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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
75 EXPORT_SYMBOL(memory_cgrp_subsys
);
77 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup
*, int_active_memcg
);
81 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg
);
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket __ro_after_init
;
86 /* Kernel memory accounting disabled? */
87 bool cgroup_memory_nokmem __ro_after_init
;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 bool cgroup_memory_noswap __ro_after_init
;
93 #define cgroup_memory_noswap 1
96 #ifdef CONFIG_CGROUP_WRITEBACK
97 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
100 /* Whether legacy memory+swap accounting is active */
101 static bool do_memsw_account(void)
103 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_noswap
;
106 /* memcg and lruvec stats flushing */
107 static void flush_memcg_stats_dwork(struct work_struct
*w
);
108 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork
, flush_memcg_stats_dwork
);
109 static DEFINE_SPINLOCK(stats_flush_lock
);
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
115 * Cgroups above their limits are maintained in a RB-Tree, independent of
116 * their hierarchy representation
119 struct mem_cgroup_tree_per_node
{
120 struct rb_root rb_root
;
121 struct rb_node
*rb_rightmost
;
125 struct mem_cgroup_tree
{
126 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
129 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
132 struct mem_cgroup_eventfd_list
{
133 struct list_head list
;
134 struct eventfd_ctx
*eventfd
;
138 * cgroup_event represents events which userspace want to receive.
140 struct mem_cgroup_event
{
142 * memcg which the event belongs to.
144 struct mem_cgroup
*memcg
;
146 * eventfd to signal userspace about the event.
148 struct eventfd_ctx
*eventfd
;
150 * Each of these stored in a list by the cgroup.
152 struct list_head list
;
154 * register_event() callback will be used to add new userspace
155 * waiter for changes related to this event. Use eventfd_signal()
156 * on eventfd to send notification to userspace.
158 int (*register_event
)(struct mem_cgroup
*memcg
,
159 struct eventfd_ctx
*eventfd
, const char *args
);
161 * unregister_event() callback will be called when userspace closes
162 * the eventfd or on cgroup removing. This callback must be set,
163 * if you want provide notification functionality.
165 void (*unregister_event
)(struct mem_cgroup
*memcg
,
166 struct eventfd_ctx
*eventfd
);
168 * All fields below needed to unregister event when
169 * userspace closes eventfd.
172 wait_queue_head_t
*wqh
;
173 wait_queue_entry_t wait
;
174 struct work_struct remove
;
177 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
178 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
180 /* Stuffs for move charges at task migration. */
182 * Types of charges to be moved.
184 #define MOVE_ANON 0x1U
185 #define MOVE_FILE 0x2U
186 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
188 /* "mc" and its members are protected by cgroup_mutex */
189 static struct move_charge_struct
{
190 spinlock_t lock
; /* for from, to */
191 struct mm_struct
*mm
;
192 struct mem_cgroup
*from
;
193 struct mem_cgroup
*to
;
195 unsigned long precharge
;
196 unsigned long moved_charge
;
197 unsigned long moved_swap
;
198 struct task_struct
*moving_task
; /* a task moving charges */
199 wait_queue_head_t waitq
; /* a waitq for other context */
201 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
202 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
206 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
207 * limit reclaim to prevent infinite loops, if they ever occur.
209 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
210 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
212 /* for encoding cft->private value on file */
221 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
222 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
223 #define MEMFILE_ATTR(val) ((val) & 0xffff)
224 /* Used for OOM notifier */
225 #define OOM_CONTROL (0)
228 * Iteration constructs for visiting all cgroups (under a tree). If
229 * loops are exited prematurely (break), mem_cgroup_iter_break() must
230 * be used for reference counting.
232 #define for_each_mem_cgroup_tree(iter, root) \
233 for (iter = mem_cgroup_iter(root, NULL, NULL); \
235 iter = mem_cgroup_iter(root, iter, NULL))
237 #define for_each_mem_cgroup(iter) \
238 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
240 iter = mem_cgroup_iter(NULL, iter, NULL))
242 static inline bool should_force_charge(void)
244 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
245 (current
->flags
& PF_EXITING
);
248 /* Some nice accessors for the vmpressure. */
249 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
252 memcg
= root_mem_cgroup
;
253 return &memcg
->vmpressure
;
256 struct mem_cgroup
*vmpressure_to_memcg(struct vmpressure
*vmpr
)
258 return container_of(vmpr
, struct mem_cgroup
, vmpressure
);
261 #ifdef CONFIG_MEMCG_KMEM
262 extern spinlock_t css_set_lock
;
264 bool mem_cgroup_kmem_disabled(void)
266 return cgroup_memory_nokmem
;
269 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
270 unsigned int nr_pages
);
272 static void obj_cgroup_release(struct percpu_ref
*ref
)
274 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
275 unsigned int nr_bytes
;
276 unsigned int nr_pages
;
280 * At this point all allocated objects are freed, and
281 * objcg->nr_charged_bytes can't have an arbitrary byte value.
282 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
284 * The following sequence can lead to it:
285 * 1) CPU0: objcg == stock->cached_objcg
286 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
287 * PAGE_SIZE bytes are charged
288 * 3) CPU1: a process from another memcg is allocating something,
289 * the stock if flushed,
290 * objcg->nr_charged_bytes = PAGE_SIZE - 92
291 * 5) CPU0: we do release this object,
292 * 92 bytes are added to stock->nr_bytes
293 * 6) CPU0: stock is flushed,
294 * 92 bytes are added to objcg->nr_charged_bytes
296 * In the result, nr_charged_bytes == PAGE_SIZE.
297 * This page will be uncharged in obj_cgroup_release().
299 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
300 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
301 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
304 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
306 spin_lock_irqsave(&css_set_lock
, flags
);
307 list_del(&objcg
->list
);
308 spin_unlock_irqrestore(&css_set_lock
, flags
);
310 percpu_ref_exit(ref
);
311 kfree_rcu(objcg
, rcu
);
314 static struct obj_cgroup
*obj_cgroup_alloc(void)
316 struct obj_cgroup
*objcg
;
319 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
323 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
329 INIT_LIST_HEAD(&objcg
->list
);
333 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
334 struct mem_cgroup
*parent
)
336 struct obj_cgroup
*objcg
, *iter
;
338 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
340 spin_lock_irq(&css_set_lock
);
342 /* 1) Ready to reparent active objcg. */
343 list_add(&objcg
->list
, &memcg
->objcg_list
);
344 /* 2) Reparent active objcg and already reparented objcgs to parent. */
345 list_for_each_entry(iter
, &memcg
->objcg_list
, list
)
346 WRITE_ONCE(iter
->memcg
, parent
);
347 /* 3) Move already reparented objcgs to the parent's list */
348 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
350 spin_unlock_irq(&css_set_lock
);
352 percpu_ref_kill(&objcg
->refcnt
);
356 * This will be used as a shrinker list's index.
357 * The main reason for not using cgroup id for this:
358 * this works better in sparse environments, where we have a lot of memcgs,
359 * but only a few kmem-limited. Or also, if we have, for instance, 200
360 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
361 * 200 entry array for that.
363 * The current size of the caches array is stored in memcg_nr_cache_ids. It
364 * will double each time we have to increase it.
366 static DEFINE_IDA(memcg_cache_ida
);
367 int memcg_nr_cache_ids
;
369 /* Protects memcg_nr_cache_ids */
370 static DECLARE_RWSEM(memcg_cache_ids_sem
);
372 void memcg_get_cache_ids(void)
374 down_read(&memcg_cache_ids_sem
);
377 void memcg_put_cache_ids(void)
379 up_read(&memcg_cache_ids_sem
);
383 * MIN_SIZE is different than 1, because we would like to avoid going through
384 * the alloc/free process all the time. In a small machine, 4 kmem-limited
385 * cgroups is a reasonable guess. In the future, it could be a parameter or
386 * tunable, but that is strictly not necessary.
388 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
389 * this constant directly from cgroup, but it is understandable that this is
390 * better kept as an internal representation in cgroup.c. In any case, the
391 * cgrp_id space is not getting any smaller, and we don't have to necessarily
392 * increase ours as well if it increases.
394 #define MEMCG_CACHES_MIN_SIZE 4
395 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
398 * A lot of the calls to the cache allocation functions are expected to be
399 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
400 * conditional to this static branch, we'll have to allow modules that does
401 * kmem_cache_alloc and the such to see this symbol as well
403 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
404 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
408 * mem_cgroup_css_from_page - css of the memcg associated with a page
409 * @page: page of interest
411 * If memcg is bound to the default hierarchy, css of the memcg associated
412 * with @page is returned. The returned css remains associated with @page
413 * until it is released.
415 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
418 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
420 struct mem_cgroup
*memcg
;
422 memcg
= page_memcg(page
);
424 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
425 memcg
= root_mem_cgroup
;
431 * page_cgroup_ino - return inode number of the memcg a page is charged to
434 * Look up the closest online ancestor of the memory cgroup @page is charged to
435 * and return its inode number or 0 if @page is not charged to any cgroup. It
436 * is safe to call this function without holding a reference to @page.
438 * Note, this function is inherently racy, because there is nothing to prevent
439 * the cgroup inode from getting torn down and potentially reallocated a moment
440 * after page_cgroup_ino() returns, so it only should be used by callers that
441 * do not care (such as procfs interfaces).
443 ino_t
page_cgroup_ino(struct page
*page
)
445 struct mem_cgroup
*memcg
;
446 unsigned long ino
= 0;
449 memcg
= page_memcg_check(page
);
451 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
452 memcg
= parent_mem_cgroup(memcg
);
454 ino
= cgroup_ino(memcg
->css
.cgroup
);
459 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
460 struct mem_cgroup_tree_per_node
*mctz
,
461 unsigned long new_usage_in_excess
)
463 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
464 struct rb_node
*parent
= NULL
;
465 struct mem_cgroup_per_node
*mz_node
;
466 bool rightmost
= true;
471 mz
->usage_in_excess
= new_usage_in_excess
;
472 if (!mz
->usage_in_excess
)
476 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
478 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
487 mctz
->rb_rightmost
= &mz
->tree_node
;
489 rb_link_node(&mz
->tree_node
, parent
, p
);
490 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
494 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
495 struct mem_cgroup_tree_per_node
*mctz
)
500 if (&mz
->tree_node
== mctz
->rb_rightmost
)
501 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
503 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
507 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
508 struct mem_cgroup_tree_per_node
*mctz
)
512 spin_lock_irqsave(&mctz
->lock
, flags
);
513 __mem_cgroup_remove_exceeded(mz
, mctz
);
514 spin_unlock_irqrestore(&mctz
->lock
, flags
);
517 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
519 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
520 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
521 unsigned long excess
= 0;
523 if (nr_pages
> soft_limit
)
524 excess
= nr_pages
- soft_limit
;
529 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, int nid
)
531 unsigned long excess
;
532 struct mem_cgroup_per_node
*mz
;
533 struct mem_cgroup_tree_per_node
*mctz
;
535 mctz
= soft_limit_tree
.rb_tree_per_node
[nid
];
539 * Necessary to update all ancestors when hierarchy is used.
540 * because their event counter is not touched.
542 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
543 mz
= memcg
->nodeinfo
[nid
];
544 excess
= soft_limit_excess(memcg
);
546 * We have to update the tree if mz is on RB-tree or
547 * mem is over its softlimit.
549 if (excess
|| mz
->on_tree
) {
552 spin_lock_irqsave(&mctz
->lock
, flags
);
553 /* if on-tree, remove it */
555 __mem_cgroup_remove_exceeded(mz
, mctz
);
557 * Insert again. mz->usage_in_excess will be updated.
558 * If excess is 0, no tree ops.
560 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
561 spin_unlock_irqrestore(&mctz
->lock
, flags
);
566 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
568 struct mem_cgroup_tree_per_node
*mctz
;
569 struct mem_cgroup_per_node
*mz
;
573 mz
= memcg
->nodeinfo
[nid
];
574 mctz
= soft_limit_tree
.rb_tree_per_node
[nid
];
576 mem_cgroup_remove_exceeded(mz
, mctz
);
580 static struct mem_cgroup_per_node
*
581 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
583 struct mem_cgroup_per_node
*mz
;
587 if (!mctz
->rb_rightmost
)
588 goto done
; /* Nothing to reclaim from */
590 mz
= rb_entry(mctz
->rb_rightmost
,
591 struct mem_cgroup_per_node
, tree_node
);
593 * Remove the node now but someone else can add it back,
594 * we will to add it back at the end of reclaim to its correct
595 * position in the tree.
597 __mem_cgroup_remove_exceeded(mz
, mctz
);
598 if (!soft_limit_excess(mz
->memcg
) ||
599 !css_tryget(&mz
->memcg
->css
))
605 static struct mem_cgroup_per_node
*
606 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
608 struct mem_cgroup_per_node
*mz
;
610 spin_lock_irq(&mctz
->lock
);
611 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
612 spin_unlock_irq(&mctz
->lock
);
617 * __mod_memcg_state - update cgroup memory statistics
618 * @memcg: the memory cgroup
619 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
620 * @val: delta to add to the counter, can be negative
622 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
624 if (mem_cgroup_disabled())
627 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
628 cgroup_rstat_updated(memcg
->css
.cgroup
, smp_processor_id());
631 /* idx can be of type enum memcg_stat_item or node_stat_item. */
632 static unsigned long memcg_page_state_local(struct mem_cgroup
*memcg
, int idx
)
637 for_each_possible_cpu(cpu
)
638 x
+= per_cpu(memcg
->vmstats_percpu
->state
[idx
], cpu
);
646 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
649 struct mem_cgroup_per_node
*pn
;
650 struct mem_cgroup
*memcg
;
652 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
656 __mod_memcg_state(memcg
, idx
, val
);
659 __this_cpu_add(pn
->lruvec_stats_percpu
->state
[idx
], val
);
663 * __mod_lruvec_state - update lruvec memory statistics
664 * @lruvec: the lruvec
665 * @idx: the stat item
666 * @val: delta to add to the counter, can be negative
668 * The lruvec is the intersection of the NUMA node and a cgroup. This
669 * function updates the all three counters that are affected by a
670 * change of state at this level: per-node, per-cgroup, per-lruvec.
672 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
676 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
678 /* Update memcg and lruvec */
679 if (!mem_cgroup_disabled())
680 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
683 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
686 struct page
*head
= compound_head(page
); /* rmap on tail pages */
687 struct mem_cgroup
*memcg
;
688 pg_data_t
*pgdat
= page_pgdat(page
);
689 struct lruvec
*lruvec
;
692 memcg
= page_memcg(head
);
693 /* Untracked pages have no memcg, no lruvec. Update only the node */
696 __mod_node_page_state(pgdat
, idx
, val
);
700 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
701 __mod_lruvec_state(lruvec
, idx
, val
);
704 EXPORT_SYMBOL(__mod_lruvec_page_state
);
706 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
708 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
709 struct mem_cgroup
*memcg
;
710 struct lruvec
*lruvec
;
713 memcg
= mem_cgroup_from_obj(p
);
716 * Untracked pages have no memcg, no lruvec. Update only the
717 * node. If we reparent the slab objects to the root memcg,
718 * when we free the slab object, we need to update the per-memcg
719 * vmstats to keep it correct for the root memcg.
722 __mod_node_page_state(pgdat
, idx
, val
);
724 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
725 __mod_lruvec_state(lruvec
, idx
, val
);
731 * mod_objcg_mlstate() may be called with irq enabled, so
732 * mod_memcg_lruvec_state() should be used.
734 static inline void mod_objcg_mlstate(struct obj_cgroup
*objcg
,
735 struct pglist_data
*pgdat
,
736 enum node_stat_item idx
, int nr
)
738 struct mem_cgroup
*memcg
;
739 struct lruvec
*lruvec
;
742 memcg
= obj_cgroup_memcg(objcg
);
743 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
744 mod_memcg_lruvec_state(lruvec
, idx
, nr
);
749 * __count_memcg_events - account VM events in a cgroup
750 * @memcg: the memory cgroup
751 * @idx: the event item
752 * @count: the number of events that occurred
754 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
757 if (mem_cgroup_disabled())
760 __this_cpu_add(memcg
->vmstats_percpu
->events
[idx
], count
);
761 cgroup_rstat_updated(memcg
->css
.cgroup
, smp_processor_id());
764 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
766 return READ_ONCE(memcg
->vmstats
.events
[event
]);
769 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
774 for_each_possible_cpu(cpu
)
775 x
+= per_cpu(memcg
->vmstats_percpu
->events
[event
], cpu
);
779 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
782 /* pagein of a big page is an event. So, ignore page size */
784 __count_memcg_events(memcg
, PGPGIN
, 1);
786 __count_memcg_events(memcg
, PGPGOUT
, 1);
787 nr_pages
= -nr_pages
; /* for event */
790 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
793 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
794 enum mem_cgroup_events_target target
)
796 unsigned long val
, next
;
798 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
799 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
800 /* from time_after() in jiffies.h */
801 if ((long)(next
- val
) < 0) {
803 case MEM_CGROUP_TARGET_THRESH
:
804 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
806 case MEM_CGROUP_TARGET_SOFTLIMIT
:
807 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
812 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
819 * Check events in order.
822 static void memcg_check_events(struct mem_cgroup
*memcg
, int nid
)
824 /* threshold event is triggered in finer grain than soft limit */
825 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
826 MEM_CGROUP_TARGET_THRESH
))) {
829 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
830 MEM_CGROUP_TARGET_SOFTLIMIT
);
831 mem_cgroup_threshold(memcg
);
832 if (unlikely(do_softlimit
))
833 mem_cgroup_update_tree(memcg
, nid
);
837 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
840 * mm_update_next_owner() may clear mm->owner to NULL
841 * if it races with swapoff, page migration, etc.
842 * So this can be called with p == NULL.
847 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
849 EXPORT_SYMBOL(mem_cgroup_from_task
);
851 static __always_inline
struct mem_cgroup
*active_memcg(void)
854 return this_cpu_read(int_active_memcg
);
856 return current
->active_memcg
;
860 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
861 * @mm: mm from which memcg should be extracted. It can be NULL.
863 * Obtain a reference on mm->memcg and returns it if successful. If mm
864 * is NULL, then the memcg is chosen as follows:
865 * 1) The active memcg, if set.
866 * 2) current->mm->memcg, if available
868 * If mem_cgroup is disabled, NULL is returned.
870 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
872 struct mem_cgroup
*memcg
;
874 if (mem_cgroup_disabled())
878 * Page cache insertions can happen without an
879 * actual mm context, e.g. during disk probing
880 * on boot, loopback IO, acct() writes etc.
882 * No need to css_get on root memcg as the reference
883 * counting is disabled on the root level in the
884 * cgroup core. See CSS_NO_REF.
887 memcg
= active_memcg();
888 if (unlikely(memcg
)) {
889 /* remote memcg must hold a ref */
890 css_get(&memcg
->css
);
895 return root_mem_cgroup
;
900 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
901 if (unlikely(!memcg
))
902 memcg
= root_mem_cgroup
;
903 } while (!css_tryget(&memcg
->css
));
907 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
909 static __always_inline
bool memcg_kmem_bypass(void)
911 /* Allow remote memcg charging from any context. */
912 if (unlikely(active_memcg()))
915 /* Memcg to charge can't be determined. */
916 if (!in_task() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
923 * mem_cgroup_iter - iterate over memory cgroup hierarchy
924 * @root: hierarchy root
925 * @prev: previously returned memcg, NULL on first invocation
926 * @reclaim: cookie for shared reclaim walks, NULL for full walks
928 * Returns references to children of the hierarchy below @root, or
929 * @root itself, or %NULL after a full round-trip.
931 * Caller must pass the return value in @prev on subsequent
932 * invocations for reference counting, or use mem_cgroup_iter_break()
933 * to cancel a hierarchy walk before the round-trip is complete.
