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(sizeof(*new) + size
, GFP_KERNEL
);
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(sizeof(*map
) + size
, GFP_KERNEL
);
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_online(&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 struct page
*page
= virt_to_head_page(p
);
763 pg_data_t
*pgdat
= page_pgdat(page
);
764 struct mem_cgroup
*memcg
;
765 struct lruvec
*lruvec
;
768 memcg
= memcg_from_slab_page(page
);
770 /* Untracked pages have no memcg, no lruvec. Update only the node */
771 if (!memcg
|| memcg
== root_mem_cgroup
) {
772 __mod_node_page_state(pgdat
, idx
, val
);
774 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
775 __mod_lruvec_state(lruvec
, idx
, val
);
781 * __count_memcg_events - account VM events in a cgroup
782 * @memcg: the memory cgroup
783 * @idx: the event item
784 * @count: the number of events that occured
786 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
791 if (mem_cgroup_disabled())
794 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
795 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
796 struct mem_cgroup
*mi
;
799 * Batch local counters to keep them in sync with
800 * the hierarchical ones.
802 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
803 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
804 atomic_long_add(x
, &mi
->vmevents
[idx
]);
807 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
810 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
812 return atomic_long_read(&memcg
->vmevents
[event
]);
815 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
820 for_each_possible_cpu(cpu
)
821 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
825 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
827 bool compound
, int nr_pages
)
830 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
831 * counted as CACHE even if it's on ANON LRU.
834 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
836 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
837 if (PageSwapBacked(page
))
838 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
842 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
843 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
846 /* pagein of a big page is an event. So, ignore page size */
848 __count_memcg_events(memcg
, PGPGIN
, 1);
850 __count_memcg_events(memcg
, PGPGOUT
, 1);
851 nr_pages
= -nr_pages
; /* for event */
854 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
857 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
858 enum mem_cgroup_events_target target
)
860 unsigned long val
, next
;
862 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
863 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
864 /* from time_after() in jiffies.h */
865 if ((long)(next
- val
) < 0) {
867 case MEM_CGROUP_TARGET_THRESH
:
868 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
870 case MEM_CGROUP_TARGET_SOFTLIMIT
:
871 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
876 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
883 * Check events in order.
886 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
888 /* threshold event is triggered in finer grain than soft limit */
889 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
890 MEM_CGROUP_TARGET_THRESH
))) {
893 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
894 MEM_CGROUP_TARGET_SOFTLIMIT
);
895 mem_cgroup_threshold(memcg
);
896 if (unlikely(do_softlimit
))
897 mem_cgroup_update_tree(memcg
, page
);
901 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
904 * mm_update_next_owner() may clear mm->owner to NULL
905 * if it races with swapoff, page migration, etc.
906 * So this can be called with p == NULL.
911 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
913 EXPORT_SYMBOL(mem_cgroup_from_task
);
916 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
917 * @mm: mm from which memcg should be extracted. It can be NULL.
919 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
920 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
923 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
925 struct mem_cgroup
*memcg
;
927 if (mem_cgroup_disabled())
933 * Page cache insertions can happen withou an
934 * actual mm context, e.g. during disk probing
935 * on boot, loopback IO, acct() writes etc.
938 memcg
= root_mem_cgroup
;
940 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
941 if (unlikely(!memcg
))
942 memcg
= root_mem_cgroup
;
944 } while (!css_tryget(&memcg
->css
));
948 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
951 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
952 * @page: page from which memcg should be extracted.
954 * Obtain a reference on page->memcg and returns it if successful. Otherwise
955 * root_mem_cgroup is returned.
957 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
959 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
961 if (mem_cgroup_disabled())
965 if (!memcg
|| !css_tryget_online(&memcg
->css
))
966 memcg
= root_mem_cgroup
;
970 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
973 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
975 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
977 if (unlikely(current
->active_memcg
)) {
978 struct mem_cgroup
*memcg
= root_mem_cgroup
;
981 if (css_tryget_online(¤t
->active_memcg
->css
))
982 memcg
= current
->active_memcg
;
986 return get_mem_cgroup_from_mm(current
->mm
);
990 * mem_cgroup_iter - iterate over memory cgroup hierarchy
991 * @root: hierarchy root
992 * @prev: previously returned memcg, NULL on first invocation
993 * @reclaim: cookie for shared reclaim walks, NULL for full walks
995 * Returns references to children of the hierarchy below @root, or
996 * @root itself, or %NULL after a full round-trip.
998 * Caller must pass the return value in @prev on subsequent
999 * invocations for reference counting, or use mem_cgroup_iter_break()
1000 * to cancel a hierarchy walk before the round-trip is complete.
1002 * Reclaimers can specify a node and a priority level in @reclaim to
1003 * divide up the memcgs in the hierarchy among all concurrent
1004 * reclaimers operating on the same node and priority.
1006 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1007 struct mem_cgroup
*prev
,
1008 struct mem_cgroup_reclaim_cookie
*reclaim
)
1010 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1011 struct cgroup_subsys_state
*css
= NULL
;
1012 struct mem_cgroup
*memcg
= NULL
;
1013 struct mem_cgroup
*pos
= NULL
;
1015 if (mem_cgroup_disabled())
1019 root
= root_mem_cgroup
;
1021 if (prev
&& !reclaim
)
1024 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1033 struct mem_cgroup_per_node
*mz
;
1035 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1038 if (prev
&& reclaim
->generation
!= iter
->generation
)
1042 pos
= READ_ONCE(iter
->position
);
1043 if (!pos
|| css_tryget(&pos
->css
))
1046 * css reference reached zero, so iter->position will
1047 * be cleared by ->css_released. However, we should not
1048 * rely on this happening soon, because ->css_released
1049 * is called from a work queue, and by busy-waiting we
1050 * might block it. So we clear iter->position right
1053 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1061 css
= css_next_descendant_pre(css
, &root
->css
);
1064 * Reclaimers share the hierarchy walk, and a
1065 * new one might jump in right at the end of
1066 * the hierarchy - make sure they see at least
1067 * one group and restart from the beginning.
1075 * Verify the css and acquire a reference. The root
1076 * is provided by the caller, so we know it's alive
1077 * and kicking, and don't take an extra reference.
1079 memcg
= mem_cgroup_from_css(css
);
1081 if (css
== &root
->css
)
1084 if (css_tryget(css
))
1092 * The position could have already been updated by a competing
1093 * thread, so check that the value hasn't changed since we read
1094 * it to avoid reclaiming from the same cgroup twice.
1096 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1104 reclaim
->generation
= iter
->generation
;
1110 if (prev
&& prev
!= root
)
1111 css_put(&prev
->css
);
1117 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1118 * @root: hierarchy root
1119 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1121 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1122 struct mem_cgroup
*prev
)
1125 root
= root_mem_cgroup
;
1126 if (prev
&& prev
!= root
)
1127 css_put(&prev
->css
);
1130 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1131 struct mem_cgroup
*dead_memcg
)
1133 struct mem_cgroup_reclaim_iter
*iter
;
1134 struct mem_cgroup_per_node
*mz
;
1137 for_each_node(nid
) {
1138 mz
= mem_cgroup_nodeinfo(from
, nid
);
1140 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1144 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1146 struct mem_cgroup
*memcg
= dead_memcg
;
1147 struct mem_cgroup
*last
;
1150 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1152 } while ((memcg
= parent_mem_cgroup(memcg
)));
1155 * When cgruop1 non-hierarchy mode is used,
1156 * parent_mem_cgroup() does not walk all the way up to the
1157 * cgroup root (root_mem_cgroup). So we have to handle
1158 * dead_memcg from cgroup root separately.
1160 if (last
!= root_mem_cgroup
)
1161 __invalidate_reclaim_iterators(root_mem_cgroup
,
1166 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1167 * @memcg: hierarchy root
1168 * @fn: function to call for each task
1169 * @arg: argument passed to @fn
1171 * This function iterates over tasks attached to @memcg or to any of its
1172 * descendants and calls @fn for each task. If @fn returns a non-zero
1173 * value, the function breaks the iteration loop and returns the value.
1174 * Otherwise, it will iterate over all tasks and return 0.
1176 * This function must not be called for the root memory cgroup.
1178 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1179 int (*fn
)(struct task_struct
*, void *), void *arg
)
1181 struct mem_cgroup
*iter
;
1184 BUG_ON(memcg
== root_mem_cgroup
);
1186 for_each_mem_cgroup_tree(iter
, memcg
) {
1187 struct css_task_iter it
;
1188 struct task_struct
*task
;
1190 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1191 while (!ret
&& (task
= css_task_iter_next(&it
)))
1192 ret
= fn(task
, arg
);
1193 css_task_iter_end(&it
);
1195 mem_cgroup_iter_break(memcg
, iter
);
1203 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1205 * @pgdat: pgdat of the page
1207 * This function is only safe when following the LRU page isolation
1208 * and putback protocol: the LRU lock must be held, and the page must
1209 * either be PageLRU() or the caller must have isolated/allocated it.
1211 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1213 struct mem_cgroup_per_node
*mz
;
1214 struct mem_cgroup
*memcg
;
1215 struct lruvec
*lruvec
;
1217 if (mem_cgroup_disabled()) {
1218 lruvec
= &pgdat
->__lruvec
;
1222 memcg
= page
->mem_cgroup
;
1224 * Swapcache readahead pages are added to the LRU - and
1225 * possibly migrated - before they are charged.
1228 memcg
= root_mem_cgroup
;
1230 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1231 lruvec
= &mz
->lruvec
;
1234 * Since a node can be onlined after the mem_cgroup was created,
1235 * we have to be prepared to initialize lruvec->zone here;
1236 * and if offlined then reonlined, we need to reinitialize it.
1238 if (unlikely(lruvec
->pgdat
!= pgdat
))
1239 lruvec
->pgdat
= pgdat
;
1244 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1245 * @lruvec: mem_cgroup per zone lru vector
1246 * @lru: index of lru list the page is sitting on
1247 * @zid: zone id of the accounted pages
1248 * @nr_pages: positive when adding or negative when removing
1250 * This function must be called under lru_lock, just before a page is added
1251 * to or just after a page is removed from an lru list (that ordering being
1252 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1254 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1255 int zid
, int nr_pages
)
1257 struct mem_cgroup_per_node
*mz
;
1258 unsigned long *lru_size
;
1261 if (mem_cgroup_disabled())
1264 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1265 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1268 *lru_size
+= nr_pages
;
1271 if (WARN_ONCE(size
< 0,
1272 "%s(%p, %d, %d): lru_size %ld\n",
1273 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1279 *lru_size
+= nr_pages
;
1283 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1284 * @memcg: the memory cgroup
1286 * Returns the maximum amount of memory @mem can be charged with, in
1289 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1291 unsigned long margin
= 0;
1292 unsigned long count
;
1293 unsigned long limit
;
1295 count
= page_counter_read(&memcg
->memory
);
1296 limit
= READ_ONCE(memcg
->memory
.max
);
1298 margin
= limit
- count
;
1300 if (do_memsw_account()) {
1301 count
= page_counter_read(&memcg
->memsw
);
1302 limit
= READ_ONCE(memcg
->memsw
.max
);
1304 margin
= min(margin
, limit
- count
);
1313 * A routine for checking "mem" is under move_account() or not.
1315 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1316 * moving cgroups. This is for waiting at high-memory pressure
1319 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1321 struct mem_cgroup
*from
;
1322 struct mem_cgroup
*to
;
1325 * Unlike task_move routines, we access mc.to, mc.from not under
1326 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1328 spin_lock(&mc
.lock
);
1334 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1335 mem_cgroup_is_descendant(to
, memcg
);
1337 spin_unlock(&mc
.lock
);
1341 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1343 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1344 if (mem_cgroup_under_move(memcg
)) {
1346 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1347 /* moving charge context might have finished. */
1350 finish_wait(&mc
.waitq
, &wait
);
1357 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1362 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1367 * Provide statistics on the state of the memory subsystem as
1368 * well as cumulative event counters that show past behavior.
1370 * This list is ordered following a combination of these gradients:
1371 * 1) generic big picture -> specifics and details
1372 * 2) reflecting userspace activity -> reflecting kernel heuristics
1374 * Current memory state:
1377 seq_buf_printf(&s
, "anon %llu\n",
1378 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) *
1380 seq_buf_printf(&s
, "file %llu\n",
1381 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) *
1383 seq_buf_printf(&s
, "kernel_stack %llu\n",
1384 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) *
1386 seq_buf_printf(&s
, "slab %llu\n",
1387 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
1388 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
1390 seq_buf_printf(&s
, "sock %llu\n",
1391 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) *
1394 seq_buf_printf(&s
, "shmem %llu\n",
1395 (u64
)memcg_page_state(memcg
, NR_SHMEM
) *
1397 seq_buf_printf(&s
, "file_mapped %llu\n",
1398 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) *
1400 seq_buf_printf(&s
, "file_dirty %llu\n",
1401 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) *
1403 seq_buf_printf(&s
, "file_writeback %llu\n",
1404 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) *
1408 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1409 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1410 * arse because it requires migrating the work out of rmap to a place
1411 * where the page->mem_cgroup is set up and stable.
1413 seq_buf_printf(&s
, "anon_thp %llu\n",
1414 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) *
1417 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1418 seq_buf_printf(&s
, "%s %llu\n", lru_list_name(i
),
1419 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
1422 seq_buf_printf(&s
, "slab_reclaimable %llu\n",
1423 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
1425 seq_buf_printf(&s
, "slab_unreclaimable %llu\n",
1426 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
1429 /* Accumulated memory events */
1431 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1432 memcg_events(memcg
, PGFAULT
));
1433 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1434 memcg_events(memcg
, PGMAJFAULT
));
1436 seq_buf_printf(&s
, "workingset_refault %lu\n",
1437 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
1438 seq_buf_printf(&s
, "workingset_activate %lu\n",
1439 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
1440 seq_buf_printf(&s
, "workingset_nodereclaim %lu\n",
1441 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
1443 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1444 memcg_events(memcg
, PGREFILL
));
1445 seq_buf_printf(&s
, "pgscan %lu\n",
1446 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1447 memcg_events(memcg
, PGSCAN_DIRECT
));
1448 seq_buf_printf(&s
, "pgsteal %lu\n",
1449 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1450 memcg_events(memcg
, PGSTEAL_DIRECT
));
1451 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1452 memcg_events(memcg
, PGACTIVATE
));
1453 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1454 memcg_events(memcg
, PGDEACTIVATE
));
1455 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1456 memcg_events(memcg
, PGLAZYFREE
));
1457 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1458 memcg_events(memcg
, PGLAZYFREED
));
1460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1461 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1462 memcg_events(memcg
, THP_FAULT_ALLOC
));
1463 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1464 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1465 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1467 /* The above should easily fit into one page */
1468 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1473 #define K(x) ((x) << (PAGE_SHIFT-10))
1475 * mem_cgroup_print_oom_context: Print OOM information relevant to
1476 * memory controller.
1477 * @memcg: The memory cgroup that went over limit
1478 * @p: Task that is going to be killed
1480 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1483 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1488 pr_cont(",oom_memcg=");
1489 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1491 pr_cont(",global_oom");
1493 pr_cont(",task_memcg=");
1494 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1500 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1501 * memory controller.
1502 * @memcg: The memory cgroup that went over limit
1504 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1508 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1509 K((u64
)page_counter_read(&memcg
->memory
)),
1510 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1511 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1512 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1513 K((u64
)page_counter_read(&memcg
->swap
)),
1514 K((u64
)memcg
->swap
.max
), memcg
->swap
.failcnt
);
1516 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1517 K((u64
)page_counter_read(&memcg
->memsw
)),
1518 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1519 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1520 K((u64
)page_counter_read(&memcg
->kmem
)),
1521 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1524 pr_info("Memory cgroup stats for ");
1525 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1527 buf
= memory_stat_format(memcg
);
1535 * Return the memory (and swap, if configured) limit for a memcg.
1537 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1541 max
= memcg
->memory
.max
;
1542 if (mem_cgroup_swappiness(memcg
)) {
1543 unsigned long memsw_max
;
1544 unsigned long swap_max
;
1546 memsw_max
= memcg
->memsw
.max
;
1547 swap_max
= memcg
->swap
.max
;
1548 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1549 max
= min(max
+ swap_max
, memsw_max
);
1554 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1556 return page_counter_read(&memcg
->memory
);
1559 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1562 struct oom_control oc
= {
1566 .gfp_mask
= gfp_mask
,
1571 if (mutex_lock_killable(&oom_lock
))
1574 * A few threads which were not waiting at mutex_lock_killable() can
1575 * fail to bail out. Therefore, check again after holding oom_lock.
1577 ret
= should_force_charge() || out_of_memory(&oc
);
1578 mutex_unlock(&oom_lock
);
1582 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1585 unsigned long *total_scanned
)
1587 struct mem_cgroup
*victim
= NULL
;
1590 unsigned long excess
;
1591 unsigned long nr_scanned
;
1592 struct mem_cgroup_reclaim_cookie reclaim
= {
1596 excess
= soft_limit_excess(root_memcg
);
1599 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1604 * If we have not been able to reclaim
1605 * anything, it might because there are
1606 * no reclaimable pages under this hierarchy
1611 * We want to do more targeted reclaim.
