1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket
;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem
;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly
;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
) && do_swap_account
;
101 static const char * const mem_cgroup_stat_names
[] = {
111 static const char * const mem_cgroup_events_names
[] = {
118 static const char * const mem_cgroup_lru_names
[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone
{
136 struct rb_root rb_root
;
140 struct mem_cgroup_tree_per_node
{
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
144 struct mem_cgroup_tree
{
145 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
151 struct mem_cgroup_eventfd_list
{
152 struct list_head list
;
153 struct eventfd_ctx
*eventfd
;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event
{
161 * memcg which the event belongs to.
163 struct mem_cgroup
*memcg
;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx
*eventfd
;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list
;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event
)(struct mem_cgroup
*memcg
,
178 struct eventfd_ctx
*eventfd
, const char *args
);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event
)(struct mem_cgroup
*memcg
,
185 struct eventfd_ctx
*eventfd
);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t
*wqh
;
193 struct work_struct remove
;
196 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
197 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct
{
209 spinlock_t lock
; /* for from, to */
210 struct mem_cgroup
*from
;
211 struct mem_cgroup
*to
;
213 unsigned long precharge
;
214 unsigned long moved_charge
;
215 unsigned long moved_swap
;
216 struct task_struct
*moving_task
; /* a task moving charges */
217 wait_queue_head_t waitq
; /* a waitq for other context */
219 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
220 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON
,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
257 memcg
= root_mem_cgroup
;
258 return &memcg
->vmpressure
;
261 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
263 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
268 return (memcg
== root_mem_cgroup
);
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida
);
284 int memcg_nr_cache_ids
;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem
);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem
);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem
);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key
);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
323 #endif /* !CONFIG_SLOB */
325 static struct mem_cgroup_per_zone
*
326 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
328 int nid
= zone_to_nid(zone
);
329 int zid
= zone_idx(zone
);
331 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
335 * mem_cgroup_css_from_page - css of the memcg associated with a page
336 * @page: page of interest
338 * If memcg is bound to the default hierarchy, css of the memcg associated
339 * with @page is returned. The returned css remains associated with @page
340 * until it is released.
342 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
345 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
347 struct mem_cgroup
*memcg
;
349 memcg
= page
->mem_cgroup
;
351 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
352 memcg
= root_mem_cgroup
;
358 * page_cgroup_ino - return inode number of the memcg a page is charged to
361 * Look up the closest online ancestor of the memory cgroup @page is charged to
362 * and return its inode number or 0 if @page is not charged to any cgroup. It
363 * is safe to call this function without holding a reference to @page.
365 * Note, this function is inherently racy, because there is nothing to prevent
366 * the cgroup inode from getting torn down and potentially reallocated a moment
367 * after page_cgroup_ino() returns, so it only should be used by callers that
368 * do not care (such as procfs interfaces).
370 ino_t
page_cgroup_ino(struct page
*page
)
372 struct mem_cgroup
*memcg
;
373 unsigned long ino
= 0;
376 memcg
= READ_ONCE(page
->mem_cgroup
);
377 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
378 memcg
= parent_mem_cgroup(memcg
);
380 ino
= cgroup_ino(memcg
->css
.cgroup
);
385 static struct mem_cgroup_per_zone
*
386 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
388 int nid
= page_to_nid(page
);
389 int zid
= page_zonenum(page
);
391 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
394 static struct mem_cgroup_tree_per_zone
*
395 soft_limit_tree_node_zone(int nid
, int zid
)
397 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
400 static struct mem_cgroup_tree_per_zone
*
401 soft_limit_tree_from_page(struct page
*page
)
403 int nid
= page_to_nid(page
);
404 int zid
= page_zonenum(page
);
406 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
409 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
410 struct mem_cgroup_tree_per_zone
*mctz
,
411 unsigned long new_usage_in_excess
)
413 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
414 struct rb_node
*parent
= NULL
;
415 struct mem_cgroup_per_zone
*mz_node
;
420 mz
->usage_in_excess
= new_usage_in_excess
;
421 if (!mz
->usage_in_excess
)
425 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
427 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
430 * We can't avoid mem cgroups that are over their soft
431 * limit by the same amount
433 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
436 rb_link_node(&mz
->tree_node
, parent
, p
);
437 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
441 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
442 struct mem_cgroup_tree_per_zone
*mctz
)
446 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
450 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
451 struct mem_cgroup_tree_per_zone
*mctz
)
455 spin_lock_irqsave(&mctz
->lock
, flags
);
456 __mem_cgroup_remove_exceeded(mz
, mctz
);
457 spin_unlock_irqrestore(&mctz
->lock
, flags
);
460 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
462 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
463 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
464 unsigned long excess
= 0;
466 if (nr_pages
> soft_limit
)
467 excess
= nr_pages
- soft_limit
;
472 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
474 unsigned long excess
;
475 struct mem_cgroup_per_zone
*mz
;
476 struct mem_cgroup_tree_per_zone
*mctz
;
478 mctz
= soft_limit_tree_from_page(page
);
480 * Necessary to update all ancestors when hierarchy is used.
481 * because their event counter is not touched.
483 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
484 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
485 excess
= soft_limit_excess(memcg
);
487 * We have to update the tree if mz is on RB-tree or
488 * mem is over its softlimit.
490 if (excess
|| mz
->on_tree
) {
493 spin_lock_irqsave(&mctz
->lock
, flags
);
494 /* if on-tree, remove it */
496 __mem_cgroup_remove_exceeded(mz
, mctz
);
498 * Insert again. mz->usage_in_excess will be updated.
499 * If excess is 0, no tree ops.
501 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
502 spin_unlock_irqrestore(&mctz
->lock
, flags
);
507 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
509 struct mem_cgroup_tree_per_zone
*mctz
;
510 struct mem_cgroup_per_zone
*mz
;
514 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
515 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
516 mctz
= soft_limit_tree_node_zone(nid
, zid
);
517 mem_cgroup_remove_exceeded(mz
, mctz
);
522 static struct mem_cgroup_per_zone
*
523 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
525 struct rb_node
*rightmost
= NULL
;
526 struct mem_cgroup_per_zone
*mz
;
530 rightmost
= rb_last(&mctz
->rb_root
);
532 goto done
; /* Nothing to reclaim from */
534 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
536 * Remove the node now but someone else can add it back,
537 * we will to add it back at the end of reclaim to its correct
538 * position in the tree.
