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
);
780 void mod_memcg_obj_state(void *p
, int idx
, int val
)
782 struct mem_cgroup
*memcg
;
785 memcg
= mem_cgroup_from_obj(p
);
787 mod_memcg_state(memcg
, idx
, val
);
792 * __count_memcg_events - account VM events in a cgroup
793 * @memcg: the memory cgroup
794 * @idx: the event item
795 * @count: the number of events that occured
797 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
802 if (mem_cgroup_disabled())
805 x
= count
+ __this_cpu_read(memcg
->vmstats_percpu
->events
[idx
]);
806 if (unlikely(x
> MEMCG_CHARGE_BATCH
)) {
807 struct mem_cgroup
*mi
;
810 * Batch local counters to keep them in sync with
811 * the hierarchical ones.
813 __this_cpu_add(memcg
->vmstats_local
->events
[idx
], x
);
814 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
815 atomic_long_add(x
, &mi
->vmevents
[idx
]);
818 __this_cpu_write(memcg
->vmstats_percpu
->events
[idx
], x
);
821 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
823 return atomic_long_read(&memcg
->vmevents
[event
]);
826 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
831 for_each_possible_cpu(cpu
)
832 x
+= per_cpu(memcg
->vmstats_local
->events
[event
], cpu
);
836 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
838 bool compound
, int nr_pages
)
841 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
842 * counted as CACHE even if it's on ANON LRU.
845 __mod_memcg_state(memcg
, MEMCG_RSS
, nr_pages
);
847 __mod_memcg_state(memcg
, MEMCG_CACHE
, nr_pages
);
848 if (PageSwapBacked(page
))
849 __mod_memcg_state(memcg
, NR_SHMEM
, nr_pages
);
853 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
854 __mod_memcg_state(memcg
, MEMCG_RSS_HUGE
, nr_pages
);
857 /* pagein of a big page is an event. So, ignore page size */
859 __count_memcg_events(memcg
, PGPGIN
, 1);
861 __count_memcg_events(memcg
, PGPGOUT
, 1);
862 nr_pages
= -nr_pages
; /* for event */
865 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
868 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
869 enum mem_cgroup_events_target target
)
871 unsigned long val
, next
;
873 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
874 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
875 /* from time_after() in jiffies.h */
876 if ((long)(next
- val
) < 0) {
878 case MEM_CGROUP_TARGET_THRESH
:
879 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
881 case MEM_CGROUP_TARGET_SOFTLIMIT
:
882 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
887 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
894 * Check events in order.
897 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
899 /* threshold event is triggered in finer grain than soft limit */
900 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
901 MEM_CGROUP_TARGET_THRESH
))) {
904 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
905 MEM_CGROUP_TARGET_SOFTLIMIT
);
906 mem_cgroup_threshold(memcg
);
907 if (unlikely(do_softlimit
))
908 mem_cgroup_update_tree(memcg
, page
);
912 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
915 * mm_update_next_owner() may clear mm->owner to NULL
916 * if it races with swapoff, page migration, etc.
917 * So this can be called with p == NULL.
922 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
924 EXPORT_SYMBOL(mem_cgroup_from_task
);
927 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
928 * @mm: mm from which memcg should be extracted. It can be NULL.
930 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
931 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
934 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
936 struct mem_cgroup
*memcg
;
938 if (mem_cgroup_disabled())
944 * Page cache insertions can happen withou an
945 * actual mm context, e.g. during disk probing
946 * on boot, loopback IO, acct() writes etc.
949 memcg
= root_mem_cgroup
;
951 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
952 if (unlikely(!memcg
))
953 memcg
= root_mem_cgroup
;
955 } while (!css_tryget(&memcg
->css
));
959 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
962 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
963 * @page: page from which memcg should be extracted.
965 * Obtain a reference on page->memcg and returns it if successful. Otherwise
966 * root_mem_cgroup is returned.
968 struct mem_cgroup
*get_mem_cgroup_from_page(struct page
*page
)
970 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
972 if (mem_cgroup_disabled())
976 if (!memcg
|| !css_tryget_online(&memcg
->css
))
977 memcg
= root_mem_cgroup
;
981 EXPORT_SYMBOL(get_mem_cgroup_from_page
);
984 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
986 static __always_inline
struct mem_cgroup
*get_mem_cgroup_from_current(void)
988 if (unlikely(current
->active_memcg
)) {
989 struct mem_cgroup
*memcg
= root_mem_cgroup
;
992 if (css_tryget_online(¤t
->active_memcg
->css
))
993 memcg
= current
->active_memcg
;
997 return get_mem_cgroup_from_mm(current
->mm
);
1001 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1002 * @root: hierarchy root
1003 * @prev: previously returned memcg, NULL on first invocation
1004 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1006 * Returns references to children of the hierarchy below @root, or
1007 * @root itself, or %NULL after a full round-trip.
1009 * Caller must pass the return value in @prev on subsequent
1010 * invocations for reference counting, or use mem_cgroup_iter_break()
1011 * to cancel a hierarchy walk before the round-trip is complete.
1013 * Reclaimers can specify a node and a priority level in @reclaim to
1014 * divide up the memcgs in the hierarchy among all concurrent
1015 * reclaimers operating on the same node and priority.
1017 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1018 struct mem_cgroup
*prev
,
1019 struct mem_cgroup_reclaim_cookie
*reclaim
)
1021 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1022 struct cgroup_subsys_state
*css
= NULL
;
1023 struct mem_cgroup
*memcg
= NULL
;
1024 struct mem_cgroup
*pos
= NULL
;
1026 if (mem_cgroup_disabled())
1030 root
= root_mem_cgroup
;
1032 if (prev
&& !reclaim
)
1035 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1044 struct mem_cgroup_per_node
*mz
;
1046 mz
= mem_cgroup_nodeinfo(root
, reclaim
->pgdat
->node_id
);
1049 if (prev
&& reclaim
->generation
!= iter
->generation
)
1053 pos
= READ_ONCE(iter
->position
);
1054 if (!pos
|| css_tryget(&pos
->css
))
1057 * css reference reached zero, so iter->position will
1058 * be cleared by ->css_released. However, we should not
1059 * rely on this happening soon, because ->css_released
1060 * is called from a work queue, and by busy-waiting we
1061 * might block it. So we clear iter->position right
1064 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1072 css
= css_next_descendant_pre(css
, &root
->css
);
1075 * Reclaimers share the hierarchy walk, and a
1076 * new one might jump in right at the end of
1077 * the hierarchy - make sure they see at least
1078 * one group and restart from the beginning.
1086 * Verify the css and acquire a reference. The root
1087 * is provided by the caller, so we know it's alive
1088 * and kicking, and don't take an extra reference.
1090 memcg
= mem_cgroup_from_css(css
);
1092 if (css
== &root
->css
)
1095 if (css_tryget(css
))
1103 * The position could have already been updated by a competing
1104 * thread, so check that the value hasn't changed since we read
1105 * it to avoid reclaiming from the same cgroup twice.
1107 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1115 reclaim
->generation
= iter
->generation
;
1121 if (prev
&& prev
!= root
)
1122 css_put(&prev
->css
);
1128 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1129 * @root: hierarchy root
1130 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1132 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1133 struct mem_cgroup
*prev
)
1136 root
= root_mem_cgroup
;
1137 if (prev
&& prev
!= root
)
1138 css_put(&prev
->css
);
1141 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1142 struct mem_cgroup
*dead_memcg
)
1144 struct mem_cgroup_reclaim_iter
*iter
;
1145 struct mem_cgroup_per_node
*mz
;
1148 for_each_node(nid
) {
1149 mz
= mem_cgroup_nodeinfo(from
, nid
);
1151 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1155 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1157 struct mem_cgroup
*memcg
= dead_memcg
;
1158 struct mem_cgroup
*last
;
1161 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1163 } while ((memcg
= parent_mem_cgroup(memcg
)));
1166 * When cgruop1 non-hierarchy mode is used,
1167 * parent_mem_cgroup() does not walk all the way up to the
1168 * cgroup root (root_mem_cgroup). So we have to handle
1169 * dead_memcg from cgroup root separately.
1171 if (last
!= root_mem_cgroup
)
1172 __invalidate_reclaim_iterators(root_mem_cgroup
,
1177 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1178 * @memcg: hierarchy root
1179 * @fn: function to call for each task
1180 * @arg: argument passed to @fn
1182 * This function iterates over tasks attached to @memcg or to any of its
1183 * descendants and calls @fn for each task. If @fn returns a non-zero
1184 * value, the function breaks the iteration loop and returns the value.
1185 * Otherwise, it will iterate over all tasks and return 0.
1187 * This function must not be called for the root memory cgroup.
1189 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1190 int (*fn
)(struct task_struct
*, void *), void *arg
)
1192 struct mem_cgroup
*iter
;
1195 BUG_ON(memcg
== root_mem_cgroup
);
1197 for_each_mem_cgroup_tree(iter
, memcg
) {
1198 struct css_task_iter it
;
1199 struct task_struct
*task
;
1201 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1202 while (!ret
&& (task
= css_task_iter_next(&it
)))
1203 ret
= fn(task
, arg
);
1204 css_task_iter_end(&it
);
1206 mem_cgroup_iter_break(memcg
, iter
);
1214 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1216 * @pgdat: pgdat of the page
1218 * This function is only safe when following the LRU page isolation
1219 * and putback protocol: the LRU lock must be held, and the page must
1220 * either be PageLRU() or the caller must have isolated/allocated it.
1222 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct pglist_data
*pgdat
)
1224 struct mem_cgroup_per_node
*mz
;
1225 struct mem_cgroup
*memcg
;
1226 struct lruvec
*lruvec
;
1228 if (mem_cgroup_disabled()) {
1229 lruvec
= &pgdat
->__lruvec
;
1233 memcg
= page
->mem_cgroup
;
1235 * Swapcache readahead pages are added to the LRU - and
1236 * possibly migrated - before they are charged.
1239 memcg
= root_mem_cgroup
;
1241 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
1242 lruvec
= &mz
->lruvec
;
1245 * Since a node can be onlined after the mem_cgroup was created,
1246 * we have to be prepared to initialize lruvec->zone here;
1247 * and if offlined then reonlined, we need to reinitialize it.
1249 if (unlikely(lruvec
->pgdat
!= pgdat
))
1250 lruvec
->pgdat
= pgdat
;
1255 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1256 * @lruvec: mem_cgroup per zone lru vector
1257 * @lru: index of lru list the page is sitting on
1258 * @zid: zone id of the accounted pages
1259 * @nr_pages: positive when adding or negative when removing
1261 * This function must be called under lru_lock, just before a page is added
1262 * to or just after a page is removed from an lru list (that ordering being
1263 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1265 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1266 int zid
, int nr_pages
)
1268 struct mem_cgroup_per_node
*mz
;
1269 unsigned long *lru_size
;
1272 if (mem_cgroup_disabled())
1275 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1276 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1279 *lru_size
+= nr_pages
;
1282 if (WARN_ONCE(size
< 0,
1283 "%s(%p, %d, %d): lru_size %ld\n",
1284 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1290 *lru_size
+= nr_pages
;
1294 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1295 * @memcg: the memory cgroup
1297 * Returns the maximum amount of memory @mem can be charged with, in
1300 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1302 unsigned long margin
= 0;
1303 unsigned long count
;
1304 unsigned long limit
;
1306 count
= page_counter_read(&memcg
->memory
);
1307 limit
= READ_ONCE(memcg
->memory
.max
);
1309 margin
= limit
- count
;
1311 if (do_memsw_account()) {
1312 count
= page_counter_read(&memcg
->memsw
);
1313 limit
= READ_ONCE(memcg
->memsw
.max
);
1315 margin
= min(margin
, limit
- count
);
1324 * A routine for checking "mem" is under move_account() or not.
1326 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1327 * moving cgroups. This is for waiting at high-memory pressure
1330 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1332 struct mem_cgroup
*from
;
1333 struct mem_cgroup
*to
;
1336 * Unlike task_move routines, we access mc.to, mc.from not under
1337 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1339 spin_lock(&mc
.lock
);
1345 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1346 mem_cgroup_is_descendant(to
, memcg
);
1348 spin_unlock(&mc
.lock
);
1352 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1354 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1355 if (mem_cgroup_under_move(memcg
)) {
1357 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1358 /* moving charge context might have finished. */
1361 finish_wait(&mc
.waitq
, &wait
);
1368 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1373 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1378 * Provide statistics on the state of the memory subsystem as
1379 * well as cumulative event counters that show past behavior.
1381 * This list is ordered following a combination of these gradients:
1382 * 1) generic big picture -> specifics and details
1383 * 2) reflecting userspace activity -> reflecting kernel heuristics
1385 * Current memory state:
1388 seq_buf_printf(&s
, "anon %llu\n",
1389 (u64
)memcg_page_state(memcg
, MEMCG_RSS
) *
1391 seq_buf_printf(&s
, "file %llu\n",
1392 (u64
)memcg_page_state(memcg
, MEMCG_CACHE
) *
1394 seq_buf_printf(&s
, "kernel_stack %llu\n",
1395 (u64
)memcg_page_state(memcg
, MEMCG_KERNEL_STACK_KB
) *
1397 seq_buf_printf(&s
, "slab %llu\n",
1398 (u64
)(memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) +
1399 memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
)) *
1401 seq_buf_printf(&s
, "sock %llu\n",
1402 (u64
)memcg_page_state(memcg
, MEMCG_SOCK
) *
1405 seq_buf_printf(&s
, "shmem %llu\n",
1406 (u64
)memcg_page_state(memcg
, NR_SHMEM
) *
1408 seq_buf_printf(&s
, "file_mapped %llu\n",
1409 (u64
)memcg_page_state(memcg
, NR_FILE_MAPPED
) *
1411 seq_buf_printf(&s
, "file_dirty %llu\n",
1412 (u64
)memcg_page_state(memcg
, NR_FILE_DIRTY
) *
1414 seq_buf_printf(&s
, "file_writeback %llu\n",
1415 (u64
)memcg_page_state(memcg
, NR_WRITEBACK
) *
1419 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1420 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1421 * arse because it requires migrating the work out of rmap to a place
1422 * where the page->mem_cgroup is set up and stable.
1424 seq_buf_printf(&s
, "anon_thp %llu\n",
1425 (u64
)memcg_page_state(memcg
, MEMCG_RSS_HUGE
) *
1428 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1429 seq_buf_printf(&s
, "%s %llu\n", lru_list_name(i
),
1430 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
1433 seq_buf_printf(&s
, "slab_reclaimable %llu\n",
1434 (u64
)memcg_page_state(memcg
, NR_SLAB_RECLAIMABLE
) *
1436 seq_buf_printf(&s
, "slab_unreclaimable %llu\n",
1437 (u64
)memcg_page_state(memcg
, NR_SLAB_UNRECLAIMABLE
) *
1440 /* Accumulated memory events */
1442 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1443 memcg_events(memcg
, PGFAULT
));
1444 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1445 memcg_events(memcg
, PGMAJFAULT
));
1447 seq_buf_printf(&s
, "workingset_refault %lu\n",
1448 memcg_page_state(memcg
, WORKINGSET_REFAULT
));
1449 seq_buf_printf(&s
, "workingset_activate %lu\n",
1450 memcg_page_state(memcg
, WORKINGSET_ACTIVATE
));
1451 seq_buf_printf(&s
, "workingset_nodereclaim %lu\n",
1452 memcg_page_state(memcg
, WORKINGSET_NODERECLAIM
));
1454 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1455 memcg_events(memcg
, PGREFILL
));
1456 seq_buf_printf(&s
, "pgscan %lu\n",
1457 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1458 memcg_events(memcg
, PGSCAN_DIRECT
));
1459 seq_buf_printf(&s
, "pgsteal %lu\n",
1460 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1461 memcg_events(memcg
, PGSTEAL_DIRECT
));
1462 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1463 memcg_events(memcg
, PGACTIVATE
));
1464 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1465 memcg_events(memcg
, PGDEACTIVATE
));
1466 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1467 memcg_events(memcg
, PGLAZYFREE
));
1468 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1469 memcg_events(memcg
, PGLAZYFREED
));
1471 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1472 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1473 memcg_events(memcg
, THP_FAULT_ALLOC
));
1474 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1475 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1476 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1478 /* The above should easily fit into one page */
1479 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1484 #define K(x) ((x) << (PAGE_SHIFT-10))
1486 * mem_cgroup_print_oom_context: Print OOM information relevant to
1487 * memory controller.
1488 * @memcg: The memory cgroup that went over limit
1489 * @p: Task that is going to be killed
1491 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1494 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1499 pr_cont(",oom_memcg=");
1500 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1502 pr_cont(",global_oom");
1504 pr_cont(",task_memcg=");
1505 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1511 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1512 * memory controller.
1513 * @memcg: The memory cgroup that went over limit
1515 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1519 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1520 K((u64
)page_counter_read(&memcg
->memory
)),
1521 K((u64
)memcg
->memory
.max
), memcg
->memory
.failcnt
);
1522 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1523 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1524 K((u64
)page_counter_read(&memcg
->swap
)),
1525 K((u64
)memcg
->swap
.max
), memcg
->swap
.failcnt
);
1527 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1528 K((u64
)page_counter_read(&memcg
->memsw
)),
1529 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1530 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64
)page_counter_read(&memcg
->kmem
)),
1532 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1535 pr_info("Memory cgroup stats for ");
1536 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1538 buf
= memory_stat_format(memcg
);
1546 * Return the memory (and swap, if configured) limit for a memcg.