935 * Reclaimers can specify a node in @reclaim to divide up the memcgs
936 * in the hierarchy among all concurrent reclaimers operating on the
939 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
940 struct mem_cgroup
*prev
,
941 struct mem_cgroup_reclaim_cookie
*reclaim
)
943 struct mem_cgroup_reclaim_iter
*iter
;
944 struct cgroup_subsys_state
*css
= NULL
;
945 struct mem_cgroup
*memcg
= NULL
;
946 struct mem_cgroup
*pos
= NULL
;
948 if (mem_cgroup_disabled())
952 root
= root_mem_cgroup
;
954 if (prev
&& !reclaim
)
960 struct mem_cgroup_per_node
*mz
;
962 mz
= root
->nodeinfo
[reclaim
->pgdat
->node_id
];
965 if (prev
&& reclaim
->generation
!= iter
->generation
)
969 pos
= READ_ONCE(iter
->position
);
970 if (!pos
|| css_tryget(&pos
->css
))
973 * css reference reached zero, so iter->position will
974 * be cleared by ->css_released. However, we should not
975 * rely on this happening soon, because ->css_released
976 * is called from a work queue, and by busy-waiting we
977 * might block it. So we clear iter->position right
980 (void)cmpxchg(&iter
->position
, pos
, NULL
);
988 css
= css_next_descendant_pre(css
, &root
->css
);
991 * Reclaimers share the hierarchy walk, and a
992 * new one might jump in right at the end of
993 * the hierarchy - make sure they see at least
994 * one group and restart from the beginning.
1002 * Verify the css and acquire a reference. The root
1003 * is provided by the caller, so we know it's alive
1004 * and kicking, and don't take an extra reference.
1006 memcg
= mem_cgroup_from_css(css
);
1008 if (css
== &root
->css
)
1011 if (css_tryget(css
))
1019 * The position could have already been updated by a competing
1020 * thread, so check that the value hasn't changed since we read
1021 * it to avoid reclaiming from the same cgroup twice.
1023 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1031 reclaim
->generation
= iter
->generation
;
1036 if (prev
&& prev
!= root
)
1037 css_put(&prev
->css
);
1043 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1044 * @root: hierarchy root
1045 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1047 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1048 struct mem_cgroup
*prev
)
1051 root
= root_mem_cgroup
;
1052 if (prev
&& prev
!= root
)
1053 css_put(&prev
->css
);
1056 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1057 struct mem_cgroup
*dead_memcg
)
1059 struct mem_cgroup_reclaim_iter
*iter
;
1060 struct mem_cgroup_per_node
*mz
;
1063 for_each_node(nid
) {
1064 mz
= from
->nodeinfo
[nid
];
1066 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1070 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1072 struct mem_cgroup
*memcg
= dead_memcg
;
1073 struct mem_cgroup
*last
;
1076 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1078 } while ((memcg
= parent_mem_cgroup(memcg
)));
1081 * When cgruop1 non-hierarchy mode is used,
1082 * parent_mem_cgroup() does not walk all the way up to the
1083 * cgroup root (root_mem_cgroup). So we have to handle
1084 * dead_memcg from cgroup root separately.
1086 if (last
!= root_mem_cgroup
)
1087 __invalidate_reclaim_iterators(root_mem_cgroup
,
1092 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1093 * @memcg: hierarchy root
1094 * @fn: function to call for each task
1095 * @arg: argument passed to @fn
1097 * This function iterates over tasks attached to @memcg or to any of its
1098 * descendants and calls @fn for each task. If @fn returns a non-zero
1099 * value, the function breaks the iteration loop and returns the value.
1100 * Otherwise, it will iterate over all tasks and return 0.
1102 * This function must not be called for the root memory cgroup.
1104 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1105 int (*fn
)(struct task_struct
*, void *), void *arg
)
1107 struct mem_cgroup
*iter
;
1110 BUG_ON(memcg
== root_mem_cgroup
);
1112 for_each_mem_cgroup_tree(iter
, memcg
) {
1113 struct css_task_iter it
;
1114 struct task_struct
*task
;
1116 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1117 while (!ret
&& (task
= css_task_iter_next(&it
)))
1118 ret
= fn(task
, arg
);
1119 css_task_iter_end(&it
);
1121 mem_cgroup_iter_break(memcg
, iter
);
1128 #ifdef CONFIG_DEBUG_VM
1129 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct page
*page
)
1131 struct mem_cgroup
*memcg
;
1133 if (mem_cgroup_disabled())
1136 memcg
= page_memcg(page
);
1139 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != root_mem_cgroup
, page
);
1141 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != memcg
, page
);
1146 * lock_page_lruvec - lock and return lruvec for a given page.
1149 * These functions are safe to use under any of the following conditions:
1152 * - lock_page_memcg()
1153 * - page->_refcount is zero
1155 struct lruvec
*lock_page_lruvec(struct page
*page
)
1157 struct lruvec
*lruvec
;
1159 lruvec
= mem_cgroup_page_lruvec(page
);
1160 spin_lock(&lruvec
->lru_lock
);
1162 lruvec_memcg_debug(lruvec
, page
);
1167 struct lruvec
*lock_page_lruvec_irq(struct page
*page
)
1169 struct lruvec
*lruvec
;
1171 lruvec
= mem_cgroup_page_lruvec(page
);
1172 spin_lock_irq(&lruvec
->lru_lock
);
1174 lruvec_memcg_debug(lruvec
, page
);
1179 struct lruvec
*lock_page_lruvec_irqsave(struct page
*page
, unsigned long *flags
)
1181 struct lruvec
*lruvec
;
1183 lruvec
= mem_cgroup_page_lruvec(page
);
1184 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1186 lruvec_memcg_debug(lruvec
, page
);
1192 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1193 * @lruvec: mem_cgroup per zone lru vector
1194 * @lru: index of lru list the page is sitting on
1195 * @zid: zone id of the accounted pages
1196 * @nr_pages: positive when adding or negative when removing
1198 * This function must be called under lru_lock, just before a page is added
1199 * to or just after a page is removed from an lru list (that ordering being
1200 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1202 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1203 int zid
, int nr_pages
)
1205 struct mem_cgroup_per_node
*mz
;
1206 unsigned long *lru_size
;
1209 if (mem_cgroup_disabled())
1212 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1213 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1216 *lru_size
+= nr_pages
;
1219 if (WARN_ONCE(size
< 0,
1220 "%s(%p, %d, %d): lru_size %ld\n",
1221 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1227 *lru_size
+= nr_pages
;
1231 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1232 * @memcg: the memory cgroup
1234 * Returns the maximum amount of memory @mem can be charged with, in
1237 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1239 unsigned long margin
= 0;
1240 unsigned long count
;
1241 unsigned long limit
;
1243 count
= page_counter_read(&memcg
->memory
);
1244 limit
= READ_ONCE(memcg
->memory
.max
);
1246 margin
= limit
- count
;
1248 if (do_memsw_account()) {
1249 count
= page_counter_read(&memcg
->memsw
);
1250 limit
= READ_ONCE(memcg
->memsw
.max
);
1252 margin
= min(margin
, limit
- count
);
1261 * A routine for checking "mem" is under move_account() or not.
1263 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1264 * moving cgroups. This is for waiting at high-memory pressure
1267 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1269 struct mem_cgroup
*from
;
1270 struct mem_cgroup
*to
;
1273 * Unlike task_move routines, we access mc.to, mc.from not under
1274 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1276 spin_lock(&mc
.lock
);
1282 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1283 mem_cgroup_is_descendant(to
, memcg
);
1285 spin_unlock(&mc
.lock
);
1289 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1291 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1292 if (mem_cgroup_under_move(memcg
)) {
1294 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1295 /* moving charge context might have finished. */
1298 finish_wait(&mc
.waitq
, &wait
);
1305 struct memory_stat
{
1310 static const struct memory_stat memory_stats
[] = {
1311 { "anon", NR_ANON_MAPPED
},
1312 { "file", NR_FILE_PAGES
},
1313 { "kernel_stack", NR_KERNEL_STACK_KB
},
1314 { "pagetables", NR_PAGETABLE
},
1315 { "percpu", MEMCG_PERCPU_B
},
1316 { "sock", MEMCG_SOCK
},
1317 { "shmem", NR_SHMEM
},
1318 { "file_mapped", NR_FILE_MAPPED
},
1319 { "file_dirty", NR_FILE_DIRTY
},
1320 { "file_writeback", NR_WRITEBACK
},
1322 { "swapcached", NR_SWAPCACHE
},
1324 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1325 { "anon_thp", NR_ANON_THPS
},
1326 { "file_thp", NR_FILE_THPS
},
1327 { "shmem_thp", NR_SHMEM_THPS
},
1329 { "inactive_anon", NR_INACTIVE_ANON
},
1330 { "active_anon", NR_ACTIVE_ANON
},
1331 { "inactive_file", NR_INACTIVE_FILE
},
1332 { "active_file", NR_ACTIVE_FILE
},
1333 { "unevictable", NR_UNEVICTABLE
},
1334 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B
},
1335 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B
},
1337 /* The memory events */
1338 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON
},
1339 { "workingset_refault_file", WORKINGSET_REFAULT_FILE
},
1340 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON
},
1341 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE
},
1342 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON
},
1343 { "workingset_restore_file", WORKINGSET_RESTORE_FILE
},
1344 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM
},
1347 /* Translate stat items to the correct unit for memory.stat output */
1348 static int memcg_page_state_unit(int item
)
1351 case MEMCG_PERCPU_B
:
1352 case NR_SLAB_RECLAIMABLE_B
:
1353 case NR_SLAB_UNRECLAIMABLE_B
:
1354 case WORKINGSET_REFAULT_ANON
:
1355 case WORKINGSET_REFAULT_FILE
:
1356 case WORKINGSET_ACTIVATE_ANON
:
1357 case WORKINGSET_ACTIVATE_FILE
:
1358 case WORKINGSET_RESTORE_ANON
:
1359 case WORKINGSET_RESTORE_FILE
:
1360 case WORKINGSET_NODERECLAIM
:
1362 case NR_KERNEL_STACK_KB
:
1369 static inline unsigned long memcg_page_state_output(struct mem_cgroup
*memcg
,
1372 return memcg_page_state(memcg
, item
) * memcg_page_state_unit(item
);
1375 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1380 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1385 * Provide statistics on the state of the memory subsystem as
1386 * well as cumulative event counters that show past behavior.
1388 * This list is ordered following a combination of these gradients:
1389 * 1) generic big picture -> specifics and details
1390 * 2) reflecting userspace activity -> reflecting kernel heuristics
1392 * Current memory state:
1394 cgroup_rstat_flush(memcg
->css
.cgroup
);
1396 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1399 size
= memcg_page_state_output(memcg
, memory_stats
[i
].idx
);
1400 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1402 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1403 size
+= memcg_page_state_output(memcg
,
1404 NR_SLAB_RECLAIMABLE_B
);
1405 seq_buf_printf(&s
, "slab %llu\n", size
);
1409 /* Accumulated memory events */
1411 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1412 memcg_events(memcg
, PGFAULT
));
1413 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1414 memcg_events(memcg
, PGMAJFAULT
));
1415 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1416 memcg_events(memcg
, PGREFILL
));
1417 seq_buf_printf(&s
, "pgscan %lu\n",
1418 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1419 memcg_events(memcg
, PGSCAN_DIRECT
));
1420 seq_buf_printf(&s
, "pgsteal %lu\n",
1421 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1422 memcg_events(memcg
, PGSTEAL_DIRECT
));
1423 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1424 memcg_events(memcg
, PGACTIVATE
));
1425 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1426 memcg_events(memcg
, PGDEACTIVATE
));
1427 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1428 memcg_events(memcg
, PGLAZYFREE
));
1429 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1430 memcg_events(memcg
, PGLAZYFREED
));
1432 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1433 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1434 memcg_events(memcg
, THP_FAULT_ALLOC
));
1435 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1436 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1437 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1439 /* The above should easily fit into one page */
1440 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1445 #define K(x) ((x) << (PAGE_SHIFT-10))
1447 * mem_cgroup_print_oom_context: Print OOM information relevant to
1448 * memory controller.
1449 * @memcg: The memory cgroup that went over limit
1450 * @p: Task that is going to be killed
1452 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1455 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1460 pr_cont(",oom_memcg=");
1461 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1463 pr_cont(",global_oom");
1465 pr_cont(",task_memcg=");
1466 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1472 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1473 * memory controller.
1474 * @memcg: The memory cgroup that went over limit
1476 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1480 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1481 K((u64
)page_counter_read(&memcg
->memory
)),
1482 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1483 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1484 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1485 K((u64
)page_counter_read(&memcg
->swap
)),
1486 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1488 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1489 K((u64
)page_counter_read(&memcg
->memsw
)),
1490 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1491 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1492 K((u64
)page_counter_read(&memcg
->kmem
)),
1493 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1496 pr_info("Memory cgroup stats for ");
1497 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1499 buf
= memory_stat_format(memcg
);
1507 * Return the memory (and swap, if configured) limit for a memcg.
1509 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1511 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1513 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1514 if (mem_cgroup_swappiness(memcg
))
1515 max
+= min(READ_ONCE(memcg
->swap
.max
),
1516 (unsigned long)total_swap_pages
);
1518 if (mem_cgroup_swappiness(memcg
)) {
1519 /* Calculate swap excess capacity from memsw limit */
1520 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1522 max
+= min(swap
, (unsigned long)total_swap_pages
);
1528 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1530 return page_counter_read(&memcg
->memory
);
1533 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1536 struct oom_control oc
= {
1540 .gfp_mask
= gfp_mask
,
1545 if (mutex_lock_killable(&oom_lock
))
1548 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1552 * A few threads which were not waiting at mutex_lock_killable() can
1553 * fail to bail out. Therefore, check again after holding oom_lock.
1555 ret
= should_force_charge() || out_of_memory(&oc
);
1558 mutex_unlock(&oom_lock
);
1562 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1565 unsigned long *total_scanned
)
1567 struct mem_cgroup
*victim
= NULL
;
1570 unsigned long excess
;
1571 unsigned long nr_scanned
;
1572 struct mem_cgroup_reclaim_cookie reclaim
= {
1576 excess
= soft_limit_excess(root_memcg
);
1579 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1584 * If we have not been able to reclaim
1585 * anything, it might because there are
1586 * no reclaimable pages under this hierarchy
1591 * We want to do more targeted reclaim.
1592 * excess >> 2 is not to excessive so as to
1593 * reclaim too much, nor too less that we keep
1594 * coming back to reclaim from this cgroup
1596 if (total
>= (excess
>> 2) ||
1597 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1602 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1603 pgdat
, &nr_scanned
);
1604 *total_scanned
+= nr_scanned
;
1605 if (!soft_limit_excess(root_memcg
))
1608 mem_cgroup_iter_break(root_memcg
, victim
);
1612 #ifdef CONFIG_LOCKDEP
1613 static struct lockdep_map memcg_oom_lock_dep_map
= {
1614 .name
= "memcg_oom_lock",
1618 static DEFINE_SPINLOCK(memcg_oom_lock
);
1621 * Check OOM-Killer is already running under our hierarchy.
1622 * If someone is running, return false.
1624 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1626 struct mem_cgroup
*iter
, *failed
= NULL
;
1628 spin_lock(&memcg_oom_lock
);
1630 for_each_mem_cgroup_tree(iter
, memcg
) {
1631 if (iter
->oom_lock
) {
1633 * this subtree of our hierarchy is already locked
1634 * so we cannot give a lock.
1637 mem_cgroup_iter_break(memcg
, iter
);
1640 iter
->oom_lock
= true;
1645 * OK, we failed to lock the whole subtree so we have
1646 * to clean up what we set up to the failing subtree
1648 for_each_mem_cgroup_tree(iter
, memcg
) {
1649 if (iter
== failed
) {
1650 mem_cgroup_iter_break(memcg
, iter
);
1653 iter
->oom_lock
= false;
1656 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1658 spin_unlock(&memcg_oom_lock
);
1663 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1665 struct mem_cgroup
*iter
;
1667 spin_lock(&memcg_oom_lock
);
1668 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1669 for_each_mem_cgroup_tree(iter
, memcg
)
1670 iter
->oom_lock
= false;
1671 spin_unlock(&memcg_oom_lock
);
1674 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1676 struct mem_cgroup
*iter
;
1678 spin_lock(&memcg_oom_lock
);
1679 for_each_mem_cgroup_tree(iter
, memcg
)
1681 spin_unlock(&memcg_oom_lock
);
1684 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1686 struct mem_cgroup
*iter
;
1689 * Be careful about under_oom underflows because a child memcg
1690 * could have been added after mem_cgroup_mark_under_oom.
1692 spin_lock(&memcg_oom_lock
);
1693 for_each_mem_cgroup_tree(iter
, memcg
)
1694 if (iter
->under_oom
> 0)
1696 spin_unlock(&memcg_oom_lock
);
1699 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1701 struct oom_wait_info
{
1702 struct mem_cgroup
*memcg
;
1703 wait_queue_entry_t wait
;
1706 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1707 unsigned mode
, int sync
, void *arg
)
1709 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1710 struct mem_cgroup
*oom_wait_memcg
;
1711 struct oom_wait_info
*oom_wait_info
;
1713 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1714 oom_wait_memcg
= oom_wait_info
->memcg
;
1716 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1717 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1719 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1722 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1725 * For the following lockless ->under_oom test, the only required
1726 * guarantee is that it must see the state asserted by an OOM when
1727 * this function is called as a result of userland actions
1728 * triggered by the notification of the OOM. This is trivially
1729 * achieved by invoking mem_cgroup_mark_under_oom() before
1730 * triggering notification.
1732 if (memcg
&& memcg
->under_oom
)
1733 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1743 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1745 enum oom_status ret
;
1748 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1751 memcg_memory_event(memcg
, MEMCG_OOM
);
1754 * We are in the middle of the charge context here, so we
1755 * don't want to block when potentially sitting on a callstack
1756 * that holds all kinds of filesystem and mm locks.
1758 * cgroup1 allows disabling the OOM killer and waiting for outside
1759 * handling until the charge can succeed; remember the context and put
1760 * the task to sleep at the end of the page fault when all locks are
1763 * On the other hand, in-kernel OOM killer allows for an async victim
1764 * memory reclaim (oom_reaper) and that means that we are not solely
1765 * relying on the oom victim to make a forward progress and we can
1766 * invoke the oom killer here.
1768 * Please note that mem_cgroup_out_of_memory might fail to find a
1769 * victim and then we have to bail out from the charge path.
1771 if (memcg
->oom_kill_disable
) {
1772 if (!current
->in_user_fault
)
1774 css_get(&memcg
->css
);
1775 current
->memcg_in_oom
= memcg
;
1776 current
->memcg_oom_gfp_mask
= mask
;
1777 current
->memcg_oom_order
= order
;
1782 mem_cgroup_mark_under_oom(memcg
);
1784 locked
= mem_cgroup_oom_trylock(memcg
);
1787 mem_cgroup_oom_notify(memcg
);
1789 mem_cgroup_unmark_under_oom(memcg
);
1790 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1796 mem_cgroup_oom_unlock(memcg
);
1802 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1803 * @handle: actually kill/wait or just clean up the OOM state
1805 * This has to be called at the end of a page fault if the memcg OOM
1806 * handler was enabled.
1808 * Memcg supports userspace OOM handling where failed allocations must
1809 * sleep on a waitqueue until the userspace task resolves the
1810 * situation. Sleeping directly in the charge context with all kinds
1811 * of locks held is not a good idea, instead we remember an OOM state
1812 * in the task and mem_cgroup_oom_synchronize() has to be called at
1813 * the end of the page fault to complete the OOM handling.
1815 * Returns %true if an ongoing memcg OOM situation was detected and
1816 * completed, %false otherwise.
1818 bool mem_cgroup_oom_synchronize(bool handle
)
1820 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1821 struct oom_wait_info owait
;
1824 /* OOM is global, do not handle */
1831 owait
.memcg
= memcg
;
1832 owait
.wait
.flags
= 0;
1833 owait
.wait
.func
= memcg_oom_wake_function
;
1834 owait
.wait
.private = current
;
1835 INIT_LIST_HEAD(&owait
.wait
.entry
);
1837 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1838 mem_cgroup_mark_under_oom(memcg
);
1840 locked
= mem_cgroup_oom_trylock(memcg
);
1843 mem_cgroup_oom_notify(memcg
);
1845 if (locked
&& !memcg
->oom_kill_disable
) {
1846 mem_cgroup_unmark_under_oom(memcg
);
1847 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1848 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1849 current
->memcg_oom_order
);
1852 mem_cgroup_unmark_under_oom(memcg
);
1853 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1857 mem_cgroup_oom_unlock(memcg
);
1859 * There is no guarantee that an OOM-lock contender
1860 * sees the wakeups triggered by the OOM kill
1861 * uncharges. Wake any sleepers explicitly.
1863 memcg_oom_recover(memcg
);
1866 current
->memcg_in_oom
= NULL
;
1867 css_put(&memcg
->css
);
1872 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1873 * @victim: task to be killed by the OOM killer
1874 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1876 * Returns a pointer to a memory cgroup, which has to be cleaned up
1877 * by killing all belonging OOM-killable tasks.