1612 * excess >> 2 is not to excessive so as to
1613 * reclaim too much, nor too less that we keep
1614 * coming back to reclaim from this cgroup
1616 if (total
>= (excess
>> 2) ||
1617 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1622 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1623 pgdat
, &nr_scanned
);
1624 *total_scanned
+= nr_scanned
;
1625 if (!soft_limit_excess(root_memcg
))
1628 mem_cgroup_iter_break(root_memcg
, victim
);
1632 #ifdef CONFIG_LOCKDEP
1633 static struct lockdep_map memcg_oom_lock_dep_map
= {
1634 .name
= "memcg_oom_lock",
1638 static DEFINE_SPINLOCK(memcg_oom_lock
);
1641 * Check OOM-Killer is already running under our hierarchy.
1642 * If someone is running, return false.
1644 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1646 struct mem_cgroup
*iter
, *failed
= NULL
;
1648 spin_lock(&memcg_oom_lock
);
1650 for_each_mem_cgroup_tree(iter
, memcg
) {
1651 if (iter
->oom_lock
) {
1653 * this subtree of our hierarchy is already locked
1654 * so we cannot give a lock.
1657 mem_cgroup_iter_break(memcg
, iter
);
1660 iter
->oom_lock
= true;
1665 * OK, we failed to lock the whole subtree so we have
1666 * to clean up what we set up to the failing subtree
1668 for_each_mem_cgroup_tree(iter
, memcg
) {
1669 if (iter
== failed
) {
1670 mem_cgroup_iter_break(memcg
, iter
);
1673 iter
->oom_lock
= false;
1676 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1678 spin_unlock(&memcg_oom_lock
);
1683 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1685 struct mem_cgroup
*iter
;
1687 spin_lock(&memcg_oom_lock
);
1688 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1689 for_each_mem_cgroup_tree(iter
, memcg
)
1690 iter
->oom_lock
= false;
1691 spin_unlock(&memcg_oom_lock
);
1694 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1696 struct mem_cgroup
*iter
;
1698 spin_lock(&memcg_oom_lock
);
1699 for_each_mem_cgroup_tree(iter
, memcg
)
1701 spin_unlock(&memcg_oom_lock
);
1704 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1706 struct mem_cgroup
*iter
;
1709 * When a new child is created while the hierarchy is under oom,
1710 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1712 spin_lock(&memcg_oom_lock
);
1713 for_each_mem_cgroup_tree(iter
, memcg
)
1714 if (iter
->under_oom
> 0)
1716 spin_unlock(&memcg_oom_lock
);
1719 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1721 struct oom_wait_info
{
1722 struct mem_cgroup
*memcg
;
1723 wait_queue_entry_t wait
;
1726 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1727 unsigned mode
, int sync
, void *arg
)
1729 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1730 struct mem_cgroup
*oom_wait_memcg
;
1731 struct oom_wait_info
*oom_wait_info
;
1733 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1734 oom_wait_memcg
= oom_wait_info
->memcg
;
1736 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1737 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1739 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1742 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1745 * For the following lockless ->under_oom test, the only required
1746 * guarantee is that it must see the state asserted by an OOM when
1747 * this function is called as a result of userland actions
1748 * triggered by the notification of the OOM. This is trivially
1749 * achieved by invoking mem_cgroup_mark_under_oom() before
1750 * triggering notification.
1752 if (memcg
&& memcg
->under_oom
)
1753 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1763 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1765 enum oom_status ret
;
1768 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1771 memcg_memory_event(memcg
, MEMCG_OOM
);
1774 * We are in the middle of the charge context here, so we
1775 * don't want to block when potentially sitting on a callstack
1776 * that holds all kinds of filesystem and mm locks.
1778 * cgroup1 allows disabling the OOM killer and waiting for outside
1779 * handling until the charge can succeed; remember the context and put
1780 * the task to sleep at the end of the page fault when all locks are
1783 * On the other hand, in-kernel OOM killer allows for an async victim
1784 * memory reclaim (oom_reaper) and that means that we are not solely
1785 * relying on the oom victim to make a forward progress and we can
1786 * invoke the oom killer here.
1788 * Please note that mem_cgroup_out_of_memory might fail to find a
1789 * victim and then we have to bail out from the charge path.
1791 if (memcg
->oom_kill_disable
) {
1792 if (!current
->in_user_fault
)
1794 css_get(&memcg
->css
);
1795 current
->memcg_in_oom
= memcg
;
1796 current
->memcg_oom_gfp_mask
= mask
;
1797 current
->memcg_oom_order
= order
;
1802 mem_cgroup_mark_under_oom(memcg
);
1804 locked
= mem_cgroup_oom_trylock(memcg
);
1807 mem_cgroup_oom_notify(memcg
);
1809 mem_cgroup_unmark_under_oom(memcg
);
1810 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1816 mem_cgroup_oom_unlock(memcg
);
1822 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1823 * @handle: actually kill/wait or just clean up the OOM state
1825 * This has to be called at the end of a page fault if the memcg OOM
1826 * handler was enabled.
1828 * Memcg supports userspace OOM handling where failed allocations must
1829 * sleep on a waitqueue until the userspace task resolves the
1830 * situation. Sleeping directly in the charge context with all kinds
1831 * of locks held is not a good idea, instead we remember an OOM state
1832 * in the task and mem_cgroup_oom_synchronize() has to be called at
1833 * the end of the page fault to complete the OOM handling.
1835 * Returns %true if an ongoing memcg OOM situation was detected and
1836 * completed, %false otherwise.
1838 bool mem_cgroup_oom_synchronize(bool handle
)
1840 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1841 struct oom_wait_info owait
;
1844 /* OOM is global, do not handle */
1851 owait
.memcg
= memcg
;
1852 owait
.wait
.flags
= 0;
1853 owait
.wait
.func
= memcg_oom_wake_function
;
1854 owait
.wait
.private = current
;
1855 INIT_LIST_HEAD(&owait
.wait
.entry
);
1857 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1858 mem_cgroup_mark_under_oom(memcg
);
1860 locked
= mem_cgroup_oom_trylock(memcg
);
1863 mem_cgroup_oom_notify(memcg
);
1865 if (locked
&& !memcg
->oom_kill_disable
) {
1866 mem_cgroup_unmark_under_oom(memcg
);
1867 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1868 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1869 current
->memcg_oom_order
);
1872 mem_cgroup_unmark_under_oom(memcg
);
1873 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1877 mem_cgroup_oom_unlock(memcg
);
1879 * There is no guarantee that an OOM-lock contender
1880 * sees the wakeups triggered by the OOM kill
1881 * uncharges. Wake any sleepers explicitely.
1883 memcg_oom_recover(memcg
);
1886 current
->memcg_in_oom
= NULL
;
1887 css_put(&memcg
->css
);
1892 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1893 * @victim: task to be killed by the OOM killer
1894 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1896 * Returns a pointer to a memory cgroup, which has to be cleaned up
1897 * by killing all belonging OOM-killable tasks.
1899 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1901 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1902 struct mem_cgroup
*oom_domain
)
1904 struct mem_cgroup
*oom_group
= NULL
;
1905 struct mem_cgroup
*memcg
;
1907 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1911 oom_domain
= root_mem_cgroup
;
1915 memcg
= mem_cgroup_from_task(victim
);
1916 if (memcg
== root_mem_cgroup
)
1920 * Traverse the memory cgroup hierarchy from the victim task's
1921 * cgroup up to the OOMing cgroup (or root) to find the
1922 * highest-level memory cgroup with oom.group set.
1924 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1925 if (memcg
->oom_group
)
1928 if (memcg
== oom_domain
)
1933 css_get(&oom_group
->css
);
1940 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1942 pr_info("Tasks in ");
1943 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1944 pr_cont(" are going to be killed due to memory.oom.group set\n");
1948 * lock_page_memcg - lock a page->mem_cgroup binding
1951 * This function protects unlocked LRU pages from being moved to
1954 * It ensures lifetime of the returned memcg. Caller is responsible
1955 * for the lifetime of the page; __unlock_page_memcg() is available
1956 * when @page might get freed inside the locked section.
1958 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1960 struct mem_cgroup
*memcg
;
1961 unsigned long flags
;
1964 * The RCU lock is held throughout the transaction. The fast
1965 * path can get away without acquiring the memcg->move_lock
1966 * because page moving starts with an RCU grace period.
1968 * The RCU lock also protects the memcg from being freed when
1969 * the page state that is going to change is the only thing
1970 * preventing the page itself from being freed. E.g. writeback
1971 * doesn't hold a page reference and relies on PG_writeback to
1972 * keep off truncation, migration and so forth.
1976 if (mem_cgroup_disabled())
1979 memcg
= page
->mem_cgroup
;
1980 if (unlikely(!memcg
))
1983 if (atomic_read(&memcg
->moving_account
) <= 0)
1986 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1987 if (memcg
!= page
->mem_cgroup
) {
1988 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1993 * When charge migration first begins, we can have locked and
1994 * unlocked page stat updates happening concurrently. Track
1995 * the task who has the lock for unlock_page_memcg().
1997 memcg
->move_lock_task
= current
;
1998 memcg
->move_lock_flags
= flags
;
2002 EXPORT_SYMBOL(lock_page_memcg
);
2005 * __unlock_page_memcg - unlock and unpin a memcg
2008 * Unlock and unpin a memcg returned by lock_page_memcg().
2010 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2012 if (memcg
&& memcg
->move_lock_task
== current
) {
2013 unsigned long flags
= memcg
->move_lock_flags
;
2015 memcg
->move_lock_task
= NULL
;
2016 memcg
->move_lock_flags
= 0;
2018 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2025 * unlock_page_memcg - unlock a page->mem_cgroup binding
2028 void unlock_page_memcg(struct page
*page
)
2030 __unlock_page_memcg(page
->mem_cgroup
);
2032 EXPORT_SYMBOL(unlock_page_memcg
);
2034 struct memcg_stock_pcp
{
2035 struct mem_cgroup
*cached
; /* this never be root cgroup */
2036 unsigned int nr_pages
;
2037 struct work_struct work
;
2038 unsigned long flags
;
2039 #define FLUSHING_CACHED_CHARGE 0
2041 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2042 static DEFINE_MUTEX(percpu_charge_mutex
);
2045 * consume_stock: Try to consume stocked charge on this cpu.
2046 * @memcg: memcg to consume from.
2047 * @nr_pages: how many pages to charge.
2049 * The charges will only happen if @memcg matches the current cpu's memcg
2050 * stock, and at least @nr_pages are available in that stock. Failure to
2051 * service an allocation will refill the stock.
2053 * returns true if successful, false otherwise.
2055 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2057 struct memcg_stock_pcp
*stock
;
2058 unsigned long flags
;
2061 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2064 local_irq_save(flags
);
2066 stock
= this_cpu_ptr(&memcg_stock
);
2067 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2068 stock
->nr_pages
-= nr_pages
;
2072 local_irq_restore(flags
);
2078 * Returns stocks cached in percpu and reset cached information.
2080 static void drain_stock(struct memcg_stock_pcp
*stock
)
2082 struct mem_cgroup
*old
= stock
->cached
;
2084 if (stock
->nr_pages
) {
2085 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2086 if (do_memsw_account())
2087 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2088 css_put_many(&old
->css
, stock
->nr_pages
);
2089 stock
->nr_pages
= 0;
2091 stock
->cached
= NULL
;
2094 static void drain_local_stock(struct work_struct
*dummy
)
2096 struct memcg_stock_pcp
*stock
;
2097 unsigned long flags
;
2100 * The only protection from memory hotplug vs. drain_stock races is
2101 * that we always operate on local CPU stock here with IRQ disabled
2103 local_irq_save(flags
);
2105 stock
= this_cpu_ptr(&memcg_stock
);
2107 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2109 local_irq_restore(flags
);
2113 * Cache charges(val) to local per_cpu area.
2114 * This will be consumed by consume_stock() function, later.
2116 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2118 struct memcg_stock_pcp
*stock
;
2119 unsigned long flags
;
2121 local_irq_save(flags
);
2123 stock
= this_cpu_ptr(&memcg_stock
);
2124 if (stock
->cached
!= memcg
) { /* reset if necessary */
2126 stock
->cached
= memcg
;
2128 stock
->nr_pages
+= nr_pages
;
2130 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2133 local_irq_restore(flags
);
2137 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2138 * of the hierarchy under it.
2140 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2144 /* If someone's already draining, avoid adding running more workers. */
2145 if (!mutex_trylock(&percpu_charge_mutex
))
2148 * Notify other cpus that system-wide "drain" is running
2149 * We do not care about races with the cpu hotplug because cpu down
2150 * as well as workers from this path always operate on the local
2151 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2154 for_each_online_cpu(cpu
) {
2155 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2156 struct mem_cgroup
*memcg
;
2160 memcg
= stock
->cached
;
2161 if (memcg
&& stock
->nr_pages
&&
2162 mem_cgroup_is_descendant(memcg
, root_memcg
))
2167 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2169 drain_local_stock(&stock
->work
);
2171 schedule_work_on(cpu
, &stock
->work
);
2175 mutex_unlock(&percpu_charge_mutex
);
2178 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2180 struct memcg_stock_pcp
*stock
;
2181 struct mem_cgroup
*memcg
, *mi
;
2183 stock
= &per_cpu(memcg_stock
, cpu
);
2186 for_each_mem_cgroup(memcg
) {
2189 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2193 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2195 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2196 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2198 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2201 for_each_node(nid
) {
2202 struct mem_cgroup_per_node
*pn
;
2204 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2205 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2208 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2209 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2213 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2216 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2218 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2219 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2226 static void reclaim_high(struct mem_cgroup
*memcg
,
2227 unsigned int nr_pages
,
2231 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2233 memcg_memory_event(memcg
, MEMCG_HIGH
);
2234 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2235 } while ((memcg
= parent_mem_cgroup(memcg
)));
2238 static void high_work_func(struct work_struct
*work
)
2240 struct mem_cgroup
*memcg
;
2242 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2243 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2247 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2248 * enough to still cause a significant slowdown in most cases, while still
2249 * allowing diagnostics and tracing to proceed without becoming stuck.
2251 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2254 * When calculating the delay, we use these either side of the exponentiation to
2255 * maintain precision and scale to a reasonable number of jiffies (see the table
2258 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2259 * overage ratio to a delay.
2260 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2261 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2262 * to produce a reasonable delay curve.
2264 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2265 * reasonable delay curve compared to precision-adjusted overage, not
2266 * penalising heavily at first, but still making sure that growth beyond the
2267 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2268 * example, with a high of 100 megabytes:
2270 * +-------+------------------------+
2271 * | usage | time to allocate in ms |
2272 * +-------+------------------------+
2294 * +-------+------------------------+
2296 #define MEMCG_DELAY_PRECISION_SHIFT 20
2297 #define MEMCG_DELAY_SCALING_SHIFT 14
2300 * Scheduled by try_charge() to be executed from the userland return path
2301 * and reclaims memory over the high limit.
2303 void mem_cgroup_handle_over_high(void)
2305 unsigned long usage
, high
, clamped_high
;
2306 unsigned long pflags
;
2307 unsigned long penalty_jiffies
, overage
;
2308 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2309 struct mem_cgroup
*memcg
;
2311 if (likely(!nr_pages
))
2314 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2315 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2316 current
->memcg_nr_pages_over_high
= 0;
2319 * memory.high is breached and reclaim is unable to keep up. Throttle
2320 * allocators proactively to slow down excessive growth.
2322 * We use overage compared to memory.high to calculate the number of
2323 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2324 * fairly lenient on small overages, and increasingly harsh when the
2325 * memcg in question makes it clear that it has no intention of stopping
2326 * its crazy behaviour, so we exponentially increase the delay based on
2330 usage
= page_counter_read(&memcg
->memory
);
2331 high
= READ_ONCE(memcg
->high
);
2337 * Prevent division by 0 in overage calculation by acting as if it was a
2338 * threshold of 1 page
2340 clamped_high
= max(high
, 1UL);
2342 overage
= div_u64((u64
)(usage
- high
) << MEMCG_DELAY_PRECISION_SHIFT
,
2345 penalty_jiffies
= ((u64
)overage
* overage
* HZ
)
2346 >> (MEMCG_DELAY_PRECISION_SHIFT
+ MEMCG_DELAY_SCALING_SHIFT
);
2349 * Factor in the task's own contribution to the overage, such that four
2350 * N-sized allocations are throttled approximately the same as one
2351 * 4N-sized allocation.
2353 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2354 * larger the current charge patch is than that.