540 __mem_cgroup_remove_exceeded(mz
, mctz
);
541 if (!soft_limit_excess(mz
->memcg
) ||
542 !css_tryget_online(&mz
->memcg
->css
))
548 static struct mem_cgroup_per_zone
*
549 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
551 struct mem_cgroup_per_zone
*mz
;
553 spin_lock_irq(&mctz
->lock
);
554 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
555 spin_unlock_irq(&mctz
->lock
);
560 * Return page count for single (non recursive) @memcg.
562 * Implementation Note: reading percpu statistics for memcg.
564 * Both of vmstat[] and percpu_counter has threshold and do periodic
565 * synchronization to implement "quick" read. There are trade-off between
566 * reading cost and precision of value. Then, we may have a chance to implement
567 * a periodic synchronization of counter in memcg's counter.
569 * But this _read() function is used for user interface now. The user accounts
570 * memory usage by memory cgroup and he _always_ requires exact value because
571 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
572 * have to visit all online cpus and make sum. So, for now, unnecessary
573 * synchronization is not implemented. (just implemented for cpu hotplug)
575 * If there are kernel internal actions which can make use of some not-exact
576 * value, and reading all cpu value can be performance bottleneck in some
577 * common workload, threshold and synchronization as vmstat[] should be
581 mem_cgroup_read_stat(struct mem_cgroup
*memcg
, enum mem_cgroup_stat_index idx
)
586 /* Per-cpu values can be negative, use a signed accumulator */
587 for_each_possible_cpu(cpu
)
588 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
590 * Summing races with updates, so val may be negative. Avoid exposing
591 * transient negative values.
598 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
599 enum mem_cgroup_events_index idx
)
601 unsigned long val
= 0;
604 for_each_possible_cpu(cpu
)
605 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
609 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
611 bool compound
, int nr_pages
)
614 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
615 * counted as CACHE even if it's on ANON LRU.
618 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
621 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
625 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
626 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
630 /* pagein of a big page is an event. So, ignore page size */
632 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
634 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
635 nr_pages
= -nr_pages
; /* for event */
638 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
641 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
642 int nid
, unsigned int lru_mask
)
644 unsigned long nr
= 0;
647 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
649 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
650 struct mem_cgroup_per_zone
*mz
;
654 if (!(BIT(lru
) & lru_mask
))
656 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
657 nr
+= mz
->lru_size
[lru
];
663 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
664 unsigned int lru_mask
)
666 unsigned long nr
= 0;
669 for_each_node_state(nid
, N_MEMORY
)
670 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
674 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
675 enum mem_cgroup_events_target target
)
677 unsigned long val
, next
;
679 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
680 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
681 /* from time_after() in jiffies.h */
682 if ((long)next
- (long)val
< 0) {
684 case MEM_CGROUP_TARGET_THRESH
:
685 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
687 case MEM_CGROUP_TARGET_SOFTLIMIT
:
688 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
690 case MEM_CGROUP_TARGET_NUMAINFO
:
691 next
= val
+ NUMAINFO_EVENTS_TARGET
;
696 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
703 * Check events in order.
706 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
708 /* threshold event is triggered in finer grain than soft limit */
709 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
710 MEM_CGROUP_TARGET_THRESH
))) {
712 bool do_numainfo __maybe_unused
;
714 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
715 MEM_CGROUP_TARGET_SOFTLIMIT
);
717 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
718 MEM_CGROUP_TARGET_NUMAINFO
);
720 mem_cgroup_threshold(memcg
);
721 if (unlikely(do_softlimit
))
722 mem_cgroup_update_tree(memcg
, page
);
724 if (unlikely(do_numainfo
))
725 atomic_inc(&memcg
->numainfo_events
);
730 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
733 * mm_update_next_owner() may clear mm->owner to NULL
734 * if it races with swapoff, page migration, etc.
735 * So this can be called with p == NULL.
740 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
742 EXPORT_SYMBOL(mem_cgroup_from_task
);
744 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
746 struct mem_cgroup
*memcg
= NULL
;
751 * Page cache insertions can happen withou an
752 * actual mm context, e.g. during disk probing
753 * on boot, loopback IO, acct() writes etc.
756 memcg
= root_mem_cgroup
;
758 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
759 if (unlikely(!memcg
))
760 memcg
= root_mem_cgroup
;
762 } while (!css_tryget_online(&memcg
->css
));
768 * mem_cgroup_iter - iterate over memory cgroup hierarchy
769 * @root: hierarchy root
770 * @prev: previously returned memcg, NULL on first invocation
771 * @reclaim: cookie for shared reclaim walks, NULL for full walks
773 * Returns references to children of the hierarchy below @root, or
774 * @root itself, or %NULL after a full round-trip.
776 * Caller must pass the return value in @prev on subsequent
777 * invocations for reference counting, or use mem_cgroup_iter_break()
778 * to cancel a hierarchy walk before the round-trip is complete.
780 * Reclaimers can specify a zone and a priority level in @reclaim to
781 * divide up the memcgs in the hierarchy among all concurrent
782 * reclaimers operating on the same zone and priority.
784 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
785 struct mem_cgroup
*prev
,
786 struct mem_cgroup_reclaim_cookie
*reclaim
)
788 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
789 struct cgroup_subsys_state
*css
= NULL
;
790 struct mem_cgroup
*memcg
= NULL
;
791 struct mem_cgroup
*pos
= NULL
;
793 if (mem_cgroup_disabled())
797 root
= root_mem_cgroup
;
799 if (prev
&& !reclaim
)
802 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
811 struct mem_cgroup_per_zone
*mz
;
813 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
814 iter
= &mz
->iter
[reclaim
->priority
];
816 if (prev
&& reclaim
->generation
!= iter
->generation
)
820 pos
= READ_ONCE(iter
->position
);
821 if (!pos
|| css_tryget(&pos
->css
))
824 * css reference reached zero, so iter->position will
825 * be cleared by ->css_released. However, we should not
826 * rely on this happening soon, because ->css_released
827 * is called from a work queue, and by busy-waiting we
828 * might block it. So we clear iter->position right
831 (void)cmpxchg(&iter
->position
, pos
, NULL
);
839 css
= css_next_descendant_pre(css
, &root
->css
);
842 * Reclaimers share the hierarchy walk, and a
843 * new one might jump in right at the end of
844 * the hierarchy - make sure they see at least
845 * one group and restart from the beginning.
853 * Verify the css and acquire a reference. The root
854 * is provided by the caller, so we know it's alive
855 * and kicking, and don't take an extra reference.
857 memcg
= mem_cgroup_from_css(css
);
859 if (css
== &root
->css
)
870 * The position could have already been updated by a competing
871 * thread, so check that the value hasn't changed since we read
872 * it to avoid reclaiming from the same cgroup twice.