1879 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1881 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1882 struct mem_cgroup
*oom_domain
)
1884 struct mem_cgroup
*oom_group
= NULL
;
1885 struct mem_cgroup
*memcg
;
1887 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1891 oom_domain
= root_mem_cgroup
;
1895 memcg
= mem_cgroup_from_task(victim
);
1896 if (memcg
== root_mem_cgroup
)
1900 * If the victim task has been asynchronously moved to a different
1901 * memory cgroup, we might end up killing tasks outside oom_domain.
1902 * In this case it's better to ignore memory.group.oom.
1904 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
1908 * Traverse the memory cgroup hierarchy from the victim task's
1909 * cgroup up to the OOMing cgroup (or root) to find the
1910 * highest-level memory cgroup with oom.group set.
1912 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1913 if (memcg
->oom_group
)
1916 if (memcg
== oom_domain
)
1921 css_get(&oom_group
->css
);
1928 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
1930 pr_info("Tasks in ");
1931 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1932 pr_cont(" are going to be killed due to memory.oom.group set\n");
1936 * folio_memcg_lock - Bind a folio to its memcg.
1937 * @folio: The folio.
1939 * This function prevents unlocked LRU folios from being moved to
1942 * It ensures lifetime of the bound memcg. The caller is responsible
1943 * for the lifetime of the folio.
1945 void folio_memcg_lock(struct folio
*folio
)
1947 struct mem_cgroup
*memcg
;
1948 unsigned long flags
;
1951 * The RCU lock is held throughout the transaction. The fast
1952 * path can get away without acquiring the memcg->move_lock
1953 * because page moving starts with an RCU grace period.
1957 if (mem_cgroup_disabled())
1960 memcg
= folio_memcg(folio
);
1961 if (unlikely(!memcg
))
1964 #ifdef CONFIG_PROVE_LOCKING
1965 local_irq_save(flags
);
1966 might_lock(&memcg
->move_lock
);
1967 local_irq_restore(flags
);
1970 if (atomic_read(&memcg
->moving_account
) <= 0)
1973 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1974 if (memcg
!= folio_memcg(folio
)) {
1975 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1980 * When charge migration first begins, we can have multiple
1981 * critical sections holding the fast-path RCU lock and one
1982 * holding the slowpath move_lock. Track the task who has the
1983 * move_lock for unlock_page_memcg().
1985 memcg
->move_lock_task
= current
;
1986 memcg
->move_lock_flags
= flags
;
1988 EXPORT_SYMBOL(folio_memcg_lock
);
1990 void lock_page_memcg(struct page
*page
)
1992 folio_memcg_lock(page_folio(page
));
1994 EXPORT_SYMBOL(lock_page_memcg
);
1996 static void __folio_memcg_unlock(struct mem_cgroup
*memcg
)
1998 if (memcg
&& memcg
->move_lock_task
== current
) {
1999 unsigned long flags
= memcg
->move_lock_flags
;
2001 memcg
->move_lock_task
= NULL
;
2002 memcg
->move_lock_flags
= 0;
2004 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2011 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2012 * @folio: The folio.
2014 * This releases the binding created by folio_memcg_lock(). This does
2015 * not change the accounting of this folio to its memcg, but it does
2016 * permit others to change it.
2018 void folio_memcg_unlock(struct folio
*folio
)
2020 __folio_memcg_unlock(folio_memcg(folio
));
2022 EXPORT_SYMBOL(folio_memcg_unlock
);
2024 void unlock_page_memcg(struct page
*page
)
2026 folio_memcg_unlock(page_folio(page
));
2028 EXPORT_SYMBOL(unlock_page_memcg
);
2031 #ifdef CONFIG_MEMCG_KMEM
2032 struct obj_cgroup
*cached_objcg
;
2033 struct pglist_data
*cached_pgdat
;
2034 unsigned int nr_bytes
;
2035 int nr_slab_reclaimable_b
;
2036 int nr_slab_unreclaimable_b
;
2042 struct memcg_stock_pcp
{
2043 struct mem_cgroup
*cached
; /* this never be root cgroup */
2044 unsigned int nr_pages
;
2045 struct obj_stock task_obj
;
2046 struct obj_stock irq_obj
;
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
);
2055 #ifdef CONFIG_MEMCG_KMEM
2056 static void drain_obj_stock(struct obj_stock
*stock
);
2057 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2058 struct mem_cgroup
*root_memcg
);
2061 static inline void drain_obj_stock(struct obj_stock
*stock
)
2064 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2065 struct mem_cgroup
*root_memcg
)
2072 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2073 * sequence used in this case to access content from object stock is slow.
2074 * To optimize for user context access, there are now two object stocks for
2075 * task context and interrupt context access respectively.
2077 * The task context object stock can be accessed by disabling preemption only
2078 * which is cheap in non-preempt kernel. The interrupt context object stock
2079 * can only be accessed after disabling interrupt. User context code can
2080 * access interrupt object stock, but not vice versa.
2082 static inline struct obj_stock
*get_obj_stock(unsigned long *pflags
)
2084 struct memcg_stock_pcp
*stock
;
2086 if (likely(in_task())) {
2089 stock
= this_cpu_ptr(&memcg_stock
);
2090 return &stock
->task_obj
;
2093 local_irq_save(*pflags
);
2094 stock
= this_cpu_ptr(&memcg_stock
);
2095 return &stock
->irq_obj
;
2098 static inline void put_obj_stock(unsigned long flags
)
2100 if (likely(in_task()))
2103 local_irq_restore(flags
);
2107 * consume_stock: Try to consume stocked charge on this cpu.
2108 * @memcg: memcg to consume from.
2109 * @nr_pages: how many pages to charge.
2111 * The charges will only happen if @memcg matches the current cpu's memcg
2112 * stock, and at least @nr_pages are available in that stock. Failure to
2113 * service an allocation will refill the stock.
2115 * returns true if successful, false otherwise.
2117 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2119 struct memcg_stock_pcp
*stock
;
2120 unsigned long flags
;
2123 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2126 local_irq_save(flags
);
2128 stock
= this_cpu_ptr(&memcg_stock
);
2129 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2130 stock
->nr_pages
-= nr_pages
;
2134 local_irq_restore(flags
);
2140 * Returns stocks cached in percpu and reset cached information.
2142 static void drain_stock(struct memcg_stock_pcp
*stock
)
2144 struct mem_cgroup
*old
= stock
->cached
;
2149 if (stock
->nr_pages
) {
2150 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2151 if (do_memsw_account())
2152 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2153 stock
->nr_pages
= 0;
2157 stock
->cached
= NULL
;
2160 static void drain_local_stock(struct work_struct
*dummy
)
2162 struct memcg_stock_pcp
*stock
;
2163 unsigned long flags
;
2166 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2167 * drain_stock races is that we always operate on local CPU stock
2168 * here with IRQ disabled
2170 local_irq_save(flags
);
2172 stock
= this_cpu_ptr(&memcg_stock
);
2173 drain_obj_stock(&stock
->irq_obj
);
2175 drain_obj_stock(&stock
->task_obj
);
2177 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2179 local_irq_restore(flags
);
2183 * Cache charges(val) to local per_cpu area.
2184 * This will be consumed by consume_stock() function, later.
2186 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2188 struct memcg_stock_pcp
*stock
;
2189 unsigned long flags
;
2191 local_irq_save(flags
);
2193 stock
= this_cpu_ptr(&memcg_stock
);
2194 if (stock
->cached
!= memcg
) { /* reset if necessary */
2196 css_get(&memcg
->css
);
2197 stock
->cached
= memcg
;
2199 stock
->nr_pages
+= nr_pages
;
2201 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2204 local_irq_restore(flags
);
2208 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2209 * of the hierarchy under it.
2211 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2215 /* If someone's already draining, avoid adding running more workers. */
2216 if (!mutex_trylock(&percpu_charge_mutex
))
2219 * Notify other cpus that system-wide "drain" is running
2220 * We do not care about races with the cpu hotplug because cpu down
2221 * as well as workers from this path always operate on the local
2222 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2225 for_each_online_cpu(cpu
) {
2226 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2227 struct mem_cgroup
*memcg
;
2231 memcg
= stock
->cached
;
2232 if (memcg
&& stock
->nr_pages
&&
2233 mem_cgroup_is_descendant(memcg
, root_memcg
))
2235 else if (obj_stock_flush_required(stock
, root_memcg
))
2240 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2242 drain_local_stock(&stock
->work
);
2244 schedule_work_on(cpu
, &stock
->work
);
2248 mutex_unlock(&percpu_charge_mutex
);
2251 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2253 struct memcg_stock_pcp
*stock
;
2255 stock
= &per_cpu(memcg_stock
, cpu
);
2261 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2262 unsigned int nr_pages
,
2265 unsigned long nr_reclaimed
= 0;
2268 unsigned long pflags
;
2270 if (page_counter_read(&memcg
->memory
) <=
2271 READ_ONCE(memcg
->memory
.high
))
2274 memcg_memory_event(memcg
, MEMCG_HIGH
);
2276 psi_memstall_enter(&pflags
);
2277 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2279 psi_memstall_leave(&pflags
);
2280 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2281 !mem_cgroup_is_root(memcg
));
2283 return nr_reclaimed
;
2286 static void high_work_func(struct work_struct
*work
)
2288 struct mem_cgroup
*memcg
;
2290 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2291 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2295 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2296 * enough to still cause a significant slowdown in most cases, while still
2297 * allowing diagnostics and tracing to proceed without becoming stuck.
2299 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2302 * When calculating the delay, we use these either side of the exponentiation to
2303 * maintain precision and scale to a reasonable number of jiffies (see the table
2306 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2307 * overage ratio to a delay.
2308 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2309 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2310 * to produce a reasonable delay curve.
2312 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2313 * reasonable delay curve compared to precision-adjusted overage, not
2314 * penalising heavily at first, but still making sure that growth beyond the
2315 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2316 * example, with a high of 100 megabytes:
2318 * +-------+------------------------+
2319 * | usage | time to allocate in ms |
2320 * +-------+------------------------+
2342 * +-------+------------------------+
2344 #define MEMCG_DELAY_PRECISION_SHIFT 20
2345 #define MEMCG_DELAY_SCALING_SHIFT 14
2347 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2355 * Prevent division by 0 in overage calculation by acting as if
2356 * it was a threshold of 1 page
2358 high
= max(high
, 1UL);
2360 overage
= usage
- high
;
2361 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2362 return div64_u64(overage
, high
);
2365 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2367 u64 overage
, max_overage
= 0;
2370 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2371 READ_ONCE(memcg
->memory
.high
));
2372 max_overage
= max(overage
, max_overage
);
2373 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2374 !mem_cgroup_is_root(memcg
));
2379 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2381 u64 overage
, max_overage
= 0;
2384 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2385 READ_ONCE(memcg
->swap
.high
));
2387 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2388 max_overage
= max(overage
, max_overage
);
2389 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2390 !mem_cgroup_is_root(memcg
));
2396 * Get the number of jiffies that we should penalise a mischievous cgroup which
2397 * is exceeding its memory.high by checking both it and its ancestors.
2399 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2400 unsigned int nr_pages
,
2403 unsigned long penalty_jiffies
;
2409 * We use overage compared to memory.high to calculate the number of
2410 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2411 * fairly lenient on small overages, and increasingly harsh when the
2412 * memcg in question makes it clear that it has no intention of stopping
2413 * its crazy behaviour, so we exponentially increase the delay based on
2416 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2417 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2418 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2421 * Factor in the task's own contribution to the overage, such that four
2422 * N-sized allocations are throttled approximately the same as one
2423 * 4N-sized allocation.
2425 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2426 * larger the current charge patch is than that.
2428 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2432 * Scheduled by try_charge() to be executed from the userland return path
2433 * and reclaims memory over the high limit.
2435 void mem_cgroup_handle_over_high(void)
2437 unsigned long penalty_jiffies
;
2438 unsigned long pflags
;
2439 unsigned long nr_reclaimed
;
2440 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2441 int nr_retries
= MAX_RECLAIM_RETRIES
;
2442 struct mem_cgroup
*memcg
;
2443 bool in_retry
= false;
2445 if (likely(!nr_pages
))
2448 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2449 current
->memcg_nr_pages_over_high
= 0;
2453 * The allocating task should reclaim at least the batch size, but for
2454 * subsequent retries we only want to do what's necessary to prevent oom
2455 * or breaching resource isolation.
2457 * This is distinct from memory.max or page allocator behaviour because
2458 * memory.high is currently batched, whereas memory.max and the page
2459 * allocator run every time an allocation is made.
2461 nr_reclaimed
= reclaim_high(memcg
,
2462 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2466 * memory.high is breached and reclaim is unable to keep up. Throttle
2467 * allocators proactively to slow down excessive growth.
2469 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2470 mem_find_max_overage(memcg
));
2472 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2473 swap_find_max_overage(memcg
));
2476 * Clamp the max delay per usermode return so as to still keep the
2477 * application moving forwards and also permit diagnostics, albeit
2480 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2483 * Don't sleep if the amount of jiffies this memcg owes us is so low
2484 * that it's not even worth doing, in an attempt to be nice to those who
2485 * go only a small amount over their memory.high value and maybe haven't
2486 * been aggressively reclaimed enough yet.
2488 if (penalty_jiffies
<= HZ
/ 100)
2492 * If reclaim is making forward progress but we're still over
2493 * memory.high, we want to encourage that rather than doing allocator
2496 if (nr_reclaimed
|| nr_retries
--) {
2502 * If we exit early, we're guaranteed to die (since
2503 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2504 * need to account for any ill-begotten jiffies to pay them off later.
2506 psi_memstall_enter(&pflags
);
2507 schedule_timeout_killable(penalty_jiffies
);
2508 psi_memstall_leave(&pflags
);
2511 css_put(&memcg
->css
);
2514 static int try_charge_memcg(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2515 unsigned int nr_pages
)
2517 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2518 int nr_retries
= MAX_RECLAIM_RETRIES
;
2519 struct mem_cgroup
*mem_over_limit
;
2520 struct page_counter
*counter
;
2521 enum oom_status oom_status
;
2522 unsigned long nr_reclaimed
;
2523 bool may_swap
= true;
2524 bool drained
= false;
2525 unsigned long pflags
;
2528 if (consume_stock(memcg
, nr_pages
))
2531 if (!do_memsw_account() ||
2532 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2533 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2535 if (do_memsw_account())
2536 page_counter_uncharge(&memcg
->memsw
, batch
);
2537 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2539 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2543 if (batch
> nr_pages
) {
2549 * Memcg doesn't have a dedicated reserve for atomic
2550 * allocations. But like the global atomic pool, we need to
2551 * put the burden of reclaim on regular allocation requests
2552 * and let these go through as privileged allocations.
2554 if (gfp_mask
& __GFP_ATOMIC
)
2558 * Unlike in global OOM situations, memcg is not in a physical
2559 * memory shortage. Allow dying and OOM-killed tasks to
2560 * bypass the last charges so that they can exit quickly and
2561 * free their memory.
2563 if (unlikely(should_force_charge()))
2567 * Prevent unbounded recursion when reclaim operations need to
2568 * allocate memory. This might exceed the limits temporarily,
2569 * but we prefer facilitating memory reclaim and getting back
2570 * under the limit over triggering OOM kills in these cases.
2572 if (unlikely(current
->flags
& PF_MEMALLOC
))
2575 if (unlikely(task_in_memcg_oom(current
)))
2578 if (!gfpflags_allow_blocking(gfp_mask
))
2581 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2583 psi_memstall_enter(&pflags
);
2584 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2585 gfp_mask
, may_swap
);
2586 psi_memstall_leave(&pflags
);
2588 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2592 drain_all_stock(mem_over_limit
);
2597 if (gfp_mask
& __GFP_NORETRY
)
2600 * Even though the limit is exceeded at this point, reclaim
2601 * may have been able to free some pages. Retry the charge
2602 * before killing the task.
2604 * Only for regular pages, though: huge pages are rather
2605 * unlikely to succeed so close to the limit, and we fall back
2606 * to regular pages anyway in case of failure.
2608 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2611 * At task move, charge accounts can be doubly counted. So, it's
2612 * better to wait until the end of task_move if something is going on.
2614 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2620 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2623 if (fatal_signal_pending(current
))
2627 * keep retrying as long as the memcg oom killer is able to make
2628 * a forward progress or bypass the charge if the oom killer
2629 * couldn't make any progress.
2631 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2632 get_order(nr_pages
* PAGE_SIZE
));
2633 switch (oom_status
) {
2635 nr_retries
= MAX_RECLAIM_RETRIES
;
2643 if (!(gfp_mask
& __GFP_NOFAIL
))
2647 * The allocation either can't fail or will lead to more memory
2648 * being freed very soon. Allow memory usage go over the limit
2649 * temporarily by force charging it.
2651 page_counter_charge(&memcg
->memory
, nr_pages
);
2652 if (do_memsw_account())
2653 page_counter_charge(&memcg
->memsw
, nr_pages
);
2658 if (batch
> nr_pages
)
2659 refill_stock(memcg
, batch
- nr_pages
);
2662 * If the hierarchy is above the normal consumption range, schedule
2663 * reclaim on returning to userland. We can perform reclaim here
2664 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2665 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2666 * not recorded as it most likely matches current's and won't
2667 * change in the meantime. As high limit is checked again before
2668 * reclaim, the cost of mismatch is negligible.
2671 bool mem_high
, swap_high
;
2673 mem_high
= page_counter_read(&memcg
->memory
) >
2674 READ_ONCE(memcg
->memory
.high
);
2675 swap_high
= page_counter_read(&memcg
->swap
) >
2676 READ_ONCE(memcg
->swap
.high
);
2678 /* Don't bother a random interrupted task */
2679 if (in_interrupt()) {
2681 schedule_work(&memcg
->high_work
);
2687 if (mem_high
|| swap_high
) {
2689 * The allocating tasks in this cgroup will need to do
2690 * reclaim or be throttled to prevent further growth
2691 * of the memory or swap footprints.
2693 * Target some best-effort fairness between the tasks,
2694 * and distribute reclaim work and delay penalties
2695 * based on how much each task is actually allocating.
2697 current
->memcg_nr_pages_over_high
+= batch
;
2698 set_notify_resume(current
);
2701 } while ((memcg
= parent_mem_cgroup(memcg
)));
2706 static inline int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2707 unsigned int nr_pages
)
2709 if (mem_cgroup_is_root(memcg
))
2712 return try_charge_memcg(memcg
, gfp_mask
, nr_pages
);
2715 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2716 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2718 if (mem_cgroup_is_root(memcg
))
2721 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2722 if (do_memsw_account())
2723 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2727 static void commit_charge(struct folio
*folio
, struct mem_cgroup
*memcg
)
2729 VM_BUG_ON_FOLIO(folio_memcg(folio
), folio
);
2731 * Any of the following ensures page's memcg stability:
2735 * - lock_page_memcg()
2736 * - exclusive reference
2738 folio
->memcg_data
= (unsigned long)memcg
;
2741 static struct mem_cgroup
*get_mem_cgroup_from_objcg(struct obj_cgroup
*objcg
)
2743 struct mem_cgroup
*memcg
;
2747 memcg
= obj_cgroup_memcg(objcg
);
2748 if (unlikely(!css_tryget(&memcg
->css
)))
2755 #ifdef CONFIG_MEMCG_KMEM
2757 * The allocated objcg pointers array is not accounted directly.
2758 * Moreover, it should not come from DMA buffer and is not readily
2759 * reclaimable. So those GFP bits should be masked off.
2761 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2763 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2764 gfp_t gfp
, bool new_page
)
2766 unsigned int objects
= objs_per_slab_page(s
, page
);
2767 unsigned long memcg_data
;
2770 gfp
&= ~OBJCGS_CLEAR_MASK
;
2771 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2776 memcg_data
= (unsigned long) vec
| MEMCG_DATA_OBJCGS
;
2779 * If the slab page is brand new and nobody can yet access
2780 * it's memcg_data, no synchronization is required and
2781 * memcg_data can be simply assigned.
2783 page
->memcg_data
= memcg_data
;
2784 } else if (cmpxchg(&page
->memcg_data
, 0, memcg_data
)) {
2786 * If the slab page is already in use, somebody can allocate
2787 * and assign obj_cgroups in parallel. In this case the existing
2788 * objcg vector should be reused.