2356 penalty_jiffies
= penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2359 * Clamp the max delay per usermode return so as to still keep the
2360 * application moving forwards and also permit diagnostics, albeit
2363 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2366 * Don't sleep if the amount of jiffies this memcg owes us is so low
2367 * that it's not even worth doing, in an attempt to be nice to those who
2368 * go only a small amount over their memory.high value and maybe haven't
2369 * been aggressively reclaimed enough yet.
2371 if (penalty_jiffies
<= HZ
/ 100)
2375 * If we exit early, we're guaranteed to die (since
2376 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2377 * need to account for any ill-begotten jiffies to pay them off later.
2379 psi_memstall_enter(&pflags
);
2380 schedule_timeout_killable(penalty_jiffies
);
2381 psi_memstall_leave(&pflags
);
2384 css_put(&memcg
->css
);
2387 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2388 unsigned int nr_pages
)
2390 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2391 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2392 struct mem_cgroup
*mem_over_limit
;
2393 struct page_counter
*counter
;
2394 unsigned long nr_reclaimed
;
2395 bool may_swap
= true;
2396 bool drained
= false;
2397 enum oom_status oom_status
;
2399 if (mem_cgroup_is_root(memcg
))
2402 if (consume_stock(memcg
, nr_pages
))
2405 if (!do_memsw_account() ||
2406 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2407 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2409 if (do_memsw_account())
2410 page_counter_uncharge(&memcg
->memsw
, batch
);
2411 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2413 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2417 if (batch
> nr_pages
) {
2423 * Memcg doesn't have a dedicated reserve for atomic
2424 * allocations. But like the global atomic pool, we need to
2425 * put the burden of reclaim on regular allocation requests
2426 * and let these go through as privileged allocations.
2428 if (gfp_mask
& __GFP_ATOMIC
)
2432 * Unlike in global OOM situations, memcg is not in a physical
2433 * memory shortage. Allow dying and OOM-killed tasks to
2434 * bypass the last charges so that they can exit quickly and
2435 * free their memory.
2437 if (unlikely(should_force_charge()))
2441 * Prevent unbounded recursion when reclaim operations need to
2442 * allocate memory. This might exceed the limits temporarily,
2443 * but we prefer facilitating memory reclaim and getting back
2444 * under the limit over triggering OOM kills in these cases.
2446 if (unlikely(current
->flags
& PF_MEMALLOC
))
2449 if (unlikely(task_in_memcg_oom(current
)))
2452 if (!gfpflags_allow_blocking(gfp_mask
))
2455 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2457 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2458 gfp_mask
, may_swap
);
2460 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2464 drain_all_stock(mem_over_limit
);
2469 if (gfp_mask
& __GFP_NORETRY
)
2472 * Even though the limit is exceeded at this point, reclaim
2473 * may have been able to free some pages. Retry the charge
2474 * before killing the task.
2476 * Only for regular pages, though: huge pages are rather
2477 * unlikely to succeed so close to the limit, and we fall back
2478 * to regular pages anyway in case of failure.
2480 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2483 * At task move, charge accounts can be doubly counted. So, it's
2484 * better to wait until the end of task_move if something is going on.
2486 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2492 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2495 if (gfp_mask
& __GFP_NOFAIL
)
2498 if (fatal_signal_pending(current
))
2502 * keep retrying as long as the memcg oom killer is able to make
2503 * a forward progress or bypass the charge if the oom killer
2504 * couldn't make any progress.
2506 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2507 get_order(nr_pages
* PAGE_SIZE
));
2508 switch (oom_status
) {
2510 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2518 if (!(gfp_mask
& __GFP_NOFAIL
))
2522 * The allocation either can't fail or will lead to more memory
2523 * being freed very soon. Allow memory usage go over the limit
2524 * temporarily by force charging it.
2526 page_counter_charge(&memcg
->memory
, nr_pages
);
2527 if (do_memsw_account())
2528 page_counter_charge(&memcg
->memsw
, nr_pages
);
2529 css_get_many(&memcg
->css
, nr_pages
);
2534 css_get_many(&memcg
->css
, batch
);
2535 if (batch
> nr_pages
)
2536 refill_stock(memcg
, batch
- nr_pages
);
2539 * If the hierarchy is above the normal consumption range, schedule
2540 * reclaim on returning to userland. We can perform reclaim here
2541 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2542 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2543 * not recorded as it most likely matches current's and won't
2544 * change in the meantime. As high limit is checked again before
2545 * reclaim, the cost of mismatch is negligible.
2548 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2549 /* Don't bother a random interrupted task */
2550 if (in_interrupt()) {
2551 schedule_work(&memcg
->high_work
);
2554 current
->memcg_nr_pages_over_high
+= batch
;
2555 set_notify_resume(current
);
2558 } while ((memcg
= parent_mem_cgroup(memcg
)));
2563 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2565 if (mem_cgroup_is_root(memcg
))
2568 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2569 if (do_memsw_account())
2570 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2572 css_put_many(&memcg
->css
, nr_pages
);
2575 static void lock_page_lru(struct page
*page
, int *isolated
)
2577 pg_data_t
*pgdat
= page_pgdat(page
);
2579 spin_lock_irq(&pgdat
->lru_lock
);
2580 if (PageLRU(page
)) {
2581 struct lruvec
*lruvec
;
2583 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2585 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2591 static void unlock_page_lru(struct page
*page
, int isolated
)
2593 pg_data_t
*pgdat
= page_pgdat(page
);
2596 struct lruvec
*lruvec
;
2598 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2599 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2601 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2603 spin_unlock_irq(&pgdat
->lru_lock
);
2606 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2611 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2614 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2615 * may already be on some other mem_cgroup's LRU. Take care of it.
2618 lock_page_lru(page
, &isolated
);
2621 * Nobody should be changing or seriously looking at
2622 * page->mem_cgroup at this point:
2624 * - the page is uncharged
2626 * - the page is off-LRU
2628 * - an anonymous fault has exclusive page access, except for
2629 * a locked page table
2631 * - a page cache insertion, a swapin fault, or a migration
2632 * have the page locked
2634 page
->mem_cgroup
= memcg
;
2637 unlock_page_lru(page
, isolated
);
2640 #ifdef CONFIG_MEMCG_KMEM
2641 static int memcg_alloc_cache_id(void)
2646 id
= ida_simple_get(&memcg_cache_ida
,
2647 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2651 if (id
< memcg_nr_cache_ids
)
2655 * There's no space for the new id in memcg_caches arrays,
2656 * so we have to grow them.
2658 down_write(&memcg_cache_ids_sem
);
2660 size
= 2 * (id
+ 1);
2661 if (size
< MEMCG_CACHES_MIN_SIZE
)
2662 size
= MEMCG_CACHES_MIN_SIZE
;
2663 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2664 size
= MEMCG_CACHES_MAX_SIZE
;
2666 err
= memcg_update_all_caches(size
);
2668 err
= memcg_update_all_list_lrus(size
);
2670 memcg_nr_cache_ids
= size
;
2672 up_write(&memcg_cache_ids_sem
);
2675 ida_simple_remove(&memcg_cache_ida
, id
);
2681 static void memcg_free_cache_id(int id
)
2683 ida_simple_remove(&memcg_cache_ida
, id
);
2686 struct memcg_kmem_cache_create_work
{
2687 struct mem_cgroup
*memcg
;
2688 struct kmem_cache
*cachep
;
2689 struct work_struct work
;
2692 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2694 struct memcg_kmem_cache_create_work
*cw
=
2695 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2696 struct mem_cgroup
*memcg
= cw
->memcg
;
2697 struct kmem_cache
*cachep
= cw
->cachep
;
2699 memcg_create_kmem_cache(memcg
, cachep
);
2701 css_put(&memcg
->css
);
2706 * Enqueue the creation of a per-memcg kmem_cache.
2708 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2709 struct kmem_cache
*cachep
)
2711 struct memcg_kmem_cache_create_work
*cw
;
2713 if (!css_tryget_online(&memcg
->css
))
2716 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2721 cw
->cachep
= cachep
;
2722 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2724 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2727 static inline bool memcg_kmem_bypass(void)
2729 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2735 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2736 * @cachep: the original global kmem cache
2738 * Return the kmem_cache we're supposed to use for a slab allocation.
2739 * We try to use the current memcg's version of the cache.
2741 * If the cache does not exist yet, if we are the first user of it, we
2742 * create it asynchronously in a workqueue and let the current allocation
2743 * go through with the original cache.
2745 * This function takes a reference to the cache it returns to assure it
2746 * won't get destroyed while we are working with it. Once the caller is
2747 * done with it, memcg_kmem_put_cache() must be called to release the
2750 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2752 struct mem_cgroup
*memcg
;
2753 struct kmem_cache
*memcg_cachep
;
2754 struct memcg_cache_array
*arr
;
2757 VM_BUG_ON(!is_root_cache(cachep
));
2759 if (memcg_kmem_bypass())
2764 if (unlikely(current
->active_memcg
))
2765 memcg
= current
->active_memcg
;
2767 memcg
= mem_cgroup_from_task(current
);
2769 if (!memcg
|| memcg
== root_mem_cgroup
)
2772 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2776 arr
= rcu_dereference(cachep
->memcg_params
.memcg_caches
);
2779 * Make sure we will access the up-to-date value. The code updating
2780 * memcg_caches issues a write barrier to match the data dependency
2781 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2783 memcg_cachep
= READ_ONCE(arr
->entries
[kmemcg_id
]);
2786 * If we are in a safe context (can wait, and not in interrupt
2787 * context), we could be be predictable and return right away.
2788 * This would guarantee that the allocation being performed
2789 * already belongs in the new cache.
2791 * However, there are some clashes that can arrive from locking.
2792 * For instance, because we acquire the slab_mutex while doing
2793 * memcg_create_kmem_cache, this means no further allocation
2794 * could happen with the slab_mutex held. So it's better to
2797 * If the memcg is dying or memcg_cache is about to be released,
2798 * don't bother creating new kmem_caches. Because memcg_cachep
2799 * is ZEROed as the fist step of kmem offlining, we don't need
2800 * percpu_ref_tryget_live() here. css_tryget_online() check in
2801 * memcg_schedule_kmem_cache_create() will prevent us from
2802 * creation of a new kmem_cache.
2804 if (unlikely(!memcg_cachep
))
2805 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2806 else if (percpu_ref_tryget(&memcg_cachep
->memcg_params
.refcnt
))
2807 cachep
= memcg_cachep
;
2814 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2815 * @cachep: the cache returned by memcg_kmem_get_cache
2817 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2819 if (!is_root_cache(cachep
))
2820 percpu_ref_put(&cachep
->memcg_params
.refcnt
);
2824 * __memcg_kmem_charge_memcg: charge a kmem page
2825 * @page: page to charge
2826 * @gfp: reclaim mode
2827 * @order: allocation order
2828 * @memcg: memory cgroup to charge
2830 * Returns 0 on success, an error code on failure.
2832 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2833 struct mem_cgroup
*memcg
)
2835 unsigned int nr_pages
= 1 << order
;
2836 struct page_counter
*counter
;
2839 ret
= try_charge(memcg
, gfp
, nr_pages
);
2843 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2844 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2847 * Enforce __GFP_NOFAIL allocation because callers are not
2848 * prepared to see failures and likely do not have any failure
2851 if (gfp
& __GFP_NOFAIL
) {
2852 page_counter_charge(&memcg
->kmem
, nr_pages
);
2855 cancel_charge(memcg
, nr_pages
);
2862 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2863 * @page: page to charge
2864 * @gfp: reclaim mode
2865 * @order: allocation order
2867 * Returns 0 on success, an error code on failure.
2869 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2871 struct mem_cgroup
*memcg
;
2874 if (memcg_kmem_bypass())
2877 memcg
= get_mem_cgroup_from_current();
2878 if (!mem_cgroup_is_root(memcg
)) {
2879 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2881 page
->mem_cgroup
= memcg
;
2882 __SetPageKmemcg(page
);
2885 css_put(&memcg
->css
);
2890 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2891 * @memcg: memcg to uncharge
2892 * @nr_pages: number of pages to uncharge
2894 void __memcg_kmem_uncharge_memcg(struct mem_cgroup
*memcg
,
2895 unsigned int nr_pages
)
2897 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2898 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2900 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2901 if (do_memsw_account())
2902 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2905 * __memcg_kmem_uncharge: uncharge a kmem page
2906 * @page: page to uncharge
2907 * @order: allocation order
2909 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2911 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2912 unsigned int nr_pages
= 1 << order
;
2917 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2918 __memcg_kmem_uncharge_memcg(memcg
, nr_pages
);
2919 page
->mem_cgroup
= NULL
;
2921 /* slab pages do not have PageKmemcg flag set */
2922 if (PageKmemcg(page
))
2923 __ClearPageKmemcg(page
);
2925 css_put_many(&memcg
->css
, nr_pages
);
2927 #endif /* CONFIG_MEMCG_KMEM */
2929 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2932 * Because tail pages are not marked as "used", set it. We're under
2933 * pgdat->lru_lock and migration entries setup in all page mappings.
2935 void mem_cgroup_split_huge_fixup(struct page
*head
)
2939 if (mem_cgroup_disabled())
2942 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2943 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2945 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
2947 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2949 #ifdef CONFIG_MEMCG_SWAP
2951 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2952 * @entry: swap entry to be moved
2953 * @from: mem_cgroup which the entry is moved from
2954 * @to: mem_cgroup which the entry is moved to
2956 * It succeeds only when the swap_cgroup's record for this entry is the same
2957 * as the mem_cgroup's id of @from.
2959 * Returns 0 on success, -EINVAL on failure.
2961 * The caller must have charged to @to, IOW, called page_counter_charge() about
2962 * both res and memsw, and called css_get().
2964 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2965 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2967 unsigned short old_id
, new_id
;
2969 old_id
= mem_cgroup_id(from
);
2970 new_id
= mem_cgroup_id(to
);
2972 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2973 mod_memcg_state(from
, MEMCG_SWAP
, -1);
2974 mod_memcg_state(to
, MEMCG_SWAP
, 1);
2980 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2981 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2987 static DEFINE_MUTEX(memcg_max_mutex
);
2989 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
2990 unsigned long max
, bool memsw
)
2992 bool enlarge
= false;
2993 bool drained
= false;
2995 bool limits_invariant
;
2996 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
2999 if (signal_pending(current
)) {
3004 mutex_lock(&memcg_max_mutex
);
3006 * Make sure that the new limit (memsw or memory limit) doesn't
3007 * break our basic invariant rule memory.max <= memsw.max.
3009 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
3010 max
<= memcg
->memsw
.max
;
3011 if (!limits_invariant
) {
3012 mutex_unlock(&memcg_max_mutex
);
3016 if (max
> counter
->max
)
3018 ret
= page_counter_set_max(counter
, max
);
3019 mutex_unlock(&memcg_max_mutex
);
3025 drain_all_stock(memcg
);
3030 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3031 GFP_KERNEL
, !memsw
)) {
3037 if (!ret
&& enlarge
)
3038 memcg_oom_recover(memcg
);
3043 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3045 unsigned long *total_scanned
)
3047 unsigned long nr_reclaimed
= 0;
3048 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3049 unsigned long reclaimed
;
3051 struct mem_cgroup_tree_per_node
*mctz
;
3052 unsigned long excess
;
3053 unsigned long nr_scanned
;
3058 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3061 * Do not even bother to check the largest node if the root
3062 * is empty. Do it lockless to prevent lock bouncing. Races
3063 * are acceptable as soft limit is best effort anyway.
3065 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3069 * This loop can run a while, specially if mem_cgroup's continuously
3070 * keep exceeding their soft limit and putting the system under
3077 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3082 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3083 gfp_mask
, &nr_scanned
);
3084 nr_reclaimed
+= reclaimed
;
3085 *total_scanned
+= nr_scanned
;
3086 spin_lock_irq(&mctz
->lock
);
3087 __mem_cgroup_remove_exceeded(mz
, mctz
);
3090 * If we failed to reclaim anything from this memory cgroup
3091 * it is time to move on to the next cgroup
3095 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3097 excess
= soft_limit_excess(mz
->memcg
);
3099 * One school of thought says that we should not add
3100 * back the node to the tree if reclaim returns 0.
3101 * But our reclaim could return 0, simply because due
3102 * to priority we are exposing a smaller subset of
3103 * memory to reclaim from. Consider this as a longer
3106 /* If excess == 0, no tree ops */
3107 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3108 spin_unlock_irq(&mctz
->lock
);
3109 css_put(&mz
->memcg
->css
);
3112 * Could not reclaim anything and there are no more
3113 * mem cgroups to try or we seem to be looping without
3114 * reclaiming anything.
3116 if (!nr_reclaimed
&&
3118 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3120 } while (!nr_reclaimed
);
3122 css_put(&next_mz
->memcg
->css
);
3123 return nr_reclaimed
;
3127 * Test whether @memcg has children, dead or alive. Note that this
3128 * function doesn't care whether @memcg has use_hierarchy enabled and
3129 * returns %true if there are child csses according to the cgroup
3130 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3132 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3137 ret
= css_next_child(NULL
, &memcg
->css
);
3143 * Reclaims as many pages from the given memcg as possible.