874 (void)cmpxchg(&iter
->position
, pos
, memcg
);
882 reclaim
->generation
= iter
->generation
;
888 if (prev
&& prev
!= root
)
895 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
896 * @root: hierarchy root
897 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
899 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
900 struct mem_cgroup
*prev
)
903 root
= root_mem_cgroup
;
904 if (prev
&& prev
!= root
)
908 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
910 struct mem_cgroup
*memcg
= dead_memcg
;
911 struct mem_cgroup_reclaim_iter
*iter
;
912 struct mem_cgroup_per_zone
*mz
;
916 while ((memcg
= parent_mem_cgroup(memcg
))) {
918 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
919 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
920 for (i
= 0; i
<= DEF_PRIORITY
; i
++) {
922 cmpxchg(&iter
->position
,
931 * Iteration constructs for visiting all cgroups (under a tree). If
932 * loops are exited prematurely (break), mem_cgroup_iter_break() must
933 * be used for reference counting.
935 #define for_each_mem_cgroup_tree(iter, root) \
936 for (iter = mem_cgroup_iter(root, NULL, NULL); \
938 iter = mem_cgroup_iter(root, iter, NULL))
940 #define for_each_mem_cgroup(iter) \
941 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
943 iter = mem_cgroup_iter(NULL, iter, NULL))
946 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
947 * @zone: zone of the wanted lruvec
948 * @memcg: memcg of the wanted lruvec
950 * Returns the lru list vector holding pages for the given @zone and
951 * @mem. This can be the global zone lruvec, if the memory controller
954 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
955 struct mem_cgroup
*memcg
)
957 struct mem_cgroup_per_zone
*mz
;
958 struct lruvec
*lruvec
;
960 if (mem_cgroup_disabled()) {
961 lruvec
= &zone
->lruvec
;
965 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
966 lruvec
= &mz
->lruvec
;
969 * Since a node can be onlined after the mem_cgroup was created,
970 * we have to be prepared to initialize lruvec->zone here;
971 * and if offlined then reonlined, we need to reinitialize it.
973 if (unlikely(lruvec
->zone
!= zone
))
979 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
981 * @zone: zone of the page
983 * This function is only safe when following the LRU page isolation
984 * and putback protocol: the LRU lock must be held, and the page must
985 * either be PageLRU() or the caller must have isolated/allocated it.
987 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
989 struct mem_cgroup_per_zone
*mz
;
990 struct mem_cgroup
*memcg
;
991 struct lruvec
*lruvec
;
993 if (mem_cgroup_disabled()) {
994 lruvec
= &zone
->lruvec
;
998 memcg
= page
->mem_cgroup
;
1000 * Swapcache readahead pages are added to the LRU - and
1001 * possibly migrated - before they are charged.
1004 memcg
= root_mem_cgroup
;
1006 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1007 lruvec
= &mz
->lruvec
;
1010 * Since a node can be onlined after the mem_cgroup was created,
1011 * we have to be prepared to initialize lruvec->zone here;
1012 * and if offlined then reonlined, we need to reinitialize it.
1014 if (unlikely(lruvec
->zone
!= zone
))
1015 lruvec
->zone
= zone
;
1020 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1021 * @lruvec: mem_cgroup per zone lru vector
1022 * @lru: index of lru list the page is sitting on
1023 * @nr_pages: positive when adding or negative when removing
1025 * This function must be called when a page is added to or removed from an
1028 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1031 struct mem_cgroup_per_zone
*mz
;
1032 unsigned long *lru_size
;
1034 if (mem_cgroup_disabled())
1037 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1038 lru_size
= mz
->lru_size
+ lru
;
1039 *lru_size
+= nr_pages
;
1040 VM_BUG_ON((long)(*lru_size
) < 0);
1043 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1045 struct mem_cgroup
*task_memcg
;
1046 struct task_struct
*p
;
1049 p
= find_lock_task_mm(task
);
1051 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1055 * All threads may have already detached their mm's, but the oom
1056 * killer still needs to detect if they have already been oom
1057 * killed to prevent needlessly killing additional tasks.
1060 task_memcg
= mem_cgroup_from_task(task
);
1061 css_get(&task_memcg
->css
);
1064 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1065 css_put(&task_memcg
->css
);
1070 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1071 * @memcg: the memory cgroup
1073 * Returns the maximum amount of memory @mem can be charged with, in
1076 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1078 unsigned long margin
= 0;
1079 unsigned long count
;
1080 unsigned long limit
;
1082 count
= page_counter_read(&memcg
->memory
);
1083 limit
= READ_ONCE(memcg
->memory
.limit
);
1085 margin
= limit
- count
;
1087 if (do_memsw_account()) {
1088 count
= page_counter_read(&memcg
->memsw
);
1089 limit
= READ_ONCE(memcg
->memsw
.limit
);
1091 margin
= min(margin
, limit
- count
);
1098 * A routine for checking "mem" is under move_account() or not.
1100 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1101 * moving cgroups. This is for waiting at high-memory pressure
1104 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1106 struct mem_cgroup
*from
;
1107 struct mem_cgroup
*to
;
1110 * Unlike task_move routines, we access mc.to, mc.from not under
1111 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1113 spin_lock(&mc
.lock
);
1119 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1120 mem_cgroup_is_descendant(to
, memcg
);
1122 spin_unlock(&mc
.lock
);
1126 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1128 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1129 if (mem_cgroup_under_move(memcg
)) {
1131 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1132 /* moving charge context might have finished. */
1135 finish_wait(&mc
.waitq
, &wait
);
1142 #define K(x) ((x) << (PAGE_SHIFT-10))
1144 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1145 * @memcg: The memory cgroup that went over limit
1146 * @p: Task that is going to be killed
1148 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1151 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1153 /* oom_info_lock ensures that parallel ooms do not interleave */
1154 static DEFINE_MUTEX(oom_info_lock
);
1155 struct mem_cgroup
*iter
;
1158 mutex_lock(&oom_info_lock
);
1162 pr_info("Task in ");
1163 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1164 pr_cont(" killed as a result of limit of ");
1166 pr_info("Memory limit reached of cgroup ");
1169 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1174 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64
)page_counter_read(&memcg
->memory
)),
1176 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1177 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1178 K((u64
)page_counter_read(&memcg
->memsw
)),
1179 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1180 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1181 K((u64
)page_counter_read(&memcg
->kmem
)),
1182 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1184 for_each_mem_cgroup_tree(iter
, memcg
) {
1185 pr_info("Memory cgroup stats for ");
1186 pr_cont_cgroup_path(iter
->css
.cgroup
);
1189 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1190 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1192 pr_cont(" %s:%luKB", mem_cgroup_stat_names
[i
],
1193 K(mem_cgroup_read_stat(iter
, i
)));
1196 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1197 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1198 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1202 mutex_unlock(&oom_info_lock
);
1206 * This function returns the number of memcg under hierarchy tree. Returns
1207 * 1(self count) if no children.