2794 kmemleak_not_leak(vec
);
2799 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2801 * A passed kernel object can be a slab object or a generic kernel page, so
2802 * different mechanisms for getting the memory cgroup pointer should be used.
2803 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2804 * can not know for sure how the kernel object is implemented.
2805 * mem_cgroup_from_obj() can be safely used in such cases.
2807 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2808 * cgroup_mutex, etc.
2810 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2814 if (mem_cgroup_disabled())
2817 page
= virt_to_head_page(p
);
2820 * Slab objects are accounted individually, not per-page.
2821 * Memcg membership data for each individual object is saved in
2822 * the page->obj_cgroups.
2824 if (page_objcgs_check(page
)) {
2825 struct obj_cgroup
*objcg
;
2828 off
= obj_to_index(page
->slab_cache
, page
, p
);
2829 objcg
= page_objcgs(page
)[off
];
2831 return obj_cgroup_memcg(objcg
);
2837 * page_memcg_check() is used here, because page_has_obj_cgroups()
2838 * check above could fail because the object cgroups vector wasn't set
2839 * at that moment, but it can be set concurrently.
2840 * page_memcg_check(page) will guarantee that a proper memory
2841 * cgroup pointer or NULL will be returned.
2843 return page_memcg_check(page
);
2846 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2848 struct obj_cgroup
*objcg
= NULL
;
2849 struct mem_cgroup
*memcg
;
2851 if (memcg_kmem_bypass())
2855 if (unlikely(active_memcg()))
2856 memcg
= active_memcg();
2858 memcg
= mem_cgroup_from_task(current
);
2860 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
2861 objcg
= rcu_dereference(memcg
->objcg
);
2862 if (objcg
&& obj_cgroup_tryget(objcg
))
2871 static int memcg_alloc_cache_id(void)
2876 id
= ida_simple_get(&memcg_cache_ida
,
2877 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2881 if (id
< memcg_nr_cache_ids
)
2885 * There's no space for the new id in memcg_caches arrays,
2886 * so we have to grow them.
2888 down_write(&memcg_cache_ids_sem
);
2890 size
= 2 * (id
+ 1);
2891 if (size
< MEMCG_CACHES_MIN_SIZE
)
2892 size
= MEMCG_CACHES_MIN_SIZE
;
2893 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2894 size
= MEMCG_CACHES_MAX_SIZE
;
2896 err
= memcg_update_all_list_lrus(size
);
2898 memcg_nr_cache_ids
= size
;
2900 up_write(&memcg_cache_ids_sem
);
2903 ida_simple_remove(&memcg_cache_ida
, id
);
2909 static void memcg_free_cache_id(int id
)
2911 ida_simple_remove(&memcg_cache_ida
, id
);
2915 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2916 * @objcg: object cgroup to uncharge
2917 * @nr_pages: number of pages to uncharge
2919 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
2920 unsigned int nr_pages
)
2922 struct mem_cgroup
*memcg
;
2924 memcg
= get_mem_cgroup_from_objcg(objcg
);
2926 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2927 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2928 refill_stock(memcg
, nr_pages
);
2930 css_put(&memcg
->css
);
2934 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2935 * @objcg: object cgroup to charge
2936 * @gfp: reclaim mode
2937 * @nr_pages: number of pages to charge
2939 * Returns 0 on success, an error code on failure.
2941 static int obj_cgroup_charge_pages(struct obj_cgroup
*objcg
, gfp_t gfp
,
2942 unsigned int nr_pages
)
2944 struct page_counter
*counter
;
2945 struct mem_cgroup
*memcg
;
2948 memcg
= get_mem_cgroup_from_objcg(objcg
);
2950 ret
= try_charge_memcg(memcg
, gfp
, nr_pages
);
2954 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2955 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2958 * Enforce __GFP_NOFAIL allocation because callers are not
2959 * prepared to see failures and likely do not have any failure
2962 if (gfp
& __GFP_NOFAIL
) {
2963 page_counter_charge(&memcg
->kmem
, nr_pages
);
2966 cancel_charge(memcg
, nr_pages
);
2970 css_put(&memcg
->css
);
2976 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
2977 * @page: page to charge
2978 * @gfp: reclaim mode
2979 * @order: allocation order
2981 * Returns 0 on success, an error code on failure.
2983 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
2985 struct obj_cgroup
*objcg
;
2988 objcg
= get_obj_cgroup_from_current();
2990 ret
= obj_cgroup_charge_pages(objcg
, gfp
, 1 << order
);
2992 page
->memcg_data
= (unsigned long)objcg
|
2996 obj_cgroup_put(objcg
);
3002 * __memcg_kmem_uncharge_page: uncharge a kmem page
3003 * @page: page to uncharge
3004 * @order: allocation order
3006 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3008 struct folio
*folio
= page_folio(page
);
3009 struct obj_cgroup
*objcg
;
3010 unsigned int nr_pages
= 1 << order
;
3012 if (!folio_memcg_kmem(folio
))
3015 objcg
= __folio_objcg(folio
);
3016 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3017 folio
->memcg_data
= 0;
3018 obj_cgroup_put(objcg
);
3021 void mod_objcg_state(struct obj_cgroup
*objcg
, struct pglist_data
*pgdat
,
3022 enum node_stat_item idx
, int nr
)
3024 unsigned long flags
;
3025 struct obj_stock
*stock
= get_obj_stock(&flags
);
3029 * Save vmstat data in stock and skip vmstat array update unless
3030 * accumulating over a page of vmstat data or when pgdat or idx
3033 if (stock
->cached_objcg
!= objcg
) {
3034 drain_obj_stock(stock
);
3035 obj_cgroup_get(objcg
);
3036 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3037 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3038 stock
->cached_objcg
= objcg
;
3039 stock
->cached_pgdat
= pgdat
;
3040 } else if (stock
->cached_pgdat
!= pgdat
) {
3041 /* Flush the existing cached vmstat data */
3042 struct pglist_data
*oldpg
= stock
->cached_pgdat
;
3044 if (stock
->nr_slab_reclaimable_b
) {
3045 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_RECLAIMABLE_B
,
3046 stock
->nr_slab_reclaimable_b
);
3047 stock
->nr_slab_reclaimable_b
= 0;
3049 if (stock
->nr_slab_unreclaimable_b
) {
3050 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_UNRECLAIMABLE_B
,
3051 stock
->nr_slab_unreclaimable_b
);
3052 stock
->nr_slab_unreclaimable_b
= 0;
3054 stock
->cached_pgdat
= pgdat
;
3057 bytes
= (idx
== NR_SLAB_RECLAIMABLE_B
) ? &stock
->nr_slab_reclaimable_b
3058 : &stock
->nr_slab_unreclaimable_b
;
3060 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3061 * cached locally at least once before pushing it out.
3068 if (abs(*bytes
) > PAGE_SIZE
) {
3076 mod_objcg_mlstate(objcg
, pgdat
, idx
, nr
);
3078 put_obj_stock(flags
);
3081 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3083 unsigned long flags
;
3084 struct obj_stock
*stock
= get_obj_stock(&flags
);
3087 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3088 stock
->nr_bytes
-= nr_bytes
;
3092 put_obj_stock(flags
);
3097 static void drain_obj_stock(struct obj_stock
*stock
)
3099 struct obj_cgroup
*old
= stock
->cached_objcg
;
3104 if (stock
->nr_bytes
) {
3105 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3106 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3109 obj_cgroup_uncharge_pages(old
, nr_pages
);
3112 * The leftover is flushed to the centralized per-memcg value.
3113 * On the next attempt to refill obj stock it will be moved
3114 * to a per-cpu stock (probably, on an other CPU), see
3115 * refill_obj_stock().
3117 * How often it's flushed is a trade-off between the memory
3118 * limit enforcement accuracy and potential CPU contention,
3119 * so it might be changed in the future.
3121 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3122 stock
->nr_bytes
= 0;
3126 * Flush the vmstat data in current stock
3128 if (stock
->nr_slab_reclaimable_b
|| stock
->nr_slab_unreclaimable_b
) {
3129 if (stock
->nr_slab_reclaimable_b
) {
3130 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3131 NR_SLAB_RECLAIMABLE_B
,
3132 stock
->nr_slab_reclaimable_b
);
3133 stock
->nr_slab_reclaimable_b
= 0;
3135 if (stock
->nr_slab_unreclaimable_b
) {
3136 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3137 NR_SLAB_UNRECLAIMABLE_B
,
3138 stock
->nr_slab_unreclaimable_b
);
3139 stock
->nr_slab_unreclaimable_b
= 0;
3141 stock
->cached_pgdat
= NULL
;
3144 obj_cgroup_put(old
);
3145 stock
->cached_objcg
= NULL
;
3148 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3149 struct mem_cgroup
*root_memcg
)
3151 struct mem_cgroup
*memcg
;
3153 if (in_task() && stock
->task_obj
.cached_objcg
) {
3154 memcg
= obj_cgroup_memcg(stock
->task_obj
.cached_objcg
);
3155 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3158 if (stock
->irq_obj
.cached_objcg
) {
3159 memcg
= obj_cgroup_memcg(stock
->irq_obj
.cached_objcg
);
3160 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3167 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
,
3168 bool allow_uncharge
)
3170 unsigned long flags
;
3171 struct obj_stock
*stock
= get_obj_stock(&flags
);
3172 unsigned int nr_pages
= 0;
3174 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3175 drain_obj_stock(stock
);
3176 obj_cgroup_get(objcg
);
3177 stock
->cached_objcg
= objcg
;
3178 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3179 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3180 allow_uncharge
= true; /* Allow uncharge when objcg changes */
3182 stock
->nr_bytes
+= nr_bytes
;
3184 if (allow_uncharge
&& (stock
->nr_bytes
> PAGE_SIZE
)) {
3185 nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3186 stock
->nr_bytes
&= (PAGE_SIZE
- 1);
3189 put_obj_stock(flags
);
3192 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3195 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3197 unsigned int nr_pages
, nr_bytes
;
3200 if (consume_obj_stock(objcg
, size
))
3204 * In theory, objcg->nr_charged_bytes can have enough
3205 * pre-charged bytes to satisfy the allocation. However,
3206 * flushing objcg->nr_charged_bytes requires two atomic
3207 * operations, and objcg->nr_charged_bytes can't be big.
3208 * The shared objcg->nr_charged_bytes can also become a
3209 * performance bottleneck if all tasks of the same memcg are
3210 * trying to update it. So it's better to ignore it and try
3211 * grab some new pages. The stock's nr_bytes will be flushed to
3212 * objcg->nr_charged_bytes later on when objcg changes.
3214 * The stock's nr_bytes may contain enough pre-charged bytes
3215 * to allow one less page from being charged, but we can't rely
3216 * on the pre-charged bytes not being changed outside of
3217 * consume_obj_stock() or refill_obj_stock(). So ignore those
3218 * pre-charged bytes as well when charging pages. To avoid a
3219 * page uncharge right after a page charge, we set the
3220 * allow_uncharge flag to false when calling refill_obj_stock()
3221 * to temporarily allow the pre-charged bytes to exceed the page
3222 * size limit. The maximum reachable value of the pre-charged
3223 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3226 nr_pages
= size
>> PAGE_SHIFT
;
3227 nr_bytes
= size
& (PAGE_SIZE
- 1);
3232 ret
= obj_cgroup_charge_pages(objcg
, gfp
, nr_pages
);
3233 if (!ret
&& nr_bytes
)
3234 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
, false);
3239 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3241 refill_obj_stock(objcg
, size
, true);
3244 #endif /* CONFIG_MEMCG_KMEM */
3247 * Because page_memcg(head) is not set on tails, set it now.
3249 void split_page_memcg(struct page
*head
, unsigned int nr
)
3251 struct folio
*folio
= page_folio(head
);
3252 struct mem_cgroup
*memcg
= folio_memcg(folio
);
3255 if (mem_cgroup_disabled() || !memcg
)
3258 for (i
= 1; i
< nr
; i
++)
3259 folio_page(folio
, i
)->memcg_data
= folio
->memcg_data
;
3261 if (folio_memcg_kmem(folio
))
3262 obj_cgroup_get_many(__folio_objcg(folio
), nr
- 1);
3264 css_get_many(&memcg
->css
, nr
- 1);
3267 #ifdef CONFIG_MEMCG_SWAP
3269 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3270 * @entry: swap entry to be moved
3271 * @from: mem_cgroup which the entry is moved from
3272 * @to: mem_cgroup which the entry is moved to
3274 * It succeeds only when the swap_cgroup's record for this entry is the same
3275 * as the mem_cgroup's id of @from.
3277 * Returns 0 on success, -EINVAL on failure.
3279 * The caller must have charged to @to, IOW, called page_counter_charge() about
3280 * both res and memsw, and called css_get().
3282 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3283 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3285 unsigned short old_id
, new_id
;
3287 old_id
= mem_cgroup_id(from
);
3288 new_id
= mem_cgroup_id(to
);
3290 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3291 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3292 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3298 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3299 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3305 static DEFINE_MUTEX(memcg_max_mutex
);
3307 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3308 unsigned long max
, bool memsw
)
3310 bool enlarge
= false;
3311 bool drained
= false;
3313 bool limits_invariant
;
3314 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3317 if (signal_pending(current
)) {
3322 mutex_lock(&memcg_max_mutex
);
3324 * Make sure that the new limit (memsw or memory limit) doesn't
3325 * break our basic invariant rule memory.max <= memsw.max.
3327 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3328 max
<= memcg
->memsw
.max
;
3329 if (!limits_invariant
) {
3330 mutex_unlock(&memcg_max_mutex
);
3334 if (max
> counter
->max
)
3336 ret
= page_counter_set_max(counter
, max
);
3337 mutex_unlock(&memcg_max_mutex
);
3343 drain_all_stock(memcg
);
3348 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3349 GFP_KERNEL
, !memsw
)) {
3355 if (!ret
&& enlarge
)
3356 memcg_oom_recover(memcg
);
3361 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3363 unsigned long *total_scanned
)
3365 unsigned long nr_reclaimed
= 0;
3366 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3367 unsigned long reclaimed
;
3369 struct mem_cgroup_tree_per_node
*mctz
;
3370 unsigned long excess
;
3371 unsigned long nr_scanned
;
3376 mctz
= soft_limit_tree
.rb_tree_per_node
[pgdat
->node_id
];
3379 * Do not even bother to check the largest node if the root
3380 * is empty. Do it lockless to prevent lock bouncing. Races
3381 * are acceptable as soft limit is best effort anyway.
3383 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3387 * This loop can run a while, specially if mem_cgroup's continuously
3388 * keep exceeding their soft limit and putting the system under
3395 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3400 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3401 gfp_mask
, &nr_scanned
);
3402 nr_reclaimed
+= reclaimed
;
3403 *total_scanned
+= nr_scanned
;
3404 spin_lock_irq(&mctz
->lock
);
3405 __mem_cgroup_remove_exceeded(mz
, mctz
);
3408 * If we failed to reclaim anything from this memory cgroup
3409 * it is time to move on to the next cgroup
3413 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3415 excess
= soft_limit_excess(mz
->memcg
);
3417 * One school of thought says that we should not add
3418 * back the node to the tree if reclaim returns 0.
3419 * But our reclaim could return 0, simply because due
3420 * to priority we are exposing a smaller subset of
3421 * memory to reclaim from. Consider this as a longer
3424 /* If excess == 0, no tree ops */
3425 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3426 spin_unlock_irq(&mctz
->lock
);
3427 css_put(&mz
->memcg
->css
);
3430 * Could not reclaim anything and there are no more
3431 * mem cgroups to try or we seem to be looping without
3432 * reclaiming anything.
3434 if (!nr_reclaimed
&&
3436 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3438 } while (!nr_reclaimed
);
3440 css_put(&next_mz
->memcg
->css
);
3441 return nr_reclaimed
;
3445 * Reclaims as many pages from the given memcg as possible.
3447 * Caller is responsible for holding css reference for memcg.
3449 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3451 int nr_retries
= MAX_RECLAIM_RETRIES
;
3453 /* we call try-to-free pages for make this cgroup empty */
3454 lru_add_drain_all();
3456 drain_all_stock(memcg
);
3458 /* try to free all pages in this cgroup */
3459 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3462 if (signal_pending(current
))
3465 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3469 /* maybe some writeback is necessary */
3470 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3478 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3479 char *buf
, size_t nbytes
,
3482 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3484 if (mem_cgroup_is_root(memcg
))
3486 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3489 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3495 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3496 struct cftype
*cft
, u64 val
)
3501 pr_warn_once("Non-hierarchical mode is deprecated. "
3502 "Please report your usecase to linux-mm@kvack.org if you "
3503 "depend on this functionality.\n");
3508 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3512 if (mem_cgroup_is_root(memcg
)) {
3513 /* mem_cgroup_threshold() calls here from irqsafe context */
3514 cgroup_rstat_flush_irqsafe(memcg
->css
.cgroup
);
3515 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3516 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3518 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3521 val
= page_counter_read(&memcg
->memory
);
3523 val
= page_counter_read(&memcg
->memsw
);
3536 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3539 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3540 struct page_counter
*counter
;
3542 switch (MEMFILE_TYPE(cft
->private)) {
3544 counter
= &memcg
->memory
;
3547 counter
= &memcg
->memsw
;
3550 counter
= &memcg
->kmem
;
3553 counter
= &memcg
->tcpmem
;
3559 switch (MEMFILE_ATTR(cft
->private)) {
3561 if (counter
== &memcg
->memory
)
3562 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3563 if (counter
== &memcg
->memsw
)
3564 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3565 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3567 return (u64
)counter
->max
* PAGE_SIZE
;
3569 return (u64
)counter
->watermark
* PAGE_SIZE
;
3571 return counter
->failcnt
;
3572 case RES_SOFT_LIMIT
:
3573 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3579 #ifdef CONFIG_MEMCG_KMEM
3580 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3582 struct obj_cgroup
*objcg
;
3585 if (cgroup_memory_nokmem
)
3588 BUG_ON(memcg
->kmemcg_id
>= 0);
3589 BUG_ON(memcg
->kmem_state
);
3591 memcg_id
= memcg_alloc_cache_id();
3595 objcg
= obj_cgroup_alloc();
3597 memcg_free_cache_id(memcg_id
);
3600 objcg
->memcg
= memcg
;
3601 rcu_assign_pointer(memcg
->objcg
, objcg
);
3603 static_branch_enable(&memcg_kmem_enabled_key
);
3605 memcg
->kmemcg_id
= memcg_id
;
3606 memcg
->kmem_state
= KMEM_ONLINE
;
3611 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3613 struct cgroup_subsys_state
*css
;
3614 struct mem_cgroup
*parent
, *child
;
3617 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3620 memcg
->kmem_state
= KMEM_ALLOCATED
;
3622 parent
= parent_mem_cgroup(memcg
);
3624 parent
= root_mem_cgroup
;
3626 memcg_reparent_objcgs(memcg
, parent
);
3628 kmemcg_id
= memcg
->kmemcg_id
;
3629 BUG_ON(kmemcg_id
< 0);
3632 * Change kmemcg_id of this cgroup and all its descendants to the
3633 * parent's id, and then move all entries from this cgroup's list_lrus
3634 * to ones of the parent. After we have finished, all list_lrus
3635 * corresponding to this cgroup are guaranteed to remain empty. The
3636 * ordering is imposed by list_lru_node->lock taken by
3637 * memcg_drain_all_list_lrus().
3639 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3640 css_for_each_descendant_pre(css
, &memcg
->css
) {
3641 child
= mem_cgroup_from_css(css
);
3642 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3643 child
->kmemcg_id
= parent
->kmemcg_id
;
3647 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3649 memcg_free_cache_id(kmemcg_id
);
3652 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3654 /* css_alloc() failed, offlining didn't happen */
3655 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3656 memcg_offline_kmem(memcg
);
3659 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3663 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3666 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3669 #endif /* CONFIG_MEMCG_KMEM */
3671 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3676 mutex_lock(&memcg_max_mutex
);
3677 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3678 mutex_unlock(&memcg_max_mutex
);
3682 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3686 mutex_lock(&memcg_max_mutex
);
3688 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3692 if (!memcg
->tcpmem_active
) {
3694 * The active flag needs to be written after the static_key
3695 * update. This is what guarantees that the socket activation
3696 * function is the last one to run. See mem_cgroup_sk_alloc()
3697 * for details, and note that we don't mark any socket as
3698 * belonging to this memcg until that flag is up.