3145 * Caller is responsible for holding css reference for memcg.
3147 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3149 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3151 /* we call try-to-free pages for make this cgroup empty */
3152 lru_add_drain_all();
3154 drain_all_stock(memcg
);
3156 /* try to free all pages in this cgroup */
3157 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3160 if (signal_pending(current
))
3163 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3167 /* maybe some writeback is necessary */
3168 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3176 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3177 char *buf
, size_t nbytes
,
3180 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3182 if (mem_cgroup_is_root(memcg
))
3184 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3187 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3190 return mem_cgroup_from_css(css
)->use_hierarchy
;
3193 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3194 struct cftype
*cft
, u64 val
)
3197 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3198 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3200 if (memcg
->use_hierarchy
== val
)
3204 * If parent's use_hierarchy is set, we can't make any modifications
3205 * in the child subtrees. If it is unset, then the change can
3206 * occur, provided the current cgroup has no children.
3208 * For the root cgroup, parent_mem is NULL, we allow value to be
3209 * set if there are no children.
3211 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3212 (val
== 1 || val
== 0)) {
3213 if (!memcg_has_children(memcg
))
3214 memcg
->use_hierarchy
= val
;
3223 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3227 if (mem_cgroup_is_root(memcg
)) {
3228 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3229 memcg_page_state(memcg
, MEMCG_RSS
);
3231 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3234 val
= page_counter_read(&memcg
->memory
);
3236 val
= page_counter_read(&memcg
->memsw
);
3249 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3252 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3253 struct page_counter
*counter
;
3255 switch (MEMFILE_TYPE(cft
->private)) {
3257 counter
= &memcg
->memory
;
3260 counter
= &memcg
->memsw
;
3263 counter
= &memcg
->kmem
;
3266 counter
= &memcg
->tcpmem
;
3272 switch (MEMFILE_ATTR(cft
->private)) {
3274 if (counter
== &memcg
->memory
)
3275 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3276 if (counter
== &memcg
->memsw
)
3277 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3278 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3280 return (u64
)counter
->max
* PAGE_SIZE
;
3282 return (u64
)counter
->watermark
* PAGE_SIZE
;
3284 return counter
->failcnt
;
3285 case RES_SOFT_LIMIT
:
3286 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3292 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3294 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3295 struct mem_cgroup
*mi
;
3298 for_each_online_cpu(cpu
)
3299 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3300 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3302 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3303 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3304 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3306 for_each_node(node
) {
3307 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3308 struct mem_cgroup_per_node
*pi
;
3310 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3313 for_each_online_cpu(cpu
)
3314 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3316 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3318 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3319 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3320 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3324 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3326 unsigned long events
[NR_VM_EVENT_ITEMS
];
3327 struct mem_cgroup
*mi
;
3330 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3333 for_each_online_cpu(cpu
)
3334 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3335 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3338 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3339 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3340 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3343 #ifdef CONFIG_MEMCG_KMEM
3344 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3348 if (cgroup_memory_nokmem
)
3351 BUG_ON(memcg
->kmemcg_id
>= 0);
3352 BUG_ON(memcg
->kmem_state
);
3354 memcg_id
= memcg_alloc_cache_id();
3358 static_branch_inc(&memcg_kmem_enabled_key
);
3360 * A memory cgroup is considered kmem-online as soon as it gets
3361 * kmemcg_id. Setting the id after enabling static branching will
3362 * guarantee no one starts accounting before all call sites are
3365 memcg
->kmemcg_id
= memcg_id
;
3366 memcg
->kmem_state
= KMEM_ONLINE
;
3367 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3372 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3374 struct cgroup_subsys_state
*css
;
3375 struct mem_cgroup
*parent
, *child
;
3378 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3381 * Clear the online state before clearing memcg_caches array
3382 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3383 * guarantees that no cache will be created for this cgroup
3384 * after we are done (see memcg_create_kmem_cache()).
3386 memcg
->kmem_state
= KMEM_ALLOCATED
;
3388 parent
= parent_mem_cgroup(memcg
);
3390 parent
= root_mem_cgroup
;
3393 * Deactivate and reparent kmem_caches.
3395 memcg_deactivate_kmem_caches(memcg
, parent
);
3397 kmemcg_id
= memcg
->kmemcg_id
;
3398 BUG_ON(kmemcg_id
< 0);
3401 * Change kmemcg_id of this cgroup and all its descendants to the
3402 * parent's id, and then move all entries from this cgroup's list_lrus
3403 * to ones of the parent. After we have finished, all list_lrus
3404 * corresponding to this cgroup are guaranteed to remain empty. The
3405 * ordering is imposed by list_lru_node->lock taken by
3406 * memcg_drain_all_list_lrus().
3408 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3409 css_for_each_descendant_pre(css
, &memcg
->css
) {
3410 child
= mem_cgroup_from_css(css
);
3411 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3412 child
->kmemcg_id
= parent
->kmemcg_id
;
3413 if (!memcg
->use_hierarchy
)
3418 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3420 memcg_free_cache_id(kmemcg_id
);
3423 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3425 /* css_alloc() failed, offlining didn't happen */
3426 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3427 memcg_offline_kmem(memcg
);
3429 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3430 WARN_ON(!list_empty(&memcg
->kmem_caches
));
3431 static_branch_dec(&memcg_kmem_enabled_key
);
3435 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3439 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3442 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3445 #endif /* CONFIG_MEMCG_KMEM */
3447 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3452 mutex_lock(&memcg_max_mutex
);
3453 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3454 mutex_unlock(&memcg_max_mutex
);
3458 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3462 mutex_lock(&memcg_max_mutex
);
3464 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3468 if (!memcg
->tcpmem_active
) {
3470 * The active flag needs to be written after the static_key
3471 * update. This is what guarantees that the socket activation
3472 * function is the last one to run. See mem_cgroup_sk_alloc()
3473 * for details, and note that we don't mark any socket as
3474 * belonging to this memcg until that flag is up.
3476 * We need to do this, because static_keys will span multiple
3477 * sites, but we can't control their order. If we mark a socket
3478 * as accounted, but the accounting functions are not patched in
3479 * yet, we'll lose accounting.
3481 * We never race with the readers in mem_cgroup_sk_alloc(),
3482 * because when this value change, the code to process it is not
3485 static_branch_inc(&memcg_sockets_enabled_key
);
3486 memcg
->tcpmem_active
= true;
3489 mutex_unlock(&memcg_max_mutex
);
3494 * The user of this function is...
3497 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3498 char *buf
, size_t nbytes
, loff_t off
)
3500 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3501 unsigned long nr_pages
;
3504 buf
= strstrip(buf
);
3505 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3509 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3511 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3515 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3517 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3520 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3523 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3524 "Please report your usecase to linux-mm@kvack.org if you "
3525 "depend on this functionality.\n");
3526 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3529 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3533 case RES_SOFT_LIMIT
:
3534 memcg
->soft_limit
= nr_pages
;
3538 return ret
?: nbytes
;
3541 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3542 size_t nbytes
, loff_t off
)
3544 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3545 struct page_counter
*counter
;
3547 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3549 counter
= &memcg
->memory
;
3552 counter
= &memcg
->memsw
;
3555 counter
= &memcg
->kmem
;
3558 counter
= &memcg
->tcpmem
;
3564 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3566 page_counter_reset_watermark(counter
);
3569 counter
->failcnt
= 0;
3578 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3581 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3585 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3586 struct cftype
*cft
, u64 val
)
3588 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3590 if (val
& ~MOVE_MASK
)
3594 * No kind of locking is needed in here, because ->can_attach() will
3595 * check this value once in the beginning of the process, and then carry
3596 * on with stale data. This means that changes to this value will only
3597 * affect task migrations starting after the change.
3599 memcg
->move_charge_at_immigrate
= val
;
3603 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3604 struct cftype
*cft
, u64 val
)
3612 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3613 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3614 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3616 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3617 int nid
, unsigned int lru_mask
)
3619 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3620 unsigned long nr
= 0;
3623 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3626 if (!(BIT(lru
) & lru_mask
))
3628 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3633 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3634 unsigned int lru_mask
)
3636 unsigned long nr
= 0;
3640 if (!(BIT(lru
) & lru_mask
))
3642 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3647 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3651 unsigned int lru_mask
;
3654 static const struct numa_stat stats
[] = {
3655 { "total", LRU_ALL
},
3656 { "file", LRU_ALL_FILE
},
3657 { "anon", LRU_ALL_ANON
},
3658 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3660 const struct numa_stat
*stat
;
3663 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3665 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3666 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3667 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3668 for_each_node_state(nid
, N_MEMORY
) {
3669 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3671 seq_printf(m
, " N%d=%lu", nid
, nr
);
3676 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3677 struct mem_cgroup
*iter
;
3680 for_each_mem_cgroup_tree(iter
, memcg
)
3681 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3682 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3683 for_each_node_state(nid
, N_MEMORY
) {
3685 for_each_mem_cgroup_tree(iter
, memcg
)
3686 nr
+= mem_cgroup_node_nr_lru_pages(
3687 iter
, nid
, stat
->lru_mask
);
3688 seq_printf(m
, " N%d=%lu", nid
, nr
);
3695 #endif /* CONFIG_NUMA */
3697 static const unsigned int memcg1_stats
[] = {
3708 static const char *const memcg1_stat_names
[] = {
3719 /* Universal VM events cgroup1 shows, original sort order */
3720 static const unsigned int memcg1_events
[] = {
3727 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3729 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3730 unsigned long memory
, memsw
;
3731 struct mem_cgroup
*mi
;
3734 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3736 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3737 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3739 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3740 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3744 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3745 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
3746 memcg_events_local(memcg
, memcg1_events
[i
]));
3748 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3749 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
3750 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3753 /* Hierarchical information */
3754 memory
= memsw
= PAGE_COUNTER_MAX
;
3755 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3756 memory
= min(memory
, mi
->memory
.max
);
3757 memsw
= min(memsw
, mi
->memsw
.max
);
3759 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3760 (u64
)memory
* PAGE_SIZE
);
3761 if (do_memsw_account())
3762 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3763 (u64
)memsw
* PAGE_SIZE
);
3765 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3766 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3768 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3769 (u64
)memcg_page_state(memcg
, memcg1_stats
[i
]) *
3773 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3774 seq_printf(m
, "total_%s %llu\n",
3775 vm_event_name(memcg1_events
[i
]),
3776 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
3778 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3779 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
3780 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3783 #ifdef CONFIG_DEBUG_VM
3786 struct mem_cgroup_per_node
*mz
;
3787 struct zone_reclaim_stat
*rstat
;
3788 unsigned long recent_rotated
[2] = {0, 0};
3789 unsigned long recent_scanned
[2] = {0, 0};
3791 for_each_online_pgdat(pgdat
) {
3792 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3793 rstat
= &mz
->lruvec
.reclaim_stat
;
3795 recent_rotated
[0] += rstat
->recent_rotated
[0];
3796 recent_rotated
[1] += rstat
->recent_rotated
[1];
3797 recent_scanned
[0] += rstat
->recent_scanned
[0];
3798 recent_scanned
[1] += rstat
->recent_scanned
[1];
3800 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3801 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3802 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3803 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3810 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3813 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3815 return mem_cgroup_swappiness(memcg
);
3818 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3819 struct cftype
*cft
, u64 val
)
3821 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3827 memcg
->swappiness
= val
;
3829 vm_swappiness
= val
;
3834 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3836 struct mem_cgroup_threshold_ary
*t
;
3837 unsigned long usage
;
3842 t
= rcu_dereference(memcg
->thresholds
.primary
);
3844 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3849 usage
= mem_cgroup_usage(memcg
, swap
);
3852 * current_threshold points to threshold just below or equal to usage.
3853 * If it's not true, a threshold was crossed after last
3854 * call of __mem_cgroup_threshold().
3856 i
= t
->current_threshold
;
3859 * Iterate backward over array of thresholds starting from
3860 * current_threshold and check if a threshold is crossed.
3861 * If none of thresholds below usage is crossed, we read
3862 * only one element of the array here.
3864 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3865 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3867 /* i = current_threshold + 1 */
3871 * Iterate forward over array of thresholds starting from
3872 * current_threshold+1 and check if a threshold is crossed.
3873 * If none of thresholds above usage is crossed, we read
3874 * only one element of the array here.
3876 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3877 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3879 /* Update current_threshold */
3880 t
->current_threshold
= i
- 1;
3885 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3888 __mem_cgroup_threshold(memcg
, false);
3889 if (do_memsw_account())
3890 __mem_cgroup_threshold(memcg
, true);
3892 memcg
= parent_mem_cgroup(memcg
);
3896 static int compare_thresholds(const void *a
, const void *b
)
3898 const struct mem_cgroup_threshold
*_a
= a
;
3899 const struct mem_cgroup_threshold
*_b
= b
;
3901 if (_a
->threshold
> _b
->threshold
)
3904 if (_a
->threshold
< _b
->threshold
)
3910 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3912 struct mem_cgroup_eventfd_list
*ev
;
3914 spin_lock(&memcg_oom_lock
);
3916 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3917 eventfd_signal(ev
->eventfd
, 1);
3919 spin_unlock(&memcg_oom_lock
);
3923 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3925 struct mem_cgroup
*iter
;
3927 for_each_mem_cgroup_tree(iter
, memcg
)
3928 mem_cgroup_oom_notify_cb(iter
);
3931 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3932 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3934 struct mem_cgroup_thresholds
*thresholds
;
3935 struct mem_cgroup_threshold_ary
*new;
3936 unsigned long threshold
;
3937 unsigned long usage
;
3940 ret
= page_counter_memparse(args
, "-1", &threshold
);
3944 mutex_lock(&memcg
->thresholds_lock
);
3947 thresholds
= &memcg
->thresholds
;
3948 usage
= mem_cgroup_usage(memcg
, false);
3949 } else if (type
== _MEMSWAP
) {
3950 thresholds
= &memcg
->memsw_thresholds
;
3951 usage
= mem_cgroup_usage(memcg
, true);
3955 /* Check if a threshold crossed before adding a new one */
3956 if (thresholds
->primary
)
3957 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3959 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3961 /* Allocate memory for new array of thresholds */
3962 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
3969 /* Copy thresholds (if any) to new array */
3970 if (thresholds
->primary
) {
3971 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3972 sizeof(struct mem_cgroup_threshold
));
3975 /* Add new threshold */
3976 new->entries
[size
- 1].eventfd
= eventfd
;
3977 new->entries
[size
- 1].threshold
= threshold
;
3979 /* Sort thresholds. Registering of new threshold isn't time-critical */
3980 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3981 compare_thresholds
, NULL
);
3983 /* Find current threshold */
3984 new->current_threshold
= -1;
3985 for (i
= 0; i
< size
; i
++) {
3986 if (new->entries
[i
].threshold
<= usage
) {
3988 * new->current_threshold will not be used until
3989 * rcu_assign_pointer(), so it's safe to increment
3992 ++new->current_threshold
;
3997 /* Free old spare buffer and save old primary buffer as spare */
3998 kfree(thresholds
->spare
);
3999 thresholds
->spare
= thresholds
->primary
;
4001 rcu_assign_pointer(thresholds
->primary
, new);
4003 /* To be sure that nobody uses thresholds */
4007 mutex_unlock(&memcg
->thresholds_lock
);
4012 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4013 struct eventfd_ctx
*eventfd
, const char *args
)
4015 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4018 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4019 struct eventfd_ctx
*eventfd
, const char *args
)
4021 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4024 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4025 struct eventfd_ctx
*eventfd
, enum res_type type
)
4027 struct mem_cgroup_thresholds
*thresholds
;
4028 struct mem_cgroup_threshold_ary
*new;
4029 unsigned long usage
;
4032 mutex_lock(&memcg
->thresholds_lock
);
4035 thresholds
= &memcg
->thresholds
;
4036 usage
= mem_cgroup_usage(memcg
, false);
4037 } else if (type
== _MEMSWAP
) {
4038 thresholds
= &memcg
->memsw_thresholds
;
4039 usage
= mem_cgroup_usage(memcg
, true);
4043 if (!thresholds
->primary
)
4046 /* Check if a threshold crossed before removing */
4047 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4049 /* Calculate new number of threshold */
4051 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4052 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4056 new = thresholds
->spare
;
4058 /* Set thresholds array to NULL if we don't have thresholds */
4067 /* Copy thresholds and find current threshold */
4068 new->current_threshold
= -1;
4069 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4070 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4073 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4074 if (new->entries
[j
].threshold
<= usage
) {
4076 * new->current_threshold will not be used
4077 * until rcu_assign_pointer(), so it's safe to increment
4080 ++new->current_threshold
;
4086 /* Swap primary and spare array */
4087 thresholds
->spare
= thresholds
->primary
;
4089 rcu_assign_pointer(thresholds
->primary
, new);
4091 /* To be sure that nobody uses thresholds */
4094 /* If all events are unregistered, free the spare array */
4096 kfree(thresholds
->spare
);
4097 thresholds
->spare
= NULL
;
4100 mutex_unlock(&memcg
->thresholds_lock
);
4103 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4104 struct eventfd_ctx
*eventfd
)
4106 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4109 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4110 struct eventfd_ctx
*eventfd
)
4112 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4115 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4116 struct eventfd_ctx
*eventfd
, const char *args
)
4118 struct mem_cgroup_eventfd_list
*event
;
4120 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4124 spin_lock(&memcg_oom_lock
);
4126 event
->eventfd
= eventfd
;
4127 list_add(&event
->list
, &memcg
->oom_notify
);
4129 /* already in OOM ? */
4130 if (memcg
->under_oom
)
4131 eventfd_signal(eventfd
, 1);
4132 spin_unlock(&memcg_oom_lock
);
4137 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4138 struct eventfd_ctx
*eventfd
)
4140 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4142 spin_lock(&memcg_oom_lock
);
4144 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4145 if (ev
->eventfd
== eventfd
) {
4146 list_del(&ev
->list
);
4151 spin_unlock(&memcg_oom_lock
);
4154 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4156 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4158 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4159 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4160 seq_printf(sf
, "oom_kill %lu\n",
4161 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4165 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4166 struct cftype
*cft
, u64 val
)
4168 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4170 /* cannot set to root cgroup and only 0 and 1 are allowed */
4171 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4174 memcg
->oom_kill_disable
= val
;
4176 memcg_oom_recover(memcg
);
4181 #ifdef CONFIG_CGROUP_WRITEBACK
4183 #include <trace/events/writeback.h>
4185 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4187 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4190 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4192 wb_domain_exit(&memcg
->cgwb_domain
);
4195 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4197 wb_domain_size_changed(&memcg
->cgwb_domain
);
4200 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4202 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4204 if (!memcg
->css
.parent
)
4207 return &memcg
->cgwb_domain
;
4211 * idx can be of type enum memcg_stat_item or node_stat_item.