1209 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1212 struct mem_cgroup
*iter
;
1214 for_each_mem_cgroup_tree(iter
, memcg
)
1220 * Return the memory (and swap, if configured) limit for a memcg.
1222 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1224 unsigned long limit
;
1226 limit
= memcg
->memory
.limit
;
1227 if (mem_cgroup_swappiness(memcg
)) {
1228 unsigned long memsw_limit
;
1229 unsigned long swap_limit
;
1231 memsw_limit
= memcg
->memsw
.limit
;
1232 swap_limit
= memcg
->swap
.limit
;
1233 swap_limit
= min(swap_limit
, (unsigned long)total_swap_pages
);
1234 limit
= min(limit
+ swap_limit
, memsw_limit
);
1239 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1242 struct oom_control oc
= {
1245 .gfp_mask
= gfp_mask
,
1248 struct mem_cgroup
*iter
;
1249 unsigned long chosen_points
= 0;
1250 unsigned long totalpages
;
1251 unsigned int points
= 0;
1252 struct task_struct
*chosen
= NULL
;
1254 mutex_lock(&oom_lock
);
1257 * If current has a pending SIGKILL or is exiting, then automatically
1258 * select it. The goal is to allow it to allocate so that it may
1259 * quickly exit and free its memory.
1261 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1262 mark_oom_victim(current
);
1266 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1267 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1268 for_each_mem_cgroup_tree(iter
, memcg
) {
1269 struct css_task_iter it
;
1270 struct task_struct
*task
;
1272 css_task_iter_start(&iter
->css
, &it
);
1273 while ((task
= css_task_iter_next(&it
))) {
1274 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1275 case OOM_SCAN_SELECT
:
1277 put_task_struct(chosen
);
1279 chosen_points
= ULONG_MAX
;
1280 get_task_struct(chosen
);
1282 case OOM_SCAN_CONTINUE
:
1284 case OOM_SCAN_ABORT
:
1285 css_task_iter_end(&it
);
1286 mem_cgroup_iter_break(memcg
, iter
);
1288 put_task_struct(chosen
);
1293 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1294 if (!points
|| points
< chosen_points
)
1296 /* Prefer thread group leaders for display purposes */
1297 if (points
== chosen_points
&&
1298 thread_group_leader(chosen
))
1302 put_task_struct(chosen
);
1304 chosen_points
= points
;
1305 get_task_struct(chosen
);
1307 css_task_iter_end(&it
);
1311 points
= chosen_points
* 1000 / totalpages
;
1312 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1313 "Memory cgroup out of memory");
1316 mutex_unlock(&oom_lock
);
1319 #if MAX_NUMNODES > 1
1322 * test_mem_cgroup_node_reclaimable
1323 * @memcg: the target memcg
1324 * @nid: the node ID to be checked.
1325 * @noswap : specify true here if the user wants flle only information.
1327 * This function returns whether the specified memcg contains any
1328 * reclaimable pages on a node. Returns true if there are any reclaimable
1329 * pages in the node.
1331 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1332 int nid
, bool noswap
)
1334 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1336 if (noswap
|| !total_swap_pages
)
1338 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1345 * Always updating the nodemask is not very good - even if we have an empty
1346 * list or the wrong list here, we can start from some node and traverse all
1347 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1350 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1354 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1355 * pagein/pageout changes since the last update.
1357 if (!atomic_read(&memcg
->numainfo_events
))
1359 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1362 /* make a nodemask where this memcg uses memory from */
1363 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1365 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1367 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1368 node_clear(nid
, memcg
->scan_nodes
);
1371 atomic_set(&memcg
->numainfo_events
, 0);
1372 atomic_set(&memcg
->numainfo_updating
, 0);
1376 * Selecting a node where we start reclaim from. Because what we need is just
1377 * reducing usage counter, start from anywhere is O,K. Considering
1378 * memory reclaim from current node, there are pros. and cons.
1380 * Freeing memory from current node means freeing memory from a node which
1381 * we'll use or we've used. So, it may make LRU bad. And if several threads
1382 * hit limits, it will see a contention on a node. But freeing from remote
1383 * node means more costs for memory reclaim because of memory latency.
1385 * Now, we use round-robin. Better algorithm is welcomed.
1387 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1391 mem_cgroup_may_update_nodemask(memcg
);
1392 node
= memcg
->last_scanned_node
;
1394 node
= next_node(node
, memcg
->scan_nodes
);
1395 if (node
== MAX_NUMNODES
)
1396 node
= first_node(memcg
->scan_nodes
);
1398 * We call this when we hit limit, not when pages are added to LRU.
1399 * No LRU may hold pages because all pages are UNEVICTABLE or
1400 * memcg is too small and all pages are not on LRU. In that case,
1401 * we use curret node.
1403 if (unlikely(node
== MAX_NUMNODES
))
1404 node
= numa_node_id();
1406 memcg
->last_scanned_node
= node
;
1410 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1416 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1419 unsigned long *total_scanned
)
1421 struct mem_cgroup
*victim
= NULL
;
1424 unsigned long excess
;
1425 unsigned long nr_scanned
;
1426 struct mem_cgroup_reclaim_cookie reclaim
= {
1431 excess
= soft_limit_excess(root_memcg
);
1434 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1439 * If we have not been able to reclaim
1440 * anything, it might because there are
1441 * no reclaimable pages under this hierarchy
1446 * We want to do more targeted reclaim.
1447 * excess >> 2 is not to excessive so as to
1448 * reclaim too much, nor too less that we keep
1449 * coming back to reclaim from this cgroup
1451 if (total
>= (excess
>> 2) ||
1452 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1457 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1459 *total_scanned
+= nr_scanned
;
1460 if (!soft_limit_excess(root_memcg
))
1463 mem_cgroup_iter_break(root_memcg
, victim
);
1467 #ifdef CONFIG_LOCKDEP
1468 static struct lockdep_map memcg_oom_lock_dep_map
= {
1469 .name
= "memcg_oom_lock",
1473 static DEFINE_SPINLOCK(memcg_oom_lock
);
1476 * Check OOM-Killer is already running under our hierarchy.
1477 * If someone is running, return false.