3700 * We need to do this, because static_keys will span multiple
3701 * sites, but we can't control their order. If we mark a socket
3702 * as accounted, but the accounting functions are not patched in
3703 * yet, we'll lose accounting.
3705 * We never race with the readers in mem_cgroup_sk_alloc(),
3706 * because when this value change, the code to process it is not
3709 static_branch_inc(&memcg_sockets_enabled_key
);
3710 memcg
->tcpmem_active
= true;
3713 mutex_unlock(&memcg_max_mutex
);
3718 * The user of this function is...
3721 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3722 char *buf
, size_t nbytes
, loff_t off
)
3724 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3725 unsigned long nr_pages
;
3728 buf
= strstrip(buf
);
3729 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3733 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3735 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3739 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3741 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3744 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3747 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3748 "Please report your usecase to linux-mm@kvack.org if you "
3749 "depend on this functionality.\n");
3750 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3753 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3757 case RES_SOFT_LIMIT
:
3758 memcg
->soft_limit
= nr_pages
;
3762 return ret
?: nbytes
;
3765 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3766 size_t nbytes
, loff_t off
)
3768 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3769 struct page_counter
*counter
;
3771 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3773 counter
= &memcg
->memory
;
3776 counter
= &memcg
->memsw
;
3779 counter
= &memcg
->kmem
;
3782 counter
= &memcg
->tcpmem
;
3788 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3790 page_counter_reset_watermark(counter
);
3793 counter
->failcnt
= 0;
3802 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3805 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3809 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3810 struct cftype
*cft
, u64 val
)
3812 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3814 if (val
& ~MOVE_MASK
)
3818 * No kind of locking is needed in here, because ->can_attach() will
3819 * check this value once in the beginning of the process, and then carry
3820 * on with stale data. This means that changes to this value will only
3821 * affect task migrations starting after the change.
3823 memcg
->move_charge_at_immigrate
= val
;
3827 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3828 struct cftype
*cft
, u64 val
)
3836 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3837 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3838 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3840 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3841 int nid
, unsigned int lru_mask
, bool tree
)
3843 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3844 unsigned long nr
= 0;
3847 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3850 if (!(BIT(lru
) & lru_mask
))
3853 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3855 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3860 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3861 unsigned int lru_mask
,
3864 unsigned long nr
= 0;
3868 if (!(BIT(lru
) & lru_mask
))
3871 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3873 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3878 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3882 unsigned int lru_mask
;
3885 static const struct numa_stat stats
[] = {
3886 { "total", LRU_ALL
},
3887 { "file", LRU_ALL_FILE
},
3888 { "anon", LRU_ALL_ANON
},
3889 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3891 const struct numa_stat
*stat
;
3893 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3895 cgroup_rstat_flush(memcg
->css
.cgroup
);
3897 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3898 seq_printf(m
, "%s=%lu", stat
->name
,
3899 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3901 for_each_node_state(nid
, N_MEMORY
)
3902 seq_printf(m
, " N%d=%lu", nid
,
3903 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3904 stat
->lru_mask
, false));
3908 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3910 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
3911 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3913 for_each_node_state(nid
, N_MEMORY
)
3914 seq_printf(m
, " N%d=%lu", nid
,
3915 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3916 stat
->lru_mask
, true));
3922 #endif /* CONFIG_NUMA */
3924 static const unsigned int memcg1_stats
[] = {
3927 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3937 static const char *const memcg1_stat_names
[] = {
3940 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3950 /* Universal VM events cgroup1 shows, original sort order */
3951 static const unsigned int memcg1_events
[] = {
3958 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3960 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3961 unsigned long memory
, memsw
;
3962 struct mem_cgroup
*mi
;
3965 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
3967 cgroup_rstat_flush(memcg
->css
.cgroup
);
3969 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
3972 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
3974 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
3975 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
3978 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
3979 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
3980 memcg_events_local(memcg
, memcg1_events
[i
]));
3982 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3983 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
3984 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
3987 /* Hierarchical information */
3988 memory
= memsw
= PAGE_COUNTER_MAX
;
3989 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3990 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
3991 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
3993 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3994 (u64
)memory
* PAGE_SIZE
);
3995 if (do_memsw_account())
3996 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3997 (u64
)memsw
* PAGE_SIZE
);
3999 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4002 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4004 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4005 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4006 (u64
)nr
* PAGE_SIZE
);
4009 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4010 seq_printf(m
, "total_%s %llu\n",
4011 vm_event_name(memcg1_events
[i
]),
4012 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4014 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4015 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4016 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4019 #ifdef CONFIG_DEBUG_VM
4022 struct mem_cgroup_per_node
*mz
;
4023 unsigned long anon_cost
= 0;
4024 unsigned long file_cost
= 0;
4026 for_each_online_pgdat(pgdat
) {
4027 mz
= memcg
->nodeinfo
[pgdat
->node_id
];
4029 anon_cost
+= mz
->lruvec
.anon_cost
;
4030 file_cost
+= mz
->lruvec
.file_cost
;
4032 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4033 seq_printf(m
, "file_cost %lu\n", file_cost
);
4040 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4043 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4045 return mem_cgroup_swappiness(memcg
);
4048 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4049 struct cftype
*cft
, u64 val
)
4051 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4056 if (!mem_cgroup_is_root(memcg
))
4057 memcg
->swappiness
= val
;
4059 vm_swappiness
= val
;
4064 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4066 struct mem_cgroup_threshold_ary
*t
;
4067 unsigned long usage
;
4072 t
= rcu_dereference(memcg
->thresholds
.primary
);
4074 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4079 usage
= mem_cgroup_usage(memcg
, swap
);
4082 * current_threshold points to threshold just below or equal to usage.
4083 * If it's not true, a threshold was crossed after last
4084 * call of __mem_cgroup_threshold().
4086 i
= t
->current_threshold
;
4089 * Iterate backward over array of thresholds starting from
4090 * current_threshold and check if a threshold is crossed.
4091 * If none of thresholds below usage is crossed, we read
4092 * only one element of the array here.
4094 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4095 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4097 /* i = current_threshold + 1 */
4101 * Iterate forward over array of thresholds starting from
4102 * current_threshold+1 and check if a threshold is crossed.
4103 * If none of thresholds above usage is crossed, we read
4104 * only one element of the array here.
4106 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4107 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4109 /* Update current_threshold */
4110 t
->current_threshold
= i
- 1;
4115 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4118 __mem_cgroup_threshold(memcg
, false);
4119 if (do_memsw_account())
4120 __mem_cgroup_threshold(memcg
, true);
4122 memcg
= parent_mem_cgroup(memcg
);
4126 static int compare_thresholds(const void *a
, const void *b
)
4128 const struct mem_cgroup_threshold
*_a
= a
;
4129 const struct mem_cgroup_threshold
*_b
= b
;
4131 if (_a
->threshold
> _b
->threshold
)
4134 if (_a
->threshold
< _b
->threshold
)
4140 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4142 struct mem_cgroup_eventfd_list
*ev
;
4144 spin_lock(&memcg_oom_lock
);
4146 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4147 eventfd_signal(ev
->eventfd
, 1);
4149 spin_unlock(&memcg_oom_lock
);
4153 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4155 struct mem_cgroup
*iter
;
4157 for_each_mem_cgroup_tree(iter
, memcg
)
4158 mem_cgroup_oom_notify_cb(iter
);
4161 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4162 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4164 struct mem_cgroup_thresholds
*thresholds
;
4165 struct mem_cgroup_threshold_ary
*new;
4166 unsigned long threshold
;
4167 unsigned long usage
;
4170 ret
= page_counter_memparse(args
, "-1", &threshold
);
4174 mutex_lock(&memcg
->thresholds_lock
);
4177 thresholds
= &memcg
->thresholds
;
4178 usage
= mem_cgroup_usage(memcg
, false);
4179 } else if (type
== _MEMSWAP
) {
4180 thresholds
= &memcg
->memsw_thresholds
;
4181 usage
= mem_cgroup_usage(memcg
, true);
4185 /* Check if a threshold crossed before adding a new one */
4186 if (thresholds
->primary
)
4187 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4189 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4191 /* Allocate memory for new array of thresholds */
4192 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4199 /* Copy thresholds (if any) to new array */
4200 if (thresholds
->primary
)
4201 memcpy(new->entries
, thresholds
->primary
->entries
,
4202 flex_array_size(new, entries
, size
- 1));
4204 /* Add new threshold */
4205 new->entries
[size
- 1].eventfd
= eventfd
;
4206 new->entries
[size
- 1].threshold
= threshold
;
4208 /* Sort thresholds. Registering of new threshold isn't time-critical */
4209 sort(new->entries
, size
, sizeof(*new->entries
),
4210 compare_thresholds
, NULL
);
4212 /* Find current threshold */
4213 new->current_threshold
= -1;
4214 for (i
= 0; i
< size
; i
++) {
4215 if (new->entries
[i
].threshold
<= usage
) {
4217 * new->current_threshold will not be used until
4218 * rcu_assign_pointer(), so it's safe to increment
4221 ++new->current_threshold
;
4226 /* Free old spare buffer and save old primary buffer as spare */
4227 kfree(thresholds
->spare
);
4228 thresholds
->spare
= thresholds
->primary
;
4230 rcu_assign_pointer(thresholds
->primary
, new);
4232 /* To be sure that nobody uses thresholds */
4236 mutex_unlock(&memcg
->thresholds_lock
);
4241 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4242 struct eventfd_ctx
*eventfd
, const char *args
)
4244 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4247 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4248 struct eventfd_ctx
*eventfd
, const char *args
)
4250 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4253 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4254 struct eventfd_ctx
*eventfd
, enum res_type type
)
4256 struct mem_cgroup_thresholds
*thresholds
;
4257 struct mem_cgroup_threshold_ary
*new;
4258 unsigned long usage
;
4259 int i
, j
, size
, entries
;
4261 mutex_lock(&memcg
->thresholds_lock
);
4264 thresholds
= &memcg
->thresholds
;
4265 usage
= mem_cgroup_usage(memcg
, false);
4266 } else if (type
== _MEMSWAP
) {
4267 thresholds
= &memcg
->memsw_thresholds
;
4268 usage
= mem_cgroup_usage(memcg
, true);
4272 if (!thresholds
->primary
)
4275 /* Check if a threshold crossed before removing */
4276 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4278 /* Calculate new number of threshold */
4280 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4281 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4287 new = thresholds
->spare
;
4289 /* If no items related to eventfd have been cleared, nothing to do */
4293 /* Set thresholds array to NULL if we don't have thresholds */
4302 /* Copy thresholds and find current threshold */
4303 new->current_threshold
= -1;
4304 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4305 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4308 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4309 if (new->entries
[j
].threshold
<= usage
) {
4311 * new->current_threshold will not be used
4312 * until rcu_assign_pointer(), so it's safe to increment
4315 ++new->current_threshold
;
4321 /* Swap primary and spare array */
4322 thresholds
->spare
= thresholds
->primary
;
4324 rcu_assign_pointer(thresholds
->primary
, new);
4326 /* To be sure that nobody uses thresholds */
4329 /* If all events are unregistered, free the spare array */
4331 kfree(thresholds
->spare
);
4332 thresholds
->spare
= NULL
;
4335 mutex_unlock(&memcg
->thresholds_lock
);
4338 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4339 struct eventfd_ctx
*eventfd
)
4341 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4344 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4345 struct eventfd_ctx
*eventfd
)
4347 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4350 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4351 struct eventfd_ctx
*eventfd
, const char *args
)
4353 struct mem_cgroup_eventfd_list
*event
;
4355 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4359 spin_lock(&memcg_oom_lock
);
4361 event
->eventfd
= eventfd
;
4362 list_add(&event
->list
, &memcg
->oom_notify
);
4364 /* already in OOM ? */
4365 if (memcg
->under_oom
)
4366 eventfd_signal(eventfd
, 1);
4367 spin_unlock(&memcg_oom_lock
);
4372 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4373 struct eventfd_ctx
*eventfd
)
4375 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4377 spin_lock(&memcg_oom_lock
);
4379 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4380 if (ev
->eventfd
== eventfd
) {
4381 list_del(&ev
->list
);
4386 spin_unlock(&memcg_oom_lock
);
4389 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4391 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4393 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4394 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4395 seq_printf(sf
, "oom_kill %lu\n",
4396 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4400 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4401 struct cftype
*cft
, u64 val
)
4403 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4405 /* cannot set to root cgroup and only 0 and 1 are allowed */
4406 if (mem_cgroup_is_root(memcg
) || !((val
== 0) || (val
== 1)))
4409 memcg
->oom_kill_disable
= val
;
4411 memcg_oom_recover(memcg
);
4416 #ifdef CONFIG_CGROUP_WRITEBACK
4418 #include <trace/events/writeback.h>
4420 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4422 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4425 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4427 wb_domain_exit(&memcg
->cgwb_domain
);
4430 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4432 wb_domain_size_changed(&memcg
->cgwb_domain
);
4435 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4437 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4439 if (!memcg
->css
.parent
)
4442 return &memcg
->cgwb_domain
;
4446 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4447 * @wb: bdi_writeback in question
4448 * @pfilepages: out parameter for number of file pages
4449 * @pheadroom: out parameter for number of allocatable pages according to memcg
4450 * @pdirty: out parameter for number of dirty pages
4451 * @pwriteback: out parameter for number of pages under writeback
4453 * Determine the numbers of file, headroom, dirty, and writeback pages in
4454 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4455 * is a bit more involved.
4457 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4458 * headroom is calculated as the lowest headroom of itself and the
4459 * ancestors. Note that this doesn't consider the actual amount of
4460 * available memory in the system. The caller should further cap
4461 * *@pheadroom accordingly.
4463 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4464 unsigned long *pheadroom
, unsigned long *pdirty
,
4465 unsigned long *pwriteback
)
4467 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4468 struct mem_cgroup
*parent
;
4470 cgroup_rstat_flush_irqsafe(memcg
->css
.cgroup
);
4472 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
4473 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
4474 *pfilepages
= memcg_page_state(memcg
, NR_INACTIVE_FILE
) +
4475 memcg_page_state(memcg
, NR_ACTIVE_FILE
);
4477 *pheadroom
= PAGE_COUNTER_MAX
;
4478 while ((parent
= parent_mem_cgroup(memcg
))) {
4479 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4480 READ_ONCE(memcg
->memory
.high
));
4481 unsigned long used
= page_counter_read(&memcg
->memory
);
4483 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4489 * Foreign dirty flushing
4491 * There's an inherent mismatch between memcg and writeback. The former
4492 * tracks ownership per-page while the latter per-inode. This was a
4493 * deliberate design decision because honoring per-page ownership in the
4494 * writeback path is complicated, may lead to higher CPU and IO overheads
4495 * and deemed unnecessary given that write-sharing an inode across
4496 * different cgroups isn't a common use-case.
4498 * Combined with inode majority-writer ownership switching, this works well
4499 * enough in most cases but there are some pathological cases. For
4500 * example, let's say there are two cgroups A and B which keep writing to
4501 * different but confined parts of the same inode. B owns the inode and
4502 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4503 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4504 * triggering background writeback. A will be slowed down without a way to
4505 * make writeback of the dirty pages happen.
4507 * Conditions like the above can lead to a cgroup getting repeatedly and
4508 * severely throttled after making some progress after each
4509 * dirty_expire_interval while the underlying IO device is almost
4512 * Solving this problem completely requires matching the ownership tracking
4513 * granularities between memcg and writeback in either direction. However,
4514 * the more egregious behaviors can be avoided by simply remembering the
4515 * most recent foreign dirtying events and initiating remote flushes on
4516 * them when local writeback isn't enough to keep the memory clean enough.
4518 * The following two functions implement such mechanism. When a foreign
4519 * page - a page whose memcg and writeback ownerships don't match - is
4520 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4521 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4522 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4523 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4524 * foreign bdi_writebacks which haven't expired. Both the numbers of
4525 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4526 * limited to MEMCG_CGWB_FRN_CNT.
4528 * The mechanism only remembers IDs and doesn't hold any object references.
4529 * As being wrong occasionally doesn't matter, updates and accesses to the
4530 * records are lockless and racy.
4532 void mem_cgroup_track_foreign_dirty_slowpath(struct folio
*folio
,
4533 struct bdi_writeback
*wb
)
4535 struct mem_cgroup
*memcg
= folio_memcg(folio
);
4536 struct memcg_cgwb_frn
*frn
;
4537 u64 now
= get_jiffies_64();
4538 u64 oldest_at
= now
;
4542 trace_track_foreign_dirty(folio
, wb
);
4545 * Pick the slot to use. If there is already a slot for @wb, keep
4546 * using it. If not replace the oldest one which isn't being
4549 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4550 frn
= &memcg
->cgwb_frn
[i
];
4551 if (frn
->bdi_id
== wb
->bdi
->id
&&
4552 frn
->memcg_id
== wb
->memcg_css
->id
)
4554 if (time_before64(frn
->at
, oldest_at
) &&
4555 atomic_read(&frn
->done
.cnt
) == 1) {
4557 oldest_at
= frn
->at
;
4561 if (i
< MEMCG_CGWB_FRN_CNT
) {
4563 * Re-using an existing one. Update timestamp lazily to
4564 * avoid making the cacheline hot. We want them to be
4565 * reasonably up-to-date and significantly shorter than
4566 * dirty_expire_interval as that's what expires the record.
4567 * Use the shorter of 1s and dirty_expire_interval / 8.
4569 unsigned long update_intv
=
4570 min_t(unsigned long, HZ
,
4571 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4573 if (time_before64(frn
->at
, now
- update_intv
))
4575 } else if (oldest
>= 0) {
4576 /* replace the oldest free one */
4577 frn
= &memcg
->cgwb_frn
[oldest
];
4578 frn
->bdi_id
= wb
->bdi
->id
;
4579 frn
->memcg_id
= wb
->memcg_css
->id
;
4584 /* issue foreign writeback flushes for recorded foreign dirtying events */
4585 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4587 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4588 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4589 u64 now
= jiffies_64
;
4592 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4593 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4596 * If the record is older than dirty_expire_interval,
4597 * writeback on it has already started. No need to kick it
4598 * off again. Also, don't start a new one if there's
4599 * already one in flight.
4601 if (time_after64(frn
->at
, now
- intv
) &&
4602 atomic_read(&frn
->done
.cnt
) == 1) {
4604 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4605 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
,
4606 WB_REASON_FOREIGN_FLUSH
,
4612 #else /* CONFIG_CGROUP_WRITEBACK */
4614 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4619 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4623 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4627 #endif /* CONFIG_CGROUP_WRITEBACK */
4630 * DO NOT USE IN NEW FILES.
4632 * "cgroup.event_control" implementation.
4634 * This is way over-engineered. It tries to support fully configurable
4635 * events for each user. Such level of flexibility is completely
4636 * unnecessary especially in the light of the planned unified hierarchy.
4638 * Please deprecate this and replace with something simpler if at all
4643 * Unregister event and free resources.
4645 * Gets called from workqueue.
4647 static void memcg_event_remove(struct work_struct
*work
)
4649 struct mem_cgroup_event
*event
=
4650 container_of(work
, struct mem_cgroup_event
, remove
);
4651 struct mem_cgroup
*memcg
= event
->memcg
;
4653 remove_wait_queue(event
->wqh
, &event
->wait
);
4655 event
->unregister_event(memcg
, event
->eventfd
);
4657 /* Notify userspace the event is going away. */
4658 eventfd_signal(event
->eventfd
, 1);
4660 eventfd_ctx_put(event
->eventfd
);
4662 css_put(&memcg
->css
);
4666 * Gets called on EPOLLHUP on eventfd when user closes it.
4668 * Called with wqh->lock held and interrupts disabled.
4670 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4671 int sync
, void *key
)
4673 struct mem_cgroup_event
*event
=
4674 container_of(wait
, struct mem_cgroup_event
, wait
);
4675 struct mem_cgroup
*memcg
= event
->memcg
;
4676 __poll_t flags
= key_to_poll(key
);
4678 if (flags
& EPOLLHUP
) {
4680 * If the event has been detached at cgroup removal, we
4681 * can simply return knowing the other side will cleanup
4684 * We can't race against event freeing since the other
4685 * side will require wqh->lock via remove_wait_queue(),
4688 spin_lock(&memcg
->event_list_lock
);
4689 if (!list_empty(&event
->list
)) {
4690 list_del_init(&event
->list
);
4692 * We are in atomic context, but cgroup_event_remove()
4693 * may sleep, so we have to call it in workqueue.