1548 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1552 max
= memcg
->memory
.max
;
1553 if (mem_cgroup_swappiness(memcg
)) {
1554 unsigned long memsw_max
;
1555 unsigned long swap_max
;
1557 memsw_max
= memcg
->memsw
.max
;
1558 swap_max
= memcg
->swap
.max
;
1559 swap_max
= min(swap_max
, (unsigned long)total_swap_pages
);
1560 max
= min(max
+ swap_max
, memsw_max
);
1565 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1567 return page_counter_read(&memcg
->memory
);
1570 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1573 struct oom_control oc
= {
1577 .gfp_mask
= gfp_mask
,
1582 if (mutex_lock_killable(&oom_lock
))
1585 * A few threads which were not waiting at mutex_lock_killable() can
1586 * fail to bail out. Therefore, check again after holding oom_lock.
1588 ret
= should_force_charge() || out_of_memory(&oc
);
1589 mutex_unlock(&oom_lock
);
1593 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1596 unsigned long *total_scanned
)
1598 struct mem_cgroup
*victim
= NULL
;
1601 unsigned long excess
;
1602 unsigned long nr_scanned
;
1603 struct mem_cgroup_reclaim_cookie reclaim
= {
1607 excess
= soft_limit_excess(root_memcg
);
1610 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1615 * If we have not been able to reclaim
1616 * anything, it might because there are
1617 * no reclaimable pages under this hierarchy
1622 * We want to do more targeted reclaim.
1623 * excess >> 2 is not to excessive so as to
1624 * reclaim too much, nor too less that we keep
1625 * coming back to reclaim from this cgroup
1627 if (total
>= (excess
>> 2) ||
1628 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1633 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1634 pgdat
, &nr_scanned
);
1635 *total_scanned
+= nr_scanned
;
1636 if (!soft_limit_excess(root_memcg
))
1639 mem_cgroup_iter_break(root_memcg
, victim
);
1643 #ifdef CONFIG_LOCKDEP
1644 static struct lockdep_map memcg_oom_lock_dep_map
= {
1645 .name
= "memcg_oom_lock",
1649 static DEFINE_SPINLOCK(memcg_oom_lock
);
1652 * Check OOM-Killer is already running under our hierarchy.
1653 * If someone is running, return false.
1655 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1657 struct mem_cgroup
*iter
, *failed
= NULL
;
1659 spin_lock(&memcg_oom_lock
);
1661 for_each_mem_cgroup_tree(iter
, memcg
) {
1662 if (iter
->oom_lock
) {
1664 * this subtree of our hierarchy is already locked
1665 * so we cannot give a lock.
1668 mem_cgroup_iter_break(memcg
, iter
);
1671 iter
->oom_lock
= true;
1676 * OK, we failed to lock the whole subtree so we have
1677 * to clean up what we set up to the failing subtree
1679 for_each_mem_cgroup_tree(iter
, memcg
) {
1680 if (iter
== failed
) {
1681 mem_cgroup_iter_break(memcg
, iter
);
1684 iter
->oom_lock
= false;
1687 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1689 spin_unlock(&memcg_oom_lock
);
1694 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1696 struct mem_cgroup
*iter
;
1698 spin_lock(&memcg_oom_lock
);
1699 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1700 for_each_mem_cgroup_tree(iter
, memcg
)
1701 iter
->oom_lock
= false;
1702 spin_unlock(&memcg_oom_lock
);
1705 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1707 struct mem_cgroup
*iter
;
1709 spin_lock(&memcg_oom_lock
);
1710 for_each_mem_cgroup_tree(iter
, memcg
)
1712 spin_unlock(&memcg_oom_lock
);
1715 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1717 struct mem_cgroup
*iter
;
1720 * When a new child is created while the hierarchy is under oom,
1721 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1723 spin_lock(&memcg_oom_lock
);
1724 for_each_mem_cgroup_tree(iter
, memcg
)
1725 if (iter
->under_oom
> 0)
1727 spin_unlock(&memcg_oom_lock
);
1730 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1732 struct oom_wait_info
{
1733 struct mem_cgroup
*memcg
;
1734 wait_queue_entry_t wait
;
1737 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1738 unsigned mode
, int sync
, void *arg
)
1740 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1741 struct mem_cgroup
*oom_wait_memcg
;
1742 struct oom_wait_info
*oom_wait_info
;
1744 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1745 oom_wait_memcg
= oom_wait_info
->memcg
;
1747 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1748 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1750 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1753 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1756 * For the following lockless ->under_oom test, the only required
1757 * guarantee is that it must see the state asserted by an OOM when
1758 * this function is called as a result of userland actions
1759 * triggered by the notification of the OOM. This is trivially
1760 * achieved by invoking mem_cgroup_mark_under_oom() before
1761 * triggering notification.
1763 if (memcg
&& memcg
->under_oom
)
1764 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1774 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1776 enum oom_status ret
;
1779 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1782 memcg_memory_event(memcg
, MEMCG_OOM
);
1785 * We are in the middle of the charge context here, so we
1786 * don't want to block when potentially sitting on a callstack
1787 * that holds all kinds of filesystem and mm locks.
1789 * cgroup1 allows disabling the OOM killer and waiting for outside
1790 * handling until the charge can succeed; remember the context and put
1791 * the task to sleep at the end of the page fault when all locks are
1794 * On the other hand, in-kernel OOM killer allows for an async victim
1795 * memory reclaim (oom_reaper) and that means that we are not solely
1796 * relying on the oom victim to make a forward progress and we can
1797 * invoke the oom killer here.
1799 * Please note that mem_cgroup_out_of_memory might fail to find a
1800 * victim and then we have to bail out from the charge path.
1802 if (memcg
->oom_kill_disable
) {
1803 if (!current
->in_user_fault
)
1805 css_get(&memcg
->css
);
1806 current
->memcg_in_oom
= memcg
;
1807 current
->memcg_oom_gfp_mask
= mask
;
1808 current
->memcg_oom_order
= order
;
1813 mem_cgroup_mark_under_oom(memcg
);
1815 locked
= mem_cgroup_oom_trylock(memcg
);
1818 mem_cgroup_oom_notify(memcg
);
1820 mem_cgroup_unmark_under_oom(memcg
);
1821 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1827 mem_cgroup_oom_unlock(memcg
);
1833 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1834 * @handle: actually kill/wait or just clean up the OOM state
1836 * This has to be called at the end of a page fault if the memcg OOM
1837 * handler was enabled.
1839 * Memcg supports userspace OOM handling where failed allocations must
1840 * sleep on a waitqueue until the userspace task resolves the
1841 * situation. Sleeping directly in the charge context with all kinds
1842 * of locks held is not a good idea, instead we remember an OOM state
1843 * in the task and mem_cgroup_oom_synchronize() has to be called at
1844 * the end of the page fault to complete the OOM handling.
1846 * Returns %true if an ongoing memcg OOM situation was detected and
1847 * completed, %false otherwise.
1849 bool mem_cgroup_oom_synchronize(bool handle
)
1851 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1852 struct oom_wait_info owait
;
1855 /* OOM is global, do not handle */
1862 owait
.memcg
= memcg
;
1863 owait
.wait
.flags
= 0;
1864 owait
.wait
.func
= memcg_oom_wake_function
;
1865 owait
.wait
.private = current
;
1866 INIT_LIST_HEAD(&owait
.wait
.entry
);
1868 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1869 mem_cgroup_mark_under_oom(memcg
);
1871 locked
= mem_cgroup_oom_trylock(memcg
);
1874 mem_cgroup_oom_notify(memcg
);
1876 if (locked
&& !memcg
->oom_kill_disable
) {
1877 mem_cgroup_unmark_under_oom(memcg
);
1878 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1879 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1880 current
->memcg_oom_order
);
1883 mem_cgroup_unmark_under_oom(memcg
);
1884 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1888 mem_cgroup_oom_unlock(memcg
);
1890 * There is no guarantee that an OOM-lock contender
1891 * sees the wakeups triggered by the OOM kill
1892 * uncharges. Wake any sleepers explicitely.
1894 memcg_oom_recover(memcg
);
1897 current
->memcg_in_oom
= NULL
;
1898 css_put(&memcg
->css
);
1903 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1904 * @victim: task to be killed by the OOM killer
1905 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1907 * Returns a pointer to a memory cgroup, which has to be cleaned up
1908 * by killing all belonging OOM-killable tasks.
1910 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1912 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1913 struct mem_cgroup
*oom_domain
)
1915 struct mem_cgroup
*oom_group
= NULL
;
1916 struct mem_cgroup
*memcg
;
1918 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1922 oom_domain
= root_mem_cgroup
;
1926 memcg
= mem_cgroup_from_task(victim
);
1927 if (memcg
== root_mem_cgroup
)
1931 * Traverse the memory cgroup hierarchy from the victim task's
1932 * cgroup up to the OOMing cgroup (or root) to find the
1933 * highest-level memory cgroup with oom.group set.
1935 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1936 if (memcg
->oom_group
)
1939 if (memcg
== oom_domain
)
1944 css_get(&oom_group
->css
);
1951 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1953 pr_info("Tasks in ");
1954 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1955 pr_cont(" are going to be killed due to memory.oom.group set\n");
1959 * lock_page_memcg - lock a page->mem_cgroup binding
1962 * This function protects unlocked LRU pages from being moved to
1965 * It ensures lifetime of the returned memcg. Caller is responsible
1966 * for the lifetime of the page; __unlock_page_memcg() is available
1967 * when @page might get freed inside the locked section.
1969 struct mem_cgroup
*lock_page_memcg(struct page
*page
)
1971 struct mem_cgroup
*memcg
;
1972 unsigned long flags
;
1975 * The RCU lock is held throughout the transaction. The fast
1976 * path can get away without acquiring the memcg->move_lock
1977 * because page moving starts with an RCU grace period.
1979 * The RCU lock also protects the memcg from being freed when
1980 * the page state that is going to change is the only thing
1981 * preventing the page itself from being freed. E.g. writeback
1982 * doesn't hold a page reference and relies on PG_writeback to
1983 * keep off truncation, migration and so forth.
1987 if (mem_cgroup_disabled())
1990 memcg
= page
->mem_cgroup
;
1991 if (unlikely(!memcg
))
1994 if (atomic_read(&memcg
->moving_account
) <= 0)
1997 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1998 if (memcg
!= page
->mem_cgroup
) {
1999 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2004 * When charge migration first begins, we can have locked and
2005 * unlocked page stat updates happening concurrently. Track
2006 * the task who has the lock for unlock_page_memcg().
2008 memcg
->move_lock_task
= current
;
2009 memcg
->move_lock_flags
= flags
;
2013 EXPORT_SYMBOL(lock_page_memcg
);
2016 * __unlock_page_memcg - unlock and unpin a memcg
2019 * Unlock and unpin a memcg returned by lock_page_memcg().
2021 void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2023 if (memcg
&& memcg
->move_lock_task
== current
) {
2024 unsigned long flags
= memcg
->move_lock_flags
;
2026 memcg
->move_lock_task
= NULL
;
2027 memcg
->move_lock_flags
= 0;
2029 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2036 * unlock_page_memcg - unlock a page->mem_cgroup binding
2039 void unlock_page_memcg(struct page
*page
)
2041 __unlock_page_memcg(page
->mem_cgroup
);
2043 EXPORT_SYMBOL(unlock_page_memcg
);
2045 struct memcg_stock_pcp
{
2046 struct mem_cgroup
*cached
; /* this never be root cgroup */
2047 unsigned int nr_pages
;
2048 struct work_struct work
;
2049 unsigned long flags
;
2050 #define FLUSHING_CACHED_CHARGE 0
2052 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2053 static DEFINE_MUTEX(percpu_charge_mutex
);
2056 * consume_stock: Try to consume stocked charge on this cpu.
2057 * @memcg: memcg to consume from.
2058 * @nr_pages: how many pages to charge.
2060 * The charges will only happen if @memcg matches the current cpu's memcg
2061 * stock, and at least @nr_pages are available in that stock. Failure to
2062 * service an allocation will refill the stock.
2064 * returns true if successful, false otherwise.
2066 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2068 struct memcg_stock_pcp
*stock
;
2069 unsigned long flags
;
2072 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2075 local_irq_save(flags
);
2077 stock
= this_cpu_ptr(&memcg_stock
);
2078 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2079 stock
->nr_pages
-= nr_pages
;
2083 local_irq_restore(flags
);
2089 * Returns stocks cached in percpu and reset cached information.
2091 static void drain_stock(struct memcg_stock_pcp
*stock
)
2093 struct mem_cgroup
*old
= stock
->cached
;
2095 if (stock
->nr_pages
) {
2096 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2097 if (do_memsw_account())
2098 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2099 css_put_many(&old
->css
, stock
->nr_pages
);
2100 stock
->nr_pages
= 0;
2102 stock
->cached
= NULL
;
2105 static void drain_local_stock(struct work_struct
*dummy
)
2107 struct memcg_stock_pcp
*stock
;
2108 unsigned long flags
;
2111 * The only protection from memory hotplug vs. drain_stock races is
2112 * that we always operate on local CPU stock here with IRQ disabled
2114 local_irq_save(flags
);
2116 stock
= this_cpu_ptr(&memcg_stock
);
2118 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2120 local_irq_restore(flags
);
2124 * Cache charges(val) to local per_cpu area.
2125 * This will be consumed by consume_stock() function, later.
2127 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2129 struct memcg_stock_pcp
*stock
;
2130 unsigned long flags
;
2132 local_irq_save(flags
);
2134 stock
= this_cpu_ptr(&memcg_stock
);
2135 if (stock
->cached
!= memcg
) { /* reset if necessary */
2137 stock
->cached
= memcg
;
2139 stock
->nr_pages
+= nr_pages
;
2141 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2144 local_irq_restore(flags
);
2148 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2149 * of the hierarchy under it.
2151 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2155 /* If someone's already draining, avoid adding running more workers. */
2156 if (!mutex_trylock(&percpu_charge_mutex
))
2159 * Notify other cpus that system-wide "drain" is running
2160 * We do not care about races with the cpu hotplug because cpu down
2161 * as well as workers from this path always operate on the local
2162 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2165 for_each_online_cpu(cpu
) {
2166 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2167 struct mem_cgroup
*memcg
;
2171 memcg
= stock
->cached
;
2172 if (memcg
&& stock
->nr_pages
&&
2173 mem_cgroup_is_descendant(memcg
, root_memcg
))
2178 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2180 drain_local_stock(&stock
->work
);
2182 schedule_work_on(cpu
, &stock
->work
);
2186 mutex_unlock(&percpu_charge_mutex
);
2189 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2191 struct memcg_stock_pcp
*stock
;
2192 struct mem_cgroup
*memcg
, *mi
;
2194 stock
= &per_cpu(memcg_stock
, cpu
);
2197 for_each_mem_cgroup(memcg
) {
2200 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
2204 x
= this_cpu_xchg(memcg
->vmstats_percpu
->stat
[i
], 0);
2206 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2207 atomic_long_add(x
, &memcg
->vmstats
[i
]);
2209 if (i
>= NR_VM_NODE_STAT_ITEMS
)
2212 for_each_node(nid
) {
2213 struct mem_cgroup_per_node
*pn
;
2215 pn
= mem_cgroup_nodeinfo(memcg
, nid
);
2216 x
= this_cpu_xchg(pn
->lruvec_stat_cpu
->count
[i
], 0);
2219 atomic_long_add(x
, &pn
->lruvec_stat
[i
]);
2220 } while ((pn
= parent_nodeinfo(pn
, nid
)));
2224 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
2227 x
= this_cpu_xchg(memcg
->vmstats_percpu
->events
[i
], 0);
2229 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
2230 atomic_long_add(x
, &memcg
->vmevents
[i
]);
2237 static void reclaim_high(struct mem_cgroup
*memcg
,
2238 unsigned int nr_pages
,
2242 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2244 memcg_memory_event(memcg
, MEMCG_HIGH
);
2245 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2246 } while ((memcg
= parent_mem_cgroup(memcg
)));
2249 static void high_work_func(struct work_struct
*work
)
2251 struct mem_cgroup
*memcg
;
2253 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2254 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2258 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2259 * enough to still cause a significant slowdown in most cases, while still
2260 * allowing diagnostics and tracing to proceed without becoming stuck.
2262 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2265 * When calculating the delay, we use these either side of the exponentiation to
2266 * maintain precision and scale to a reasonable number of jiffies (see the table
2269 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2270 * overage ratio to a delay.
2271 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2272 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2273 * to produce a reasonable delay curve.
2275 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2276 * reasonable delay curve compared to precision-adjusted overage, not
2277 * penalising heavily at first, but still making sure that growth beyond the
2278 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2279 * example, with a high of 100 megabytes:
2281 * +-------+------------------------+
2282 * | usage | time to allocate in ms |
2283 * +-------+------------------------+
2305 * +-------+------------------------+
2307 #define MEMCG_DELAY_PRECISION_SHIFT 20
2308 #define MEMCG_DELAY_SCALING_SHIFT 14
2311 * Get the number of jiffies that we should penalise a mischievous cgroup which
2312 * is exceeding its memory.high by checking both it and its ancestors.
2314 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2315 unsigned int nr_pages
)
2317 unsigned long penalty_jiffies
;
2318 u64 max_overage
= 0;
2321 unsigned long usage
, high
;
2324 usage
= page_counter_read(&memcg
->memory
);
2325 high
= READ_ONCE(memcg
->high
);
2328 * Prevent division by 0 in overage calculation by acting as if
2329 * it was a threshold of 1 page
2331 high
= max(high
, 1UL);
2333 overage
= usage
- high
;
2334 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2335 overage
= div64_u64(overage
, high
);
2337 if (overage
> max_overage
)
2338 max_overage
= overage
;
2339 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2340 !mem_cgroup_is_root(memcg
));
2346 * We use overage compared to memory.high to calculate the number of
2347 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2348 * fairly lenient on small overages, and increasingly harsh when the
2349 * memcg in question makes it clear that it has no intention of stopping
2350 * its crazy behaviour, so we exponentially increase the delay based on
2353 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2354 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2355 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2358 * Factor in the task's own contribution to the overage, such that four
2359 * N-sized allocations are throttled approximately the same as one
2360 * 4N-sized allocation.
2362 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2363 * larger the current charge patch is than that.
