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 #define THRESHOLDS_EVENTS_TARGET 128
107 #define SOFTLIMIT_EVENTS_TARGET 1024
110 * Cgroups above their limits are maintained in a RB-Tree, independent of
111 * their hierarchy representation
114 struct mem_cgroup_tree_per_node
{
115 struct rb_root rb_root
;
116 struct rb_node
*rb_rightmost
;
120 struct mem_cgroup_tree
{
121 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
124 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
127 struct mem_cgroup_eventfd_list
{
128 struct list_head list
;
129 struct eventfd_ctx
*eventfd
;
133 * cgroup_event represents events which userspace want to receive.
135 struct mem_cgroup_event
{
137 * memcg which the event belongs to.
139 struct mem_cgroup
*memcg
;
141 * eventfd to signal userspace about the event.
143 struct eventfd_ctx
*eventfd
;
145 * Each of these stored in a list by the cgroup.
147 struct list_head list
;
149 * register_event() callback will be used to add new userspace
150 * waiter for changes related to this event. Use eventfd_signal()
151 * on eventfd to send notification to userspace.
153 int (*register_event
)(struct mem_cgroup
*memcg
,
154 struct eventfd_ctx
*eventfd
, const char *args
);
156 * unregister_event() callback will be called when userspace closes
157 * the eventfd or on cgroup removing. This callback must be set,
158 * if you want provide notification functionality.
160 void (*unregister_event
)(struct mem_cgroup
*memcg
,
161 struct eventfd_ctx
*eventfd
);
163 * All fields below needed to unregister event when
164 * userspace closes eventfd.
167 wait_queue_head_t
*wqh
;
168 wait_queue_entry_t wait
;
169 struct work_struct remove
;
172 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
173 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
175 /* Stuffs for move charges at task migration. */
177 * Types of charges to be moved.
179 #define MOVE_ANON 0x1U
180 #define MOVE_FILE 0x2U
181 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
183 /* "mc" and its members are protected by cgroup_mutex */
184 static struct move_charge_struct
{
185 spinlock_t lock
; /* for from, to */
186 struct mm_struct
*mm
;
187 struct mem_cgroup
*from
;
188 struct mem_cgroup
*to
;
190 unsigned long precharge
;
191 unsigned long moved_charge
;
192 unsigned long moved_swap
;
193 struct task_struct
*moving_task
; /* a task moving charges */
194 wait_queue_head_t waitq
; /* a waitq for other context */
196 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
197 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
201 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
202 * limit reclaim to prevent infinite loops, if they ever occur.
204 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
205 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
207 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 /* Used for OOM notifier */
220 #define OOM_CONTROL (0)
223 * Iteration constructs for visiting all cgroups (under a tree). If
224 * loops are exited prematurely (break), mem_cgroup_iter_break() must
225 * be used for reference counting.
227 #define for_each_mem_cgroup_tree(iter, root) \
228 for (iter = mem_cgroup_iter(root, NULL, NULL); \
230 iter = mem_cgroup_iter(root, iter, NULL))
232 #define for_each_mem_cgroup(iter) \
233 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
235 iter = mem_cgroup_iter(NULL, iter, NULL))
237 static inline bool task_is_dying(void)
239 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
240 (current
->flags
& PF_EXITING
);
243 /* Some nice accessors for the vmpressure. */
244 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
247 memcg
= root_mem_cgroup
;
248 return &memcg
->vmpressure
;
251 struct mem_cgroup
*vmpressure_to_memcg(struct vmpressure
*vmpr
)
253 return container_of(vmpr
, struct mem_cgroup
, vmpressure
);
256 #ifdef CONFIG_MEMCG_KMEM
257 static DEFINE_SPINLOCK(objcg_lock
);
259 bool mem_cgroup_kmem_disabled(void)
261 return cgroup_memory_nokmem
;
264 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
265 unsigned int nr_pages
);
267 static void obj_cgroup_release(struct percpu_ref
*ref
)
269 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
270 unsigned int nr_bytes
;
271 unsigned int nr_pages
;
275 * At this point all allocated objects are freed, and
276 * objcg->nr_charged_bytes can't have an arbitrary byte value.
277 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
279 * The following sequence can lead to it:
280 * 1) CPU0: objcg == stock->cached_objcg
281 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
282 * PAGE_SIZE bytes are charged
283 * 3) CPU1: a process from another memcg is allocating something,
284 * the stock if flushed,
285 * objcg->nr_charged_bytes = PAGE_SIZE - 92
286 * 5) CPU0: we do release this object,
287 * 92 bytes are added to stock->nr_bytes
288 * 6) CPU0: stock is flushed,
289 * 92 bytes are added to objcg->nr_charged_bytes
291 * In the result, nr_charged_bytes == PAGE_SIZE.
292 * This page will be uncharged in obj_cgroup_release().
294 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
295 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
296 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
299 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
301 spin_lock_irqsave(&objcg_lock
, flags
);
302 list_del(&objcg
->list
);
303 spin_unlock_irqrestore(&objcg_lock
, flags
);
305 percpu_ref_exit(ref
);
306 kfree_rcu(objcg
, rcu
);
309 static struct obj_cgroup
*obj_cgroup_alloc(void)
311 struct obj_cgroup
*objcg
;
314 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
318 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
324 INIT_LIST_HEAD(&objcg
->list
);
328 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
329 struct mem_cgroup
*parent
)
331 struct obj_cgroup
*objcg
, *iter
;
333 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
335 spin_lock_irq(&objcg_lock
);
337 /* 1) Ready to reparent active objcg. */
338 list_add(&objcg
->list
, &memcg
->objcg_list
);
339 /* 2) Reparent active objcg and already reparented objcgs to parent. */
340 list_for_each_entry(iter
, &memcg
->objcg_list
, list
)
341 WRITE_ONCE(iter
->memcg
, parent
);
342 /* 3) Move already reparented objcgs to the parent's list */
343 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
345 spin_unlock_irq(&objcg_lock
);
347 percpu_ref_kill(&objcg
->refcnt
);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida
);
362 int memcg_nr_cache_ids
;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem
);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem
);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem
);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
415 struct mem_cgroup
*memcg
;
417 memcg
= page_memcg(page
);
419 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
420 memcg
= root_mem_cgroup
;
426 * page_cgroup_ino - return inode number of the memcg a page is charged to
429 * Look up the closest online ancestor of the memory cgroup @page is charged to
430 * and return its inode number or 0 if @page is not charged to any cgroup. It
431 * is safe to call this function without holding a reference to @page.
433 * Note, this function is inherently racy, because there is nothing to prevent
434 * the cgroup inode from getting torn down and potentially reallocated a moment
435 * after page_cgroup_ino() returns, so it only should be used by callers that
436 * do not care (such as procfs interfaces).
438 ino_t
page_cgroup_ino(struct page
*page
)
440 struct mem_cgroup
*memcg
;
441 unsigned long ino
= 0;
444 memcg
= page_memcg_check(page
);
446 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
447 memcg
= parent_mem_cgroup(memcg
);
449 ino
= cgroup_ino(memcg
->css
.cgroup
);
454 static struct mem_cgroup_per_node
*
455 mem_cgroup_page_nodeinfo(struct mem_cgroup
*memcg
, struct page
*page
)
457 int nid
= page_to_nid(page
);
459 return memcg
->nodeinfo
[nid
];
462 static struct mem_cgroup_tree_per_node
*
463 soft_limit_tree_node(int nid
)
465 return soft_limit_tree
.rb_tree_per_node
[nid
];
468 static struct mem_cgroup_tree_per_node
*
469 soft_limit_tree_from_page(struct page
*page
)
471 int nid
= page_to_nid(page
);
473 return soft_limit_tree
.rb_tree_per_node
[nid
];
476 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
477 struct mem_cgroup_tree_per_node
*mctz
,
478 unsigned long new_usage_in_excess
)
480 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
481 struct rb_node
*parent
= NULL
;
482 struct mem_cgroup_per_node
*mz_node
;
483 bool rightmost
= true;
488 mz
->usage_in_excess
= new_usage_in_excess
;
489 if (!mz
->usage_in_excess
)
493 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
495 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
504 mctz
->rb_rightmost
= &mz
->tree_node
;
506 rb_link_node(&mz
->tree_node
, parent
, p
);
507 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
511 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
512 struct mem_cgroup_tree_per_node
*mctz
)
517 if (&mz
->tree_node
== mctz
->rb_rightmost
)
518 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
520 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
524 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
525 struct mem_cgroup_tree_per_node
*mctz
)
529 spin_lock_irqsave(&mctz
->lock
, flags
);
530 __mem_cgroup_remove_exceeded(mz
, mctz
);
531 spin_unlock_irqrestore(&mctz
->lock
, flags
);
534 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
536 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
537 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
538 unsigned long excess
= 0;
540 if (nr_pages
> soft_limit
)
541 excess
= nr_pages
- soft_limit
;
546 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
548 unsigned long excess
;
549 struct mem_cgroup_per_node
*mz
;
550 struct mem_cgroup_tree_per_node
*mctz
;
552 mctz
= soft_limit_tree_from_page(page
);
556 * Necessary to update all ancestors when hierarchy is used.
557 * because their event counter is not touched.
559 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
560 mz
= mem_cgroup_page_nodeinfo(memcg
, page
);
561 excess
= soft_limit_excess(memcg
);
563 * We have to update the tree if mz is on RB-tree or
564 * mem is over its softlimit.
566 if (excess
|| mz
->on_tree
) {
569 spin_lock_irqsave(&mctz
->lock
, flags
);
570 /* if on-tree, remove it */
572 __mem_cgroup_remove_exceeded(mz
, mctz
);
574 * Insert again. mz->usage_in_excess will be updated.
575 * If excess is 0, no tree ops.
577 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
578 spin_unlock_irqrestore(&mctz
->lock
, flags
);
583 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
585 struct mem_cgroup_tree_per_node
*mctz
;
586 struct mem_cgroup_per_node
*mz
;
590 mz
= memcg
->nodeinfo
[nid
];
591 mctz
= soft_limit_tree_node(nid
);
593 mem_cgroup_remove_exceeded(mz
, mctz
);
597 static struct mem_cgroup_per_node
*
598 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
600 struct mem_cgroup_per_node
*mz
;
604 if (!mctz
->rb_rightmost
)
605 goto done
; /* Nothing to reclaim from */
607 mz
= rb_entry(mctz
->rb_rightmost
,
608 struct mem_cgroup_per_node
, tree_node
);
610 * Remove the node now but someone else can add it back,
611 * we will to add it back at the end of reclaim to its correct
612 * position in the tree.
614 __mem_cgroup_remove_exceeded(mz
, mctz
);
615 if (!soft_limit_excess(mz
->memcg
) ||
616 !css_tryget(&mz
->memcg
->css
))
622 static struct mem_cgroup_per_node
*
623 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
625 struct mem_cgroup_per_node
*mz
;
627 spin_lock_irq(&mctz
->lock
);
628 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
629 spin_unlock_irq(&mctz
->lock
);
634 * memcg and lruvec stats flushing
636 * Many codepaths leading to stats update or read are performance sensitive and
637 * adding stats flushing in such codepaths is not desirable. So, to optimize the
638 * flushing the kernel does:
640 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
641 * rstat update tree grow unbounded.
643 * 2) Flush the stats synchronously on reader side only when there are more than
644 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
645 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
646 * only for 2 seconds due to (1).
648 static void flush_memcg_stats_dwork(struct work_struct
*w
);
649 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork
, flush_memcg_stats_dwork
);
650 static DEFINE_SPINLOCK(stats_flush_lock
);
651 static DEFINE_PER_CPU(unsigned int, stats_updates
);
652 static atomic_t stats_flush_threshold
= ATOMIC_INIT(0);
654 static inline void memcg_rstat_updated(struct mem_cgroup
*memcg
, int val
)
658 cgroup_rstat_updated(memcg
->css
.cgroup
, smp_processor_id());
660 x
= __this_cpu_add_return(stats_updates
, abs(val
));
661 if (x
> MEMCG_CHARGE_BATCH
) {
662 atomic_add(x
/ MEMCG_CHARGE_BATCH
, &stats_flush_threshold
);
663 __this_cpu_write(stats_updates
, 0);
667 static void __mem_cgroup_flush_stats(void)
671 if (!spin_trylock_irqsave(&stats_flush_lock
, flag
))
674 cgroup_rstat_flush_irqsafe(root_mem_cgroup
->css
.cgroup
);
675 atomic_set(&stats_flush_threshold
, 0);
676 spin_unlock_irqrestore(&stats_flush_lock
, flag
);
679 void mem_cgroup_flush_stats(void)
681 if (atomic_read(&stats_flush_threshold
) > num_online_cpus())
682 __mem_cgroup_flush_stats();
685 static void flush_memcg_stats_dwork(struct work_struct
*w
)
687 __mem_cgroup_flush_stats();
688 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
, 2UL*HZ
);
692 * __mod_memcg_state - update cgroup memory statistics
693 * @memcg: the memory cgroup
694 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
695 * @val: delta to add to the counter, can be negative
697 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
699 if (mem_cgroup_disabled())
702 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
703 memcg_rstat_updated(memcg
, val
);
706 /* idx can be of type enum memcg_stat_item or node_stat_item. */
707 static unsigned long memcg_page_state_local(struct mem_cgroup
*memcg
, int idx
)
712 for_each_possible_cpu(cpu
)
713 x
+= per_cpu(memcg
->vmstats_percpu
->state
[idx
], cpu
);
721 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
724 struct mem_cgroup_per_node
*pn
;
725 struct mem_cgroup
*memcg
;
727 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
731 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
734 __this_cpu_add(pn
->lruvec_stats_percpu
->state
[idx
], val
);
736 memcg_rstat_updated(memcg
, val
);
740 * __mod_lruvec_state - update lruvec memory statistics
741 * @lruvec: the lruvec
742 * @idx: the stat item
743 * @val: delta to add to the counter, can be negative
745 * The lruvec is the intersection of the NUMA node and a cgroup. This
746 * function updates the all three counters that are affected by a
747 * change of state at this level: per-node, per-cgroup, per-lruvec.
749 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
753 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
755 /* Update memcg and lruvec */
756 if (!mem_cgroup_disabled())
757 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
760 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
763 struct page
*head
= compound_head(page
); /* rmap on tail pages */
764 struct mem_cgroup
*memcg
;
765 pg_data_t
*pgdat
= page_pgdat(page
);
766 struct lruvec
*lruvec
;
769 memcg
= page_memcg(head
);
770 /* Untracked pages have no memcg, no lruvec. Update only the node */
773 __mod_node_page_state(pgdat
, idx
, val
);
777 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
778 __mod_lruvec_state(lruvec
, idx
, val
);
781 EXPORT_SYMBOL(__mod_lruvec_page_state
);
783 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
785 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
786 struct mem_cgroup
*memcg
;
787 struct lruvec
*lruvec
;
790 memcg
= mem_cgroup_from_obj(p
);
793 * Untracked pages have no memcg, no lruvec. Update only the
794 * node. If we reparent the slab objects to the root memcg,
795 * when we free the slab object, we need to update the per-memcg
796 * vmstats to keep it correct for the root memcg.
799 __mod_node_page_state(pgdat
, idx
, val
);
801 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
802 __mod_lruvec_state(lruvec
, idx
, val
);
808 * mod_objcg_mlstate() may be called with irq enabled, so
809 * mod_memcg_lruvec_state() should be used.
811 static inline void mod_objcg_mlstate(struct obj_cgroup
*objcg
,
812 struct pglist_data
*pgdat
,
813 enum node_stat_item idx
, int nr
)
815 struct mem_cgroup
*memcg
;
816 struct lruvec
*lruvec
;
819 memcg
= obj_cgroup_memcg(objcg
);
820 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
821 mod_memcg_lruvec_state(lruvec
, idx
, nr
);
826 * __count_memcg_events - account VM events in a cgroup
827 * @memcg: the memory cgroup
828 * @idx: the event item
829 * @count: the number of events that occurred
831 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
834 if (mem_cgroup_disabled())
837 __this_cpu_add(memcg
->vmstats_percpu
->events
[idx
], count
);
838 memcg_rstat_updated(memcg
, count
);
841 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
843 return READ_ONCE(memcg
->vmstats
.events
[event
]);
846 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
851 for_each_possible_cpu(cpu
)
852 x
+= per_cpu(memcg
->vmstats_percpu
->events
[event
], cpu
);
856 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
860 /* pagein of a big page is an event. So, ignore page size */
862 __count_memcg_events(memcg
, PGPGIN
, 1);
864 __count_memcg_events(memcg
, PGPGOUT
, 1);
865 nr_pages
= -nr_pages
; /* for event */
868 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
871 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
872 enum mem_cgroup_events_target target
)
874 unsigned long val
, next
;
876 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
877 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
878 /* from time_after() in jiffies.h */
879 if ((long)(next
- val
) < 0) {
881 case MEM_CGROUP_TARGET_THRESH
:
882 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
884 case MEM_CGROUP_TARGET_SOFTLIMIT
:
885 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
890 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
897 * Check events in order.
900 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
902 /* threshold event is triggered in finer grain than soft limit */
903 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
904 MEM_CGROUP_TARGET_THRESH
))) {
907 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
908 MEM_CGROUP_TARGET_SOFTLIMIT
);
909 mem_cgroup_threshold(memcg
);
910 if (unlikely(do_softlimit
))
911 mem_cgroup_update_tree(memcg
, page
);
915 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
918 * mm_update_next_owner() may clear mm->owner to NULL
919 * if it races with swapoff, page migration, etc.
920 * So this can be called with p == NULL.
925 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
927 EXPORT_SYMBOL(mem_cgroup_from_task
);
929 static __always_inline
struct mem_cgroup
*active_memcg(void)
932 return this_cpu_read(int_active_memcg
);
934 return current
->active_memcg
;
938 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
939 * @mm: mm from which memcg should be extracted. It can be NULL.
941 * Obtain a reference on mm->memcg and returns it if successful. If mm
942 * is NULL, then the memcg is chosen as follows:
943 * 1) The active memcg, if set.
944 * 2) current->mm->memcg, if available
946 * If mem_cgroup is disabled, NULL is returned.
948 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
950 struct mem_cgroup
*memcg
;
952 if (mem_cgroup_disabled())
956 * Page cache insertions can happen without an
957 * actual mm context, e.g. during disk probing
958 * on boot, loopback IO, acct() writes etc.
960 * No need to css_get on root memcg as the reference
961 * counting is disabled on the root level in the
962 * cgroup core. See CSS_NO_REF.
965 memcg
= active_memcg();
966 if (unlikely(memcg
)) {
967 /* remote memcg must hold a ref */
968 css_get(&memcg
->css
);
973 return root_mem_cgroup
;
978 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
979 if (unlikely(!memcg
))
980 memcg
= root_mem_cgroup
;
981 } while (!css_tryget(&memcg
->css
));
985 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
987 static __always_inline
bool memcg_kmem_bypass(void)
989 /* Allow remote memcg charging from any context. */
990 if (unlikely(active_memcg()))
993 /* Memcg to charge can't be determined. */
994 if (!in_task() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
1001 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1002 * @root: hierarchy root
1003 * @prev: previously returned memcg, NULL on first invocation
1004 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1006 * Returns references to children of the hierarchy below @root, or
1007 * @root itself, or %NULL after a full round-trip.
1009 * Caller must pass the return value in @prev on subsequent
1010 * invocations for reference counting, or use mem_cgroup_iter_break()
1011 * to cancel a hierarchy walk before the round-trip is complete.
