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);
653 static u64 flush_next_time
;
655 #define FLUSH_TIME (2UL*HZ)
657 static inline void memcg_rstat_updated(struct mem_cgroup
*memcg
, int val
)
661 cgroup_rstat_updated(memcg
->css
.cgroup
, smp_processor_id());
663 x
= __this_cpu_add_return(stats_updates
, abs(val
));
664 if (x
> MEMCG_CHARGE_BATCH
) {
665 atomic_add(x
/ MEMCG_CHARGE_BATCH
, &stats_flush_threshold
);
666 __this_cpu_write(stats_updates
, 0);
670 static void __mem_cgroup_flush_stats(void)
674 if (!spin_trylock_irqsave(&stats_flush_lock
, flag
))
677 flush_next_time
= jiffies_64
+ 2*FLUSH_TIME
;
678 cgroup_rstat_flush_irqsafe(root_mem_cgroup
->css
.cgroup
);
679 atomic_set(&stats_flush_threshold
, 0);
680 spin_unlock_irqrestore(&stats_flush_lock
, flag
);
683 void mem_cgroup_flush_stats(void)
685 if (atomic_read(&stats_flush_threshold
) > num_online_cpus())
686 __mem_cgroup_flush_stats();
689 void mem_cgroup_flush_stats_delayed(void)
691 if (time_after64(jiffies_64
, flush_next_time
))
692 mem_cgroup_flush_stats();
695 static void flush_memcg_stats_dwork(struct work_struct
*w
)
697 __mem_cgroup_flush_stats();
698 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
, FLUSH_TIME
);
702 * __mod_memcg_state - update cgroup memory statistics
703 * @memcg: the memory cgroup
704 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
705 * @val: delta to add to the counter, can be negative
707 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
709 if (mem_cgroup_disabled())
712 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
713 memcg_rstat_updated(memcg
, val
);
716 /* idx can be of type enum memcg_stat_item or node_stat_item. */
717 static unsigned long memcg_page_state_local(struct mem_cgroup
*memcg
, int idx
)
722 for_each_possible_cpu(cpu
)
723 x
+= per_cpu(memcg
->vmstats_percpu
->state
[idx
], cpu
);
731 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
734 struct mem_cgroup_per_node
*pn
;
735 struct mem_cgroup
*memcg
;
737 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
741 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
744 __this_cpu_add(pn
->lruvec_stats_percpu
->state
[idx
], val
);
746 memcg_rstat_updated(memcg
, val
);
750 * __mod_lruvec_state - update lruvec memory statistics
751 * @lruvec: the lruvec
752 * @idx: the stat item
753 * @val: delta to add to the counter, can be negative
755 * The lruvec is the intersection of the NUMA node and a cgroup. This
756 * function updates the all three counters that are affected by a
757 * change of state at this level: per-node, per-cgroup, per-lruvec.
759 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
763 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
765 /* Update memcg and lruvec */
766 if (!mem_cgroup_disabled())
767 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
770 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
773 struct page
*head
= compound_head(page
); /* rmap on tail pages */
774 struct mem_cgroup
*memcg
;
775 pg_data_t
*pgdat
= page_pgdat(page
);
776 struct lruvec
*lruvec
;
779 memcg
= page_memcg(head
);
780 /* Untracked pages have no memcg, no lruvec. Update only the node */
783 __mod_node_page_state(pgdat
, idx
, val
);
787 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
788 __mod_lruvec_state(lruvec
, idx
, val
);
791 EXPORT_SYMBOL(__mod_lruvec_page_state
);
793 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
795 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
796 struct mem_cgroup
*memcg
;
797 struct lruvec
*lruvec
;
800 memcg
= mem_cgroup_from_obj(p
);
803 * Untracked pages have no memcg, no lruvec. Update only the
804 * node. If we reparent the slab objects to the root memcg,
805 * when we free the slab object, we need to update the per-memcg
806 * vmstats to keep it correct for the root memcg.
809 __mod_node_page_state(pgdat
, idx
, val
);
811 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
812 __mod_lruvec_state(lruvec
, idx
, val
);
818 * mod_objcg_mlstate() may be called with irq enabled, so
819 * mod_memcg_lruvec_state() should be used.
821 static inline void mod_objcg_mlstate(struct obj_cgroup
*objcg
,
822 struct pglist_data
*pgdat
,
823 enum node_stat_item idx
, int nr
)
825 struct mem_cgroup
*memcg
;
826 struct lruvec
*lruvec
;
829 memcg
= obj_cgroup_memcg(objcg
);
830 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
831 mod_memcg_lruvec_state(lruvec
, idx
, nr
);
836 * __count_memcg_events - account VM events in a cgroup
837 * @memcg: the memory cgroup
838 * @idx: the event item
839 * @count: the number of events that occurred
841 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
844 if (mem_cgroup_disabled())
847 __this_cpu_add(memcg
->vmstats_percpu
->events
[idx
], count
);
848 memcg_rstat_updated(memcg
, count
);
851 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
853 return READ_ONCE(memcg
->vmstats
.events
[event
]);
856 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
861 for_each_possible_cpu(cpu
)
862 x
+= per_cpu(memcg
->vmstats_percpu
->events
[event
], cpu
);
866 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
870 /* pagein of a big page is an event. So, ignore page size */
872 __count_memcg_events(memcg
, PGPGIN
, 1);
874 __count_memcg_events(memcg
, PGPGOUT
, 1);
875 nr_pages
= -nr_pages
; /* for event */
878 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
881 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
882 enum mem_cgroup_events_target target
)
884 unsigned long val
, next
;
886 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
887 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
888 /* from time_after() in jiffies.h */
889 if ((long)(next
- val
) < 0) {
891 case MEM_CGROUP_TARGET_THRESH
:
892 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
894 case MEM_CGROUP_TARGET_SOFTLIMIT
:
895 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
900 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
907 * Check events in order.
910 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
912 /* threshold event is triggered in finer grain than soft limit */
913 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
914 MEM_CGROUP_TARGET_THRESH
))) {
917 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
918 MEM_CGROUP_TARGET_SOFTLIMIT
);
919 mem_cgroup_threshold(memcg
);
920 if (unlikely(do_softlimit
))
921 mem_cgroup_update_tree(memcg
, page
);
925 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
928 * mm_update_next_owner() may clear mm->owner to NULL
929 * if it races with swapoff, page migration, etc.
930 * So this can be called with p == NULL.
935 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
937 EXPORT_SYMBOL(mem_cgroup_from_task
);
939 static __always_inline
struct mem_cgroup
*active_memcg(void)
942 return this_cpu_read(int_active_memcg
);
944 return current
->active_memcg
;
948 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
949 * @mm: mm from which memcg should be extracted. It can be NULL.
951 * Obtain a reference on mm->memcg and returns it if successful. If mm
952 * is NULL, then the memcg is chosen as follows:
953 * 1) The active memcg, if set.
954 * 2) current->mm->memcg, if available
956 * If mem_cgroup is disabled, NULL is returned.
958 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
960 struct mem_cgroup
*memcg
;
962 if (mem_cgroup_disabled())
966 * Page cache insertions can happen without an
967 * actual mm context, e.g. during disk probing
968 * on boot, loopback IO, acct() writes etc.
970 * No need to css_get on root memcg as the reference
971 * counting is disabled on the root level in the
972 * cgroup core. See CSS_NO_REF.
975 memcg
= active_memcg();
976 if (unlikely(memcg
)) {
977 /* remote memcg must hold a ref */
978 css_get(&memcg
->css
);
983 return root_mem_cgroup
;
988 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
989 if (unlikely(!memcg
))
990 memcg
= root_mem_cgroup
;
991 } while (!css_tryget(&memcg
->css
));
995 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
997 static __always_inline
bool memcg_kmem_bypass(void)
999 /* Allow remote memcg charging from any context. */
1000 if (unlikely(active_memcg()))
1003 /* Memcg to charge can't be determined. */
1004 if (!in_task() || !current
->mm
|| (current
->flags
& PF_KTHREAD
))
1011 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1012 * @root: hierarchy root
1013 * @prev: previously returned memcg, NULL on first invocation
1014 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1016 * Returns references to children of the hierarchy below @root, or
1017 * @root itself, or %NULL after a full round-trip.
1019 * Caller must pass the return value in @prev on subsequent
1020 * invocations for reference counting, or use mem_cgroup_iter_break()
1021 * to cancel a hierarchy walk before the round-trip is complete.
1023 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1024 * in the hierarchy among all concurrent reclaimers operating on the
1027 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1028 struct mem_cgroup
*prev
,
1029 struct mem_cgroup_reclaim_cookie
*reclaim
)
1031 struct mem_cgroup_reclaim_iter
*iter
;
1032 struct cgroup_subsys_state
*css
= NULL
;
1033 struct mem_cgroup
*memcg
= NULL
;
1034 struct mem_cgroup
*pos
= NULL
;
1036 if (mem_cgroup_disabled())
1040 root
= root_mem_cgroup
;
1042 if (prev
&& !reclaim
)
1048 struct mem_cgroup_per_node
*mz
;
1050 mz
= root
->nodeinfo
[reclaim
->pgdat
->node_id
];
1053 if (prev
&& reclaim
->generation
!= iter
->generation
)
1057 pos
= READ_ONCE(iter
->position
);
1058 if (!pos
|| css_tryget(&pos
->css
))
1061 * css reference reached zero, so iter->position will
1062 * be cleared by ->css_released. However, we should not
1063 * rely on this happening soon, because ->css_released
1064 * is called from a work queue, and by busy-waiting we
1065 * might block it. So we clear iter->position right
1068 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1076 css
= css_next_descendant_pre(css
, &root
->css
);
1079 * Reclaimers share the hierarchy walk, and a
1080 * new one might jump in right at the end of
1081 * the hierarchy - make sure they see at least
1082 * one group and restart from the beginning.
1090 * Verify the css and acquire a reference. The root
1091 * is provided by the caller, so we know it's alive
1092 * and kicking, and don't take an extra reference.
1094 memcg
= mem_cgroup_from_css(css
);
1096 if (css
== &root
->css
)
1099 if (css_tryget(css
))
1107 * The position could have already been updated by a competing
1108 * thread, so check that the value hasn't changed since we read
1109 * it to avoid reclaiming from the same cgroup twice.
1111 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1119 reclaim
->generation
= iter
->generation
;
1124 if (prev
&& prev
!= root
)
1125 css_put(&prev
->css
);
1131 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1132 * @root: hierarchy root
1133 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1135 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1136 struct mem_cgroup
*prev
)
1139 root
= root_mem_cgroup
;
1140 if (prev
&& prev
!= root
)
1141 css_put(&prev
->css
);
1144 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1145 struct mem_cgroup
*dead_memcg
)
1147 struct mem_cgroup_reclaim_iter
*iter
;
1148 struct mem_cgroup_per_node
*mz
;
1151 for_each_node(nid
) {
1152 mz
= from
->nodeinfo
[nid
];
1154 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1158 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1160 struct mem_cgroup
*memcg
= dead_memcg
;
1161 struct mem_cgroup
*last
;
1164 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1166 } while ((memcg
= parent_mem_cgroup(memcg
)));
1169 * When cgruop1 non-hierarchy mode is used,
1170 * parent_mem_cgroup() does not walk all the way up to the
1171 * cgroup root (root_mem_cgroup). So we have to handle
1172 * dead_memcg from cgroup root separately.
1174 if (last
!= root_mem_cgroup
)
1175 __invalidate_reclaim_iterators(root_mem_cgroup
,
1180 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1181 * @memcg: hierarchy root
1182 * @fn: function to call for each task
1183 * @arg: argument passed to @fn
1185 * This function iterates over tasks attached to @memcg or to any of its
1186 * descendants and calls @fn for each task. If @fn returns a non-zero
1187 * value, the function breaks the iteration loop and returns the value.
1188 * Otherwise, it will iterate over all tasks and return 0.
1190 * This function must not be called for the root memory cgroup.
1192 int mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1193 int (*fn
)(struct task_struct
*, void *), void *arg
)
1195 struct mem_cgroup
*iter
;
1198 BUG_ON(memcg
== root_mem_cgroup
);
1200 for_each_mem_cgroup_tree(iter
, memcg
) {
1201 struct css_task_iter it
;
1202 struct task_struct
*task
;
1204 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1205 while (!ret
&& (task
= css_task_iter_next(&it
)))
1206 ret
= fn(task
, arg
);
1207 css_task_iter_end(&it
);
1209 mem_cgroup_iter_break(memcg
, iter
);
1216 #ifdef CONFIG_DEBUG_VM
1217 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct page
*page
)
1219 struct mem_cgroup
*memcg
;
1221 if (mem_cgroup_disabled())
1224 memcg
= page_memcg(page
);
1227 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != root_mem_cgroup
, page
);
1229 VM_BUG_ON_PAGE(lruvec_memcg(lruvec
) != memcg
, page
);
1234 * lock_page_lruvec - lock and return lruvec for a given page.
1237 * These functions are safe to use under any of the following conditions:
1240 * - lock_page_memcg()
1241 * - page->_refcount is zero
1243 struct lruvec
*lock_page_lruvec(struct page
*page
)
1245 struct lruvec
*lruvec
;
1247 lruvec
= mem_cgroup_page_lruvec(page
);
1248 spin_lock(&lruvec
->lru_lock
);
1250 lruvec_memcg_debug(lruvec
, page
);
1255 struct lruvec
*lock_page_lruvec_irq(struct page
*page
)
1257 struct lruvec
*lruvec
;
1259 lruvec
= mem_cgroup_page_lruvec(page
);
1260 spin_lock_irq(&lruvec
->lru_lock
);
1262 lruvec_memcg_debug(lruvec
, page
);
1267 struct lruvec
*lock_page_lruvec_irqsave(struct page
*page
, unsigned long *flags
)
1269 struct lruvec
*lruvec
;
1271 lruvec
= mem_cgroup_page_lruvec(page
);
1272 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1274 lruvec_memcg_debug(lruvec
, page
);
1280 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1281 * @lruvec: mem_cgroup per zone lru vector
1282 * @lru: index of lru list the page is sitting on
1283 * @zid: zone id of the accounted pages
1284 * @nr_pages: positive when adding or negative when removing
1286 * This function must be called under lru_lock, just before a page is added
1287 * to or just after a page is removed from an lru list (that ordering being
1288 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1290 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1291 int zid
, int nr_pages
)
1293 struct mem_cgroup_per_node
*mz
;
1294 unsigned long *lru_size
;
1297 if (mem_cgroup_disabled())
1300 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1301 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1304 *lru_size
+= nr_pages
;
1307 if (WARN_ONCE(size
< 0,
1308 "%s(%p, %d, %d): lru_size %ld\n",
1309 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1315 *lru_size
+= nr_pages
;
1319 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1320 * @memcg: the memory cgroup
1322 * Returns the maximum amount of memory @mem can be charged with, in
1325 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1327 unsigned long margin
= 0;
1328 unsigned long count
;
1329 unsigned long limit
;
1331 count
= page_counter_read(&memcg
->memory
);
1332 limit
= READ_ONCE(memcg
->memory
.max
);
1334 margin
= limit
- count
;
1336 if (do_memsw_account()) {
1337 count
= page_counter_read(&memcg
->memsw
);
1338 limit
= READ_ONCE(memcg
->memsw
.max
);
1340 margin
= min(margin
, limit
- count
);
1349 * A routine for checking "mem" is under move_account() or not.
1351 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1352 * moving cgroups. This is for waiting at high-memory pressure
1355 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1357 struct mem_cgroup
*from
;
1358 struct mem_cgroup
*to
;
1361 * Unlike task_move routines, we access mc.to, mc.from not under
1362 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1364 spin_lock(&mc
.lock
);
1370 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1371 mem_cgroup_is_descendant(to
, memcg
);
1373 spin_unlock(&mc
.lock
);
1377 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1379 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1380 if (mem_cgroup_under_move(memcg
)) {
1382 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1383 /* moving charge context might have finished. */
1386 finish_wait(&mc
.waitq
, &wait
);
1393 struct memory_stat
{
1398 static const struct memory_stat memory_stats
[] = {
1399 { "anon", NR_ANON_MAPPED
},
1400 { "file", NR_FILE_PAGES
},
1401 { "kernel_stack", NR_KERNEL_STACK_KB
},
1402 { "pagetables", NR_PAGETABLE
},
1403 { "percpu", MEMCG_PERCPU_B
},
1404 { "sock", MEMCG_SOCK
},
1405 { "shmem", NR_SHMEM
},
1406 { "file_mapped", NR_FILE_MAPPED
},
1407 { "file_dirty", NR_FILE_DIRTY
},
1408 { "file_writeback", NR_WRITEBACK
},
1410 { "swapcached", NR_SWAPCACHE
},
1412 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1413 { "anon_thp", NR_ANON_THPS
},
1414 { "file_thp", NR_FILE_THPS
},
1415 { "shmem_thp", NR_SHMEM_THPS
},
1417 { "inactive_anon", NR_INACTIVE_ANON
},
1418 { "active_anon", NR_ACTIVE_ANON
},
1419 { "inactive_file", NR_INACTIVE_FILE
},
1420 { "active_file", NR_ACTIVE_FILE
},
1421 { "unevictable", NR_UNEVICTABLE
},
1422 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B
},
1423 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B
},
1425 /* The memory events */
1426 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON
},
1427 { "workingset_refault_file", WORKINGSET_REFAULT_FILE
},
1428 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON
},
1429 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE
},
1430 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON
},
1431 { "workingset_restore_file", WORKINGSET_RESTORE_FILE
},
1432 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM
},
1435 /* Translate stat items to the correct unit for memory.stat output */
1436 static int memcg_page_state_unit(int item
)
1439 case MEMCG_PERCPU_B
:
1440 case NR_SLAB_RECLAIMABLE_B
:
1441 case NR_SLAB_UNRECLAIMABLE_B
:
1442 case WORKINGSET_REFAULT_ANON
:
1443 case WORKINGSET_REFAULT_FILE
:
1444 case WORKINGSET_ACTIVATE_ANON
:
1445 case WORKINGSET_ACTIVATE_FILE
:
1446 case WORKINGSET_RESTORE_ANON
:
1447 case WORKINGSET_RESTORE_FILE
:
1448 case WORKINGSET_NODERECLAIM
:
1450 case NR_KERNEL_STACK_KB
:
1457 static inline unsigned long memcg_page_state_output(struct mem_cgroup
*memcg
,
1460 return memcg_page_state(memcg
, item
) * memcg_page_state_unit(item
);
1463 static char *memory_stat_format(struct mem_cgroup
*memcg
)
1468 seq_buf_init(&s
, kmalloc(PAGE_SIZE
, GFP_KERNEL
), PAGE_SIZE
);
1473 * Provide statistics on the state of the memory subsystem as
1474 * well as cumulative event counters that show past behavior.