2365 penalty_jiffies
= penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2368 * Clamp the max delay per usermode return so as to still keep the
2369 * application moving forwards and also permit diagnostics, albeit
2372 return min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2376 * Scheduled by try_charge() to be executed from the userland return path
2377 * and reclaims memory over the high limit.
2379 void mem_cgroup_handle_over_high(void)
2381 unsigned long penalty_jiffies
;
2382 unsigned long pflags
;
2383 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2384 struct mem_cgroup
*memcg
;
2386 if (likely(!nr_pages
))
2389 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2390 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
2391 current
->memcg_nr_pages_over_high
= 0;
2394 * memory.high is breached and reclaim is unable to keep up. Throttle
2395 * allocators proactively to slow down excessive growth.
2397 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
);
2400 * Don't sleep if the amount of jiffies this memcg owes us is so low
2401 * that it's not even worth doing, in an attempt to be nice to those who
2402 * go only a small amount over their memory.high value and maybe haven't
2403 * been aggressively reclaimed enough yet.
2405 if (penalty_jiffies
<= HZ
/ 100)
2409 * If we exit early, we're guaranteed to die (since
2410 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2411 * need to account for any ill-begotten jiffies to pay them off later.
2413 psi_memstall_enter(&pflags
);
2414 schedule_timeout_killable(penalty_jiffies
);
2415 psi_memstall_leave(&pflags
);
2418 css_put(&memcg
->css
);
2421 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2422 unsigned int nr_pages
)
2424 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2425 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2426 struct mem_cgroup
*mem_over_limit
;
2427 struct page_counter
*counter
;
2428 unsigned long nr_reclaimed
;
2429 bool may_swap
= true;
2430 bool drained
= false;
2431 enum oom_status oom_status
;
2433 if (mem_cgroup_is_root(memcg
))
2436 if (consume_stock(memcg
, nr_pages
))
2439 if (!do_memsw_account() ||
2440 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2441 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2443 if (do_memsw_account())
2444 page_counter_uncharge(&memcg
->memsw
, batch
);
2445 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2447 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2451 if (batch
> nr_pages
) {
2457 * Memcg doesn't have a dedicated reserve for atomic
2458 * allocations. But like the global atomic pool, we need to
2459 * put the burden of reclaim on regular allocation requests
2460 * and let these go through as privileged allocations.
2462 if (gfp_mask
& __GFP_ATOMIC
)
2466 * Unlike in global OOM situations, memcg is not in a physical
2467 * memory shortage. Allow dying and OOM-killed tasks to
2468 * bypass the last charges so that they can exit quickly and
2469 * free their memory.
2471 if (unlikely(should_force_charge()))
2475 * Prevent unbounded recursion when reclaim operations need to
2476 * allocate memory. This might exceed the limits temporarily,
2477 * but we prefer facilitating memory reclaim and getting back
2478 * under the limit over triggering OOM kills in these cases.
2480 if (unlikely(current
->flags
& PF_MEMALLOC
))
2483 if (unlikely(task_in_memcg_oom(current
)))
2486 if (!gfpflags_allow_blocking(gfp_mask
))
2489 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2491 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2492 gfp_mask
, may_swap
);
2494 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2498 drain_all_stock(mem_over_limit
);
2503 if (gfp_mask
& __GFP_NORETRY
)
2506 * Even though the limit is exceeded at this point, reclaim
2507 * may have been able to free some pages. Retry the charge
2508 * before killing the task.
2510 * Only for regular pages, though: huge pages are rather
2511 * unlikely to succeed so close to the limit, and we fall back
2512 * to regular pages anyway in case of failure.
2514 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2517 * At task move, charge accounts can be doubly counted. So, it's
2518 * better to wait until the end of task_move if something is going on.
2520 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2526 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2529 if (gfp_mask
& __GFP_NOFAIL
)
2532 if (fatal_signal_pending(current
))
2536 * keep retrying as long as the memcg oom killer is able to make
2537 * a forward progress or bypass the charge if the oom killer
2538 * couldn't make any progress.
2540 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2541 get_order(nr_pages
* PAGE_SIZE
));
2542 switch (oom_status
) {
2544 nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2552 if (!(gfp_mask
& __GFP_NOFAIL
))
2556 * The allocation either can't fail or will lead to more memory
2557 * being freed very soon. Allow memory usage go over the limit
2558 * temporarily by force charging it.
2560 page_counter_charge(&memcg
->memory
, nr_pages
);
2561 if (do_memsw_account())
2562 page_counter_charge(&memcg
->memsw
, nr_pages
);
2563 css_get_many(&memcg
->css
, nr_pages
);
2568 css_get_many(&memcg
->css
, batch
);
2569 if (batch
> nr_pages
)
2570 refill_stock(memcg
, batch
- nr_pages
);
2573 * If the hierarchy is above the normal consumption range, schedule
2574 * reclaim on returning to userland. We can perform reclaim here
2575 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2576 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2577 * not recorded as it most likely matches current's and won't
2578 * change in the meantime. As high limit is checked again before
2579 * reclaim, the cost of mismatch is negligible.
2582 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2583 /* Don't bother a random interrupted task */
2584 if (in_interrupt()) {
2585 schedule_work(&memcg
->high_work
);
2588 current
->memcg_nr_pages_over_high
+= batch
;
2589 set_notify_resume(current
);
2592 } while ((memcg
= parent_mem_cgroup(memcg
)));
2597 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2599 if (mem_cgroup_is_root(memcg
))
2602 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2603 if (do_memsw_account())
2604 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2606 css_put_many(&memcg
->css
, nr_pages
);
2609 static void lock_page_lru(struct page
*page
, int *isolated
)
2611 pg_data_t
*pgdat
= page_pgdat(page
);
2613 spin_lock_irq(&pgdat
->lru_lock
);
2614 if (PageLRU(page
)) {
2615 struct lruvec
*lruvec
;
2617 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2619 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2625 static void unlock_page_lru(struct page
*page
, int isolated
)
2627 pg_data_t
*pgdat
= page_pgdat(page
);
2630 struct lruvec
*lruvec
;
2632 lruvec
= mem_cgroup_page_lruvec(page
, pgdat
);
2633 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2635 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2637 spin_unlock_irq(&pgdat
->lru_lock
);
2640 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2645 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2648 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2649 * may already be on some other mem_cgroup's LRU. Take care of it.
2652 lock_page_lru(page
, &isolated
);
2655 * Nobody should be changing or seriously looking at
2656 * page->mem_cgroup at this point:
2658 * - the page is uncharged
2660 * - the page is off-LRU
2662 * - an anonymous fault has exclusive page access, except for
2663 * a locked page table
2665 * - a page cache insertion, a swapin fault, or a migration
2666 * have the page locked
2668 page
->mem_cgroup
= memcg
;
2671 unlock_page_lru(page
, isolated
);
2674 #ifdef CONFIG_MEMCG_KMEM
2676 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2678 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2679 * cgroup_mutex, etc.
2681 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2685 if (mem_cgroup_disabled())
2688 page
= virt_to_head_page(p
);
2691 * Slab pages don't have page->mem_cgroup set because corresponding
2692 * kmem caches can be reparented during the lifetime. That's why
2693 * memcg_from_slab_page() should be used instead.
2696 return memcg_from_slab_page(page
);
2698 /* All other pages use page->mem_cgroup */
2699 return page
->mem_cgroup
;
2702 static int memcg_alloc_cache_id(void)
2707 id
= ida_simple_get(&memcg_cache_ida
,
2708 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2712 if (id
< memcg_nr_cache_ids
)
2716 * There's no space for the new id in memcg_caches arrays,
2717 * so we have to grow them.
2719 down_write(&memcg_cache_ids_sem
);
2721 size
= 2 * (id
+ 1);
2722 if (size
< MEMCG_CACHES_MIN_SIZE
)
2723 size
= MEMCG_CACHES_MIN_SIZE
;
2724 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2725 size
= MEMCG_CACHES_MAX_SIZE
;
2727 err
= memcg_update_all_caches(size
);
2729 err
= memcg_update_all_list_lrus(size
);
2731 memcg_nr_cache_ids
= size
;
2733 up_write(&memcg_cache_ids_sem
);
2736 ida_simple_remove(&memcg_cache_ida
, id
);
2742 static void memcg_free_cache_id(int id
)
2744 ida_simple_remove(&memcg_cache_ida
, id
);
2747 struct memcg_kmem_cache_create_work
{
2748 struct mem_cgroup
*memcg
;
2749 struct kmem_cache
*cachep
;
2750 struct work_struct work
;
2753 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2755 struct memcg_kmem_cache_create_work
*cw
=
2756 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2757 struct mem_cgroup
*memcg
= cw
->memcg
;
2758 struct kmem_cache
*cachep
= cw
->cachep
;
2760 memcg_create_kmem_cache(memcg
, cachep
);
2762 css_put(&memcg
->css
);
2767 * Enqueue the creation of a per-memcg kmem_cache.
2769 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2770 struct kmem_cache
*cachep
)
2772 struct memcg_kmem_cache_create_work
*cw
;
2774 if (!css_tryget_online(&memcg
->css
))
2777 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
| __GFP_NOWARN
);
2782 cw
->cachep
= cachep
;
2783 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2785 queue_work(memcg_kmem_cache_wq
, &cw
->work
);
2788 static inline bool memcg_kmem_bypass(void)
2790 if (in_interrupt() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
2796 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2797 * @cachep: the original global kmem cache
2799 * Return the kmem_cache we're supposed to use for a slab allocation.
2800 * We try to use the current memcg's version of the cache.
2802 * If the cache does not exist yet, if we are the first user of it, we
2803 * create it asynchronously in a workqueue and let the current allocation
2804 * go through with the original cache.
2806 * This function takes a reference to the cache it returns to assure it
2807 * won't get destroyed while we are working with it. Once the caller is
2808 * done with it, memcg_kmem_put_cache() must be called to release the
2811 struct kmem_cache
*memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2813 struct mem_cgroup
*memcg
;
2814 struct kmem_cache
*memcg_cachep
;
2815 struct memcg_cache_array
*arr
;
2818 VM_BUG_ON(!is_root_cache(cachep
));
2820 if (memcg_kmem_bypass())
2825 if (unlikely(current
->active_memcg
))
2826 memcg
= current
->active_memcg
;
2828 memcg
= mem_cgroup_from_task(current
);
2830 if (!memcg
|| memcg
== root_mem_cgroup
)
2833 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2837 arr
= rcu_dereference(cachep
->memcg_params
.memcg_caches
);
2840 * Make sure we will access the up-to-date value. The code updating
2841 * memcg_caches issues a write barrier to match the data dependency
2842 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2844 memcg_cachep
= READ_ONCE(arr
->entries
[kmemcg_id
]);
2847 * If we are in a safe context (can wait, and not in interrupt
2848 * context), we could be be predictable and return right away.
2849 * This would guarantee that the allocation being performed
2850 * already belongs in the new cache.
2852 * However, there are some clashes that can arrive from locking.
2853 * For instance, because we acquire the slab_mutex while doing
2854 * memcg_create_kmem_cache, this means no further allocation
2855 * could happen with the slab_mutex held. So it's better to
2858 * If the memcg is dying or memcg_cache is about to be released,
2859 * don't bother creating new kmem_caches. Because memcg_cachep
2860 * is ZEROed as the fist step of kmem offlining, we don't need
2861 * percpu_ref_tryget_live() here. css_tryget_online() check in
2862 * memcg_schedule_kmem_cache_create() will prevent us from
2863 * creation of a new kmem_cache.
2865 if (unlikely(!memcg_cachep
))
2866 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2867 else if (percpu_ref_tryget(&memcg_cachep
->memcg_params
.refcnt
))
2868 cachep
= memcg_cachep
;
2875 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2876 * @cachep: the cache returned by memcg_kmem_get_cache
2878 void memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2880 if (!is_root_cache(cachep
))
2881 percpu_ref_put(&cachep
->memcg_params
.refcnt
);
2885 * __memcg_kmem_charge_memcg: charge a kmem page
2886 * @page: page to charge
2887 * @gfp: reclaim mode
2888 * @order: allocation order
2889 * @memcg: memory cgroup to charge
2891 * Returns 0 on success, an error code on failure.
2893 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2894 struct mem_cgroup
*memcg
)
2896 unsigned int nr_pages
= 1 << order
;
2897 struct page_counter
*counter
;
2900 ret
= try_charge(memcg
, gfp
, nr_pages
);
2904 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2905 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2908 * Enforce __GFP_NOFAIL allocation because callers are not
2909 * prepared to see failures and likely do not have any failure
2912 if (gfp
& __GFP_NOFAIL
) {
2913 page_counter_charge(&memcg
->kmem
, nr_pages
);
2916 cancel_charge(memcg
, nr_pages
);
2923 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2924 * @page: page to charge
2925 * @gfp: reclaim mode
2926 * @order: allocation order
2928 * Returns 0 on success, an error code on failure.
2930 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2932 struct mem_cgroup
*memcg
;
2935 if (memcg_kmem_bypass())
2938 memcg
= get_mem_cgroup_from_current();
2939 if (!mem_cgroup_is_root(memcg
)) {
2940 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2942 page
->mem_cgroup
= memcg
;
2943 __SetPageKmemcg(page
);
2946 css_put(&memcg
->css
);
2951 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2952 * @memcg: memcg to uncharge
2953 * @nr_pages: number of pages to uncharge
2955 void __memcg_kmem_uncharge_memcg(struct mem_cgroup
*memcg
,
2956 unsigned int nr_pages
)
2958 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2959 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2961 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2962 if (do_memsw_account())
2963 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2966 * __memcg_kmem_uncharge: uncharge a kmem page
2967 * @page: page to uncharge
2968 * @order: allocation order
2970 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2972 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2973 unsigned int nr_pages
= 1 << order
;
2978 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2979 __memcg_kmem_uncharge_memcg(memcg
, nr_pages
);
2980 page
->mem_cgroup
= NULL
;
2982 /* slab pages do not have PageKmemcg flag set */
2983 if (PageKmemcg(page
))
2984 __ClearPageKmemcg(page
);
2986 css_put_many(&memcg
->css
, nr_pages
);
2988 #endif /* CONFIG_MEMCG_KMEM */
2990 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2993 * Because tail pages are not marked as "used", set it. We're under
2994 * pgdat->lru_lock and migration entries setup in all page mappings.
2996 void mem_cgroup_split_huge_fixup(struct page
*head
)
3000 if (mem_cgroup_disabled())
3003 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
3004 head
[i
].mem_cgroup
= head
->mem_cgroup
;
3006 __mod_memcg_state(head
->mem_cgroup
, MEMCG_RSS_HUGE
, -HPAGE_PMD_NR
);
3008 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3010 #ifdef CONFIG_MEMCG_SWAP
3012 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3013 * @entry: swap entry to be moved
3014 * @from: mem_cgroup which the entry is moved from
3015 * @to: mem_cgroup which the entry is moved to
3017 * It succeeds only when the swap_cgroup's record for this entry is the same
3018 * as the mem_cgroup's id of @from.
3020 * Returns 0 on success, -EINVAL on failure.
3022 * The caller must have charged to @to, IOW, called page_counter_charge() about
3023 * both res and memsw, and called css_get().
3025 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3026 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3028 unsigned short old_id
, new_id
;
3030 old_id
= mem_cgroup_id(from
);
3031 new_id
= mem_cgroup_id(to
);
3033 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3034 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3035 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3041 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3042 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3048 static DEFINE_MUTEX(memcg_max_mutex
);
3050 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3051 unsigned long max
, bool memsw
)
3053 bool enlarge
= false;
3054 bool drained
= false;
3056 bool limits_invariant
;
3057 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3060 if (signal_pending(current
)) {
3065 mutex_lock(&memcg_max_mutex
);
3067 * Make sure that the new limit (memsw or memory limit) doesn't
3068 * break our basic invariant rule memory.max <= memsw.max.
3070 limits_invariant
= memsw
? max
>= memcg
->memory
.max
:
3071 max
<= memcg
->memsw
.max
;
3072 if (!limits_invariant
) {
3073 mutex_unlock(&memcg_max_mutex
);
3077 if (max
> counter
->max
)
3079 ret
= page_counter_set_max(counter
, max
);
3080 mutex_unlock(&memcg_max_mutex
);
3086 drain_all_stock(memcg
);
3091 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3092 GFP_KERNEL
, !memsw
)) {
3098 if (!ret
&& enlarge
)
3099 memcg_oom_recover(memcg
);
3104 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3106 unsigned long *total_scanned
)
3108 unsigned long nr_reclaimed
= 0;
3109 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3110 unsigned long reclaimed
;
3112 struct mem_cgroup_tree_per_node
*mctz
;
3113 unsigned long excess
;
3114 unsigned long nr_scanned
;
3119 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3122 * Do not even bother to check the largest node if the root
3123 * is empty. Do it lockless to prevent lock bouncing. Races
3124 * are acceptable as soft limit is best effort anyway.
3126 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3130 * This loop can run a while, specially if mem_cgroup's continuously
3131 * keep exceeding their soft limit and putting the system under
3138 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3143 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3144 gfp_mask
, &nr_scanned
);
3145 nr_reclaimed
+= reclaimed
;
3146 *total_scanned
+= nr_scanned
;
3147 spin_lock_irq(&mctz
->lock
);
3148 __mem_cgroup_remove_exceeded(mz
, mctz
);
3151 * If we failed to reclaim anything from this memory cgroup
3152 * it is time to move on to the next cgroup
3156 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3158 excess
= soft_limit_excess(mz
->memcg
);
3160 * One school of thought says that we should not add
3161 * back the node to the tree if reclaim returns 0.
3162 * But our reclaim could return 0, simply because due
3163 * to priority we are exposing a smaller subset of
3164 * memory to reclaim from. Consider this as a longer
3167 /* If excess == 0, no tree ops */
3168 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3169 spin_unlock_irq(&mctz
->lock
);
3170 css_put(&mz
->memcg
->css
);
3173 * Could not reclaim anything and there are no more
3174 * mem cgroups to try or we seem to be looping without
3175 * reclaiming anything.