1013 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1014 * in the hierarchy among all concurrent reclaimers operating on the
1017 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1018 struct mem_cgroup
*prev
,
1019 struct mem_cgroup_reclaim_cookie
*reclaim
)
1021 struct mem_cgroup_reclaim_iter
*iter
;
1022 struct cgroup_subsys_state
*css
= NULL
;
1023 struct mem_cgroup
*memcg
= NULL
;
1024 struct mem_cgroup
*pos
= NULL
;
1026 if (mem_cgroup_disabled())
1030 root
= root_mem_cgroup
;
1032 if (prev
&& !reclaim
)
1038 struct mem_cgroup_per_node
*mz
;
1040 mz
= root
->nodeinfo
[reclaim
->pgdat
->node_id
];
1043 if (prev
&& reclaim
->generation
!= iter
->generation
)
1047 pos
= READ_ONCE(iter
->position
);
1048 if (!pos
|| css_tryget(&pos
->css
))
1051 * css reference reached zero, so iter->position will
1052 * be cleared by ->css_released. However, we should not
1053 * rely on this happening soon, because ->css_released
1054 * is called from a work queue, and by busy-waiting we
1055 * might block it. So we clear iter->position right
1058 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1066 css
= css_next_descendant_pre(css
, &root
->css
);
1069 * Reclaimers share the hierarchy walk, and a
1070 * new one might jump in right at the end of
1071 * the hierarchy - make sure they see at least
1072 * one group and restart from the beginning.
1080 * Verify the css and acquire a reference. The root
1081 * is provided by the caller, so we know it's alive
1082 * and kicking, and don't take an extra reference.
1084 memcg
= mem_cgroup_from_css(css
);
1086 if (css
== &root
->css
)
1089 if (css_tryget(css
))
1097 * The position could have already been updated by a competing
1098 * thread, so check that the value hasn't changed since we read
1099 * it to avoid reclaiming from the same cgroup twice.
1101 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1109 reclaim
->generation
= iter
->generation
;
1114 if (prev
&& prev
!= root
)
1115 css_put(&prev
->css
);
1121 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1122 * @root: hierarchy root
1123 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1125 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1126 struct mem_cgroup
*prev
)
1129 root
= root_mem_cgroup
;
1130 if (prev
&& prev
!= root
)
1131 css_put(&prev
->css
);
1134 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1135 struct mem_cgroup
*dead_memcg
)
1137 struct mem_cgroup_reclaim_iter
*iter
;
1138 struct mem_cgroup_per_node
*mz
;
1141 for_each_node(nid
) {
1142 mz
= from
->nodeinfo
[nid
];
1144 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1148 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1150 struct mem_cgroup
*memcg
= dead_memcg
;
1151 struct mem_cgroup
*last
;
1154 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1156 } while ((memcg
= parent_mem_cgroup(memcg
)));
1159 * When cgruop1 non-hierarchy mode is used,
1160 * parent_mem_cgroup() does not walk all the way up to the
1161 * cgroup root (root_mem_cgroup). So we have to handle
1162 * dead_memcg from cgroup root separately.
1164 if (last
!= root_mem_cgroup
)
1165 __invalidate_reclaim_iterators(root_mem_cgroup
,
1170 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1171 * @memcg: hierarchy root
1172 * @fn: function to call for each task
1173 * @arg: argument passed to @fn
1175 * This function iterates over tasks attached to @memcg or to any of its
1176 * descendants and calls @fn for each task. If @fn returns a non-zero
1177 * value, the function breaks the iteration loop and returns the value.
1178 * Otherwise, it will iterate over all tasks and return 0.
1180 * This function must not be called for the root memory cgroup.
1182 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1183 int (*fn
)(struct task_struct
*, void *), void *arg
)
1185 struct mem_cgroup
*iter
;
1188 BUG_ON(memcg
== root_mem_cgroup
);
1190 for_each_mem_cgroup_tree(iter
, memcg
) {
1191 struct css_task_iter it
;
1192 struct task_struct
*task
;
1194 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1195 while (!ret
&& (task
= css_task_iter_next(&it
)))
1196 ret
= fn(task
, arg
);
1197 css_task_iter_end(&it
);
1199 mem_cgroup_iter_break(memcg
, iter
);
1206 #ifdef CONFIG_DEBUG_VM
1207 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct page
*page
)
1209 struct mem_cgroup
*memcg
;
1211 if (mem_cgroup_disabled())
1214 memcg
= page_memcg(page
);
1217 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != root_mem_cgroup
, page
);
1219 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != memcg
, page
);
1224 * lock_page_lruvec - lock and return lruvec for a given page.
1227 * These functions are safe to use under any of the following conditions:
1230 * - lock_page_memcg()
1231 * - page->_refcount is zero
1233 struct lruvec
*lock_page_lruvec(struct page
*page
)
1235 struct lruvec
*lruvec
;
1237 lruvec
= mem_cgroup_page_lruvec(page
);
1238 spin_lock(&lruvec
->lru_lock
);
1240 lruvec_memcg_debug(lruvec
, page
);
1245 struct lruvec
*lock_page_lruvec_irq(struct page
*page
)
1247 struct lruvec
*lruvec
;
1249 lruvec
= mem_cgroup_page_lruvec(page
);
1250 spin_lock_irq(&lruvec
->lru_lock
);
1252 lruvec_memcg_debug(lruvec
, page
);
1257 struct lruvec
*lock_page_lruvec_irqsave(struct page
*page
, unsigned long *flags
)
1259 struct lruvec
*lruvec
;
1261 lruvec
= mem_cgroup_page_lruvec(page
);
1262 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1264 lruvec_memcg_debug(lruvec
, page
);
1270 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1271 * @lruvec: mem_cgroup per zone lru vector
1272 * @lru: index of lru list the page is sitting on
1273 * @zid: zone id of the accounted pages
1274 * @nr_pages: positive when adding or negative when removing
1276 * This function must be called under lru_lock, just before a page is added
1277 * to or just after a page is removed from an lru list (that ordering being
1278 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1280 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1281 int zid
, int nr_pages
)
1283 struct mem_cgroup_per_node
*mz
;
1284 unsigned long *lru_size
;
1287 if (mem_cgroup_disabled())
1290 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1291 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1294 *lru_size
+= nr_pages
;
1297 if (WARN_ONCE(size
< 0,
1298 "%s(%p, %d, %d): lru_size %ld\n",
1299 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1305 *lru_size
+= nr_pages
;
1309 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1310 * @memcg: the memory cgroup
1312 * Returns the maximum amount of memory @mem can be charged with, in
1315 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1317 unsigned long margin
= 0;
1318 unsigned long count
;
1319 unsigned long limit
;
1321 count
= page_counter_read(&memcg
->memory
);
1322 limit
= READ_ONCE(memcg
->memory
.max
);
1324 margin
= limit
- count
;
1326 if (do_memsw_account()) {
1327 count
= page_counter_read(&memcg
->memsw
);
1328 limit
= READ_ONCE(memcg
->memsw
.max
);
1330 margin
= min(margin
, limit
- count
);
1339 * A routine for checking "mem" is under move_account() or not.
1341 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1342 * moving cgroups. This is for waiting at high-memory pressure
1345 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1347 struct mem_cgroup
*from
;
1348 struct mem_cgroup
*to
;
1351 * Unlike task_move routines, we access mc.to, mc.from not under
1352 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1354 spin_lock(&mc
.lock
);
1360 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1361 mem_cgroup_is_descendant(to
, memcg
);
1363 spin_unlock(&mc
.lock
);
1367 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1369 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1370 if (mem_cgroup_under_move(memcg
)) {
1372 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1373 /* moving charge context might have finished. */
1376 finish_wait(&mc
.waitq
, &wait
);
1383 struct memory_stat
{
1388 static const struct memory_stat memory_stats
[] = {
1389 { "anon", NR_ANON_MAPPED
},
1390 { "file", NR_FILE_PAGES
},
1391 { "kernel_stack", NR_KERNEL_STACK_KB
},
1392 { "pagetables", NR_PAGETABLE
},
1393 { "percpu", MEMCG_PERCPU_B
},
1394 { "sock", MEMCG_SOCK
},
1395 { "shmem", NR_SHMEM
},
1396 { "file_mapped", NR_FILE_MAPPED
},
1397 { "file_dirty", NR_FILE_DIRTY
},
1398 { "file_writeback", NR_WRITEBACK
},
1400 { "swapcached", NR_SWAPCACHE
},
1402 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1403 { "anon_thp", NR_ANON_THPS
},
1404 { "file_thp", NR_FILE_THPS
},
1405 { "shmem_thp", NR_SHMEM_THPS
},
1407 { "inactive_anon", NR_INACTIVE_ANON
},
1408 { "active_anon", NR_ACTIVE_ANON
},
1409 { "inactive_file", NR_INACTIVE_FILE
},
1410 { "active_file", NR_ACTIVE_FILE
},
1411 { "unevictable", NR_UNEVICTABLE
},
1412 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B
},
1413 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B
},
1415 /* The memory events */
1416 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON
},
1417 { "workingset_refault_file", WORKINGSET_REFAULT_FILE
},
1418 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON
},
1419 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE
},
1420 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON
},
1421 { "workingset_restore_file", WORKINGSET_RESTORE_FILE
},
1422 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM
},
1425 /* Translate stat items to the correct unit for memory.stat output */
1426 static int memcg_page_state_unit(int item
)
1429 case MEMCG_PERCPU_B
:
1430 case NR_SLAB_RECLAIMABLE_B
:
1431 case NR_SLAB_UNRECLAIMABLE_B
:
1432 case WORKINGSET_REFAULT_ANON
:
1433 case WORKINGSET_REFAULT_FILE
:
1434 case WORKINGSET_ACTIVATE_ANON
:
1435 case WORKINGSET_ACTIVATE_FILE
:
1436 case WORKINGSET_RESTORE_ANON
:
1437 case WORKINGSET_RESTORE_FILE
:
1438 case WORKINGSET_NODERECLAIM
:
1440 case NR_KERNEL_STACK_KB
:
1447 static inline unsigned long memcg_page_state_output(struct mem_cgroup
*memcg
,
1450 return memcg_page_state(memcg
, item
) * memcg_page_state_unit(item
);
1453 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1458 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1463 * Provide statistics on the state of the memory subsystem as
1464 * well as cumulative event counters that show past behavior.
1466 * This list is ordered following a combination of these gradients:
1467 * 1) generic big picture -> specifics and details
1468 * 2) reflecting userspace activity -> reflecting kernel heuristics
1470 * Current memory state:
1472 mem_cgroup_flush_stats();
1474 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1477 size
= memcg_page_state_output(memcg
, memory_stats
[i
].idx
);
1478 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1480 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1481 size
+= memcg_page_state_output(memcg
,
1482 NR_SLAB_RECLAIMABLE_B
);
1483 seq_buf_printf(&s
, "slab %llu\n", size
);
1487 /* Accumulated memory events */
1489 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1490 memcg_events(memcg
, PGFAULT
));
1491 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1492 memcg_events(memcg
, PGMAJFAULT
));
1493 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1494 memcg_events(memcg
, PGREFILL
));
1495 seq_buf_printf(&s
, "pgscan %lu\n",
1496 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1497 memcg_events(memcg
, PGSCAN_DIRECT
));
1498 seq_buf_printf(&s
, "pgsteal %lu\n",
1499 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1500 memcg_events(memcg
, PGSTEAL_DIRECT
));
1501 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1502 memcg_events(memcg
, PGACTIVATE
));
1503 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1504 memcg_events(memcg
, PGDEACTIVATE
));
1505 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1506 memcg_events(memcg
, PGLAZYFREE
));
1507 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1508 memcg_events(memcg
, PGLAZYFREED
));
1510 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1511 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1512 memcg_events(memcg
, THP_FAULT_ALLOC
));
1513 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1514 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1515 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1517 /* The above should easily fit into one page */
1518 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1523 #define K(x) ((x) << (PAGE_SHIFT-10))
1525 * mem_cgroup_print_oom_context: Print OOM information relevant to
1526 * memory controller.
1527 * @memcg: The memory cgroup that went over limit
1528 * @p: Task that is going to be killed
1530 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1533 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1538 pr_cont(",oom_memcg=");
1539 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1541 pr_cont(",global_oom");
1543 pr_cont(",task_memcg=");
1544 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1550 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1551 * memory controller.
1552 * @memcg: The memory cgroup that went over limit
1554 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1558 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1559 K((u64
)page_counter_read(&memcg
->memory
)),
1560 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1561 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1562 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1563 K((u64
)page_counter_read(&memcg
->swap
)),
1564 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1566 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1567 K((u64
)page_counter_read(&memcg
->memsw
)),
1568 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1569 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1570 K((u64
)page_counter_read(&memcg
->kmem
)),
1571 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1574 pr_info("Memory cgroup stats for ");
1575 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1577 buf
= memory_stat_format(memcg
);
1585 * Return the memory (and swap, if configured) limit for a memcg.
1587 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1589 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1591 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1592 if (mem_cgroup_swappiness(memcg
))
1593 max
+= min(READ_ONCE(memcg
->swap
.max
),
1594 (unsigned long)total_swap_pages
);
1596 if (mem_cgroup_swappiness(memcg
)) {
1597 /* Calculate swap excess capacity from memsw limit */
1598 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1600 max
+= min(swap
, (unsigned long)total_swap_pages
);
1606 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1608 return page_counter_read(&memcg
->memory
);
1611 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1614 struct oom_control oc
= {
1618 .gfp_mask
= gfp_mask
,
1623 if (mutex_lock_killable(&oom_lock
))
1626 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1630 * A few threads which were not waiting at mutex_lock_killable() can
1631 * fail to bail out. Therefore, check again after holding oom_lock.
1633 ret
= task_is_dying() || out_of_memory(&oc
);
1636 mutex_unlock(&oom_lock
);
1640 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1643 unsigned long *total_scanned
)
1645 struct mem_cgroup
*victim
= NULL
;
1648 unsigned long excess
;
1649 unsigned long nr_scanned
;
1650 struct mem_cgroup_reclaim_cookie reclaim
= {
1654 excess
= soft_limit_excess(root_memcg
);
1657 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1662 * If we have not been able to reclaim
1663 * anything, it might because there are
1664 * no reclaimable pages under this hierarchy
1669 * We want to do more targeted reclaim.
1670 * excess >> 2 is not to excessive so as to
1671 * reclaim too much, nor too less that we keep
1672 * coming back to reclaim from this cgroup
1674 if (total
>= (excess
>> 2) ||
1675 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1680 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1681 pgdat
, &nr_scanned
);
1682 *total_scanned
+= nr_scanned
;
1683 if (!soft_limit_excess(root_memcg
))
1686 mem_cgroup_iter_break(root_memcg
, victim
);
1690 #ifdef CONFIG_LOCKDEP
1691 static struct lockdep_map memcg_oom_lock_dep_map
= {
1692 .name
= "memcg_oom_lock",
1696 static DEFINE_SPINLOCK(memcg_oom_lock
);
1699 * Check OOM-Killer is already running under our hierarchy.
1700 * If someone is running, return false.
1702 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1704 struct mem_cgroup
*iter
, *failed
= NULL
;
1706 spin_lock(&memcg_oom_lock
);
1708 for_each_mem_cgroup_tree(iter
, memcg
) {
1709 if (iter
->oom_lock
) {
1711 * this subtree of our hierarchy is already locked
1712 * so we cannot give a lock.
1715 mem_cgroup_iter_break(memcg
, iter
);
1718 iter
->oom_lock
= true;
1723 * OK, we failed to lock the whole subtree so we have
1724 * to clean up what we set up to the failing subtree
1726 for_each_mem_cgroup_tree(iter
, memcg
) {
1727 if (iter
== failed
) {
1728 mem_cgroup_iter_break(memcg
, iter
);
1731 iter
->oom_lock
= false;
1734 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1736 spin_unlock(&memcg_oom_lock
);
1741 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1743 struct mem_cgroup
*iter
;
1745 spin_lock(&memcg_oom_lock
);
1746 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1747 for_each_mem_cgroup_tree(iter
, memcg
)
1748 iter
->oom_lock
= false;
1749 spin_unlock(&memcg_oom_lock
);
1752 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1754 struct mem_cgroup
*iter
;
1756 spin_lock(&memcg_oom_lock
);
1757 for_each_mem_cgroup_tree(iter
, memcg
)
1759 spin_unlock(&memcg_oom_lock
);
1762 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1764 struct mem_cgroup
*iter
;
1767 * Be careful about under_oom underflows because a child memcg
1768 * could have been added after mem_cgroup_mark_under_oom.
1770 spin_lock(&memcg_oom_lock
);
1771 for_each_mem_cgroup_tree(iter
, memcg
)
1772 if (iter
->under_oom
> 0)
1774 spin_unlock(&memcg_oom_lock
);
1777 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1779 struct oom_wait_info
{
1780 struct mem_cgroup
*memcg
;
1781 wait_queue_entry_t wait
;
1784 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1785 unsigned mode
, int sync
, void *arg
)
1787 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1788 struct mem_cgroup
*oom_wait_memcg
;
1789 struct oom_wait_info
*oom_wait_info
;
1791 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1792 oom_wait_memcg
= oom_wait_info
->memcg
;
1794 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1795 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1797 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1800 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1803 * For the following lockless ->under_oom test, the only required
1804 * guarantee is that it must see the state asserted by an OOM when
1805 * this function is called as a result of userland actions
1806 * triggered by the notification of the OOM. This is trivially
1807 * achieved by invoking mem_cgroup_mark_under_oom() before
1808 * triggering notification.
1810 if (memcg
&& memcg
->under_oom
)
1811 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1821 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1823 enum oom_status ret
;
1826 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1829 memcg_memory_event(memcg
, MEMCG_OOM
);
1832 * We are in the middle of the charge context here, so we
1833 * don't want to block when potentially sitting on a callstack
1834 * that holds all kinds of filesystem and mm locks.
1836 * cgroup1 allows disabling the OOM killer and waiting for outside
1837 * handling until the charge can succeed; remember the context and put
1838 * the task to sleep at the end of the page fault when all locks are
1841 * On the other hand, in-kernel OOM killer allows for an async victim
1842 * memory reclaim (oom_reaper) and that means that we are not solely
1843 * relying on the oom victim to make a forward progress and we can
1844 * invoke the oom killer here.
1846 * Please note that mem_cgroup_out_of_memory might fail to find a
1847 * victim and then we have to bail out from the charge path.
1849 if (memcg
->oom_kill_disable
) {
1850 if (!current
->in_user_fault
)
1852 css_get(&memcg
->css
);
1853 current
->memcg_in_oom
= memcg
;
1854 current
->memcg_oom_gfp_mask
= mask
;
1855 current
->memcg_oom_order
= order
;
1860 mem_cgroup_mark_under_oom(memcg
);
1862 locked
= mem_cgroup_oom_trylock(memcg
);
1865 mem_cgroup_oom_notify(memcg
);
1867 mem_cgroup_unmark_under_oom(memcg
);
1868 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1874 mem_cgroup_oom_unlock(memcg
);
1880 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1881 * @handle: actually kill/wait or just clean up the OOM state
1883 * This has to be called at the end of a page fault if the memcg OOM
1884 * handler was enabled.
1886 * Memcg supports userspace OOM handling where failed allocations must
1887 * sleep on a waitqueue until the userspace task resolves the
1888 * situation. Sleeping directly in the charge context with all kinds
1889 * of locks held is not a good idea, instead we remember an OOM state
1890 * in the task and mem_cgroup_oom_synchronize() has to be called at
1891 * the end of the page fault to complete the OOM handling.
1893 * Returns %true if an ongoing memcg OOM situation was detected and
1894 * completed, %false otherwise.
1896 bool mem_cgroup_oom_synchronize(bool handle
)
1898 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1899 struct oom_wait_info owait
;
1902 /* OOM is global, do not handle */
1909 owait
.memcg
= memcg
;
1910 owait
.wait
.flags
= 0;
1911 owait
.wait
.func
= memcg_oom_wake_function
;
1912 owait
.wait
.private = current
;
1913 INIT_LIST_HEAD(&owait
.wait
.entry
);
1915 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1916 mem_cgroup_mark_under_oom(memcg
);
1918 locked
= mem_cgroup_oom_trylock(memcg
);
1921 mem_cgroup_oom_notify(memcg
);
1923 if (locked
&& !memcg
->oom_kill_disable
) {
1924 mem_cgroup_unmark_under_oom(memcg
);
1925 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1926 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1927 current
->memcg_oom_order
);
1930 mem_cgroup_unmark_under_oom(memcg
);
1931 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1935 mem_cgroup_oom_unlock(memcg
);
1937 * There is no guarantee that an OOM-lock contender
1938 * sees the wakeups triggered by the OOM kill
1939 * uncharges. Wake any sleepers explicitly.