1476 * This list is ordered following a combination of these gradients:
1477 * 1) generic big picture -> specifics and details
1478 * 2) reflecting userspace activity -> reflecting kernel heuristics
1480 * Current memory state:
1482 mem_cgroup_flush_stats();
1484 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1487 size
= memcg_page_state_output(memcg
, memory_stats
[i
].idx
);
1488 seq_buf_printf(&s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1490 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1491 size
+= memcg_page_state_output(memcg
,
1492 NR_SLAB_RECLAIMABLE_B
);
1493 seq_buf_printf(&s
, "slab %llu\n", size
);
1497 /* Accumulated memory events */
1499 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGFAULT
),
1500 memcg_events(memcg
, PGFAULT
));
1501 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGMAJFAULT
),
1502 memcg_events(memcg
, PGMAJFAULT
));
1503 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGREFILL
),
1504 memcg_events(memcg
, PGREFILL
));
1505 seq_buf_printf(&s
, "pgscan %lu\n",
1506 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1507 memcg_events(memcg
, PGSCAN_DIRECT
));
1508 seq_buf_printf(&s
, "pgsteal %lu\n",
1509 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1510 memcg_events(memcg
, PGSTEAL_DIRECT
));
1511 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGACTIVATE
),
1512 memcg_events(memcg
, PGACTIVATE
));
1513 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGDEACTIVATE
),
1514 memcg_events(memcg
, PGDEACTIVATE
));
1515 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREE
),
1516 memcg_events(memcg
, PGLAZYFREE
));
1517 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(PGLAZYFREED
),
1518 memcg_events(memcg
, PGLAZYFREED
));
1520 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1521 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC
),
1522 memcg_events(memcg
, THP_FAULT_ALLOC
));
1523 seq_buf_printf(&s
, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC
),
1524 memcg_events(memcg
, THP_COLLAPSE_ALLOC
));
1525 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1527 /* The above should easily fit into one page */
1528 WARN_ON_ONCE(seq_buf_has_overflowed(&s
));
1533 #define K(x) ((x) << (PAGE_SHIFT-10))
1535 * mem_cgroup_print_oom_context: Print OOM information relevant to
1536 * memory controller.
1537 * @memcg: The memory cgroup that went over limit
1538 * @p: Task that is going to be killed
1540 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1543 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1548 pr_cont(",oom_memcg=");
1549 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1551 pr_cont(",global_oom");
1553 pr_cont(",task_memcg=");
1554 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1560 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1561 * memory controller.
1562 * @memcg: The memory cgroup that went over limit
1564 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1568 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1569 K((u64
)page_counter_read(&memcg
->memory
)),
1570 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1571 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1572 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1573 K((u64
)page_counter_read(&memcg
->swap
)),
1574 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1576 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1577 K((u64
)page_counter_read(&memcg
->memsw
)),
1578 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1579 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1580 K((u64
)page_counter_read(&memcg
->kmem
)),
1581 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1584 pr_info("Memory cgroup stats for ");
1585 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1587 buf
= memory_stat_format(memcg
);
1595 * Return the memory (and swap, if configured) limit for a memcg.
1597 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1599 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1601 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
1602 if (mem_cgroup_swappiness(memcg
))
1603 max
+= min(READ_ONCE(memcg
->swap
.max
),
1604 (unsigned long)total_swap_pages
);
1606 if (mem_cgroup_swappiness(memcg
)) {
1607 /* Calculate swap excess capacity from memsw limit */
1608 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1610 max
+= min(swap
, (unsigned long)total_swap_pages
);
1616 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1618 return page_counter_read(&memcg
->memory
);
1621 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1624 struct oom_control oc
= {
1628 .gfp_mask
= gfp_mask
,
1633 if (mutex_lock_killable(&oom_lock
))
1636 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1640 * A few threads which were not waiting at mutex_lock_killable() can
1641 * fail to bail out. Therefore, check again after holding oom_lock.
1643 ret
= task_is_dying() || out_of_memory(&oc
);
1646 mutex_unlock(&oom_lock
);
1650 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1653 unsigned long *total_scanned
)
1655 struct mem_cgroup
*victim
= NULL
;
1658 unsigned long excess
;
1659 unsigned long nr_scanned
;
1660 struct mem_cgroup_reclaim_cookie reclaim
= {
1664 excess
= soft_limit_excess(root_memcg
);
1667 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1672 * If we have not been able to reclaim
1673 * anything, it might because there are
1674 * no reclaimable pages under this hierarchy
1679 * We want to do more targeted reclaim.
1680 * excess >> 2 is not to excessive so as to
1681 * reclaim too much, nor too less that we keep
1682 * coming back to reclaim from this cgroup
1684 if (total
>= (excess
>> 2) ||
1685 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1690 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1691 pgdat
, &nr_scanned
);
1692 *total_scanned
+= nr_scanned
;
1693 if (!soft_limit_excess(root_memcg
))
1696 mem_cgroup_iter_break(root_memcg
, victim
);
1700 #ifdef CONFIG_LOCKDEP
1701 static struct lockdep_map memcg_oom_lock_dep_map
= {
1702 .name
= "memcg_oom_lock",
1706 static DEFINE_SPINLOCK(memcg_oom_lock
);
1709 * Check OOM-Killer is already running under our hierarchy.
1710 * If someone is running, return false.
1712 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1714 struct mem_cgroup
*iter
, *failed
= NULL
;
1716 spin_lock(&memcg_oom_lock
);
1718 for_each_mem_cgroup_tree(iter
, memcg
) {
1719 if (iter
->oom_lock
) {
1721 * this subtree of our hierarchy is already locked
1722 * so we cannot give a lock.
1725 mem_cgroup_iter_break(memcg
, iter
);
1728 iter
->oom_lock
= true;
1733 * OK, we failed to lock the whole subtree so we have
1734 * to clean up what we set up to the failing subtree
1736 for_each_mem_cgroup_tree(iter
, memcg
) {
1737 if (iter
== failed
) {
1738 mem_cgroup_iter_break(memcg
, iter
);
1741 iter
->oom_lock
= false;
1744 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1746 spin_unlock(&memcg_oom_lock
);
1751 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1753 struct mem_cgroup
*iter
;
1755 spin_lock(&memcg_oom_lock
);
1756 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1757 for_each_mem_cgroup_tree(iter
, memcg
)
1758 iter
->oom_lock
= false;
1759 spin_unlock(&memcg_oom_lock
);
1762 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1764 struct mem_cgroup
*iter
;
1766 spin_lock(&memcg_oom_lock
);
1767 for_each_mem_cgroup_tree(iter
, memcg
)
1769 spin_unlock(&memcg_oom_lock
);
1772 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1774 struct mem_cgroup
*iter
;
1777 * Be careful about under_oom underflows because a child memcg
1778 * could have been added after mem_cgroup_mark_under_oom.
1780 spin_lock(&memcg_oom_lock
);
1781 for_each_mem_cgroup_tree(iter
, memcg
)
1782 if (iter
->under_oom
> 0)
1784 spin_unlock(&memcg_oom_lock
);
1787 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1789 struct oom_wait_info
{
1790 struct mem_cgroup
*memcg
;
1791 wait_queue_entry_t wait
;
1794 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1795 unsigned mode
, int sync
, void *arg
)
1797 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1798 struct mem_cgroup
*oom_wait_memcg
;
1799 struct oom_wait_info
*oom_wait_info
;
1801 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1802 oom_wait_memcg
= oom_wait_info
->memcg
;
1804 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1805 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1807 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1810 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1813 * For the following lockless ->under_oom test, the only required
1814 * guarantee is that it must see the state asserted by an OOM when
1815 * this function is called as a result of userland actions
1816 * triggered by the notification of the OOM. This is trivially
1817 * achieved by invoking mem_cgroup_mark_under_oom() before
1818 * triggering notification.
1820 if (memcg
&& memcg
->under_oom
)
1821 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1831 static enum oom_status
mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1833 enum oom_status ret
;
1836 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1839 memcg_memory_event(memcg
, MEMCG_OOM
);
1842 * We are in the middle of the charge context here, so we
1843 * don't want to block when potentially sitting on a callstack
1844 * that holds all kinds of filesystem and mm locks.
1846 * cgroup1 allows disabling the OOM killer and waiting for outside
1847 * handling until the charge can succeed; remember the context and put
1848 * the task to sleep at the end of the page fault when all locks are
1851 * On the other hand, in-kernel OOM killer allows for an async victim
1852 * memory reclaim (oom_reaper) and that means that we are not solely
1853 * relying on the oom victim to make a forward progress and we can
1854 * invoke the oom killer here.
1856 * Please note that mem_cgroup_out_of_memory might fail to find a
1857 * victim and then we have to bail out from the charge path.
1859 if (memcg
->oom_kill_disable
) {
1860 if (!current
->in_user_fault
)
1862 css_get(&memcg
->css
);
1863 current
->memcg_in_oom
= memcg
;
1864 current
->memcg_oom_gfp_mask
= mask
;
1865 current
->memcg_oom_order
= order
;
1870 mem_cgroup_mark_under_oom(memcg
);
1872 locked
= mem_cgroup_oom_trylock(memcg
);
1875 mem_cgroup_oom_notify(memcg
);
1877 mem_cgroup_unmark_under_oom(memcg
);
1878 if (mem_cgroup_out_of_memory(memcg
, mask
, order
))
1884 mem_cgroup_oom_unlock(memcg
);
1890 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1891 * @handle: actually kill/wait or just clean up the OOM state
1893 * This has to be called at the end of a page fault if the memcg OOM
1894 * handler was enabled.
1896 * Memcg supports userspace OOM handling where failed allocations must
1897 * sleep on a waitqueue until the userspace task resolves the
1898 * situation. Sleeping directly in the charge context with all kinds
1899 * of locks held is not a good idea, instead we remember an OOM state
1900 * in the task and mem_cgroup_oom_synchronize() has to be called at
1901 * the end of the page fault to complete the OOM handling.
1903 * Returns %true if an ongoing memcg OOM situation was detected and
1904 * completed, %false otherwise.
1906 bool mem_cgroup_oom_synchronize(bool handle
)
1908 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1909 struct oom_wait_info owait
;
1912 /* OOM is global, do not handle */
1919 owait
.memcg
= memcg
;
1920 owait
.wait
.flags
= 0;
1921 owait
.wait
.func
= memcg_oom_wake_function
;
1922 owait
.wait
.private = current
;
1923 INIT_LIST_HEAD(&owait
.wait
.entry
);
1925 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1926 mem_cgroup_mark_under_oom(memcg
);
1928 locked
= mem_cgroup_oom_trylock(memcg
);
1931 mem_cgroup_oom_notify(memcg
);
1933 if (locked
&& !memcg
->oom_kill_disable
) {
1934 mem_cgroup_unmark_under_oom(memcg
);
1935 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1936 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1937 current
->memcg_oom_order
);
1940 mem_cgroup_unmark_under_oom(memcg
);
1941 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1945 mem_cgroup_oom_unlock(memcg
);
1947 * There is no guarantee that an OOM-lock contender
1948 * sees the wakeups triggered by the OOM kill
1949 * uncharges. Wake any sleepers explicitly.
1951 memcg_oom_recover(memcg
);
1954 current
->memcg_in_oom
= NULL
;
1955 css_put(&memcg
->css
);
1960 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1961 * @victim: task to be killed by the OOM killer
1962 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1964 * Returns a pointer to a memory cgroup, which has to be cleaned up
1965 * by killing all belonging OOM-killable tasks.
1967 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1969 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
1970 struct mem_cgroup
*oom_domain
)
1972 struct mem_cgroup
*oom_group
= NULL
;
1973 struct mem_cgroup
*memcg
;
1975 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1979 oom_domain
= root_mem_cgroup
;
1983 memcg
= mem_cgroup_from_task(victim
);
1984 if (memcg
== root_mem_cgroup
)
1988 * If the victim task has been asynchronously moved to a different
1989 * memory cgroup, we might end up killing tasks outside oom_domain.
1990 * In this case it's better to ignore memory.group.oom.
1992 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
1996 * Traverse the memory cgroup hierarchy from the victim task's
1997 * cgroup up to the OOMing cgroup (or root) to find the
1998 * highest-level memory cgroup with oom.group set.
2000 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2001 if (memcg
->oom_group
)
2004 if (memcg
== oom_domain
)
2009 css_get(&oom_group
->css
);
2016 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2018 pr_info("Tasks in ");
2019 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2020 pr_cont(" are going to be killed due to memory.oom.group set\n");
2024 * lock_page_memcg - lock a page and memcg binding
2027 * This function protects unlocked LRU pages from being moved to
2030 * It ensures lifetime of the locked memcg. Caller is responsible
2031 * for the lifetime of the page.
2033 void lock_page_memcg(struct page
*page
)
2035 struct page
*head
= compound_head(page
); /* rmap on tail pages */
2036 struct mem_cgroup
*memcg
;
2037 unsigned long flags
;
2040 * The RCU lock is held throughout the transaction. The fast
2041 * path can get away without acquiring the memcg->move_lock
2042 * because page moving starts with an RCU grace period.
2046 if (mem_cgroup_disabled())
2049 memcg
= page_memcg(head
);
2050 if (unlikely(!memcg
))
2053 #ifdef CONFIG_PROVE_LOCKING
2054 local_irq_save(flags
);
2055 might_lock(&memcg
->move_lock
);
2056 local_irq_restore(flags
);
2059 if (atomic_read(&memcg
->moving_account
) <= 0)
2062 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2063 if (memcg
!= page_memcg(head
)) {
2064 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2069 * When charge migration first begins, we can have multiple
2070 * critical sections holding the fast-path RCU lock and one
2071 * holding the slowpath move_lock. Track the task who has the
2072 * move_lock for unlock_page_memcg().
2074 memcg
->move_lock_task
= current
;
2075 memcg
->move_lock_flags
= flags
;
2077 EXPORT_SYMBOL(lock_page_memcg
);
2079 static void __unlock_page_memcg(struct mem_cgroup
*memcg
)
2081 if (memcg
&& memcg
->move_lock_task
== current
) {
2082 unsigned long flags
= memcg
->move_lock_flags
;
2084 memcg
->move_lock_task
= NULL
;
2085 memcg
->move_lock_flags
= 0;
2087 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2094 * unlock_page_memcg - unlock a page and memcg binding
2097 void unlock_page_memcg(struct page
*page
)
2099 struct page
*head
= compound_head(page
);
2101 __unlock_page_memcg(page_memcg(head
));
2103 EXPORT_SYMBOL(unlock_page_memcg
);
2106 #ifdef CONFIG_MEMCG_KMEM
2107 struct obj_cgroup
*cached_objcg
;
2108 struct pglist_data
*cached_pgdat
;
2109 unsigned int nr_bytes
;
2110 int nr_slab_reclaimable_b
;
2111 int nr_slab_unreclaimable_b
;
2117 struct memcg_stock_pcp
{
2118 struct mem_cgroup
*cached
; /* this never be root cgroup */
2119 unsigned int nr_pages
;
2120 struct obj_stock task_obj
;
2121 struct obj_stock irq_obj
;
2123 struct work_struct work
;
2124 unsigned long flags
;
2125 #define FLUSHING_CACHED_CHARGE 0
2127 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2128 static DEFINE_MUTEX(percpu_charge_mutex
);
2130 #ifdef CONFIG_MEMCG_KMEM
2131 static void drain_obj_stock(struct obj_stock
*stock
);
2132 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2133 struct mem_cgroup
*root_memcg
);
2136 static inline void drain_obj_stock(struct obj_stock
*stock
)
2139 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2140 struct mem_cgroup
*root_memcg
)
2147 * Most kmem_cache_alloc() calls are from user context. The irq disable/enable
2148 * sequence used in this case to access content from object stock is slow.
2149 * To optimize for user context access, there are now two object stocks for
2150 * task context and interrupt context access respectively.
2152 * The task context object stock can be accessed by disabling preemption only
2153 * which is cheap in non-preempt kernel. The interrupt context object stock
2154 * can only be accessed after disabling interrupt. User context code can
2155 * access interrupt object stock, but not vice versa.
2157 static inline struct obj_stock
*get_obj_stock(unsigned long *pflags
)
2159 struct memcg_stock_pcp
*stock
;
2161 if (likely(in_task())) {
2164 stock
= this_cpu_ptr(&memcg_stock
);
2165 return &stock
->task_obj
;
2168 local_irq_save(*pflags
);
2169 stock
= this_cpu_ptr(&memcg_stock
);
2170 return &stock
->irq_obj
;
2173 static inline void put_obj_stock(unsigned long flags
)
2175 if (likely(in_task()))
2178 local_irq_restore(flags
);
2182 * consume_stock: Try to consume stocked charge on this cpu.
2183 * @memcg: memcg to consume from.
2184 * @nr_pages: how many pages to charge.
2186 * The charges will only happen if @memcg matches the current cpu's memcg
2187 * stock, and at least @nr_pages are available in that stock. Failure to
2188 * service an allocation will refill the stock.
2190 * returns true if successful, false otherwise.
2192 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2194 struct memcg_stock_pcp
*stock
;
2195 unsigned long flags
;
2198 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2201 local_irq_save(flags
);
2203 stock
= this_cpu_ptr(&memcg_stock
);
2204 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2205 stock
->nr_pages
-= nr_pages
;
2209 local_irq_restore(flags
);
2215 * Returns stocks cached in percpu and reset cached information.
2217 static void drain_stock(struct memcg_stock_pcp
*stock
)
2219 struct mem_cgroup
*old
= stock
->cached
;
2224 if (stock
->nr_pages
) {
2225 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2226 if (do_memsw_account())
2227 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2228 stock
->nr_pages
= 0;
2232 stock
->cached
= NULL
;
2235 static void drain_local_stock(struct work_struct
*dummy
)
2237 struct memcg_stock_pcp
*stock
;
2238 unsigned long flags
;
2241 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2242 * drain_stock races is that we always operate on local CPU stock
2243 * here with IRQ disabled
2245 local_irq_save(flags
);
2247 stock
= this_cpu_ptr(&memcg_stock
);
2248 drain_obj_stock(&stock
->irq_obj
);
2250 drain_obj_stock(&stock
->task_obj
);
2252 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2254 local_irq_restore(flags
);
2258 * Cache charges(val) to local per_cpu area.
2259 * This will be consumed by consume_stock() function, later.
2261 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2263 struct memcg_stock_pcp
*stock
;
2264 unsigned long flags
;
2266 local_irq_save(flags
);
2268 stock
= this_cpu_ptr(&memcg_stock
);
2269 if (stock
->cached
!= memcg
) { /* reset if necessary */
2271 css_get(&memcg
->css
);
2272 stock
->cached
= memcg
;
2274 stock
->nr_pages
+= nr_pages
;
2276 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2279 local_irq_restore(flags
);
2283 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2284 * of the hierarchy under it.
2286 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2290 /* If someone's already draining, avoid adding running more workers. */
2291 if (!mutex_trylock(&percpu_charge_mutex
))
2294 * Notify other cpus that system-wide "drain" is running
2295 * We do not care about races with the cpu hotplug because cpu down
2296 * as well as workers from this path always operate on the local
2297 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2300 for_each_online_cpu(cpu
) {
2301 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2302 struct mem_cgroup
*memcg
;
2306 memcg
= stock
->cached
;
2307 if (memcg
&& stock
->nr_pages
&&
2308 mem_cgroup_is_descendant(memcg
, root_memcg
))
2310 else if (obj_stock_flush_required(stock
, root_memcg
))
2315 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2317 drain_local_stock(&stock
->work
);
2319 schedule_work_on(cpu
, &stock
->work
);
2323 mutex_unlock(&percpu_charge_mutex
);
2326 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2328 struct memcg_stock_pcp
*stock
;
2330 stock
= &per_cpu(memcg_stock
, cpu
);
2336 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2337 unsigned int nr_pages
,
2340 unsigned long nr_reclaimed
= 0;
2343 unsigned long pflags
;
2345 if (page_counter_read(&memcg
->memory
) <=
2346 READ_ONCE(memcg
->memory
.high
))
2349 memcg_memory_event(memcg
, MEMCG_HIGH
);
2351 psi_memstall_enter(&pflags
);
2352 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2354 psi_memstall_leave(&pflags
);
2355 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2356 !mem_cgroup_is_root(memcg
));
2358 return nr_reclaimed
;
2361 static void high_work_func(struct work_struct
*work
)
2363 struct mem_cgroup
*memcg
;
2365 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2366 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2370 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2371 * enough to still cause a significant slowdown in most cases, while still
2372 * allowing diagnostics and tracing to proceed without becoming stuck.