3177 if (!nr_reclaimed
&&
3179 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3181 } while (!nr_reclaimed
);
3183 css_put(&next_mz
->memcg
->css
);
3184 return nr_reclaimed
;
3188 * Test whether @memcg has children, dead or alive. Note that this
3189 * function doesn't care whether @memcg has use_hierarchy enabled and
3190 * returns %true if there are child csses according to the cgroup
3191 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3193 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3198 ret
= css_next_child(NULL
, &memcg
->css
);
3204 * Reclaims as many pages from the given memcg as possible.
3206 * Caller is responsible for holding css reference for memcg.
3208 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3210 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3212 /* we call try-to-free pages for make this cgroup empty */
3213 lru_add_drain_all();
3215 drain_all_stock(memcg
);
3217 /* try to free all pages in this cgroup */
3218 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3221 if (signal_pending(current
))
3224 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3228 /* maybe some writeback is necessary */
3229 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3237 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3238 char *buf
, size_t nbytes
,
3241 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3243 if (mem_cgroup_is_root(memcg
))
3245 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3248 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3251 return mem_cgroup_from_css(css
)->use_hierarchy
;
3254 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3255 struct cftype
*cft
, u64 val
)
3258 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3259 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3261 if (memcg
->use_hierarchy
== val
)
3265 * If parent's use_hierarchy is set, we can't make any modifications
3266 * in the child subtrees. If it is unset, then the change can
3267 * occur, provided the current cgroup has no children.
3269 * For the root cgroup, parent_mem is NULL, we allow value to be
3270 * set if there are no children.
3272 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3273 (val
== 1 || val
== 0)) {
3274 if (!memcg_has_children(memcg
))
3275 memcg
->use_hierarchy
= val
;
3284 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3288 if (mem_cgroup_is_root(memcg
)) {
3289 val
= memcg_page_state(memcg
, MEMCG_CACHE
) +
3290 memcg_page_state(memcg
, MEMCG_RSS
);
3292 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3295 val
= page_counter_read(&memcg
->memory
);
3297 val
= page_counter_read(&memcg
->memsw
);
3310 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3313 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3314 struct page_counter
*counter
;
3316 switch (MEMFILE_TYPE(cft
->private)) {
3318 counter
= &memcg
->memory
;
3321 counter
= &memcg
->memsw
;
3324 counter
= &memcg
->kmem
;
3327 counter
= &memcg
->tcpmem
;
3333 switch (MEMFILE_ATTR(cft
->private)) {
3335 if (counter
== &memcg
->memory
)
3336 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3337 if (counter
== &memcg
->memsw
)
3338 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3339 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3341 return (u64
)counter
->max
* PAGE_SIZE
;
3343 return (u64
)counter
->watermark
* PAGE_SIZE
;
3345 return counter
->failcnt
;
3346 case RES_SOFT_LIMIT
:
3347 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3353 static void memcg_flush_percpu_vmstats(struct mem_cgroup
*memcg
)
3355 unsigned long stat
[MEMCG_NR_STAT
] = {0};
3356 struct mem_cgroup
*mi
;
3359 for_each_online_cpu(cpu
)
3360 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3361 stat
[i
] += per_cpu(memcg
->vmstats_percpu
->stat
[i
], cpu
);
3363 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3364 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
3365 atomic_long_add(stat
[i
], &mi
->vmstats
[i
]);
3367 for_each_node(node
) {
3368 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
3369 struct mem_cgroup_per_node
*pi
;
3371 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3374 for_each_online_cpu(cpu
)
3375 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3377 pn
->lruvec_stat_cpu
->count
[i
], cpu
);
3379 for (pi
= pn
; pi
; pi
= parent_nodeinfo(pi
, node
))
3380 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++)
3381 atomic_long_add(stat
[i
], &pi
->lruvec_stat
[i
]);
3385 static void memcg_flush_percpu_vmevents(struct mem_cgroup
*memcg
)
3387 unsigned long events
[NR_VM_EVENT_ITEMS
];
3388 struct mem_cgroup
*mi
;
3391 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3394 for_each_online_cpu(cpu
)
3395 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3396 events
[i
] += per_cpu(memcg
->vmstats_percpu
->events
[i
],
3399 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
))
3400 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++)
3401 atomic_long_add(events
[i
], &mi
->vmevents
[i
]);
3404 #ifdef CONFIG_MEMCG_KMEM
3405 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3409 if (cgroup_memory_nokmem
)
3412 BUG_ON(memcg
->kmemcg_id
>= 0);
3413 BUG_ON(memcg
->kmem_state
);
3415 memcg_id
= memcg_alloc_cache_id();
3419 static_branch_inc(&memcg_kmem_enabled_key
);
3421 * A memory cgroup is considered kmem-online as soon as it gets
3422 * kmemcg_id. Setting the id after enabling static branching will
3423 * guarantee no one starts accounting before all call sites are
3426 memcg
->kmemcg_id
= memcg_id
;
3427 memcg
->kmem_state
= KMEM_ONLINE
;
3428 INIT_LIST_HEAD(&memcg
->kmem_caches
);
3433 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3435 struct cgroup_subsys_state
*css
;
3436 struct mem_cgroup
*parent
, *child
;
3439 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3442 * Clear the online state before clearing memcg_caches array
3443 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3444 * guarantees that no cache will be created for this cgroup
3445 * after we are done (see memcg_create_kmem_cache()).
3447 memcg
->kmem_state
= KMEM_ALLOCATED
;
3449 parent
= parent_mem_cgroup(memcg
);
3451 parent
= root_mem_cgroup
;
3454 * Deactivate and reparent kmem_caches.
3456 memcg_deactivate_kmem_caches(memcg
, parent
);
3458 kmemcg_id
= memcg
->kmemcg_id
;
3459 BUG_ON(kmemcg_id
< 0);
3462 * Change kmemcg_id of this cgroup and all its descendants to the
3463 * parent's id, and then move all entries from this cgroup's list_lrus
3464 * to ones of the parent. After we have finished, all list_lrus
3465 * corresponding to this cgroup are guaranteed to remain empty. The
3466 * ordering is imposed by list_lru_node->lock taken by
3467 * memcg_drain_all_list_lrus().
3469 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3470 css_for_each_descendant_pre(css
, &memcg
->css
) {
3471 child
= mem_cgroup_from_css(css
);
3472 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3473 child
->kmemcg_id
= parent
->kmemcg_id
;
3474 if (!memcg
->use_hierarchy
)
3479 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3481 memcg_free_cache_id(kmemcg_id
);
3484 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3486 /* css_alloc() failed, offlining didn't happen */
3487 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3488 memcg_offline_kmem(memcg
);
3490 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
3491 WARN_ON(!list_empty(&memcg
->kmem_caches
));
3492 static_branch_dec(&memcg_kmem_enabled_key
);
3496 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3500 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3503 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3506 #endif /* CONFIG_MEMCG_KMEM */
3508 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3513 mutex_lock(&memcg_max_mutex
);
3514 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3515 mutex_unlock(&memcg_max_mutex
);
3519 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3523 mutex_lock(&memcg_max_mutex
);
3525 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3529 if (!memcg
->tcpmem_active
) {
3531 * The active flag needs to be written after the static_key
3532 * update. This is what guarantees that the socket activation
3533 * function is the last one to run. See mem_cgroup_sk_alloc()
3534 * for details, and note that we don't mark any socket as
3535 * belonging to this memcg until that flag is up.
3537 * We need to do this, because static_keys will span multiple
3538 * sites, but we can't control their order. If we mark a socket
3539 * as accounted, but the accounting functions are not patched in
3540 * yet, we'll lose accounting.
3542 * We never race with the readers in mem_cgroup_sk_alloc(),
3543 * because when this value change, the code to process it is not
3546 static_branch_inc(&memcg_sockets_enabled_key
);
3547 memcg
->tcpmem_active
= true;
3550 mutex_unlock(&memcg_max_mutex
);
3555 * The user of this function is...
3558 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3559 char *buf
, size_t nbytes
, loff_t off
)
3561 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3562 unsigned long nr_pages
;
3565 buf
= strstrip(buf
);
3566 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3570 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3572 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3576 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3578 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3581 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3584 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3585 "Please report your usecase to linux-mm@kvack.org if you "
3586 "depend on this functionality.\n");
3587 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3590 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3594 case RES_SOFT_LIMIT
:
3595 memcg
->soft_limit
= nr_pages
;
3599 return ret
?: nbytes
;
3602 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3603 size_t nbytes
, loff_t off
)
3605 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3606 struct page_counter
*counter
;
3608 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3610 counter
= &memcg
->memory
;
3613 counter
= &memcg
->memsw
;
3616 counter
= &memcg
->kmem
;
3619 counter
= &memcg
->tcpmem
;
3625 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3627 page_counter_reset_watermark(counter
);
3630 counter
->failcnt
= 0;
3639 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3642 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3646 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3647 struct cftype
*cft
, u64 val
)
3649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3651 if (val
& ~MOVE_MASK
)
3655 * No kind of locking is needed in here, because ->can_attach() will
3656 * check this value once in the beginning of the process, and then carry
3657 * on with stale data. This means that changes to this value will only
3658 * affect task migrations starting after the change.
3660 memcg
->move_charge_at_immigrate
= val
;
3664 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3665 struct cftype
*cft
, u64 val
)
3673 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3674 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3675 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3677 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3678 int nid
, unsigned int lru_mask
)
3680 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3681 unsigned long nr
= 0;
3684 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3687 if (!(BIT(lru
) & lru_mask
))
3689 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3694 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3695 unsigned int lru_mask
)
3697 unsigned long nr
= 0;
3701 if (!(BIT(lru
) & lru_mask
))
3703 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3708 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3712 unsigned int lru_mask
;
3715 static const struct numa_stat stats
[] = {
3716 { "total", LRU_ALL
},
3717 { "file", LRU_ALL_FILE
},
3718 { "anon", LRU_ALL_ANON
},
3719 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3721 const struct numa_stat
*stat
;
3724 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3726 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3727 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3728 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3729 for_each_node_state(nid
, N_MEMORY
) {
3730 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3732 seq_printf(m
, " N%d=%lu", nid
, nr
);
3737 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3738 struct mem_cgroup
*iter
;
3741 for_each_mem_cgroup_tree(iter
, memcg
)
3742 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3743 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3744 for_each_node_state(nid
, N_MEMORY
) {
3746 for_each_mem_cgroup_tree(iter
, memcg
)
3747 nr
+= mem_cgroup_node_nr_lru_pages(
3748 iter
, nid
, stat
->lru_mask
);
3749 seq_printf(m
, " N%d=%lu", nid
, nr
);
3756 #endif /* CONFIG_NUMA */
3758 static const unsigned int memcg1_stats
[] = {
3769 static const char *const memcg1_stat_names
[] = {
3780 /* Universal VM events cgroup1 shows, original sort order */
3781 static const unsigned int memcg1_events
[] = {
3788 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3790 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3791 unsigned long memory
, memsw
;
3792 struct mem_cgroup
*mi
;
3795 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3797 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3798 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3800 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
],
3801 memcg_page_state_local(memcg
, memcg1_stats
[i
]) *
3805 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3806 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
3807 memcg_events_local(memcg
, memcg1_events
[i
]));
3809 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3810 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
3811 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3814 /* Hierarchical information */
3815 memory
= memsw
= PAGE_COUNTER_MAX
;
3816 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3817 memory
= min(memory
, mi
->memory
.max
);
3818 memsw
= min(memsw
, mi
->memsw
.max
);
3820 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3821 (u64
)memory
* PAGE_SIZE
);
3822 if (do_memsw_account())
3823 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3824 (u64
)memsw
* PAGE_SIZE
);
3826 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3827 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3829 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
3830 (u64
)memcg_page_state(memcg
, memcg1_stats
[i
]) *
3834 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3835 seq_printf(m
, "total_%s %llu\n",
3836 vm_event_name(memcg1_events
[i
]),
3837 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
3839 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3840 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
3841 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
3844 #ifdef CONFIG_DEBUG_VM
3847 struct mem_cgroup_per_node
*mz
;
3848 struct zone_reclaim_stat
*rstat
;
3849 unsigned long recent_rotated
[2] = {0, 0};
3850 unsigned long recent_scanned
[2] = {0, 0};
3852 for_each_online_pgdat(pgdat
) {
3853 mz
= mem_cgroup_nodeinfo(memcg
, pgdat
->node_id
);
3854 rstat
= &mz
->lruvec
.reclaim_stat
;
3856 recent_rotated
[0] += rstat
->recent_rotated
[0];
3857 recent_rotated
[1] += rstat
->recent_rotated
[1];
3858 recent_scanned
[0] += rstat
->recent_scanned
[0];
3859 recent_scanned
[1] += rstat
->recent_scanned
[1];
3861 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3862 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3863 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3864 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3871 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3874 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3876 return mem_cgroup_swappiness(memcg
);
3879 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3880 struct cftype
*cft
, u64 val
)
3882 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3888 memcg
->swappiness
= val
;
3890 vm_swappiness
= val
;
3895 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3897 struct mem_cgroup_threshold_ary
*t
;
3898 unsigned long usage
;
3903 t
= rcu_dereference(memcg
->thresholds
.primary
);
3905 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3910 usage
= mem_cgroup_usage(memcg
, swap
);
3913 * current_threshold points to threshold just below or equal to usage.
3914 * If it's not true, a threshold was crossed after last
3915 * call of __mem_cgroup_threshold().
3917 i
= t
->current_threshold
;
3920 * Iterate backward over array of thresholds starting from
3921 * current_threshold and check if a threshold is crossed.
3922 * If none of thresholds below usage is crossed, we read
3923 * only one element of the array here.
3925 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3926 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3928 /* i = current_threshold + 1 */
3932 * Iterate forward over array of thresholds starting from
3933 * current_threshold+1 and check if a threshold is crossed.
3934 * If none of thresholds above usage is crossed, we read
3935 * only one element of the array here.
3937 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3938 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3940 /* Update current_threshold */
3941 t
->current_threshold
= i
- 1;
3946 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3949 __mem_cgroup_threshold(memcg
, false);
3950 if (do_memsw_account())
3951 __mem_cgroup_threshold(memcg
, true);
3953 memcg
= parent_mem_cgroup(memcg
);
3957 static int compare_thresholds(const void *a
, const void *b
)
3959 const struct mem_cgroup_threshold
*_a
= a
;
3960 const struct mem_cgroup_threshold
*_b
= b
;
3962 if (_a
->threshold
> _b
->threshold
)
3965 if (_a
->threshold
< _b
->threshold
)
3971 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3973 struct mem_cgroup_eventfd_list
*ev
;
3975 spin_lock(&memcg_oom_lock
);
3977 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3978 eventfd_signal(ev
->eventfd
, 1);
3980 spin_unlock(&memcg_oom_lock
);
3984 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3986 struct mem_cgroup
*iter
;
3988 for_each_mem_cgroup_tree(iter
, memcg
)
3989 mem_cgroup_oom_notify_cb(iter
);
3992 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3993 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3995 struct mem_cgroup_thresholds
*thresholds
;
3996 struct mem_cgroup_threshold_ary
*new;
3997 unsigned long threshold
;
3998 unsigned long usage
;
4001 ret
= page_counter_memparse(args
, "-1", &threshold
);
4005 mutex_lock(&memcg
->thresholds_lock
);
4008 thresholds
= &memcg
->thresholds
;
4009 usage
= mem_cgroup_usage(memcg
, false);
4010 } else if (type
== _MEMSWAP
) {
4011 thresholds
= &memcg
->memsw_thresholds
;
4012 usage
= mem_cgroup_usage(memcg
, true);
4016 /* Check if a threshold crossed before adding a new one */
4017 if (thresholds
->primary
)
4018 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4020 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4022 /* Allocate memory for new array of thresholds */
4023 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4030 /* Copy thresholds (if any) to new array */
4031 if (thresholds
->primary
) {
4032 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4033 sizeof(struct mem_cgroup_threshold
));
4036 /* Add new threshold */
4037 new->entries
[size
- 1].eventfd
= eventfd
;
4038 new->entries
[size
- 1].threshold
= threshold
;
4040 /* Sort thresholds. Registering of new threshold isn't time-critical */
4041 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4042 compare_thresholds
, NULL
);
4044 /* Find current threshold */
4045 new->current_threshold
= -1;
4046 for (i
= 0; i
< size
; i
++) {
4047 if (new->entries
[i
].threshold
<= usage
) {
4049 * new->current_threshold will not be used until
4050 * rcu_assign_pointer(), so it's safe to increment
4053 ++new->current_threshold
;
4058 /* Free old spare buffer and save old primary buffer as spare */
4059 kfree(thresholds
->spare
);
4060 thresholds
->spare
= thresholds
->primary
;
4062 rcu_assign_pointer(thresholds
->primary
, new);
4064 /* To be sure that nobody uses thresholds */
4068 mutex_unlock(&memcg
->thresholds_lock
);
4073 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4074 struct eventfd_ctx
*eventfd
, const char *args
)
4076 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4079 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4080 struct eventfd_ctx
*eventfd
, const char *args
)
4082 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4085 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4086 struct eventfd_ctx
*eventfd
, enum res_type type
)
4088 struct mem_cgroup_thresholds
*thresholds
;
4089 struct mem_cgroup_threshold_ary
*new;
4090 unsigned long usage
;
4091 int i
, j
, size
, entries
;
4093 mutex_lock(&memcg
->thresholds_lock
);
4096 thresholds
= &memcg
->thresholds
;
4097 usage
= mem_cgroup_usage(memcg
, false);
4098 } else if (type
== _MEMSWAP
) {
4099 thresholds
= &memcg
->memsw_thresholds
;
4100 usage
= mem_cgroup_usage(memcg
, true);
4104 if (!thresholds
->primary
)
4107 /* Check if a threshold crossed before removing */
4108 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4110 /* Calculate new number of threshold */
4112 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4113 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4119 new = thresholds
->spare
;
4121 /* If no items related to eventfd have been cleared, nothing to do */
4125 /* Set thresholds array to NULL if we don't have thresholds */
4134 /* Copy thresholds and find current threshold */
4135 new->current_threshold
= -1;
4136 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4137 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4140 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4141 if (new->entries
[j
].threshold
<= usage
) {
4143 * new->current_threshold will not be used
4144 * until rcu_assign_pointer(), so it's safe to increment
4147 ++new->current_threshold
;
4153 /* Swap primary and spare array */
4154 thresholds
->spare
= thresholds
->primary
;
4156 rcu_assign_pointer(thresholds
->primary
, new);
4158 /* To be sure that nobody uses thresholds */
4161 /* If all events are unregistered, free the spare array */
4163 kfree(thresholds
->spare
);
4164 thresholds
->spare
= NULL
;
4167 mutex_unlock(&memcg
->thresholds_lock
);
4170 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4171 struct eventfd_ctx
*eventfd
)
4173 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4176 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4177 struct eventfd_ctx
*eventfd
)
4179 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4182 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4183 struct eventfd_ctx
*eventfd
, const char *args
)
4185 struct mem_cgroup_eventfd_list
*event
;
4187 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4191 spin_lock(&memcg_oom_lock
);
4193 event
->eventfd
= eventfd
;
4194 list_add(&event
->list
, &memcg
->oom_notify
);
4196 /* already in OOM ? */
4197 if (memcg
->under_oom
)
4198 eventfd_signal(eventfd
, 1);
4199 spin_unlock(&memcg_oom_lock
);
4204 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4205 struct eventfd_ctx
*eventfd
)
4207 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4209 spin_lock(&memcg_oom_lock
);
4211 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4212 if (ev
->eventfd
== eventfd
) {
4213 list_del(&ev
->list
);
4218 spin_unlock(&memcg_oom_lock
);
4221 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4223 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4225 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4226 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4227 seq_printf(sf
, "oom_kill %lu\n",
4228 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4232 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4233 struct cftype
*cft
, u64 val
)
4235 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4237 /* cannot set to root cgroup and only 0 and 1 are allowed */
4238 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4241 memcg
->oom_kill_disable
= val
;
4243 memcg_oom_recover(memcg
);
4248 #ifdef CONFIG_CGROUP_WRITEBACK
4250 #include <trace/events/writeback.h>
4252 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4254 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4257 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4259 wb_domain_exit(&memcg
->cgwb_domain
);
4262 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4264 wb_domain_size_changed(&memcg
->cgwb_domain
);
4267 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4269 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4271 if (!memcg
->css
.parent
)
4274 return &memcg
->cgwb_domain
;
4278 * idx can be of type enum memcg_stat_item or node_stat_item.