4695 schedule_work(&event
->remove
);
4697 spin_unlock(&memcg
->event_list_lock
);
4703 static void memcg_event_ptable_queue_proc(struct file
*file
,
4704 wait_queue_head_t
*wqh
, poll_table
*pt
)
4706 struct mem_cgroup_event
*event
=
4707 container_of(pt
, struct mem_cgroup_event
, pt
);
4710 add_wait_queue(wqh
, &event
->wait
);
4714 * DO NOT USE IN NEW FILES.
4716 * Parse input and register new cgroup event handler.
4718 * Input must be in format '<event_fd> <control_fd> <args>'.
4719 * Interpretation of args is defined by control file implementation.
4721 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4722 char *buf
, size_t nbytes
, loff_t off
)
4724 struct cgroup_subsys_state
*css
= of_css(of
);
4725 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4726 struct mem_cgroup_event
*event
;
4727 struct cgroup_subsys_state
*cfile_css
;
4728 unsigned int efd
, cfd
;
4735 buf
= strstrip(buf
);
4737 efd
= simple_strtoul(buf
, &endp
, 10);
4742 cfd
= simple_strtoul(buf
, &endp
, 10);
4743 if ((*endp
!= ' ') && (*endp
!= '\0'))
4747 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4751 event
->memcg
= memcg
;
4752 INIT_LIST_HEAD(&event
->list
);
4753 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4754 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4755 INIT_WORK(&event
->remove
, memcg_event_remove
);
4763 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4764 if (IS_ERR(event
->eventfd
)) {
4765 ret
= PTR_ERR(event
->eventfd
);
4772 goto out_put_eventfd
;
4775 /* the process need read permission on control file */
4776 /* AV: shouldn't we check that it's been opened for read instead? */
4777 ret
= file_permission(cfile
.file
, MAY_READ
);
4782 * Determine the event callbacks and set them in @event. This used
4783 * to be done via struct cftype but cgroup core no longer knows
4784 * about these events. The following is crude but the whole thing
4785 * is for compatibility anyway.
4787 * DO NOT ADD NEW FILES.
4789 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4791 if (!strcmp(name
, "memory.usage_in_bytes")) {
4792 event
->register_event
= mem_cgroup_usage_register_event
;
4793 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4794 } else if (!strcmp(name
, "memory.oom_control")) {
4795 event
->register_event
= mem_cgroup_oom_register_event
;
4796 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4797 } else if (!strcmp(name
, "memory.pressure_level")) {
4798 event
->register_event
= vmpressure_register_event
;
4799 event
->unregister_event
= vmpressure_unregister_event
;
4800 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4801 event
->register_event
= memsw_cgroup_usage_register_event
;
4802 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4809 * Verify @cfile should belong to @css. Also, remaining events are
4810 * automatically removed on cgroup destruction but the removal is
4811 * asynchronous, so take an extra ref on @css.
4813 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4814 &memory_cgrp_subsys
);
4816 if (IS_ERR(cfile_css
))
4818 if (cfile_css
!= css
) {
4823 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4827 vfs_poll(efile
.file
, &event
->pt
);
4829 spin_lock_irq(&memcg
->event_list_lock
);
4830 list_add(&event
->list
, &memcg
->event_list
);
4831 spin_unlock_irq(&memcg
->event_list_lock
);
4843 eventfd_ctx_put(event
->eventfd
);
4852 static struct cftype mem_cgroup_legacy_files
[] = {
4854 .name
= "usage_in_bytes",
4855 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4856 .read_u64
= mem_cgroup_read_u64
,
4859 .name
= "max_usage_in_bytes",
4860 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4861 .write
= mem_cgroup_reset
,
4862 .read_u64
= mem_cgroup_read_u64
,
4865 .name
= "limit_in_bytes",
4866 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4867 .write
= mem_cgroup_write
,
4868 .read_u64
= mem_cgroup_read_u64
,
4871 .name
= "soft_limit_in_bytes",
4872 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4873 .write
= mem_cgroup_write
,
4874 .read_u64
= mem_cgroup_read_u64
,
4878 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4879 .write
= mem_cgroup_reset
,
4880 .read_u64
= mem_cgroup_read_u64
,
4884 .seq_show
= memcg_stat_show
,
4887 .name
= "force_empty",
4888 .write
= mem_cgroup_force_empty_write
,
4891 .name
= "use_hierarchy",
4892 .write_u64
= mem_cgroup_hierarchy_write
,
4893 .read_u64
= mem_cgroup_hierarchy_read
,
4896 .name
= "cgroup.event_control", /* XXX: for compat */
4897 .write
= memcg_write_event_control
,
4898 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4901 .name
= "swappiness",
4902 .read_u64
= mem_cgroup_swappiness_read
,
4903 .write_u64
= mem_cgroup_swappiness_write
,
4906 .name
= "move_charge_at_immigrate",
4907 .read_u64
= mem_cgroup_move_charge_read
,
4908 .write_u64
= mem_cgroup_move_charge_write
,
4911 .name
= "oom_control",
4912 .seq_show
= mem_cgroup_oom_control_read
,
4913 .write_u64
= mem_cgroup_oom_control_write
,
4914 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4917 .name
= "pressure_level",
4921 .name
= "numa_stat",
4922 .seq_show
= memcg_numa_stat_show
,
4926 .name
= "kmem.limit_in_bytes",
4927 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4928 .write
= mem_cgroup_write
,
4929 .read_u64
= mem_cgroup_read_u64
,
4932 .name
= "kmem.usage_in_bytes",
4933 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4934 .read_u64
= mem_cgroup_read_u64
,
4937 .name
= "kmem.failcnt",
4938 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4939 .write
= mem_cgroup_reset
,
4940 .read_u64
= mem_cgroup_read_u64
,
4943 .name
= "kmem.max_usage_in_bytes",
4944 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4945 .write
= mem_cgroup_reset
,
4946 .read_u64
= mem_cgroup_read_u64
,
4948 #if defined(CONFIG_MEMCG_KMEM) && \
4949 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4951 .name
= "kmem.slabinfo",
4952 .seq_show
= memcg_slab_show
,
4956 .name
= "kmem.tcp.limit_in_bytes",
4957 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4958 .write
= mem_cgroup_write
,
4959 .read_u64
= mem_cgroup_read_u64
,
4962 .name
= "kmem.tcp.usage_in_bytes",
4963 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4964 .read_u64
= mem_cgroup_read_u64
,
4967 .name
= "kmem.tcp.failcnt",
4968 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4969 .write
= mem_cgroup_reset
,
4970 .read_u64
= mem_cgroup_read_u64
,
4973 .name
= "kmem.tcp.max_usage_in_bytes",
4974 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4975 .write
= mem_cgroup_reset
,
4976 .read_u64
= mem_cgroup_read_u64
,
4978 { }, /* terminate */
4982 * Private memory cgroup IDR
4984 * Swap-out records and page cache shadow entries need to store memcg
4985 * references in constrained space, so we maintain an ID space that is
4986 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4987 * memory-controlled cgroups to 64k.
4989 * However, there usually are many references to the offline CSS after
4990 * the cgroup has been destroyed, such as page cache or reclaimable
4991 * slab objects, that don't need to hang on to the ID. We want to keep
4992 * those dead CSS from occupying IDs, or we might quickly exhaust the
4993 * relatively small ID space and prevent the creation of new cgroups
4994 * even when there are much fewer than 64k cgroups - possibly none.
4996 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4997 * be freed and recycled when it's no longer needed, which is usually
4998 * when the CSS is offlined.
5000 * The only exception to that are records of swapped out tmpfs/shmem
5001 * pages that need to be attributed to live ancestors on swapin. But
5002 * those references are manageable from userspace.
5005 static DEFINE_IDR(mem_cgroup_idr
);
5007 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5009 if (memcg
->id
.id
> 0) {
5010 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5015 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5018 refcount_add(n
, &memcg
->id
.ref
);
5021 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5023 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5024 mem_cgroup_id_remove(memcg
);
5026 /* Memcg ID pins CSS */
5027 css_put(&memcg
->css
);
5031 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5033 mem_cgroup_id_put_many(memcg
, 1);
5037 * mem_cgroup_from_id - look up a memcg from a memcg id
5038 * @id: the memcg id to look up
5040 * Caller must hold rcu_read_lock().
5042 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5044 WARN_ON_ONCE(!rcu_read_lock_held());
5045 return idr_find(&mem_cgroup_idr
, id
);
5048 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5050 struct mem_cgroup_per_node
*pn
;
5053 * This routine is called against possible nodes.
5054 * But it's BUG to call kmalloc() against offline node.
5056 * TODO: this routine can waste much memory for nodes which will
5057 * never be onlined. It's better to use memory hotplug callback
5060 if (!node_state(node
, N_NORMAL_MEMORY
))
5062 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5066 pn
->lruvec_stats_percpu
= alloc_percpu_gfp(struct lruvec_stats_percpu
,
5067 GFP_KERNEL_ACCOUNT
);
5068 if (!pn
->lruvec_stats_percpu
) {
5073 lruvec_init(&pn
->lruvec
);
5074 pn
->usage_in_excess
= 0;
5075 pn
->on_tree
= false;
5078 memcg
->nodeinfo
[node
] = pn
;
5082 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5084 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5089 free_percpu(pn
->lruvec_stats_percpu
);
5093 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5098 free_mem_cgroup_per_node_info(memcg
, node
);
5099 free_percpu(memcg
->vmstats_percpu
);
5103 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5105 memcg_wb_domain_exit(memcg
);
5106 __mem_cgroup_free(memcg
);
5109 static struct mem_cgroup
*mem_cgroup_alloc(void)
5111 struct mem_cgroup
*memcg
;
5114 int __maybe_unused i
;
5115 long error
= -ENOMEM
;
5117 size
= sizeof(struct mem_cgroup
);
5118 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5120 memcg
= kzalloc(size
, GFP_KERNEL
);
5122 return ERR_PTR(error
);
5124 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5125 1, MEM_CGROUP_ID_MAX
,
5127 if (memcg
->id
.id
< 0) {
5128 error
= memcg
->id
.id
;
5132 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5133 GFP_KERNEL_ACCOUNT
);
5134 if (!memcg
->vmstats_percpu
)
5138 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5141 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5144 INIT_WORK(&memcg
->high_work
, high_work_func
);
5145 INIT_LIST_HEAD(&memcg
->oom_notify
);
5146 mutex_init(&memcg
->thresholds_lock
);
5147 spin_lock_init(&memcg
->move_lock
);
5148 vmpressure_init(&memcg
->vmpressure
);
5149 INIT_LIST_HEAD(&memcg
->event_list
);
5150 spin_lock_init(&memcg
->event_list_lock
);
5151 memcg
->socket_pressure
= jiffies
;
5152 #ifdef CONFIG_MEMCG_KMEM
5153 memcg
->kmemcg_id
= -1;
5154 INIT_LIST_HEAD(&memcg
->objcg_list
);
5156 #ifdef CONFIG_CGROUP_WRITEBACK
5157 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5158 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5159 memcg
->cgwb_frn
[i
].done
=
5160 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5162 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5163 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5164 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5165 memcg
->deferred_split_queue
.split_queue_len
= 0;
5167 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5170 mem_cgroup_id_remove(memcg
);
5171 __mem_cgroup_free(memcg
);
5172 return ERR_PTR(error
);
5175 static struct cgroup_subsys_state
* __ref
5176 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5178 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5179 struct mem_cgroup
*memcg
, *old_memcg
;
5180 long error
= -ENOMEM
;
5182 old_memcg
= set_active_memcg(parent
);
5183 memcg
= mem_cgroup_alloc();
5184 set_active_memcg(old_memcg
);
5186 return ERR_CAST(memcg
);
5188 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5189 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5190 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5192 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5193 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5195 page_counter_init(&memcg
->memory
, &parent
->memory
);
5196 page_counter_init(&memcg
->swap
, &parent
->swap
);
5197 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5198 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5200 page_counter_init(&memcg
->memory
, NULL
);
5201 page_counter_init(&memcg
->swap
, NULL
);
5202 page_counter_init(&memcg
->kmem
, NULL
);
5203 page_counter_init(&memcg
->tcpmem
, NULL
);
5205 root_mem_cgroup
= memcg
;
5209 /* The following stuff does not apply to the root */
5210 error
= memcg_online_kmem(memcg
);
5214 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5215 static_branch_inc(&memcg_sockets_enabled_key
);
5219 mem_cgroup_id_remove(memcg
);
5220 mem_cgroup_free(memcg
);
5221 return ERR_PTR(error
);
5224 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5226 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5229 * A memcg must be visible for expand_shrinker_info()
5230 * by the time the maps are allocated. So, we allocate maps
5231 * here, when for_each_mem_cgroup() can't skip it.
5233 if (alloc_shrinker_info(memcg
)) {
5234 mem_cgroup_id_remove(memcg
);
5238 /* Online state pins memcg ID, memcg ID pins CSS */
5239 refcount_set(&memcg
->id
.ref
, 1);
5242 if (unlikely(mem_cgroup_is_root(memcg
)))
5243 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
,
5248 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5250 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5251 struct mem_cgroup_event
*event
, *tmp
;
5254 * Unregister events and notify userspace.
5255 * Notify userspace about cgroup removing only after rmdir of cgroup
5256 * directory to avoid race between userspace and kernelspace.
5258 spin_lock_irq(&memcg
->event_list_lock
);
5259 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5260 list_del_init(&event
->list
);
5261 schedule_work(&event
->remove
);
5263 spin_unlock_irq(&memcg
->event_list_lock
);
5265 page_counter_set_min(&memcg
->memory
, 0);
5266 page_counter_set_low(&memcg
->memory
, 0);
5268 memcg_offline_kmem(memcg
);
5269 reparent_shrinker_deferred(memcg
);
5270 wb_memcg_offline(memcg
);
5272 drain_all_stock(memcg
);
5274 mem_cgroup_id_put(memcg
);
5277 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5279 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5281 invalidate_reclaim_iterators(memcg
);
5284 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5286 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5287 int __maybe_unused i
;
5289 #ifdef CONFIG_CGROUP_WRITEBACK
5290 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5291 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5293 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5294 static_branch_dec(&memcg_sockets_enabled_key
);
5296 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5297 static_branch_dec(&memcg_sockets_enabled_key
);
5299 vmpressure_cleanup(&memcg
->vmpressure
);
5300 cancel_work_sync(&memcg
->high_work
);
5301 mem_cgroup_remove_from_trees(memcg
);
5302 free_shrinker_info(memcg
);
5303 memcg_free_kmem(memcg
);
5304 mem_cgroup_free(memcg
);
5308 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5309 * @css: the target css
5311 * Reset the states of the mem_cgroup associated with @css. This is
5312 * invoked when the userland requests disabling on the default hierarchy
5313 * but the memcg is pinned through dependency. The memcg should stop
5314 * applying policies and should revert to the vanilla state as it may be
5315 * made visible again.
5317 * The current implementation only resets the essential configurations.
5318 * This needs to be expanded to cover all the visible parts.
5320 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5322 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5324 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5325 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5326 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5327 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5328 page_counter_set_min(&memcg
->memory
, 0);
5329 page_counter_set_low(&memcg
->memory
, 0);
5330 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5331 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5332 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5333 memcg_wb_domain_size_changed(memcg
);
5336 void mem_cgroup_flush_stats(void)
5338 if (!spin_trylock(&stats_flush_lock
))
5341 cgroup_rstat_flush_irqsafe(root_mem_cgroup
->css
.cgroup
);
5342 spin_unlock(&stats_flush_lock
);
5345 static void flush_memcg_stats_dwork(struct work_struct
*w
)
5347 mem_cgroup_flush_stats();
5348 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
, 2UL*HZ
);
5351 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
5353 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5354 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5355 struct memcg_vmstats_percpu
*statc
;
5359 statc
= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
);
5361 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
5363 * Collect the aggregated propagation counts of groups
5364 * below us. We're in a per-cpu loop here and this is
5365 * a global counter, so the first cycle will get them.
5367 delta
= memcg
->vmstats
.state_pending
[i
];
5369 memcg
->vmstats
.state_pending
[i
] = 0;
5371 /* Add CPU changes on this level since the last flush */
5372 v
= READ_ONCE(statc
->state
[i
]);
5373 if (v
!= statc
->state_prev
[i
]) {
5374 delta
+= v
- statc
->state_prev
[i
];
5375 statc
->state_prev
[i
] = v
;
5381 /* Aggregate counts on this level and propagate upwards */
5382 memcg
->vmstats
.state
[i
] += delta
;
5384 parent
->vmstats
.state_pending
[i
] += delta
;
5387 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
5388 delta
= memcg
->vmstats
.events_pending
[i
];
5390 memcg
->vmstats
.events_pending
[i
] = 0;
5392 v
= READ_ONCE(statc
->events
[i
]);
5393 if (v
!= statc
->events_prev
[i
]) {
5394 delta
+= v
- statc
->events_prev
[i
];
5395 statc
->events_prev
[i
] = v
;
5401 memcg
->vmstats
.events
[i
] += delta
;
5403 parent
->vmstats
.events_pending
[i
] += delta
;
5406 for_each_node_state(nid
, N_MEMORY
) {
5407 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[nid
];
5408 struct mem_cgroup_per_node
*ppn
= NULL
;
5409 struct lruvec_stats_percpu
*lstatc
;
5412 ppn
= parent
->nodeinfo
[nid
];
5414 lstatc
= per_cpu_ptr(pn
->lruvec_stats_percpu
, cpu
);
5416 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
5417 delta
= pn
->lruvec_stats
.state_pending
[i
];
5419 pn
->lruvec_stats
.state_pending
[i
] = 0;
5421 v
= READ_ONCE(lstatc
->state
[i
]);
5422 if (v
!= lstatc
->state_prev
[i
]) {
5423 delta
+= v
- lstatc
->state_prev
[i
];
5424 lstatc
->state_prev
[i
] = v
;
5430 pn
->lruvec_stats
.state
[i
] += delta
;
5432 ppn
->lruvec_stats
.state_pending
[i
] += delta
;
5438 /* Handlers for move charge at task migration. */
5439 static int mem_cgroup_do_precharge(unsigned long count
)
5443 /* Try a single bulk charge without reclaim first, kswapd may wake */
5444 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5446 mc
.precharge
+= count
;
5450 /* Try charges one by one with reclaim, but do not retry */
5452 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5466 enum mc_target_type
{
5473 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5474 unsigned long addr
, pte_t ptent
)
5476 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5478 if (!page
|| !page_mapped(page
))
5480 if (PageAnon(page
)) {
5481 if (!(mc
.flags
& MOVE_ANON
))
5484 if (!(mc
.flags
& MOVE_FILE
))
5487 if (!get_page_unless_zero(page
))
5493 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5494 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5495 pte_t ptent
, swp_entry_t
*entry
)
5497 struct page
*page
= NULL
;
5498 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5500 if (!(mc
.flags
& MOVE_ANON
))
5504 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5505 * a device and because they are not accessible by CPU they are store
5506 * as special swap entry in the CPU page table.
5508 if (is_device_private_entry(ent
)) {
5509 page
= pfn_swap_entry_to_page(ent
);
5511 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5512 * a refcount of 1 when free (unlike normal page)
5514 if (!page_ref_add_unless(page
, 1, 1))
5519 if (non_swap_entry(ent
))
5523 * Because lookup_swap_cache() updates some statistics counter,
5524 * we call find_get_page() with swapper_space directly.
5526 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5527 entry
->val
= ent
.val
;
5532 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5533 pte_t ptent
, swp_entry_t
*entry
)
5539 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5540 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5542 if (!vma
->vm_file
) /* anonymous vma */
5544 if (!(mc
.flags
& MOVE_FILE
))
5547 /* page is moved even if it's not RSS of this task(page-faulted). */
5548 /* shmem/tmpfs may report page out on swap: account for that too. */
5549 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5550 linear_page_index(vma
, addr
));
5554 * mem_cgroup_move_account - move account of the page
5556 * @compound: charge the page as compound or small page
5557 * @from: mem_cgroup which the page is moved from.
5558 * @to: mem_cgroup which the page is moved to. @from != @to.
5560 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5562 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5565 static int mem_cgroup_move_account(struct page
*page
,
5567 struct mem_cgroup
*from
,
5568 struct mem_cgroup
*to
)
5570 struct lruvec
*from_vec
, *to_vec
;
5571 struct pglist_data
*pgdat
;
5572 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5575 VM_BUG_ON(from
== to
);
5576 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5577 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5580 * Prevent mem_cgroup_migrate() from looking at
5581 * page's memory cgroup of its source page while we change it.