4212 * Keep in sync with memcg_exact_page().
4214 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4216 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4219 for_each_online_cpu(cpu
)
4220 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4227 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4228 * @wb: bdi_writeback in question
4229 * @pfilepages: out parameter for number of file pages
4230 * @pheadroom: out parameter for number of allocatable pages according to memcg
4231 * @pdirty: out parameter for number of dirty pages
4232 * @pwriteback: out parameter for number of pages under writeback
4234 * Determine the numbers of file, headroom, dirty, and writeback pages in
4235 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4236 * is a bit more involved.
4238 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4239 * headroom is calculated as the lowest headroom of itself and the
4240 * ancestors. Note that this doesn't consider the actual amount of
4241 * available memory in the system. The caller should further cap
4242 * *@pheadroom accordingly.
4244 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4245 unsigned long *pheadroom
, unsigned long *pdirty
,
4246 unsigned long *pwriteback
)
4248 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4249 struct mem_cgroup
*parent
;
4251 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4253 /* this should eventually include NR_UNSTABLE_NFS */
4254 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4255 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4256 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4257 *pheadroom
= PAGE_COUNTER_MAX
;
4259 while ((parent
= parent_mem_cgroup(memcg
))) {
4260 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
4261 unsigned long used
= page_counter_read(&memcg
->memory
);
4263 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4269 * Foreign dirty flushing
4271 * There's an inherent mismatch between memcg and writeback. The former
4272 * trackes ownership per-page while the latter per-inode. This was a
4273 * deliberate design decision because honoring per-page ownership in the
4274 * writeback path is complicated, may lead to higher CPU and IO overheads
4275 * and deemed unnecessary given that write-sharing an inode across
4276 * different cgroups isn't a common use-case.
4278 * Combined with inode majority-writer ownership switching, this works well
4279 * enough in most cases but there are some pathological cases. For
4280 * example, let's say there are two cgroups A and B which keep writing to
4281 * different but confined parts of the same inode. B owns the inode and
4282 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4283 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4284 * triggering background writeback. A will be slowed down without a way to
4285 * make writeback of the dirty pages happen.
4287 * Conditions like the above can lead to a cgroup getting repatedly and
4288 * severely throttled after making some progress after each
4289 * dirty_expire_interval while the underyling IO device is almost
4292 * Solving this problem completely requires matching the ownership tracking
4293 * granularities between memcg and writeback in either direction. However,
4294 * the more egregious behaviors can be avoided by simply remembering the
4295 * most recent foreign dirtying events and initiating remote flushes on
4296 * them when local writeback isn't enough to keep the memory clean enough.
4298 * The following two functions implement such mechanism. When a foreign
4299 * page - a page whose memcg and writeback ownerships don't match - is
4300 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4301 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4302 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4303 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4304 * foreign bdi_writebacks which haven't expired. Both the numbers of
4305 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4306 * limited to MEMCG_CGWB_FRN_CNT.
4308 * The mechanism only remembers IDs and doesn't hold any object references.
4309 * As being wrong occasionally doesn't matter, updates and accesses to the
4310 * records are lockless and racy.
4312 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4313 struct bdi_writeback
*wb
)
4315 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4316 struct memcg_cgwb_frn
*frn
;
4317 u64 now
= get_jiffies_64();
4318 u64 oldest_at
= now
;
4322 trace_track_foreign_dirty(page
, wb
);
4325 * Pick the slot to use. If there is already a slot for @wb, keep
4326 * using it. If not replace the oldest one which isn't being
4329 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4330 frn
= &memcg
->cgwb_frn
[i
];
4331 if (frn
->bdi_id
== wb
->bdi
->id
&&
4332 frn
->memcg_id
== wb
->memcg_css
->id
)
4334 if (time_before64(frn
->at
, oldest_at
) &&
4335 atomic_read(&frn
->done
.cnt
) == 1) {
4337 oldest_at
= frn
->at
;
4341 if (i
< MEMCG_CGWB_FRN_CNT
) {
4343 * Re-using an existing one. Update timestamp lazily to
4344 * avoid making the cacheline hot. We want them to be
4345 * reasonably up-to-date and significantly shorter than
4346 * dirty_expire_interval as that's what expires the record.
4347 * Use the shorter of 1s and dirty_expire_interval / 8.
4349 unsigned long update_intv
=
4350 min_t(unsigned long, HZ
,
4351 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4353 if (time_before64(frn
->at
, now
- update_intv
))
4355 } else if (oldest
>= 0) {
4356 /* replace the oldest free one */
4357 frn
= &memcg
->cgwb_frn
[oldest
];
4358 frn
->bdi_id
= wb
->bdi
->id
;
4359 frn
->memcg_id
= wb
->memcg_css
->id
;
4364 /* issue foreign writeback flushes for recorded foreign dirtying events */
4365 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4367 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4368 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4369 u64 now
= jiffies_64
;
4372 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4373 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4376 * If the record is older than dirty_expire_interval,
4377 * writeback on it has already started. No need to kick it
4378 * off again. Also, don't start a new one if there's
4379 * already one in flight.
4381 if (time_after64(frn
->at
, now
- intv
) &&
4382 atomic_read(&frn
->done
.cnt
) == 1) {
4384 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4385 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4386 WB_REASON_FOREIGN_FLUSH
,
4392 #else /* CONFIG_CGROUP_WRITEBACK */
4394 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4399 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4403 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4407 #endif /* CONFIG_CGROUP_WRITEBACK */
4410 * DO NOT USE IN NEW FILES.
4412 * "cgroup.event_control" implementation.
4414 * This is way over-engineered. It tries to support fully configurable
4415 * events for each user. Such level of flexibility is completely
4416 * unnecessary especially in the light of the planned unified hierarchy.
4418 * Please deprecate this and replace with something simpler if at all
4423 * Unregister event and free resources.
4425 * Gets called from workqueue.
4427 static void memcg_event_remove(struct work_struct
*work
)
4429 struct mem_cgroup_event
*event
=
4430 container_of(work
, struct mem_cgroup_event
, remove
);
4431 struct mem_cgroup
*memcg
= event
->memcg
;
4433 remove_wait_queue(event
->wqh
, &event
->wait
);
4435 event
->unregister_event(memcg
, event
->eventfd
);
4437 /* Notify userspace the event is going away. */
4438 eventfd_signal(event
->eventfd
, 1);
4440 eventfd_ctx_put(event
->eventfd
);
4442 css_put(&memcg
->css
);
4446 * Gets called on EPOLLHUP on eventfd when user closes it.
4448 * Called with wqh->lock held and interrupts disabled.
4450 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4451 int sync
, void *key
)
4453 struct mem_cgroup_event
*event
=
4454 container_of(wait
, struct mem_cgroup_event
, wait
);
4455 struct mem_cgroup
*memcg
= event
->memcg
;
4456 __poll_t flags
= key_to_poll(key
);
4458 if (flags
& EPOLLHUP
) {
4460 * If the event has been detached at cgroup removal, we
4461 * can simply return knowing the other side will cleanup
4464 * We can't race against event freeing since the other
4465 * side will require wqh->lock via remove_wait_queue(),
4468 spin_lock(&memcg
->event_list_lock
);
4469 if (!list_empty(&event
->list
)) {
4470 list_del_init(&event
->list
);
4472 * We are in atomic context, but cgroup_event_remove()
4473 * may sleep, so we have to call it in workqueue.
4475 schedule_work(&event
->remove
);
4477 spin_unlock(&memcg
->event_list_lock
);
4483 static void memcg_event_ptable_queue_proc(struct file
*file
,
4484 wait_queue_head_t
*wqh
, poll_table
*pt
)
4486 struct mem_cgroup_event
*event
=
4487 container_of(pt
, struct mem_cgroup_event
, pt
);
4490 add_wait_queue(wqh
, &event
->wait
);
4494 * DO NOT USE IN NEW FILES.
4496 * Parse input and register new cgroup event handler.
4498 * Input must be in format '<event_fd> <control_fd> <args>'.
4499 * Interpretation of args is defined by control file implementation.
4501 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4502 char *buf
, size_t nbytes
, loff_t off
)
4504 struct cgroup_subsys_state
*css
= of_css(of
);
4505 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4506 struct mem_cgroup_event
*event
;
4507 struct cgroup_subsys_state
*cfile_css
;
4508 unsigned int efd
, cfd
;
4515 buf
= strstrip(buf
);
4517 efd
= simple_strtoul(buf
, &endp
, 10);
4522 cfd
= simple_strtoul(buf
, &endp
, 10);
4523 if ((*endp
!= ' ') && (*endp
!= '\0'))
4527 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4531 event
->memcg
= memcg
;
4532 INIT_LIST_HEAD(&event
->list
);
4533 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4534 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4535 INIT_WORK(&event
->remove
, memcg_event_remove
);
4543 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4544 if (IS_ERR(event
->eventfd
)) {
4545 ret
= PTR_ERR(event
->eventfd
);
4552 goto out_put_eventfd
;
4555 /* the process need read permission on control file */
4556 /* AV: shouldn't we check that it's been opened for read instead? */
4557 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4562 * Determine the event callbacks and set them in @event. This used
4563 * to be done via struct cftype but cgroup core no longer knows
4564 * about these events. The following is crude but the whole thing
4565 * is for compatibility anyway.
4567 * DO NOT ADD NEW FILES.
4569 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4571 if (!strcmp(name
, "memory.usage_in_bytes")) {
4572 event
->register_event
= mem_cgroup_usage_register_event
;
4573 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4574 } else if (!strcmp(name
, "memory.oom_control")) {
4575 event
->register_event
= mem_cgroup_oom_register_event
;
4576 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4577 } else if (!strcmp(name
, "memory.pressure_level")) {
4578 event
->register_event
= vmpressure_register_event
;
4579 event
->unregister_event
= vmpressure_unregister_event
;
4580 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4581 event
->register_event
= memsw_cgroup_usage_register_event
;
4582 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4589 * Verify @cfile should belong to @css. Also, remaining events are
4590 * automatically removed on cgroup destruction but the removal is
4591 * asynchronous, so take an extra ref on @css.
4593 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4594 &memory_cgrp_subsys
);
4596 if (IS_ERR(cfile_css
))
4598 if (cfile_css
!= css
) {
4603 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4607 vfs_poll(efile
.file
, &event
->pt
);
4609 spin_lock(&memcg
->event_list_lock
);
4610 list_add(&event
->list
, &memcg
->event_list
);
4611 spin_unlock(&memcg
->event_list_lock
);
4623 eventfd_ctx_put(event
->eventfd
);
4632 static struct cftype mem_cgroup_legacy_files
[] = {
4634 .name
= "usage_in_bytes",
4635 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4636 .read_u64
= mem_cgroup_read_u64
,
4639 .name
= "max_usage_in_bytes",
4640 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4641 .write
= mem_cgroup_reset
,
4642 .read_u64
= mem_cgroup_read_u64
,
4645 .name
= "limit_in_bytes",
4646 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4647 .write
= mem_cgroup_write
,
4648 .read_u64
= mem_cgroup_read_u64
,
4651 .name
= "soft_limit_in_bytes",
4652 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4653 .write
= mem_cgroup_write
,
4654 .read_u64
= mem_cgroup_read_u64
,
4658 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4659 .write
= mem_cgroup_reset
,
4660 .read_u64
= mem_cgroup_read_u64
,
4664 .seq_show
= memcg_stat_show
,
4667 .name
= "force_empty",
4668 .write
= mem_cgroup_force_empty_write
,
4671 .name
= "use_hierarchy",
4672 .write_u64
= mem_cgroup_hierarchy_write
,
4673 .read_u64
= mem_cgroup_hierarchy_read
,
4676 .name
= "cgroup.event_control", /* XXX: for compat */
4677 .write
= memcg_write_event_control
,
4678 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4681 .name
= "swappiness",
4682 .read_u64
= mem_cgroup_swappiness_read
,
4683 .write_u64
= mem_cgroup_swappiness_write
,
4686 .name
= "move_charge_at_immigrate",
4687 .read_u64
= mem_cgroup_move_charge_read
,
4688 .write_u64
= mem_cgroup_move_charge_write
,
4691 .name
= "oom_control",
4692 .seq_show
= mem_cgroup_oom_control_read
,
4693 .write_u64
= mem_cgroup_oom_control_write
,
4694 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4697 .name
= "pressure_level",
4701 .name
= "numa_stat",
4702 .seq_show
= memcg_numa_stat_show
,
4706 .name
= "kmem.limit_in_bytes",
4707 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4708 .write
= mem_cgroup_write
,
4709 .read_u64
= mem_cgroup_read_u64
,
4712 .name
= "kmem.usage_in_bytes",
4713 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4714 .read_u64
= mem_cgroup_read_u64
,
4717 .name
= "kmem.failcnt",
4718 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4719 .write
= mem_cgroup_reset
,
4720 .read_u64
= mem_cgroup_read_u64
,
4723 .name
= "kmem.max_usage_in_bytes",
4724 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4725 .write
= mem_cgroup_reset
,
4726 .read_u64
= mem_cgroup_read_u64
,
4728 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4730 .name
= "kmem.slabinfo",
4731 .seq_start
= memcg_slab_start
,
4732 .seq_next
= memcg_slab_next
,
4733 .seq_stop
= memcg_slab_stop
,
4734 .seq_show
= memcg_slab_show
,
4738 .name
= "kmem.tcp.limit_in_bytes",
4739 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4740 .write
= mem_cgroup_write
,
4741 .read_u64
= mem_cgroup_read_u64
,
4744 .name
= "kmem.tcp.usage_in_bytes",
4745 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4746 .read_u64
= mem_cgroup_read_u64
,
4749 .name
= "kmem.tcp.failcnt",
4750 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4751 .write
= mem_cgroup_reset
,
4752 .read_u64
= mem_cgroup_read_u64
,
4755 .name
= "kmem.tcp.max_usage_in_bytes",
4756 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4757 .write
= mem_cgroup_reset
,
4758 .read_u64
= mem_cgroup_read_u64
,
4760 { }, /* terminate */
4764 * Private memory cgroup IDR
4766 * Swap-out records and page cache shadow entries need to store memcg
4767 * references in constrained space, so we maintain an ID space that is
4768 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4769 * memory-controlled cgroups to 64k.
4771 * However, there usually are many references to the oflline CSS after
4772 * the cgroup has been destroyed, such as page cache or reclaimable
4773 * slab objects, that don't need to hang on to the ID. We want to keep
4774 * those dead CSS from occupying IDs, or we might quickly exhaust the
4775 * relatively small ID space and prevent the creation of new cgroups
4776 * even when there are much fewer than 64k cgroups - possibly none.
4778 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4779 * be freed and recycled when it's no longer needed, which is usually
4780 * when the CSS is offlined.
4782 * The only exception to that are records of swapped out tmpfs/shmem
4783 * pages that need to be attributed to live ancestors on swapin. But
4784 * those references are manageable from userspace.