4279 * Keep in sync with memcg_exact_page().
4281 static unsigned long memcg_exact_page_state(struct mem_cgroup
*memcg
, int idx
)
4283 long x
= atomic_long_read(&memcg
->vmstats
[idx
]);
4286 for_each_online_cpu(cpu
)
4287 x
+= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
)->stat
[idx
];
4294 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4295 * @wb: bdi_writeback in question
4296 * @pfilepages: out parameter for number of file pages
4297 * @pheadroom: out parameter for number of allocatable pages according to memcg
4298 * @pdirty: out parameter for number of dirty pages
4299 * @pwriteback: out parameter for number of pages under writeback
4301 * Determine the numbers of file, headroom, dirty, and writeback pages in
4302 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4303 * is a bit more involved.
4305 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4306 * headroom is calculated as the lowest headroom of itself and the
4307 * ancestors. Note that this doesn't consider the actual amount of
4308 * available memory in the system. The caller should further cap
4309 * *@pheadroom accordingly.
4311 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4312 unsigned long *pheadroom
, unsigned long *pdirty
,
4313 unsigned long *pwriteback
)
4315 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4316 struct mem_cgroup
*parent
;
4318 *pdirty
= memcg_exact_page_state(memcg
, NR_FILE_DIRTY
);
4320 /* this should eventually include NR_UNSTABLE_NFS */
4321 *pwriteback
= memcg_exact_page_state(memcg
, NR_WRITEBACK
);
4322 *pfilepages
= memcg_exact_page_state(memcg
, NR_INACTIVE_FILE
) +
4323 memcg_exact_page_state(memcg
, NR_ACTIVE_FILE
);
4324 *pheadroom
= PAGE_COUNTER_MAX
;
4326 while ((parent
= parent_mem_cgroup(memcg
))) {
4327 unsigned long ceiling
= min(memcg
->memory
.max
, memcg
->high
);
4328 unsigned long used
= page_counter_read(&memcg
->memory
);
4330 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4336 * Foreign dirty flushing
4338 * There's an inherent mismatch between memcg and writeback. The former
4339 * trackes ownership per-page while the latter per-inode. This was a
4340 * deliberate design decision because honoring per-page ownership in the
4341 * writeback path is complicated, may lead to higher CPU and IO overheads
4342 * and deemed unnecessary given that write-sharing an inode across
4343 * different cgroups isn't a common use-case.
4345 * Combined with inode majority-writer ownership switching, this works well
4346 * enough in most cases but there are some pathological cases. For
4347 * example, let's say there are two cgroups A and B which keep writing to
4348 * different but confined parts of the same inode. B owns the inode and
4349 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4350 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4351 * triggering background writeback. A will be slowed down without a way to
4352 * make writeback of the dirty pages happen.
4354 * Conditions like the above can lead to a cgroup getting repatedly and
4355 * severely throttled after making some progress after each
4356 * dirty_expire_interval while the underyling IO device is almost
4359 * Solving this problem completely requires matching the ownership tracking
4360 * granularities between memcg and writeback in either direction. However,
4361 * the more egregious behaviors can be avoided by simply remembering the
4362 * most recent foreign dirtying events and initiating remote flushes on
4363 * them when local writeback isn't enough to keep the memory clean enough.
4365 * The following two functions implement such mechanism. When a foreign
4366 * page - a page whose memcg and writeback ownerships don't match - is
4367 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4368 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4369 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4370 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4371 * foreign bdi_writebacks which haven't expired. Both the numbers of
4372 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4373 * limited to MEMCG_CGWB_FRN_CNT.
4375 * The mechanism only remembers IDs and doesn't hold any object references.
4376 * As being wrong occasionally doesn't matter, updates and accesses to the
4377 * records are lockless and racy.
4379 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4380 struct bdi_writeback
*wb
)
4382 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
4383 struct memcg_cgwb_frn
*frn
;
4384 u64 now
= get_jiffies_64();
4385 u64 oldest_at
= now
;
4389 trace_track_foreign_dirty(page
, wb
);
4392 * Pick the slot to use. If there is already a slot for @wb, keep
4393 * using it. If not replace the oldest one which isn't being
4396 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4397 frn
= &memcg
->cgwb_frn
[i
];
4398 if (frn
->bdi_id
== wb
->bdi
->id
&&
4399 frn
->memcg_id
== wb
->memcg_css
->id
)
4401 if (time_before64(frn
->at
, oldest_at
) &&
4402 atomic_read(&frn
->done
.cnt
) == 1) {
4404 oldest_at
= frn
->at
;
4408 if (i
< MEMCG_CGWB_FRN_CNT
) {
4410 * Re-using an existing one. Update timestamp lazily to
4411 * avoid making the cacheline hot. We want them to be
4412 * reasonably up-to-date and significantly shorter than
4413 * dirty_expire_interval as that's what expires the record.
4414 * Use the shorter of 1s and dirty_expire_interval / 8.
4416 unsigned long update_intv
=
4417 min_t(unsigned long, HZ
,
4418 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4420 if (time_before64(frn
->at
, now
- update_intv
))
4422 } else if (oldest
>= 0) {
4423 /* replace the oldest free one */
4424 frn
= &memcg
->cgwb_frn
[oldest
];
4425 frn
->bdi_id
= wb
->bdi
->id
;
4426 frn
->memcg_id
= wb
->memcg_css
->id
;
4431 /* issue foreign writeback flushes for recorded foreign dirtying events */
4432 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4434 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4435 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4436 u64 now
= jiffies_64
;
4439 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4440 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4443 * If the record is older than dirty_expire_interval,
4444 * writeback on it has already started. No need to kick it
4445 * off again. Also, don't start a new one if there's
4446 * already one in flight.
4448 if (time_after64(frn
->at
, now
- intv
) &&
4449 atomic_read(&frn
->done
.cnt
) == 1) {
4451 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4452 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
, 0,
4453 WB_REASON_FOREIGN_FLUSH
,
4459 #else /* CONFIG_CGROUP_WRITEBACK */
4461 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4466 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4470 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4474 #endif /* CONFIG_CGROUP_WRITEBACK */
4477 * DO NOT USE IN NEW FILES.
4479 * "cgroup.event_control" implementation.
4481 * This is way over-engineered. It tries to support fully configurable
4482 * events for each user. Such level of flexibility is completely
4483 * unnecessary especially in the light of the planned unified hierarchy.
4485 * Please deprecate this and replace with something simpler if at all
4490 * Unregister event and free resources.
4492 * Gets called from workqueue.
4494 static void memcg_event_remove(struct work_struct
*work
)
4496 struct mem_cgroup_event
*event
=
4497 container_of(work
, struct mem_cgroup_event
, remove
);
4498 struct mem_cgroup
*memcg
= event
->memcg
;
4500 remove_wait_queue(event
->wqh
, &event
->wait
);
4502 event
->unregister_event(memcg
, event
->eventfd
);
4504 /* Notify userspace the event is going away. */
4505 eventfd_signal(event
->eventfd
, 1);
4507 eventfd_ctx_put(event
->eventfd
);
4509 css_put(&memcg
->css
);
4513 * Gets called on EPOLLHUP on eventfd when user closes it.
4515 * Called with wqh->lock held and interrupts disabled.
4517 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4518 int sync
, void *key
)
4520 struct mem_cgroup_event
*event
=
4521 container_of(wait
, struct mem_cgroup_event
, wait
);
4522 struct mem_cgroup
*memcg
= event
->memcg
;
4523 __poll_t flags
= key_to_poll(key
);
4525 if (flags
& EPOLLHUP
) {
4527 * If the event has been detached at cgroup removal, we
4528 * can simply return knowing the other side will cleanup
4531 * We can't race against event freeing since the other
4532 * side will require wqh->lock via remove_wait_queue(),
4535 spin_lock(&memcg
->event_list_lock
);
4536 if (!list_empty(&event
->list
)) {
4537 list_del_init(&event
->list
);
4539 * We are in atomic context, but cgroup_event_remove()
4540 * may sleep, so we have to call it in workqueue.
4542 schedule_work(&event
->remove
);
4544 spin_unlock(&memcg
->event_list_lock
);
4550 static void memcg_event_ptable_queue_proc(struct file
*file
,
4551 wait_queue_head_t
*wqh
, poll_table
*pt
)
4553 struct mem_cgroup_event
*event
=
4554 container_of(pt
, struct mem_cgroup_event
, pt
);
4557 add_wait_queue(wqh
, &event
->wait
);
4561 * DO NOT USE IN NEW FILES.
4563 * Parse input and register new cgroup event handler.
4565 * Input must be in format '<event_fd> <control_fd> <args>'.
4566 * Interpretation of args is defined by control file implementation.
4568 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4569 char *buf
, size_t nbytes
, loff_t off
)
4571 struct cgroup_subsys_state
*css
= of_css(of
);
4572 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4573 struct mem_cgroup_event
*event
;
4574 struct cgroup_subsys_state
*cfile_css
;
4575 unsigned int efd
, cfd
;
4582 buf
= strstrip(buf
);
4584 efd
= simple_strtoul(buf
, &endp
, 10);
4589 cfd
= simple_strtoul(buf
, &endp
, 10);
4590 if ((*endp
!= ' ') && (*endp
!= '\0'))
4594 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4598 event
->memcg
= memcg
;
4599 INIT_LIST_HEAD(&event
->list
);
4600 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4601 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4602 INIT_WORK(&event
->remove
, memcg_event_remove
);
4610 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4611 if (IS_ERR(event
->eventfd
)) {
4612 ret
= PTR_ERR(event
->eventfd
);
4619 goto out_put_eventfd
;
4622 /* the process need read permission on control file */
4623 /* AV: shouldn't we check that it's been opened for read instead? */
4624 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4629 * Determine the event callbacks and set them in @event. This used
4630 * to be done via struct cftype but cgroup core no longer knows
4631 * about these events. The following is crude but the whole thing
4632 * is for compatibility anyway.
4634 * DO NOT ADD NEW FILES.
4636 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4638 if (!strcmp(name
, "memory.usage_in_bytes")) {
4639 event
->register_event
= mem_cgroup_usage_register_event
;
4640 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4641 } else if (!strcmp(name
, "memory.oom_control")) {
4642 event
->register_event
= mem_cgroup_oom_register_event
;
4643 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4644 } else if (!strcmp(name
, "memory.pressure_level")) {
4645 event
->register_event
= vmpressure_register_event
;
4646 event
->unregister_event
= vmpressure_unregister_event
;
4647 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4648 event
->register_event
= memsw_cgroup_usage_register_event
;
4649 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4656 * Verify @cfile should belong to @css. Also, remaining events are
4657 * automatically removed on cgroup destruction but the removal is
4658 * asynchronous, so take an extra ref on @css.
4660 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4661 &memory_cgrp_subsys
);
4663 if (IS_ERR(cfile_css
))
4665 if (cfile_css
!= css
) {
4670 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4674 vfs_poll(efile
.file
, &event
->pt
);
4676 spin_lock(&memcg
->event_list_lock
);
4677 list_add(&event
->list
, &memcg
->event_list
);
4678 spin_unlock(&memcg
->event_list_lock
);
4690 eventfd_ctx_put(event
->eventfd
);
4699 static struct cftype mem_cgroup_legacy_files
[] = {
4701 .name
= "usage_in_bytes",
4702 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4703 .read_u64
= mem_cgroup_read_u64
,
4706 .name
= "max_usage_in_bytes",
4707 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4708 .write
= mem_cgroup_reset
,
4709 .read_u64
= mem_cgroup_read_u64
,
4712 .name
= "limit_in_bytes",
4713 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4714 .write
= mem_cgroup_write
,
4715 .read_u64
= mem_cgroup_read_u64
,
4718 .name
= "soft_limit_in_bytes",
4719 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4720 .write
= mem_cgroup_write
,
4721 .read_u64
= mem_cgroup_read_u64
,
4725 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4726 .write
= mem_cgroup_reset
,
4727 .read_u64
= mem_cgroup_read_u64
,
4731 .seq_show
= memcg_stat_show
,
4734 .name
= "force_empty",
4735 .write
= mem_cgroup_force_empty_write
,
4738 .name
= "use_hierarchy",
4739 .write_u64
= mem_cgroup_hierarchy_write
,
4740 .read_u64
= mem_cgroup_hierarchy_read
,
4743 .name
= "cgroup.event_control", /* XXX: for compat */
4744 .write
= memcg_write_event_control
,
4745 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4748 .name
= "swappiness",
4749 .read_u64
= mem_cgroup_swappiness_read
,
4750 .write_u64
= mem_cgroup_swappiness_write
,
4753 .name
= "move_charge_at_immigrate",
4754 .read_u64
= mem_cgroup_move_charge_read
,
4755 .write_u64
= mem_cgroup_move_charge_write
,
4758 .name
= "oom_control",
4759 .seq_show
= mem_cgroup_oom_control_read
,
4760 .write_u64
= mem_cgroup_oom_control_write
,
4761 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4764 .name
= "pressure_level",
4768 .name
= "numa_stat",
4769 .seq_show
= memcg_numa_stat_show
,
4773 .name
= "kmem.limit_in_bytes",
4774 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4775 .write
= mem_cgroup_write
,
4776 .read_u64
= mem_cgroup_read_u64
,
4779 .name
= "kmem.usage_in_bytes",
4780 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4781 .read_u64
= mem_cgroup_read_u64
,
4784 .name
= "kmem.failcnt",
4785 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4786 .write
= mem_cgroup_reset
,
4787 .read_u64
= mem_cgroup_read_u64
,
4790 .name
= "kmem.max_usage_in_bytes",
4791 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4792 .write
= mem_cgroup_reset
,
4793 .read_u64
= mem_cgroup_read_u64
,
4795 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4797 .name
= "kmem.slabinfo",
4798 .seq_start
= memcg_slab_start
,
4799 .seq_next
= memcg_slab_next
,
4800 .seq_stop
= memcg_slab_stop
,
4801 .seq_show
= memcg_slab_show
,
4805 .name
= "kmem.tcp.limit_in_bytes",
4806 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4807 .write
= mem_cgroup_write
,
4808 .read_u64
= mem_cgroup_read_u64
,
4811 .name
= "kmem.tcp.usage_in_bytes",
4812 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4813 .read_u64
= mem_cgroup_read_u64
,
4816 .name
= "kmem.tcp.failcnt",
4817 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4818 .write
= mem_cgroup_reset
,
4819 .read_u64
= mem_cgroup_read_u64
,
4822 .name
= "kmem.tcp.max_usage_in_bytes",
4823 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4824 .write
= mem_cgroup_reset
,
4825 .read_u64
= mem_cgroup_read_u64
,
4827 { }, /* terminate */
4831 * Private memory cgroup IDR
4833 * Swap-out records and page cache shadow entries need to store memcg
4834 * references in constrained space, so we maintain an ID space that is
4835 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4836 * memory-controlled cgroups to 64k.
4838 * However, there usually are many references to the oflline CSS after
4839 * the cgroup has been destroyed, such as page cache or reclaimable
4840 * slab objects, that don't need to hang on to the ID. We want to keep
4841 * those dead CSS from occupying IDs, or we might quickly exhaust the
4842 * relatively small ID space and prevent the creation of new cgroups
4843 * even when there are much fewer than 64k cgroups - possibly none.