2374 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2377 * When calculating the delay, we use these either side of the exponentiation to
2378 * maintain precision and scale to a reasonable number of jiffies (see the table
2381 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2382 * overage ratio to a delay.
2383 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2384 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2385 * to produce a reasonable delay curve.
2387 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2388 * reasonable delay curve compared to precision-adjusted overage, not
2389 * penalising heavily at first, but still making sure that growth beyond the
2390 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2391 * example, with a high of 100 megabytes:
2393 * +-------+------------------------+
2394 * | usage | time to allocate in ms |
2395 * +-------+------------------------+
2417 * +-------+------------------------+
2419 #define MEMCG_DELAY_PRECISION_SHIFT 20
2420 #define MEMCG_DELAY_SCALING_SHIFT 14
2422 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2430 * Prevent division by 0 in overage calculation by acting as if
2431 * it was a threshold of 1 page
2433 high
= max(high
, 1UL);
2435 overage
= usage
- high
;
2436 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2437 return div64_u64(overage
, high
);
2440 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2442 u64 overage
, max_overage
= 0;
2445 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2446 READ_ONCE(memcg
->memory
.high
));
2447 max_overage
= max(overage
, max_overage
);
2448 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2449 !mem_cgroup_is_root(memcg
));
2454 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2456 u64 overage
, max_overage
= 0;
2459 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2460 READ_ONCE(memcg
->swap
.high
));
2462 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2463 max_overage
= max(overage
, max_overage
);
2464 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2465 !mem_cgroup_is_root(memcg
));
2471 * Get the number of jiffies that we should penalise a mischievous cgroup which
2472 * is exceeding its memory.high by checking both it and its ancestors.
2474 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2475 unsigned int nr_pages
,
2478 unsigned long penalty_jiffies
;
2484 * We use overage compared to memory.high to calculate the number of
2485 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2486 * fairly lenient on small overages, and increasingly harsh when the
2487 * memcg in question makes it clear that it has no intention of stopping
2488 * its crazy behaviour, so we exponentially increase the delay based on
2491 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2492 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2493 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2496 * Factor in the task's own contribution to the overage, such that four
2497 * N-sized allocations are throttled approximately the same as one
2498 * 4N-sized allocation.
2500 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2501 * larger the current charge patch is than that.
2503 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2507 * Scheduled by try_charge() to be executed from the userland return path
2508 * and reclaims memory over the high limit.
2510 void mem_cgroup_handle_over_high(void)
2512 unsigned long penalty_jiffies
;
2513 unsigned long pflags
;
2514 unsigned long nr_reclaimed
;
2515 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2516 int nr_retries
= MAX_RECLAIM_RETRIES
;
2517 struct mem_cgroup
*memcg
;
2518 bool in_retry
= false;
2520 if (likely(!nr_pages
))
2523 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2524 current
->memcg_nr_pages_over_high
= 0;
2528 * The allocating task should reclaim at least the batch size, but for
2529 * subsequent retries we only want to do what's necessary to prevent oom
2530 * or breaching resource isolation.
2532 * This is distinct from memory.max or page allocator behaviour because
2533 * memory.high is currently batched, whereas memory.max and the page
2534 * allocator run every time an allocation is made.
2536 nr_reclaimed
= reclaim_high(memcg
,
2537 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2541 * memory.high is breached and reclaim is unable to keep up. Throttle
2542 * allocators proactively to slow down excessive growth.
2544 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2545 mem_find_max_overage(memcg
));
2547 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2548 swap_find_max_overage(memcg
));
2551 * Clamp the max delay per usermode return so as to still keep the
2552 * application moving forwards and also permit diagnostics, albeit
2555 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2558 * Don't sleep if the amount of jiffies this memcg owes us is so low
2559 * that it's not even worth doing, in an attempt to be nice to those who
2560 * go only a small amount over their memory.high value and maybe haven't
2561 * been aggressively reclaimed enough yet.
2563 if (penalty_jiffies
<= HZ
/ 100)
2567 * If reclaim is making forward progress but we're still over
2568 * memory.high, we want to encourage that rather than doing allocator
2571 if (nr_reclaimed
|| nr_retries
--) {
2577 * If we exit early, we're guaranteed to die (since
2578 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2579 * need to account for any ill-begotten jiffies to pay them off later.
2581 psi_memstall_enter(&pflags
);
2582 schedule_timeout_killable(penalty_jiffies
);
2583 psi_memstall_leave(&pflags
);
2586 css_put(&memcg
->css
);
2589 static int try_charge_memcg(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2590 unsigned int nr_pages
)
2592 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2593 int nr_retries
= MAX_RECLAIM_RETRIES
;
2594 struct mem_cgroup
*mem_over_limit
;
2595 struct page_counter
*counter
;
2596 enum oom_status oom_status
;
2597 unsigned long nr_reclaimed
;
2598 bool passed_oom
= false;
2599 bool may_swap
= true;
2600 bool drained
= false;
2601 unsigned long pflags
;
2604 if (consume_stock(memcg
, nr_pages
))
2607 if (!do_memsw_account() ||
2608 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2609 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2611 if (do_memsw_account())
2612 page_counter_uncharge(&memcg
->memsw
, batch
);
2613 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2615 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2619 if (batch
> nr_pages
) {
2625 * Memcg doesn't have a dedicated reserve for atomic
2626 * allocations. But like the global atomic pool, we need to
2627 * put the burden of reclaim on regular allocation requests
2628 * and let these go through as privileged allocations.
2630 if (gfp_mask
& __GFP_ATOMIC
)
2634 * Prevent unbounded recursion when reclaim operations need to
2635 * allocate memory. This might exceed the limits temporarily,
2636 * but we prefer facilitating memory reclaim and getting back
2637 * under the limit over triggering OOM kills in these cases.
2639 if (unlikely(current
->flags
& PF_MEMALLOC
))
2642 if (unlikely(task_in_memcg_oom(current
)))
2645 if (!gfpflags_allow_blocking(gfp_mask
))
2648 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2650 psi_memstall_enter(&pflags
);
2651 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2652 gfp_mask
, may_swap
);
2653 psi_memstall_leave(&pflags
);
2655 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2659 drain_all_stock(mem_over_limit
);
2664 if (gfp_mask
& __GFP_NORETRY
)
2667 * Even though the limit is exceeded at this point, reclaim
2668 * may have been able to free some pages. Retry the charge
2669 * before killing the task.
2671 * Only for regular pages, though: huge pages are rather
2672 * unlikely to succeed so close to the limit, and we fall back
2673 * to regular pages anyway in case of failure.
2675 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2678 * At task move, charge accounts can be doubly counted. So, it's
2679 * better to wait until the end of task_move if something is going on.
2681 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2687 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2690 /* Avoid endless loop for tasks bypassed by the oom killer */
2691 if (passed_oom
&& task_is_dying())
2695 * keep retrying as long as the memcg oom killer is able to make
2696 * a forward progress or bypass the charge if the oom killer
2697 * couldn't make any progress.
2699 oom_status
= mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2700 get_order(nr_pages
* PAGE_SIZE
));
2701 if (oom_status
== OOM_SUCCESS
) {
2703 nr_retries
= MAX_RECLAIM_RETRIES
;
2707 if (!(gfp_mask
& __GFP_NOFAIL
))
2711 * The allocation either can't fail or will lead to more memory
2712 * being freed very soon. Allow memory usage go over the limit
2713 * temporarily by force charging it.
2715 page_counter_charge(&memcg
->memory
, nr_pages
);
2716 if (do_memsw_account())
2717 page_counter_charge(&memcg
->memsw
, nr_pages
);
2722 if (batch
> nr_pages
)
2723 refill_stock(memcg
, batch
- nr_pages
);
2726 * If the hierarchy is above the normal consumption range, schedule
2727 * reclaim on returning to userland. We can perform reclaim here
2728 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2729 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2730 * not recorded as it most likely matches current's and won't
2731 * change in the meantime. As high limit is checked again before
2732 * reclaim, the cost of mismatch is negligible.
2735 bool mem_high
, swap_high
;
2737 mem_high
= page_counter_read(&memcg
->memory
) >
2738 READ_ONCE(memcg
->memory
.high
);
2739 swap_high
= page_counter_read(&memcg
->swap
) >
2740 READ_ONCE(memcg
->swap
.high
);
2742 /* Don't bother a random interrupted task */
2743 if (in_interrupt()) {
2745 schedule_work(&memcg
->high_work
);
2751 if (mem_high
|| swap_high
) {
2753 * The allocating tasks in this cgroup will need to do
2754 * reclaim or be throttled to prevent further growth
2755 * of the memory or swap footprints.
2757 * Target some best-effort fairness between the tasks,
2758 * and distribute reclaim work and delay penalties
2759 * based on how much each task is actually allocating.
2761 current
->memcg_nr_pages_over_high
+= batch
;
2762 set_notify_resume(current
);
2765 } while ((memcg
= parent_mem_cgroup(memcg
)));
2770 static inline int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2771 unsigned int nr_pages
)
2773 if (mem_cgroup_is_root(memcg
))
2776 return try_charge_memcg(memcg
, gfp_mask
, nr_pages
);
2779 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2780 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2782 if (mem_cgroup_is_root(memcg
))
2785 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2786 if (do_memsw_account())
2787 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2791 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
)
2793 VM_BUG_ON_PAGE(page_memcg(page
), page
);
2795 * Any of the following ensures page's memcg stability:
2799 * - lock_page_memcg()
2800 * - exclusive reference
2802 page
->memcg_data
= (unsigned long)memcg
;
2805 static struct mem_cgroup
*get_mem_cgroup_from_objcg(struct obj_cgroup
*objcg
)
2807 struct mem_cgroup
*memcg
;
2811 memcg
= obj_cgroup_memcg(objcg
);
2812 if (unlikely(!css_tryget(&memcg
->css
)))
2819 #ifdef CONFIG_MEMCG_KMEM
2821 * The allocated objcg pointers array is not accounted directly.
2822 * Moreover, it should not come from DMA buffer and is not readily
2823 * reclaimable. So those GFP bits should be masked off.
2825 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2827 int memcg_alloc_page_obj_cgroups(struct page
*page
, struct kmem_cache
*s
,
2828 gfp_t gfp
, bool new_page
)
2830 unsigned int objects
= objs_per_slab_page(s
, page
);
2831 unsigned long memcg_data
;
2834 gfp
&= ~OBJCGS_CLEAR_MASK
;
2835 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2840 memcg_data
= (unsigned long) vec
| MEMCG_DATA_OBJCGS
;
2843 * If the slab page is brand new and nobody can yet access
2844 * it's memcg_data, no synchronization is required and
2845 * memcg_data can be simply assigned.
2847 page
->memcg_data
= memcg_data
;
2848 } else if (cmpxchg(&page
->memcg_data
, 0, memcg_data
)) {
2850 * If the slab page is already in use, somebody can allocate
2851 * and assign obj_cgroups in parallel. In this case the existing
2852 * objcg vector should be reused.
2858 kmemleak_not_leak(vec
);
2863 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2865 * A passed kernel object can be a slab object or a generic kernel page, so
2866 * different mechanisms for getting the memory cgroup pointer should be used.
2867 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2868 * can not know for sure how the kernel object is implemented.
2869 * mem_cgroup_from_obj() can be safely used in such cases.
2871 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2872 * cgroup_mutex, etc.
2874 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
2878 if (mem_cgroup_disabled())
2881 page
= virt_to_head_page(p
);
2884 * Slab objects are accounted individually, not per-page.
2885 * Memcg membership data for each individual object is saved in
2886 * the page->obj_cgroups.
2888 if (page_objcgs_check(page
)) {
2889 struct obj_cgroup
*objcg
;
2892 off
= obj_to_index(page
->slab_cache
, page
, p
);
2893 objcg
= page_objcgs(page
)[off
];
2895 return obj_cgroup_memcg(objcg
);
2901 * page_memcg_check() is used here, because page_has_obj_cgroups()
2902 * check above could fail because the object cgroups vector wasn't set
2903 * at that moment, but it can be set concurrently.
2904 * page_memcg_check(page) will guarantee that a proper memory
2905 * cgroup pointer or NULL will be returned.
2907 return page_memcg_check(page
);
2910 __always_inline
struct obj_cgroup
*get_obj_cgroup_from_current(void)
2912 struct obj_cgroup
*objcg
= NULL
;
2913 struct mem_cgroup
*memcg
;
2915 if (memcg_kmem_bypass())
2919 if (unlikely(active_memcg()))
2920 memcg
= active_memcg();
2922 memcg
= mem_cgroup_from_task(current
);
2924 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
2925 objcg
= rcu_dereference(memcg
->objcg
);
2926 if (objcg
&& obj_cgroup_tryget(objcg
))
2935 static int memcg_alloc_cache_id(void)
2940 id
= ida_simple_get(&memcg_cache_ida
,
2941 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2945 if (id
< memcg_nr_cache_ids
)
2949 * There's no space for the new id in memcg_caches arrays,
2950 * so we have to grow them.
2952 down_write(&memcg_cache_ids_sem
);
2954 size
= 2 * (id
+ 1);
2955 if (size
< MEMCG_CACHES_MIN_SIZE
)
2956 size
= MEMCG_CACHES_MIN_SIZE
;
2957 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2958 size
= MEMCG_CACHES_MAX_SIZE
;
2960 err
= memcg_update_all_list_lrus(size
);
2962 memcg_nr_cache_ids
= size
;
2964 up_write(&memcg_cache_ids_sem
);
2967 ida_simple_remove(&memcg_cache_ida
, id
);
2973 static void memcg_free_cache_id(int id
)
2975 ida_simple_remove(&memcg_cache_ida
, id
);
2979 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
2980 * @objcg: object cgroup to uncharge
2981 * @nr_pages: number of pages to uncharge
2983 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
2984 unsigned int nr_pages
)
2986 struct mem_cgroup
*memcg
;
2988 memcg
= get_mem_cgroup_from_objcg(objcg
);
2990 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2991 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2992 refill_stock(memcg
, nr_pages
);
2994 css_put(&memcg
->css
);
2998 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
2999 * @objcg: object cgroup to charge
3000 * @gfp: reclaim mode
3001 * @nr_pages: number of pages to charge
3003 * Returns 0 on success, an error code on failure.
3005 static int obj_cgroup_charge_pages(struct obj_cgroup
*objcg
, gfp_t gfp
,
3006 unsigned int nr_pages
)
3008 struct page_counter
*counter
;
3009 struct mem_cgroup
*memcg
;
3012 memcg
= get_mem_cgroup_from_objcg(objcg
);
3014 ret
= try_charge_memcg(memcg
, gfp
, nr_pages
);
3018 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
3019 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
3022 * Enforce __GFP_NOFAIL allocation because callers are not
3023 * prepared to see failures and likely do not have any failure
3026 if (gfp
& __GFP_NOFAIL
) {
3027 page_counter_charge(&memcg
->kmem
, nr_pages
);
3030 cancel_charge(memcg
, nr_pages
);
3034 css_put(&memcg
->css
);
3040 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3041 * @page: page to charge
3042 * @gfp: reclaim mode
3043 * @order: allocation order
3045 * Returns 0 on success, an error code on failure.
3047 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3049 struct obj_cgroup
*objcg
;
3052 objcg
= get_obj_cgroup_from_current();
3054 ret
= obj_cgroup_charge_pages(objcg
, gfp
, 1 << order
);
3056 page
->memcg_data
= (unsigned long)objcg
|
3060 obj_cgroup_put(objcg
);
3066 * __memcg_kmem_uncharge_page: uncharge a kmem page
3067 * @page: page to uncharge
3068 * @order: allocation order
3070 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3072 struct obj_cgroup
*objcg
;
3073 unsigned int nr_pages
= 1 << order
;
3075 if (!PageMemcgKmem(page
))
3078 objcg
= __page_objcg(page
);
3079 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3080 page
->memcg_data
= 0;
3081 obj_cgroup_put(objcg
);
3084 void mod_objcg_state(struct obj_cgroup
*objcg
, struct pglist_data
*pgdat
,
3085 enum node_stat_item idx
, int nr
)
3087 unsigned long flags
;
3088 struct obj_stock
*stock
= get_obj_stock(&flags
);
3092 * Save vmstat data in stock and skip vmstat array update unless
3093 * accumulating over a page of vmstat data or when pgdat or idx
3096 if (stock
->cached_objcg
!= objcg
) {
3097 drain_obj_stock(stock
);
3098 obj_cgroup_get(objcg
);
3099 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3100 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3101 stock
->cached_objcg
= objcg
;
3102 stock
->cached_pgdat
= pgdat
;
3103 } else if (stock
->cached_pgdat
!= pgdat
) {
3104 /* Flush the existing cached vmstat data */
3105 struct pglist_data
*oldpg
= stock
->cached_pgdat
;
3107 if (stock
->nr_slab_reclaimable_b
) {
3108 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_RECLAIMABLE_B
,
3109 stock
->nr_slab_reclaimable_b
);
3110 stock
->nr_slab_reclaimable_b
= 0;
3112 if (stock
->nr_slab_unreclaimable_b
) {
3113 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_UNRECLAIMABLE_B
,
3114 stock
->nr_slab_unreclaimable_b
);
3115 stock
->nr_slab_unreclaimable_b
= 0;
3117 stock
->cached_pgdat
= pgdat
;
3120 bytes
= (idx
== NR_SLAB_RECLAIMABLE_B
) ? &stock
->nr_slab_reclaimable_b
3121 : &stock
->nr_slab_unreclaimable_b
;
3123 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3124 * cached locally at least once before pushing it out.
3131 if (abs(*bytes
) > PAGE_SIZE
) {
3139 mod_objcg_mlstate(objcg
, pgdat
, idx
, nr
);
3141 put_obj_stock(flags
);
3144 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3146 unsigned long flags
;
3147 struct obj_stock
*stock
= get_obj_stock(&flags
);
3150 if (objcg
== stock
->cached_objcg
&& stock
->nr_bytes
>= nr_bytes
) {
3151 stock
->nr_bytes
-= nr_bytes
;
3155 put_obj_stock(flags
);
3160 static void drain_obj_stock(struct obj_stock
*stock
)
3162 struct obj_cgroup
*old
= stock
->cached_objcg
;
3167 if (stock
->nr_bytes
) {
3168 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3169 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3172 obj_cgroup_uncharge_pages(old
, nr_pages
);
3175 * The leftover is flushed to the centralized per-memcg value.
3176 * On the next attempt to refill obj stock it will be moved
3177 * to a per-cpu stock (probably, on an other CPU), see
3178 * refill_obj_stock().