1941 memcg_oom_recover(memcg
);
1944 current
->memcg_in_oom
= NULL
;
1945 css_put(&memcg
->css
);
1950 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1951 * @victim: task to be killed by the OOM killer
1952 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1954 * Returns a pointer to a memory cgroup, which has to be cleaned up
1955 * by killing all belonging OOM-killable tasks.
1957 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1959 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1960 struct mem_cgroup
*oom_domain
)
1962 struct mem_cgroup
*oom_group
= NULL
;
1963 struct mem_cgroup
*memcg
;
1965 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1969 oom_domain
= root_mem_cgroup
;
1973 memcg
= mem_cgroup_from_task(victim
);
1974 if (memcg
== root_mem_cgroup
)
1978 * If the victim task has been asynchronously moved to a different
1979 * memory cgroup, we might end up killing tasks outside oom_domain.
1980 * In this case it's better to ignore memory.group.oom.
1982 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
1986 * Traverse the memory cgroup hierarchy from the victim task's
1987 * cgroup up to the OOMing cgroup (or root) to find the
1988 * highest-level memory cgroup with oom.group set.
1990 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
1991 if (memcg
->oom_group
)
1994 if (memcg
== oom_domain
)
1999 css_get(&oom_group
->css
);
2006 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2008 pr_info("Tasks in ");
2009 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2010 pr_cont(" are going to be killed due to memory.oom.group set\n");
2014 * lock_page_memcg - lock a page and memcg binding
2017 * This function protects unlocked LRU pages from being moved to
2020 * It ensures lifetime of the locked memcg. Caller is responsible
2021 * for the lifetime of the page.
2023 void lock_page_memcg(struct page
*page
)
2025 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2026 struct mem_cgroup
*memcg
;
2027 unsigned long flags
;
2030 * The RCU lock is held throughout the transaction. The fast
2031 * path can get away without acquiring the memcg->move_lock
2032 * because page moving starts with an RCU grace period.
2036 if (mem_cgroup_disabled())
2039 memcg
= page_memcg(head
);
2040 if (unlikely(!memcg
))
2043 #ifdef CONFIG_PROVE_LOCKING
2044 local_irq_save(flags
);
2045 might_lock(&memcg
->move_lock
);
2046 local_irq_restore(flags
);
2049 if (atomic_read(&memcg
->moving_account
) <= 0)
2052 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2053 if (memcg
!= page_memcg(head
)) {
2054 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2059 * When charge migration first begins, we can have multiple
2060 * critical sections holding the fast-path RCU lock and one
2061 * holding the slowpath move_lock. Track the task who has the
2062 * move_lock for unlock_page_memcg().
2064 memcg
->move_lock_task
= current
;
2065 memcg
->move_lock_flags
= flags
;
2067 EXPORT_SYMBOL(lock_page_memcg
);
2069 static void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2071 if (memcg
&& memcg
->move_lock_task
== current
) {
2072 unsigned long flags
= memcg
->move_lock_flags
;
2074 memcg
->move_lock_task
= NULL
;
2075 memcg
->move_lock_flags
= 0;
2077 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2084 * unlock_page_memcg - unlock a page and memcg binding
2087 void unlock_page_memcg(struct page
*page
)
2089 struct page
*head
= compound_head(page
);
2091 __unlock_page_memcg(page_memcg(head
));
2093 EXPORT_SYMBOL(unlock_page_memcg
);
2096 #ifdef CONFIG_MEMCG_KMEM
2097 struct obj_cgroup
*cached_objcg
;
2098 struct pglist_data
*cached_pgdat
;
2099 unsigned int nr_bytes
;
2100 int nr_slab_reclaimable_b
;
2101 int nr_slab_unreclaimable_b
;
2107 struct memcg_stock_pcp
{
2108 struct mem_cgroup
*cached
; /* this never be root cgroup */
2109 unsigned int nr_pages
;
2110 struct obj_stock task_obj
;
2111 struct obj_stock irq_obj
;
2113 struct work_struct work
;
2114 unsigned long flags
;
2115 #define FLUSHING_CACHED_CHARGE 0
2117 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2118 static DEFINE_MUTEX(percpu_charge_mutex
);
2120 #ifdef CONFIG_MEMCG_KMEM
2121 static void drain_obj_stock(struct obj_stock
*stock
);
2122 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2123 struct mem_cgroup
*root_memcg
);
2126 static inline void drain_obj_stock(struct obj_stock
*stock
)
2129 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2130 struct mem_cgroup
*root_memcg
)
2137 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2138 * sequence used in this case to access content from object stock is slow.
2139 * To optimize for user context access, there are now two object stocks for
2140 * task context and interrupt context access respectively.
2142 * The task context object stock can be accessed by disabling preemption only
2143 * which is cheap in non-preempt kernel. The interrupt context object stock
2144 * can only be accessed after disabling interrupt. User context code can
2145 * access interrupt object stock, but not vice versa.
2147 static inline struct obj_stock
*get_obj_stock(unsigned long *pflags
)
2149 struct memcg_stock_pcp
*stock
;
2151 if (likely(in_task())) {
2154 stock
= this_cpu_ptr(&memcg_stock
);
2155 return &stock
->task_obj
;
2158 local_irq_save(*pflags
);
2159 stock
= this_cpu_ptr(&memcg_stock
);
2160 return &stock
->irq_obj
;
2163 static inline void put_obj_stock(unsigned long flags
)
2165 if (likely(in_task()))
2168 local_irq_restore(flags
);
2172 * consume_stock: Try to consume stocked charge on this cpu.
2173 * @memcg: memcg to consume from.
2174 * @nr_pages: how many pages to charge.
2176 * The charges will only happen if @memcg matches the current cpu's memcg
2177 * stock, and at least @nr_pages are available in that stock. Failure to
2178 * service an allocation will refill the stock.
2180 * returns true if successful, false otherwise.
2182 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2184 struct memcg_stock_pcp
*stock
;
2185 unsigned long flags
;
2188 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2191 local_irq_save(flags
);
2193 stock
= this_cpu_ptr(&memcg_stock
);
2194 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2195 stock
->nr_pages
-= nr_pages
;
2199 local_irq_restore(flags
);
2205 * Returns stocks cached in percpu and reset cached information.
2207 static void drain_stock(struct memcg_stock_pcp
*stock
)
2209 struct mem_cgroup
*old
= stock
->cached
;
2214 if (stock
->nr_pages
) {
2215 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2216 if (do_memsw_account())
2217 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2218 stock
->nr_pages
= 0;
2222 stock
->cached
= NULL
;
2225 static void drain_local_stock(struct work_struct
*dummy
)
2227 struct memcg_stock_pcp
*stock
;
2228 unsigned long flags
;
2231 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2232 * drain_stock races is that we always operate on local CPU stock
2233 * here with IRQ disabled
2235 local_irq_save(flags
);
2237 stock
= this_cpu_ptr(&memcg_stock
);
2238 drain_obj_stock(&stock
->irq_obj
);
2240 drain_obj_stock(&stock
->task_obj
);
2242 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2244 local_irq_restore(flags
);
2248 * Cache charges(val) to local per_cpu area.
2249 * This will be consumed by consume_stock() function, later.
2251 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2253 struct memcg_stock_pcp
*stock
;
2254 unsigned long flags
;
2256 local_irq_save(flags
);
2258 stock
= this_cpu_ptr(&memcg_stock
);
2259 if (stock
->cached
!= memcg
) { /* reset if necessary */
2261 css_get(&memcg
->css
);
2262 stock
->cached
= memcg
;
2264 stock
->nr_pages
+= nr_pages
;
2266 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2269 local_irq_restore(flags
);
2273 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2274 * of the hierarchy under it.
2276 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2280 /* If someone's already draining, avoid adding running more workers. */
2281 if (!mutex_trylock(&percpu_charge_mutex
))
2284 * Notify other cpus that system-wide "drain" is running
2285 * We do not care about races with the cpu hotplug because cpu down
2286 * as well as workers from this path always operate on the local
2287 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2290 for_each_online_cpu(cpu
) {
2291 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2292 struct mem_cgroup
*memcg
;
2296 memcg
= stock
->cached
;
2297 if (memcg
&& stock
->nr_pages
&&
2298 mem_cgroup_is_descendant(memcg
, root_memcg
))
2300 else if (obj_stock_flush_required(stock
, root_memcg
))
2305 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2307 drain_local_stock(&stock
->work
);
2309 schedule_work_on(cpu
, &stock
->work
);
2313 mutex_unlock(&percpu_charge_mutex
);
2316 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2318 struct memcg_stock_pcp
*stock
;
2320 stock
= &per_cpu(memcg_stock
, cpu
);
2326 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2327 unsigned int nr_pages
,
2330 unsigned long nr_reclaimed
= 0;
2333 unsigned long pflags
;
2335 if (page_counter_read(&memcg
->memory
) <=
2336 READ_ONCE(memcg
->memory
.high
))
2339 memcg_memory_event(memcg
, MEMCG_HIGH
);
2341 psi_memstall_enter(&pflags
);
2342 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2344 psi_memstall_leave(&pflags
);
2345 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2346 !mem_cgroup_is_root(memcg
));
2348 return nr_reclaimed
;
2351 static void high_work_func(struct work_struct
*work
)
2353 struct mem_cgroup
*memcg
;
2355 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2356 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2360 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2361 * enough to still cause a significant slowdown in most cases, while still
2362 * allowing diagnostics and tracing to proceed without becoming stuck.
2364 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2367 * When calculating the delay, we use these either side of the exponentiation to
2368 * maintain precision and scale to a reasonable number of jiffies (see the table
2371 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2372 * overage ratio to a delay.
2373 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2374 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2375 * to produce a reasonable delay curve.
2377 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2378 * reasonable delay curve compared to precision-adjusted overage, not
2379 * penalising heavily at first, but still making sure that growth beyond the
2380 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2381 * example, with a high of 100 megabytes:
2383 * +-------+------------------------+
2384 * | usage | time to allocate in ms |
2385 * +-------+------------------------+
2407 * +-------+------------------------+
2409 #define MEMCG_DELAY_PRECISION_SHIFT 20
2410 #define MEMCG_DELAY_SCALING_SHIFT 14
2412 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2420 * Prevent division by 0 in overage calculation by acting as if
2421 * it was a threshold of 1 page
2423 high
= max(high
, 1UL);
2425 overage
= usage
- high
;
2426 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2427 return div64_u64(overage
, high
);
2430 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2432 u64 overage
, max_overage
= 0;
2435 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2436 READ_ONCE(memcg
->memory
.high
));
2437 max_overage
= max(overage
, max_overage
);
2438 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2439 !mem_cgroup_is_root(memcg
));
2444 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2446 u64 overage
, max_overage
= 0;
2449 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2450 READ_ONCE(memcg
->swap
.high
));
2452 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2453 max_overage
= max(overage
, max_overage
);
2454 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2455 !mem_cgroup_is_root(memcg
));
2461 * Get the number of jiffies that we should penalise a mischievous cgroup which
2462 * is exceeding its memory.high by checking both it and its ancestors.
2464 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2465 unsigned int nr_pages
,
2468 unsigned long penalty_jiffies
;
2474 * We use overage compared to memory.high to calculate the number of
2475 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2476 * fairly lenient on small overages, and increasingly harsh when the
2477 * memcg in question makes it clear that it has no intention of stopping
2478 * its crazy behaviour, so we exponentially increase the delay based on
2481 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2482 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2483 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2486 * Factor in the task's own contribution to the overage, such that four
2487 * N-sized allocations are throttled approximately the same as one
2488 * 4N-sized allocation.
2490 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2491 * larger the current charge patch is than that.
2493 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2497 * Scheduled by try_charge() to be executed from the userland return path
2498 * and reclaims memory over the high limit.
2500 void mem_cgroup_handle_over_high(void)
2502 unsigned long penalty_jiffies
;
2503 unsigned long pflags
;
2504 unsigned long nr_reclaimed
;
2505 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2506 int nr_retries
= MAX_RECLAIM_RETRIES
;
2507 struct mem_cgroup
*memcg
;
2508 bool in_retry
= false;
2510 if (likely(!nr_pages
))
2513 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2514 current
->memcg_nr_pages_over_high
= 0;
2518 * The allocating task should reclaim at least the batch size, but for
2519 * subsequent retries we only want to do what's necessary to prevent oom
2520 * or breaching resource isolation.
2522 * This is distinct from memory.max or page allocator behaviour because
2523 * memory.high is currently batched, whereas memory.max and the page
2524 * allocator run every time an allocation is made.
2526 nr_reclaimed
= reclaim_high(memcg
,
2527 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2531 * memory.high is breached and reclaim is unable to keep up. Throttle
2532 * allocators proactively to slow down excessive growth.
2534 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2535 mem_find_max_overage(memcg
));
2537 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2538 swap_find_max_overage(memcg
));
2541 * Clamp the max delay per usermode return so as to still keep the
2542 * application moving forwards and also permit diagnostics, albeit
2545 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2548 * Don't sleep if the amount of jiffies this memcg owes us is so low
2549 * that it's not even worth doing, in an attempt to be nice to those who
2550 * go only a small amount over their memory.high value and maybe haven't
2551 * been aggressively reclaimed enough yet.
2553 if (penalty_jiffies
<= HZ
/ 100)
2557 * If reclaim is making forward progress but we're still over
2558 * memory.high, we want to encourage that rather than doing allocator
2561 if (nr_reclaimed
|| nr_retries
--) {
2567 * If we exit early, we're guaranteed to die (since
2568 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2569 * need to account for any ill-begotten jiffies to pay them off later.
2571 psi_memstall_enter(&pflags
);
2572 schedule_timeout_killable(penalty_jiffies
);
2573 psi_memstall_leave(&pflags
);
2576 css_put(&memcg
->css
);
2579 static int try_charge_memcg(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2580 unsigned int nr_pages
)
2582 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2583 int nr_retries
= MAX_RECLAIM_RETRIES
;
2584 struct mem_cgroup
*mem_over_limit
;
2585 struct page_counter
*counter
;
2586 enum oom_status oom_status
;
2587 unsigned long nr_reclaimed
;
2588 bool passed_oom
= false;
2589 bool may_swap
= true;
2590 bool drained
= false;
2591 unsigned long pflags
;
2594 if (consume_stock(memcg
, nr_pages
))
2597 if (!do_memsw_account() ||
2598 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2599 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2601 if (do_memsw_account())
2602 page_counter_uncharge(&memcg
->memsw
, batch
);
2603 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2605 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2609 if (batch
> nr_pages
) {
2615 * Memcg doesn't have a dedicated reserve for atomic
2616 * allocations. But like the global atomic pool, we need to
2617 * put the burden of reclaim on regular allocation requests
2618 * and let these go through as privileged allocations.
2620 if (gfp_mask
& __GFP_ATOMIC
)
2624 * Prevent unbounded recursion when reclaim operations need to
2625 * allocate memory. This might exceed the limits temporarily,
2626 * but we prefer facilitating memory reclaim and getting back
2627 * under the limit over triggering OOM kills in these cases.
2629 if (unlikely(current
->flags
& PF_MEMALLOC
))
2632 if (unlikely(task_in_memcg_oom(current
)))
2635 if (!gfpflags_allow_blocking(gfp_mask
))
2638 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2640 psi_memstall_enter(&pflags
);
2641 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2642 gfp_mask
, may_swap
);
2643 psi_memstall_leave(&pflags
);
2645 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2649 drain_all_stock(mem_over_limit
);
2654 if (gfp_mask
& __GFP_NORETRY
)
2657 * Even though the limit is exceeded at this point, reclaim
2658 * may have been able to free some pages. Retry the charge
2659 * before killing the task.
2661 * Only for regular pages, though: huge pages are rather
2662 * unlikely to succeed so close to the limit, and we fall back
2663 * to regular pages anyway in case of failure.
2665 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2668 * At task move, charge accounts can be doubly counted. So, it's
2669 * better to wait until the end of task_move if something is going on.
2671 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2677 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2680 /* Avoid endless loop for tasks bypassed by the oom killer */
2681 if (passed_oom
&& task_is_dying())
2685 * keep retrying as long as the memcg oom killer is able to make
2686 * a forward progress or bypass the charge if the oom killer
2687 * couldn't make any progress.
2689 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2690 get_order(nr_pages
* PAGE_SIZE
));
2691 if (oom_status
== OOM_SUCCESS
) {
2693 nr_retries
= MAX_RECLAIM_RETRIES
;
2697 if (!(gfp_mask
& __GFP_NOFAIL
))
2701 * The allocation either can't fail or will lead to more memory
2702 * being freed very soon. Allow memory usage go over the limit
2703 * temporarily by force charging it.
2705 page_counter_charge(&memcg
->memory
, nr_pages
);
2706 if (do_memsw_account())
2707 page_counter_charge(&memcg
->memsw
, nr_pages
);
2712 if (batch
> nr_pages
)
2713 refill_stock(memcg
, batch
- nr_pages
);
2716 * If the hierarchy is above the normal consumption range, schedule
2717 * reclaim on returning to userland. We can perform reclaim here
2718 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2719 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2720 * not recorded as it most likely matches current's and won't
2721 * change in the meantime. As high limit is checked again before
2722 * reclaim, the cost of mismatch is negligible.
2725 bool mem_high
, swap_high
;
2727 mem_high
= page_counter_read(&memcg
->memory
) >
2728 READ_ONCE(memcg
->memory
.high
);
2729 swap_high
= page_counter_read(&memcg
->swap
) >
2730 READ_ONCE(memcg
->swap
.high
);
2732 /* Don't bother a random interrupted task */
2733 if (in_interrupt()) {
2735 schedule_work(&memcg
->high_work
);
2741 if (mem_high
|| swap_high
) {
2743 * The allocating tasks in this cgroup will need to do
2744 * reclaim or be throttled to prevent further growth
2745 * of the memory or swap footprints.
2747 * Target some best-effort fairness between the tasks,
2748 * and distribute reclaim work and delay penalties
2749 * based on how much each task is actually allocating.
2751 current
->memcg_nr_pages_over_high
+= batch
;
2752 set_notify_resume(current
);
2755 } while ((memcg
= parent_mem_cgroup(memcg
)));
2760 static inline int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2761 unsigned int nr_pages
)
2763 if (mem_cgroup_is_root(memcg
))
2766 return try_charge_memcg(memcg
, gfp_mask
, nr_pages
);
2769 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2770 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2772 if (mem_cgroup_is_root(memcg
))
2775 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2776 if (do_memsw_account())
2777 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2781 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2783 VM_BUG_ON_PAGE(page_memcg(page
), page
);
2785 * Any of the following ensures page's memcg stability:
2789 * - lock_page_memcg()
2790 * - exclusive reference
2792 page
->memcg_data
= (unsigned long)memcg
;
2795 static struct mem_cgroup
*get_mem_cgroup_from_objcg(struct obj_cgroup
*objcg
)
2797 struct mem_cgroup
*memcg
;
2801 memcg
= obj_cgroup_memcg(objcg
);
2802 if (unlikely(!css_tryget(&memcg
->css
)))
2809 #ifdef CONFIG_MEMCG_KMEM
2811 * The allocated objcg pointers array is not accounted directly.
2812 * Moreover, it should not come from DMA buffer and is not readily
2813 * reclaimable. So those GFP bits should be masked off.
2815 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2817 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2818 gfp_t gfp
, bool new_page
)
2820 unsigned int objects
= objs_per_slab_page(s
, page
);
2821 unsigned long memcg_data
;
2824 gfp
&= ~OBJCGS_CLEAR_MASK
;
2825 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2830 memcg_data
= (unsigned long) vec
| MEMCG_DATA_OBJCGS
;
2833 * If the slab page is brand new and nobody can yet access
2834 * it's memcg_data, no synchronization is required and
2835 * memcg_data can be simply assigned.
2837 page
->memcg_data
= memcg_data
;
2838 } else if (cmpxchg(&page
->memcg_data
, 0, memcg_data
)) {
2840 * If the slab page is already in use, somebody can allocate
2841 * and assign obj_cgroups in parallel. In this case the existing
2842 * objcg vector should be reused.