4787 static DEFINE_IDR(mem_cgroup_idr
);
4789 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4791 if (memcg
->id
.id
> 0) {
4792 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4797 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4799 refcount_add(n
, &memcg
->id
.ref
);
4802 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4804 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4805 mem_cgroup_id_remove(memcg
);
4807 /* Memcg ID pins CSS */
4808 css_put(&memcg
->css
);
4812 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4814 mem_cgroup_id_put_many(memcg
, 1);
4818 * mem_cgroup_from_id - look up a memcg from a memcg id
4819 * @id: the memcg id to look up
4821 * Caller must hold rcu_read_lock().
4823 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4825 WARN_ON_ONCE(!rcu_read_lock_held());
4826 return idr_find(&mem_cgroup_idr
, id
);
4829 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4831 struct mem_cgroup_per_node
*pn
;
4834 * This routine is called against possible nodes.
4835 * But it's BUG to call kmalloc() against offline node.
4837 * TODO: this routine can waste much memory for nodes which will
4838 * never be onlined. It's better to use memory hotplug callback
4841 if (!node_state(node
, N_NORMAL_MEMORY
))
4843 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4847 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
4848 if (!pn
->lruvec_stat_local
) {
4853 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4854 if (!pn
->lruvec_stat_cpu
) {
4855 free_percpu(pn
->lruvec_stat_local
);
4860 lruvec_init(&pn
->lruvec
);
4861 pn
->usage_in_excess
= 0;
4862 pn
->on_tree
= false;
4865 memcg
->nodeinfo
[node
] = pn
;
4869 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4871 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4876 free_percpu(pn
->lruvec_stat_cpu
);
4877 free_percpu(pn
->lruvec_stat_local
);
4881 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4886 free_mem_cgroup_per_node_info(memcg
, node
);
4887 free_percpu(memcg
->vmstats_percpu
);
4888 free_percpu(memcg
->vmstats_local
);
4892 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4894 memcg_wb_domain_exit(memcg
);
4896 * Flush percpu vmstats and vmevents to guarantee the value correctness
4897 * on parent's and all ancestor levels.
4899 memcg_flush_percpu_vmstats(memcg
);
4900 memcg_flush_percpu_vmevents(memcg
);
4901 __mem_cgroup_free(memcg
);
4904 static struct mem_cgroup
*mem_cgroup_alloc(void)
4906 struct mem_cgroup
*memcg
;
4909 int __maybe_unused i
;
4911 size
= sizeof(struct mem_cgroup
);
4912 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4914 memcg
= kzalloc(size
, GFP_KERNEL
);
4918 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4919 1, MEM_CGROUP_ID_MAX
,
4921 if (memcg
->id
.id
< 0)
4924 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
4925 if (!memcg
->vmstats_local
)
4928 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
4929 if (!memcg
->vmstats_percpu
)
4933 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
4936 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4939 INIT_WORK(&memcg
->high_work
, high_work_func
);
4940 INIT_LIST_HEAD(&memcg
->oom_notify
);
4941 mutex_init(&memcg
->thresholds_lock
);
4942 spin_lock_init(&memcg
->move_lock
);
4943 vmpressure_init(&memcg
->vmpressure
);
4944 INIT_LIST_HEAD(&memcg
->event_list
);
4945 spin_lock_init(&memcg
->event_list_lock
);
4946 memcg
->socket_pressure
= jiffies
;
4947 #ifdef CONFIG_MEMCG_KMEM
4948 memcg
->kmemcg_id
= -1;
4950 #ifdef CONFIG_CGROUP_WRITEBACK
4951 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4952 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
4953 memcg
->cgwb_frn
[i
].done
=
4954 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
4956 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4957 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
4958 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
4959 memcg
->deferred_split_queue
.split_queue_len
= 0;
4961 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
4964 mem_cgroup_id_remove(memcg
);
4965 __mem_cgroup_free(memcg
);
4969 static struct cgroup_subsys_state
* __ref
4970 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4972 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4973 struct mem_cgroup
*memcg
;
4974 long error
= -ENOMEM
;
4976 memcg
= mem_cgroup_alloc();
4978 return ERR_PTR(error
);
4980 memcg
->high
= PAGE_COUNTER_MAX
;
4981 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4983 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4984 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4986 if (parent
&& parent
->use_hierarchy
) {
4987 memcg
->use_hierarchy
= true;
4988 page_counter_init(&memcg
->memory
, &parent
->memory
);
4989 page_counter_init(&memcg
->swap
, &parent
->swap
);
4990 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4991 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4992 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4994 page_counter_init(&memcg
->memory
, NULL
);
4995 page_counter_init(&memcg
->swap
, NULL
);
4996 page_counter_init(&memcg
->memsw
, NULL
);
4997 page_counter_init(&memcg
->kmem
, NULL
);
4998 page_counter_init(&memcg
->tcpmem
, NULL
);
5000 * Deeper hierachy with use_hierarchy == false doesn't make
5001 * much sense so let cgroup subsystem know about this
5002 * unfortunate state in our controller.
5004 if (parent
!= root_mem_cgroup
)
5005 memory_cgrp_subsys
.broken_hierarchy
= true;
5008 /* The following stuff does not apply to the root */
5010 #ifdef CONFIG_MEMCG_KMEM
5011 INIT_LIST_HEAD(&memcg
->kmem_caches
);
5013 root_mem_cgroup
= memcg
;
5017 error
= memcg_online_kmem(memcg
);
5021 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5022 static_branch_inc(&memcg_sockets_enabled_key
);
5026 mem_cgroup_id_remove(memcg
);
5027 mem_cgroup_free(memcg
);
5028 return ERR_PTR(-ENOMEM
);
5031 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5033 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5036 * A memcg must be visible for memcg_expand_shrinker_maps()
5037 * by the time the maps are allocated. So, we allocate maps
5038 * here, when for_each_mem_cgroup() can't skip it.
5040 if (memcg_alloc_shrinker_maps(memcg
)) {
5041 mem_cgroup_id_remove(memcg
);
5045 /* Online state pins memcg ID, memcg ID pins CSS */
5046 refcount_set(&memcg
->id
.ref
, 1);
5051 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5054 struct mem_cgroup_event
*event
, *tmp
;
5057 * Unregister events and notify userspace.
5058 * Notify userspace about cgroup removing only after rmdir of cgroup
5059 * directory to avoid race between userspace and kernelspace.
5061 spin_lock(&memcg
->event_list_lock
);
5062 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5063 list_del_init(&event
->list
);
5064 schedule_work(&event
->remove
);
5066 spin_unlock(&memcg
->event_list_lock
);
5068 page_counter_set_min(&memcg
->memory
, 0);
5069 page_counter_set_low(&memcg
->memory
, 0);
5071 memcg_offline_kmem(memcg
);
5072 wb_memcg_offline(memcg
);
5074 drain_all_stock(memcg
);
5076 mem_cgroup_id_put(memcg
);
5079 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5081 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5083 invalidate_reclaim_iterators(memcg
);
5086 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5088 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5089 int __maybe_unused i
;
5091 #ifdef CONFIG_CGROUP_WRITEBACK
5092 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5093 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5095 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5096 static_branch_dec(&memcg_sockets_enabled_key
);
5098 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5099 static_branch_dec(&memcg_sockets_enabled_key
);
5101 vmpressure_cleanup(&memcg
->vmpressure
);
5102 cancel_work_sync(&memcg
->high_work
);
5103 mem_cgroup_remove_from_trees(memcg
);
5104 memcg_free_shrinker_maps(memcg
);
5105 memcg_free_kmem(memcg
);
5106 mem_cgroup_free(memcg
);
5110 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5111 * @css: the target css
5113 * Reset the states of the mem_cgroup associated with @css. This is
5114 * invoked when the userland requests disabling on the default hierarchy
5115 * but the memcg is pinned through dependency. The memcg should stop
5116 * applying policies and should revert to the vanilla state as it may be
5117 * made visible again.
5119 * The current implementation only resets the essential configurations.
5120 * This needs to be expanded to cover all the visible parts.
5122 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5124 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5126 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5127 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5128 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
5129 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5130 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5131 page_counter_set_min(&memcg
->memory
, 0);
5132 page_counter_set_low(&memcg
->memory
, 0);
5133 memcg
->high
= PAGE_COUNTER_MAX
;
5134 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5135 memcg_wb_domain_size_changed(memcg
);
5139 /* Handlers for move charge at task migration. */
5140 static int mem_cgroup_do_precharge(unsigned long count
)
5144 /* Try a single bulk charge without reclaim first, kswapd may wake */
5145 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5147 mc
.precharge
+= count
;
5151 /* Try charges one by one with reclaim, but do not retry */
5153 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5167 enum mc_target_type
{
5174 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5175 unsigned long addr
, pte_t ptent
)
5177 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5179 if (!page
|| !page_mapped(page
))
5181 if (PageAnon(page
)) {
5182 if (!(mc
.flags
& MOVE_ANON
))
5185 if (!(mc
.flags
& MOVE_FILE
))
5188 if (!get_page_unless_zero(page
))
5194 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5195 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5196 pte_t ptent
, swp_entry_t
*entry
)
5198 struct page
*page
= NULL
;
5199 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5201 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
5205 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5206 * a device and because they are not accessible by CPU they are store
5207 * as special swap entry in the CPU page table.
5209 if (is_device_private_entry(ent
)) {
5210 page
= device_private_entry_to_page(ent
);
5212 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5213 * a refcount of 1 when free (unlike normal page)
5215 if (!page_ref_add_unless(page
, 1, 1))
5221 * Because lookup_swap_cache() updates some statistics counter,
5222 * we call find_get_page() with swapper_space directly.
5224 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5225 if (do_memsw_account())
5226 entry
->val
= ent
.val
;
5231 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5232 pte_t ptent
, swp_entry_t
*entry
)
5238 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5239 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5241 struct page
*page
= NULL
;
5242 struct address_space
*mapping
;
5245 if (!vma
->vm_file
) /* anonymous vma */
5247 if (!(mc
.flags
& MOVE_FILE
))
5250 mapping
= vma
->vm_file
->f_mapping
;
5251 pgoff
= linear_page_index(vma
, addr
);
5253 /* page is moved even if it's not RSS of this task(page-faulted). */
5255 /* shmem/tmpfs may report page out on swap: account for that too. */
5256 if (shmem_mapping(mapping
)) {
5257 page
= find_get_entry(mapping
, pgoff
);
5258 if (xa_is_value(page
)) {
5259 swp_entry_t swp
= radix_to_swp_entry(page
);
5260 if (do_memsw_account())
5262 page
= find_get_page(swap_address_space(swp
),
5266 page
= find_get_page(mapping
, pgoff
);
5268 page
= find_get_page(mapping
, pgoff
);
5274 * mem_cgroup_move_account - move account of the page
5276 * @compound: charge the page as compound or small page
5277 * @from: mem_cgroup which the page is moved from.
5278 * @to: mem_cgroup which the page is moved to. @from != @to.
5280 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5282 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5285 static int mem_cgroup_move_account(struct page
*page
,
5287 struct mem_cgroup
*from
,
5288 struct mem_cgroup
*to
)
5290 struct lruvec
*from_vec
, *to_vec
;
5291 struct pglist_data
*pgdat
;
5292 unsigned long flags
;
5293 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5297 VM_BUG_ON(from
== to
);
5298 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5299 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5302 * Prevent mem_cgroup_migrate() from looking at
5303 * page->mem_cgroup of its source page while we change it.
5306 if (!trylock_page(page
))
5310 if (page
->mem_cgroup
!= from
)
5313 anon
= PageAnon(page
);
5315 pgdat
= page_pgdat(page
);
5316 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5317 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5319 spin_lock_irqsave(&from
->move_lock
, flags
);
5321 if (!anon
&& page_mapped(page
)) {
5322 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5323 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5327 * move_lock grabbed above and caller set from->moving_account, so
5328 * mod_memcg_page_state will serialize updates to PageDirty.
5329 * So mapping should be stable for dirty pages.
5331 if (!anon
&& PageDirty(page
)) {
5332 struct address_space
*mapping
= page_mapping(page
);
5334 if (mapping_cap_account_dirty(mapping
)) {
5335 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
, -nr_pages
);
5336 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
, nr_pages
);
5340 if (PageWriteback(page
)) {
5341 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5342 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5346 * It is safe to change page->mem_cgroup here because the page
5347 * is referenced, charged, and isolated - we can't race with
5348 * uncharging, charging, migration, or LRU putback.
5351 /* caller should have done css_get */
5352 page
->mem_cgroup
= to
;
5354 spin_unlock_irqrestore(&from
->move_lock
, flags
);
5358 local_irq_disable();
5359 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
5360 memcg_check_events(to
, page
);
5361 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
5362 memcg_check_events(from
, page
);
5371 * get_mctgt_type - get target type of moving charge
5372 * @vma: the vma the pte to be checked belongs
5373 * @addr: the address corresponding to the pte to be checked
5374 * @ptent: the pte to be checked
5375 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5378 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5379 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5380 * move charge. if @target is not NULL, the page is stored in target->page
5381 * with extra refcnt got(Callers should handle it).
5382 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5383 * target for charge migration. if @target is not NULL, the entry is stored
5385 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5386 * (so ZONE_DEVICE page and thus not on the lru).
5387 * For now we such page is charge like a regular page would be as for all
5388 * intent and purposes it is just special memory taking the place of a
5391 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5393 * Called with pte lock held.
5396 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5397 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5399 struct page
*page
= NULL
;
5400 enum mc_target_type ret
= MC_TARGET_NONE
;
5401 swp_entry_t ent
= { .val
= 0 };
5403 if (pte_present(ptent
))
5404 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5405 else if (is_swap_pte(ptent
))
5406 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5407 else if (pte_none(ptent
))
5408 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5410 if (!page
&& !ent
.val
)
5414 * Do only loose check w/o serialization.
5415 * mem_cgroup_move_account() checks the page is valid or
5416 * not under LRU exclusion.
5418 if (page
->mem_cgroup
== mc
.from
) {
5419 ret
= MC_TARGET_PAGE
;
5420 if (is_device_private_page(page
))
5421 ret
= MC_TARGET_DEVICE
;
5423 target
->page
= page
;
5425 if (!ret
|| !target
)
5429 * There is a swap entry and a page doesn't exist or isn't charged.
5430 * But we cannot move a tail-page in a THP.
5432 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5433 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5434 ret
= MC_TARGET_SWAP
;
5441 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5443 * We don't consider PMD mapped swapping or file mapped pages because THP does
5444 * not support them for now.
5445 * Caller should make sure that pmd_trans_huge(pmd) is true.
5447 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5448 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5450 struct page
*page
= NULL
;
5451 enum mc_target_type ret
= MC_TARGET_NONE
;
5453 if (unlikely(is_swap_pmd(pmd
))) {
5454 VM_BUG_ON(thp_migration_supported() &&
5455 !is_pmd_migration_entry(pmd
));
5458 page
= pmd_page(pmd
);
5459 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5460 if (!(mc
.flags
& MOVE_ANON
))
5462 if (page
->mem_cgroup
== mc
.from
) {
5463 ret
= MC_TARGET_PAGE
;
5466 target
->page
= page
;
5472 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5473 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5475 return MC_TARGET_NONE
;
5479 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5480 unsigned long addr
, unsigned long end
,
5481 struct mm_walk
*walk
)
5483 struct vm_area_struct
*vma
= walk
->vma
;
5487 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5490 * Note their can not be MC_TARGET_DEVICE for now as we do not
5491 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5492 * this might change.
5494 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5495 mc
.precharge
+= HPAGE_PMD_NR
;
5500 if (pmd_trans_unstable(pmd
))
5502 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5503 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5504 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5505 mc
.precharge
++; /* increment precharge temporarily */
5506 pte_unmap_unlock(pte
- 1, ptl
);
5512 static const struct mm_walk_ops precharge_walk_ops
= {
5513 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5516 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5518 unsigned long precharge
;
5520 down_read(&mm
->mmap_sem
);
5521 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5522 up_read(&mm
->mmap_sem
);
5524 precharge
= mc
.precharge
;
5530 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5532 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5534 VM_BUG_ON(mc
.moving_task
);
5535 mc
.moving_task
= current
;
5536 return mem_cgroup_do_precharge(precharge
);
5539 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5540 static void __mem_cgroup_clear_mc(void)
5542 struct mem_cgroup
*from
= mc
.from
;
5543 struct mem_cgroup
*to
= mc
.to
;
5545 /* we must uncharge all the leftover precharges from mc.to */
5547 cancel_charge(mc
.to
, mc
.precharge
);
5551 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5552 * we must uncharge here.
5554 if (mc
.moved_charge
) {
5555 cancel_charge(mc
.from
, mc
.moved_charge
);
5556 mc
.moved_charge
= 0;
5558 /* we must fixup refcnts and charges */
5559 if (mc
.moved_swap
) {
5560 /* uncharge swap account from the old cgroup */
5561 if (!mem_cgroup_is_root(mc
.from
))
5562 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5564 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5567 * we charged both to->memory and to->memsw, so we
5568 * should uncharge to->memory.