3180 * How often it's flushed is a trade-off between the memory
3181 * limit enforcement accuracy and potential CPU contention,
3182 * so it might be changed in the future.
3184 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3185 stock
->nr_bytes
= 0;
3189 * Flush the vmstat data in current stock
3191 if (stock
->nr_slab_reclaimable_b
|| stock
->nr_slab_unreclaimable_b
) {
3192 if (stock
->nr_slab_reclaimable_b
) {
3193 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3194 NR_SLAB_RECLAIMABLE_B
,
3195 stock
->nr_slab_reclaimable_b
);
3196 stock
->nr_slab_reclaimable_b
= 0;
3198 if (stock
->nr_slab_unreclaimable_b
) {
3199 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3200 NR_SLAB_UNRECLAIMABLE_B
,
3201 stock
->nr_slab_unreclaimable_b
);
3202 stock
->nr_slab_unreclaimable_b
= 0;
3204 stock
->cached_pgdat
= NULL
;
3207 obj_cgroup_put(old
);
3208 stock
->cached_objcg
= NULL
;
3211 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3212 struct mem_cgroup
*root_memcg
)
3214 struct mem_cgroup
*memcg
;
3216 if (in_task() && stock
->task_obj
.cached_objcg
) {
3217 memcg
= obj_cgroup_memcg(stock
->task_obj
.cached_objcg
);
3218 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3221 if (stock
->irq_obj
.cached_objcg
) {
3222 memcg
= obj_cgroup_memcg(stock
->irq_obj
.cached_objcg
);
3223 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3230 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
,
3231 bool allow_uncharge
)
3233 unsigned long flags
;
3234 struct obj_stock
*stock
= get_obj_stock(&flags
);
3235 unsigned int nr_pages
= 0;
3237 if (stock
->cached_objcg
!= objcg
) { /* reset if necessary */
3238 drain_obj_stock(stock
);
3239 obj_cgroup_get(objcg
);
3240 stock
->cached_objcg
= objcg
;
3241 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3242 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3243 allow_uncharge
= true; /* Allow uncharge when objcg changes */
3245 stock
->nr_bytes
+= nr_bytes
;
3247 if (allow_uncharge
&& (stock
->nr_bytes
> PAGE_SIZE
)) {
3248 nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3249 stock
->nr_bytes
&= (PAGE_SIZE
- 1);
3252 put_obj_stock(flags
);
3255 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3258 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3260 unsigned int nr_pages
, nr_bytes
;
3263 if (consume_obj_stock(objcg
, size
))
3267 * In theory, objcg->nr_charged_bytes can have enough
3268 * pre-charged bytes to satisfy the allocation. However,
3269 * flushing objcg->nr_charged_bytes requires two atomic
3270 * operations, and objcg->nr_charged_bytes can't be big.
3271 * The shared objcg->nr_charged_bytes can also become a
3272 * performance bottleneck if all tasks of the same memcg are
3273 * trying to update it. So it's better to ignore it and try
3274 * grab some new pages. The stock's nr_bytes will be flushed to
3275 * objcg->nr_charged_bytes later on when objcg changes.
3277 * The stock's nr_bytes may contain enough pre-charged bytes
3278 * to allow one less page from being charged, but we can't rely
3279 * on the pre-charged bytes not being changed outside of
3280 * consume_obj_stock() or refill_obj_stock(). So ignore those
3281 * pre-charged bytes as well when charging pages. To avoid a
3282 * page uncharge right after a page charge, we set the
3283 * allow_uncharge flag to false when calling refill_obj_stock()
3284 * to temporarily allow the pre-charged bytes to exceed the page
3285 * size limit. The maximum reachable value of the pre-charged
3286 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3289 nr_pages
= size
>> PAGE_SHIFT
;
3290 nr_bytes
= size
& (PAGE_SIZE
- 1);
3295 ret
= obj_cgroup_charge_pages(objcg
, gfp
, nr_pages
);
3296 if (!ret
&& nr_bytes
)
3297 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
, false);
3302 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3304 refill_obj_stock(objcg
, size
, true);
3307 #endif /* CONFIG_MEMCG_KMEM */
3310 * Because page_memcg(head) is not set on tails, set it now.
3312 void split_page_memcg(struct page
*head
, unsigned int nr
)
3314 struct mem_cgroup
*memcg
= page_memcg(head
);
3317 if (mem_cgroup_disabled() || !memcg
)
3320 for (i
= 1; i
< nr
; i
++)
3321 head
[i
].memcg_data
= head
->memcg_data
;
3323 if (PageMemcgKmem(head
))
3324 obj_cgroup_get_many(__page_objcg(head
), nr
- 1);
3326 css_get_many(&memcg
->css
, nr
- 1);
3329 #ifdef CONFIG_MEMCG_SWAP
3331 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3332 * @entry: swap entry to be moved
3333 * @from: mem_cgroup which the entry is moved from
3334 * @to: mem_cgroup which the entry is moved to
3336 * It succeeds only when the swap_cgroup's record for this entry is the same
3337 * as the mem_cgroup's id of @from.
3339 * Returns 0 on success, -EINVAL on failure.
3341 * The caller must have charged to @to, IOW, called page_counter_charge() about
3342 * both res and memsw, and called css_get().
3344 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3345 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3347 unsigned short old_id
, new_id
;
3349 old_id
= mem_cgroup_id(from
);
3350 new_id
= mem_cgroup_id(to
);
3352 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3353 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3354 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3360 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3361 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3367 static DEFINE_MUTEX(memcg_max_mutex
);
3369 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3370 unsigned long max
, bool memsw
)
3372 bool enlarge
= false;
3373 bool drained
= false;
3375 bool limits_invariant
;
3376 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3379 if (signal_pending(current
)) {
3384 mutex_lock(&memcg_max_mutex
);
3386 * Make sure that the new limit (memsw or memory limit) doesn't
3387 * break our basic invariant rule memory.max <= memsw.max.
3389 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3390 max
<= memcg
->memsw
.max
;
3391 if (!limits_invariant
) {
3392 mutex_unlock(&memcg_max_mutex
);
3396 if (max
> counter
->max
)
3398 ret
= page_counter_set_max(counter
, max
);
3399 mutex_unlock(&memcg_max_mutex
);
3405 drain_all_stock(memcg
);
3410 if (!try_to_free_mem_cgroup_pages(memcg
, 1,
3411 GFP_KERNEL
, !memsw
)) {
3417 if (!ret
&& enlarge
)
3418 memcg_oom_recover(memcg
);
3423 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3425 unsigned long *total_scanned
)
3427 unsigned long nr_reclaimed
= 0;
3428 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3429 unsigned long reclaimed
;
3431 struct mem_cgroup_tree_per_node
*mctz
;
3432 unsigned long excess
;
3433 unsigned long nr_scanned
;
3438 mctz
= soft_limit_tree_node(pgdat
->node_id
);
3441 * Do not even bother to check the largest node if the root
3442 * is empty. Do it lockless to prevent lock bouncing. Races
3443 * are acceptable as soft limit is best effort anyway.
3445 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3449 * This loop can run a while, specially if mem_cgroup's continuously
3450 * keep exceeding their soft limit and putting the system under
3457 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3462 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3463 gfp_mask
, &nr_scanned
);
3464 nr_reclaimed
+= reclaimed
;
3465 *total_scanned
+= nr_scanned
;
3466 spin_lock_irq(&mctz
->lock
);
3467 __mem_cgroup_remove_exceeded(mz
, mctz
);
3470 * If we failed to reclaim anything from this memory cgroup
3471 * it is time to move on to the next cgroup
3475 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3477 excess
= soft_limit_excess(mz
->memcg
);
3479 * One school of thought says that we should not add
3480 * back the node to the tree if reclaim returns 0.
3481 * But our reclaim could return 0, simply because due
3482 * to priority we are exposing a smaller subset of
3483 * memory to reclaim from. Consider this as a longer
3486 /* If excess == 0, no tree ops */
3487 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3488 spin_unlock_irq(&mctz
->lock
);
3489 css_put(&mz
->memcg
->css
);
3492 * Could not reclaim anything and there are no more
3493 * mem cgroups to try or we seem to be looping without
3494 * reclaiming anything.
3496 if (!nr_reclaimed
&&
3498 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3500 } while (!nr_reclaimed
);
3502 css_put(&next_mz
->memcg
->css
);
3503 return nr_reclaimed
;
3507 * Reclaims as many pages from the given memcg as possible.
3509 * Caller is responsible for holding css reference for memcg.
3511 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3513 int nr_retries
= MAX_RECLAIM_RETRIES
;
3515 /* we call try-to-free pages for make this cgroup empty */
3516 lru_add_drain_all();
3518 drain_all_stock(memcg
);
3520 /* try to free all pages in this cgroup */
3521 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3524 if (signal_pending(current
))
3527 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3531 /* maybe some writeback is necessary */
3532 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3540 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3541 char *buf
, size_t nbytes
,
3544 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3546 if (mem_cgroup_is_root(memcg
))
3548 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3551 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3557 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3558 struct cftype
*cft
, u64 val
)
3563 pr_warn_once("Non-hierarchical mode is deprecated. "
3564 "Please report your usecase to linux-mm@kvack.org if you "
3565 "depend on this functionality.\n");
3570 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3574 if (mem_cgroup_is_root(memcg
)) {
3575 mem_cgroup_flush_stats();
3576 val
= memcg_page_state(memcg
, NR_FILE_PAGES
) +
3577 memcg_page_state(memcg
, NR_ANON_MAPPED
);
3579 val
+= memcg_page_state(memcg
, MEMCG_SWAP
);
3582 val
= page_counter_read(&memcg
->memory
);
3584 val
= page_counter_read(&memcg
->memsw
);
3597 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3600 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3601 struct page_counter
*counter
;
3603 switch (MEMFILE_TYPE(cft
->private)) {
3605 counter
= &memcg
->memory
;
3608 counter
= &memcg
->memsw
;
3611 counter
= &memcg
->kmem
;
3614 counter
= &memcg
->tcpmem
;
3620 switch (MEMFILE_ATTR(cft
->private)) {
3622 if (counter
== &memcg
->memory
)
3623 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3624 if (counter
== &memcg
->memsw
)
3625 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3626 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3628 return (u64
)counter
->max
* PAGE_SIZE
;
3630 return (u64
)counter
->watermark
* PAGE_SIZE
;
3632 return counter
->failcnt
;
3633 case RES_SOFT_LIMIT
:
3634 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3640 #ifdef CONFIG_MEMCG_KMEM
3641 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3643 struct obj_cgroup
*objcg
;
3646 if (cgroup_memory_nokmem
)
3649 BUG_ON(memcg
->kmemcg_id
>= 0);
3650 BUG_ON(memcg
->kmem_state
);
3652 memcg_id
= memcg_alloc_cache_id();
3656 objcg
= obj_cgroup_alloc();
3658 memcg_free_cache_id(memcg_id
);
3661 objcg
->memcg
= memcg
;
3662 rcu_assign_pointer(memcg
->objcg
, objcg
);
3664 static_branch_enable(&memcg_kmem_enabled_key
);
3666 memcg
->kmemcg_id
= memcg_id
;
3667 memcg
->kmem_state
= KMEM_ONLINE
;
3672 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3674 struct cgroup_subsys_state
*css
;
3675 struct mem_cgroup
*parent
, *child
;
3678 if (memcg
->kmem_state
!= KMEM_ONLINE
)
3681 memcg
->kmem_state
= KMEM_ALLOCATED
;
3683 parent
= parent_mem_cgroup(memcg
);
3685 parent
= root_mem_cgroup
;
3687 memcg_reparent_objcgs(memcg
, parent
);
3689 kmemcg_id
= memcg
->kmemcg_id
;
3690 BUG_ON(kmemcg_id
< 0);
3693 * Change kmemcg_id of this cgroup and all its descendants to the
3694 * parent's id, and then move all entries from this cgroup's list_lrus
3695 * to ones of the parent. After we have finished, all list_lrus
3696 * corresponding to this cgroup are guaranteed to remain empty. The
3697 * ordering is imposed by list_lru_node->lock taken by
3698 * memcg_drain_all_list_lrus().
3700 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3701 css_for_each_descendant_pre(css
, &memcg
->css
) {
3702 child
= mem_cgroup_from_css(css
);
3703 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3704 child
->kmemcg_id
= parent
->kmemcg_id
;
3708 memcg_drain_all_list_lrus(kmemcg_id
, parent
);
3710 memcg_free_cache_id(kmemcg_id
);
3713 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3715 /* css_alloc() failed, offlining didn't happen */
3716 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
3717 memcg_offline_kmem(memcg
);
3720 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3724 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3727 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
3730 #endif /* CONFIG_MEMCG_KMEM */
3732 static int memcg_update_kmem_max(struct mem_cgroup
*memcg
,
3737 mutex_lock(&memcg_max_mutex
);
3738 ret
= page_counter_set_max(&memcg
->kmem
, max
);
3739 mutex_unlock(&memcg_max_mutex
);
3743 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3747 mutex_lock(&memcg_max_mutex
);
3749 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3753 if (!memcg
->tcpmem_active
) {
3755 * The active flag needs to be written after the static_key
3756 * update. This is what guarantees that the socket activation
3757 * function is the last one to run. See mem_cgroup_sk_alloc()
3758 * for details, and note that we don't mark any socket as
3759 * belonging to this memcg until that flag is up.
3761 * We need to do this, because static_keys will span multiple
3762 * sites, but we can't control their order. If we mark a socket
3763 * as accounted, but the accounting functions are not patched in
3764 * yet, we'll lose accounting.
3766 * We never race with the readers in mem_cgroup_sk_alloc(),
3767 * because when this value change, the code to process it is not
3770 static_branch_inc(&memcg_sockets_enabled_key
);
3771 memcg
->tcpmem_active
= true;
3774 mutex_unlock(&memcg_max_mutex
);
3779 * The user of this function is...
3782 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3783 char *buf
, size_t nbytes
, loff_t off
)
3785 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3786 unsigned long nr_pages
;
3789 buf
= strstrip(buf
);
3790 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3794 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3796 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3800 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3802 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
3805 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
3808 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3809 "Please report your usecase to linux-mm@kvack.org if you "
3810 "depend on this functionality.\n");
3811 ret
= memcg_update_kmem_max(memcg
, nr_pages
);
3814 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
3818 case RES_SOFT_LIMIT
:
3819 memcg
->soft_limit
= nr_pages
;
3823 return ret
?: nbytes
;
3826 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3827 size_t nbytes
, loff_t off
)
3829 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3830 struct page_counter
*counter
;
3832 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3834 counter
= &memcg
->memory
;
3837 counter
= &memcg
->memsw
;
3840 counter
= &memcg
->kmem
;
3843 counter
= &memcg
->tcpmem
;
3849 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3851 page_counter_reset_watermark(counter
);
3854 counter
->failcnt
= 0;
3863 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3866 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3870 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3871 struct cftype
*cft
, u64 val
)
3873 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3875 if (val
& ~MOVE_MASK
)
3879 * No kind of locking is needed in here, because ->can_attach() will
3880 * check this value once in the beginning of the process, and then carry
3881 * on with stale data. This means that changes to this value will only
3882 * affect task migrations starting after the change.
3884 memcg
->move_charge_at_immigrate
= val
;
3888 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3889 struct cftype
*cft
, u64 val
)
3897 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3898 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3899 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3901 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
3902 int nid
, unsigned int lru_mask
, bool tree
)
3904 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
3905 unsigned long nr
= 0;
3908 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
3911 if (!(BIT(lru
) & lru_mask
))
3914 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
3916 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
3921 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
3922 unsigned int lru_mask
,
3925 unsigned long nr
= 0;
3929 if (!(BIT(lru
) & lru_mask
))
3932 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
3934 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
3939 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3943 unsigned int lru_mask
;
3946 static const struct numa_stat stats
[] = {
3947 { "total", LRU_ALL
},
3948 { "file", LRU_ALL_FILE
},
3949 { "anon", LRU_ALL_ANON
},
3950 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3952 const struct numa_stat
*stat
;
3954 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
3956 mem_cgroup_flush_stats();
3958 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3959 seq_printf(m
, "%s=%lu", stat
->name
,
3960 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3962 for_each_node_state(nid
, N_MEMORY
)
3963 seq_printf(m
, " N%d=%lu", nid
,
3964 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3965 stat
->lru_mask
, false));
3969 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3971 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
3972 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
3974 for_each_node_state(nid
, N_MEMORY
)
3975 seq_printf(m
, " N%d=%lu", nid
,
3976 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3977 stat
->lru_mask
, true));
3983 #endif /* CONFIG_NUMA */
3985 static const unsigned int memcg1_stats
[] = {
3988 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3998 static const char *const memcg1_stat_names
[] = {
4001 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4011 /* Universal VM events cgroup1 shows, original sort order */
4012 static const unsigned int memcg1_events
[] = {
4019 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4021 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4022 unsigned long memory
, memsw
;
4023 struct mem_cgroup
*mi
;
4026 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4028 mem_cgroup_flush_stats();
4030 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4033 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4035 nr
= memcg_page_state_local(memcg
, memcg1_stats
[i
]);
4036 seq_printf(m
, "%s %lu\n", memcg1_stat_names
[i
], nr
* PAGE_SIZE
);
4039 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4040 seq_printf(m
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4041 memcg_events_local(memcg
, memcg1_events
[i
]));
4043 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4044 seq_printf(m
, "%s %lu\n", lru_list_name(i
),
4045 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4048 /* Hierarchical information */
4049 memory
= memsw
= PAGE_COUNTER_MAX
;
4050 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4051 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4052 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4054 seq_printf(m
, "hierarchical_memory_limit %llu\n",
4055 (u64
)memory
* PAGE_SIZE
);
4056 if (do_memsw_account())
4057 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4058 (u64
)memsw
* PAGE_SIZE
);
4060 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4063 if (memcg1_stats
[i
] == MEMCG_SWAP
&& !do_memsw_account())
4065 nr
= memcg_page_state(memcg
, memcg1_stats
[i
]);
4066 seq_printf(m
, "total_%s %llu\n", memcg1_stat_names
[i
],
4067 (u64
)nr
* PAGE_SIZE
);
4070 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4071 seq_printf(m
, "total_%s %llu\n",
4072 vm_event_name(memcg1_events
[i
]),
4073 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4075 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4076 seq_printf(m
, "total_%s %llu\n", lru_list_name(i
),
4077 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4080 #ifdef CONFIG_DEBUG_VM
4083 struct mem_cgroup_per_node
*mz
;
4084 unsigned long anon_cost
= 0;
4085 unsigned long file_cost
= 0;
4087 for_each_online_pgdat(pgdat
) {
4088 mz
= memcg
->nodeinfo
[pgdat
->node_id
];
4090 anon_cost
+= mz
->lruvec
.anon_cost
;
4091 file_cost
+= mz
->lruvec
.file_cost
;
4093 seq_printf(m
, "anon_cost %lu\n", anon_cost
);
4094 seq_printf(m
, "file_cost %lu\n", file_cost
);
4101 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4104 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4106 return mem_cgroup_swappiness(memcg
);
4109 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4110 struct cftype
*cft
, u64 val
)
4112 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4117 if (!mem_cgroup_is_root(memcg
))
4118 memcg
->swappiness
= val
;
4120 vm_swappiness
= val
;
4125 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4127 struct mem_cgroup_threshold_ary
*t
;
4128 unsigned long usage
;
4133 t
= rcu_dereference(memcg
->thresholds
.primary
);
4135 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4140 usage
= mem_cgroup_usage(memcg
, swap
);
4143 * current_threshold points to threshold just below or equal to usage.