2848 kmemleak_not_leak(vec
);
2853 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2855 * A passed kernel object can be a slab object or a generic kernel page, so
2856 * different mechanisms for getting the memory cgroup pointer should be used.
2857 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2858 * can not know for sure how the kernel object is implemented.
2859 * mem_cgroup_from_obj() can be safely used in such cases.
2861 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2862 * cgroup_mutex, etc.
2864 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2868 if (mem_cgroup_disabled())
2871 page
= virt_to_head_page(p
);
2874 * Slab objects are accounted individually, not per-page.
2875 * Memcg membership data for each individual object is saved in
2876 * the page->obj_cgroups.
2878 if (page_objcgs_check(page
)) {
2879 struct obj_cgroup
*objcg
;
2882 off
= obj_to_index(page
->slab_cache
, page
, p
);
2883 objcg
= page_objcgs(page
)[off
];
2885 return obj_cgroup_memcg(objcg
);
2891 * page_memcg_check() is used here, because page_has_obj_cgroups()
2892 * check above could fail because the object cgroups vector wasn't set
2893 * at that moment, but it can be set concurrently.
2894 * page_memcg_check(page) will guarantee that a proper memory
2895 * cgroup pointer or NULL will be returned.
2897 return page_memcg_check(page
);
2900 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2902 struct obj_cgroup
*objcg
= NULL
;
2903 struct mem_cgroup
*memcg
;
2905 if (memcg_kmem_bypass())
2909 if (unlikely(active_memcg()))
2910 memcg
= active_memcg();
2912 memcg
= mem_cgroup_from_task(current
);
2914 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
2915 objcg
= rcu_dereference(memcg
->objcg
);
2916 if (objcg
&& obj_cgroup_tryget(objcg
))
2925 static int memcg_alloc_cache_id(void)
2930 id
= ida_simple_get(&memcg_cache_ida
,
2931 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2935 if (id
< memcg_nr_cache_ids
)
2939 * There's no space for the new id in memcg_caches arrays,
2940 * so we have to grow them.
2942 down_write(&memcg_cache_ids_sem
);
2944 size
= 2 * (id
+ 1);
2945 if (size
< MEMCG_CACHES_MIN_SIZE
)
2946 size
= MEMCG_CACHES_MIN_SIZE
;
2947 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2948 size
= MEMCG_CACHES_MAX_SIZE
;
2950 err
= memcg_update_all_list_lrus(size
);
2952 memcg_nr_cache_ids
= size
;
2954 up_write(&memcg_cache_ids_sem
);
2957 ida_simple_remove(&memcg_cache_ida
, id
);
2963 static void memcg_free_cache_id(int id
)
2965 ida_simple_remove(&memcg_cache_ida
, id
);
2969 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2970 * @objcg: object cgroup to uncharge
2971 * @nr_pages: number of pages to uncharge
2973 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
2974 unsigned int nr_pages
)
2976 struct mem_cgroup
*memcg
;
2978 memcg
= get_mem_cgroup_from_objcg(objcg
);
2980 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2981 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2982 refill_stock(memcg
, nr_pages
);
2984 css_put(&memcg
->css
);
2988 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2989 * @objcg: object cgroup to charge
2990 * @gfp: reclaim mode
2991 * @nr_pages: number of pages to charge
2993 * Returns 0 on success, an error code on failure.
2995 static int obj_cgroup_charge_pages(struct obj_cgroup
*objcg
, gfp_t gfp
,
2996 unsigned int nr_pages
)
2998 struct page_counter
*counter
;
2999 struct mem_cgroup
*memcg
;
3002 memcg
= get_mem_cgroup_from_objcg(objcg
);
3004 ret
= try_charge_memcg(memcg
, gfp
, nr_pages
);
3008 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
3009 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
3012 * Enforce __GFP_NOFAIL allocation because callers are not
3013 * prepared to see failures and likely do not have any failure
3016 if (gfp
& __GFP_NOFAIL
) {
3017 page_counter_charge(&memcg
->kmem
, nr_pages
);
3020 cancel_charge(memcg
, nr_pages
);
3024 css_put(&memcg
->css
);
3030 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3031 * @page: page to charge
3032 * @gfp: reclaim mode
3033 * @order: allocation order
3035 * Returns 0 on success, an error code on failure.
3037 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3039 struct obj_cgroup
*objcg
;
3042 objcg
= get_obj_cgroup_from_current();
3044 ret
= obj_cgroup_charge_pages(objcg
, gfp
, 1 << order
);
3046 page
->memcg_data
= (unsigned long)objcg
|
3050 obj_cgroup_put(objcg
);
3056 * __memcg_kmem_uncharge_page: uncharge a kmem page
3057 * @page: page to uncharge
3058 * @order: allocation order
3060 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3062 struct obj_cgroup
*objcg
;
3063 unsigned int nr_pages
= 1 << order
;
3065 if (!PageMemcgKmem(page
))
3068 objcg
= __page_objcg(page
);
3069 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3070 page
->memcg_data
= 0;
3071 obj_cgroup_put(objcg
);
3074 void mod_objcg_state(struct obj_cgroup
*objcg
, struct pglist_data
*pgdat
,
3075 enum node_stat_item idx
, int nr
)
3077 unsigned long flags
;
3078 struct obj_stock
*stock
= get_obj_stock(&flags
);
3082 * Save vmstat data in stock and skip vmstat array update unless
3083 * accumulating over a page of vmstat data or when pgdat or idx
3086 if (stock
->cached_objcg
!= objcg
) {
3087 drain_obj_stock(stock
);
3088 obj_cgroup_get(objcg
);
3089 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3090 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3091 stock
->cached_objcg
= objcg
;
3092 stock
->cached_pgdat
= pgdat
;
3093 } else if (stock
->cached_pgdat
!= pgdat
) {
3094 /* Flush the existing cached vmstat data */
3095 struct pglist_data
*oldpg
= stock
->cached_pgdat
;
3097 if (stock
->nr_slab_reclaimable_b
) {
3098 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_RECLAIMABLE_B
,
3099 stock
->nr_slab_reclaimable_b
);
3100 stock
->nr_slab_reclaimable_b
= 0;
3102 if (stock
->nr_slab_unreclaimable_b
) {
3103 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_UNRECLAIMABLE_B
,
3104 stock
->nr_slab_unreclaimable_b
);
3105 stock
->nr_slab_unreclaimable_b
= 0;
3107 stock
->cached_pgdat
= pgdat
;
3110 bytes
= (idx
== NR_SLAB_RECLAIMABLE_B
) ? &stock
->nr_slab_reclaimable_b
3111 : &stock
->nr_slab_unreclaimable_b
;
3113 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3114 * cached locally at least once before pushing it out.
3121 if (abs(*bytes
) > PAGE_SIZE
) {
3129 mod_objcg_mlstate(objcg
, pgdat
, idx
, nr
);
3131 put_obj_stock(flags
);
3134 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3136 unsigned long flags
;
3137 struct obj_stock
*stock
= get_obj_stock(&flags
);
3140 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3141 stock
->nr_bytes
-= nr_bytes
;
3145 put_obj_stock(flags
);
3150 static void drain_obj_stock(struct obj_stock
*stock
)
3152 struct obj_cgroup
*old
= stock
->cached_objcg
;
3157 if (stock
->nr_bytes
) {
3158 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3159 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3162 obj_cgroup_uncharge_pages(old
, nr_pages
);
3165 * The leftover is flushed to the centralized per-memcg value.
3166 * On the next attempt to refill obj stock it will be moved
3167 * to a per-cpu stock (probably, on an other CPU), see
3168 * refill_obj_stock().
3170 * How often it's flushed is a trade-off between the memory
3171 * limit enforcement accuracy and potential CPU contention,
3172 * so it might be changed in the future.
3174 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3175 stock
->nr_bytes
= 0;
3179 * Flush the vmstat data in current stock
3181 if (stock
->nr_slab_reclaimable_b
|| stock
->nr_slab_unreclaimable_b
) {
3182 if (stock
->nr_slab_reclaimable_b
) {
3183 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3184 NR_SLAB_RECLAIMABLE_B
,
3185 stock
->nr_slab_reclaimable_b
);
3186 stock
->nr_slab_reclaimable_b
= 0;
3188 if (stock
->nr_slab_unreclaimable_b
) {
3189 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3190 NR_SLAB_UNRECLAIMABLE_B
,
3191 stock
->nr_slab_unreclaimable_b
);
3192 stock
->nr_slab_unreclaimable_b
= 0;
3194 stock
->cached_pgdat
= NULL
;
3197 obj_cgroup_put(old
);
3198 stock
->cached_objcg
= NULL
;
3201 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3202 struct mem_cgroup
*root_memcg
)
3204 struct mem_cgroup
*memcg
;
3206 if (in_task() && stock
->task_obj
.cached_objcg
) {
3207 memcg
= obj_cgroup_memcg(stock
->task_obj
.cached_objcg
);
3208 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3211 if (stock
->irq_obj
.cached_objcg
) {
3212 memcg
= obj_cgroup_memcg(stock
->irq_obj
.cached_objcg
);
3213 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3220 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
,
3221 bool allow_uncharge
)
3223 unsigned long flags
;
3224 struct obj_stock
*stock
= get_obj_stock(&flags
);
3225 unsigned int nr_pages
= 0;
3227 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3228 drain_obj_stock(stock
);
3229 obj_cgroup_get(objcg
);
3230 stock
->cached_objcg
= objcg
;
3231 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3232 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3233 allow_uncharge
= true; /* Allow uncharge when objcg changes */
3235 stock
->nr_bytes
+= nr_bytes
;
3237 if (allow_uncharge
&& (stock
->nr_bytes
> PAGE_SIZE
)) {
3238 nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3239 stock
->nr_bytes
&= (PAGE_SIZE
- 1);
3242 put_obj_stock(flags
);
3245 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3248 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3250 unsigned int nr_pages
, nr_bytes
;
3253 if (consume_obj_stock(objcg
, size
))
3257 * In theory, objcg->nr_charged_bytes can have enough
3258 * pre-charged bytes to satisfy the allocation. However,
3259 * flushing objcg->nr_charged_bytes requires two atomic
3260 * operations, and objcg->nr_charged_bytes can't be big.
3261 * The shared objcg->nr_charged_bytes can also become a
3262 * performance bottleneck if all tasks of the same memcg are
3263 * trying to update it. So it's better to ignore it and try
3264 * grab some new pages. The stock's nr_bytes will be flushed to
3265 * objcg->nr_charged_bytes later on when objcg changes.
3267 * The stock's nr_bytes may contain enough pre-charged bytes
3268 * to allow one less page from being charged, but we can't rely
3269 * on the pre-charged bytes not being changed outside of
3270 * consume_obj_stock() or refill_obj_stock(). So ignore those
3271 * pre-charged bytes as well when charging pages. To avoid a
3272 * page uncharge right after a page charge, we set the
3273 * allow_uncharge flag to false when calling refill_obj_stock()
3274 * to temporarily allow the pre-charged bytes to exceed the page
3275 * size limit. The maximum reachable value of the pre-charged
3276 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3279 nr_pages
= size
>> PAGE_SHIFT
;
3280 nr_bytes
= size
& (PAGE_SIZE
- 1);
3285 ret
= obj_cgroup_charge_pages(objcg
, gfp
, nr_pages
);
3286 if (!ret
&& nr_bytes
)
3287 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
, false);
3292 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3294 refill_obj_stock(objcg
, size
, true);
3297 #endif /* CONFIG_MEMCG_KMEM */
3300 * Because page_memcg(head) is not set on tails, set it now.
3302 void split_page_memcg(struct page
*head
, unsigned int nr
)
3304 struct mem_cgroup
*memcg
= page_memcg(head
);
3307 if (mem_cgroup_disabled() || !memcg
)
3310 for (i
= 1; i
< nr
; i
++)
3311 head
[i
].memcg_data
= head
->memcg_data
;
3313 if (PageMemcgKmem(head
))
3314 obj_cgroup_get_many(__page_objcg(head
), nr
- 1);
3316 css_get_many(&memcg
->css
, nr
- 1);
3319 #ifdef CONFIG_MEMCG_SWAP
3321 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3322 * @entry: swap entry to be moved
3323 * @from: mem_cgroup which the entry is moved from
3324 * @to: mem_cgroup which the entry is moved to
3326 * It succeeds only when the swap_cgroup's record for this entry is the same
3327 * as the mem_cgroup's id of @from.
3329 * Returns 0 on success, -EINVAL on failure.
3331 * The caller must have charged to @to, IOW, called page_counter_charge() about
3332 * both res and memsw, and called css_get().
3334 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3335 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3337 unsigned short old_id
, new_id
;
3339 old_id
= mem_cgroup_id(from
);
3340 new_id
= mem_cgroup_id(to
);
3342 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3343 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3344 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3350 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3351 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3357 static DEFINE_MUTEX(memcg_max_mutex
);
3359 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3360 unsigned long max
, bool memsw
)
3362 bool enlarge
= false;
3363 bool drained
= false;
3365 bool limits_invariant
;
3366 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3369 if (signal_pending(current
)) {
3374 mutex_lock(&memcg_max_mutex
);
3376 * Make sure that the new limit (memsw or memory limit) doesn't
3377 * break our basic invariant rule memory.max <= memsw.max.
3379 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3380 max
<= memcg
->memsw
.max
;
3381 if (!limits_invariant
) {
3382 mutex_unlock(&memcg_max_mutex
);
3386 if (max
> counter
->max
)
3388 ret
= page_counter_set_max(counter
, max
);
3389 mutex_unlock(&memcg_max_mutex
);
3395 drain_all_stock(memcg
);
3400 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3401 GFP_KERNEL
, !memsw
)) {
3407 if (!ret
&& enlarge
)
3408 memcg_oom_recover(memcg
);
3413 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3415 unsigned long *total_scanned
)
3417 unsigned long nr_reclaimed
= 0;
3418 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3419 unsigned long reclaimed
;
3421 struct mem_cgroup_tree_per_node
*mctz
;
3422 unsigned long excess
;
3423 unsigned long nr_scanned
;
3428 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3431 * Do not even bother to check the largest node if the root
3432 * is empty. Do it lockless to prevent lock bouncing. Races
3433 * are acceptable as soft limit is best effort anyway.
3435 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3439 * This loop can run a while, specially if mem_cgroup's continuously
3440 * keep exceeding their soft limit and putting the system under
3447 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3452 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3453 gfp_mask
, &nr_scanned
);
3454 nr_reclaimed
+= reclaimed
;
3455 *total_scanned
+= nr_scanned
;
3456 spin_lock_irq(&mctz
->lock
);
3457 __mem_cgroup_remove_exceeded(mz
, mctz
);
3460 * If we failed to reclaim anything from this memory cgroup
3461 * it is time to move on to the next cgroup
3465 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3467 excess
= soft_limit_excess(mz
->memcg
);
3469 * One school of thought says that we should not add
3470 * back the node to the tree if reclaim returns 0.
3471 * But our reclaim could return 0, simply because due
3472 * to priority we are exposing a smaller subset of
3473 * memory to reclaim from. Consider this as a longer
3476 /* If excess == 0, no tree ops */
3477 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3478 spin_unlock_irq(&mctz
->lock
);
3479 css_put(&mz
->memcg
->css
);
3482 * Could not reclaim anything and there are no more
3483 * mem cgroups to try or we seem to be looping without
3484 * reclaiming anything.
3486 if (!nr_reclaimed
&&
3488 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3490 } while (!nr_reclaimed
);
3492 css_put(&next_mz
->memcg
->css
);
3493 return nr_reclaimed
;
3497 * Reclaims as many pages from the given memcg as possible.
3499 * Caller is responsible for holding css reference for memcg.
3501 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3503 int nr_retries
= MAX_RECLAIM_RETRIES
;
3505 /* we call try-to-free pages for make this cgroup empty */
3506 lru_add_drain_all();
3508 drain_all_stock(memcg
);
3510 /* try to free all pages in this cgroup */
3511 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3514 if (signal_pending(current
))
3517 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3521 /* maybe some writeback is necessary */
3522 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3530 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3531 char *buf
, size_t nbytes
,
3534 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3536 if (mem_cgroup_is_root(memcg
))
3538 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3541 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3547 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3548 struct cftype
*cft
, u64 val
)
3553 pr_warn_once("Non-hierarchical mode is deprecated. "
3554 "Please report your usecase to linux-mm@kvack.org if you "
3555 "depend on this functionality.\n");
3560 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3564 if (mem_cgroup_is_root(memcg
)) {
3565 mem_cgroup_flush_stats();
3566 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3567 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3569 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3572 val
= page_counter_read(&memcg
->memory
);
3574 val
= page_counter_read(&memcg
->memsw
);
3587 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3590 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3591 struct page_counter
*counter
;
3593 switch (MEMFILE_TYPE(cft
->private)) {
3595 counter
= &memcg
->memory
;
3598 counter
= &memcg
->memsw
;
3601 counter
= &memcg
->kmem
;
3604 counter
= &memcg
->tcpmem
;
3610 switch (MEMFILE_ATTR(cft
->private)) {
3612 if (counter
== &memcg
->memory
)
3613 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3614 if (counter
== &memcg
->memsw
)
3615 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3616 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3618 return (u64
)counter
->max
* PAGE_SIZE
;
3620 return (u64
)counter
->watermark
* PAGE_SIZE
;
3622 return counter
->failcnt
;
3623 case RES_SOFT_LIMIT
:
3624 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3630 #ifdef CONFIG_MEMCG_KMEM
3631 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3633 struct obj_cgroup
*objcg
;
3636 if (cgroup_memory_nokmem
)
3639 BUG_ON(memcg
->kmemcg_id
>= 0);
3640 BUG_ON(memcg
->kmem_state
);
3642 memcg_id
= memcg_alloc_cache_id();
3646 objcg
= obj_cgroup_alloc();
3648 memcg_free_cache_id(memcg_id
);
3651 objcg
->memcg
= memcg
;
3652 rcu_assign_pointer(memcg
->objcg
, objcg
);
3654 static_branch_enable(&memcg_kmem_enabled_key
);
3656 memcg
->kmemcg_id
= memcg_id
;
3657 memcg
->kmem_state
= KMEM_ONLINE
;
3662 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3664 struct cgroup_subsys_state
*css
;
3665 struct mem_cgroup
*parent
, *child
;
3668 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3671 memcg
->kmem_state
= KMEM_ALLOCATED
;
3673 parent
= parent_mem_cgroup(memcg
);
3675 parent
= root_mem_cgroup
;
3677 memcg_reparent_objcgs(memcg
, parent
);
3679 kmemcg_id
= memcg
->kmemcg_id
;
3680 BUG_ON(kmemcg_id
< 0);
3683 * Change kmemcg_id of this cgroup and all its descendants to the
3684 * parent's id, and then move all entries from this cgroup's list_lrus
3685 * to ones of the parent. After we have finished, all list_lrus
3686 * corresponding to this cgroup are guaranteed to remain empty. The
3687 * ordering is imposed by list_lru_node->lock taken by
3688 * memcg_drain_all_list_lrus().
3690 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3691 css_for_each_descendant_pre(css
, &memcg
->css
) {
3692 child
= mem_cgroup_from_css(css
);
3693 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3694 child
->kmemcg_id
= parent
->kmemcg_id
;
3698 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3700 memcg_free_cache_id(kmemcg_id
);
3703 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3705 /* css_alloc() failed, offlining didn't happen */
3706 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3707 memcg_offline_kmem(memcg
);
3710 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3714 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3717 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3720 #endif /* CONFIG_MEMCG_KMEM */
3722 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3727 mutex_lock(&memcg_max_mutex
);
3728 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3729 mutex_unlock(&memcg_max_mutex
);
3733 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3737 mutex_lock(&memcg_max_mutex
);
3739 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3743 if (!memcg
->tcpmem_active
) {
3745 * The active flag needs to be written after the static_key
3746 * update. This is what guarantees that the socket activation
3747 * function is the last one to run. See mem_cgroup_sk_alloc()
3748 * for details, and note that we don't mark any socket as
3749 * belonging to this memcg until that flag is up.