4845 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4846 * be freed and recycled when it's no longer needed, which is usually
4847 * when the CSS is offlined.
4849 * The only exception to that are records of swapped out tmpfs/shmem
4850 * pages that need to be attributed to live ancestors on swapin. But
4851 * those references are manageable from userspace.
4854 static DEFINE_IDR(mem_cgroup_idr
);
4856 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
4858 if (memcg
->id
.id
> 0) {
4859 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
4864 static void mem_cgroup_id_get_many(struct mem_cgroup
*memcg
, unsigned int n
)
4866 refcount_add(n
, &memcg
->id
.ref
);
4869 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
4871 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
4872 mem_cgroup_id_remove(memcg
);
4874 /* Memcg ID pins CSS */
4875 css_put(&memcg
->css
);
4879 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
4881 mem_cgroup_id_put_many(memcg
, 1);
4885 * mem_cgroup_from_id - look up a memcg from a memcg id
4886 * @id: the memcg id to look up
4888 * Caller must hold rcu_read_lock().
4890 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
4892 WARN_ON_ONCE(!rcu_read_lock_held());
4893 return idr_find(&mem_cgroup_idr
, id
);
4896 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4898 struct mem_cgroup_per_node
*pn
;
4901 * This routine is called against possible nodes.
4902 * But it's BUG to call kmalloc() against offline node.
4904 * TODO: this routine can waste much memory for nodes which will
4905 * never be onlined. It's better to use memory hotplug callback
4908 if (!node_state(node
, N_NORMAL_MEMORY
))
4910 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4914 pn
->lruvec_stat_local
= alloc_percpu(struct lruvec_stat
);
4915 if (!pn
->lruvec_stat_local
) {
4920 pn
->lruvec_stat_cpu
= alloc_percpu(struct lruvec_stat
);
4921 if (!pn
->lruvec_stat_cpu
) {
4922 free_percpu(pn
->lruvec_stat_local
);
4927 lruvec_init(&pn
->lruvec
);
4928 pn
->usage_in_excess
= 0;
4929 pn
->on_tree
= false;
4932 memcg
->nodeinfo
[node
] = pn
;
4936 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
4938 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
4943 free_percpu(pn
->lruvec_stat_cpu
);
4944 free_percpu(pn
->lruvec_stat_local
);
4948 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4953 free_mem_cgroup_per_node_info(memcg
, node
);
4954 free_percpu(memcg
->vmstats_percpu
);
4955 free_percpu(memcg
->vmstats_local
);
4959 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4961 memcg_wb_domain_exit(memcg
);
4963 * Flush percpu vmstats and vmevents to guarantee the value correctness
4964 * on parent's and all ancestor levels.
4966 memcg_flush_percpu_vmstats(memcg
);
4967 memcg_flush_percpu_vmevents(memcg
);
4968 __mem_cgroup_free(memcg
);
4971 static struct mem_cgroup
*mem_cgroup_alloc(void)
4973 struct mem_cgroup
*memcg
;
4976 int __maybe_unused i
;
4978 size
= sizeof(struct mem_cgroup
);
4979 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4981 memcg
= kzalloc(size
, GFP_KERNEL
);
4985 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
4986 1, MEM_CGROUP_ID_MAX
,
4988 if (memcg
->id
.id
< 0)
4991 memcg
->vmstats_local
= alloc_percpu(struct memcg_vmstats_percpu
);
4992 if (!memcg
->vmstats_local
)
4995 memcg
->vmstats_percpu
= alloc_percpu(struct memcg_vmstats_percpu
);
4996 if (!memcg
->vmstats_percpu
)
5000 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5003 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5006 INIT_WORK(&memcg
->high_work
, high_work_func
);
5007 INIT_LIST_HEAD(&memcg
->oom_notify
);
5008 mutex_init(&memcg
->thresholds_lock
);
5009 spin_lock_init(&memcg
->move_lock
);
5010 vmpressure_init(&memcg
->vmpressure
);
5011 INIT_LIST_HEAD(&memcg
->event_list
);
5012 spin_lock_init(&memcg
->event_list_lock
);
5013 memcg
->socket_pressure
= jiffies
;
5014 #ifdef CONFIG_MEMCG_KMEM
5015 memcg
->kmemcg_id
= -1;
5017 #ifdef CONFIG_CGROUP_WRITEBACK
5018 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5019 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5020 memcg
->cgwb_frn
[i
].done
=
5021 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5023 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5024 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5025 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5026 memcg
->deferred_split_queue
.split_queue_len
= 0;
5028 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5031 mem_cgroup_id_remove(memcg
);
5032 __mem_cgroup_free(memcg
);
5036 static struct cgroup_subsys_state
* __ref
5037 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5039 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5040 struct mem_cgroup
*memcg
;
5041 long error
= -ENOMEM
;
5043 memcg
= mem_cgroup_alloc();
5045 return ERR_PTR(error
);
5047 memcg
->high
= PAGE_COUNTER_MAX
;
5048 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5050 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5051 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5053 if (parent
&& parent
->use_hierarchy
) {
5054 memcg
->use_hierarchy
= true;
5055 page_counter_init(&memcg
->memory
, &parent
->memory
);
5056 page_counter_init(&memcg
->swap
, &parent
->swap
);
5057 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
5058 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5059 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5061 page_counter_init(&memcg
->memory
, NULL
);
5062 page_counter_init(&memcg
->swap
, NULL
);
5063 page_counter_init(&memcg
->memsw
, NULL
);
5064 page_counter_init(&memcg
->kmem
, NULL
);
5065 page_counter_init(&memcg
->tcpmem
, NULL
);
5067 * Deeper hierachy with use_hierarchy == false doesn't make
5068 * much sense so let cgroup subsystem know about this
5069 * unfortunate state in our controller.
5071 if (parent
!= root_mem_cgroup
)
5072 memory_cgrp_subsys
.broken_hierarchy
= true;
5075 /* The following stuff does not apply to the root */
5077 #ifdef CONFIG_MEMCG_KMEM
5078 INIT_LIST_HEAD(&memcg
->kmem_caches
);
5080 root_mem_cgroup
= memcg
;
5084 error
= memcg_online_kmem(memcg
);
5088 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5089 static_branch_inc(&memcg_sockets_enabled_key
);
5093 mem_cgroup_id_remove(memcg
);
5094 mem_cgroup_free(memcg
);
5095 return ERR_PTR(-ENOMEM
);
5098 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5100 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5103 * A memcg must be visible for memcg_expand_shrinker_maps()
5104 * by the time the maps are allocated. So, we allocate maps
5105 * here, when for_each_mem_cgroup() can't skip it.
5107 if (memcg_alloc_shrinker_maps(memcg
)) {
5108 mem_cgroup_id_remove(memcg
);
5112 /* Online state pins memcg ID, memcg ID pins CSS */
5113 refcount_set(&memcg
->id
.ref
, 1);
5118 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5120 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5121 struct mem_cgroup_event
*event
, *tmp
;
5124 * Unregister events and notify userspace.
5125 * Notify userspace about cgroup removing only after rmdir of cgroup
5126 * directory to avoid race between userspace and kernelspace.
5128 spin_lock(&memcg
->event_list_lock
);
5129 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5130 list_del_init(&event
->list
);
5131 schedule_work(&event
->remove
);
5133 spin_unlock(&memcg
->event_list_lock
);
5135 page_counter_set_min(&memcg
->memory
, 0);
5136 page_counter_set_low(&memcg
->memory
, 0);
5138 memcg_offline_kmem(memcg
);
5139 wb_memcg_offline(memcg
);
5141 drain_all_stock(memcg
);
5143 mem_cgroup_id_put(memcg
);
5146 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5148 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5150 invalidate_reclaim_iterators(memcg
);
5153 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5155 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5156 int __maybe_unused i
;
5158 #ifdef CONFIG_CGROUP_WRITEBACK
5159 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5160 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5162 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5163 static_branch_dec(&memcg_sockets_enabled_key
);
5165 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5166 static_branch_dec(&memcg_sockets_enabled_key
);
5168 vmpressure_cleanup(&memcg
->vmpressure
);
5169 cancel_work_sync(&memcg
->high_work
);
5170 mem_cgroup_remove_from_trees(memcg
);
5171 memcg_free_shrinker_maps(memcg
);
5172 memcg_free_kmem(memcg
);
5173 mem_cgroup_free(memcg
);
5177 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5178 * @css: the target css
5180 * Reset the states of the mem_cgroup associated with @css. This is
5181 * invoked when the userland requests disabling on the default hierarchy
5182 * but the memcg is pinned through dependency. The memcg should stop
5183 * applying policies and should revert to the vanilla state as it may be
5184 * made visible again.
5186 * The current implementation only resets the essential configurations.
5187 * This needs to be expanded to cover all the visible parts.
5189 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5191 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5193 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5194 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5195 page_counter_set_max(&memcg
->memsw
, PAGE_COUNTER_MAX
);
5196 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5197 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5198 page_counter_set_min(&memcg
->memory
, 0);
5199 page_counter_set_low(&memcg
->memory
, 0);
5200 memcg
->high
= PAGE_COUNTER_MAX
;
5201 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5202 memcg_wb_domain_size_changed(memcg
);
5206 /* Handlers for move charge at task migration. */
5207 static int mem_cgroup_do_precharge(unsigned long count
)
5211 /* Try a single bulk charge without reclaim first, kswapd may wake */
5212 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5214 mc
.precharge
+= count
;
5218 /* Try charges one by one with reclaim, but do not retry */
5220 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5234 enum mc_target_type
{
5241 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5242 unsigned long addr
, pte_t ptent
)
5244 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5246 if (!page
|| !page_mapped(page
))
5248 if (PageAnon(page
)) {
5249 if (!(mc
.flags
& MOVE_ANON
))
5252 if (!(mc
.flags
& MOVE_FILE
))
5255 if (!get_page_unless_zero(page
))
5261 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5262 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5263 pte_t ptent
, swp_entry_t
*entry
)
5265 struct page
*page
= NULL
;
5266 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5268 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
5272 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5273 * a device and because they are not accessible by CPU they are store
5274 * as special swap entry in the CPU page table.
5276 if (is_device_private_entry(ent
)) {
5277 page
= device_private_entry_to_page(ent
);
5279 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5280 * a refcount of 1 when free (unlike normal page)
5282 if (!page_ref_add_unless(page
, 1, 1))
5288 * Because lookup_swap_cache() updates some statistics counter,
5289 * we call find_get_page() with swapper_space directly.
5291 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5292 if (do_memsw_account())
5293 entry
->val
= ent
.val
;
5298 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5299 pte_t ptent
, swp_entry_t
*entry
)
5305 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5306 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5308 struct page
*page
= NULL
;
5309 struct address_space
*mapping
;
5312 if (!vma
->vm_file
) /* anonymous vma */
5314 if (!(mc
.flags
& MOVE_FILE
))
5317 mapping
= vma
->vm_file
->f_mapping
;
5318 pgoff
= linear_page_index(vma
, addr
);
5320 /* page is moved even if it's not RSS of this task(page-faulted). */
5322 /* shmem/tmpfs may report page out on swap: account for that too. */
5323 if (shmem_mapping(mapping
)) {
5324 page
= find_get_entry(mapping
, pgoff
);
5325 if (xa_is_value(page
)) {
5326 swp_entry_t swp
= radix_to_swp_entry(page
);
5327 if (do_memsw_account())
5329 page
= find_get_page(swap_address_space(swp
),
5333 page
= find_get_page(mapping
, pgoff
);
5335 page
= find_get_page(mapping
, pgoff
);
5341 * mem_cgroup_move_account - move account of the page
5343 * @compound: charge the page as compound or small page
5344 * @from: mem_cgroup which the page is moved from.
5345 * @to: mem_cgroup which the page is moved to. @from != @to.
5347 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5349 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5352 static int mem_cgroup_move_account(struct page
*page
,
5354 struct mem_cgroup
*from
,
5355 struct mem_cgroup
*to
)
5357 struct lruvec
*from_vec
, *to_vec
;
5358 struct pglist_data
*pgdat
;
5359 unsigned long flags
;
5360 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5364 VM_BUG_ON(from
== to
);
5365 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5366 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5369 * Prevent mem_cgroup_migrate() from looking at
5370 * page->mem_cgroup of its source page while we change it.
5373 if (!trylock_page(page
))
5377 if (page
->mem_cgroup
!= from
)
5380 anon
= PageAnon(page
);
5382 pgdat
= page_pgdat(page
);
5383 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5384 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5386 spin_lock_irqsave(&from
->move_lock
, flags
);
5388 if (!anon
&& page_mapped(page
)) {
5389 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5390 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5394 * move_lock grabbed above and caller set from->moving_account, so
5395 * mod_memcg_page_state will serialize updates to PageDirty.
5396 * So mapping should be stable for dirty pages.
5398 if (!anon
&& PageDirty(page
)) {
5399 struct address_space
*mapping
= page_mapping(page
);
5401 if (mapping_cap_account_dirty(mapping
)) {
5402 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
, -nr_pages
);
5403 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
, nr_pages
);
5407 if (PageWriteback(page
)) {
5408 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5409 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5413 * It is safe to change page->mem_cgroup here because the page
5414 * is referenced, charged, and isolated - we can't race with
5415 * uncharging, charging, migration, or LRU putback.
5418 /* caller should have done css_get */
5419 page
->mem_cgroup
= to
;
5421 spin_unlock_irqrestore(&from
->move_lock
, flags
);
5425 local_irq_disable();
5426 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
5427 memcg_check_events(to
, page
);
5428 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
5429 memcg_check_events(from
, page
);
5438 * get_mctgt_type - get target type of moving charge
5439 * @vma: the vma the pte to be checked belongs
5440 * @addr: the address corresponding to the pte to be checked
5441 * @ptent: the pte to be checked
5442 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5445 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5446 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5447 * move charge. if @target is not NULL, the page is stored in target->page
5448 * with extra refcnt got(Callers should handle it).
5449 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5450 * target for charge migration. if @target is not NULL, the entry is stored
5452 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5453 * (so ZONE_DEVICE page and thus not on the lru).
5454 * For now we such page is charge like a regular page would be as for all
5455 * intent and purposes it is just special memory taking the place of a
5458 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5460 * Called with pte lock held.
5463 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5464 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5466 struct page
*page
= NULL
;
5467 enum mc_target_type ret
= MC_TARGET_NONE
;
5468 swp_entry_t ent
= { .val
= 0 };
5470 if (pte_present(ptent
))
5471 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5472 else if (is_swap_pte(ptent
))
5473 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5474 else if (pte_none(ptent
))
5475 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5477 if (!page
&& !ent
.val
)
5481 * Do only loose check w/o serialization.
5482 * mem_cgroup_move_account() checks the page is valid or
5483 * not under LRU exclusion.
5485 if (page
->mem_cgroup
== mc
.from
) {
5486 ret
= MC_TARGET_PAGE
;
5487 if (is_device_private_page(page
))
5488 ret
= MC_TARGET_DEVICE
;
5490 target
->page
= page
;
5492 if (!ret
|| !target
)
5496 * There is a swap entry and a page doesn't exist or isn't charged.
5497 * But we cannot move a tail-page in a THP.
5499 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5500 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5501 ret
= MC_TARGET_SWAP
;
5508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5510 * We don't consider PMD mapped swapping or file mapped pages because THP does
5511 * not support them for now.
5512 * Caller should make sure that pmd_trans_huge(pmd) is true.
5514 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5515 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5517 struct page
*page
= NULL
;
5518 enum mc_target_type ret
= MC_TARGET_NONE
;
5520 if (unlikely(is_swap_pmd(pmd
))) {
5521 VM_BUG_ON(thp_migration_supported() &&
5522 !is_pmd_migration_entry(pmd
));
5525 page
= pmd_page(pmd
);
5526 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5527 if (!(mc
.flags
& MOVE_ANON
))
5529 if (page
->mem_cgroup
== mc
.from
) {
5530 ret
= MC_TARGET_PAGE
;
5533 target
->page
= page
;
5539 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5540 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5542 return MC_TARGET_NONE
;
5546 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5547 unsigned long addr
, unsigned long end
,
5548 struct mm_walk
*walk
)
5550 struct vm_area_struct
*vma
= walk
->vma
;
5554 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5557 * Note their can not be MC_TARGET_DEVICE for now as we do not
5558 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5559 * this might change.
5561 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5562 mc
.precharge
+= HPAGE_PMD_NR
;
5567 if (pmd_trans_unstable(pmd
))
5569 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5570 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5571 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5572 mc
.precharge
++; /* increment precharge temporarily */
5573 pte_unmap_unlock(pte
- 1, ptl
);
5579 static const struct mm_walk_ops precharge_walk_ops
= {
5580 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5583 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5585 unsigned long precharge
;
5587 down_read(&mm
->mmap_sem
);
5588 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5589 up_read(&mm
->mmap_sem
);
5591 precharge
= mc
.precharge
;
5597 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5599 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5601 VM_BUG_ON(mc
.moving_task
);
5602 mc
.moving_task
= current
;
5603 return mem_cgroup_do_precharge(precharge
);
5606 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5607 static void __mem_cgroup_clear_mc(void)
5609 struct mem_cgroup
*from
= mc
.from
;
5610 struct mem_cgroup
*to
= mc
.to
;
5612 /* we must uncharge all the leftover precharges from mc.to */
5614 cancel_charge(mc
.to
, mc
.precharge
);
5618 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5619 * we must uncharge here.
5621 if (mc
.moved_charge
) {
5622 cancel_charge(mc
.from
, mc
.moved_charge
);
5623 mc
.moved_charge
= 0;
5625 /* we must fixup refcnts and charges */
5626 if (mc
.moved_swap
) {
5627 /* uncharge swap account from the old cgroup */
5628 if (!mem_cgroup_is_root(mc
.from
))
5629 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5631 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5634 * we charged both to->memory and to->memsw, so we
5635 * should uncharge to->memory.