1479 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1481 struct mem_cgroup
*iter
, *failed
= NULL
;
1483 spin_lock(&memcg_oom_lock
);
1485 for_each_mem_cgroup_tree(iter
, memcg
) {
1486 if (iter
->oom_lock
) {
1488 * this subtree of our hierarchy is already locked
1489 * so we cannot give a lock.
1492 mem_cgroup_iter_break(memcg
, iter
);
1495 iter
->oom_lock
= true;
1500 * OK, we failed to lock the whole subtree so we have
1501 * to clean up what we set up to the failing subtree
1503 for_each_mem_cgroup_tree(iter
, memcg
) {
1504 if (iter
== failed
) {
1505 mem_cgroup_iter_break(memcg
, iter
);
1508 iter
->oom_lock
= false;
1511 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1513 spin_unlock(&memcg_oom_lock
);
1518 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1520 struct mem_cgroup
*iter
;
1522 spin_lock(&memcg_oom_lock
);
1523 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1524 for_each_mem_cgroup_tree(iter
, memcg
)
1525 iter
->oom_lock
= false;
1526 spin_unlock(&memcg_oom_lock
);
1529 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1531 struct mem_cgroup
*iter
;
1533 spin_lock(&memcg_oom_lock
);
1534 for_each_mem_cgroup_tree(iter
, memcg
)
1536 spin_unlock(&memcg_oom_lock
);
1539 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1541 struct mem_cgroup
*iter
;
1544 * When a new child is created while the hierarchy is under oom,
1545 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1547 spin_lock(&memcg_oom_lock
);
1548 for_each_mem_cgroup_tree(iter
, memcg
)
1549 if (iter
->under_oom
> 0)
1551 spin_unlock(&memcg_oom_lock
);
1554 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1556 struct oom_wait_info
{
1557 struct mem_cgroup
*memcg
;
1561 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1562 unsigned mode
, int sync
, void *arg
)
1564 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1565 struct mem_cgroup
*oom_wait_memcg
;
1566 struct oom_wait_info
*oom_wait_info
;
1568 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1569 oom_wait_memcg
= oom_wait_info
->memcg
;
1571 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1572 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1574 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1577 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1580 * For the following lockless ->under_oom test, the only required
1581 * guarantee is that it must see the state asserted by an OOM when
1582 * this function is called as a result of userland actions
1583 * triggered by the notification of the OOM. This is trivially
1584 * achieved by invoking mem_cgroup_mark_under_oom() before
1585 * triggering notification.
1587 if (memcg
&& memcg
->under_oom
)
1588 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1591 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1593 if (!current
->memcg_may_oom
)
1596 * We are in the middle of the charge context here, so we
1597 * don't want to block when potentially sitting on a callstack
1598 * that holds all kinds of filesystem and mm locks.
1600 * Also, the caller may handle a failed allocation gracefully
1601 * (like optional page cache readahead) and so an OOM killer
1602 * invocation might not even be necessary.
1604 * That's why we don't do anything here except remember the
1605 * OOM context and then deal with it at the end of the page
1606 * fault when the stack is unwound, the locks are released,
1607 * and when we know whether the fault was overall successful.
1609 css_get(&memcg
->css
);
1610 current
->memcg_in_oom
= memcg
;
1611 current
->memcg_oom_gfp_mask
= mask
;
1612 current
->memcg_oom_order
= order
;
1616 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1617 * @handle: actually kill/wait or just clean up the OOM state
1619 * This has to be called at the end of a page fault if the memcg OOM
1620 * handler was enabled.
1622 * Memcg supports userspace OOM handling where failed allocations must
1623 * sleep on a waitqueue until the userspace task resolves the
1624 * situation. Sleeping directly in the charge context with all kinds
1625 * of locks held is not a good idea, instead we remember an OOM state
1626 * in the task and mem_cgroup_oom_synchronize() has to be called at
1627 * the end of the page fault to complete the OOM handling.
1629 * Returns %true if an ongoing memcg OOM situation was detected and
1630 * completed, %false otherwise.
1632 bool mem_cgroup_oom_synchronize(bool handle
)
1634 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
1635 struct oom_wait_info owait
;
1638 /* OOM is global, do not handle */
1642 if (!handle
|| oom_killer_disabled
)
1645 owait
.memcg
= memcg
;
1646 owait
.wait
.flags
= 0;
1647 owait
.wait
.func
= memcg_oom_wake_function
;
1648 owait
.wait
.private = current
;
1649 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1651 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1652 mem_cgroup_mark_under_oom(memcg
);
1654 locked
= mem_cgroup_oom_trylock(memcg
);
1657 mem_cgroup_oom_notify(memcg
);
1659 if (locked
&& !memcg
->oom_kill_disable
) {
1660 mem_cgroup_unmark_under_oom(memcg
);
1661 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1662 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom_gfp_mask
,
1663 current
->memcg_oom_order
);
1666 mem_cgroup_unmark_under_oom(memcg
);
1667 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1671 mem_cgroup_oom_unlock(memcg
);
1673 * There is no guarantee that an OOM-lock contender
1674 * sees the wakeups triggered by the OOM kill
1675 * uncharges. Wake any sleepers explicitely.
1677 memcg_oom_recover(memcg
);
1680 current
->memcg_in_oom
= NULL
;
1681 css_put(&memcg
->css
);
1686 * lock_page_memcg - lock a page->mem_cgroup binding
1689 * This function protects unlocked LRU pages from being moved to
1690 * another cgroup and stabilizes their page->mem_cgroup binding.
1692 void lock_page_memcg(struct page
*page
)
1694 struct mem_cgroup
*memcg
;
1695 unsigned long flags
;
1698 * The RCU lock is held throughout the transaction. The fast
1699 * path can get away without acquiring the memcg->move_lock
1700 * because page moving starts with an RCU grace period.
1704 if (mem_cgroup_disabled())
1707 memcg
= page
->mem_cgroup
;
1708 if (unlikely(!memcg
))
1711 if (atomic_read(&memcg
->moving_account
) <= 0)
1714 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1715 if (memcg
!= page
->mem_cgroup
) {
1716 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1721 * When charge migration first begins, we can have locked and
1722 * unlocked page stat updates happening concurrently. Track
1723 * the task who has the lock for unlock_page_memcg().