3751 * We need to do this, because static_keys will span multiple
3752 * sites, but we can't control their order. If we mark a socket
3753 * as accounted, but the accounting functions are not patched in
3754 * yet, we'll lose accounting.
3756 * We never race with the readers in mem_cgroup_sk_alloc(),
3757 * because when this value change, the code to process it is not
3760 static_branch_inc(&memcg_sockets_enabled_key
);
3761 memcg
->tcpmem_active
= true;
3764 mutex_unlock(&memcg_max_mutex
);
3769 * The user of this function is...
3772 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3773 char *buf
, size_t nbytes
, loff_t off
)
3775 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3776 unsigned long nr_pages
;
3779 buf
= strstrip(buf
);
3780 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3784 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3786 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3790 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3792 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3795 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3798 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3799 "Please report your usecase to linux-mm@kvack.org if you "
3800 "depend on this functionality.\n");
3801 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3804 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3808 case RES_SOFT_LIMIT
:
3809 memcg
->soft_limit
= nr_pages
;
3813 return ret
?: nbytes
;
3816 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3817 size_t nbytes
, loff_t off
)
3819 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3820 struct page_counter
*counter
;
3822 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3824 counter
= &memcg
->memory
;
3827 counter
= &memcg
->memsw
;
3830 counter
= &memcg
->kmem
;
3833 counter
= &memcg
->tcpmem
;
3839 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3841 page_counter_reset_watermark(counter
);
3844 counter
->failcnt
= 0;
3853 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3856 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3860 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3861 struct cftype
*cft
, u64 val
)
3863 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3865 if (val
& ~MOVE_MASK
)
3869 * No kind of locking is needed in here, because ->can_attach() will
3870 * check this value once in the beginning of the process, and then carry
3871 * on with stale data. This means that changes to this value will only
3872 * affect task migrations starting after the change.
3874 memcg
->move_charge_at_immigrate
= val
;
3878 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3879 struct cftype
*cft
, u64 val
)
3887 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3888 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3889 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3891 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3892 int nid
, unsigned int lru_mask
, bool tree
)
3894 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3895 unsigned long nr
= 0;
3898 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3901 if (!(BIT(lru
) & lru_mask
))
3904 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3906 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3911 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3912 unsigned int lru_mask
,
3915 unsigned long nr
= 0;
3919 if (!(BIT(lru
) & lru_mask
))
3922 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3924 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3929 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3933 unsigned int lru_mask
;
3936 static const struct numa_stat stats
[] = {
3937 { "total", LRU_ALL
},
3938 { "file", LRU_ALL_FILE
},
3939 { "anon", LRU_ALL_ANON
},
3940 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3942 const struct numa_stat
*stat
;
3944 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3946 mem_cgroup_flush_stats();
3948 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3949 seq_printf(m
, "%s=%lu", stat
->name
,
3950 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3952 for_each_node_state(nid
, N_MEMORY
)
3953 seq_printf(m
, " N%d=%lu", nid
,
3954 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3955 stat
->lru_mask
, false));
3959 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3961 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
3962 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3964 for_each_node_state(nid
, N_MEMORY
)
3965 seq_printf(m
, " N%d=%lu", nid
,
3966 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3967 stat
->lru_mask
, true));
3973 #endif /* CONFIG_NUMA */
3975 static const unsigned int memcg1_stats
[] = {
3978 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3988 static const char *const memcg1_stat_names
[] = {
3991 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4001 /* Universal VM events cgroup1 shows, original sort order */
4002 static const unsigned int memcg1_events
[] = {
4009 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4011 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4012 unsigned long memory
, memsw
;
4013 struct mem_cgroup
*mi
;
4016 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4018 mem_cgroup_flush_stats();
4020 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4023 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4025 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4026 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4029 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4030 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4031 memcg_events_local(memcg
, memcg1_events
[i
]));
4033 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4034 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4035 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4038 /* Hierarchical information */
4039 memory
= memsw
= PAGE_COUNTER_MAX
;
4040 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4041 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4042 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4044 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4045 (u64
)memory
* PAGE_SIZE
);
4046 if (do_memsw_account())
4047 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4048 (u64
)memsw
* PAGE_SIZE
);
4050 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4053 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4055 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4056 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4057 (u64
)nr
* PAGE_SIZE
);
4060 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4061 seq_printf(m
, "total_%s %llu\n",
4062 vm_event_name(memcg1_events
[i
]),
4063 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4065 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4066 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4067 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4070 #ifdef CONFIG_DEBUG_VM
4073 struct mem_cgroup_per_node
*mz
;
4074 unsigned long anon_cost
= 0;
4075 unsigned long file_cost
= 0;
4077 for_each_online_pgdat(pgdat
) {
4078 mz
= memcg
->nodeinfo
[pgdat
->node_id
];
4080 anon_cost
+= mz
->lruvec
.anon_cost
;
4081 file_cost
+= mz
->lruvec
.file_cost
;
4083 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4084 seq_printf(m
, "file_cost %lu\n", file_cost
);
4091 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4094 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4096 return mem_cgroup_swappiness(memcg
);
4099 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4100 struct cftype
*cft
, u64 val
)
4102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4107 if (!mem_cgroup_is_root(memcg
))
4108 memcg
->swappiness
= val
;
4110 vm_swappiness
= val
;
4115 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4117 struct mem_cgroup_threshold_ary
*t
;
4118 unsigned long usage
;
4123 t
= rcu_dereference(memcg
->thresholds
.primary
);
4125 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4130 usage
= mem_cgroup_usage(memcg
, swap
);
4133 * current_threshold points to threshold just below or equal to usage.
4134 * If it's not true, a threshold was crossed after last
4135 * call of __mem_cgroup_threshold().
4137 i
= t
->current_threshold
;
4140 * Iterate backward over array of thresholds starting from
4141 * current_threshold and check if a threshold is crossed.
4142 * If none of thresholds below usage is crossed, we read
4143 * only one element of the array here.
4145 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4146 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4148 /* i = current_threshold + 1 */
4152 * Iterate forward over array of thresholds starting from
4153 * current_threshold+1 and check if a threshold is crossed.
4154 * If none of thresholds above usage is crossed, we read
4155 * only one element of the array here.
4157 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4158 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4160 /* Update current_threshold */
4161 t
->current_threshold
= i
- 1;
4166 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4169 __mem_cgroup_threshold(memcg
, false);
4170 if (do_memsw_account())
4171 __mem_cgroup_threshold(memcg
, true);
4173 memcg
= parent_mem_cgroup(memcg
);
4177 static int compare_thresholds(const void *a
, const void *b
)
4179 const struct mem_cgroup_threshold
*_a
= a
;
4180 const struct mem_cgroup_threshold
*_b
= b
;
4182 if (_a
->threshold
> _b
->threshold
)
4185 if (_a
->threshold
< _b
->threshold
)
4191 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4193 struct mem_cgroup_eventfd_list
*ev
;
4195 spin_lock(&memcg_oom_lock
);
4197 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4198 eventfd_signal(ev
->eventfd
, 1);
4200 spin_unlock(&memcg_oom_lock
);
4204 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4206 struct mem_cgroup
*iter
;
4208 for_each_mem_cgroup_tree(iter
, memcg
)
4209 mem_cgroup_oom_notify_cb(iter
);
4212 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4213 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4215 struct mem_cgroup_thresholds
*thresholds
;
4216 struct mem_cgroup_threshold_ary
*new;
4217 unsigned long threshold
;
4218 unsigned long usage
;
4221 ret
= page_counter_memparse(args
, "-1", &threshold
);
4225 mutex_lock(&memcg
->thresholds_lock
);
4228 thresholds
= &memcg
->thresholds
;
4229 usage
= mem_cgroup_usage(memcg
, false);
4230 } else if (type
== _MEMSWAP
) {
4231 thresholds
= &memcg
->memsw_thresholds
;
4232 usage
= mem_cgroup_usage(memcg
, true);
4236 /* Check if a threshold crossed before adding a new one */
4237 if (thresholds
->primary
)
4238 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4240 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4242 /* Allocate memory for new array of thresholds */
4243 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4250 /* Copy thresholds (if any) to new array */
4251 if (thresholds
->primary
)
4252 memcpy(new->entries
, thresholds
->primary
->entries
,
4253 flex_array_size(new, entries
, size
- 1));
4255 /* Add new threshold */
4256 new->entries
[size
- 1].eventfd
= eventfd
;
4257 new->entries
[size
- 1].threshold
= threshold
;
4259 /* Sort thresholds. Registering of new threshold isn't time-critical */
4260 sort(new->entries
, size
, sizeof(*new->entries
),
4261 compare_thresholds
, NULL
);
4263 /* Find current threshold */
4264 new->current_threshold
= -1;
4265 for (i
= 0; i
< size
; i
++) {
4266 if (new->entries
[i
].threshold
<= usage
) {
4268 * new->current_threshold will not be used until
4269 * rcu_assign_pointer(), so it's safe to increment
4272 ++new->current_threshold
;
4277 /* Free old spare buffer and save old primary buffer as spare */
4278 kfree(thresholds
->spare
);
4279 thresholds
->spare
= thresholds
->primary
;
4281 rcu_assign_pointer(thresholds
->primary
, new);
4283 /* To be sure that nobody uses thresholds */
4287 mutex_unlock(&memcg
->thresholds_lock
);
4292 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4293 struct eventfd_ctx
*eventfd
, const char *args
)
4295 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4298 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4299 struct eventfd_ctx
*eventfd
, const char *args
)
4301 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4304 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4305 struct eventfd_ctx
*eventfd
, enum res_type type
)
4307 struct mem_cgroup_thresholds
*thresholds
;
4308 struct mem_cgroup_threshold_ary
*new;
4309 unsigned long usage
;
4310 int i
, j
, size
, entries
;
4312 mutex_lock(&memcg
->thresholds_lock
);
4315 thresholds
= &memcg
->thresholds
;
4316 usage
= mem_cgroup_usage(memcg
, false);
4317 } else if (type
== _MEMSWAP
) {
4318 thresholds
= &memcg
->memsw_thresholds
;
4319 usage
= mem_cgroup_usage(memcg
, true);
4323 if (!thresholds
->primary
)
4326 /* Check if a threshold crossed before removing */
4327 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4329 /* Calculate new number of threshold */
4331 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4332 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4338 new = thresholds
->spare
;
4340 /* If no items related to eventfd have been cleared, nothing to do */
4344 /* Set thresholds array to NULL if we don't have thresholds */
4353 /* Copy thresholds and find current threshold */
4354 new->current_threshold
= -1;
4355 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4356 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4359 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4360 if (new->entries
[j
].threshold
<= usage
) {
4362 * new->current_threshold will not be used
4363 * until rcu_assign_pointer(), so it's safe to increment
4366 ++new->current_threshold
;
4372 /* Swap primary and spare array */
4373 thresholds
->spare
= thresholds
->primary
;
4375 rcu_assign_pointer(thresholds
->primary
, new);
4377 /* To be sure that nobody uses thresholds */
4380 /* If all events are unregistered, free the spare array */
4382 kfree(thresholds
->spare
);
4383 thresholds
->spare
= NULL
;
4386 mutex_unlock(&memcg
->thresholds_lock
);
4389 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4390 struct eventfd_ctx
*eventfd
)
4392 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4395 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4396 struct eventfd_ctx
*eventfd
)
4398 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4401 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4402 struct eventfd_ctx
*eventfd
, const char *args
)
4404 struct mem_cgroup_eventfd_list
*event
;
4406 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4410 spin_lock(&memcg_oom_lock
);
4412 event
->eventfd
= eventfd
;
4413 list_add(&event
->list
, &memcg
->oom_notify
);
4415 /* already in OOM ? */
4416 if (memcg
->under_oom
)
4417 eventfd_signal(eventfd
, 1);
4418 spin_unlock(&memcg_oom_lock
);
4423 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4424 struct eventfd_ctx
*eventfd
)
4426 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4428 spin_lock(&memcg_oom_lock
);
4430 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4431 if (ev
->eventfd
== eventfd
) {
4432 list_del(&ev
->list
);
4437 spin_unlock(&memcg_oom_lock
);
4440 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4442 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4444 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4445 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4446 seq_printf(sf
, "oom_kill %lu\n",
4447 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4451 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4452 struct cftype
*cft
, u64 val
)
4454 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4456 /* cannot set to root cgroup and only 0 and 1 are allowed */
4457 if (mem_cgroup_is_root(memcg
) || !((val
== 0) || (val
== 1)))
4460 memcg
->oom_kill_disable
= val
;
4462 memcg_oom_recover(memcg
);
4467 #ifdef CONFIG_CGROUP_WRITEBACK
4469 #include <trace/events/writeback.h>
4471 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4473 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4476 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4478 wb_domain_exit(&memcg
->cgwb_domain
);
4481 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4483 wb_domain_size_changed(&memcg
->cgwb_domain
);
4486 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4488 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4490 if (!memcg
->css
.parent
)
4493 return &memcg
->cgwb_domain
;
4497 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4498 * @wb: bdi_writeback in question
4499 * @pfilepages: out parameter for number of file pages
4500 * @pheadroom: out parameter for number of allocatable pages according to memcg
4501 * @pdirty: out parameter for number of dirty pages
4502 * @pwriteback: out parameter for number of pages under writeback
4504 * Determine the numbers of file, headroom, dirty, and writeback pages in
4505 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4506 * is a bit more involved.
4508 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4509 * headroom is calculated as the lowest headroom of itself and the
4510 * ancestors. Note that this doesn't consider the actual amount of
4511 * available memory in the system. The caller should further cap
4512 * *@pheadroom accordingly.
4514 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4515 unsigned long *pheadroom
, unsigned long *pdirty
,
4516 unsigned long *pwriteback
)
4518 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4519 struct mem_cgroup
*parent
;
4521 mem_cgroup_flush_stats();
4523 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
4524 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
4525 *pfilepages
= memcg_page_state(memcg
, NR_INACTIVE_FILE
) +
4526 memcg_page_state(memcg
, NR_ACTIVE_FILE
);
4528 *pheadroom
= PAGE_COUNTER_MAX
;
4529 while ((parent
= parent_mem_cgroup(memcg
))) {
4530 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4531 READ_ONCE(memcg
->memory
.high
));
4532 unsigned long used
= page_counter_read(&memcg
->memory
);
4534 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4540 * Foreign dirty flushing
4542 * There's an inherent mismatch between memcg and writeback. The former
4543 * tracks ownership per-page while the latter per-inode. This was a
4544 * deliberate design decision because honoring per-page ownership in the
4545 * writeback path is complicated, may lead to higher CPU and IO overheads
4546 * and deemed unnecessary given that write-sharing an inode across
4547 * different cgroups isn't a common use-case.
4549 * Combined with inode majority-writer ownership switching, this works well
4550 * enough in most cases but there are some pathological cases. For
4551 * example, let's say there are two cgroups A and B which keep writing to
4552 * different but confined parts of the same inode. B owns the inode and
4553 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4554 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4555 * triggering background writeback. A will be slowed down without a way to
4556 * make writeback of the dirty pages happen.
4558 * Conditions like the above can lead to a cgroup getting repeatedly and
4559 * severely throttled after making some progress after each
4560 * dirty_expire_interval while the underlying IO device is almost
4563 * Solving this problem completely requires matching the ownership tracking
4564 * granularities between memcg and writeback in either direction. However,
4565 * the more egregious behaviors can be avoided by simply remembering the
4566 * most recent foreign dirtying events and initiating remote flushes on
4567 * them when local writeback isn't enough to keep the memory clean enough.
4569 * The following two functions implement such mechanism. When a foreign
4570 * page - a page whose memcg and writeback ownerships don't match - is
4571 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4572 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4573 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4574 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4575 * foreign bdi_writebacks which haven't expired. Both the numbers of
4576 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4577 * limited to MEMCG_CGWB_FRN_CNT.
4579 * The mechanism only remembers IDs and doesn't hold any object references.
4580 * As being wrong occasionally doesn't matter, updates and accesses to the
4581 * records are lockless and racy.
4583 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4584 struct bdi_writeback
*wb
)
4586 struct mem_cgroup
*memcg
= page_memcg(page
);
4587 struct memcg_cgwb_frn
*frn
;
4588 u64 now
= get_jiffies_64();
4589 u64 oldest_at
= now
;
4593 trace_track_foreign_dirty(page
, wb
);
4596 * Pick the slot to use. If there is already a slot for @wb, keep
4597 * using it. If not replace the oldest one which isn't being
4600 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4601 frn
= &memcg
->cgwb_frn
[i
];
4602 if (frn
->bdi_id
== wb
->bdi
->id
&&
4603 frn
->memcg_id
== wb
->memcg_css
->id
)
4605 if (time_before64(frn
->at
, oldest_at
) &&
4606 atomic_read(&frn
->done
.cnt
) == 1) {
4608 oldest_at
= frn
->at
;
4612 if (i
< MEMCG_CGWB_FRN_CNT
) {
4614 * Re-using an existing one. Update timestamp lazily to
4615 * avoid making the cacheline hot. We want them to be
4616 * reasonably up-to-date and significantly shorter than
4617 * dirty_expire_interval as that's what expires the record.
4618 * Use the shorter of 1s and dirty_expire_interval / 8.
4620 unsigned long update_intv
=
4621 min_t(unsigned long, HZ
,
4622 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4624 if (time_before64(frn
->at
, now
- update_intv
))
4626 } else if (oldest
>= 0) {
4627 /* replace the oldest free one */
4628 frn
= &memcg
->cgwb_frn
[oldest
];
4629 frn
->bdi_id
= wb
->bdi
->id
;
4630 frn
->memcg_id
= wb
->memcg_css
->id
;
4635 /* issue foreign writeback flushes for recorded foreign dirtying events */
4636 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4638 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4639 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4640 u64 now
= jiffies_64
;
4643 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4644 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4647 * If the record is older than dirty_expire_interval,
4648 * writeback on it has already started. No need to kick it
4649 * off again. Also, don't start a new one if there's
4650 * already one in flight.
4652 if (time_after64(frn
->at
, now
- intv
) &&
4653 atomic_read(&frn
->done
.cnt
) == 1) {
4655 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4656 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
,
4657 WB_REASON_FOREIGN_FLUSH
,
4663 #else /* CONFIG_CGROUP_WRITEBACK */
4665 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4670 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4674 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4678 #endif /* CONFIG_CGROUP_WRITEBACK */
4681 * DO NOT USE IN NEW FILES.
4683 * "cgroup.event_control" implementation.
4685 * This is way over-engineered. It tries to support fully configurable
4686 * events for each user. Such level of flexibility is completely
4687 * unnecessary especially in the light of the planned unified hierarchy.
4689 * Please deprecate this and replace with something simpler if at all
4694 * Unregister event and free resources.
4696 * Gets called from workqueue.
4698 static void memcg_event_remove(struct work_struct
*work
)
4700 struct mem_cgroup_event
*event
=
4701 container_of(work
, struct mem_cgroup_event
, remove
);
4702 struct mem_cgroup
*memcg
= event
->memcg
;
4704 remove_wait_queue(event
->wqh
, &event
->wait
);
4706 event
->unregister_event(memcg
, event
->eventfd
);
4708 /* Notify userspace the event is going away. */
4709 eventfd_signal(event
->eventfd
, 1);
4711 eventfd_ctx_put(event
->eventfd
);
4713 css_put(&memcg
->css
);
4717 * Gets called on EPOLLHUP on eventfd when user closes it.
4719 * Called with wqh->lock held and interrupts disabled.
4721 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4722 int sync
, void *key
)
4724 struct mem_cgroup_event
*event
=
4725 container_of(wait
, struct mem_cgroup_event
, wait
);
4726 struct mem_cgroup
*memcg
= event
->memcg
;
4727 __poll_t flags
= key_to_poll(key
);
4729 if (flags
& EPOLLHUP
) {
4731 * If the event has been detached at cgroup removal, we
4732 * can simply return knowing the other side will cleanup
4735 * We can't race against event freeing since the other
4736 * side will require wqh->lock via remove_wait_queue(),
4739 spin_lock(&memcg
->event_list_lock
);
4740 if (!list_empty(&event
->list
)) {
4741 list_del_init(&event
->list
);
4743 * We are in atomic context, but cgroup_event_remove()
4744 * may sleep, so we have to call it in workqueue.