5570 if (!mem_cgroup_is_root(mc
.to
))
5571 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5573 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5574 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5578 memcg_oom_recover(from
);
5579 memcg_oom_recover(to
);
5580 wake_up_all(&mc
.waitq
);
5583 static void mem_cgroup_clear_mc(void)
5585 struct mm_struct
*mm
= mc
.mm
;
5588 * we must clear moving_task before waking up waiters at the end of
5591 mc
.moving_task
= NULL
;
5592 __mem_cgroup_clear_mc();
5593 spin_lock(&mc
.lock
);
5597 spin_unlock(&mc
.lock
);
5602 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5604 struct cgroup_subsys_state
*css
;
5605 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5606 struct mem_cgroup
*from
;
5607 struct task_struct
*leader
, *p
;
5608 struct mm_struct
*mm
;
5609 unsigned long move_flags
;
5612 /* charge immigration isn't supported on the default hierarchy */
5613 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5617 * Multi-process migrations only happen on the default hierarchy
5618 * where charge immigration is not used. Perform charge
5619 * immigration if @tset contains a leader and whine if there are
5623 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5626 memcg
= mem_cgroup_from_css(css
);
5632 * We are now commited to this value whatever it is. Changes in this
5633 * tunable will only affect upcoming migrations, not the current one.
5634 * So we need to save it, and keep it going.
5636 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5640 from
= mem_cgroup_from_task(p
);
5642 VM_BUG_ON(from
== memcg
);
5644 mm
= get_task_mm(p
);
5647 /* We move charges only when we move a owner of the mm */
5648 if (mm
->owner
== p
) {
5651 VM_BUG_ON(mc
.precharge
);
5652 VM_BUG_ON(mc
.moved_charge
);
5653 VM_BUG_ON(mc
.moved_swap
);
5655 spin_lock(&mc
.lock
);
5659 mc
.flags
= move_flags
;
5660 spin_unlock(&mc
.lock
);
5661 /* We set mc.moving_task later */
5663 ret
= mem_cgroup_precharge_mc(mm
);
5665 mem_cgroup_clear_mc();
5672 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5675 mem_cgroup_clear_mc();
5678 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5679 unsigned long addr
, unsigned long end
,
5680 struct mm_walk
*walk
)
5683 struct vm_area_struct
*vma
= walk
->vma
;
5686 enum mc_target_type target_type
;
5687 union mc_target target
;
5690 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5692 if (mc
.precharge
< HPAGE_PMD_NR
) {
5696 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5697 if (target_type
== MC_TARGET_PAGE
) {
5699 if (!isolate_lru_page(page
)) {
5700 if (!mem_cgroup_move_account(page
, true,
5702 mc
.precharge
-= HPAGE_PMD_NR
;
5703 mc
.moved_charge
+= HPAGE_PMD_NR
;
5705 putback_lru_page(page
);
5708 } else if (target_type
== MC_TARGET_DEVICE
) {
5710 if (!mem_cgroup_move_account(page
, true,
5712 mc
.precharge
-= HPAGE_PMD_NR
;
5713 mc
.moved_charge
+= HPAGE_PMD_NR
;
5721 if (pmd_trans_unstable(pmd
))
5724 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5725 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5726 pte_t ptent
= *(pte
++);
5727 bool device
= false;
5733 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5734 case MC_TARGET_DEVICE
:
5737 case MC_TARGET_PAGE
:
5740 * We can have a part of the split pmd here. Moving it
5741 * can be done but it would be too convoluted so simply
5742 * ignore such a partial THP and keep it in original
5743 * memcg. There should be somebody mapping the head.
5745 if (PageTransCompound(page
))
5747 if (!device
&& isolate_lru_page(page
))
5749 if (!mem_cgroup_move_account(page
, false,
5752 /* we uncharge from mc.from later. */
5756 putback_lru_page(page
);
5757 put
: /* get_mctgt_type() gets the page */
5760 case MC_TARGET_SWAP
:
5762 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5764 /* we fixup refcnts and charges later. */
5772 pte_unmap_unlock(pte
- 1, ptl
);
5777 * We have consumed all precharges we got in can_attach().
5778 * We try charge one by one, but don't do any additional
5779 * charges to mc.to if we have failed in charge once in attach()
5782 ret
= mem_cgroup_do_precharge(1);
5790 static const struct mm_walk_ops charge_walk_ops
= {
5791 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5794 static void mem_cgroup_move_charge(void)
5796 lru_add_drain_all();
5798 * Signal lock_page_memcg() to take the memcg's move_lock
5799 * while we're moving its pages to another memcg. Then wait
5800 * for already started RCU-only updates to finish.
5802 atomic_inc(&mc
.from
->moving_account
);
5805 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5807 * Someone who are holding the mmap_sem might be waiting in
5808 * waitq. So we cancel all extra charges, wake up all waiters,
5809 * and retry. Because we cancel precharges, we might not be able
5810 * to move enough charges, but moving charge is a best-effort
5811 * feature anyway, so it wouldn't be a big problem.
5813 __mem_cgroup_clear_mc();
5818 * When we have consumed all precharges and failed in doing
5819 * additional charge, the page walk just aborts.
5821 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
5824 up_read(&mc
.mm
->mmap_sem
);
5825 atomic_dec(&mc
.from
->moving_account
);
5828 static void mem_cgroup_move_task(void)
5831 mem_cgroup_move_charge();
5832 mem_cgroup_clear_mc();
5835 #else /* !CONFIG_MMU */
5836 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5840 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5843 static void mem_cgroup_move_task(void)
5849 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5850 * to verify whether we're attached to the default hierarchy on each mount
5853 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5856 * use_hierarchy is forced on the default hierarchy. cgroup core
5857 * guarantees that @root doesn't have any children, so turning it
5858 * on for the root memcg is enough.
5860 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5861 root_mem_cgroup
->use_hierarchy
= true;
5863 root_mem_cgroup
->use_hierarchy
= false;
5866 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
5868 if (value
== PAGE_COUNTER_MAX
)
5869 seq_puts(m
, "max\n");
5871 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
5876 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5879 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5881 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5884 static int memory_min_show(struct seq_file
*m
, void *v
)
5886 return seq_puts_memcg_tunable(m
,
5887 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
5890 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
5891 char *buf
, size_t nbytes
, loff_t off
)
5893 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5897 buf
= strstrip(buf
);
5898 err
= page_counter_memparse(buf
, "max", &min
);
5902 page_counter_set_min(&memcg
->memory
, min
);
5907 static int memory_low_show(struct seq_file
*m
, void *v
)
5909 return seq_puts_memcg_tunable(m
,
5910 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
5913 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5914 char *buf
, size_t nbytes
, loff_t off
)
5916 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5920 buf
= strstrip(buf
);
5921 err
= page_counter_memparse(buf
, "max", &low
);
5925 page_counter_set_low(&memcg
->memory
, low
);
5930 static int memory_high_show(struct seq_file
*m
, void *v
)
5932 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
5935 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5936 char *buf
, size_t nbytes
, loff_t off
)
5938 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5939 unsigned int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
5940 bool drained
= false;
5944 buf
= strstrip(buf
);
5945 err
= page_counter_memparse(buf
, "max", &high
);
5952 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
5953 unsigned long reclaimed
;
5955 if (nr_pages
<= high
)
5958 if (signal_pending(current
))
5962 drain_all_stock(memcg
);
5967 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
5970 if (!reclaimed
&& !nr_retries
--)
5977 static int memory_max_show(struct seq_file
*m
, void *v
)
5979 return seq_puts_memcg_tunable(m
,
5980 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
5983 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5984 char *buf
, size_t nbytes
, loff_t off
)
5986 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5987 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
5988 bool drained
= false;
5992 buf
= strstrip(buf
);
5993 err
= page_counter_memparse(buf
, "max", &max
);
5997 xchg(&memcg
->memory
.max
, max
);
6000 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6002 if (nr_pages
<= max
)
6005 if (signal_pending(current
))
6009 drain_all_stock(memcg
);
6015 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6021 memcg_memory_event(memcg
, MEMCG_OOM
);
6022 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6026 memcg_wb_domain_size_changed(memcg
);
6030 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6032 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6033 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6034 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6035 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6036 seq_printf(m
, "oom_kill %lu\n",
6037 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6040 static int memory_events_show(struct seq_file
*m
, void *v
)
6042 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6044 __memory_events_show(m
, memcg
->memory_events
);
6048 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6050 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6052 __memory_events_show(m
, memcg
->memory_events_local
);
6056 static int memory_stat_show(struct seq_file
*m
, void *v
)
6058 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6061 buf
= memory_stat_format(memcg
);
6069 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6071 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6073 seq_printf(m
, "%d\n", memcg
->oom_group
);
6078 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6079 char *buf
, size_t nbytes
, loff_t off
)
6081 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6084 buf
= strstrip(buf
);
6088 ret
= kstrtoint(buf
, 0, &oom_group
);
6092 if (oom_group
!= 0 && oom_group
!= 1)
6095 memcg
->oom_group
= oom_group
;
6100 static struct cftype memory_files
[] = {
6103 .flags
= CFTYPE_NOT_ON_ROOT
,
6104 .read_u64
= memory_current_read
,
6108 .flags
= CFTYPE_NOT_ON_ROOT
,
6109 .seq_show
= memory_min_show
,
6110 .write
= memory_min_write
,
6114 .flags
= CFTYPE_NOT_ON_ROOT
,
6115 .seq_show
= memory_low_show
,
6116 .write
= memory_low_write
,
6120 .flags
= CFTYPE_NOT_ON_ROOT
,
6121 .seq_show
= memory_high_show
,
6122 .write
= memory_high_write
,
6126 .flags
= CFTYPE_NOT_ON_ROOT
,
6127 .seq_show
= memory_max_show
,
6128 .write
= memory_max_write
,
6132 .flags
= CFTYPE_NOT_ON_ROOT
,
6133 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6134 .seq_show
= memory_events_show
,
6137 .name
= "events.local",
6138 .flags
= CFTYPE_NOT_ON_ROOT
,
6139 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6140 .seq_show
= memory_events_local_show
,
6144 .flags
= CFTYPE_NOT_ON_ROOT
,
6145 .seq_show
= memory_stat_show
,
6148 .name
= "oom.group",
6149 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6150 .seq_show
= memory_oom_group_show
,
6151 .write
= memory_oom_group_write
,
6156 struct cgroup_subsys memory_cgrp_subsys
= {
6157 .css_alloc
= mem_cgroup_css_alloc
,
6158 .css_online
= mem_cgroup_css_online
,
6159 .css_offline
= mem_cgroup_css_offline
,
6160 .css_released
= mem_cgroup_css_released
,
6161 .css_free
= mem_cgroup_css_free
,
6162 .css_reset
= mem_cgroup_css_reset
,
6163 .can_attach
= mem_cgroup_can_attach
,
6164 .cancel_attach
= mem_cgroup_cancel_attach
,
6165 .post_attach
= mem_cgroup_move_task
,
6166 .bind
= mem_cgroup_bind
,
6167 .dfl_cftypes
= memory_files
,
6168 .legacy_cftypes
= mem_cgroup_legacy_files
,
6173 * mem_cgroup_protected - check if memory consumption is in the normal range
6174 * @root: the top ancestor of the sub-tree being checked
6175 * @memcg: the memory cgroup to check
6177 * WARNING: This function is not stateless! It can only be used as part
6178 * of a top-down tree iteration, not for isolated queries.
6180 * Returns one of the following:
6181 * MEMCG_PROT_NONE: cgroup memory is not protected
6182 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6183 * an unprotected supply of reclaimable memory from other cgroups.
6184 * MEMCG_PROT_MIN: cgroup memory is protected
6186 * @root is exclusive; it is never protected when looked at directly
6188 * To provide a proper hierarchical behavior, effective memory.min/low values
6189 * are used. Below is the description of how effective memory.low is calculated.
6190 * Effective memory.min values is calculated in the same way.
6192 * Effective memory.low is always equal or less than the original memory.low.
6193 * If there is no memory.low overcommittment (which is always true for
6194 * top-level memory cgroups), these two values are equal.
6195 * Otherwise, it's a part of parent's effective memory.low,
6196 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6197 * memory.low usages, where memory.low usage is the size of actually
6201 * elow = min( memory.low, parent->elow * ------------------ ),
6202 * siblings_low_usage
6204 * | memory.current, if memory.current < memory.low
6209 * Such definition of the effective memory.low provides the expected
6210 * hierarchical behavior: parent's memory.low value is limiting
6211 * children, unprotected memory is reclaimed first and cgroups,
6212 * which are not using their guarantee do not affect actual memory
6215 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6217 * A A/memory.low = 2G, A/memory.current = 6G
6219 * BC DE B/memory.low = 3G B/memory.current = 2G
6220 * C/memory.low = 1G C/memory.current = 2G
6221 * D/memory.low = 0 D/memory.current = 2G
6222 * E/memory.low = 10G E/memory.current = 0
6224 * and the memory pressure is applied, the following memory distribution
6225 * is expected (approximately):
6227 * A/memory.current = 2G
6229 * B/memory.current = 1.3G
6230 * C/memory.current = 0.6G
6231 * D/memory.current = 0
6232 * E/memory.current = 0
6234 * These calculations require constant tracking of the actual low usages
6235 * (see propagate_protected_usage()), as well as recursive calculation of
6236 * effective memory.low values. But as we do call mem_cgroup_protected()
6237 * path for each memory cgroup top-down from the reclaim,
6238 * it's possible to optimize this part, and save calculated elow
6239 * for next usage. This part is intentionally racy, but it's ok,
6240 * as memory.low is a best-effort mechanism.
6242 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
6243 struct mem_cgroup
*memcg
)
6245 struct mem_cgroup
*parent
;
6246 unsigned long emin
, parent_emin
;
6247 unsigned long elow
, parent_elow
;
6248 unsigned long usage
;
6250 if (mem_cgroup_disabled())
6251 return MEMCG_PROT_NONE
;
6254 root
= root_mem_cgroup
;
6256 return MEMCG_PROT_NONE
;
6258 usage
= page_counter_read(&memcg
->memory
);
6260 return MEMCG_PROT_NONE
;
6262 emin
= memcg
->memory
.min
;
6263 elow
= memcg
->memory
.low
;
6265 parent
= parent_mem_cgroup(memcg
);
6266 /* No parent means a non-hierarchical mode on v1 memcg */
6268 return MEMCG_PROT_NONE
;
6273 parent_emin
= READ_ONCE(parent
->memory
.emin
);
6274 emin
= min(emin
, parent_emin
);
6275 if (emin
&& parent_emin
) {
6276 unsigned long min_usage
, siblings_min_usage
;
6278 min_usage
= min(usage
, memcg
->memory
.min
);
6279 siblings_min_usage
= atomic_long_read(
6280 &parent
->memory
.children_min_usage
);
6282 if (min_usage
&& siblings_min_usage
)
6283 emin
= min(emin
, parent_emin
* min_usage
/
6284 siblings_min_usage
);
6287 parent_elow
= READ_ONCE(parent
->memory
.elow
);
6288 elow
= min(elow
, parent_elow
);
6289 if (elow
&& parent_elow
) {
6290 unsigned long low_usage
, siblings_low_usage
;
6292 low_usage
= min(usage
, memcg
->memory
.low
);
6293 siblings_low_usage
= atomic_long_read(
6294 &parent
->memory
.children_low_usage
);
6296 if (low_usage
&& siblings_low_usage
)
6297 elow
= min(elow
, parent_elow
* low_usage
/
6298 siblings_low_usage
);
6302 memcg
->memory
.emin
= emin
;
6303 memcg
->memory
.elow
= elow
;
6306 return MEMCG_PROT_MIN
;
6307 else if (usage
<= elow
)
6308 return MEMCG_PROT_LOW
;
6310 return MEMCG_PROT_NONE
;
6314 * mem_cgroup_try_charge - try charging a page
6315 * @page: page to charge
6316 * @mm: mm context of the victim
6317 * @gfp_mask: reclaim mode
6318 * @memcgp: charged memcg return
6319 * @compound: charge the page as compound or small page
6321 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6322 * pages according to @gfp_mask if necessary.
6324 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6325 * Otherwise, an error code is returned.
6327 * After page->mapping has been set up, the caller must finalize the
6328 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6329 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6331 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6332 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6335 struct mem_cgroup
*memcg
= NULL
;
6336 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6339 if (mem_cgroup_disabled())
6342 if (PageSwapCache(page
)) {
6344 * Every swap fault against a single page tries to charge the
6345 * page, bail as early as possible. shmem_unuse() encounters
6346 * already charged pages, too. The USED bit is protected by
6347 * the page lock, which serializes swap cache removal, which
6348 * in turn serializes uncharging.