1725 memcg
->move_lock_task
= current
;
1726 memcg
->move_lock_flags
= flags
;
1730 EXPORT_SYMBOL(lock_page_memcg
);
1733 * unlock_page_memcg - unlock a page->mem_cgroup binding
1736 void unlock_page_memcg(struct page
*page
)
1738 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
1740 if (memcg
&& memcg
->move_lock_task
== current
) {
1741 unsigned long flags
= memcg
->move_lock_flags
;
1743 memcg
->move_lock_task
= NULL
;
1744 memcg
->move_lock_flags
= 0;
1746 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1751 EXPORT_SYMBOL(unlock_page_memcg
);
1754 * size of first charge trial. "32" comes from vmscan.c's magic value.
1755 * TODO: maybe necessary to use big numbers in big irons.
1757 #define CHARGE_BATCH 32U
1758 struct memcg_stock_pcp
{
1759 struct mem_cgroup
*cached
; /* this never be root cgroup */
1760 unsigned int nr_pages
;
1761 struct work_struct work
;
1762 unsigned long flags
;
1763 #define FLUSHING_CACHED_CHARGE 0
1765 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1766 static DEFINE_MUTEX(percpu_charge_mutex
);
1769 * consume_stock: Try to consume stocked charge on this cpu.
1770 * @memcg: memcg to consume from.
1771 * @nr_pages: how many pages to charge.
1773 * The charges will only happen if @memcg matches the current cpu's memcg
1774 * stock, and at least @nr_pages are available in that stock. Failure to
1775 * service an allocation will refill the stock.
1777 * returns true if successful, false otherwise.
1779 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1781 struct memcg_stock_pcp
*stock
;
1784 if (nr_pages
> CHARGE_BATCH
)
1787 stock
= &get_cpu_var(memcg_stock
);
1788 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1789 stock
->nr_pages
-= nr_pages
;
1792 put_cpu_var(memcg_stock
);
1797 * Returns stocks cached in percpu and reset cached information.
1799 static void drain_stock(struct memcg_stock_pcp
*stock
)
1801 struct mem_cgroup
*old
= stock
->cached
;
1803 if (stock
->nr_pages
) {
1804 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1805 if (do_memsw_account())
1806 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1807 css_put_many(&old
->css
, stock
->nr_pages
);
1808 stock
->nr_pages
= 0;
1810 stock
->cached
= NULL
;
1814 * This must be called under preempt disabled or must be called by
1815 * a thread which is pinned to local cpu.
1817 static void drain_local_stock(struct work_struct
*dummy
)
1819 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1821 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1825 * Cache charges(val) to local per_cpu area.
1826 * This will be consumed by consume_stock() function, later.
1828 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1830 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1832 if (stock
->cached
!= memcg
) { /* reset if necessary */
1834 stock
->cached
= memcg
;
1836 stock
->nr_pages
+= nr_pages
;
1837 put_cpu_var(memcg_stock
);
1841 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1842 * of the hierarchy under it.
1844 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1848 /* If someone's already draining, avoid adding running more workers. */
1849 if (!mutex_trylock(&percpu_charge_mutex
))
1851 /* Notify other cpus that system-wide "drain" is running */
1854 for_each_online_cpu(cpu
) {
1855 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1856 struct mem_cgroup
*memcg
;
1858 memcg
= stock
->cached
;
1859 if (!memcg
|| !stock
->nr_pages
)
1861 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1863 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1865 drain_local_stock(&stock
->work
);
1867 schedule_work_on(cpu
, &stock
->work
);
1872 mutex_unlock(&percpu_charge_mutex
);
1875 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1876 unsigned long action
,
1879 int cpu
= (unsigned long)hcpu
;
1880 struct memcg_stock_pcp
*stock
;
1882 if (action
== CPU_ONLINE
)
1885 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1888 stock
= &per_cpu(memcg_stock
, cpu
);
1893 static void reclaim_high(struct mem_cgroup
*memcg
,
1894 unsigned int nr_pages
,
1898 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
1900 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
1901 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
1902 } while ((memcg
= parent_mem_cgroup(memcg
)));
1905 static void high_work_func(struct work_struct
*work
)
1907 struct mem_cgroup
*memcg
;
1909 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
1910 reclaim_high(memcg
, CHARGE_BATCH
, GFP_KERNEL
);
1914 * Scheduled by try_charge() to be executed from the userland return path
1915 * and reclaims memory over the high limit.
1917 void mem_cgroup_handle_over_high(void)
1919 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
1920 struct mem_cgroup
*memcg
;
1922 if (likely(!nr_pages
))
1925 memcg
= get_mem_cgroup_from_mm(current
->mm
);
1926 reclaim_high(memcg
, nr_pages
, GFP_KERNEL
);
1927 css_put(&memcg
->css
);
1928 current
->memcg_nr_pages_over_high
= 0;
1931 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1932 unsigned int nr_pages
)
1934 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1935 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1936 struct mem_cgroup
*mem_over_limit
;
1937 struct page_counter
*counter
;
1938 unsigned long nr_reclaimed
;
1939 bool may_swap
= true;
1940 bool drained
= false;
1942 if (mem_cgroup_is_root(memcg
))
1945 if (consume_stock(memcg
, nr_pages
))
1948 if (!do_memsw_account() ||
1949 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1950 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1952 if (do_memsw_account())
1953 page_counter_uncharge(&memcg
->memsw
, batch
);
1954 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1956 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1960 if (batch
> nr_pages
) {
1966 * Unlike in global OOM situations, memcg is not in a physical
1967 * memory shortage. Allow dying and OOM-killed tasks to
1968 * bypass the last charges so that they can exit quickly and
1969 * free their memory.
1971 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
1972 fatal_signal_pending(current
) ||
1973 current
->flags
& PF_EXITING
))
1976 if (unlikely(task_in_memcg_oom(current
)))
1979 if (!gfpflags_allow_blocking(gfp_mask
))
1982 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
1984 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
1985 gfp_mask
, may_swap
);
1987 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
1991 drain_all_stock(mem_over_limit
);
1996 if (gfp_mask
& __GFP_NORETRY
)
1999 * Even though the limit is exceeded at this point, reclaim
2000 * may have been able to free some pages. Retry the charge
2001 * before killing the task.
2003 * Only for regular pages, though: huge pages are rather
2004 * unlikely to succeed so close to the limit, and we fall back
2005 * to regular pages anyway in case of failure.
2007 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2010 * At task move, charge accounts can be doubly counted. So, it's
2011 * better to wait until the end of task_move if something is going on.
2013 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2019 if (gfp_mask
& __GFP_NOFAIL
)
2022 if (fatal_signal_pending(current
))
2025 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2027 mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2028 get_order(nr_pages
* PAGE_SIZE
));
2030 if (!(gfp_mask
& __GFP_NOFAIL
))
2034 * The allocation either can't fail or will lead to more memory
2035 * being freed very soon. Allow memory usage go over the limit
2036 * temporarily by force charging it.