4746 schedule_work(&event
->remove
);
4748 spin_unlock(&memcg
->event_list_lock
);
4754 static void memcg_event_ptable_queue_proc(struct file
*file
,
4755 wait_queue_head_t
*wqh
, poll_table
*pt
)
4757 struct mem_cgroup_event
*event
=
4758 container_of(pt
, struct mem_cgroup_event
, pt
);
4761 add_wait_queue(wqh
, &event
->wait
);
4765 * DO NOT USE IN NEW FILES.
4767 * Parse input and register new cgroup event handler.
4769 * Input must be in format '<event_fd> <control_fd> <args>'.
4770 * Interpretation of args is defined by control file implementation.
4772 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4773 char *buf
, size_t nbytes
, loff_t off
)
4775 struct cgroup_subsys_state
*css
= of_css(of
);
4776 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4777 struct mem_cgroup_event
*event
;
4778 struct cgroup_subsys_state
*cfile_css
;
4779 unsigned int efd
, cfd
;
4786 buf
= strstrip(buf
);
4788 efd
= simple_strtoul(buf
, &endp
, 10);
4793 cfd
= simple_strtoul(buf
, &endp
, 10);
4794 if ((*endp
!= ' ') && (*endp
!= '\0'))
4798 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4802 event
->memcg
= memcg
;
4803 INIT_LIST_HEAD(&event
->list
);
4804 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4805 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4806 INIT_WORK(&event
->remove
, memcg_event_remove
);
4814 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4815 if (IS_ERR(event
->eventfd
)) {
4816 ret
= PTR_ERR(event
->eventfd
);
4823 goto out_put_eventfd
;
4826 /* the process need read permission on control file */
4827 /* AV: shouldn't we check that it's been opened for read instead? */
4828 ret
= file_permission(cfile
.file
, MAY_READ
);
4833 * Determine the event callbacks and set them in @event. This used
4834 * to be done via struct cftype but cgroup core no longer knows
4835 * about these events. The following is crude but the whole thing
4836 * is for compatibility anyway.
4838 * DO NOT ADD NEW FILES.
4840 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4842 if (!strcmp(name
, "memory.usage_in_bytes")) {
4843 event
->register_event
= mem_cgroup_usage_register_event
;
4844 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4845 } else if (!strcmp(name
, "memory.oom_control")) {
4846 event
->register_event
= mem_cgroup_oom_register_event
;
4847 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4848 } else if (!strcmp(name
, "memory.pressure_level")) {
4849 event
->register_event
= vmpressure_register_event
;
4850 event
->unregister_event
= vmpressure_unregister_event
;
4851 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4852 event
->register_event
= memsw_cgroup_usage_register_event
;
4853 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4860 * Verify @cfile should belong to @css. Also, remaining events are
4861 * automatically removed on cgroup destruction but the removal is
4862 * asynchronous, so take an extra ref on @css.
4864 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4865 &memory_cgrp_subsys
);
4867 if (IS_ERR(cfile_css
))
4869 if (cfile_css
!= css
) {
4874 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4878 vfs_poll(efile
.file
, &event
->pt
);
4880 spin_lock_irq(&memcg
->event_list_lock
);
4881 list_add(&event
->list
, &memcg
->event_list
);
4882 spin_unlock_irq(&memcg
->event_list_lock
);
4894 eventfd_ctx_put(event
->eventfd
);
4903 static struct cftype mem_cgroup_legacy_files
[] = {
4905 .name
= "usage_in_bytes",
4906 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4907 .read_u64
= mem_cgroup_read_u64
,
4910 .name
= "max_usage_in_bytes",
4911 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4912 .write
= mem_cgroup_reset
,
4913 .read_u64
= mem_cgroup_read_u64
,
4916 .name
= "limit_in_bytes",
4917 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4918 .write
= mem_cgroup_write
,
4919 .read_u64
= mem_cgroup_read_u64
,
4922 .name
= "soft_limit_in_bytes",
4923 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4924 .write
= mem_cgroup_write
,
4925 .read_u64
= mem_cgroup_read_u64
,
4929 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4930 .write
= mem_cgroup_reset
,
4931 .read_u64
= mem_cgroup_read_u64
,
4935 .seq_show
= memcg_stat_show
,
4938 .name
= "force_empty",
4939 .write
= mem_cgroup_force_empty_write
,
4942 .name
= "use_hierarchy",
4943 .write_u64
= mem_cgroup_hierarchy_write
,
4944 .read_u64
= mem_cgroup_hierarchy_read
,
4947 .name
= "cgroup.event_control", /* XXX: for compat */
4948 .write
= memcg_write_event_control
,
4949 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4952 .name
= "swappiness",
4953 .read_u64
= mem_cgroup_swappiness_read
,
4954 .write_u64
= mem_cgroup_swappiness_write
,
4957 .name
= "move_charge_at_immigrate",
4958 .read_u64
= mem_cgroup_move_charge_read
,
4959 .write_u64
= mem_cgroup_move_charge_write
,
4962 .name
= "oom_control",
4963 .seq_show
= mem_cgroup_oom_control_read
,
4964 .write_u64
= mem_cgroup_oom_control_write
,
4965 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4968 .name
= "pressure_level",
4972 .name
= "numa_stat",
4973 .seq_show
= memcg_numa_stat_show
,
4977 .name
= "kmem.limit_in_bytes",
4978 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4979 .write
= mem_cgroup_write
,
4980 .read_u64
= mem_cgroup_read_u64
,
4983 .name
= "kmem.usage_in_bytes",
4984 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4985 .read_u64
= mem_cgroup_read_u64
,
4988 .name
= "kmem.failcnt",
4989 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4990 .write
= mem_cgroup_reset
,
4991 .read_u64
= mem_cgroup_read_u64
,
4994 .name
= "kmem.max_usage_in_bytes",
4995 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4996 .write
= mem_cgroup_reset
,
4997 .read_u64
= mem_cgroup_read_u64
,
4999 #if defined(CONFIG_MEMCG_KMEM) && \
5000 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5002 .name
= "kmem.slabinfo",
5003 .seq_show
= memcg_slab_show
,
5007 .name
= "kmem.tcp.limit_in_bytes",
5008 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5009 .write
= mem_cgroup_write
,
5010 .read_u64
= mem_cgroup_read_u64
,
5013 .name
= "kmem.tcp.usage_in_bytes",
5014 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5015 .read_u64
= mem_cgroup_read_u64
,
5018 .name
= "kmem.tcp.failcnt",
5019 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5020 .write
= mem_cgroup_reset
,
5021 .read_u64
= mem_cgroup_read_u64
,
5024 .name
= "kmem.tcp.max_usage_in_bytes",
5025 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5026 .write
= mem_cgroup_reset
,
5027 .read_u64
= mem_cgroup_read_u64
,
5029 { }, /* terminate */
5033 * Private memory cgroup IDR
5035 * Swap-out records and page cache shadow entries need to store memcg
5036 * references in constrained space, so we maintain an ID space that is
5037 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5038 * memory-controlled cgroups to 64k.
5040 * However, there usually are many references to the offline CSS after
5041 * the cgroup has been destroyed, such as page cache or reclaimable
5042 * slab objects, that don't need to hang on to the ID. We want to keep
5043 * those dead CSS from occupying IDs, or we might quickly exhaust the
5044 * relatively small ID space and prevent the creation of new cgroups
5045 * even when there are much fewer than 64k cgroups - possibly none.
5047 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5048 * be freed and recycled when it's no longer needed, which is usually
5049 * when the CSS is offlined.
5051 * The only exception to that are records of swapped out tmpfs/shmem
5052 * pages that need to be attributed to live ancestors on swapin. But
5053 * those references are manageable from userspace.
5056 static DEFINE_IDR(mem_cgroup_idr
);
5058 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5060 if (memcg
->id
.id
> 0) {
5061 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5066 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5069 refcount_add(n
, &memcg
->id
.ref
);
5072 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5074 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5075 mem_cgroup_id_remove(memcg
);
5077 /* Memcg ID pins CSS */
5078 css_put(&memcg
->css
);
5082 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5084 mem_cgroup_id_put_many(memcg
, 1);
5088 * mem_cgroup_from_id - look up a memcg from a memcg id
5089 * @id: the memcg id to look up
5091 * Caller must hold rcu_read_lock().
5093 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5095 WARN_ON_ONCE(!rcu_read_lock_held());
5096 return idr_find(&mem_cgroup_idr
, id
);
5099 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5101 struct mem_cgroup_per_node
*pn
;
5104 * This routine is called against possible nodes.
5105 * But it's BUG to call kmalloc() against offline node.
5107 * TODO: this routine can waste much memory for nodes which will
5108 * never be onlined. It's better to use memory hotplug callback
5111 if (!node_state(node
, N_NORMAL_MEMORY
))
5113 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5117 pn
->lruvec_stats_percpu
= alloc_percpu_gfp(struct lruvec_stats_percpu
,
5118 GFP_KERNEL_ACCOUNT
);
5119 if (!pn
->lruvec_stats_percpu
) {
5124 lruvec_init(&pn
->lruvec
);
5125 pn
->usage_in_excess
= 0;
5126 pn
->on_tree
= false;
5129 memcg
->nodeinfo
[node
] = pn
;
5133 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5135 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5140 free_percpu(pn
->lruvec_stats_percpu
);
5144 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5149 free_mem_cgroup_per_node_info(memcg
, node
);
5150 free_percpu(memcg
->vmstats_percpu
);
5154 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5156 memcg_wb_domain_exit(memcg
);
5157 __mem_cgroup_free(memcg
);
5160 static struct mem_cgroup
*mem_cgroup_alloc(void)
5162 struct mem_cgroup
*memcg
;
5165 int __maybe_unused i
;
5166 long error
= -ENOMEM
;
5168 size
= sizeof(struct mem_cgroup
);
5169 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5171 memcg
= kzalloc(size
, GFP_KERNEL
);
5173 return ERR_PTR(error
);
5175 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5176 1, MEM_CGROUP_ID_MAX
,
5178 if (memcg
->id
.id
< 0) {
5179 error
= memcg
->id
.id
;
5183 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5184 GFP_KERNEL_ACCOUNT
);
5185 if (!memcg
->vmstats_percpu
)
5189 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5192 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5195 INIT_WORK(&memcg
->high_work
, high_work_func
);
5196 INIT_LIST_HEAD(&memcg
->oom_notify
);
5197 mutex_init(&memcg
->thresholds_lock
);
5198 spin_lock_init(&memcg
->move_lock
);
5199 vmpressure_init(&memcg
->vmpressure
);
5200 INIT_LIST_HEAD(&memcg
->event_list
);
5201 spin_lock_init(&memcg
->event_list_lock
);
5202 memcg
->socket_pressure
= jiffies
;
5203 #ifdef CONFIG_MEMCG_KMEM
5204 memcg
->kmemcg_id
= -1;
5205 INIT_LIST_HEAD(&memcg
->objcg_list
);
5207 #ifdef CONFIG_CGROUP_WRITEBACK
5208 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5209 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5210 memcg
->cgwb_frn
[i
].done
=
5211 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5213 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5214 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5215 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5216 memcg
->deferred_split_queue
.split_queue_len
= 0;
5218 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5221 mem_cgroup_id_remove(memcg
);
5222 __mem_cgroup_free(memcg
);
5223 return ERR_PTR(error
);
5226 static struct cgroup_subsys_state
* __ref
5227 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5229 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5230 struct mem_cgroup
*memcg
, *old_memcg
;
5231 long error
= -ENOMEM
;
5233 old_memcg
= set_active_memcg(parent
);
5234 memcg
= mem_cgroup_alloc();
5235 set_active_memcg(old_memcg
);
5237 return ERR_CAST(memcg
);
5239 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5240 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5241 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5243 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5244 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5246 page_counter_init(&memcg
->memory
, &parent
->memory
);
5247 page_counter_init(&memcg
->swap
, &parent
->swap
);
5248 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5249 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5251 page_counter_init(&memcg
->memory
, NULL
);
5252 page_counter_init(&memcg
->swap
, NULL
);
5253 page_counter_init(&memcg
->kmem
, NULL
);
5254 page_counter_init(&memcg
->tcpmem
, NULL
);
5256 root_mem_cgroup
= memcg
;
5260 /* The following stuff does not apply to the root */
5261 error
= memcg_online_kmem(memcg
);
5265 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5266 static_branch_inc(&memcg_sockets_enabled_key
);
5270 mem_cgroup_id_remove(memcg
);
5271 mem_cgroup_free(memcg
);
5272 return ERR_PTR(error
);
5275 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5277 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5280 * A memcg must be visible for expand_shrinker_info()
5281 * by the time the maps are allocated. So, we allocate maps
5282 * here, when for_each_mem_cgroup() can't skip it.
5284 if (alloc_shrinker_info(memcg
)) {
5285 mem_cgroup_id_remove(memcg
);
5289 /* Online state pins memcg ID, memcg ID pins CSS */
5290 refcount_set(&memcg
->id
.ref
, 1);
5293 if (unlikely(mem_cgroup_is_root(memcg
)))
5294 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
,
5299 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5301 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5302 struct mem_cgroup_event
*event
, *tmp
;
5305 * Unregister events and notify userspace.
5306 * Notify userspace about cgroup removing only after rmdir of cgroup
5307 * directory to avoid race between userspace and kernelspace.
5309 spin_lock_irq(&memcg
->event_list_lock
);
5310 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5311 list_del_init(&event
->list
);
5312 schedule_work(&event
->remove
);
5314 spin_unlock_irq(&memcg
->event_list_lock
);
5316 page_counter_set_min(&memcg
->memory
, 0);
5317 page_counter_set_low(&memcg
->memory
, 0);
5319 memcg_offline_kmem(memcg
);
5320 reparent_shrinker_deferred(memcg
);
5321 wb_memcg_offline(memcg
);
5323 drain_all_stock(memcg
);
5325 mem_cgroup_id_put(memcg
);
5328 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5330 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5332 invalidate_reclaim_iterators(memcg
);
5335 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5337 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5338 int __maybe_unused i
;
5340 #ifdef CONFIG_CGROUP_WRITEBACK
5341 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5342 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5344 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5345 static_branch_dec(&memcg_sockets_enabled_key
);
5347 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5348 static_branch_dec(&memcg_sockets_enabled_key
);
5350 vmpressure_cleanup(&memcg
->vmpressure
);
5351 cancel_work_sync(&memcg
->high_work
);
5352 mem_cgroup_remove_from_trees(memcg
);
5353 free_shrinker_info(memcg
);
5354 memcg_free_kmem(memcg
);
5355 mem_cgroup_free(memcg
);
5359 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5360 * @css: the target css
5362 * Reset the states of the mem_cgroup associated with @css. This is
5363 * invoked when the userland requests disabling on the default hierarchy
5364 * but the memcg is pinned through dependency. The memcg should stop
5365 * applying policies and should revert to the vanilla state as it may be
5366 * made visible again.
5368 * The current implementation only resets the essential configurations.
5369 * This needs to be expanded to cover all the visible parts.
5371 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5373 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5375 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5376 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5377 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5378 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5379 page_counter_set_min(&memcg
->memory
, 0);
5380 page_counter_set_low(&memcg
->memory
, 0);
5381 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5382 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5383 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5384 memcg_wb_domain_size_changed(memcg
);
5387 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
5389 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5390 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5391 struct memcg_vmstats_percpu
*statc
;
5395 statc
= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
);
5397 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
5399 * Collect the aggregated propagation counts of groups
5400 * below us. We're in a per-cpu loop here and this is
5401 * a global counter, so the first cycle will get them.
5403 delta
= memcg
->vmstats
.state_pending
[i
];
5405 memcg
->vmstats
.state_pending
[i
] = 0;
5407 /* Add CPU changes on this level since the last flush */
5408 v
= READ_ONCE(statc
->state
[i
]);
5409 if (v
!= statc
->state_prev
[i
]) {
5410 delta
+= v
- statc
->state_prev
[i
];
5411 statc
->state_prev
[i
] = v
;
5417 /* Aggregate counts on this level and propagate upwards */
5418 memcg
->vmstats
.state
[i
] += delta
;
5420 parent
->vmstats
.state_pending
[i
] += delta
;
5423 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
5424 delta
= memcg
->vmstats
.events_pending
[i
];
5426 memcg
->vmstats
.events_pending
[i
] = 0;
5428 v
= READ_ONCE(statc
->events
[i
]);
5429 if (v
!= statc
->events_prev
[i
]) {
5430 delta
+= v
- statc
->events_prev
[i
];
5431 statc
->events_prev
[i
] = v
;
5437 memcg
->vmstats
.events
[i
] += delta
;
5439 parent
->vmstats
.events_pending
[i
] += delta
;
5442 for_each_node_state(nid
, N_MEMORY
) {
5443 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[nid
];
5444 struct mem_cgroup_per_node
*ppn
= NULL
;
5445 struct lruvec_stats_percpu
*lstatc
;
5448 ppn
= parent
->nodeinfo
[nid
];
5450 lstatc
= per_cpu_ptr(pn
->lruvec_stats_percpu
, cpu
);
5452 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
5453 delta
= pn
->lruvec_stats
.state_pending
[i
];
5455 pn
->lruvec_stats
.state_pending
[i
] = 0;
5457 v
= READ_ONCE(lstatc
->state
[i
]);
5458 if (v
!= lstatc
->state_prev
[i
]) {
5459 delta
+= v
- lstatc
->state_prev
[i
];
5460 lstatc
->state_prev
[i
] = v
;
5466 pn
->lruvec_stats
.state
[i
] += delta
;
5468 ppn
->lruvec_stats
.state_pending
[i
] += delta
;
5474 /* Handlers for move charge at task migration. */
5475 static int mem_cgroup_do_precharge(unsigned long count
)
5479 /* Try a single bulk charge without reclaim first, kswapd may wake */
5480 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5482 mc
.precharge
+= count
;
5486 /* Try charges one by one with reclaim, but do not retry */
5488 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5502 enum mc_target_type
{
5509 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5510 unsigned long addr
, pte_t ptent
)
5512 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5514 if (!page
|| !page_mapped(page
))
5516 if (PageAnon(page
)) {
5517 if (!(mc
.flags
& MOVE_ANON
))
5520 if (!(mc
.flags
& MOVE_FILE
))
5523 if (!get_page_unless_zero(page
))
5529 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5530 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5531 pte_t ptent
, swp_entry_t
*entry
)
5533 struct page
*page
= NULL
;
5534 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5536 if (!(mc
.flags
& MOVE_ANON
))
5540 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5541 * a device and because they are not accessible by CPU they are store
5542 * as special swap entry in the CPU page table.
5544 if (is_device_private_entry(ent
)) {
5545 page
= pfn_swap_entry_to_page(ent
);
5547 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5548 * a refcount of 1 when free (unlike normal page)
5550 if (!page_ref_add_unless(page
, 1, 1))
5555 if (non_swap_entry(ent
))
5559 * Because lookup_swap_cache() updates some statistics counter,
5560 * we call find_get_page() with swapper_space directly.
5562 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5563 entry
->val
= ent
.val
;
5568 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5569 pte_t ptent
, swp_entry_t
*entry
)
5575 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5576 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5578 if (!vma
->vm_file
) /* anonymous vma */
5580 if (!(mc
.flags
& MOVE_FILE
))
5583 /* page is moved even if it's not RSS of this task(page-faulted). */
5584 /* shmem/tmpfs may report page out on swap: account for that too. */
5585 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5586 linear_page_index(vma
, addr
));
5590 * mem_cgroup_move_account - move account of the page
5592 * @compound: charge the page as compound or small page
5593 * @from: mem_cgroup which the page is moved from.
5594 * @to: mem_cgroup which the page is moved to. @from != @to.
5596 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5598 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5601 static int mem_cgroup_move_account(struct page
*page
,
5603 struct mem_cgroup
*from
,
5604 struct mem_cgroup
*to
)
5606 struct lruvec
*from_vec
, *to_vec
;
5607 struct pglist_data
*pgdat
;
5608 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5611 VM_BUG_ON(from
== to
);
5612 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5613 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5616 * Prevent mem_cgroup_migrate() from looking at
5617 * page's memory cgroup of its source page while we change it.