5584 if (!trylock_page(page
))
5588 if (page_memcg(page
) != from
)
5591 pgdat
= page_pgdat(page
);
5592 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5593 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5595 lock_page_memcg(page
);
5597 if (PageAnon(page
)) {
5598 if (page_mapped(page
)) {
5599 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5600 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5601 if (PageTransHuge(page
)) {
5602 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5604 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5609 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5610 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5612 if (PageSwapBacked(page
)) {
5613 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5614 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5617 if (page_mapped(page
)) {
5618 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5619 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5622 if (PageDirty(page
)) {
5623 struct address_space
*mapping
= page_mapping(page
);
5625 if (mapping_can_writeback(mapping
)) {
5626 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5628 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5634 if (PageWriteback(page
)) {
5635 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5636 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5640 * All state has been migrated, let's switch to the new memcg.
5642 * It is safe to change page's memcg here because the page
5643 * is referenced, charged, isolated, and locked: we can't race
5644 * with (un)charging, migration, LRU putback, or anything else
5645 * that would rely on a stable page's memory cgroup.
5647 * Note that lock_page_memcg is a memcg lock, not a page lock,
5648 * to save space. As soon as we switch page's memory cgroup to a
5649 * new memcg that isn't locked, the above state can change
5650 * concurrently again. Make sure we're truly done with it.
5655 css_put(&from
->css
);
5657 page
->memcg_data
= (unsigned long)to
;
5659 __folio_memcg_unlock(from
);
5662 nid
= page_to_nid(page
);
5664 local_irq_disable();
5665 mem_cgroup_charge_statistics(to
, nr_pages
);
5666 memcg_check_events(to
, nid
);
5667 mem_cgroup_charge_statistics(from
, -nr_pages
);
5668 memcg_check_events(from
, nid
);
5677 * get_mctgt_type - get target type of moving charge
5678 * @vma: the vma the pte to be checked belongs
5679 * @addr: the address corresponding to the pte to be checked
5680 * @ptent: the pte to be checked
5681 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5684 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5685 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5686 * move charge. if @target is not NULL, the page is stored in target->page
5687 * with extra refcnt got(Callers should handle it).
5688 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5689 * target for charge migration. if @target is not NULL, the entry is stored
5691 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5692 * (so ZONE_DEVICE page and thus not on the lru).
5693 * For now we such page is charge like a regular page would be as for all
5694 * intent and purposes it is just special memory taking the place of a
5697 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5699 * Called with pte lock held.
5702 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5703 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5705 struct page
*page
= NULL
;
5706 enum mc_target_type ret
= MC_TARGET_NONE
;
5707 swp_entry_t ent
= { .val
= 0 };
5709 if (pte_present(ptent
))
5710 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5711 else if (is_swap_pte(ptent
))
5712 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5713 else if (pte_none(ptent
))
5714 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5716 if (!page
&& !ent
.val
)
5720 * Do only loose check w/o serialization.
5721 * mem_cgroup_move_account() checks the page is valid or
5722 * not under LRU exclusion.
5724 if (page_memcg(page
) == mc
.from
) {
5725 ret
= MC_TARGET_PAGE
;
5726 if (is_device_private_page(page
))
5727 ret
= MC_TARGET_DEVICE
;
5729 target
->page
= page
;
5731 if (!ret
|| !target
)
5735 * There is a swap entry and a page doesn't exist or isn't charged.
5736 * But we cannot move a tail-page in a THP.
5738 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5739 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5740 ret
= MC_TARGET_SWAP
;
5747 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5749 * We don't consider PMD mapped swapping or file mapped pages because THP does
5750 * not support them for now.
5751 * Caller should make sure that pmd_trans_huge(pmd) is true.
5753 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5754 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5756 struct page
*page
= NULL
;
5757 enum mc_target_type ret
= MC_TARGET_NONE
;
5759 if (unlikely(is_swap_pmd(pmd
))) {
5760 VM_BUG_ON(thp_migration_supported() &&
5761 !is_pmd_migration_entry(pmd
));
5764 page
= pmd_page(pmd
);
5765 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5766 if (!(mc
.flags
& MOVE_ANON
))
5768 if (page_memcg(page
) == mc
.from
) {
5769 ret
= MC_TARGET_PAGE
;
5772 target
->page
= page
;
5778 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5779 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5781 return MC_TARGET_NONE
;
5785 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5786 unsigned long addr
, unsigned long end
,
5787 struct mm_walk
*walk
)
5789 struct vm_area_struct
*vma
= walk
->vma
;
5793 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5796 * Note their can not be MC_TARGET_DEVICE for now as we do not
5797 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5798 * this might change.
5800 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5801 mc
.precharge
+= HPAGE_PMD_NR
;
5806 if (pmd_trans_unstable(pmd
))
5808 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5809 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5810 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5811 mc
.precharge
++; /* increment precharge temporarily */
5812 pte_unmap_unlock(pte
- 1, ptl
);
5818 static const struct mm_walk_ops precharge_walk_ops
= {
5819 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5822 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5824 unsigned long precharge
;
5827 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5828 mmap_read_unlock(mm
);
5830 precharge
= mc
.precharge
;
5836 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5838 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5840 VM_BUG_ON(mc
.moving_task
);
5841 mc
.moving_task
= current
;
5842 return mem_cgroup_do_precharge(precharge
);
5845 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5846 static void __mem_cgroup_clear_mc(void)
5848 struct mem_cgroup
*from
= mc
.from
;
5849 struct mem_cgroup
*to
= mc
.to
;
5851 /* we must uncharge all the leftover precharges from mc.to */
5853 cancel_charge(mc
.to
, mc
.precharge
);
5857 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5858 * we must uncharge here.
5860 if (mc
.moved_charge
) {
5861 cancel_charge(mc
.from
, mc
.moved_charge
);
5862 mc
.moved_charge
= 0;
5864 /* we must fixup refcnts and charges */
5865 if (mc
.moved_swap
) {
5866 /* uncharge swap account from the old cgroup */
5867 if (!mem_cgroup_is_root(mc
.from
))
5868 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5870 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5873 * we charged both to->memory and to->memsw, so we
5874 * should uncharge to->memory.
5876 if (!mem_cgroup_is_root(mc
.to
))
5877 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5881 memcg_oom_recover(from
);
5882 memcg_oom_recover(to
);
5883 wake_up_all(&mc
.waitq
);
5886 static void mem_cgroup_clear_mc(void)
5888 struct mm_struct
*mm
= mc
.mm
;
5891 * we must clear moving_task before waking up waiters at the end of
5894 mc
.moving_task
= NULL
;
5895 __mem_cgroup_clear_mc();
5896 spin_lock(&mc
.lock
);
5900 spin_unlock(&mc
.lock
);
5905 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5907 struct cgroup_subsys_state
*css
;
5908 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5909 struct mem_cgroup
*from
;
5910 struct task_struct
*leader
, *p
;
5911 struct mm_struct
*mm
;
5912 unsigned long move_flags
;
5915 /* charge immigration isn't supported on the default hierarchy */
5916 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5920 * Multi-process migrations only happen on the default hierarchy
5921 * where charge immigration is not used. Perform charge
5922 * immigration if @tset contains a leader and whine if there are
5926 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5929 memcg
= mem_cgroup_from_css(css
);
5935 * We are now committed to this value whatever it is. Changes in this
5936 * tunable will only affect upcoming migrations, not the current one.
5937 * So we need to save it, and keep it going.
5939 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5943 from
= mem_cgroup_from_task(p
);
5945 VM_BUG_ON(from
== memcg
);
5947 mm
= get_task_mm(p
);
5950 /* We move charges only when we move a owner of the mm */
5951 if (mm
->owner
== p
) {
5954 VM_BUG_ON(mc
.precharge
);
5955 VM_BUG_ON(mc
.moved_charge
);
5956 VM_BUG_ON(mc
.moved_swap
);
5958 spin_lock(&mc
.lock
);
5962 mc
.flags
= move_flags
;
5963 spin_unlock(&mc
.lock
);
5964 /* We set mc.moving_task later */
5966 ret
= mem_cgroup_precharge_mc(mm
);
5968 mem_cgroup_clear_mc();
5975 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
5978 mem_cgroup_clear_mc();
5981 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5982 unsigned long addr
, unsigned long end
,
5983 struct mm_walk
*walk
)
5986 struct vm_area_struct
*vma
= walk
->vma
;
5989 enum mc_target_type target_type
;
5990 union mc_target target
;
5993 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5995 if (mc
.precharge
< HPAGE_PMD_NR
) {
5999 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6000 if (target_type
== MC_TARGET_PAGE
) {
6002 if (!isolate_lru_page(page
)) {
6003 if (!mem_cgroup_move_account(page
, true,
6005 mc
.precharge
-= HPAGE_PMD_NR
;
6006 mc
.moved_charge
+= HPAGE_PMD_NR
;
6008 putback_lru_page(page
);
6011 } else if (target_type
== MC_TARGET_DEVICE
) {
6013 if (!mem_cgroup_move_account(page
, true,
6015 mc
.precharge
-= HPAGE_PMD_NR
;
6016 mc
.moved_charge
+= HPAGE_PMD_NR
;
6024 if (pmd_trans_unstable(pmd
))
6027 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6028 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6029 pte_t ptent
= *(pte
++);
6030 bool device
= false;
6036 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6037 case MC_TARGET_DEVICE
:
6040 case MC_TARGET_PAGE
:
6043 * We can have a part of the split pmd here. Moving it
6044 * can be done but it would be too convoluted so simply
6045 * ignore such a partial THP and keep it in original
6046 * memcg. There should be somebody mapping the head.
6048 if (PageTransCompound(page
))
6050 if (!device
&& isolate_lru_page(page
))
6052 if (!mem_cgroup_move_account(page
, false,
6055 /* we uncharge from mc.from later. */
6059 putback_lru_page(page
);
6060 put
: /* get_mctgt_type() gets the page */
6063 case MC_TARGET_SWAP
:
6065 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6067 mem_cgroup_id_get_many(mc
.to
, 1);
6068 /* we fixup other refcnts and charges later. */
6076 pte_unmap_unlock(pte
- 1, ptl
);
6081 * We have consumed all precharges we got in can_attach().
6082 * We try charge one by one, but don't do any additional
6083 * charges to mc.to if we have failed in charge once in attach()
6086 ret
= mem_cgroup_do_precharge(1);
6094 static const struct mm_walk_ops charge_walk_ops
= {
6095 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6098 static void mem_cgroup_move_charge(void)
6100 lru_add_drain_all();
6102 * Signal lock_page_memcg() to take the memcg's move_lock
6103 * while we're moving its pages to another memcg. Then wait
6104 * for already started RCU-only updates to finish.
6106 atomic_inc(&mc
.from
->moving_account
);
6109 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6111 * Someone who are holding the mmap_lock might be waiting in
6112 * waitq. So we cancel all extra charges, wake up all waiters,
6113 * and retry. Because we cancel precharges, we might not be able
6114 * to move enough charges, but moving charge is a best-effort
6115 * feature anyway, so it wouldn't be a big problem.
6117 __mem_cgroup_clear_mc();
6122 * When we have consumed all precharges and failed in doing
6123 * additional charge, the page walk just aborts.
6125 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6128 mmap_read_unlock(mc
.mm
);
6129 atomic_dec(&mc
.from
->moving_account
);
6132 static void mem_cgroup_move_task(void)
6135 mem_cgroup_move_charge();
6136 mem_cgroup_clear_mc();
6139 #else /* !CONFIG_MMU */
6140 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6144 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6147 static void mem_cgroup_move_task(void)
6152 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6154 if (value
== PAGE_COUNTER_MAX
)
6155 seq_puts(m
, "max\n");
6157 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6162 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6165 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6167 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6170 static int memory_min_show(struct seq_file
*m
, void *v
)
6172 return seq_puts_memcg_tunable(m
,
6173 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6176 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6177 char *buf
, size_t nbytes
, loff_t off
)
6179 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6183 buf
= strstrip(buf
);
6184 err
= page_counter_memparse(buf
, "max", &min
);
6188 page_counter_set_min(&memcg
->memory
, min
);
6193 static int memory_low_show(struct seq_file
*m
, void *v
)
6195 return seq_puts_memcg_tunable(m
,
6196 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6199 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6200 char *buf
, size_t nbytes
, loff_t off
)
6202 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6206 buf
= strstrip(buf
);
6207 err
= page_counter_memparse(buf
, "max", &low
);
6211 page_counter_set_low(&memcg
->memory
, low
);
6216 static int memory_high_show(struct seq_file
*m
, void *v
)
6218 return seq_puts_memcg_tunable(m
,
6219 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6222 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6223 char *buf
, size_t nbytes
, loff_t off
)
6225 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6226 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6227 bool drained
= false;
6231 buf
= strstrip(buf
);
6232 err
= page_counter_memparse(buf
, "max", &high
);
6236 page_counter_set_high(&memcg
->memory
, high
);
6239 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6240 unsigned long reclaimed
;
6242 if (nr_pages
<= high
)
6245 if (signal_pending(current
))
6249 drain_all_stock(memcg
);
6254 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6257 if (!reclaimed
&& !nr_retries
--)
6261 memcg_wb_domain_size_changed(memcg
);
6265 static int memory_max_show(struct seq_file
*m
, void *v
)
6267 return seq_puts_memcg_tunable(m
,
6268 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6271 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6272 char *buf
, size_t nbytes
, loff_t off
)
6274 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6275 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6276 bool drained
= false;
6280 buf
= strstrip(buf
);
6281 err
= page_counter_memparse(buf
, "max", &max
);
6285 xchg(&memcg
->memory
.max
, max
);
6288 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6290 if (nr_pages
<= max
)
6293 if (signal_pending(current
))
6297 drain_all_stock(memcg
);
6303 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6309 memcg_memory_event(memcg
, MEMCG_OOM
);
6310 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6314 memcg_wb_domain_size_changed(memcg
);
6318 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6320 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6321 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6322 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6323 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6324 seq_printf(m
, "oom_kill %lu\n",
6325 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6328 static int memory_events_show(struct seq_file
*m
, void *v
)
6330 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6332 __memory_events_show(m
, memcg
->memory_events
);
6336 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6338 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6340 __memory_events_show(m
, memcg
->memory_events_local
);
6344 static int memory_stat_show(struct seq_file
*m
, void *v
)
6346 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6349 buf
= memory_stat_format(memcg
);
6358 static inline unsigned long lruvec_page_state_output(struct lruvec
*lruvec
,
6361 return lruvec_page_state(lruvec
, item
) * memcg_page_state_unit(item
);
6364 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6367 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6369 cgroup_rstat_flush(memcg
->css
.cgroup
);
6371 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6374 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6377 seq_printf(m
, "%s", memory_stats
[i
].name
);
6378 for_each_node_state(nid
, N_MEMORY
) {
6380 struct lruvec
*lruvec
;
6382 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6383 size
= lruvec_page_state_output(lruvec
,
6384 memory_stats
[i
].idx
);
6385 seq_printf(m
, " N%d=%llu", nid
, size
);
6394 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6396 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6398 seq_printf(m
, "%d\n", memcg
->oom_group
);
6403 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6404 char *buf
, size_t nbytes
, loff_t off
)
6406 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6409 buf
= strstrip(buf
);
6413 ret
= kstrtoint(buf
, 0, &oom_group
);
6417 if (oom_group
!= 0 && oom_group
!= 1)
6420 memcg
->oom_group
= oom_group
;
6425 static struct cftype memory_files
[] = {
6428 .flags
= CFTYPE_NOT_ON_ROOT
,
6429 .read_u64
= memory_current_read
,
6433 .flags
= CFTYPE_NOT_ON_ROOT
,
6434 .seq_show
= memory_min_show
,
6435 .write
= memory_min_write
,
6439 .flags
= CFTYPE_NOT_ON_ROOT
,
6440 .seq_show
= memory_low_show
,
6441 .write
= memory_low_write
,
6445 .flags
= CFTYPE_NOT_ON_ROOT
,
6446 .seq_show
= memory_high_show
,
6447 .write
= memory_high_write
,
6451 .flags
= CFTYPE_NOT_ON_ROOT
,
6452 .seq_show
= memory_max_show
,
6453 .write
= memory_max_write
,
6457 .flags
= CFTYPE_NOT_ON_ROOT
,
6458 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6459 .seq_show
= memory_events_show
,
6462 .name
= "events.local",
6463 .flags
= CFTYPE_NOT_ON_ROOT
,
6464 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6465 .seq_show
= memory_events_local_show
,
6469 .seq_show
= memory_stat_show
,
6473 .name
= "numa_stat",
6474 .seq_show
= memory_numa_stat_show
,
6478 .name
= "oom.group",
6479 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6480 .seq_show
= memory_oom_group_show
,
6481 .write
= memory_oom_group_write
,
6486 struct cgroup_subsys memory_cgrp_subsys
= {
6487 .css_alloc
= mem_cgroup_css_alloc
,
6488 .css_online
= mem_cgroup_css_online
,
6489 .css_offline
= mem_cgroup_css_offline
,
6490 .css_released
= mem_cgroup_css_released
,
6491 .css_free
= mem_cgroup_css_free
,
6492 .css_reset
= mem_cgroup_css_reset
,
6493 .css_rstat_flush
= mem_cgroup_css_rstat_flush
,
6494 .can_attach
= mem_cgroup_can_attach
,
6495 .cancel_attach
= mem_cgroup_cancel_attach
,
6496 .post_attach
= mem_cgroup_move_task
,
6497 .dfl_cftypes
= memory_files
,
6498 .legacy_cftypes
= mem_cgroup_legacy_files
,
6503 * This function calculates an individual cgroup's effective
6504 * protection which is derived from its own memory.min/low, its
6505 * parent's and siblings' settings, as well as the actual memory
6506 * distribution in the tree.
6508 * The following rules apply to the effective protection values:
6510 * 1. At the first level of reclaim, effective protection is equal to
6511 * the declared protection in memory.min and memory.low.
6513 * 2. To enable safe delegation of the protection configuration, at
6514 * subsequent levels the effective protection is capped to the
6515 * parent's effective protection.
6517 * 3. To make complex and dynamic subtrees easier to configure, the
6518 * user is allowed to overcommit the declared protection at a given
6519 * level. If that is the case, the parent's effective protection is
6520 * distributed to the children in proportion to how much protection
6521 * they have declared and how much of it they are utilizing.
6523 * This makes distribution proportional, but also work-conserving:
6524 * if one cgroup claims much more protection than it uses memory,
6525 * the unused remainder is available to its siblings.
6527 * 4. Conversely, when the declared protection is undercommitted at a
6528 * given level, the distribution of the larger parental protection
6529 * budget is NOT proportional. A cgroup's protection from a sibling
6530 * is capped to its own memory.min/low setting.
6532 * 5. However, to allow protecting recursive subtrees from each other
6533 * without having to declare each individual cgroup's fixed share
6534 * of the ancestor's claim to protection, any unutilized -
6535 * "floating" - protection from up the tree is distributed in
6536 * proportion to each cgroup's *usage*. This makes the protection
6537 * neutral wrt sibling cgroups and lets them compete freely over
6538 * the shared parental protection budget, but it protects the
6539 * subtree as a whole from neighboring subtrees.
6541 * Note that 4. and 5. are not in conflict: 4. is about protecting
6542 * against immediate siblings whereas 5. is about protecting against
6543 * neighboring subtrees.
6545 static unsigned long effective_protection(unsigned long usage
,
6546 unsigned long parent_usage
,
6547 unsigned long setting
,
6548 unsigned long parent_effective
,
6549 unsigned long siblings_protected
)
6551 unsigned long protected;
6554 protected = min(usage
, setting
);
6556 * If all cgroups at this level combined claim and use more
6557 * protection then what the parent affords them, distribute
6558 * shares in proportion to utilization.
6560 * We are using actual utilization rather than the statically
6561 * claimed protection in order to be work-conserving: claimed
6562 * but unused protection is available to siblings that would
6563 * otherwise get a smaller chunk than what they claimed.
6565 if (siblings_protected
> parent_effective
)
6566 return protected * parent_effective
/ siblings_protected
;
6569 * Ok, utilized protection of all children is within what the
6570 * parent affords them, so we know whatever this child claims
6571 * and utilizes is effectively protected.
6573 * If there is unprotected usage beyond this value, reclaim
6574 * will apply pressure in proportion to that amount.