4144 * If it's not true, a threshold was crossed after last
4145 * call of __mem_cgroup_threshold().
4147 i
= t
->current_threshold
;
4150 * Iterate backward over array of thresholds starting from
4151 * current_threshold and check if a threshold is crossed.
4152 * If none of thresholds below usage is crossed, we read
4153 * only one element of the array here.
4155 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4156 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4158 /* i = current_threshold + 1 */
4162 * Iterate forward over array of thresholds starting from
4163 * current_threshold+1 and check if a threshold is crossed.
4164 * If none of thresholds above usage is crossed, we read
4165 * only one element of the array here.
4167 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4168 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4170 /* Update current_threshold */
4171 t
->current_threshold
= i
- 1;
4176 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4179 __mem_cgroup_threshold(memcg
, false);
4180 if (do_memsw_account())
4181 __mem_cgroup_threshold(memcg
, true);
4183 memcg
= parent_mem_cgroup(memcg
);
4187 static int compare_thresholds(const void *a
, const void *b
)
4189 const struct mem_cgroup_threshold
*_a
= a
;
4190 const struct mem_cgroup_threshold
*_b
= b
;
4192 if (_a
->threshold
> _b
->threshold
)
4195 if (_a
->threshold
< _b
->threshold
)
4201 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4203 struct mem_cgroup_eventfd_list
*ev
;
4205 spin_lock(&memcg_oom_lock
);
4207 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4208 eventfd_signal(ev
->eventfd
, 1);
4210 spin_unlock(&memcg_oom_lock
);
4214 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4216 struct mem_cgroup
*iter
;
4218 for_each_mem_cgroup_tree(iter
, memcg
)
4219 mem_cgroup_oom_notify_cb(iter
);
4222 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4223 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4225 struct mem_cgroup_thresholds
*thresholds
;
4226 struct mem_cgroup_threshold_ary
*new;
4227 unsigned long threshold
;
4228 unsigned long usage
;
4231 ret
= page_counter_memparse(args
, "-1", &threshold
);
4235 mutex_lock(&memcg
->thresholds_lock
);
4238 thresholds
= &memcg
->thresholds
;
4239 usage
= mem_cgroup_usage(memcg
, false);
4240 } else if (type
== _MEMSWAP
) {
4241 thresholds
= &memcg
->memsw_thresholds
;
4242 usage
= mem_cgroup_usage(memcg
, true);
4246 /* Check if a threshold crossed before adding a new one */
4247 if (thresholds
->primary
)
4248 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4250 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4252 /* Allocate memory for new array of thresholds */
4253 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4260 /* Copy thresholds (if any) to new array */
4261 if (thresholds
->primary
)
4262 memcpy(new->entries
, thresholds
->primary
->entries
,
4263 flex_array_size(new, entries
, size
- 1));
4265 /* Add new threshold */
4266 new->entries
[size
- 1].eventfd
= eventfd
;
4267 new->entries
[size
- 1].threshold
= threshold
;
4269 /* Sort thresholds. Registering of new threshold isn't time-critical */
4270 sort(new->entries
, size
, sizeof(*new->entries
),
4271 compare_thresholds
, NULL
);
4273 /* Find current threshold */
4274 new->current_threshold
= -1;
4275 for (i
= 0; i
< size
; i
++) {
4276 if (new->entries
[i
].threshold
<= usage
) {
4278 * new->current_threshold will not be used until
4279 * rcu_assign_pointer(), so it's safe to increment
4282 ++new->current_threshold
;
4287 /* Free old spare buffer and save old primary buffer as spare */
4288 kfree(thresholds
->spare
);
4289 thresholds
->spare
= thresholds
->primary
;
4291 rcu_assign_pointer(thresholds
->primary
, new);
4293 /* To be sure that nobody uses thresholds */
4297 mutex_unlock(&memcg
->thresholds_lock
);
4302 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4303 struct eventfd_ctx
*eventfd
, const char *args
)
4305 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4308 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4309 struct eventfd_ctx
*eventfd
, const char *args
)
4311 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4314 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4315 struct eventfd_ctx
*eventfd
, enum res_type type
)
4317 struct mem_cgroup_thresholds
*thresholds
;
4318 struct mem_cgroup_threshold_ary
*new;
4319 unsigned long usage
;
4320 int i
, j
, size
, entries
;
4322 mutex_lock(&memcg
->thresholds_lock
);
4325 thresholds
= &memcg
->thresholds
;
4326 usage
= mem_cgroup_usage(memcg
, false);
4327 } else if (type
== _MEMSWAP
) {
4328 thresholds
= &memcg
->memsw_thresholds
;
4329 usage
= mem_cgroup_usage(memcg
, true);
4333 if (!thresholds
->primary
)
4336 /* Check if a threshold crossed before removing */
4337 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4339 /* Calculate new number of threshold */
4341 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4342 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4348 new = thresholds
->spare
;
4350 /* If no items related to eventfd have been cleared, nothing to do */
4354 /* Set thresholds array to NULL if we don't have thresholds */
4363 /* Copy thresholds and find current threshold */
4364 new->current_threshold
= -1;
4365 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4366 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4369 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4370 if (new->entries
[j
].threshold
<= usage
) {
4372 * new->current_threshold will not be used
4373 * until rcu_assign_pointer(), so it's safe to increment
4376 ++new->current_threshold
;
4382 /* Swap primary and spare array */
4383 thresholds
->spare
= thresholds
->primary
;
4385 rcu_assign_pointer(thresholds
->primary
, new);
4387 /* To be sure that nobody uses thresholds */
4390 /* If all events are unregistered, free the spare array */
4392 kfree(thresholds
->spare
);
4393 thresholds
->spare
= NULL
;
4396 mutex_unlock(&memcg
->thresholds_lock
);
4399 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4400 struct eventfd_ctx
*eventfd
)
4402 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4405 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4406 struct eventfd_ctx
*eventfd
)
4408 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4411 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4412 struct eventfd_ctx
*eventfd
, const char *args
)
4414 struct mem_cgroup_eventfd_list
*event
;
4416 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4420 spin_lock(&memcg_oom_lock
);
4422 event
->eventfd
= eventfd
;
4423 list_add(&event
->list
, &memcg
->oom_notify
);
4425 /* already in OOM ? */
4426 if (memcg
->under_oom
)
4427 eventfd_signal(eventfd
, 1);
4428 spin_unlock(&memcg_oom_lock
);
4433 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4434 struct eventfd_ctx
*eventfd
)
4436 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4438 spin_lock(&memcg_oom_lock
);
4440 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4441 if (ev
->eventfd
== eventfd
) {
4442 list_del(&ev
->list
);
4447 spin_unlock(&memcg_oom_lock
);
4450 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4452 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4454 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4455 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4456 seq_printf(sf
, "oom_kill %lu\n",
4457 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4461 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4462 struct cftype
*cft
, u64 val
)
4464 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4466 /* cannot set to root cgroup and only 0 and 1 are allowed */
4467 if (mem_cgroup_is_root(memcg
) || !((val
== 0) || (val
== 1)))
4470 memcg
->oom_kill_disable
= val
;
4472 memcg_oom_recover(memcg
);
4477 #ifdef CONFIG_CGROUP_WRITEBACK
4479 #include <trace/events/writeback.h>
4481 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4483 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4486 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4488 wb_domain_exit(&memcg
->cgwb_domain
);
4491 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4493 wb_domain_size_changed(&memcg
->cgwb_domain
);
4496 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4498 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4500 if (!memcg
->css
.parent
)
4503 return &memcg
->cgwb_domain
;
4507 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4508 * @wb: bdi_writeback in question
4509 * @pfilepages: out parameter for number of file pages
4510 * @pheadroom: out parameter for number of allocatable pages according to memcg
4511 * @pdirty: out parameter for number of dirty pages
4512 * @pwriteback: out parameter for number of pages under writeback
4514 * Determine the numbers of file, headroom, dirty, and writeback pages in
4515 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4516 * is a bit more involved.
4518 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4519 * headroom is calculated as the lowest headroom of itself and the
4520 * ancestors. Note that this doesn't consider the actual amount of
4521 * available memory in the system. The caller should further cap
4522 * *@pheadroom accordingly.
4524 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4525 unsigned long *pheadroom
, unsigned long *pdirty
,
4526 unsigned long *pwriteback
)
4528 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4529 struct mem_cgroup
*parent
;
4531 mem_cgroup_flush_stats();
4533 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
4534 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
4535 *pfilepages
= memcg_page_state(memcg
, NR_INACTIVE_FILE
) +
4536 memcg_page_state(memcg
, NR_ACTIVE_FILE
);
4538 *pheadroom
= PAGE_COUNTER_MAX
;
4539 while ((parent
= parent_mem_cgroup(memcg
))) {
4540 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4541 READ_ONCE(memcg
->memory
.high
));
4542 unsigned long used
= page_counter_read(&memcg
->memory
);
4544 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4550 * Foreign dirty flushing
4552 * There's an inherent mismatch between memcg and writeback. The former
4553 * tracks ownership per-page while the latter per-inode. This was a
4554 * deliberate design decision because honoring per-page ownership in the
4555 * writeback path is complicated, may lead to higher CPU and IO overheads
4556 * and deemed unnecessary given that write-sharing an inode across
4557 * different cgroups isn't a common use-case.
4559 * Combined with inode majority-writer ownership switching, this works well
4560 * enough in most cases but there are some pathological cases. For
4561 * example, let's say there are two cgroups A and B which keep writing to
4562 * different but confined parts of the same inode. B owns the inode and
4563 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4564 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4565 * triggering background writeback. A will be slowed down without a way to
4566 * make writeback of the dirty pages happen.
4568 * Conditions like the above can lead to a cgroup getting repeatedly and
4569 * severely throttled after making some progress after each
4570 * dirty_expire_interval while the underlying IO device is almost
4573 * Solving this problem completely requires matching the ownership tracking
4574 * granularities between memcg and writeback in either direction. However,
4575 * the more egregious behaviors can be avoided by simply remembering the
4576 * most recent foreign dirtying events and initiating remote flushes on
4577 * them when local writeback isn't enough to keep the memory clean enough.
4579 * The following two functions implement such mechanism. When a foreign
4580 * page - a page whose memcg and writeback ownerships don't match - is
4581 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4582 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4583 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4584 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4585 * foreign bdi_writebacks which haven't expired. Both the numbers of
4586 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4587 * limited to MEMCG_CGWB_FRN_CNT.
4589 * The mechanism only remembers IDs and doesn't hold any object references.
4590 * As being wrong occasionally doesn't matter, updates and accesses to the
4591 * records are lockless and racy.
4593 void mem_cgroup_track_foreign_dirty_slowpath(struct page
*page
,
4594 struct bdi_writeback
*wb
)
4596 struct mem_cgroup
*memcg
= page_memcg(page
);
4597 struct memcg_cgwb_frn
*frn
;
4598 u64 now
= get_jiffies_64();
4599 u64 oldest_at
= now
;
4603 trace_track_foreign_dirty(page
, wb
);
4606 * Pick the slot to use. If there is already a slot for @wb, keep
4607 * using it. If not replace the oldest one which isn't being
4610 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4611 frn
= &memcg
->cgwb_frn
[i
];
4612 if (frn
->bdi_id
== wb
->bdi
->id
&&
4613 frn
->memcg_id
== wb
->memcg_css
->id
)
4615 if (time_before64(frn
->at
, oldest_at
) &&
4616 atomic_read(&frn
->done
.cnt
) == 1) {
4618 oldest_at
= frn
->at
;
4622 if (i
< MEMCG_CGWB_FRN_CNT
) {
4624 * Re-using an existing one. Update timestamp lazily to
4625 * avoid making the cacheline hot. We want them to be
4626 * reasonably up-to-date and significantly shorter than
4627 * dirty_expire_interval as that's what expires the record.
4628 * Use the shorter of 1s and dirty_expire_interval / 8.
4630 unsigned long update_intv
=
4631 min_t(unsigned long, HZ
,
4632 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4634 if (time_before64(frn
->at
, now
- update_intv
))
4636 } else if (oldest
>= 0) {
4637 /* replace the oldest free one */
4638 frn
= &memcg
->cgwb_frn
[oldest
];
4639 frn
->bdi_id
= wb
->bdi
->id
;
4640 frn
->memcg_id
= wb
->memcg_css
->id
;
4645 /* issue foreign writeback flushes for recorded foreign dirtying events */
4646 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4648 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4649 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4650 u64 now
= jiffies_64
;
4653 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4654 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4657 * If the record is older than dirty_expire_interval,
4658 * writeback on it has already started. No need to kick it
4659 * off again. Also, don't start a new one if there's
4660 * already one in flight.
4662 if (time_after64(frn
->at
, now
- intv
) &&
4663 atomic_read(&frn
->done
.cnt
) == 1) {
4665 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4666 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
,
4667 WB_REASON_FOREIGN_FLUSH
,
4673 #else /* CONFIG_CGROUP_WRITEBACK */
4675 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4680 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4684 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4688 #endif /* CONFIG_CGROUP_WRITEBACK */
4691 * DO NOT USE IN NEW FILES.
4693 * "cgroup.event_control" implementation.
4695 * This is way over-engineered. It tries to support fully configurable
4696 * events for each user. Such level of flexibility is completely
4697 * unnecessary especially in the light of the planned unified hierarchy.
4699 * Please deprecate this and replace with something simpler if at all
4704 * Unregister event and free resources.
4706 * Gets called from workqueue.
4708 static void memcg_event_remove(struct work_struct
*work
)
4710 struct mem_cgroup_event
*event
=
4711 container_of(work
, struct mem_cgroup_event
, remove
);
4712 struct mem_cgroup
*memcg
= event
->memcg
;
4714 remove_wait_queue(event
->wqh
, &event
->wait
);
4716 event
->unregister_event(memcg
, event
->eventfd
);
4718 /* Notify userspace the event is going away. */
4719 eventfd_signal(event
->eventfd
, 1);
4721 eventfd_ctx_put(event
->eventfd
);
4723 css_put(&memcg
->css
);
4727 * Gets called on EPOLLHUP on eventfd when user closes it.
4729 * Called with wqh->lock held and interrupts disabled.
4731 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4732 int sync
, void *key
)
4734 struct mem_cgroup_event
*event
=
4735 container_of(wait
, struct mem_cgroup_event
, wait
);
4736 struct mem_cgroup
*memcg
= event
->memcg
;
4737 __poll_t flags
= key_to_poll(key
);
4739 if (flags
& EPOLLHUP
) {
4741 * If the event has been detached at cgroup removal, we
4742 * can simply return knowing the other side will cleanup
4745 * We can't race against event freeing since the other
4746 * side will require wqh->lock via remove_wait_queue(),
4749 spin_lock(&memcg
->event_list_lock
);
4750 if (!list_empty(&event
->list
)) {
4751 list_del_init(&event
->list
);
4753 * We are in atomic context, but cgroup_event_remove()
4754 * may sleep, so we have to call it in workqueue.
4756 schedule_work(&event
->remove
);
4758 spin_unlock(&memcg
->event_list_lock
);
4764 static void memcg_event_ptable_queue_proc(struct file
*file
,
4765 wait_queue_head_t
*wqh
, poll_table
*pt
)
4767 struct mem_cgroup_event
*event
=
4768 container_of(pt
, struct mem_cgroup_event
, pt
);
4771 add_wait_queue(wqh
, &event
->wait
);
4775 * DO NOT USE IN NEW FILES.
4777 * Parse input and register new cgroup event handler.
4779 * Input must be in format '<event_fd> <control_fd> <args>'.
4780 * Interpretation of args is defined by control file implementation.
4782 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4783 char *buf
, size_t nbytes
, loff_t off
)
4785 struct cgroup_subsys_state
*css
= of_css(of
);
4786 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4787 struct mem_cgroup_event
*event
;
4788 struct cgroup_subsys_state
*cfile_css
;
4789 unsigned int efd
, cfd
;
4796 buf
= strstrip(buf
);
4798 efd
= simple_strtoul(buf
, &endp
, 10);
4803 cfd
= simple_strtoul(buf
, &endp
, 10);
4804 if ((*endp
!= ' ') && (*endp
!= '\0'))
4808 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4812 event
->memcg
= memcg
;
4813 INIT_LIST_HEAD(&event
->list
);
4814 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4815 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4816 INIT_WORK(&event
->remove
, memcg_event_remove
);
4824 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4825 if (IS_ERR(event
->eventfd
)) {
4826 ret
= PTR_ERR(event
->eventfd
);
4833 goto out_put_eventfd
;
4836 /* the process need read permission on control file */
4837 /* AV: shouldn't we check that it's been opened for read instead? */
4838 ret
= file_permission(cfile
.file
, MAY_READ
);
4843 * Determine the event callbacks and set them in @event. This used
4844 * to be done via struct cftype but cgroup core no longer knows
4845 * about these events. The following is crude but the whole thing
4846 * is for compatibility anyway.
4848 * DO NOT ADD NEW FILES.
4850 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4852 if (!strcmp(name
, "memory.usage_in_bytes")) {
4853 event
->register_event
= mem_cgroup_usage_register_event
;
4854 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4855 } else if (!strcmp(name
, "memory.oom_control")) {
4856 event
->register_event
= mem_cgroup_oom_register_event
;
4857 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4858 } else if (!strcmp(name
, "memory.pressure_level")) {
4859 event
->register_event
= vmpressure_register_event
;
4860 event
->unregister_event
= vmpressure_unregister_event
;
4861 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4862 event
->register_event
= memsw_cgroup_usage_register_event
;
4863 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4870 * Verify @cfile should belong to @css. Also, remaining events are
4871 * automatically removed on cgroup destruction but the removal is
4872 * asynchronous, so take an extra ref on @css.