2038 page_counter_charge(&memcg
->memory
, nr_pages
);
2039 if (do_memsw_account())
2040 page_counter_charge(&memcg
->memsw
, nr_pages
);
2041 css_get_many(&memcg
->css
, nr_pages
);
2046 css_get_many(&memcg
->css
, batch
);
2047 if (batch
> nr_pages
)
2048 refill_stock(memcg
, batch
- nr_pages
);
2051 * If the hierarchy is above the normal consumption range, schedule
2052 * reclaim on returning to userland. We can perform reclaim here
2053 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2054 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2055 * not recorded as it most likely matches current's and won't
2056 * change in the meantime. As high limit is checked again before
2057 * reclaim, the cost of mismatch is negligible.
2060 if (page_counter_read(&memcg
->memory
) > memcg
->high
) {
2061 /* Don't bother a random interrupted task */
2062 if (in_interrupt()) {
2063 schedule_work(&memcg
->high_work
);
2066 current
->memcg_nr_pages_over_high
+= batch
;
2067 set_notify_resume(current
);
2070 } while ((memcg
= parent_mem_cgroup(memcg
)));
2075 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2077 if (mem_cgroup_is_root(memcg
))
2080 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2081 if (do_memsw_account())
2082 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2084 css_put_many(&memcg
->css
, nr_pages
);
2087 static void lock_page_lru(struct page
*page
, int *isolated
)
2089 struct zone
*zone
= page_zone(page
);
2091 spin_lock_irq(&zone
->lru_lock
);
2092 if (PageLRU(page
)) {
2093 struct lruvec
*lruvec
;
2095 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2097 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2103 static void unlock_page_lru(struct page
*page
, int isolated
)
2105 struct zone
*zone
= page_zone(page
);
2108 struct lruvec
*lruvec
;
2110 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2111 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2113 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2115 spin_unlock_irq(&zone
->lru_lock
);
2118 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2123 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2126 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2127 * may already be on some other mem_cgroup's LRU. Take care of it.
2130 lock_page_lru(page
, &isolated
);
2133 * Nobody should be changing or seriously looking at
2134 * page->mem_cgroup at this point:
2136 * - the page is uncharged
2138 * - the page is off-LRU
2140 * - an anonymous fault has exclusive page access, except for
2141 * a locked page table
2143 * - a page cache insertion, a swapin fault, or a migration
2144 * have the page locked
2146 page
->mem_cgroup
= memcg
;
2149 unlock_page_lru(page
, isolated
);
2153 static int memcg_alloc_cache_id(void)
2158 id
= ida_simple_get(&memcg_cache_ida
,
2159 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2163 if (id
< memcg_nr_cache_ids
)
2167 * There's no space for the new id in memcg_caches arrays,
2168 * so we have to grow them.
2170 down_write(&memcg_cache_ids_sem
);
2172 size
= 2 * (id
+ 1);
2173 if (size
< MEMCG_CACHES_MIN_SIZE
)
2174 size
= MEMCG_CACHES_MIN_SIZE
;
2175 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2176 size
= MEMCG_CACHES_MAX_SIZE
;
2178 err
= memcg_update_all_caches(size
);
2180 err
= memcg_update_all_list_lrus(size
);
2182 memcg_nr_cache_ids
= size
;
2184 up_write(&memcg_cache_ids_sem
);
2187 ida_simple_remove(&memcg_cache_ida
, id
);
2193 static void memcg_free_cache_id(int id
)
2195 ida_simple_remove(&memcg_cache_ida
, id
);
2198 struct memcg_kmem_cache_create_work
{
2199 struct mem_cgroup
*memcg
;
2200 struct kmem_cache
*cachep
;
2201 struct work_struct work
;
2204 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2206 struct memcg_kmem_cache_create_work
*cw
=
2207 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2208 struct mem_cgroup
*memcg
= cw
->memcg
;
2209 struct kmem_cache
*cachep
= cw
->cachep
;
2211 memcg_create_kmem_cache(memcg
, cachep
);
2213 css_put(&memcg
->css
);
2218 * Enqueue the creation of a per-memcg kmem_cache.
2220 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2221 struct kmem_cache
*cachep
)
2223 struct memcg_kmem_cache_create_work
*cw
;
2225 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2229 css_get(&memcg
->css
);
2232 cw
->cachep
= cachep
;
2233 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2235 schedule_work(&cw
->work
);
2238 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2239 struct kmem_cache
*cachep
)
2242 * We need to stop accounting when we kmalloc, because if the
2243 * corresponding kmalloc cache is not yet created, the first allocation
2244 * in __memcg_schedule_kmem_cache_create will recurse.
2246 * However, it is better to enclose the whole function. Depending on
2247 * the debugging options enabled, INIT_WORK(), for instance, can
2248 * trigger an allocation. This too, will make us recurse. Because at
2249 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2250 * the safest choice is to do it like this, wrapping the whole function.
2252 current
->memcg_kmem_skip_account
= 1;
2253 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2254 current
->memcg_kmem_skip_account
= 0;
2258 * Return the kmem_cache we're supposed to use for a slab allocation.
2259 * We try to use the current memcg's version of the cache.
2261 * If the cache does not exist yet, if we are the first user of it,
2262 * we either create it immediately, if possible, or create it asynchronously
2264 * In the latter case, we will let the current allocation go through with
2265 * the original cache.
2267 * Can't be called in interrupt context or from kernel threads.
2268 * This function needs to be called with rcu_read_lock() held.
2270 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
, gfp_t gfp
)
2272 struct mem_cgroup
*memcg
;
2273 struct kmem_cache
*memcg_cachep
;
2276 VM_BUG_ON(!is_root_cache(cachep
));
2278 if (cachep
->flags
& SLAB_ACCOUNT
)
2279 gfp
|= __GFP_ACCOUNT
;
2281 if (!(gfp
& __GFP_ACCOUNT
))
2284 if (current
->memcg_kmem_skip_account
)
2287 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2288 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2292 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2293 if (likely(memcg_cachep
))
2294 return memcg_cachep
;
2297 * If we are in a safe context (can wait, and not in interrupt
2298 * context), we could be be predictable and return right away.
2299 * This would guarantee that the allocation being performed
2300 * already belongs in the new cache.
2302 * However, there are some clashes that can arrive from locking.