5637 if (!mem_cgroup_is_root(mc
.to
))
5638 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5640 mem_cgroup_id_get_many(mc
.to
, mc
.moved_swap
);
5641 css_put_many(&mc
.to
->css
, mc
.moved_swap
);
5645 memcg_oom_recover(from
);
5646 memcg_oom_recover(to
);
5647 wake_up_all(&mc
.waitq
);
5650 static void mem_cgroup_clear_mc(void)
5652 struct mm_struct
*mm
= mc
.mm
;
5655 * we must clear moving_task before waking up waiters at the end of
5658 mc
.moving_task
= NULL
;
5659 __mem_cgroup_clear_mc();
5660 spin_lock(&mc
.lock
);
5664 spin_unlock(&mc
.lock
);
5669 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5671 struct cgroup_subsys_state
*css
;
5672 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5673 struct mem_cgroup
*from
;
5674 struct task_struct
*leader
, *p
;
5675 struct mm_struct
*mm
;
5676 unsigned long move_flags
;
5679 /* charge immigration isn't supported on the default hierarchy */
5680 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5684 * Multi-process migrations only happen on the default hierarchy
5685 * where charge immigration is not used. Perform charge
5686 * immigration if @tset contains a leader and whine if there are
5690 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5693 memcg
= mem_cgroup_from_css(css
);
5699 * We are now commited to this value whatever it is. Changes in this
5700 * tunable will only affect upcoming migrations, not the current one.
5701 * So we need to save it, and keep it going.
5703 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5707 from
= mem_cgroup_from_task(p
);
5709 VM_BUG_ON(from
== memcg
);
5711 mm
= get_task_mm(p
);
5714 /* We move charges only when we move a owner of the mm */
5715 if (mm
->owner
== p
) {
5718 VM_BUG_ON(mc
.precharge
);
5719 VM_BUG_ON(mc
.moved_charge
);
5720 VM_BUG_ON(mc
.moved_swap
);
5722 spin_lock(&mc
.lock
);
5726 mc
.flags
= move_flags
;
5727 spin_unlock(&mc
.lock
);
5728 /* We set mc.moving_task later */
5730 ret
= mem_cgroup_precharge_mc(mm
);
5732 mem_cgroup_clear_mc();
5739 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5742 mem_cgroup_clear_mc();
5745 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5746 unsigned long addr
, unsigned long end
,
5747 struct mm_walk
*walk
)
5750 struct vm_area_struct
*vma
= walk
->vma
;
5753 enum mc_target_type target_type
;
5754 union mc_target target
;
5757 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5759 if (mc
.precharge
< HPAGE_PMD_NR
) {
5763 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5764 if (target_type
== MC_TARGET_PAGE
) {
5766 if (!isolate_lru_page(page
)) {
5767 if (!mem_cgroup_move_account(page
, true,
5769 mc
.precharge
-= HPAGE_PMD_NR
;
5770 mc
.moved_charge
+= HPAGE_PMD_NR
;
5772 putback_lru_page(page
);
5775 } else if (target_type
== MC_TARGET_DEVICE
) {
5777 if (!mem_cgroup_move_account(page
, true,
5779 mc
.precharge
-= HPAGE_PMD_NR
;
5780 mc
.moved_charge
+= HPAGE_PMD_NR
;
5788 if (pmd_trans_unstable(pmd
))
5791 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5792 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5793 pte_t ptent
= *(pte
++);
5794 bool device
= false;
5800 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5801 case MC_TARGET_DEVICE
:
5804 case MC_TARGET_PAGE
:
5807 * We can have a part of the split pmd here. Moving it
5808 * can be done but it would be too convoluted so simply
5809 * ignore such a partial THP and keep it in original
5810 * memcg. There should be somebody mapping the head.
5812 if (PageTransCompound(page
))
5814 if (!device
&& isolate_lru_page(page
))
5816 if (!mem_cgroup_move_account(page
, false,
5819 /* we uncharge from mc.from later. */
5823 putback_lru_page(page
);
5824 put
: /* get_mctgt_type() gets the page */
5827 case MC_TARGET_SWAP
:
5829 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5831 /* we fixup refcnts and charges later. */
5839 pte_unmap_unlock(pte
- 1, ptl
);
5844 * We have consumed all precharges we got in can_attach().
5845 * We try charge one by one, but don't do any additional
5846 * charges to mc.to if we have failed in charge once in attach()
5849 ret
= mem_cgroup_do_precharge(1);
5857 static const struct mm_walk_ops charge_walk_ops
= {
5858 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5861 static void mem_cgroup_move_charge(void)
5863 lru_add_drain_all();
5865 * Signal lock_page_memcg() to take the memcg's move_lock
5866 * while we're moving its pages to another memcg. Then wait
5867 * for already started RCU-only updates to finish.
5869 atomic_inc(&mc
.from
->moving_account
);
5872 if (unlikely(!down_read_trylock(&mc
.mm
->mmap_sem
))) {
5874 * Someone who are holding the mmap_sem might be waiting in
5875 * waitq. So we cancel all extra charges, wake up all waiters,
5876 * and retry. Because we cancel precharges, we might not be able
5877 * to move enough charges, but moving charge is a best-effort
5878 * feature anyway, so it wouldn't be a big problem.
5880 __mem_cgroup_clear_mc();
5885 * When we have consumed all precharges and failed in doing
5886 * additional charge, the page walk just aborts.
5888 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
5891 up_read(&mc
.mm
->mmap_sem
);
5892 atomic_dec(&mc
.from
->moving_account
);
5895 static void mem_cgroup_move_task(void)
5898 mem_cgroup_move_charge();
5899 mem_cgroup_clear_mc();
5902 #else /* !CONFIG_MMU */
5903 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5907 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5910 static void mem_cgroup_move_task(void)
5916 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5917 * to verify whether we're attached to the default hierarchy on each mount
5920 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5923 * use_hierarchy is forced on the default hierarchy. cgroup core
5924 * guarantees that @root doesn't have any children, so turning it
5925 * on for the root memcg is enough.
5927 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5928 root_mem_cgroup
->use_hierarchy
= true;
5930 root_mem_cgroup
->use_hierarchy
= false;
5933 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
5935 if (value
== PAGE_COUNTER_MAX
)
5936 seq_puts(m
, "max\n");
5938 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
5943 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5946 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5948 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
5951 static int memory_min_show(struct seq_file
*m
, void *v
)
5953 return seq_puts_memcg_tunable(m
,
5954 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
5957 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
5958 char *buf
, size_t nbytes
, loff_t off
)
5960 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5964 buf
= strstrip(buf
);
5965 err
= page_counter_memparse(buf
, "max", &min
);
5969 page_counter_set_min(&memcg
->memory
, min
);
5974 static int memory_low_show(struct seq_file
*m
, void *v
)
5976 return seq_puts_memcg_tunable(m
,
5977 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
5980 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5981 char *buf
, size_t nbytes
, loff_t off
)
5983 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5987 buf
= strstrip(buf
);
5988 err
= page_counter_memparse(buf
, "max", &low
);
5992 page_counter_set_low(&memcg
->memory
, low
);
5997 static int memory_high_show(struct seq_file
*m
, void *v
)
5999 return seq_puts_memcg_tunable(m
, READ_ONCE(mem_cgroup_from_seq(m
)->high
));
6002 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6003 char *buf
, size_t nbytes
, loff_t off
)
6005 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6006 unsigned int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
6007 bool drained
= false;
6011 buf
= strstrip(buf
);
6012 err
= page_counter_memparse(buf
, "max", &high
);
6019 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6020 unsigned long reclaimed
;
6022 if (nr_pages
<= high
)
6025 if (signal_pending(current
))
6029 drain_all_stock(memcg
);
6034 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6037 if (!reclaimed
&& !nr_retries
--)
6044 static int memory_max_show(struct seq_file
*m
, void *v
)
6046 return seq_puts_memcg_tunable(m
,
6047 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6050 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6051 char *buf
, size_t nbytes
, loff_t off
)
6053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6054 unsigned int nr_reclaims
= MEM_CGROUP_RECLAIM_RETRIES
;
6055 bool drained
= false;
6059 buf
= strstrip(buf
);
6060 err
= page_counter_memparse(buf
, "max", &max
);
6064 xchg(&memcg
->memory
.max
, max
);
6067 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6069 if (nr_pages
<= max
)
6072 if (signal_pending(current
))
6076 drain_all_stock(memcg
);
6082 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6088 memcg_memory_event(memcg
, MEMCG_OOM
);
6089 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6093 memcg_wb_domain_size_changed(memcg
);
6097 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6099 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6100 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6101 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6102 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6103 seq_printf(m
, "oom_kill %lu\n",
6104 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6107 static int memory_events_show(struct seq_file
*m
, void *v
)
6109 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6111 __memory_events_show(m
, memcg
->memory_events
);
6115 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6117 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6119 __memory_events_show(m
, memcg
->memory_events_local
);
6123 static int memory_stat_show(struct seq_file
*m
, void *v
)
6125 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6128 buf
= memory_stat_format(memcg
);
6136 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6138 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6140 seq_printf(m
, "%d\n", memcg
->oom_group
);
6145 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6146 char *buf
, size_t nbytes
, loff_t off
)
6148 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6151 buf
= strstrip(buf
);
6155 ret
= kstrtoint(buf
, 0, &oom_group
);
6159 if (oom_group
!= 0 && oom_group
!= 1)
6162 memcg
->oom_group
= oom_group
;
6167 static struct cftype memory_files
[] = {
6170 .flags
= CFTYPE_NOT_ON_ROOT
,
6171 .read_u64
= memory_current_read
,
6175 .flags
= CFTYPE_NOT_ON_ROOT
,
6176 .seq_show
= memory_min_show
,
6177 .write
= memory_min_write
,
6181 .flags
= CFTYPE_NOT_ON_ROOT
,
6182 .seq_show
= memory_low_show
,
6183 .write
= memory_low_write
,
6187 .flags
= CFTYPE_NOT_ON_ROOT
,
6188 .seq_show
= memory_high_show
,
6189 .write
= memory_high_write
,
6193 .flags
= CFTYPE_NOT_ON_ROOT
,
6194 .seq_show
= memory_max_show
,
6195 .write
= memory_max_write
,
6199 .flags
= CFTYPE_NOT_ON_ROOT
,
6200 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6201 .seq_show
= memory_events_show
,
6204 .name
= "events.local",
6205 .flags
= CFTYPE_NOT_ON_ROOT
,
6206 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6207 .seq_show
= memory_events_local_show
,
6211 .flags
= CFTYPE_NOT_ON_ROOT
,
6212 .seq_show
= memory_stat_show
,
6215 .name
= "oom.group",
6216 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6217 .seq_show
= memory_oom_group_show
,
6218 .write
= memory_oom_group_write
,
6223 struct cgroup_subsys memory_cgrp_subsys
= {
6224 .css_alloc
= mem_cgroup_css_alloc
,
6225 .css_online
= mem_cgroup_css_online
,
6226 .css_offline
= mem_cgroup_css_offline
,
6227 .css_released
= mem_cgroup_css_released
,
6228 .css_free
= mem_cgroup_css_free
,
6229 .css_reset
= mem_cgroup_css_reset
,
6230 .can_attach
= mem_cgroup_can_attach
,
6231 .cancel_attach
= mem_cgroup_cancel_attach
,
6232 .post_attach
= mem_cgroup_move_task
,
6233 .bind
= mem_cgroup_bind
,
6234 .dfl_cftypes
= memory_files
,
6235 .legacy_cftypes
= mem_cgroup_legacy_files
,
6240 * mem_cgroup_protected - check if memory consumption is in the normal range
6241 * @root: the top ancestor of the sub-tree being checked
6242 * @memcg: the memory cgroup to check
6244 * WARNING: This function is not stateless! It can only be used as part
6245 * of a top-down tree iteration, not for isolated queries.
6247 * Returns one of the following:
6248 * MEMCG_PROT_NONE: cgroup memory is not protected
6249 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6250 * an unprotected supply of reclaimable memory from other cgroups.
6251 * MEMCG_PROT_MIN: cgroup memory is protected
6253 * @root is exclusive; it is never protected when looked at directly
6255 * To provide a proper hierarchical behavior, effective memory.min/low values
6256 * are used. Below is the description of how effective memory.low is calculated.
6257 * Effective memory.min values is calculated in the same way.
6259 * Effective memory.low is always equal or less than the original memory.low.
6260 * If there is no memory.low overcommittment (which is always true for
6261 * top-level memory cgroups), these two values are equal.
6262 * Otherwise, it's a part of parent's effective memory.low,
6263 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6264 * memory.low usages, where memory.low usage is the size of actually
6268 * elow = min( memory.low, parent->elow * ------------------ ),
6269 * siblings_low_usage
6271 * | memory.current, if memory.current < memory.low
6276 * Such definition of the effective memory.low provides the expected
6277 * hierarchical behavior: parent's memory.low value is limiting
6278 * children, unprotected memory is reclaimed first and cgroups,
6279 * which are not using their guarantee do not affect actual memory
6282 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6284 * A A/memory.low = 2G, A/memory.current = 6G
6286 * BC DE B/memory.low = 3G B/memory.current = 2G
6287 * C/memory.low = 1G C/memory.current = 2G
6288 * D/memory.low = 0 D/memory.current = 2G
6289 * E/memory.low = 10G E/memory.current = 0
6291 * and the memory pressure is applied, the following memory distribution
6292 * is expected (approximately):
6294 * A/memory.current = 2G
6296 * B/memory.current = 1.3G
6297 * C/memory.current = 0.6G
6298 * D/memory.current = 0
6299 * E/memory.current = 0
6301 * These calculations require constant tracking of the actual low usages
6302 * (see propagate_protected_usage()), as well as recursive calculation of
6303 * effective memory.low values. But as we do call mem_cgroup_protected()
6304 * path for each memory cgroup top-down from the reclaim,
6305 * it's possible to optimize this part, and save calculated elow
6306 * for next usage. This part is intentionally racy, but it's ok,
6307 * as memory.low is a best-effort mechanism.
6309 enum mem_cgroup_protection
mem_cgroup_protected(struct mem_cgroup
*root
,
6310 struct mem_cgroup
*memcg
)
6312 struct mem_cgroup
*parent
;
6313 unsigned long emin
, parent_emin
;
6314 unsigned long elow
, parent_elow
;
6315 unsigned long usage
;
6317 if (mem_cgroup_disabled())
6318 return MEMCG_PROT_NONE
;
6321 root
= root_mem_cgroup
;
6323 return MEMCG_PROT_NONE
;
6325 usage
= page_counter_read(&memcg
->memory
);
6327 return MEMCG_PROT_NONE
;
6329 emin
= memcg
->memory
.min
;
6330 elow
= memcg
->memory
.low
;
6332 parent
= parent_mem_cgroup(memcg
);
6333 /* No parent means a non-hierarchical mode on v1 memcg */
6335 return MEMCG_PROT_NONE
;
6340 parent_emin
= READ_ONCE(parent
->memory
.emin
);
6341 emin
= min(emin
, parent_emin
);
6342 if (emin
&& parent_emin
) {
6343 unsigned long min_usage
, siblings_min_usage
;
6345 min_usage
= min(usage
, memcg
->memory
.min
);
6346 siblings_min_usage
= atomic_long_read(
6347 &parent
->memory
.children_min_usage
);
6349 if (min_usage
&& siblings_min_usage
)
6350 emin
= min(emin
, parent_emin
* min_usage
/
6351 siblings_min_usage
);
6354 parent_elow
= READ_ONCE(parent
->memory
.elow
);
6355 elow
= min(elow
, parent_elow
);
6356 if (elow
&& parent_elow
) {
6357 unsigned long low_usage
, siblings_low_usage
;
6359 low_usage
= min(usage
, memcg
->memory
.low
);
6360 siblings_low_usage
= atomic_long_read(
6361 &parent
->memory
.children_low_usage
);
6363 if (low_usage
&& siblings_low_usage
)
6364 elow
= min(elow
, parent_elow
* low_usage
/
6365 siblings_low_usage
);
6369 memcg
->memory
.emin
= emin
;
6370 memcg
->memory
.elow
= elow
;
6373 return MEMCG_PROT_MIN
;
6374 else if (usage
<= elow
)
6375 return MEMCG_PROT_LOW
;
6377 return MEMCG_PROT_NONE
;
6381 * mem_cgroup_try_charge - try charging a page
6382 * @page: page to charge
6383 * @mm: mm context of the victim
6384 * @gfp_mask: reclaim mode
6385 * @memcgp: charged memcg return
6386 * @compound: charge the page as compound or small page
6388 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6389 * pages according to @gfp_mask if necessary.
6391 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6392 * Otherwise, an error code is returned.
6394 * After page->mapping has been set up, the caller must finalize the
6395 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6396 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6398 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6399 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6402 struct mem_cgroup
*memcg
= NULL
;
6403 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6406 if (mem_cgroup_disabled())
6409 if (PageSwapCache(page
)) {
6411 * Every swap fault against a single page tries to charge the
6412 * page, bail as early as possible. shmem_unuse() encounters
6413 * already charged pages, too. The USED bit is protected by
6414 * the page lock, which serializes swap cache removal, which
6415 * in turn serializes uncharging.