4874 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4875 &memory_cgrp_subsys
);
4877 if (IS_ERR(cfile_css
))
4879 if (cfile_css
!= css
) {
4884 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4888 vfs_poll(efile
.file
, &event
->pt
);
4890 spin_lock_irq(&memcg
->event_list_lock
);
4891 list_add(&event
->list
, &memcg
->event_list
);
4892 spin_unlock_irq(&memcg
->event_list_lock
);
4904 eventfd_ctx_put(event
->eventfd
);
4913 static struct cftype mem_cgroup_legacy_files
[] = {
4915 .name
= "usage_in_bytes",
4916 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4917 .read_u64
= mem_cgroup_read_u64
,
4920 .name
= "max_usage_in_bytes",
4921 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4922 .write
= mem_cgroup_reset
,
4923 .read_u64
= mem_cgroup_read_u64
,
4926 .name
= "limit_in_bytes",
4927 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4928 .write
= mem_cgroup_write
,
4929 .read_u64
= mem_cgroup_read_u64
,
4932 .name
= "soft_limit_in_bytes",
4933 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4934 .write
= mem_cgroup_write
,
4935 .read_u64
= mem_cgroup_read_u64
,
4939 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4940 .write
= mem_cgroup_reset
,
4941 .read_u64
= mem_cgroup_read_u64
,
4945 .seq_show
= memcg_stat_show
,
4948 .name
= "force_empty",
4949 .write
= mem_cgroup_force_empty_write
,
4952 .name
= "use_hierarchy",
4953 .write_u64
= mem_cgroup_hierarchy_write
,
4954 .read_u64
= mem_cgroup_hierarchy_read
,
4957 .name
= "cgroup.event_control", /* XXX: for compat */
4958 .write
= memcg_write_event_control
,
4959 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
4962 .name
= "swappiness",
4963 .read_u64
= mem_cgroup_swappiness_read
,
4964 .write_u64
= mem_cgroup_swappiness_write
,
4967 .name
= "move_charge_at_immigrate",
4968 .read_u64
= mem_cgroup_move_charge_read
,
4969 .write_u64
= mem_cgroup_move_charge_write
,
4972 .name
= "oom_control",
4973 .seq_show
= mem_cgroup_oom_control_read
,
4974 .write_u64
= mem_cgroup_oom_control_write
,
4975 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4978 .name
= "pressure_level",
4982 .name
= "numa_stat",
4983 .seq_show
= memcg_numa_stat_show
,
4987 .name
= "kmem.limit_in_bytes",
4988 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4989 .write
= mem_cgroup_write
,
4990 .read_u64
= mem_cgroup_read_u64
,
4993 .name
= "kmem.usage_in_bytes",
4994 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4995 .read_u64
= mem_cgroup_read_u64
,
4998 .name
= "kmem.failcnt",
4999 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5000 .write
= mem_cgroup_reset
,
5001 .read_u64
= mem_cgroup_read_u64
,
5004 .name
= "kmem.max_usage_in_bytes",
5005 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5006 .write
= mem_cgroup_reset
,
5007 .read_u64
= mem_cgroup_read_u64
,
5009 #if defined(CONFIG_MEMCG_KMEM) && \
5010 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5012 .name
= "kmem.slabinfo",
5013 .seq_show
= memcg_slab_show
,
5017 .name
= "kmem.tcp.limit_in_bytes",
5018 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5019 .write
= mem_cgroup_write
,
5020 .read_u64
= mem_cgroup_read_u64
,
5023 .name
= "kmem.tcp.usage_in_bytes",
5024 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5025 .read_u64
= mem_cgroup_read_u64
,
5028 .name
= "kmem.tcp.failcnt",
5029 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5030 .write
= mem_cgroup_reset
,
5031 .read_u64
= mem_cgroup_read_u64
,
5034 .name
= "kmem.tcp.max_usage_in_bytes",
5035 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5036 .write
= mem_cgroup_reset
,
5037 .read_u64
= mem_cgroup_read_u64
,
5039 { }, /* terminate */
5043 * Private memory cgroup IDR
5045 * Swap-out records and page cache shadow entries need to store memcg
5046 * references in constrained space, so we maintain an ID space that is
5047 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5048 * memory-controlled cgroups to 64k.
5050 * However, there usually are many references to the offline CSS after
5051 * the cgroup has been destroyed, such as page cache or reclaimable
5052 * slab objects, that don't need to hang on to the ID. We want to keep
5053 * those dead CSS from occupying IDs, or we might quickly exhaust the
5054 * relatively small ID space and prevent the creation of new cgroups
5055 * even when there are much fewer than 64k cgroups - possibly none.
5057 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5058 * be freed and recycled when it's no longer needed, which is usually
5059 * when the CSS is offlined.
5061 * The only exception to that are records of swapped out tmpfs/shmem
5062 * pages that need to be attributed to live ancestors on swapin. But
5063 * those references are manageable from userspace.
5066 static DEFINE_IDR(mem_cgroup_idr
);
5068 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5070 if (memcg
->id
.id
> 0) {
5071 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5076 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5079 refcount_add(n
, &memcg
->id
.ref
);
5082 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5084 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5085 mem_cgroup_id_remove(memcg
);
5087 /* Memcg ID pins CSS */
5088 css_put(&memcg
->css
);
5092 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5094 mem_cgroup_id_put_many(memcg
, 1);
5098 * mem_cgroup_from_id - look up a memcg from a memcg id
5099 * @id: the memcg id to look up
5101 * Caller must hold rcu_read_lock().
5103 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5105 WARN_ON_ONCE(!rcu_read_lock_held());
5106 return idr_find(&mem_cgroup_idr
, id
);
5109 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5111 struct mem_cgroup_per_node
*pn
;
5114 * This routine is called against possible nodes.
5115 * But it's BUG to call kmalloc() against offline node.
5117 * TODO: this routine can waste much memory for nodes which will
5118 * never be onlined. It's better to use memory hotplug callback
5121 if (!node_state(node
, N_NORMAL_MEMORY
))
5123 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5127 pn
->lruvec_stats_percpu
= alloc_percpu_gfp(struct lruvec_stats_percpu
,
5128 GFP_KERNEL_ACCOUNT
);
5129 if (!pn
->lruvec_stats_percpu
) {
5134 lruvec_init(&pn
->lruvec
);
5135 pn
->usage_in_excess
= 0;
5136 pn
->on_tree
= false;
5139 memcg
->nodeinfo
[node
] = pn
;
5143 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5145 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5150 free_percpu(pn
->lruvec_stats_percpu
);
5154 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5159 free_mem_cgroup_per_node_info(memcg
, node
);
5160 free_percpu(memcg
->vmstats_percpu
);
5164 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5166 memcg_wb_domain_exit(memcg
);
5167 __mem_cgroup_free(memcg
);
5170 static struct mem_cgroup
*mem_cgroup_alloc(void)
5172 struct mem_cgroup
*memcg
;
5175 int __maybe_unused i
;
5176 long error
= -ENOMEM
;
5178 size
= sizeof(struct mem_cgroup
);
5179 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5181 memcg
= kzalloc(size
, GFP_KERNEL
);
5183 return ERR_PTR(error
);
5185 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5186 1, MEM_CGROUP_ID_MAX
,
5188 if (memcg
->id
.id
< 0) {
5189 error
= memcg
->id
.id
;
5193 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5194 GFP_KERNEL_ACCOUNT
);
5195 if (!memcg
->vmstats_percpu
)
5199 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5202 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5205 INIT_WORK(&memcg
->high_work
, high_work_func
);
5206 INIT_LIST_HEAD(&memcg
->oom_notify
);
5207 mutex_init(&memcg
->thresholds_lock
);
5208 spin_lock_init(&memcg
->move_lock
);
5209 vmpressure_init(&memcg
->vmpressure
);
5210 INIT_LIST_HEAD(&memcg
->event_list
);
5211 spin_lock_init(&memcg
->event_list_lock
);
5212 memcg
->socket_pressure
= jiffies
;
5213 #ifdef CONFIG_MEMCG_KMEM
5214 memcg
->kmemcg_id
= -1;
5215 INIT_LIST_HEAD(&memcg
->objcg_list
);
5217 #ifdef CONFIG_CGROUP_WRITEBACK
5218 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5219 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5220 memcg
->cgwb_frn
[i
].done
=
5221 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5223 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5224 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5225 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5226 memcg
->deferred_split_queue
.split_queue_len
= 0;
5228 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5231 mem_cgroup_id_remove(memcg
);
5232 __mem_cgroup_free(memcg
);
5233 return ERR_PTR(error
);
5236 static struct cgroup_subsys_state
* __ref
5237 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5239 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5240 struct mem_cgroup
*memcg
, *old_memcg
;
5241 long error
= -ENOMEM
;
5243 old_memcg
= set_active_memcg(parent
);
5244 memcg
= mem_cgroup_alloc();
5245 set_active_memcg(old_memcg
);
5247 return ERR_CAST(memcg
);
5249 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5250 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5251 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5253 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5254 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5256 page_counter_init(&memcg
->memory
, &parent
->memory
);
5257 page_counter_init(&memcg
->swap
, &parent
->swap
);
5258 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5259 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5261 page_counter_init(&memcg
->memory
, NULL
);
5262 page_counter_init(&memcg
->swap
, NULL
);
5263 page_counter_init(&memcg
->kmem
, NULL
);
5264 page_counter_init(&memcg
->tcpmem
, NULL
);
5266 root_mem_cgroup
= memcg
;
5270 /* The following stuff does not apply to the root */
5271 error
= memcg_online_kmem(memcg
);
5275 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5276 static_branch_inc(&memcg_sockets_enabled_key
);
5280 mem_cgroup_id_remove(memcg
);
5281 mem_cgroup_free(memcg
);
5282 return ERR_PTR(error
);
5285 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5287 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5290 * A memcg must be visible for expand_shrinker_info()
5291 * by the time the maps are allocated. So, we allocate maps
5292 * here, when for_each_mem_cgroup() can't skip it.
5294 if (alloc_shrinker_info(memcg
)) {
5295 mem_cgroup_id_remove(memcg
);
5299 /* Online state pins memcg ID, memcg ID pins CSS */
5300 refcount_set(&memcg
->id
.ref
, 1);
5303 if (unlikely(mem_cgroup_is_root(memcg
)))
5304 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
,
5309 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5311 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5312 struct mem_cgroup_event
*event
, *tmp
;
5315 * Unregister events and notify userspace.
5316 * Notify userspace about cgroup removing only after rmdir of cgroup
5317 * directory to avoid race between userspace and kernelspace.
5319 spin_lock_irq(&memcg
->event_list_lock
);
5320 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5321 list_del_init(&event
->list
);
5322 schedule_work(&event
->remove
);
5324 spin_unlock_irq(&memcg
->event_list_lock
);
5326 page_counter_set_min(&memcg
->memory
, 0);
5327 page_counter_set_low(&memcg
->memory
, 0);
5329 memcg_offline_kmem(memcg
);
5330 reparent_shrinker_deferred(memcg
);
5331 wb_memcg_offline(memcg
);
5333 drain_all_stock(memcg
);
5335 mem_cgroup_id_put(memcg
);
5338 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5340 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5342 invalidate_reclaim_iterators(memcg
);
5345 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5347 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5348 int __maybe_unused i
;
5350 #ifdef CONFIG_CGROUP_WRITEBACK
5351 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5352 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5354 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5355 static_branch_dec(&memcg_sockets_enabled_key
);
5357 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5358 static_branch_dec(&memcg_sockets_enabled_key
);
5360 vmpressure_cleanup(&memcg
->vmpressure
);
5361 cancel_work_sync(&memcg
->high_work
);
5362 mem_cgroup_remove_from_trees(memcg
);
5363 free_shrinker_info(memcg
);
5364 memcg_free_kmem(memcg
);
5365 mem_cgroup_free(memcg
);
5369 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5370 * @css: the target css
5372 * Reset the states of the mem_cgroup associated with @css. This is
5373 * invoked when the userland requests disabling on the default hierarchy
5374 * but the memcg is pinned through dependency. The memcg should stop
5375 * applying policies and should revert to the vanilla state as it may be
5376 * made visible again.
5378 * The current implementation only resets the essential configurations.
5379 * This needs to be expanded to cover all the visible parts.
5381 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5383 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5385 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5386 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5387 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5388 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5389 page_counter_set_min(&memcg
->memory
, 0);
5390 page_counter_set_low(&memcg
->memory
, 0);
5391 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5392 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
5393 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5394 memcg_wb_domain_size_changed(memcg
);
5397 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
5399 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5400 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5401 struct memcg_vmstats_percpu
*statc
;
5405 statc
= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
);
5407 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
5409 * Collect the aggregated propagation counts of groups
5410 * below us. We're in a per-cpu loop here and this is
5411 * a global counter, so the first cycle will get them.
5413 delta
= memcg
->vmstats
.state_pending
[i
];
5415 memcg
->vmstats
.state_pending
[i
] = 0;
5417 /* Add CPU changes on this level since the last flush */
5418 v
= READ_ONCE(statc
->state
[i
]);
5419 if (v
!= statc
->state_prev
[i
]) {
5420 delta
+= v
- statc
->state_prev
[i
];
5421 statc
->state_prev
[i
] = v
;
5427 /* Aggregate counts on this level and propagate upwards */
5428 memcg
->vmstats
.state
[i
] += delta
;
5430 parent
->vmstats
.state_pending
[i
] += delta
;
5433 for (i
= 0; i
< NR_VM_EVENT_ITEMS
; i
++) {
5434 delta
= memcg
->vmstats
.events_pending
[i
];
5436 memcg
->vmstats
.events_pending
[i
] = 0;
5438 v
= READ_ONCE(statc
->events
[i
]);
5439 if (v
!= statc
->events_prev
[i
]) {
5440 delta
+= v
- statc
->events_prev
[i
];
5441 statc
->events_prev
[i
] = v
;
5447 memcg
->vmstats
.events
[i
] += delta
;
5449 parent
->vmstats
.events_pending
[i
] += delta
;
5452 for_each_node_state(nid
, N_MEMORY
) {
5453 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[nid
];
5454 struct mem_cgroup_per_node
*ppn
= NULL
;
5455 struct lruvec_stats_percpu
*lstatc
;
5458 ppn
= parent
->nodeinfo
[nid
];
5460 lstatc
= per_cpu_ptr(pn
->lruvec_stats_percpu
, cpu
);
5462 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
5463 delta
= pn
->lruvec_stats
.state_pending
[i
];
5465 pn
->lruvec_stats
.state_pending
[i
] = 0;
5467 v
= READ_ONCE(lstatc
->state
[i
]);
5468 if (v
!= lstatc
->state_prev
[i
]) {
5469 delta
+= v
- lstatc
->state_prev
[i
];
5470 lstatc
->state_prev
[i
] = v
;
5476 pn
->lruvec_stats
.state
[i
] += delta
;
5478 ppn
->lruvec_stats
.state_pending
[i
] += delta
;
5484 /* Handlers for move charge at task migration. */
5485 static int mem_cgroup_do_precharge(unsigned long count
)
5489 /* Try a single bulk charge without reclaim first, kswapd may wake */
5490 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5492 mc
.precharge
+= count
;
5496 /* Try charges one by one with reclaim, but do not retry */
5498 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5512 enum mc_target_type
{
5519 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5520 unsigned long addr
, pte_t ptent
)
5522 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5524 if (!page
|| !page_mapped(page
))
5526 if (PageAnon(page
)) {
5527 if (!(mc
.flags
& MOVE_ANON
))
5530 if (!(mc
.flags
& MOVE_FILE
))
5533 if (!get_page_unless_zero(page
))
5539 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5540 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5541 pte_t ptent
, swp_entry_t
*entry
)
5543 struct page
*page
= NULL
;
5544 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5546 if (!(mc
.flags
& MOVE_ANON
))
5550 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5551 * a device and because they are not accessible by CPU they are store
5552 * as special swap entry in the CPU page table.
5554 if (is_device_private_entry(ent
)) {
5555 page
= pfn_swap_entry_to_page(ent
);
5557 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5558 * a refcount of 1 when free (unlike normal page)
5560 if (!page_ref_add_unless(page
, 1, 1))
5565 if (non_swap_entry(ent
))
5569 * Because lookup_swap_cache() updates some statistics counter,
5570 * we call find_get_page() with swapper_space directly.
5572 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5573 entry
->val
= ent
.val
;
5578 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5579 pte_t ptent
, swp_entry_t
*entry
)
5585 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5586 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5588 if (!vma
->vm_file
) /* anonymous vma */
5590 if (!(mc
.flags
& MOVE_FILE
))
5593 /* page is moved even if it's not RSS of this task(page-faulted). */
5594 /* shmem/tmpfs may report page out on swap: account for that too. */
5595 return find_get_incore_page(vma
->vm_file
->f_mapping
,
5596 linear_page_index(vma
, addr
));
5600 * mem_cgroup_move_account - move account of the page
5602 * @compound: charge the page as compound or small page
5603 * @from: mem_cgroup which the page is moved from.
5604 * @to: mem_cgroup which the page is moved to. @from != @to.
5606 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5608 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5611 static int mem_cgroup_move_account(struct page
*page
,
5613 struct mem_cgroup
*from
,
5614 struct mem_cgroup
*to
)
5616 struct lruvec
*from_vec
, *to_vec
;
5617 struct pglist_data
*pgdat
;
5618 unsigned int nr_pages
= compound
? thp_nr_pages(page
) : 1;
5621 VM_BUG_ON(from
== to
);
5622 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5623 VM_BUG_ON(compound
&& !PageTransHuge(page
));
5626 * Prevent mem_cgroup_migrate() from looking at
5627 * page's memory cgroup of its source page while we change it.
5630 if (!trylock_page(page
))
5634 if (page_memcg(page
) != from
)
5637 pgdat
= page_pgdat(page
);
5638 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5639 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5641 lock_page_memcg(page
);
5643 if (PageAnon(page
)) {
5644 if (page_mapped(page
)) {
5645 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5646 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5647 if (PageTransHuge(page
)) {
5648 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5650 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5655 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5656 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5658 if (PageSwapBacked(page
)) {
5659 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5660 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5663 if (page_mapped(page
)) {
5664 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5665 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5668 if (PageDirty(page
)) {
5669 struct address_space
*mapping
= page_mapping(page
);
5671 if (mapping_can_writeback(mapping
)) {
5672 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5674 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5680 if (PageWriteback(page
)) {
5681 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5682 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5686 * All state has been migrated, let's switch to the new memcg.
5688 * It is safe to change page's memcg here because the page
5689 * is referenced, charged, isolated, and locked: we can't race
5690 * with (un)charging, migration, LRU putback, or anything else
5691 * that would rely on a stable page's memory cgroup.
5693 * Note that lock_page_memcg is a memcg lock, not a page lock,
5694 * to save space. As soon as we switch page's memory cgroup to a
5695 * new memcg that isn't locked, the above state can change
5696 * concurrently again. Make sure we're truly done with it.
5701 css_put(&from
->css
);
5703 page
->memcg_data
= (unsigned long)to
;
5705 __unlock_page_memcg(from
);
5709 local_irq_disable();
5710 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
5711 memcg_check_events(to
, page
);
5712 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
5713 memcg_check_events(from
, page
);
5722 * get_mctgt_type - get target type of moving charge
5723 * @vma: the vma the pte to be checked belongs
5724 * @addr: the address corresponding to the pte to be checked
5725 * @ptent: the pte to be checked
5726 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5729 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5730 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5731 * move charge. if @target is not NULL, the page is stored in target->page
5732 * with extra refcnt got(Callers should handle it).
5733 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5734 * target for charge migration. if @target is not NULL, the entry is stored
5736 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5737 * (so ZONE_DEVICE page and thus not on the lru).
5738 * For now we such page is charge like a regular page would be as for all
5739 * intent and purposes it is just special memory taking the place of a
5742 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5744 * Called with pte lock held.
5747 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5748 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5750 struct page
*page
= NULL
;
5751 enum mc_target_type ret
= MC_TARGET_NONE
;
5752 swp_entry_t ent
= { .val
= 0 };
5754 if (pte_present(ptent
))
5755 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5756 else if (is_swap_pte(ptent
))
5757 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
5758 else if (pte_none(ptent
))
5759 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5761 if (!page
&& !ent
.val
)
5765 * Do only loose check w/o serialization.
5766 * mem_cgroup_move_account() checks the page is valid or
5767 * not under LRU exclusion.
5769 if (page_memcg(page
) == mc
.from
) {
5770 ret
= MC_TARGET_PAGE
;
5771 if (is_device_private_page(page
))
5772 ret
= MC_TARGET_DEVICE
;
5774 target
->page
= page
;
5776 if (!ret
|| !target
)
5780 * There is a swap entry and a page doesn't exist or isn't charged.
