swap space at @entry.
7028 * Returns 0 on success, -ENOMEM on failure.
7030 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7032 unsigned int nr_pages
= hpage_nr_pages(page
);
7033 struct page_counter
*counter
;
7034 struct mem_cgroup
*memcg
;
7035 unsigned short oldid
;
7037 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
7040 memcg
= page
->mem_cgroup
;
7042 /* Readahead page, never charged */
7047 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7051 memcg
= mem_cgroup_id_get_online(memcg
);
7053 if (!mem_cgroup_is_root(memcg
) &&
7054 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7055 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7056 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7057 mem_cgroup_id_put(memcg
);
7061 /* Get references for the tail pages, too */
7063 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7064 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7065 VM_BUG_ON_PAGE(oldid
, page
);
7066 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7072 * mem_cgroup_uncharge_swap - uncharge swap space
7073 * @entry: swap entry to uncharge
7074 * @nr_pages: the amount of swap space to uncharge
7076 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7078 struct mem_cgroup
*memcg
;
7081 if (!do_swap_account
)
7084 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7086 memcg
= mem_cgroup_from_id(id
);
7088 if (!mem_cgroup_is_root(memcg
)) {
7089 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7090 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7092 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7094 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7095 mem_cgroup_id_put_many(memcg
, nr_pages
);
7100 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7102 long nr_swap_pages
= get_nr_swap_pages();
7104 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7105 return nr_swap_pages
;
7106 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7107 nr_swap_pages
= min_t(long, nr_swap_pages
,
7108 READ_ONCE(memcg
->swap
.max
) -
7109 page_counter_read(&memcg
->swap
));
7110 return nr_swap_pages
;
7113 bool mem_cgroup_swap_full(struct page
*page
)
7115 struct mem_cgroup
*memcg
;
7117 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7121 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7124 memcg
= page
->mem_cgroup
;
7128 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7129 if (page_counter_read(&memcg
->swap
) * 2 >=
7130 READ_ONCE(memcg
->swap
.max
))
7136 /* for remember boot option*/
7137 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7138 static int really_do_swap_account __initdata
= 1;
7140 static int really_do_swap_account __initdata
;
7143 static int __init
enable_swap_account(char *s
)
7145 if (!strcmp(s
, "1"))
7146 really_do_swap_account
= 1;
7147 else if (!strcmp(s
, "0"))
7148 really_do_swap_account
= 0;
7151 __setup("swapaccount=", enable_swap_account
);
7153 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7156 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7158 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7161 static int swap_max_show(struct seq_file
*m
, void *v
)
7163 return seq_puts_memcg_tunable(m
,
7164 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7167 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7168 char *buf
, size_t nbytes
, loff_t off
)
7170 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7174 buf
= strstrip(buf
);
7175 err
= page_counter_memparse(buf
, "max", &max
);
7179 xchg(&memcg
->swap
.max
, max
);
7184 static int swap_events_show(struct seq_file
*m
, void *v
)
7186 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7188 seq_printf(m
, "max %lu\n",
7189 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7190 seq_printf(m
, "fail %lu\n",
7191 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7196 static struct cftype swap_files
[] = {
7198 .name
= "swap.current",
7199 .flags
= CFTYPE_NOT_ON_ROOT
,
7200 .read_u64
= swap_current_read
,
7204 .flags
= CFTYPE_NOT_ON_ROOT
,
7205 .seq_show
= swap_max_show
,
7206 .write
= swap_max_write
,
7209 .name
= "swap.events",
7210 .flags
= CFTYPE_NOT_ON_ROOT
,
7211 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7212 .seq_show
= swap_events_show
,
7217 static struct cftype memsw_cgroup_files
[] = {
7219 .name
= "memsw.usage_in_bytes",
7220 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7221 .read_u64
= mem_cgroup_read_u64
,
7224 .name
= "memsw.max_usage_in_bytes",
7225 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7226 .write
= mem_cgroup_reset
,
7227 .read_u64
= mem_cgroup_read_u64
,
7230 .name
= "memsw.limit_in_bytes",
7231 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7232 .write
= mem_cgroup_write
,
7233 .read_u64
= mem_cgroup_read_u64
,
7236 .name
= "memsw.failcnt",
7237 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7238 .write
= mem_cgroup_reset
,
7239 .read_u64
= mem_cgroup_read_u64
,
7241 { }, /* terminate */
7244 static int __init
mem_cgroup_swap_init(void)
7246 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7247 do_swap_account
= 1;
7248 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
7250 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
7251 memsw_cgroup_files
));
7255 subsys_initcall(mem_cgroup_swap_init
);
7257 #endif /* CONFIG_MEMCG_SWAP */