5620 if (!trylock_page(page
))
5624 if (page_memcg(page
) != from
)
5627 pgdat
= page_pgdat(page
);
5628 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5629 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5631 lock_page_memcg(page
);
5633 if (PageAnon(page
)) {
5634 if (page_mapped(page
)) {
5635 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5636 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5637 if (PageTransHuge(page
)) {
5638 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5640 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5645 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5646 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5648 if (PageSwapBacked(page
)) {
5649 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5650 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5653 if (page_mapped(page
)) {
5654 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5655 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5658 if (PageDirty(page
)) {
5659 struct address_space
*mapping
= page_mapping(page
);
5661 if (mapping_can_writeback(mapping
)) {
5662 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5664 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5670 if (PageWriteback(page
)) {
5671 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5672 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5676 * All state has been migrated, let's switch to the new memcg.
5678 * It is safe to change page's memcg here because the page
5679 * is referenced, charged, isolated, and locked: we can't race
5680 * with (un)charging, migration, LRU putback, or anything else
5681 * that would rely on a stable page's memory cgroup.
5683 * Note that lock_page_memcg is a memcg lock, not a page lock,
5684 * to save space. As soon as we switch page's memory cgroup to a
5685 * new memcg that isn't locked, the above state can change
5686 * concurrently again. Make sure we're truly done with it.
5691 css_put(&from
->css
);
5693 page
->memcg_data
= (unsigned long)to
;
5695 __unlock_page_memcg(from
);
5699 local_irq_disable();
5700 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5701 memcg_check_events(to
, page
);
5702 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5703 memcg_check_events(from
, page
);
5712 * get_mctgt_type - get target type of moving charge
5713 * @vma: the vma the pte to be checked belongs
5714 * @addr: the address corresponding to the pte to be checked
5715 * @ptent: the pte to be checked
5716 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5719 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5720 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5721 * move charge. if @target is not NULL, the page is stored in target->page
5722 * with extra refcnt got(Callers should handle it).
5723 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5724 * target for charge migration. if @target is not NULL, the entry is stored
5726 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5727 * (so ZONE_DEVICE page and thus not on the lru).
5728 * For now we such page is charge like a regular page would be as for all
5729 * intent and purposes it is just special memory taking the place of a
5732 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5734 * Called with pte lock held.
5737 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5738 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5740 struct page
*page
= NULL
;
5741 enum mc_target_type ret
= MC_TARGET_NONE
;
5742 swp_entry_t ent
= { .val
= 0 };
5744 if (pte_present(ptent
))
5745 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5746 else if (is_swap_pte(ptent
))
5747 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5748 else if (pte_none(ptent
))
5749 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5751 if (!page
&& !ent
.val
)
5755 * Do only loose check w/o serialization.
5756 * mem_cgroup_move_account() checks the page is valid or
5757 * not under LRU exclusion.
5759 if (page_memcg(page
) == mc
.from
) {
5760 ret
= MC_TARGET_PAGE
;
5761 if (is_device_private_page(page
))
5762 ret
= MC_TARGET_DEVICE
;
5764 target
->page
= page
;
5766 if (!ret
|| !target
)
5770 * There is a swap entry and a page doesn't exist or isn't charged.
5771 * But we cannot move a tail-page in a THP.
5773 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5774 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5775 ret
= MC_TARGET_SWAP
;
5782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5784 * We don't consider PMD mapped swapping or file mapped pages because THP does
5785 * not support them for now.
5786 * Caller should make sure that pmd_trans_huge(pmd) is true.
5788 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5789 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5791 struct page
*page
= NULL
;
5792 enum mc_target_type ret
= MC_TARGET_NONE
;
5794 if (unlikely(is_swap_pmd(pmd
))) {
5795 VM_BUG_ON(thp_migration_supported() &&
5796 !is_pmd_migration_entry(pmd
));
5799 page
= pmd_page(pmd
);
5800 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5801 if (!(mc
.flags
& MOVE_ANON
))
5803 if (page_memcg(page
) == mc
.from
) {
5804 ret
= MC_TARGET_PAGE
;
5807 target
->page
= page
;
5813 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5814 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5816 return MC_TARGET_NONE
;
5820 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5821 unsigned long addr
, unsigned long end
,
5822 struct mm_walk
*walk
)
5824 struct vm_area_struct
*vma
= walk
->vma
;
5828 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5831 * Note their can not be MC_TARGET_DEVICE for now as we do not
5832 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5833 * this might change.
5835 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5836 mc
.precharge
+= HPAGE_PMD_NR
;
5841 if (pmd_trans_unstable(pmd
))
5843 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5844 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5845 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5846 mc
.precharge
++; /* increment precharge temporarily */
5847 pte_unmap_unlock(pte
- 1, ptl
);
5853 static const struct mm_walk_ops precharge_walk_ops
= {
5854 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5857 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5859 unsigned long precharge
;
5862 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5863 mmap_read_unlock(mm
);
5865 precharge
= mc
.precharge
;
5871 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5873 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5875 VM_BUG_ON(mc
.moving_task
);
5876 mc
.moving_task
= current
;
5877 return mem_cgroup_do_precharge(precharge
);
5880 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5881 static void __mem_cgroup_clear_mc(void)
5883 struct mem_cgroup
*from
= mc
.from
;
5884 struct mem_cgroup
*to
= mc
.to
;
5886 /* we must uncharge all the leftover precharges from mc.to */
5888 cancel_charge(mc
.to
, mc
.precharge
);
5892 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5893 * we must uncharge here.
5895 if (mc
.moved_charge
) {
5896 cancel_charge(mc
.from
, mc
.moved_charge
);
5897 mc
.moved_charge
= 0;
5899 /* we must fixup refcnts and charges */
5900 if (mc
.moved_swap
) {
5901 /* uncharge swap account from the old cgroup */
5902 if (!mem_cgroup_is_root(mc
.from
))
5903 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5905 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5908 * we charged both to->memory and to->memsw, so we
5909 * should uncharge to->memory.
5911 if (!mem_cgroup_is_root(mc
.to
))
5912 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5916 memcg_oom_recover(from
);
5917 memcg_oom_recover(to
);
5918 wake_up_all(&mc
.waitq
);
5921 static void mem_cgroup_clear_mc(void)
5923 struct mm_struct
*mm
= mc
.mm
;
5926 * we must clear moving_task before waking up waiters at the end of
5929 mc
.moving_task
= NULL
;
5930 __mem_cgroup_clear_mc();
5931 spin_lock(&mc
.lock
);
5935 spin_unlock(&mc
.lock
);
5940 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5942 struct cgroup_subsys_state
*css
;
5943 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5944 struct mem_cgroup
*from
;
5945 struct task_struct
*leader
, *p
;
5946 struct mm_struct
*mm
;
5947 unsigned long move_flags
;
5950 /* charge immigration isn't supported on the default hierarchy */
5951 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5955 * Multi-process migrations only happen on the default hierarchy
5956 * where charge immigration is not used. Perform charge
5957 * immigration if @tset contains a leader and whine if there are
5961 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5964 memcg
= mem_cgroup_from_css(css
);
5970 * We are now committed to this value whatever it is. Changes in this
5971 * tunable will only affect upcoming migrations, not the current one.
5972 * So we need to save it, and keep it going.
5974 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5978 from
= mem_cgroup_from_task(p
);
5980 VM_BUG_ON(from
== memcg
);
5982 mm
= get_task_mm(p
);
5985 /* We move charges only when we move a owner of the mm */
5986 if (mm
->owner
== p
) {
5989 VM_BUG_ON(mc
.precharge
);
5990 VM_BUG_ON(mc
.moved_charge
);
5991 VM_BUG_ON(mc
.moved_swap
);
5993 spin_lock(&mc
.lock
);
5997 mc
.flags
= move_flags
;
5998 spin_unlock(&mc
.lock
);
5999 /* We set mc.moving_task later */
6001 ret
= mem_cgroup_precharge_mc(mm
);
6003 mem_cgroup_clear_mc();
6010 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6013 mem_cgroup_clear_mc();
6016 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6017 unsigned long addr
, unsigned long end
,
6018 struct mm_walk
*walk
)
6021 struct vm_area_struct
*vma
= walk
->vma
;
6024 enum mc_target_type target_type
;
6025 union mc_target target
;
6028 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6030 if (mc
.precharge
< HPAGE_PMD_NR
) {
6034 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6035 if (target_type
== MC_TARGET_PAGE
) {
6037 if (!isolate_lru_page(page
)) {
6038 if (!mem_cgroup_move_account(page
, true,
6040 mc
.precharge
-= HPAGE_PMD_NR
;
6041 mc
.moved_charge
+= HPAGE_PMD_NR
;
6043 putback_lru_page(page
);
6046 } else if (target_type
== MC_TARGET_DEVICE
) {
6048 if (!mem_cgroup_move_account(page
, true,
6050 mc
.precharge
-= HPAGE_PMD_NR
;
6051 mc
.moved_charge
+= HPAGE_PMD_NR
;
6059 if (pmd_trans_unstable(pmd
))
6062 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6063 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6064 pte_t ptent
= *(pte
++);
6065 bool device
= false;
6071 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6072 case MC_TARGET_DEVICE
:
6075 case MC_TARGET_PAGE
:
6078 * We can have a part of the split pmd here. Moving it
6079 * can be done but it would be too convoluted so simply
6080 * ignore such a partial THP and keep it in original
6081 * memcg. There should be somebody mapping the head.
6083 if (PageTransCompound(page
))
6085 if (!device
&& isolate_lru_page(page
))
6087 if (!mem_cgroup_move_account(page
, false,
6090 /* we uncharge from mc.from later. */
6094 putback_lru_page(page
);
6095 put
: /* get_mctgt_type() gets the page */
6098 case MC_TARGET_SWAP
:
6100 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6102 mem_cgroup_id_get_many(mc
.to
, 1);
6103 /* we fixup other refcnts and charges later. */
6111 pte_unmap_unlock(pte
- 1, ptl
);
6116 * We have consumed all precharges we got in can_attach().
6117 * We try charge one by one, but don't do any additional
6118 * charges to mc.to if we have failed in charge once in attach()
6121 ret
= mem_cgroup_do_precharge(1);
6129 static const struct mm_walk_ops charge_walk_ops
= {
6130 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6133 static void mem_cgroup_move_charge(void)
6135 lru_add_drain_all();
6137 * Signal lock_page_memcg() to take the memcg's move_lock
6138 * while we're moving its pages to another memcg. Then wait
6139 * for already started RCU-only updates to finish.
6141 atomic_inc(&mc
.from
->moving_account
);
6144 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6146 * Someone who are holding the mmap_lock might be waiting in
6147 * waitq. So we cancel all extra charges, wake up all waiters,
6148 * and retry. Because we cancel precharges, we might not be able
6149 * to move enough charges, but moving charge is a best-effort
6150 * feature anyway, so it wouldn't be a big problem.
6152 __mem_cgroup_clear_mc();
6157 * When we have consumed all precharges and failed in doing
6158 * additional charge, the page walk just aborts.
6160 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6163 mmap_read_unlock(mc
.mm
);
6164 atomic_dec(&mc
.from
->moving_account
);
6167 static void mem_cgroup_move_task(void)
6170 mem_cgroup_move_charge();
6171 mem_cgroup_clear_mc();
6174 #else /* !CONFIG_MMU */
6175 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6179 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6182 static void mem_cgroup_move_task(void)
6187 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6189 if (value
== PAGE_COUNTER_MAX
)
6190 seq_puts(m
, "max\n");
6192 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6197 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6200 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6202 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6205 static int memory_min_show(struct seq_file
*m
, void *v
)
6207 return seq_puts_memcg_tunable(m
,
6208 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6211 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6212 char *buf
, size_t nbytes
, loff_t off
)
6214 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6218 buf
= strstrip(buf
);
6219 err
= page_counter_memparse(buf
, "max", &min
);
6223 page_counter_set_min(&memcg
->memory
, min
);
6228 static int memory_low_show(struct seq_file
*m
, void *v
)
6230 return seq_puts_memcg_tunable(m
,
6231 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6234 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6235 char *buf
, size_t nbytes
, loff_t off
)
6237 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6241 buf
= strstrip(buf
);
6242 err
= page_counter_memparse(buf
, "max", &low
);
6246 page_counter_set_low(&memcg
->memory
, low
);
6251 static int memory_high_show(struct seq_file
*m
, void *v
)
6253 return seq_puts_memcg_tunable(m
,
6254 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6257 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6258 char *buf
, size_t nbytes
, loff_t off
)
6260 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6261 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6262 bool drained
= false;
6266 buf
= strstrip(buf
);
6267 err
= page_counter_memparse(buf
, "max", &high
);
6271 page_counter_set_high(&memcg
->memory
, high
);
6274 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6275 unsigned long reclaimed
;
6277 if (nr_pages
<= high
)
6280 if (signal_pending(current
))
6284 drain_all_stock(memcg
);
6289 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6292 if (!reclaimed
&& !nr_retries
--)
6296 memcg_wb_domain_size_changed(memcg
);
6300 static int memory_max_show(struct seq_file
*m
, void *v
)
6302 return seq_puts_memcg_tunable(m
,
6303 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6306 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6307 char *buf
, size_t nbytes
, loff_t off
)
6309 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6310 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6311 bool drained
= false;
6315 buf
= strstrip(buf
);
6316 err
= page_counter_memparse(buf
, "max", &max
);
6320 xchg(&memcg
->memory
.max
, max
);
6323 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6325 if (nr_pages
<= max
)
6328 if (signal_pending(current
))
6332 drain_all_stock(memcg
);
6338 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6344 memcg_memory_event(memcg
, MEMCG_OOM
);
6345 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6349 memcg_wb_domain_size_changed(memcg
);
6353 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6355 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6356 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6357 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6358 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6359 seq_printf(m
, "oom_kill %lu\n",
6360 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6363 static int memory_events_show(struct seq_file
*m
, void *v
)
6365 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6367 __memory_events_show(m
, memcg
->memory_events
);
6371 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6373 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6375 __memory_events_show(m
, memcg
->memory_events_local
);
6379 static int memory_stat_show(struct seq_file
*m
, void *v
)
6381 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6384 buf
= memory_stat_format(memcg
);
6393 static inline unsigned long lruvec_page_state_output(struct lruvec
*lruvec
,
6396 return lruvec_page_state(lruvec
, item
) * memcg_page_state_unit(item
);
6399 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6402 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6404 mem_cgroup_flush_stats();
6406 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6409 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6412 seq_printf(m
, "%s", memory_stats
[i
].name
);
6413 for_each_node_state(nid
, N_MEMORY
) {
6415 struct lruvec
*lruvec
;
6417 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6418 size
= lruvec_page_state_output(lruvec
,
6419 memory_stats
[i
].idx
);
6420 seq_printf(m
, " N%d=%llu", nid
, size
);
6429 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6431 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6433 seq_printf(m
, "%d\n", memcg
->oom_group
);
6438 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6439 char *buf
, size_t nbytes
, loff_t off
)
6441 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6444 buf
= strstrip(buf
);
6448 ret
= kstrtoint(buf
, 0, &oom_group
);
6452 if (oom_group
!= 0 && oom_group
!= 1)
6455 memcg
->oom_group
= oom_group
;
6460 static struct cftype memory_files
[] = {
6463 .flags
= CFTYPE_NOT_ON_ROOT
,
6464 .read_u64
= memory_current_read
,
6468 .flags
= CFTYPE_NOT_ON_ROOT
,
6469 .seq_show
= memory_min_show
,
6470 .write
= memory_min_write
,
6474 .flags
= CFTYPE_NOT_ON_ROOT
,
6475 .seq_show
= memory_low_show
,
6476 .write
= memory_low_write
,
6480 .flags
= CFTYPE_NOT_ON_ROOT
,
6481 .seq_show
= memory_high_show
,
6482 .write
= memory_high_write
,
6486 .flags
= CFTYPE_NOT_ON_ROOT
,
6487 .seq_show
= memory_max_show
,
6488 .write
= memory_max_write
,
6492 .flags
= CFTYPE_NOT_ON_ROOT
,
6493 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6494 .seq_show
= memory_events_show
,
6497 .name
= "events.local",
6498 .flags
= CFTYPE_NOT_ON_ROOT
,
6499 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6500 .seq_show
= memory_events_local_show
,
6504 .seq_show
= memory_stat_show
,
6508 .name
= "numa_stat",
6509 .seq_show
= memory_numa_stat_show
,
6513 .name
= "oom.group",
6514 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6515 .seq_show
= memory_oom_group_show
,
6516 .write
= memory_oom_group_write
,
6521 struct cgroup_subsys memory_cgrp_subsys
= {
6522 .css_alloc
= mem_cgroup_css_alloc
,
6523 .css_online
= mem_cgroup_css_online
,
6524 .css_offline
= mem_cgroup_css_offline
,
6525 .css_released
= mem_cgroup_css_released
,
6526 .css_free
= mem_cgroup_css_free
,
6527 .css_reset
= mem_cgroup_css_reset
,
6528 .css_rstat_flush
= mem_cgroup_css_rstat_flush
,
6529 .can_attach
= mem_cgroup_can_attach
,
6530 .cancel_attach
= mem_cgroup_cancel_attach
,
6531 .post_attach
= mem_cgroup_move_task
,
6532 .dfl_cftypes
= memory_files
,
6533 .legacy_cftypes
= mem_cgroup_legacy_files
,
6538 * This function calculates an individual cgroup's effective
6539 * protection which is derived from its own memory.min/low, its
6540 * parent's and siblings' settings, as well as the actual memory
6541 * distribution in the tree.
6543 * The following rules apply to the effective protection values:
6545 * 1. At the first level of reclaim, effective protection is equal to
6546 * the declared protection in memory.min and memory.low.
6548 * 2. To enable safe delegation of the protection configuration, at
6549 * subsequent levels the effective protection is capped to the
6550 * parent's effective protection.
6552 * 3. To make complex and dynamic subtrees easier to configure, the
6553 * user is allowed to overcommit the declared protection at a given
6554 * level. If that is the case, the parent's effective protection is
6555 * distributed to the children in proportion to how much protection
6556 * they have declared and how much of it they are utilizing.
6558 * This makes distribution proportional, but also work-conserving:
6559 * if one cgroup claims much more protection than it uses memory,
6560 * the unused remainder is available to its siblings.
6562 * 4. Conversely, when the declared protection is undercommitted at a
6563 * given level, the distribution of the larger parental protection
6564 * budget is NOT proportional. A cgroup's protection from a sibling
6565 * is capped to its own memory.min/low setting.
6567 * 5. However, to allow protecting recursive subtrees from each other
6568 * without having to declare each individual cgroup's fixed share
6569 * of the ancestor's claim to protection, any unutilized -
6570 * "floating" - protection from up the tree is distributed in
6571 * proportion to each cgroup's *usage*. This makes the protection
6572 * neutral wrt sibling cgroups and lets them compete freely over
6573 * the shared parental protection budget, but it protects the
6574 * subtree as a whole from neighboring subtrees.
6576 * Note that 4. and 5. are not in conflict: 4. is about protecting
6577 * against immediate siblings whereas 5. is about protecting against
6578 * neighboring subtrees.
6580 static unsigned long effective_protection(unsigned long usage
,
6581 unsigned long parent_usage
,
6582 unsigned long setting
,
6583 unsigned long parent_effective
,
6584 unsigned long siblings_protected
)
6586 unsigned long protected;
6589 protected = min(usage
, setting
);
6591 * If all cgroups at this level combined claim and use more
6592 * protection then what the parent affords them, distribute
6593 * shares in proportion to utilization.
6595 * We are using actual utilization rather than the statically
6596 * claimed protection in order to be work-conserving: claimed
6597 * but unused protection is available to siblings that would
6598 * otherwise get a smaller chunk than what they claimed.
6600 if (siblings_protected
> parent_effective
)
6601 return protected * parent_effective
/ siblings_protected
;
6604 * Ok, utilized protection of all children is within what the
6605 * parent affords them, so we know whatever this child claims
6606 * and utilizes is effectively protected.