6576 * If there is unutilized protection, the cgroup will be fully
6577 * shielded from reclaim, but we do return a smaller value for
6578 * protection than what the group could enjoy in theory. This
6579 * is okay. With the overcommit distribution above, effective
6580 * protection is always dependent on how memory is actually
6581 * consumed among the siblings anyway.
6586 * If the children aren't claiming (all of) the protection
6587 * afforded to them by the parent, distribute the remainder in
6588 * proportion to the (unprotected) memory of each cgroup. That
6589 * way, cgroups that aren't explicitly prioritized wrt each
6590 * other compete freely over the allowance, but they are
6591 * collectively protected from neighboring trees.
6593 * We're using unprotected memory for the weight so that if
6594 * some cgroups DO claim explicit protection, we don't protect
6595 * the same bytes twice.
6597 * Check both usage and parent_usage against the respective
6598 * protected values. One should imply the other, but they
6599 * aren't read atomically - make sure the division is sane.
6601 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6603 if (parent_effective
> siblings_protected
&&
6604 parent_usage
> siblings_protected
&&
6605 usage
> protected) {
6606 unsigned long unclaimed
;
6608 unclaimed
= parent_effective
- siblings_protected
;
6609 unclaimed
*= usage
- protected;
6610 unclaimed
/= parent_usage
- siblings_protected
;
6619 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6620 * @root: the top ancestor of the sub-tree being checked
6621 * @memcg: the memory cgroup to check
6623 * WARNING: This function is not stateless! It can only be used as part
6624 * of a top-down tree iteration, not for isolated queries.
6626 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6627 struct mem_cgroup
*memcg
)
6629 unsigned long usage
, parent_usage
;
6630 struct mem_cgroup
*parent
;
6632 if (mem_cgroup_disabled())
6636 root
= root_mem_cgroup
;
6639 * Effective values of the reclaim targets are ignored so they
6640 * can be stale. Have a look at mem_cgroup_protection for more
6642 * TODO: calculation should be more robust so that we do not need
6643 * that special casing.
6648 usage
= page_counter_read(&memcg
->memory
);
6652 parent
= parent_mem_cgroup(memcg
);
6653 /* No parent means a non-hierarchical mode on v1 memcg */
6657 if (parent
== root
) {
6658 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6659 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6663 parent_usage
= page_counter_read(&parent
->memory
);
6665 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6666 READ_ONCE(memcg
->memory
.min
),
6667 READ_ONCE(parent
->memory
.emin
),
6668 atomic_long_read(&parent
->memory
.children_min_usage
)));
6670 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6671 READ_ONCE(memcg
->memory
.low
),
6672 READ_ONCE(parent
->memory
.elow
),
6673 atomic_long_read(&parent
->memory
.children_low_usage
)));
6676 static int charge_memcg(struct folio
*folio
, struct mem_cgroup
*memcg
,
6679 long nr_pages
= folio_nr_pages(folio
);
6682 ret
= try_charge(memcg
, gfp
, nr_pages
);
6686 css_get(&memcg
->css
);
6687 commit_charge(folio
, memcg
);
6689 local_irq_disable();
6690 mem_cgroup_charge_statistics(memcg
, nr_pages
);
6691 memcg_check_events(memcg
, folio_nid(folio
));
6697 int __mem_cgroup_charge(struct folio
*folio
, struct mm_struct
*mm
, gfp_t gfp
)
6699 struct mem_cgroup
*memcg
;
6702 memcg
= get_mem_cgroup_from_mm(mm
);
6703 ret
= charge_memcg(folio
, memcg
, gfp
);
6704 css_put(&memcg
->css
);
6710 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6711 * @page: page to charge
6712 * @mm: mm context of the victim
6713 * @gfp: reclaim mode
6714 * @entry: swap entry for which the page is allocated
6716 * This function charges a page allocated for swapin. Please call this before
6717 * adding the page to the swapcache.
6719 * Returns 0 on success. Otherwise, an error code is returned.
6721 int mem_cgroup_swapin_charge_page(struct page
*page
, struct mm_struct
*mm
,
6722 gfp_t gfp
, swp_entry_t entry
)
6724 struct folio
*folio
= page_folio(page
);
6725 struct mem_cgroup
*memcg
;
6729 if (mem_cgroup_disabled())
6732 id
= lookup_swap_cgroup_id(entry
);
6734 memcg
= mem_cgroup_from_id(id
);
6735 if (!memcg
|| !css_tryget_online(&memcg
->css
))
6736 memcg
= get_mem_cgroup_from_mm(mm
);
6739 ret
= charge_memcg(folio
, memcg
, gfp
);
6741 css_put(&memcg
->css
);
6746 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6747 * @entry: swap entry for which the page is charged
6749 * Call this function after successfully adding the charged page to swapcache.
6751 * Note: This function assumes the page for which swap slot is being uncharged
6754 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry
)
6757 * Cgroup1's unified memory+swap counter has been charged with the
6758 * new swapcache page, finish the transfer by uncharging the swap
6759 * slot. The swap slot would also get uncharged when it dies, but
6760 * it can stick around indefinitely and we'd count the page twice
6763 * Cgroup2 has separate resource counters for memory and swap,
6764 * so this is a non-issue here. Memory and swap charge lifetimes
6765 * correspond 1:1 to page and swap slot lifetimes: we charge the
6766 * page to memory here, and uncharge swap when the slot is freed.
6768 if (!mem_cgroup_disabled() && do_memsw_account()) {
6770 * The swap entry might not get freed for a long time,
6771 * let's not wait for it. The page already received a
6772 * memory+swap charge, drop the swap entry duplicate.
6774 mem_cgroup_uncharge_swap(entry
, 1);
6778 struct uncharge_gather
{
6779 struct mem_cgroup
*memcg
;
6780 unsigned long nr_memory
;
6781 unsigned long pgpgout
;
6782 unsigned long nr_kmem
;
6786 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6788 memset(ug
, 0, sizeof(*ug
));
6791 static void uncharge_batch(const struct uncharge_gather
*ug
)
6793 unsigned long flags
;
6795 if (ug
->nr_memory
) {
6796 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_memory
);
6797 if (do_memsw_account())
6798 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_memory
);
6799 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6800 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6801 memcg_oom_recover(ug
->memcg
);
6804 local_irq_save(flags
);
6805 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6806 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_memory
);
6807 memcg_check_events(ug
->memcg
, ug
->nid
);
6808 local_irq_restore(flags
);
6810 /* drop reference from uncharge_folio */
6811 css_put(&ug
->memcg
->css
);
6814 static void uncharge_folio(struct folio
*folio
, struct uncharge_gather
*ug
)
6817 struct mem_cgroup
*memcg
;
6818 struct obj_cgroup
*objcg
;
6819 bool use_objcg
= folio_memcg_kmem(folio
);
6821 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
6824 * Nobody should be changing or seriously looking at
6825 * folio memcg or objcg at this point, we have fully
6826 * exclusive access to the folio.
6829 objcg
= __folio_objcg(folio
);
6831 * This get matches the put at the end of the function and
6832 * kmem pages do not hold memcg references anymore.
6834 memcg
= get_mem_cgroup_from_objcg(objcg
);
6836 memcg
= __folio_memcg(folio
);
6842 if (ug
->memcg
!= memcg
) {
6845 uncharge_gather_clear(ug
);
6848 ug
->nid
= folio_nid(folio
);
6850 /* pairs with css_put in uncharge_batch */
6851 css_get(&memcg
->css
);
6854 nr_pages
= folio_nr_pages(folio
);
6857 ug
->nr_memory
+= nr_pages
;
6858 ug
->nr_kmem
+= nr_pages
;
6860 folio
->memcg_data
= 0;
6861 obj_cgroup_put(objcg
);
6863 /* LRU pages aren't accounted at the root level */
6864 if (!mem_cgroup_is_root(memcg
))
6865 ug
->nr_memory
+= nr_pages
;
6868 folio
->memcg_data
= 0;
6871 css_put(&memcg
->css
);
6874 void __mem_cgroup_uncharge(struct folio
*folio
)
6876 struct uncharge_gather ug
;
6878 /* Don't touch folio->lru of any random page, pre-check: */
6879 if (!folio_memcg(folio
))
6882 uncharge_gather_clear(&ug
);
6883 uncharge_folio(folio
, &ug
);
6884 uncharge_batch(&ug
);
6888 * __mem_cgroup_uncharge_list - uncharge a list of page
6889 * @page_list: list of pages to uncharge
6891 * Uncharge a list of pages previously charged with
6892 * __mem_cgroup_charge().
6894 void __mem_cgroup_uncharge_list(struct list_head
*page_list
)
6896 struct uncharge_gather ug
;
6897 struct folio
*folio
;
6899 uncharge_gather_clear(&ug
);
6900 list_for_each_entry(folio
, page_list
, lru
)
6901 uncharge_folio(folio
, &ug
);
6903 uncharge_batch(&ug
);
6907 * mem_cgroup_migrate - Charge a folio's replacement.
6908 * @old: Currently circulating folio.
6909 * @new: Replacement folio.
6911 * Charge @new as a replacement folio for @old. @old will
6912 * be uncharged upon free.
6914 * Both folios must be locked, @new->mapping must be set up.
6916 void mem_cgroup_migrate(struct folio
*old
, struct folio
*new)
6918 struct mem_cgroup
*memcg
;
6919 long nr_pages
= folio_nr_pages(new);
6920 unsigned long flags
;
6922 VM_BUG_ON_FOLIO(!folio_test_locked(old
), old
);
6923 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
6924 VM_BUG_ON_FOLIO(folio_test_anon(old
) != folio_test_anon(new), new);
6925 VM_BUG_ON_FOLIO(folio_nr_pages(old
) != nr_pages
, new);
6927 if (mem_cgroup_disabled())
6930 /* Page cache replacement: new folio already charged? */
6931 if (folio_memcg(new))
6934 memcg
= folio_memcg(old
);
6935 VM_WARN_ON_ONCE_FOLIO(!memcg
, old
);
6939 /* Force-charge the new page. The old one will be freed soon */
6940 if (!mem_cgroup_is_root(memcg
)) {
6941 page_counter_charge(&memcg
->memory
, nr_pages
);
6942 if (do_memsw_account())
6943 page_counter_charge(&memcg
->memsw
, nr_pages
);
6946 css_get(&memcg
->css
);
6947 commit_charge(new, memcg
);
6949 local_irq_save(flags
);
6950 mem_cgroup_charge_statistics(memcg
, nr_pages
);
6951 memcg_check_events(memcg
, folio_nid(new));
6952 local_irq_restore(flags
);
6955 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
6956 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
6958 void mem_cgroup_sk_alloc(struct sock
*sk
)
6960 struct mem_cgroup
*memcg
;
6962 if (!mem_cgroup_sockets_enabled
)
6965 /* Do not associate the sock with unrelated interrupted task's memcg. */
6970 memcg
= mem_cgroup_from_task(current
);
6971 if (memcg
== root_mem_cgroup
)
6973 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
6975 if (css_tryget(&memcg
->css
))
6976 sk
->sk_memcg
= memcg
;
6981 void mem_cgroup_sk_free(struct sock
*sk
)
6984 css_put(&sk
->sk_memcg
->css
);
6988 * mem_cgroup_charge_skmem - charge socket memory
6989 * @memcg: memcg to charge
6990 * @nr_pages: number of pages to charge
6991 * @gfp_mask: reclaim mode
6993 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6994 * @memcg's configured limit, %false if it doesn't.
6996 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
,
6999 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7000 struct page_counter
*fail
;
7002 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7003 memcg
->tcpmem_pressure
= 0;
7006 memcg
->tcpmem_pressure
= 1;
7007 if (gfp_mask
& __GFP_NOFAIL
) {
7008 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7014 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0) {
7015 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7023 * mem_cgroup_uncharge_skmem - uncharge socket memory
7024 * @memcg: memcg to uncharge
7025 * @nr_pages: number of pages to uncharge
7027 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7029 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7030 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7034 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7036 refill_stock(memcg
, nr_pages
);
7039 static int __init
cgroup_memory(char *s
)
7043 while ((token
= strsep(&s
, ",")) != NULL
) {
7046 if (!strcmp(token
, "nosocket"))
7047 cgroup_memory_nosocket
= true;
7048 if (!strcmp(token
, "nokmem"))
7049 cgroup_memory_nokmem
= true;
7053 __setup("cgroup.memory=", cgroup_memory
);
7056 * subsys_initcall() for memory controller.
7058 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7059 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7060 * basically everything that doesn't depend on a specific mem_cgroup structure
7061 * should be initialized from here.
7063 static int __init
mem_cgroup_init(void)
7068 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7069 * used for per-memcg-per-cpu caching of per-node statistics. In order
7070 * to work fine, we should make sure that the overfill threshold can't
7071 * exceed S32_MAX / PAGE_SIZE.
7073 BUILD_BUG_ON(MEMCG_CHARGE_BATCH
> S32_MAX
/ PAGE_SIZE
);
7075 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7076 memcg_hotplug_cpu_dead
);
7078 for_each_possible_cpu(cpu
)
7079 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7082 for_each_node(node
) {
7083 struct mem_cgroup_tree_per_node
*rtpn
;
7085 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7086 node_online(node
) ? node
: NUMA_NO_NODE
);
7088 rtpn
->rb_root
= RB_ROOT
;
7089 rtpn
->rb_rightmost
= NULL
;
7090 spin_lock_init(&rtpn
->lock
);
7091 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7096 subsys_initcall(mem_cgroup_init
);
7098 #ifdef CONFIG_MEMCG_SWAP
7099 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7101 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7103 * The root cgroup cannot be destroyed, so it's refcount must
7106 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7110 memcg
= parent_mem_cgroup(memcg
);
7112 memcg
= root_mem_cgroup
;
7118 * mem_cgroup_swapout - transfer a memsw charge to swap
7119 * @page: page whose memsw charge to transfer
7120 * @entry: swap entry to move the charge to
7122 * Transfer the memsw charge of @page to @entry.
7124 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7126 struct mem_cgroup
*memcg
, *swap_memcg
;
7127 unsigned int nr_entries
;
7128 unsigned short oldid
;
7130 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7131 VM_BUG_ON_PAGE(page_count(page
), page
);
7133 if (mem_cgroup_disabled())
7136 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7139 memcg
= page_memcg(page
);
7141 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7146 * In case the memcg owning these pages has been offlined and doesn't
7147 * have an ID allocated to it anymore, charge the closest online
7148 * ancestor for the swap instead and transfer the memory+swap charge.
7150 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7151 nr_entries
= thp_nr_pages(page
);
7152 /* Get references for the tail pages, too */
7154 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7155 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7157 VM_BUG_ON_PAGE(oldid
, page
);
7158 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7160 page
->memcg_data
= 0;
7162 if (!mem_cgroup_is_root(memcg
))
7163 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7165 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7166 if (!mem_cgroup_is_root(swap_memcg
))
7167 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7168 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7172 * Interrupts should be disabled here because the caller holds the
7173 * i_pages lock which is taken with interrupts-off. It is
7174 * important here to have the interrupts disabled because it is the
7175 * only synchronisation we have for updating the per-CPU variables.
7177 VM_BUG_ON(!irqs_disabled());
7178 mem_cgroup_charge_statistics(memcg
, -nr_entries
);
7179 memcg_check_events(memcg
, page_to_nid(page
));
7181 css_put(&memcg
->css
);
7185 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7186 * @page: page being added to swap
7187 * @entry: swap entry to charge
7189 * Try to charge @page's memcg for the swap space at @entry.
7191 * Returns 0 on success, -ENOMEM on failure.
7193 int __mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7195 unsigned int nr_pages
= thp_nr_pages(page
);
7196 struct page_counter
*counter
;
7197 struct mem_cgroup
*memcg
;
7198 unsigned short oldid
;
7200 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7203 memcg
= page_memcg(page
);
7205 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7210 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7214 memcg
= mem_cgroup_id_get_online(memcg
);
7216 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7217 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7218 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7219 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7220 mem_cgroup_id_put(memcg
);
7224 /* Get references for the tail pages, too */
7226 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7227 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7228 VM_BUG_ON_PAGE(oldid
, page
);
7229 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7235 * __mem_cgroup_uncharge_swap - uncharge swap space
7236 * @entry: swap entry to uncharge
7237 * @nr_pages: the amount of swap space to uncharge
7239 void __mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7241 struct mem_cgroup
*memcg
;
7244 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7246 memcg
= mem_cgroup_from_id(id
);
7248 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7249 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7250 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7252 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7254 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7255 mem_cgroup_id_put_many(memcg
, nr_pages
);
7260 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7262 long nr_swap_pages
= get_nr_swap_pages();
7264 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7265 return nr_swap_pages
;
7266 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7267 nr_swap_pages
= min_t(long, nr_swap_pages
,
7268 READ_ONCE(memcg
->swap
.max
) -
7269 page_counter_read(&memcg
->swap
));
7270 return nr_swap_pages
;
7273 bool mem_cgroup_swap_full(struct page
*page
)
7275 struct mem_cgroup
*memcg
;
7277 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7281 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7284 memcg
= page_memcg(page
);
7288 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7289 unsigned long usage
= page_counter_read(&memcg
->swap
);
7291 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7292 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7299 static int __init
setup_swap_account(char *s
)
7301 if (!strcmp(s
, "1"))
7302 cgroup_memory_noswap
= false;
7303 else if (!strcmp(s
, "0"))
7304 cgroup_memory_noswap
= true;
7307 __setup("swapaccount=", setup_swap_account
);
7309 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7312 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7314 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7317 static int swap_high_show(struct seq_file
*m
, void *v
)
7319 return seq_puts_memcg_tunable(m
,
7320 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7323 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7324 char *buf
, size_t nbytes
, loff_t off
)
7326 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7330 buf
= strstrip(buf
);
7331 err
= page_counter_memparse(buf
, "max", &high
);
7335 page_counter_set_high(&memcg
->swap
, high
);
7340 static int swap_max_show(struct seq_file
*m
, void *v
)
7342 return seq_puts_memcg_tunable(m
,
7343 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7346 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7347 char *buf
, size_t nbytes
, loff_t off
)
7349 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7353 buf
= strstrip(buf
);
7354 err
= page_counter_memparse(buf
, "max", &max
);
7358 xchg(&memcg
->swap
.max
, max
);
7363 static int swap_events_show(struct seq_file
*m
, void *v
)
7365 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7367 seq_printf(m
, "high %lu\n",
7368 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7369 seq_printf(m
, "max %lu\n",
7370 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7371 seq_printf(m
, "fail %lu\n",
7372 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7377 static struct cftype swap_files
[] = {
7379 .name
= "swap.current",
7380 .flags
= CFTYPE_NOT_ON_ROOT
,
7381 .read_u64
= swap_current_read
,
7384 .name
= "swap.high",
7385 .flags
= CFTYPE_NOT_ON_ROOT
,
7386 .seq_show
= swap_high_show
,
7387 .write
= swap_high_write
,
7391 .flags
= CFTYPE_NOT_ON_ROOT
,
7392 .seq_show
= swap_max_show
,
7393 .write
= swap_max_write
,
7396 .name
= "swap.events",
7397 .flags
= CFTYPE_NOT_ON_ROOT
,
7398 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7399 .seq_show
= swap_events_show
,
7404 static struct cftype memsw_files
[] = {
7406 .name
= "memsw.usage_in_bytes",
7407 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7408 .read_u64
= mem_cgroup_read_u64
,
7411 .name
= "memsw.max_usage_in_bytes",
7412 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7413 .write
= mem_cgroup_reset
,
7414 .read_u64
= mem_cgroup_read_u64
,
7417 .name
= "memsw.limit_in_bytes",
7418 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7419 .write
= mem_cgroup_write
,
7420 .read_u64
= mem_cgroup_read_u64
,
7423 .name
= "memsw.failcnt",
7424 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7425 .write
= mem_cgroup_reset
,
7426 .read_u64
= mem_cgroup_read_u64
,
7428 { }, /* terminate */
7432 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7433 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7434 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7435 * boot parameter. This may result in premature OOPS inside
7436 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7438 static int __init
mem_cgroup_swap_init(void)
7440 /* No memory control -> no swap control */
7441 if (mem_cgroup_disabled())
7442 cgroup_memory_noswap
= true;
7444 if (cgroup_memory_noswap
)
7447 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7448 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7452 core_initcall(mem_cgroup_swap_init
);
7454 #endif /* CONFIG_MEMCG_SWAP */