6350 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6351 if (compound_head(page
)->mem_cgroup
)
6354 if (do_swap_account
) {
6355 swp_entry_t ent
= { .val
= page_private(page
), };
6356 unsigned short id
= lookup_swap_cgroup_id(ent
);
6359 memcg
= mem_cgroup_from_id(id
);
6360 if (memcg
&& !css_tryget_online(&memcg
->css
))
6367 memcg
= get_mem_cgroup_from_mm(mm
);
6369 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6371 css_put(&memcg
->css
);
6377 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6378 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6381 struct mem_cgroup
*memcg
;
6384 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6386 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6391 * mem_cgroup_commit_charge - commit a page charge
6392 * @page: page to charge
6393 * @memcg: memcg to charge the page to
6394 * @lrucare: page might be on LRU already
6395 * @compound: charge the page as compound or small page
6397 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6398 * after page->mapping has been set up. This must happen atomically
6399 * as part of the page instantiation, i.e. under the page table lock
6400 * for anonymous pages, under the page lock for page and swap cache.
6402 * In addition, the page must not be on the LRU during the commit, to
6403 * prevent racing with task migration. If it might be, use @lrucare.
6405 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6407 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6408 bool lrucare
, bool compound
)
6410 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6412 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6413 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6415 if (mem_cgroup_disabled())
6418 * Swap faults will attempt to charge the same page multiple
6419 * times. But reuse_swap_page() might have removed the page
6420 * from swapcache already, so we can't check PageSwapCache().
6425 commit_charge(page
, memcg
, lrucare
);
6427 local_irq_disable();
6428 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6429 memcg_check_events(memcg
, page
);
6432 if (do_memsw_account() && PageSwapCache(page
)) {
6433 swp_entry_t entry
= { .val
= page_private(page
) };
6435 * The swap entry might not get freed for a long time,
6436 * let's not wait for it. The page already received a
6437 * memory+swap charge, drop the swap entry duplicate.
6439 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6444 * mem_cgroup_cancel_charge - cancel a page charge
6445 * @page: page to charge
6446 * @memcg: memcg to charge the page to
6447 * @compound: charge the page as compound or small page
6449 * Cancel a charge transaction started by mem_cgroup_try_charge().
6451 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6454 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6456 if (mem_cgroup_disabled())
6459 * Swap faults will attempt to charge the same page multiple
6460 * times. But reuse_swap_page() might have removed the page
6461 * from swapcache already, so we can't check PageSwapCache().
6466 cancel_charge(memcg
, nr_pages
);
6469 struct uncharge_gather
{
6470 struct mem_cgroup
*memcg
;
6471 unsigned long pgpgout
;
6472 unsigned long nr_anon
;
6473 unsigned long nr_file
;
6474 unsigned long nr_kmem
;
6475 unsigned long nr_huge
;
6476 unsigned long nr_shmem
;
6477 struct page
*dummy_page
;
6480 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6482 memset(ug
, 0, sizeof(*ug
));
6485 static void uncharge_batch(const struct uncharge_gather
*ug
)
6487 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6488 unsigned long flags
;
6490 if (!mem_cgroup_is_root(ug
->memcg
)) {
6491 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6492 if (do_memsw_account())
6493 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6494 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6495 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6496 memcg_oom_recover(ug
->memcg
);
6499 local_irq_save(flags
);
6500 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6501 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6502 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6503 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6504 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6505 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6506 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6507 local_irq_restore(flags
);
6509 if (!mem_cgroup_is_root(ug
->memcg
))
6510 css_put_many(&ug
->memcg
->css
, nr_pages
);
6513 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6515 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6516 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6517 !PageHWPoison(page
) , page
);
6519 if (!page
->mem_cgroup
)
6523 * Nobody should be changing or seriously looking at
6524 * page->mem_cgroup at this point, we have fully
6525 * exclusive access to the page.
6528 if (ug
->memcg
!= page
->mem_cgroup
) {
6531 uncharge_gather_clear(ug
);
6533 ug
->memcg
= page
->mem_cgroup
;
6536 if (!PageKmemcg(page
)) {
6537 unsigned int nr_pages
= 1;
6539 if (PageTransHuge(page
)) {
6540 nr_pages
= compound_nr(page
);
6541 ug
->nr_huge
+= nr_pages
;
6544 ug
->nr_anon
+= nr_pages
;
6546 ug
->nr_file
+= nr_pages
;
6547 if (PageSwapBacked(page
))
6548 ug
->nr_shmem
+= nr_pages
;
6552 ug
->nr_kmem
+= compound_nr(page
);
6553 __ClearPageKmemcg(page
);
6556 ug
->dummy_page
= page
;
6557 page
->mem_cgroup
= NULL
;
6560 static void uncharge_list(struct list_head
*page_list
)
6562 struct uncharge_gather ug
;
6563 struct list_head
*next
;
6565 uncharge_gather_clear(&ug
);
6568 * Note that the list can be a single page->lru; hence the
6569 * do-while loop instead of a simple list_for_each_entry().
6571 next
= page_list
->next
;
6575 page
= list_entry(next
, struct page
, lru
);
6576 next
= page
->lru
.next
;
6578 uncharge_page(page
, &ug
);
6579 } while (next
!= page_list
);
6582 uncharge_batch(&ug
);
6586 * mem_cgroup_uncharge - uncharge a page
6587 * @page: page to uncharge
6589 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6590 * mem_cgroup_commit_charge().
6592 void mem_cgroup_uncharge(struct page
*page
)
6594 struct uncharge_gather ug
;
6596 if (mem_cgroup_disabled())
6599 /* Don't touch page->lru of any random page, pre-check: */
6600 if (!page
->mem_cgroup
)
6603 uncharge_gather_clear(&ug
);
6604 uncharge_page(page
, &ug
);
6605 uncharge_batch(&ug
);
6609 * mem_cgroup_uncharge_list - uncharge a list of page
6610 * @page_list: list of pages to uncharge
6612 * Uncharge a list of pages previously charged with
6613 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6615 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6617 if (mem_cgroup_disabled())
6620 if (!list_empty(page_list
))
6621 uncharge_list(page_list
);
6625 * mem_cgroup_migrate - charge a page's replacement
6626 * @oldpage: currently circulating page
6627 * @newpage: replacement page
6629 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6630 * be uncharged upon free.
6632 * Both pages must be locked, @newpage->mapping must be set up.
6634 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6636 struct mem_cgroup
*memcg
;
6637 unsigned int nr_pages
;
6638 unsigned long flags
;
6640 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6641 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6642 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6643 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6646 if (mem_cgroup_disabled())
6649 /* Page cache replacement: new page already charged? */
6650 if (newpage
->mem_cgroup
)
6653 /* Swapcache readahead pages can get replaced before being charged */
6654 memcg
= oldpage
->mem_cgroup
;
6658 /* Force-charge the new page. The old one will be freed soon */
6659 nr_pages
= hpage_nr_pages(newpage
);
6661 page_counter_charge(&memcg
->memory
, nr_pages
);
6662 if (do_memsw_account())
6663 page_counter_charge(&memcg
->memsw
, nr_pages
);
6664 css_get_many(&memcg
->css
, nr_pages
);
6666 commit_charge(newpage
, memcg
, false);
6668 local_irq_save(flags
);
6669 mem_cgroup_charge_statistics(memcg
, newpage
, PageTransHuge(newpage
),
6671 memcg_check_events(memcg
, newpage
);
6672 local_irq_restore(flags
);
6675 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6676 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6678 void mem_cgroup_sk_alloc(struct sock
*sk
)
6680 struct mem_cgroup
*memcg
;
6682 if (!mem_cgroup_sockets_enabled
)
6685 /* Do not associate the sock with unrelated interrupted task's memcg. */
6690 memcg
= mem_cgroup_from_task(current
);
6691 if (memcg
== root_mem_cgroup
)
6693 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6695 if (css_tryget_online(&memcg
->css
))
6696 sk
->sk_memcg
= memcg
;
6701 void mem_cgroup_sk_free(struct sock
*sk
)
6704 css_put(&sk
->sk_memcg
->css
);
6708 * mem_cgroup_charge_skmem - charge socket memory
6709 * @memcg: memcg to charge
6710 * @nr_pages: number of pages to charge
6712 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6713 * @memcg's configured limit, %false if the charge had to be forced.
6715 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6717 gfp_t gfp_mask
= GFP_KERNEL
;
6719 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6720 struct page_counter
*fail
;
6722 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6723 memcg
->tcpmem_pressure
= 0;
6726 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6727 memcg
->tcpmem_pressure
= 1;
6731 /* Don't block in the packet receive path */
6733 gfp_mask
= GFP_NOWAIT
;
6735 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6737 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6740 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6745 * mem_cgroup_uncharge_skmem - uncharge socket memory
6746 * @memcg: memcg to uncharge
6747 * @nr_pages: number of pages to uncharge
6749 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6751 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6752 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6756 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6758 refill_stock(memcg
, nr_pages
);
6761 static int __init
cgroup_memory(char *s
)
6765 while ((token
= strsep(&s
, ",")) != NULL
) {
6768 if (!strcmp(token
, "nosocket"))
6769 cgroup_memory_nosocket
= true;
6770 if (!strcmp(token
, "nokmem"))
6771 cgroup_memory_nokmem
= true;
6775 __setup("cgroup.memory=", cgroup_memory
);
6778 * subsys_initcall() for memory controller.
6780 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6781 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6782 * basically everything that doesn't depend on a specific mem_cgroup structure
6783 * should be initialized from here.
6785 static int __init
mem_cgroup_init(void)
6789 #ifdef CONFIG_MEMCG_KMEM
6791 * Kmem cache creation is mostly done with the slab_mutex held,
6792 * so use a workqueue with limited concurrency to avoid stalling
6793 * all worker threads in case lots of cgroups are created and
6794 * destroyed simultaneously.
6796 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6797 BUG_ON(!memcg_kmem_cache_wq
);
6800 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6801 memcg_hotplug_cpu_dead
);
6803 for_each_possible_cpu(cpu
)
6804 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6807 for_each_node(node
) {
6808 struct mem_cgroup_tree_per_node
*rtpn
;
6810 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6811 node_online(node
) ? node
: NUMA_NO_NODE
);
6813 rtpn
->rb_root
= RB_ROOT
;
6814 rtpn
->rb_rightmost
= NULL
;
6815 spin_lock_init(&rtpn
->lock
);
6816 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6821 subsys_initcall(mem_cgroup_init
);
6823 #ifdef CONFIG_MEMCG_SWAP
6824 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6826 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6828 * The root cgroup cannot be destroyed, so it's refcount must
6831 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6835 memcg
= parent_mem_cgroup(memcg
);
6837 memcg
= root_mem_cgroup
;
6843 * mem_cgroup_swapout - transfer a memsw charge to swap
6844 * @page: page whose memsw charge to transfer
6845 * @entry: swap entry to move the charge to
6847 * Transfer the memsw charge of @page to @entry.
6849 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6851 struct mem_cgroup
*memcg
, *swap_memcg
;
6852 unsigned int nr_entries
;
6853 unsigned short oldid
;
6855 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6856 VM_BUG_ON_PAGE(page_count(page
), page
);
6858 if (!do_memsw_account())
6861 memcg
= page
->mem_cgroup
;
6863 /* Readahead page, never charged */
6868 * In case the memcg owning these pages has been offlined and doesn't
6869 * have an ID allocated to it anymore, charge the closest online
6870 * ancestor for the swap instead and transfer the memory+swap charge.
6872 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6873 nr_entries
= hpage_nr_pages(page
);
6874 /* Get references for the tail pages, too */
6876 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6877 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6879 VM_BUG_ON_PAGE(oldid
, page
);
6880 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6882 page
->mem_cgroup
= NULL
;
6884 if (!mem_cgroup_is_root(memcg
))
6885 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6887 if (memcg
!= swap_memcg
) {
6888 if (!mem_cgroup_is_root(swap_memcg
))
6889 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6890 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6894 * Interrupts should be disabled here because the caller holds the
6895 * i_pages lock which is taken with interrupts-off. It is
6896 * important here to have the interrupts disabled because it is the
6897 * only synchronisation we have for updating the per-CPU variables.
6899 VM_BUG_ON(!irqs_disabled());
6900 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6902 memcg_check_events(memcg
, page
);
6904 if (!mem_cgroup_is_root(memcg
))
6905 css_put_many(&memcg
->css
, nr_entries
);
6909 * mem_cgroup_try_charge_swap - try charging swap space for a page
6910 * @page: page being added to swap
6911 * @entry: swap entry to charge
6913 * Try to charge @page's memcg for the swap space at @entry.
6915 * Returns 0 on success, -ENOMEM on failure.
6917 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6919 unsigned int nr_pages
= hpage_nr_pages(page
);
6920 struct page_counter
*counter
;
6921 struct mem_cgroup
*memcg
;
6922 unsigned short oldid
;
6924 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6927 memcg
= page
->mem_cgroup
;
6929 /* Readahead page, never charged */
6934 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6938 memcg
= mem_cgroup_id_get_online(memcg
);
6940 if (!mem_cgroup_is_root(memcg
) &&
6941 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
6942 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
6943 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
6944 mem_cgroup_id_put(memcg
);
6948 /* Get references for the tail pages, too */
6950 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
6951 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
6952 VM_BUG_ON_PAGE(oldid
, page
);
6953 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
6959 * mem_cgroup_uncharge_swap - uncharge swap space
6960 * @entry: swap entry to uncharge
6961 * @nr_pages: the amount of swap space to uncharge
6963 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
6965 struct mem_cgroup
*memcg
;
6968 if (!do_swap_account
)
6971 id
= swap_cgroup_record(entry
, 0, nr_pages
);
6973 memcg
= mem_cgroup_from_id(id
);
6975 if (!mem_cgroup_is_root(memcg
)) {
6976 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6977 page_counter_uncharge(&memcg
->swap
, nr_pages
);
6979 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
6981 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
6982 mem_cgroup_id_put_many(memcg
, nr_pages
);
6987 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
6989 long nr_swap_pages
= get_nr_swap_pages();
6991 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6992 return nr_swap_pages
;
6993 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
6994 nr_swap_pages
= min_t(long, nr_swap_pages
,
6995 READ_ONCE(memcg
->swap
.max
) -
6996 page_counter_read(&memcg
->swap
));
6997 return nr_swap_pages
;
7000 bool mem_cgroup_swap_full(struct page
*page
)
7002 struct mem_cgroup
*memcg
;
7004 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7008 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7011 memcg
= page
->mem_cgroup
;
7015 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7016 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
7022 /* for remember boot option*/
7023 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7024 static int really_do_swap_account __initdata
= 1;
7026 static int really_do_swap_account __initdata
;
7029 static int __init
enable_swap_account(char *s
)
7031 if (!strcmp(s
, "1"))
7032 really_do_swap_account
= 1;
7033 else if (!strcmp(s
, "0"))
7034 really_do_swap_account
= 0;
7037 __setup("swapaccount=", enable_swap_account
);
7039 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7042 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7044 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7047 static int swap_max_show(struct seq_file
*m
, void *v
)
7049 return seq_puts_memcg_tunable(m
,
7050 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7053 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7054 char *buf
, size_t nbytes
, loff_t off
)
7056 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7060 buf
= strstrip(buf
);
7061 err
= page_counter_memparse(buf
, "max", &max
);
7065 xchg(&memcg
->swap
.max
, max
);
7070 static int swap_events_show(struct seq_file
*m
, void *v
)
7072 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7074 seq_printf(m
, "max %lu\n",
7075 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7076 seq_printf(m
, "fail %lu\n",
7077 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7082 static struct cftype swap_files
[] = {
7084 .name
= "swap.current",
7085 .flags
= CFTYPE_NOT_ON_ROOT
,
7086 .read_u64
= swap_current_read
,
7090 .flags
= CFTYPE_NOT_ON_ROOT
,
7091 .seq_show
= swap_max_show
,
7092 .write
= swap_max_write
,
7095 .name
= "swap.events",
7096 .flags
= CFTYPE_NOT_ON_ROOT
,
7097 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7098 .seq_show
= swap_events_show
,
7103 static struct cftype memsw_cgroup_files
[] = {
7105 .name
= "memsw.usage_in_bytes",
7106 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7107 .read_u64
= mem_cgroup_read_u64
,
7110 .name
= "memsw.max_usage_in_bytes",
7111 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7112 .write
= mem_cgroup_reset
,
7113 .read_u64
= mem_cgroup_read_u64
,
7116 .name
= "memsw.limit_in_bytes",
7117 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7118 .write
= mem_cgroup_write
,
7119 .read_u64
= mem_cgroup_read_u64
,
7122 .name
= "memsw.failcnt",
7123 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7124 .write
= mem_cgroup_reset
,
7125 .read_u64
= mem_cgroup_read_u64
,
7127 { }, /* terminate */
7130 static int __init
mem_cgroup_swap_init(void)
7132 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7133 do_swap_account
= 1;
7134 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
7136 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
7137 memsw_cgroup_files
));
7141 subsys_initcall(mem_cgroup_swap_init
);
7143 #endif /* CONFIG_MEMCG_SWAP */