6608 * If there is unprotected usage beyond this value, reclaim
6609 * will apply pressure in proportion to that amount.
6611 * If there is unutilized protection, the cgroup will be fully
6612 * shielded from reclaim, but we do return a smaller value for
6613 * protection than what the group could enjoy in theory. This
6614 * is okay. With the overcommit distribution above, effective
6615 * protection is always dependent on how memory is actually
6616 * consumed among the siblings anyway.
6621 * If the children aren't claiming (all of) the protection
6622 * afforded to them by the parent, distribute the remainder in
6623 * proportion to the (unprotected) memory of each cgroup. That
6624 * way, cgroups that aren't explicitly prioritized wrt each
6625 * other compete freely over the allowance, but they are
6626 * collectively protected from neighboring trees.
6628 * We're using unprotected memory for the weight so that if
6629 * some cgroups DO claim explicit protection, we don't protect
6630 * the same bytes twice.
6632 * Check both usage and parent_usage against the respective
6633 * protected values. One should imply the other, but they
6634 * aren't read atomically - make sure the division is sane.
6636 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6638 if (parent_effective
> siblings_protected
&&
6639 parent_usage
> siblings_protected
&&
6640 usage
> protected) {
6641 unsigned long unclaimed
;
6643 unclaimed
= parent_effective
- siblings_protected
;
6644 unclaimed
*= usage
- protected;
6645 unclaimed
/= parent_usage
- siblings_protected
;
6654 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6655 * @root: the top ancestor of the sub-tree being checked
6656 * @memcg: the memory cgroup to check
6658 * WARNING: This function is not stateless! It can only be used as part
6659 * of a top-down tree iteration, not for isolated queries.
6661 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6662 struct mem_cgroup
*memcg
)
6664 unsigned long usage
, parent_usage
;
6665 struct mem_cgroup
*parent
;
6667 if (mem_cgroup_disabled())
6671 root
= root_mem_cgroup
;
6674 * Effective values of the reclaim targets are ignored so they
6675 * can be stale. Have a look at mem_cgroup_protection for more
6677 * TODO: calculation should be more robust so that we do not need
6678 * that special casing.
6683 usage
= page_counter_read(&memcg
->memory
);
6687 parent
= parent_mem_cgroup(memcg
);
6688 /* No parent means a non-hierarchical mode on v1 memcg */
6692 if (parent
== root
) {
6693 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6694 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6698 parent_usage
= page_counter_read(&parent
->memory
);
6700 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6701 READ_ONCE(memcg
->memory
.min
),
6702 READ_ONCE(parent
->memory
.emin
),
6703 atomic_long_read(&parent
->memory
.children_min_usage
)));
6705 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6706 READ_ONCE(memcg
->memory
.low
),
6707 READ_ONCE(parent
->memory
.elow
),
6708 atomic_long_read(&parent
->memory
.children_low_usage
)));
6711 static int charge_memcg(struct page
*page
, struct mem_cgroup
*memcg
, gfp_t gfp
)
6713 unsigned int nr_pages
= thp_nr_pages(page
);
6716 ret
= try_charge(memcg
, gfp
, nr_pages
);
6720 css_get(&memcg
->css
);
6721 commit_charge(page
, memcg
);
6723 local_irq_disable();
6724 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6725 memcg_check_events(memcg
, page
);
6732 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6733 * @page: page to charge
6734 * @mm: mm context of the victim
6735 * @gfp_mask: reclaim mode
6737 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6738 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6739 * charge to the active memcg.
6741 * Do not use this for pages allocated for swapin.
6743 * Returns 0 on success. Otherwise, an error code is returned.
6745 int __mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
6748 struct mem_cgroup
*memcg
;
6751 memcg
= get_mem_cgroup_from_mm(mm
);
6752 ret
= charge_memcg(page
, memcg
, gfp_mask
);
6753 css_put(&memcg
->css
);
6759 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6760 * @page: page to charge
6761 * @mm: mm context of the victim
6762 * @gfp: reclaim mode
6763 * @entry: swap entry for which the page is allocated
6765 * This function charges a page allocated for swapin. Please call this before
6766 * adding the page to the swapcache.
6768 * Returns 0 on success. Otherwise, an error code is returned.
6770 int mem_cgroup_swapin_charge_page(struct page
*page
, struct mm_struct
*mm
,
6771 gfp_t gfp
, swp_entry_t entry
)
6773 struct mem_cgroup
*memcg
;
6777 if (mem_cgroup_disabled())
6780 id
= lookup_swap_cgroup_id(entry
);
6782 memcg
= mem_cgroup_from_id(id
);
6783 if (!memcg
|| !css_tryget_online(&memcg
->css
))
6784 memcg
= get_mem_cgroup_from_mm(mm
);
6787 ret
= charge_memcg(page
, memcg
, gfp
);
6789 css_put(&memcg
->css
);
6794 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6795 * @entry: swap entry for which the page is charged
6797 * Call this function after successfully adding the charged page to swapcache.
6799 * Note: This function assumes the page for which swap slot is being uncharged
6802 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry
)
6805 * Cgroup1's unified memory+swap counter has been charged with the
6806 * new swapcache page, finish the transfer by uncharging the swap
6807 * slot. The swap slot would also get uncharged when it dies, but
6808 * it can stick around indefinitely and we'd count the page twice
6811 * Cgroup2 has separate resource counters for memory and swap,
6812 * so this is a non-issue here. Memory and swap charge lifetimes
6813 * correspond 1:1 to page and swap slot lifetimes: we charge the
6814 * page to memory here, and uncharge swap when the slot is freed.
6816 if (!mem_cgroup_disabled() && do_memsw_account()) {
6818 * The swap entry might not get freed for a long time,
6819 * let's not wait for it. The page already received a
6820 * memory+swap charge, drop the swap entry duplicate.
6822 mem_cgroup_uncharge_swap(entry
, 1);
6826 struct uncharge_gather
{
6827 struct mem_cgroup
*memcg
;
6828 unsigned long nr_memory
;
6829 unsigned long pgpgout
;
6830 unsigned long nr_kmem
;
6831 struct page
*dummy_page
;
6834 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6836 memset(ug
, 0, sizeof(*ug
));
6839 static void uncharge_batch(const struct uncharge_gather
*ug
)
6841 unsigned long flags
;
6843 if (ug
->nr_memory
) {
6844 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_memory
);
6845 if (do_memsw_account())
6846 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_memory
);
6847 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6848 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6849 memcg_oom_recover(ug
->memcg
);
6852 local_irq_save(flags
);
6853 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6854 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_memory
);
6855 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6856 local_irq_restore(flags
);
6858 /* drop reference from uncharge_page */
6859 css_put(&ug
->memcg
->css
);
6862 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6864 unsigned long nr_pages
;
6865 struct mem_cgroup
*memcg
;
6866 struct obj_cgroup
*objcg
;
6867 bool use_objcg
= PageMemcgKmem(page
);
6869 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6872 * Nobody should be changing or seriously looking at
6873 * page memcg or objcg at this point, we have fully
6874 * exclusive access to the page.
6877 objcg
= __page_objcg(page
);
6879 * This get matches the put at the end of the function and
6880 * kmem pages do not hold memcg references anymore.
6882 memcg
= get_mem_cgroup_from_objcg(objcg
);
6884 memcg
= __page_memcg(page
);
6890 if (ug
->memcg
!= memcg
) {
6893 uncharge_gather_clear(ug
);
6896 ug
->dummy_page
= page
;
6898 /* pairs with css_put in uncharge_batch */
6899 css_get(&memcg
->css
);
6902 nr_pages
= compound_nr(page
);
6905 ug
->nr_memory
+= nr_pages
;
6906 ug
->nr_kmem
+= nr_pages
;
6908 page
->memcg_data
= 0;
6909 obj_cgroup_put(objcg
);
6911 /* LRU pages aren't accounted at the root level */
6912 if (!mem_cgroup_is_root(memcg
))
6913 ug
->nr_memory
+= nr_pages
;
6916 page
->memcg_data
= 0;
6919 css_put(&memcg
->css
);
6923 * __mem_cgroup_uncharge - uncharge a page
6924 * @page: page to uncharge
6926 * Uncharge a page previously charged with __mem_cgroup_charge().
6928 void __mem_cgroup_uncharge(struct page
*page
)
6930 struct uncharge_gather ug
;
6932 /* Don't touch page->lru of any random page, pre-check: */
6933 if (!page_memcg(page
))
6936 uncharge_gather_clear(&ug
);
6937 uncharge_page(page
, &ug
);
6938 uncharge_batch(&ug
);
6942 * __mem_cgroup_uncharge_list - uncharge a list of page
6943 * @page_list: list of pages to uncharge
6945 * Uncharge a list of pages previously charged with
6946 * __mem_cgroup_charge().
6948 void __mem_cgroup_uncharge_list(struct list_head
*page_list
)
6950 struct uncharge_gather ug
;
6953 uncharge_gather_clear(&ug
);
6954 list_for_each_entry(page
, page_list
, lru
)
6955 uncharge_page(page
, &ug
);
6957 uncharge_batch(&ug
);
6961 * mem_cgroup_migrate - charge a page's replacement
6962 * @oldpage: currently circulating page
6963 * @newpage: replacement page
6965 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6966 * be uncharged upon free.
6968 * Both pages must be locked, @newpage->mapping must be set up.
6970 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6972 struct mem_cgroup
*memcg
;
6973 unsigned int nr_pages
;
6974 unsigned long flags
;
6976 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6977 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6978 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6979 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6982 if (mem_cgroup_disabled())
6985 /* Page cache replacement: new page already charged? */
6986 if (page_memcg(newpage
))
6989 memcg
= page_memcg(oldpage
);
6990 VM_WARN_ON_ONCE_PAGE(!memcg
, oldpage
);
6994 /* Force-charge the new page. The old one will be freed soon */
6995 nr_pages
= thp_nr_pages(newpage
);
6997 if (!mem_cgroup_is_root(memcg
)) {
6998 page_counter_charge(&memcg
->memory
, nr_pages
);
6999 if (do_memsw_account())
7000 page_counter_charge(&memcg
->memsw
, nr_pages
);
7003 css_get(&memcg
->css
);
7004 commit_charge(newpage
, memcg
);
7006 local_irq_save(flags
);
7007 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
7008 memcg_check_events(memcg
, newpage
);
7009 local_irq_restore(flags
);
7012 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
7013 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
7015 void mem_cgroup_sk_alloc(struct sock
*sk
)
7017 struct mem_cgroup
*memcg
;
7019 if (!mem_cgroup_sockets_enabled
)
7022 /* Do not associate the sock with unrelated interrupted task's memcg. */
7027 memcg
= mem_cgroup_from_task(current
);
7028 if (memcg
== root_mem_cgroup
)
7030 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
7032 if (css_tryget(&memcg
->css
))
7033 sk
->sk_memcg
= memcg
;
7038 void mem_cgroup_sk_free(struct sock
*sk
)
7041 css_put(&sk
->sk_memcg
->css
);
7045 * mem_cgroup_charge_skmem - charge socket memory
7046 * @memcg: memcg to charge
7047 * @nr_pages: number of pages to charge
7048 * @gfp_mask: reclaim mode
7050 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7051 * @memcg's configured limit, %false if it doesn't.
7053 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
,
7056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7057 struct page_counter
*fail
;
7059 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7060 memcg
->tcpmem_pressure
= 0;
7063 memcg
->tcpmem_pressure
= 1;
7064 if (gfp_mask
& __GFP_NOFAIL
) {
7065 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7071 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0) {
7072 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7080 * mem_cgroup_uncharge_skmem - uncharge socket memory
7081 * @memcg: memcg to uncharge
7082 * @nr_pages: number of pages to uncharge
7084 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7086 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7087 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7091 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7093 refill_stock(memcg
, nr_pages
);
7096 static int __init
cgroup_memory(char *s
)
7100 while ((token
= strsep(&s
, ",")) != NULL
) {
7103 if (!strcmp(token
, "nosocket"))
7104 cgroup_memory_nosocket
= true;
7105 if (!strcmp(token
, "nokmem"))
7106 cgroup_memory_nokmem
= true;
7110 __setup("cgroup.memory=", cgroup_memory
);
7113 * subsys_initcall() for memory controller.
7115 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7116 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7117 * basically everything that doesn't depend on a specific mem_cgroup structure
7118 * should be initialized from here.
7120 static int __init
mem_cgroup_init(void)
7125 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7126 * used for per-memcg-per-cpu caching of per-node statistics. In order
7127 * to work fine, we should make sure that the overfill threshold can't
7128 * exceed S32_MAX / PAGE_SIZE.
7130 BUILD_BUG_ON(MEMCG_CHARGE_BATCH
> S32_MAX
/ PAGE_SIZE
);
7132 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7133 memcg_hotplug_cpu_dead
);
7135 for_each_possible_cpu(cpu
)
7136 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7139 for_each_node(node
) {
7140 struct mem_cgroup_tree_per_node
*rtpn
;
7142 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7143 node_online(node
) ? node
: NUMA_NO_NODE
);
7145 rtpn
->rb_root
= RB_ROOT
;
7146 rtpn
->rb_rightmost
= NULL
;
7147 spin_lock_init(&rtpn
->lock
);
7148 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7153 subsys_initcall(mem_cgroup_init
);
7155 #ifdef CONFIG_MEMCG_SWAP
7156 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7158 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7160 * The root cgroup cannot be destroyed, so it's refcount must
7163 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7167 memcg
= parent_mem_cgroup(memcg
);
7169 memcg
= root_mem_cgroup
;
7175 * mem_cgroup_swapout - transfer a memsw charge to swap
7176 * @page: page whose memsw charge to transfer
7177 * @entry: swap entry to move the charge to
7179 * Transfer the memsw charge of @page to @entry.
7181 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7183 struct mem_cgroup
*memcg
, *swap_memcg
;
7184 unsigned int nr_entries
;
7185 unsigned short oldid
;
7187 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7188 VM_BUG_ON_PAGE(page_count(page
), page
);
7190 if (mem_cgroup_disabled())
7193 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7196 memcg
= page_memcg(page
);
7198 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7203 * In case the memcg owning these pages has been offlined and doesn't
7204 * have an ID allocated to it anymore, charge the closest online
7205 * ancestor for the swap instead and transfer the memory+swap charge.
7207 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7208 nr_entries
= thp_nr_pages(page
);
7209 /* Get references for the tail pages, too */
7211 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7212 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7214 VM_BUG_ON_PAGE(oldid
, page
);
7215 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7217 page
->memcg_data
= 0;
7219 if (!mem_cgroup_is_root(memcg
))
7220 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7222 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7223 if (!mem_cgroup_is_root(swap_memcg
))
7224 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7225 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7229 * Interrupts should be disabled here because the caller holds the
7230 * i_pages lock which is taken with interrupts-off. It is
7231 * important here to have the interrupts disabled because it is the
7232 * only synchronisation we have for updating the per-CPU variables.
7234 VM_BUG_ON(!irqs_disabled());
7235 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7236 memcg_check_events(memcg
, page
);
7238 css_put(&memcg
->css
);
7242 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7243 * @page: page being added to swap
7244 * @entry: swap entry to charge
7246 * Try to charge @page's memcg for the swap space at @entry.
7248 * Returns 0 on success, -ENOMEM on failure.
7250 int __mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7252 unsigned int nr_pages
= thp_nr_pages(page
);
7253 struct page_counter
*counter
;
7254 struct mem_cgroup
*memcg
;
7255 unsigned short oldid
;
7257 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7260 memcg
= page_memcg(page
);
7262 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7267 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7271 memcg
= mem_cgroup_id_get_online(memcg
);
7273 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7274 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7275 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7276 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7277 mem_cgroup_id_put(memcg
);
7281 /* Get references for the tail pages, too */
7283 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7284 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7285 VM_BUG_ON_PAGE(oldid
, page
);
7286 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7292 * __mem_cgroup_uncharge_swap - uncharge swap space
7293 * @entry: swap entry to uncharge
7294 * @nr_pages: the amount of swap space to uncharge
7296 void __mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7298 struct mem_cgroup
*memcg
;
7301 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7303 memcg
= mem_cgroup_from_id(id
);
7305 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7306 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7307 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7309 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7311 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7312 mem_cgroup_id_put_many(memcg
, nr_pages
);
7317 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7319 long nr_swap_pages
= get_nr_swap_pages();
7321 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7322 return nr_swap_pages
;
7323 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7324 nr_swap_pages
= min_t(long, nr_swap_pages
,
7325 READ_ONCE(memcg
->swap
.max
) -
7326 page_counter_read(&memcg
->swap
));
7327 return nr_swap_pages
;
7330 bool mem_cgroup_swap_full(struct page
*page
)
7332 struct mem_cgroup
*memcg
;
7334 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7338 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7341 memcg
= page_memcg(page
);
7345 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7346 unsigned long usage
= page_counter_read(&memcg
->swap
);
7348 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7349 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7356 static int __init
setup_swap_account(char *s
)
7358 if (!strcmp(s
, "1"))
7359 cgroup_memory_noswap
= false;
7360 else if (!strcmp(s
, "0"))
7361 cgroup_memory_noswap
= true;
7364 __setup("swapaccount=", setup_swap_account
);
7366 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7369 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7371 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7374 static int swap_high_show(struct seq_file
*m
, void *v
)
7376 return seq_puts_memcg_tunable(m
,
7377 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7380 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7381 char *buf
, size_t nbytes
, loff_t off
)
7383 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7387 buf
= strstrip(buf
);
7388 err
= page_counter_memparse(buf
, "max", &high
);
7392 page_counter_set_high(&memcg
->swap
, high
);
7397 static int swap_max_show(struct seq_file
*m
, void *v
)
7399 return seq_puts_memcg_tunable(m
,
7400 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7403 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7404 char *buf
, size_t nbytes
, loff_t off
)
7406 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7410 buf
= strstrip(buf
);
7411 err
= page_counter_memparse(buf
, "max", &max
);
7415 xchg(&memcg
->swap
.max
, max
);
7420 static int swap_events_show(struct seq_file
*m
, void *v
)
7422 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7424 seq_printf(m
, "high %lu\n",
7425 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7426 seq_printf(m
, "max %lu\n",
7427 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7428 seq_printf(m
, "fail %lu\n",
7429 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7434 static struct cftype swap_files
[] = {
7436 .name
= "swap.current",
7437 .flags
= CFTYPE_NOT_ON_ROOT
,
7438 .read_u64
= swap_current_read
,
7441 .name
= "swap.high",
7442 .flags
= CFTYPE_NOT_ON_ROOT
,
7443 .seq_show
= swap_high_show
,
7444 .write
= swap_high_write
,
7448 .flags
= CFTYPE_NOT_ON_ROOT
,
7449 .seq_show
= swap_max_show
,
7450 .write
= swap_max_write
,
7453 .name
= "swap.events",
7454 .flags
= CFTYPE_NOT_ON_ROOT
,
7455 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7456 .seq_show
= swap_events_show
,
7461 static struct cftype memsw_files
[] = {
7463 .name
= "memsw.usage_in_bytes",
7464 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7465 .read_u64
= mem_cgroup_read_u64
,
7468 .name
= "memsw.max_usage_in_bytes",
7469 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7470 .write
= mem_cgroup_reset
,
7471 .read_u64
= mem_cgroup_read_u64
,
7474 .name
= "memsw.limit_in_bytes",
7475 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7476 .write
= mem_cgroup_write
,
7477 .read_u64
= mem_cgroup_read_u64
,
7480 .name
= "memsw.failcnt",
7481 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7482 .write
= mem_cgroup_reset
,
7483 .read_u64
= mem_cgroup_read_u64
,
7485 { }, /* terminate */
7489 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7490 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7491 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7492 * boot parameter. This may result in premature OOPS inside
7493 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7495 static int __init
mem_cgroup_swap_init(void)
7497 /* No memory control -> no swap control */
7498 if (mem_cgroup_disabled())
7499 cgroup_memory_noswap
= true;
7501 if (cgroup_memory_noswap
)
7504 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7505 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7509 core_initcall(mem_cgroup_swap_init
);
7511 #endif /* CONFIG_MEMCG_SWAP */