6417 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
6418 if (compound_head(page
)->mem_cgroup
)
6421 if (do_swap_account
) {
6422 swp_entry_t ent
= { .val
= page_private(page
), };
6423 unsigned short id
= lookup_swap_cgroup_id(ent
);
6426 memcg
= mem_cgroup_from_id(id
);
6427 if (memcg
&& !css_tryget_online(&memcg
->css
))
6434 memcg
= get_mem_cgroup_from_mm(mm
);
6436 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6438 css_put(&memcg
->css
);
6444 int mem_cgroup_try_charge_delay(struct page
*page
, struct mm_struct
*mm
,
6445 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
6448 struct mem_cgroup
*memcg
;
6451 ret
= mem_cgroup_try_charge(page
, mm
, gfp_mask
, memcgp
, compound
);
6453 mem_cgroup_throttle_swaprate(memcg
, page_to_nid(page
), gfp_mask
);
6458 * mem_cgroup_commit_charge - commit a page charge
6459 * @page: page to charge
6460 * @memcg: memcg to charge the page to
6461 * @lrucare: page might be on LRU already
6462 * @compound: charge the page as compound or small page
6464 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6465 * after page->mapping has been set up. This must happen atomically
6466 * as part of the page instantiation, i.e. under the page table lock
6467 * for anonymous pages, under the page lock for page and swap cache.
6469 * In addition, the page must not be on the LRU during the commit, to
6470 * prevent racing with task migration. If it might be, use @lrucare.
6472 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6474 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6475 bool lrucare
, bool compound
)
6477 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6479 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6480 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6482 if (mem_cgroup_disabled())
6485 * Swap faults will attempt to charge the same page multiple
6486 * times. But reuse_swap_page() might have removed the page
6487 * from swapcache already, so we can't check PageSwapCache().
6492 commit_charge(page
, memcg
, lrucare
);
6494 local_irq_disable();
6495 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
6496 memcg_check_events(memcg
, page
);
6499 if (do_memsw_account() && PageSwapCache(page
)) {
6500 swp_entry_t entry
= { .val
= page_private(page
) };
6502 * The swap entry might not get freed for a long time,
6503 * let's not wait for it. The page already received a
6504 * memory+swap charge, drop the swap entry duplicate.
6506 mem_cgroup_uncharge_swap(entry
, nr_pages
);
6511 * mem_cgroup_cancel_charge - cancel a page charge
6512 * @page: page to charge
6513 * @memcg: memcg to charge the page to
6514 * @compound: charge the page as compound or small page
6516 * Cancel a charge transaction started by mem_cgroup_try_charge().
6518 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6521 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
6523 if (mem_cgroup_disabled())
6526 * Swap faults will attempt to charge the same page multiple
6527 * times. But reuse_swap_page() might have removed the page
6528 * from swapcache already, so we can't check PageSwapCache().
6533 cancel_charge(memcg
, nr_pages
);
6536 struct uncharge_gather
{
6537 struct mem_cgroup
*memcg
;
6538 unsigned long pgpgout
;
6539 unsigned long nr_anon
;
6540 unsigned long nr_file
;
6541 unsigned long nr_kmem
;
6542 unsigned long nr_huge
;
6543 unsigned long nr_shmem
;
6544 struct page
*dummy_page
;
6547 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6549 memset(ug
, 0, sizeof(*ug
));
6552 static void uncharge_batch(const struct uncharge_gather
*ug
)
6554 unsigned long nr_pages
= ug
->nr_anon
+ ug
->nr_file
+ ug
->nr_kmem
;
6555 unsigned long flags
;
6557 if (!mem_cgroup_is_root(ug
->memcg
)) {
6558 page_counter_uncharge(&ug
->memcg
->memory
, nr_pages
);
6559 if (do_memsw_account())
6560 page_counter_uncharge(&ug
->memcg
->memsw
, nr_pages
);
6561 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6562 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6563 memcg_oom_recover(ug
->memcg
);
6566 local_irq_save(flags
);
6567 __mod_memcg_state(ug
->memcg
, MEMCG_RSS
, -ug
->nr_anon
);
6568 __mod_memcg_state(ug
->memcg
, MEMCG_CACHE
, -ug
->nr_file
);
6569 __mod_memcg_state(ug
->memcg
, MEMCG_RSS_HUGE
, -ug
->nr_huge
);
6570 __mod_memcg_state(ug
->memcg
, NR_SHMEM
, -ug
->nr_shmem
);
6571 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6572 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
6573 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6574 local_irq_restore(flags
);
6576 if (!mem_cgroup_is_root(ug
->memcg
))
6577 css_put_many(&ug
->memcg
->css
, nr_pages
);
6580 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6582 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6583 VM_BUG_ON_PAGE(page_count(page
) && !is_zone_device_page(page
) &&
6584 !PageHWPoison(page
) , page
);
6586 if (!page
->mem_cgroup
)
6590 * Nobody should be changing or seriously looking at
6591 * page->mem_cgroup at this point, we have fully
6592 * exclusive access to the page.
6595 if (ug
->memcg
!= page
->mem_cgroup
) {
6598 uncharge_gather_clear(ug
);
6600 ug
->memcg
= page
->mem_cgroup
;
6603 if (!PageKmemcg(page
)) {
6604 unsigned int nr_pages
= 1;
6606 if (PageTransHuge(page
)) {
6607 nr_pages
= compound_nr(page
);
6608 ug
->nr_huge
+= nr_pages
;
6611 ug
->nr_anon
+= nr_pages
;
6613 ug
->nr_file
+= nr_pages
;
6614 if (PageSwapBacked(page
))
6615 ug
->nr_shmem
+= nr_pages
;
6619 ug
->nr_kmem
+= compound_nr(page
);
6620 __ClearPageKmemcg(page
);
6623 ug
->dummy_page
= page
;
6624 page
->mem_cgroup
= NULL
;
6627 static void uncharge_list(struct list_head
*page_list
)
6629 struct uncharge_gather ug
;
6630 struct list_head
*next
;
6632 uncharge_gather_clear(&ug
);
6635 * Note that the list can be a single page->lru; hence the
6636 * do-while loop instead of a simple list_for_each_entry().
6638 next
= page_list
->next
;
6642 page
= list_entry(next
, struct page
, lru
);
6643 next
= page
->lru
.next
;
6645 uncharge_page(page
, &ug
);
6646 } while (next
!= page_list
);
6649 uncharge_batch(&ug
);
6653 * mem_cgroup_uncharge - uncharge a page
6654 * @page: page to uncharge
6656 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6657 * mem_cgroup_commit_charge().
6659 void mem_cgroup_uncharge(struct page
*page
)
6661 struct uncharge_gather ug
;
6663 if (mem_cgroup_disabled())
6666 /* Don't touch page->lru of any random page, pre-check: */
6667 if (!page
->mem_cgroup
)
6670 uncharge_gather_clear(&ug
);
6671 uncharge_page(page
, &ug
);
6672 uncharge_batch(&ug
);
6676 * mem_cgroup_uncharge_list - uncharge a list of page
6677 * @page_list: list of pages to uncharge
6679 * Uncharge a list of pages previously charged with
6680 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6682 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6684 if (mem_cgroup_disabled())
6687 if (!list_empty(page_list
))
6688 uncharge_list(page_list
);
6692 * mem_cgroup_migrate - charge a page's replacement
6693 * @oldpage: currently circulating page
6694 * @newpage: replacement page
6696 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6697 * be uncharged upon free.
6699 * Both pages must be locked, @newpage->mapping must be set up.
6701 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6703 struct mem_cgroup
*memcg
;
6704 unsigned int nr_pages
;
6705 unsigned long flags
;
6707 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6708 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6709 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6710 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6713 if (mem_cgroup_disabled())
6716 /* Page cache replacement: new page already charged? */
6717 if (newpage
->mem_cgroup
)
6720 /* Swapcache readahead pages can get replaced before being charged */
6721 memcg
= oldpage
->mem_cgroup
;
6725 /* Force-charge the new page. The old one will be freed soon */
6726 nr_pages
= hpage_nr_pages(newpage
);
6728 page_counter_charge(&memcg
->memory
, nr_pages
);
6729 if (do_memsw_account())
6730 page_counter_charge(&memcg
->memsw
, nr_pages
);
6731 css_get_many(&memcg
->css
, nr_pages
);
6733 commit_charge(newpage
, memcg
, false);
6735 local_irq_save(flags
);
6736 mem_cgroup_charge_statistics(memcg
, newpage
, PageTransHuge(newpage
),
6738 memcg_check_events(memcg
, newpage
);
6739 local_irq_restore(flags
);
6742 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6743 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6745 void mem_cgroup_sk_alloc(struct sock
*sk
)
6747 struct mem_cgroup
*memcg
;
6749 if (!mem_cgroup_sockets_enabled
)
6752 /* Do not associate the sock with unrelated interrupted task's memcg. */
6757 memcg
= mem_cgroup_from_task(current
);
6758 if (memcg
== root_mem_cgroup
)
6760 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6762 if (css_tryget_online(&memcg
->css
))
6763 sk
->sk_memcg
= memcg
;
6768 void mem_cgroup_sk_free(struct sock
*sk
)
6771 css_put(&sk
->sk_memcg
->css
);
6775 * mem_cgroup_charge_skmem - charge socket memory
6776 * @memcg: memcg to charge
6777 * @nr_pages: number of pages to charge
6779 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6780 * @memcg's configured limit, %false if the charge had to be forced.
6782 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6784 gfp_t gfp_mask
= GFP_KERNEL
;
6786 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6787 struct page_counter
*fail
;
6789 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
6790 memcg
->tcpmem_pressure
= 0;
6793 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
6794 memcg
->tcpmem_pressure
= 1;
6798 /* Don't block in the packet receive path */
6800 gfp_mask
= GFP_NOWAIT
;
6802 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
6804 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
6807 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
6812 * mem_cgroup_uncharge_skmem - uncharge socket memory
6813 * @memcg: memcg to uncharge
6814 * @nr_pages: number of pages to uncharge
6816 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
6818 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
6819 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
6823 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
6825 refill_stock(memcg
, nr_pages
);
6828 static int __init
cgroup_memory(char *s
)
6832 while ((token
= strsep(&s
, ",")) != NULL
) {
6835 if (!strcmp(token
, "nosocket"))
6836 cgroup_memory_nosocket
= true;
6837 if (!strcmp(token
, "nokmem"))
6838 cgroup_memory_nokmem
= true;
6842 __setup("cgroup.memory=", cgroup_memory
);
6845 * subsys_initcall() for memory controller.
6847 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6848 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6849 * basically everything that doesn't depend on a specific mem_cgroup structure
6850 * should be initialized from here.
6852 static int __init
mem_cgroup_init(void)
6856 #ifdef CONFIG_MEMCG_KMEM
6858 * Kmem cache creation is mostly done with the slab_mutex held,
6859 * so use a workqueue with limited concurrency to avoid stalling
6860 * all worker threads in case lots of cgroups are created and
6861 * destroyed simultaneously.
6863 memcg_kmem_cache_wq
= alloc_workqueue("memcg_kmem_cache", 0, 1);
6864 BUG_ON(!memcg_kmem_cache_wq
);
6867 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
6868 memcg_hotplug_cpu_dead
);
6870 for_each_possible_cpu(cpu
)
6871 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
6874 for_each_node(node
) {
6875 struct mem_cgroup_tree_per_node
*rtpn
;
6877 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
6878 node_online(node
) ? node
: NUMA_NO_NODE
);
6880 rtpn
->rb_root
= RB_ROOT
;
6881 rtpn
->rb_rightmost
= NULL
;
6882 spin_lock_init(&rtpn
->lock
);
6883 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6888 subsys_initcall(mem_cgroup_init
);
6890 #ifdef CONFIG_MEMCG_SWAP
6891 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
6893 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
6895 * The root cgroup cannot be destroyed, so it's refcount must
6898 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
6902 memcg
= parent_mem_cgroup(memcg
);
6904 memcg
= root_mem_cgroup
;
6910 * mem_cgroup_swapout - transfer a memsw charge to swap
6911 * @page: page whose memsw charge to transfer
6912 * @entry: swap entry to move the charge to
6914 * Transfer the memsw charge of @page to @entry.
6916 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6918 struct mem_cgroup
*memcg
, *swap_memcg
;
6919 unsigned int nr_entries
;
6920 unsigned short oldid
;
6922 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6923 VM_BUG_ON_PAGE(page_count(page
), page
);
6925 if (!do_memsw_account())
6928 memcg
= page
->mem_cgroup
;
6930 /* Readahead page, never charged */
6935 * In case the memcg owning these pages has been offlined and doesn't
6936 * have an ID allocated to it anymore, charge the closest online
6937 * ancestor for the swap instead and transfer the memory+swap charge.
6939 swap_memcg
= mem_cgroup_id_get_online(memcg
);
6940 nr_entries
= hpage_nr_pages(page
);
6941 /* Get references for the tail pages, too */
6943 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
6944 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
6946 VM_BUG_ON_PAGE(oldid
, page
);
6947 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
6949 page
->mem_cgroup
= NULL
;
6951 if (!mem_cgroup_is_root(memcg
))
6952 page_counter_uncharge(&memcg
->memory
, nr_entries
);
6954 if (memcg
!= swap_memcg
) {
6955 if (!mem_cgroup_is_root(swap_memcg
))
6956 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
6957 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
6961 * Interrupts should be disabled here because the caller holds the
6962 * i_pages lock which is taken with interrupts-off. It is
6963 * important here to have the interrupts disabled because it is the
6964 * only synchronisation we have for updating the per-CPU variables.
6966 VM_BUG_ON(!irqs_disabled());
6967 mem_cgroup_charge_statistics(memcg
, page
, PageTransHuge(page
),
6969 memcg_check_events(memcg
, page
);
6971 if (!mem_cgroup_is_root(memcg
))
6972 css_put_many(&memcg
->css
, nr_entries
);
6976 * mem_cgroup_try_charge_swap - try charging swap space for a page
6977 * @page: page being added to swap
6978 * @entry: swap entry to charge
6980 * Try to charge @page's memcg for the swap space at @entry.
6982 * Returns 0 on success, -ENOMEM on failure.
6984 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
6986 unsigned int nr_pages
= hpage_nr_pages(page
);
6987 struct page_counter
*counter
;
6988 struct mem_cgroup
*memcg
;
6989 unsigned short oldid
;
6991 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
6994 memcg
= page
->mem_cgroup
;
6996 /* Readahead page, never charged */
7001 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7005 memcg
= mem_cgroup_id_get_online(memcg
);
7007 if (!mem_cgroup_is_root(memcg
) &&
7008 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7009 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7010 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7011 mem_cgroup_id_put(memcg
);
7015 /* Get references for the tail pages, too */
7017 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7018 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7019 VM_BUG_ON_PAGE(oldid
, page
);
7020 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7026 * mem_cgroup_uncharge_swap - uncharge swap space
7027 * @entry: swap entry to uncharge
7028 * @nr_pages: the amount of swap space to uncharge
7030 void mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7032 struct mem_cgroup
*memcg
;
7035 if (!do_swap_account
)
7038 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7040 memcg
= mem_cgroup_from_id(id
);
7042 if (!mem_cgroup_is_root(memcg
)) {
7043 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7044 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7046 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7048 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7049 mem_cgroup_id_put_many(memcg
, nr_pages
);
7054 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7056 long nr_swap_pages
= get_nr_swap_pages();
7058 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7059 return nr_swap_pages
;
7060 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7061 nr_swap_pages
= min_t(long, nr_swap_pages
,
7062 READ_ONCE(memcg
->swap
.max
) -
7063 page_counter_read(&memcg
->swap
));
7064 return nr_swap_pages
;
7067 bool mem_cgroup_swap_full(struct page
*page
)
7069 struct mem_cgroup
*memcg
;
7071 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7075 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7078 memcg
= page
->mem_cgroup
;
7082 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7083 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.max
)
7089 /* for remember boot option*/
7090 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7091 static int really_do_swap_account __initdata
= 1;
7093 static int really_do_swap_account __initdata
;
7096 static int __init
enable_swap_account(char *s
)
7098 if (!strcmp(s
, "1"))
7099 really_do_swap_account
= 1;
7100 else if (!strcmp(s
, "0"))
7101 really_do_swap_account
= 0;
7104 __setup("swapaccount=", enable_swap_account
);
7106 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7109 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7111 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7114 static int swap_max_show(struct seq_file
*m
, void *v
)
7116 return seq_puts_memcg_tunable(m
,
7117 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7120 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7121 char *buf
, size_t nbytes
, loff_t off
)
7123 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7127 buf
= strstrip(buf
);
7128 err
= page_counter_memparse(buf
, "max", &max
);
7132 xchg(&memcg
->swap
.max
, max
);
7137 static int swap_events_show(struct seq_file
*m
, void *v
)
7139 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7141 seq_printf(m
, "max %lu\n",
7142 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7143 seq_printf(m
, "fail %lu\n",
7144 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7149 static struct cftype swap_files
[] = {
7151 .name
= "swap.current",
7152 .flags
= CFTYPE_NOT_ON_ROOT
,
7153 .read_u64
= swap_current_read
,
7157 .flags
= CFTYPE_NOT_ON_ROOT
,
7158 .seq_show
= swap_max_show
,
7159 .write
= swap_max_write
,
7162 .name
= "swap.events",
7163 .flags
= CFTYPE_NOT_ON_ROOT
,
7164 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7165 .seq_show
= swap_events_show
,
7170 static struct cftype memsw_cgroup_files
[] = {
7172 .name
= "memsw.usage_in_bytes",
7173 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7174 .read_u64
= mem_cgroup_read_u64
,
7177 .name
= "memsw.max_usage_in_bytes",
7178 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7179 .write
= mem_cgroup_reset
,
7180 .read_u64
= mem_cgroup_read_u64
,
7183 .name
= "memsw.limit_in_bytes",
7184 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7185 .write
= mem_cgroup_write
,
7186 .read_u64
= mem_cgroup_read_u64
,
7189 .name
= "memsw.failcnt",
7190 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7191 .write
= mem_cgroup_reset
,
7192 .read_u64
= mem_cgroup_read_u64
,
7194 { }, /* terminate */
7197 static int __init
mem_cgroup_swap_init(void)
7199 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7200 do_swap_account
= 1;
7201 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
7203 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
7204 memsw_cgroup_files
));
7208 subsys_initcall(mem_cgroup_swap_init
);
7210 #endif /* CONFIG_MEMCG_SWAP */