5781 * But we cannot move a tail-page in a THP.
5783 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
5784 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5785 ret
= MC_TARGET_SWAP
;
5792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5794 * We don't consider PMD mapped swapping or file mapped pages because THP does
5795 * not support them for now.
5796 * Caller should make sure that pmd_trans_huge(pmd) is true.
5798 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5799 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5801 struct page
*page
= NULL
;
5802 enum mc_target_type ret
= MC_TARGET_NONE
;
5804 if (unlikely(is_swap_pmd(pmd
))) {
5805 VM_BUG_ON(thp_migration_supported() &&
5806 !is_pmd_migration_entry(pmd
));
5809 page
= pmd_page(pmd
);
5810 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5811 if (!(mc
.flags
& MOVE_ANON
))
5813 if (page_memcg(page
) == mc
.from
) {
5814 ret
= MC_TARGET_PAGE
;
5817 target
->page
= page
;
5823 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5824 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5826 return MC_TARGET_NONE
;
5830 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5831 unsigned long addr
, unsigned long end
,
5832 struct mm_walk
*walk
)
5834 struct vm_area_struct
*vma
= walk
->vma
;
5838 ptl
= pmd_trans_huge_lock(pmd
, vma
);
5841 * Note their can not be MC_TARGET_DEVICE for now as we do not
5842 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5843 * this might change.
5845 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5846 mc
.precharge
+= HPAGE_PMD_NR
;
5851 if (pmd_trans_unstable(pmd
))
5853 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5854 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5855 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5856 mc
.precharge
++; /* increment precharge temporarily */
5857 pte_unmap_unlock(pte
- 1, ptl
);
5863 static const struct mm_walk_ops precharge_walk_ops
= {
5864 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5867 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5869 unsigned long precharge
;
5872 walk_page_range(mm
, 0, mm
->highest_vm_end
, &precharge_walk_ops
, NULL
);
5873 mmap_read_unlock(mm
);
5875 precharge
= mc
.precharge
;
5881 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5883 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5885 VM_BUG_ON(mc
.moving_task
);
5886 mc
.moving_task
= current
;
5887 return mem_cgroup_do_precharge(precharge
);
5890 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5891 static void __mem_cgroup_clear_mc(void)
5893 struct mem_cgroup
*from
= mc
.from
;
5894 struct mem_cgroup
*to
= mc
.to
;
5896 /* we must uncharge all the leftover precharges from mc.to */
5898 cancel_charge(mc
.to
, mc
.precharge
);
5902 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5903 * we must uncharge here.
5905 if (mc
.moved_charge
) {
5906 cancel_charge(mc
.from
, mc
.moved_charge
);
5907 mc
.moved_charge
= 0;
5909 /* we must fixup refcnts and charges */
5910 if (mc
.moved_swap
) {
5911 /* uncharge swap account from the old cgroup */
5912 if (!mem_cgroup_is_root(mc
.from
))
5913 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5915 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
5918 * we charged both to->memory and to->memsw, so we
5919 * should uncharge to->memory.
5921 if (!mem_cgroup_is_root(mc
.to
))
5922 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5926 memcg_oom_recover(from
);
5927 memcg_oom_recover(to
);
5928 wake_up_all(&mc
.waitq
);
5931 static void mem_cgroup_clear_mc(void)
5933 struct mm_struct
*mm
= mc
.mm
;
5936 * we must clear moving_task before waking up waiters at the end of
5939 mc
.moving_task
= NULL
;
5940 __mem_cgroup_clear_mc();
5941 spin_lock(&mc
.lock
);
5945 spin_unlock(&mc
.lock
);
5950 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
5952 struct cgroup_subsys_state
*css
;
5953 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
5954 struct mem_cgroup
*from
;
5955 struct task_struct
*leader
, *p
;
5956 struct mm_struct
*mm
;
5957 unsigned long move_flags
;
5960 /* charge immigration isn't supported on the default hierarchy */
5961 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5965 * Multi-process migrations only happen on the default hierarchy
5966 * where charge immigration is not used. Perform charge
5967 * immigration if @tset contains a leader and whine if there are
5971 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
5974 memcg
= mem_cgroup_from_css(css
);
5980 * We are now committed to this value whatever it is. Changes in this
5981 * tunable will only affect upcoming migrations, not the current one.
5982 * So we need to save it, and keep it going.
5984 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5988 from
= mem_cgroup_from_task(p
);
5990 VM_BUG_ON(from
== memcg
);
5992 mm
= get_task_mm(p
);
5995 /* We move charges only when we move a owner of the mm */
5996 if (mm
->owner
== p
) {
5999 VM_BUG_ON(mc
.precharge
);
6000 VM_BUG_ON(mc
.moved_charge
);
6001 VM_BUG_ON(mc
.moved_swap
);
6003 spin_lock(&mc
.lock
);
6007 mc
.flags
= move_flags
;
6008 spin_unlock(&mc
.lock
);
6009 /* We set mc.moving_task later */
6011 ret
= mem_cgroup_precharge_mc(mm
);
6013 mem_cgroup_clear_mc();
6020 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6023 mem_cgroup_clear_mc();
6026 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6027 unsigned long addr
, unsigned long end
,
6028 struct mm_walk
*walk
)
6031 struct vm_area_struct
*vma
= walk
->vma
;
6034 enum mc_target_type target_type
;
6035 union mc_target target
;
6038 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6040 if (mc
.precharge
< HPAGE_PMD_NR
) {
6044 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6045 if (target_type
== MC_TARGET_PAGE
) {
6047 if (!isolate_lru_page(page
)) {
6048 if (!mem_cgroup_move_account(page
, true,
6050 mc
.precharge
-= HPAGE_PMD_NR
;
6051 mc
.moved_charge
+= HPAGE_PMD_NR
;
6053 putback_lru_page(page
);
6056 } else if (target_type
== MC_TARGET_DEVICE
) {
6058 if (!mem_cgroup_move_account(page
, true,
6060 mc
.precharge
-= HPAGE_PMD_NR
;
6061 mc
.moved_charge
+= HPAGE_PMD_NR
;
6069 if (pmd_trans_unstable(pmd
))
6072 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6073 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6074 pte_t ptent
= *(pte
++);
6075 bool device
= false;
6081 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6082 case MC_TARGET_DEVICE
:
6085 case MC_TARGET_PAGE
:
6088 * We can have a part of the split pmd here. Moving it
6089 * can be done but it would be too convoluted so simply
6090 * ignore such a partial THP and keep it in original
6091 * memcg. There should be somebody mapping the head.
6093 if (PageTransCompound(page
))
6095 if (!device
&& isolate_lru_page(page
))
6097 if (!mem_cgroup_move_account(page
, false,
6100 /* we uncharge from mc.from later. */
6104 putback_lru_page(page
);
6105 put
: /* get_mctgt_type() gets the page */
6108 case MC_TARGET_SWAP
:
6110 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6112 mem_cgroup_id_get_many(mc
.to
, 1);
6113 /* we fixup other refcnts and charges later. */
6121 pte_unmap_unlock(pte
- 1, ptl
);
6126 * We have consumed all precharges we got in can_attach().
6127 * We try charge one by one, but don't do any additional
6128 * charges to mc.to if we have failed in charge once in attach()
6131 ret
= mem_cgroup_do_precharge(1);
6139 static const struct mm_walk_ops charge_walk_ops
= {
6140 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6143 static void mem_cgroup_move_charge(void)
6145 lru_add_drain_all();
6147 * Signal lock_page_memcg() to take the memcg's move_lock
6148 * while we're moving its pages to another memcg. Then wait
6149 * for already started RCU-only updates to finish.
6151 atomic_inc(&mc
.from
->moving_account
);
6154 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6156 * Someone who are holding the mmap_lock might be waiting in
6157 * waitq. So we cancel all extra charges, wake up all waiters,
6158 * and retry. Because we cancel precharges, we might not be able
6159 * to move enough charges, but moving charge is a best-effort
6160 * feature anyway, so it wouldn't be a big problem.
6162 __mem_cgroup_clear_mc();
6167 * When we have consumed all precharges and failed in doing
6168 * additional charge, the page walk just aborts.
6170 walk_page_range(mc
.mm
, 0, mc
.mm
->highest_vm_end
, &charge_walk_ops
,
6173 mmap_read_unlock(mc
.mm
);
6174 atomic_dec(&mc
.from
->moving_account
);
6177 static void mem_cgroup_move_task(void)
6180 mem_cgroup_move_charge();
6181 mem_cgroup_clear_mc();
6184 #else /* !CONFIG_MMU */
6185 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6189 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6192 static void mem_cgroup_move_task(void)
6197 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6199 if (value
== PAGE_COUNTER_MAX
)
6200 seq_puts(m
, "max\n");
6202 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6207 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6210 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6212 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6215 static int memory_min_show(struct seq_file
*m
, void *v
)
6217 return seq_puts_memcg_tunable(m
,
6218 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6221 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6222 char *buf
, size_t nbytes
, loff_t off
)
6224 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6228 buf
= strstrip(buf
);
6229 err
= page_counter_memparse(buf
, "max", &min
);
6233 page_counter_set_min(&memcg
->memory
, min
);
6238 static int memory_low_show(struct seq_file
*m
, void *v
)
6240 return seq_puts_memcg_tunable(m
,
6241 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6244 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6245 char *buf
, size_t nbytes
, loff_t off
)
6247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6251 buf
= strstrip(buf
);
6252 err
= page_counter_memparse(buf
, "max", &low
);
6256 page_counter_set_low(&memcg
->memory
, low
);
6261 static int memory_high_show(struct seq_file
*m
, void *v
)
6263 return seq_puts_memcg_tunable(m
,
6264 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6267 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6268 char *buf
, size_t nbytes
, loff_t off
)
6270 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6271 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6272 bool drained
= false;
6276 buf
= strstrip(buf
);
6277 err
= page_counter_memparse(buf
, "max", &high
);
6281 page_counter_set_high(&memcg
->memory
, high
);
6284 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6285 unsigned long reclaimed
;
6287 if (nr_pages
<= high
)
6290 if (signal_pending(current
))
6294 drain_all_stock(memcg
);
6299 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6302 if (!reclaimed
&& !nr_retries
--)
6306 memcg_wb_domain_size_changed(memcg
);
6310 static int memory_max_show(struct seq_file
*m
, void *v
)
6312 return seq_puts_memcg_tunable(m
,
6313 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6316 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6317 char *buf
, size_t nbytes
, loff_t off
)
6319 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6320 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6321 bool drained
= false;
6325 buf
= strstrip(buf
);
6326 err
= page_counter_memparse(buf
, "max", &max
);
6330 xchg(&memcg
->memory
.max
, max
);
6333 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6335 if (nr_pages
<= max
)
6338 if (signal_pending(current
))
6342 drain_all_stock(memcg
);
6348 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6354 memcg_memory_event(memcg
, MEMCG_OOM
);
6355 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6359 memcg_wb_domain_size_changed(memcg
);
6363 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6365 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6366 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6367 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6368 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6369 seq_printf(m
, "oom_kill %lu\n",
6370 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6373 static int memory_events_show(struct seq_file
*m
, void *v
)
6375 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6377 __memory_events_show(m
, memcg
->memory_events
);
6381 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6383 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6385 __memory_events_show(m
, memcg
->memory_events_local
);
6389 static int memory_stat_show(struct seq_file
*m
, void *v
)
6391 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6394 buf
= memory_stat_format(memcg
);
6403 static inline unsigned long lruvec_page_state_output(struct lruvec
*lruvec
,
6406 return lruvec_page_state(lruvec
, item
) * memcg_page_state_unit(item
);
6409 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6412 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6414 mem_cgroup_flush_stats();
6416 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6419 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6422 seq_printf(m
, "%s", memory_stats
[i
].name
);
6423 for_each_node_state(nid
, N_MEMORY
) {
6425 struct lruvec
*lruvec
;
6427 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6428 size
= lruvec_page_state_output(lruvec
,
6429 memory_stats
[i
].idx
);
6430 seq_printf(m
, " N%d=%llu", nid
, size
);
6439 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6441 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6443 seq_printf(m
, "%d\n", memcg
->oom_group
);
6448 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6449 char *buf
, size_t nbytes
, loff_t off
)
6451 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6454 buf
= strstrip(buf
);
6458 ret
= kstrtoint(buf
, 0, &oom_group
);
6462 if (oom_group
!= 0 && oom_group
!= 1)
6465 memcg
->oom_group
= oom_group
;
6470 static struct cftype memory_files
[] = {
6473 .flags
= CFTYPE_NOT_ON_ROOT
,
6474 .read_u64
= memory_current_read
,
6478 .flags
= CFTYPE_NOT_ON_ROOT
,
6479 .seq_show
= memory_min_show
,
6480 .write
= memory_min_write
,
6484 .flags
= CFTYPE_NOT_ON_ROOT
,
6485 .seq_show
= memory_low_show
,
6486 .write
= memory_low_write
,
6490 .flags
= CFTYPE_NOT_ON_ROOT
,
6491 .seq_show
= memory_high_show
,
6492 .write
= memory_high_write
,
6496 .flags
= CFTYPE_NOT_ON_ROOT
,
6497 .seq_show
= memory_max_show
,
6498 .write
= memory_max_write
,
6502 .flags
= CFTYPE_NOT_ON_ROOT
,
6503 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6504 .seq_show
= memory_events_show
,
6507 .name
= "events.local",
6508 .flags
= CFTYPE_NOT_ON_ROOT
,
6509 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6510 .seq_show
= memory_events_local_show
,
6514 .seq_show
= memory_stat_show
,
6518 .name
= "numa_stat",
6519 .seq_show
= memory_numa_stat_show
,
6523 .name
= "oom.group",
6524 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6525 .seq_show
= memory_oom_group_show
,
6526 .write
= memory_oom_group_write
,
6531 struct cgroup_subsys memory_cgrp_subsys
= {
6532 .css_alloc
= mem_cgroup_css_alloc
,
6533 .css_online
= mem_cgroup_css_online
,
6534 .css_offline
= mem_cgroup_css_offline
,
6535 .css_released
= mem_cgroup_css_released
,
6536 .css_free
= mem_cgroup_css_free
,
6537 .css_reset
= mem_cgroup_css_reset
,
6538 .css_rstat_flush
= mem_cgroup_css_rstat_flush
,
6539 .can_attach
= mem_cgroup_can_attach
,
6540 .cancel_attach
= mem_cgroup_cancel_attach
,
6541 .post_attach
= mem_cgroup_move_task
,
6542 .dfl_cftypes
= memory_files
,
6543 .legacy_cftypes
= mem_cgroup_legacy_files
,
6548 * This function calculates an individual cgroup's effective
6549 * protection which is derived from its own memory.min/low, its
6550 * parent's and siblings' settings, as well as the actual memory
6551 * distribution in the tree.
6553 * The following rules apply to the effective protection values:
6555 * 1. At the first level of reclaim, effective protection is equal to
6556 * the declared protection in memory.min and memory.low.
6558 * 2. To enable safe delegation of the protection configuration, at
6559 * subsequent levels the effective protection is capped to the
6560 * parent's effective protection.
6562 * 3. To make complex and dynamic subtrees easier to configure, the
6563 * user is allowed to overcommit the declared protection at a given
6564 * level. If that is the case, the parent's effective protection is
6565 * distributed to the children in proportion to how much protection
6566 * they have declared and how much of it they are utilizing.
6568 * This makes distribution proportional, but also work-conserving:
6569 * if one cgroup claims much more protection than it uses memory,
6570 * the unused remainder is available to its siblings.
6572 * 4. Conversely, when the declared protection is undercommitted at a
6573 * given level, the distribution of the larger parental protection
6574 * budget is NOT proportional. A cgroup's protection from a sibling
6575 * is capped to its own memory.min/low setting.
6577 * 5. However, to allow protecting recursive subtrees from each other
6578 * without having to declare each individual cgroup's fixed share
6579 * of the ancestor's claim to protection, any unutilized -
6580 * "floating" - protection from up the tree is distributed in
6581 * proportion to each cgroup's *usage*. This makes the protection
6582 * neutral wrt sibling cgroups and lets them compete freely over
6583 * the shared parental protection budget, but it protects the
6584 * subtree as a whole from neighboring subtrees.
6586 * Note that 4. and 5. are not in conflict: 4. is about protecting
6587 * against immediate siblings whereas 5. is about protecting against
6588 * neighboring subtrees.
6590 static unsigned long effective_protection(unsigned long usage
,
6591 unsigned long parent_usage
,
6592 unsigned long setting
,
6593 unsigned long parent_effective
,
6594 unsigned long siblings_protected
)
6596 unsigned long protected;
6599 protected = min(usage
, setting
);
6601 * If all cgroups at this level combined claim and use more
6602 * protection then what the parent affords them, distribute
6603 * shares in proportion to utilization.
6605 * We are using actual utilization rather than the statically
6606 * claimed protection in order to be work-conserving: claimed
6607 * but unused protection is available to siblings that would
6608 * otherwise get a smaller chunk than what they claimed.
6610 if (siblings_protected
> parent_effective
)
6611 return protected * parent_effective
/ siblings_protected
;
6614 * Ok, utilized protection of all children is within what the
6615 * parent affords them, so we know whatever this child claims
6616 * and utilizes is effectively protected.
6618 * If there is unprotected usage beyond this value, reclaim
6619 * will apply pressure in proportion to that amount.
6621 * If there is unutilized protection, the cgroup will be fully
6622 * shielded from reclaim, but we do return a smaller value for
6623 * protection than what the group could enjoy in theory. This
6624 * is okay. With the overcommit distribution above, effective
6625 * protection is always dependent on how memory is actually
6626 * consumed among the siblings anyway.
6631 * If the children aren't claiming (all of) the protection
6632 * afforded to them by the parent, distribute the remainder in
6633 * proportion to the (unprotected) memory of each cgroup. That
6634 * way, cgroups that aren't explicitly prioritized wrt each
6635 * other compete freely over the allowance, but they are
6636 * collectively protected from neighboring trees.
6638 * We're using unprotected memory for the weight so that if
6639 * some cgroups DO claim explicit protection, we don't protect
6640 * the same bytes twice.
6642 * Check both usage and parent_usage against the respective
6643 * protected values. One should imply the other, but they
6644 * aren't read atomically - make sure the division is sane.
6646 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
6648 if (parent_effective
> siblings_protected
&&
6649 parent_usage
> siblings_protected
&&
6650 usage
> protected) {
6651 unsigned long unclaimed
;
6653 unclaimed
= parent_effective
- siblings_protected
;
6654 unclaimed
*= usage
- protected;
6655 unclaimed
/= parent_usage
- siblings_protected
;
6664 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6665 * @root: the top ancestor of the sub-tree being checked
6666 * @memcg: the memory cgroup to check
6668 * WARNING: This function is not stateless! It can only be used as part
6669 * of a top-down tree iteration, not for isolated queries.
6671 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
6672 struct mem_cgroup
*memcg
)
6674 unsigned long usage
, parent_usage
;
6675 struct mem_cgroup
*parent
;
6677 if (mem_cgroup_disabled())
6681 root
= root_mem_cgroup
;
6684 * Effective values of the reclaim targets are ignored so they
6685 * can be stale. Have a look at mem_cgroup_protection for more
6687 * TODO: calculation should be more robust so that we do not need
6688 * that special casing.
6693 usage
= page_counter_read(&memcg
->memory
);
6697 parent
= parent_mem_cgroup(memcg
);
6698 /* No parent means a non-hierarchical mode on v1 memcg */
6702 if (parent
== root
) {
6703 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
6704 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
6708 parent_usage
= page_counter_read(&parent
->memory
);
6710 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
6711 READ_ONCE(memcg
->memory
.min
),
6712 READ_ONCE(parent
->memory
.emin
),
6713 atomic_long_read(&parent
->memory
.children_min_usage
)));
6715 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
6716 READ_ONCE(memcg
->memory
.low
),
6717 READ_ONCE(parent
->memory
.elow
),
6718 atomic_long_read(&parent
->memory
.children_low_usage
)));
6721 static int charge_memcg(struct page
*page
, struct mem_cgroup
*memcg
, gfp_t gfp
)
6723 unsigned int nr_pages
= thp_nr_pages(page
);
6726 ret
= try_charge(memcg
, gfp
, nr_pages
);
6730 css_get(&memcg
->css
);
6731 commit_charge(page
, memcg
);
6733 local_irq_disable();
6734 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6735 memcg_check_events(memcg
, page
);
6742 * __mem_cgroup_charge - charge a newly allocated page to a cgroup
6743 * @page: page to charge
6744 * @mm: mm context of the victim
6745 * @gfp_mask: reclaim mode
6747 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6748 * pages according to @gfp_mask if necessary. if @mm is NULL, try to
6749 * charge to the active memcg.
6751 * Do not use this for pages allocated for swapin.
6753 * Returns 0 on success. Otherwise, an error code is returned.
6755 int __mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
6758 struct mem_cgroup
*memcg
;
6761 memcg
= get_mem_cgroup_from_mm(mm
);
6762 ret
= charge_memcg(page
, memcg
, gfp_mask
);
6763 css_put(&memcg
->css
);
6769 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6770 * @page: page to charge
6771 * @mm: mm context of the victim
6772 * @gfp: reclaim mode
6773 * @entry: swap entry for which the page is allocated
6775 * This function charges a page allocated for swapin. Please call this before
6776 * adding the page to the swapcache.
6778 * Returns 0 on success. Otherwise, an error code is returned.
6780 int mem_cgroup_swapin_charge_page(struct page
*page
, struct mm_struct
*mm
,
6781 gfp_t gfp
, swp_entry_t entry
)
6783 struct mem_cgroup
*memcg
;
6787 if (mem_cgroup_disabled())
6790 id
= lookup_swap_cgroup_id(entry
);
6792 memcg
= mem_cgroup_from_id(id
);
6793 if (!memcg
|| !css_tryget_online(&memcg
->css
))
6794 memcg
= get_mem_cgroup_from_mm(mm
);
6797 ret
= charge_memcg(page
, memcg
, gfp
);
6799 css_put(&memcg
->css
);
6804 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6805 * @entry: swap entry for which the page is charged
6807 * Call this function after successfully adding the charged page to swapcache.
6809 * Note: This function assumes the page for which swap slot is being uncharged
6812 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry
)
6815 * Cgroup1's unified memory+swap counter has been charged with the
6816 * new swapcache page, finish the transfer by uncharging the swap
6817 * slot. The swap slot would also get uncharged when it dies, but
6818 * it can stick around indefinitely and we'd count the page twice
6821 * Cgroup2 has separate resource counters for memory and swap,
6822 * so this is a non-issue here. Memory and swap charge lifetimes
6823 * correspond 1:1 to page and swap slot lifetimes: we charge the
6824 * page to memory here, and uncharge swap when the slot is freed.
6826 if (!mem_cgroup_disabled() && do_memsw_account()) {
6828 * The swap entry might not get freed for a long time,
6829 * let's not wait for it. The page already received a
6830 * memory+swap charge, drop the swap entry duplicate.
6832 mem_cgroup_uncharge_swap(entry
, 1);
6836 struct uncharge_gather
{
6837 struct mem_cgroup
*memcg
;
6838 unsigned long nr_memory
;
6839 unsigned long pgpgout
;
6840 unsigned long nr_kmem
;
6841 struct page
*dummy_page
;
6844 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
6846 memset(ug
, 0, sizeof(*ug
));
6849 static void uncharge_batch(const struct uncharge_gather
*ug
)
6851 unsigned long flags
;
6853 if (ug
->nr_memory
) {
6854 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_memory
);
6855 if (do_memsw_account())
6856 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_memory
);
6857 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && ug
->nr_kmem
)
6858 page_counter_uncharge(&ug
->memcg
->kmem
, ug
->nr_kmem
);
6859 memcg_oom_recover(ug
->memcg
);
6862 local_irq_save(flags
);
6863 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
6864 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_memory
);
6865 memcg_check_events(ug
->memcg
, ug
->dummy_page
);
6866 local_irq_restore(flags
);
6868 /* drop reference from uncharge_page */
6869 css_put(&ug
->memcg
->css
);
6872 static void uncharge_page(struct page
*page
, struct uncharge_gather
*ug
)
6874 unsigned long nr_pages
;
6875 struct mem_cgroup
*memcg
;
6876 struct obj_cgroup
*objcg
;
6877 bool use_objcg
= PageMemcgKmem(page
);
6879 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6882 * Nobody should be changing or seriously looking at
6883 * page memcg or objcg at this point, we have fully
6884 * exclusive access to the page.
6887 objcg
= __page_objcg(page
);
6889 * This get matches the put at the end of the function and
6890 * kmem pages do not hold memcg references anymore.
6892 memcg
= get_mem_cgroup_from_objcg(objcg
);
6894 memcg
= __page_memcg(page
);
6900 if (ug
->memcg
!= memcg
) {
6903 uncharge_gather_clear(ug
);
6906 ug
->dummy_page
= page
;
6908 /* pairs with css_put in uncharge_batch */
6909 css_get(&memcg
->css
);
6912 nr_pages
= compound_nr(page
);
6915 ug
->nr_memory
+= nr_pages
;
6916 ug
->nr_kmem
+= nr_pages
;
6918 page
->memcg_data
= 0;
6919 obj_cgroup_put(objcg
);
6921 /* LRU pages aren't accounted at the root level */
6922 if (!mem_cgroup_is_root(memcg
))
6923 ug
->nr_memory
+= nr_pages
;
6926 page
->memcg_data
= 0;
6929 css_put(&memcg
->css
);
6933 * __mem_cgroup_uncharge - uncharge a page
6934 * @page: page to uncharge
6936 * Uncharge a page previously charged with __mem_cgroup_charge().
6938 void __mem_cgroup_uncharge(struct page
*page
)
6940 struct uncharge_gather ug
;
6942 /* Don't touch page->lru of any random page, pre-check: */
6943 if (!page_memcg(page
))
6946 uncharge_gather_clear(&ug
);
6947 uncharge_page(page
, &ug
);
6948 uncharge_batch(&ug
);
6952 * __mem_cgroup_uncharge_list - uncharge a list of page
6953 * @page_list: list of pages to uncharge
6955 * Uncharge a list of pages previously charged with
6956 * __mem_cgroup_charge().
6958 void __mem_cgroup_uncharge_list(struct list_head
*page_list
)
6960 struct uncharge_gather ug
;
6963 uncharge_gather_clear(&ug
);
6964 list_for_each_entry(page
, page_list
, lru
)
6965 uncharge_page(page
, &ug
);
6967 uncharge_batch(&ug
);
6971 * mem_cgroup_migrate - charge a page's replacement
6972 * @oldpage: currently circulating page
6973 * @newpage: replacement page
6975 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6976 * be uncharged upon free.
6978 * Both pages must be locked, @newpage->mapping must be set up.
6980 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
6982 struct mem_cgroup
*memcg
;
6983 unsigned int nr_pages
;
6984 unsigned long flags
;
6986 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6987 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6988 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6989 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6992 if (mem_cgroup_disabled())
6995 /* Page cache replacement: new page already charged? */
6996 if (page_memcg(newpage
))
6999 memcg
= page_memcg(oldpage
);
7000 VM_WARN_ON_ONCE_PAGE(!memcg
, oldpage
);
7004 /* Force-charge the new page. The old one will be freed soon */
7005 nr_pages
= thp_nr_pages(newpage
);
7007 if (!mem_cgroup_is_root(memcg
)) {
7008 page_counter_charge(&memcg
->memory
, nr_pages
);
7009 if (do_memsw_account())
7010 page_counter_charge(&memcg
->memsw
, nr_pages
);
7013 css_get(&memcg
->css
);
7014 commit_charge(newpage
, memcg
);
7016 local_irq_save(flags
);
7017 mem_cgroup_charge_statistics(memcg
, newpage
, nr_pages
);
7018 memcg_check_events(memcg
, newpage
);
7019 local_irq_restore(flags
);
7022 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
7023 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
7025 void mem_cgroup_sk_alloc(struct sock
*sk
)
7027 struct mem_cgroup
*memcg
;
7029 if (!mem_cgroup_sockets_enabled
)
7032 /* Do not associate the sock with unrelated interrupted task's memcg. */
7037 memcg
= mem_cgroup_from_task(current
);
7038 if (memcg
== root_mem_cgroup
)
7040 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
7042 if (css_tryget(&memcg
->css
))
7043 sk
->sk_memcg
= memcg
;
7048 void mem_cgroup_sk_free(struct sock
*sk
)
7051 css_put(&sk
->sk_memcg
->css
);
7055 * mem_cgroup_charge_skmem - charge socket memory
7056 * @memcg: memcg to charge
7057 * @nr_pages: number of pages to charge
7058 * @gfp_mask: reclaim mode
7060 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7061 * @memcg's configured limit, %false if it doesn't.
7063 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
,
7066 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7067 struct page_counter
*fail
;
7069 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7070 memcg
->tcpmem_pressure
= 0;
7073 memcg
->tcpmem_pressure
= 1;
7074 if (gfp_mask
& __GFP_NOFAIL
) {
7075 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7081 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0) {
7082 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7090 * mem_cgroup_uncharge_skmem - uncharge socket memory
7091 * @memcg: memcg to uncharge
7092 * @nr_pages: number of pages to uncharge
7094 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7096 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7097 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7101 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7103 refill_stock(memcg
, nr_pages
);
7106 static int __init
cgroup_memory(char *s
)
7110 while ((token
= strsep(&s
, ",")) != NULL
) {
7113 if (!strcmp(token
, "nosocket"))
7114 cgroup_memory_nosocket
= true;
7115 if (!strcmp(token
, "nokmem"))
7116 cgroup_memory_nokmem
= true;
7120 __setup("cgroup.memory=", cgroup_memory
);
7123 * subsys_initcall() for memory controller.
7125 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7126 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7127 * basically everything that doesn't depend on a specific mem_cgroup structure
7128 * should be initialized from here.
7130 static int __init
mem_cgroup_init(void)
7135 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7136 * used for per-memcg-per-cpu caching of per-node statistics. In order
7137 * to work fine, we should make sure that the overfill threshold can't
7138 * exceed S32_MAX / PAGE_SIZE.
7140 BUILD_BUG_ON(MEMCG_CHARGE_BATCH
> S32_MAX
/ PAGE_SIZE
);
7142 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7143 memcg_hotplug_cpu_dead
);
7145 for_each_possible_cpu(cpu
)
7146 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7149 for_each_node(node
) {
7150 struct mem_cgroup_tree_per_node
*rtpn
;
7152 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
7153 node_online(node
) ? node
: NUMA_NO_NODE
);
7155 rtpn
->rb_root
= RB_ROOT
;
7156 rtpn
->rb_rightmost
= NULL
;
7157 spin_lock_init(&rtpn
->lock
);
7158 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7163 subsys_initcall(mem_cgroup_init
);
7165 #ifdef CONFIG_MEMCG_SWAP
7166 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7168 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7170 * The root cgroup cannot be destroyed, so it's refcount must
7173 if (WARN_ON_ONCE(memcg
== root_mem_cgroup
)) {
7177 memcg
= parent_mem_cgroup(memcg
);
7179 memcg
= root_mem_cgroup
;
7185 * mem_cgroup_swapout - transfer a memsw charge to swap
7186 * @page: page whose memsw charge to transfer
7187 * @entry: swap entry to move the charge to
7189 * Transfer the memsw charge of @page to @entry.
7191 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
7193 struct mem_cgroup
*memcg
, *swap_memcg
;
7194 unsigned int nr_entries
;
7195 unsigned short oldid
;
7197 VM_BUG_ON_PAGE(PageLRU(page
), page
);
7198 VM_BUG_ON_PAGE(page_count(page
), page
);
7200 if (mem_cgroup_disabled())
7203 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7206 memcg
= page_memcg(page
);
7208 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7213 * In case the memcg owning these pages has been offlined and doesn't
7214 * have an ID allocated to it anymore, charge the closest online
7215 * ancestor for the swap instead and transfer the memory+swap charge.
7217 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7218 nr_entries
= thp_nr_pages(page
);
7219 /* Get references for the tail pages, too */
7221 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7222 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7224 VM_BUG_ON_PAGE(oldid
, page
);
7225 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7227 page
->memcg_data
= 0;
7229 if (!mem_cgroup_is_root(memcg
))
7230 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7232 if (!cgroup_memory_noswap
&& memcg
!= swap_memcg
) {
7233 if (!mem_cgroup_is_root(swap_memcg
))
7234 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7235 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7239 * Interrupts should be disabled here because the caller holds the
7240 * i_pages lock which is taken with interrupts-off. It is
7241 * important here to have the interrupts disabled because it is the
7242 * only synchronisation we have for updating the per-CPU variables.
7244 VM_BUG_ON(!irqs_disabled());
7245 mem_cgroup_charge_statistics(memcg
, page
, -nr_entries
);
7246 memcg_check_events(memcg
, page
);
7248 css_put(&memcg
->css
);
7252 * __mem_cgroup_try_charge_swap - try charging swap space for a page
7253 * @page: page being added to swap
7254 * @entry: swap entry to charge
7256 * Try to charge @page's memcg for the swap space at @entry.
7258 * Returns 0 on success, -ENOMEM on failure.
7260 int __mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
7262 unsigned int nr_pages
= thp_nr_pages(page
);
7263 struct page_counter
*counter
;
7264 struct mem_cgroup
*memcg
;
7265 unsigned short oldid
;
7267 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7270 memcg
= page_memcg(page
);
7272 VM_WARN_ON_ONCE_PAGE(!memcg
, page
);
7277 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7281 memcg
= mem_cgroup_id_get_online(memcg
);
7283 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
) &&
7284 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7285 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7286 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7287 mem_cgroup_id_put(memcg
);
7291 /* Get references for the tail pages, too */
7293 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7294 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7295 VM_BUG_ON_PAGE(oldid
, page
);
7296 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7302 * __mem_cgroup_uncharge_swap - uncharge swap space
7303 * @entry: swap entry to uncharge
7304 * @nr_pages: the amount of swap space to uncharge
7306 void __mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7308 struct mem_cgroup
*memcg
;
7311 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7313 memcg
= mem_cgroup_from_id(id
);
7315 if (!cgroup_memory_noswap
&& !mem_cgroup_is_root(memcg
)) {
7316 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7317 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7319 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7321 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7322 mem_cgroup_id_put_many(memcg
, nr_pages
);
7327 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7329 long nr_swap_pages
= get_nr_swap_pages();
7331 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7332 return nr_swap_pages
;
7333 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
7334 nr_swap_pages
= min_t(long, nr_swap_pages
,
7335 READ_ONCE(memcg
->swap
.max
) -
7336 page_counter_read(&memcg
->swap
));
7337 return nr_swap_pages
;
7340 bool mem_cgroup_swap_full(struct page
*page
)
7342 struct mem_cgroup
*memcg
;
7344 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
7348 if (cgroup_memory_noswap
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
7351 memcg
= page_memcg(page
);
7355 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
)) {
7356 unsigned long usage
= page_counter_read(&memcg
->swap
);
7358 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7359 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7366 static int __init
setup_swap_account(char *s
)
7368 if (!strcmp(s
, "1"))
7369 cgroup_memory_noswap
= false;
7370 else if (!strcmp(s
, "0"))
7371 cgroup_memory_noswap
= true;
7374 __setup("swapaccount=", setup_swap_account
);
7376 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7379 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7381 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7384 static int swap_high_show(struct seq_file
*m
, void *v
)
7386 return seq_puts_memcg_tunable(m
,
7387 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7390 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7391 char *buf
, size_t nbytes
, loff_t off
)
7393 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7397 buf
= strstrip(buf
);
7398 err
= page_counter_memparse(buf
, "max", &high
);
7402 page_counter_set_high(&memcg
->swap
, high
);
7407 static int swap_max_show(struct seq_file
*m
, void *v
)
7409 return seq_puts_memcg_tunable(m
,
7410 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7413 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7414 char *buf
, size_t nbytes
, loff_t off
)
7416 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7420 buf
= strstrip(buf
);
7421 err
= page_counter_memparse(buf
, "max", &max
);
7425 xchg(&memcg
->swap
.max
, max
);
7430 static int swap_events_show(struct seq_file
*m
, void *v
)
7432 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7434 seq_printf(m
, "high %lu\n",
7435 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7436 seq_printf(m
, "max %lu\n",
7437 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7438 seq_printf(m
, "fail %lu\n",
7439 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7444 static struct cftype swap_files
[] = {
7446 .name
= "swap.current",
7447 .flags
= CFTYPE_NOT_ON_ROOT
,
7448 .read_u64
= swap_current_read
,
7451 .name
= "swap.high",
7452 .flags
= CFTYPE_NOT_ON_ROOT
,
7453 .seq_show
= swap_high_show
,
7454 .write
= swap_high_write
,
7458 .flags
= CFTYPE_NOT_ON_ROOT
,
7459 .seq_show
= swap_max_show
,
7460 .write
= swap_max_write
,
7463 .name
= "swap.events",
7464 .flags
= CFTYPE_NOT_ON_ROOT
,
7465 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
7466 .seq_show
= swap_events_show
,
7471 static struct cftype memsw_files
[] = {
7473 .name
= "memsw.usage_in_bytes",
7474 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
7475 .read_u64
= mem_cgroup_read_u64
,
7478 .name
= "memsw.max_usage_in_bytes",
7479 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
7480 .write
= mem_cgroup_reset
,
7481 .read_u64
= mem_cgroup_read_u64
,
7484 .name
= "memsw.limit_in_bytes",
7485 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
7486 .write
= mem_cgroup_write
,
7487 .read_u64
= mem_cgroup_read_u64
,
7490 .name
= "memsw.failcnt",
7491 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
7492 .write
= mem_cgroup_reset
,
7493 .read_u64
= mem_cgroup_read_u64
,
7495 { }, /* terminate */
7499 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7500 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7501 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7502 * boot parameter. This may result in premature OOPS inside
7503 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7505 static int __init
mem_cgroup_swap_init(void)
7507 /* No memory control -> no swap control */
7508 if (mem_cgroup_disabled())
7509 cgroup_memory_noswap
= true;
7511 if (cgroup_memory_noswap
)
7514 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
7515 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
7519 core_initcall(mem_cgroup_swap_init
);
7521 #endif /* CONFIG_MEMCG_SWAP */