2303 * For instance, because we acquire the slab_mutex while doing
2304 * memcg_create_kmem_cache, this means no further allocation
2305 * could happen with the slab_mutex held. So it's better to
2308 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2310 css_put(&memcg
->css
);
2314 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2316 if (!is_root_cache(cachep
))
2317 css_put(&cachep
->memcg_params
.memcg
->css
);
2320 int __memcg_kmem_charge_memcg(struct page
*page
, gfp_t gfp
, int order
,
2321 struct mem_cgroup
*memcg
)
2323 unsigned int nr_pages
= 1 << order
;
2324 struct page_counter
*counter
;
2327 ret
= try_charge(memcg
, gfp
, nr_pages
);
2331 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) &&
2332 !page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
)) {
2333 cancel_charge(memcg
, nr_pages
);
2337 page
->mem_cgroup
= memcg
;
2342 int __memcg_kmem_charge(struct page
*page
, gfp_t gfp
, int order
)
2344 struct mem_cgroup
*memcg
;
2347 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2348 if (!mem_cgroup_is_root(memcg
))
2349 ret
= __memcg_kmem_charge_memcg(page
, gfp
, order
, memcg
);
2350 css_put(&memcg
->css
);
2354 void __memcg_kmem_uncharge(struct page
*page
, int order
)
2356 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2357 unsigned int nr_pages
= 1 << order
;
2362 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2364 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2365 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2367 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2368 if (do_memsw_account())
2369 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2371 page
->mem_cgroup
= NULL
;
2372 css_put_many(&memcg
->css
, nr_pages
);
2374 #endif /* !CONFIG_SLOB */
2376 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2379 * Because tail pages are not marked as "used", set it. We're under
2380 * zone->lru_lock and migration entries setup in all page mappings.
2382 void mem_cgroup_split_huge_fixup(struct page
*head
)
2386 if (mem_cgroup_disabled())
2389 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2390 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2392 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2395 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2397 #ifdef CONFIG_MEMCG_SWAP
2398 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2401 int val
= (charge
) ? 1 : -1;
2402 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2406 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2407 * @entry: swap entry to be moved
2408 * @from: mem_cgroup which the entry is moved from
2409 * @to: mem_cgroup which the entry is moved to
2411 * It succeeds only when the swap_cgroup's record for this entry is the same
2412 * as the mem_cgroup's id of @from.
2414 * Returns 0 on success, -EINVAL on failure.
2416 * The caller must have charged to @to, IOW, called page_counter_charge() about
2417 * both res and memsw, and called css_get().
2419 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2420 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2422 unsigned short old_id
, new_id
;
2424 old_id
= mem_cgroup_id(from
);
2425 new_id
= mem_cgroup_id(to
);
2427 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2428 mem_cgroup_swap_statistics(from
, false);
2429 mem_cgroup_swap_statistics(to
, true);
2435 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2436 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2442 static DEFINE_MUTEX(memcg_limit_mutex
);
2444 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2445 unsigned long limit
)
2447 unsigned long curusage
;
2448 unsigned long oldusage
;
2449 bool enlarge
= false;
2454 * For keeping hierarchical_reclaim simple, how long we should retry
2455 * is depends on callers. We set our retry-count to be function
2456 * of # of children which we should visit in this loop.
2458 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2459 mem_cgroup_count_children(memcg
);
2461 oldusage
= page_counter_read(&memcg
->memory
);
2464 if (signal_pending(current
)) {
2469 mutex_lock(&memcg_limit_mutex
);
2470 if (limit
> memcg
->memsw
.limit
) {
2471 mutex_unlock(&memcg_limit_mutex
);
2475 if (limit
> memcg
->memory
.limit
)
2477 ret
= page_counter_limit(&memcg
->memory
, limit
);
2478 mutex_unlock(&memcg_limit_mutex
);
2483 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2485 curusage
= page_counter_read(&memcg
->memory
);
2486 /* Usage is reduced ? */
2487 if (curusage
>= oldusage
)
2490 oldusage
= curusage
;
2491 } while (retry_count
);
2493 if (!ret
&& enlarge
)
2494 memcg_oom_recover(memcg
);
2499 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2500 unsigned long limit
)
2502 unsigned long curusage
;
2503 unsigned long oldusage
;
2504 bool enlarge
= false;
2508 /* see mem_cgroup_resize_res_limit */
2509 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2510 mem_cgroup_count_children(memcg
);
2512 oldusage
= page_counter_read(&memcg
->memsw
);
2515 if (signal_pending(current
)) {
2520 mutex_lock(&memcg_limit_mutex
);
2521 if (limit
< memcg
->memory
.limit
) {
2522 mutex_unlock(&memcg_limit_mutex
);
2526 if (limit
> memcg
->memsw
.limit
)
2528 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2529 mutex_unlock(&memcg_limit_mutex
);
2534 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2536 curusage
= page_counter_read(&memcg
->memsw
);
2537 /* Usage is reduced ? */
2538 if (curusage
>= oldusage
)
2541 oldusage
= curusage
;
2542 } while (retry_count
);
2544 if (!ret
&& enlarge
)
2545 memcg_oom_recover(memcg
);
2550 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2552 unsigned long *total_scanned
)
2554 unsigned long nr_reclaimed
= 0;
2555 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2556 unsigned long reclaimed
;
2558 struct mem_cgroup_tree_per_zone
*mctz
;
2559 unsigned long excess
;
2560 unsigned long nr_scanned
;
2565 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2567 * This loop can run a while, specially if mem_cgroup's continuously
2568 * keep exceeding their soft limit and putting the system under
2575 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2580 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2581 gfp_mask
, &nr_scanned
);
2582 nr_reclaimed
+= reclaimed
;
2583 *total_scanned
+= nr_scanned
;
2584 spin_lock_irq(&mctz
->lock
);
2585 __mem_cgroup_remove_exceeded(mz
, mctz
);
2588 * If we failed to reclaim anything from this memory cgroup
2589 * it is time to move on to the next cgroup
2593 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2595 excess
= soft_limit_excess(mz
->memcg
);
2597 * One school of thought says that we should not add
2598 * back the node to the tree if reclaim returns 0.
2599 * But our reclaim could return 0, simply because due
2600 * to priority we are exposing a smaller subset of
2601 * memory to reclaim from. Consider this as a longer
2604 /* If excess == 0, no tree ops */
2605 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2606 spin_unlock_irq(&mctz
->lock
);
2607 css_put(&mz
->memcg
->css
);
2610 * Could not reclaim anything and there are no more
2611 * mem cgroups to try or we seem to be looping without
2612 * reclaiming anything.
2614 if (!nr_reclaimed
&&
2616 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2618 } while (!nr_reclaimed
);
2620 css_put(&next_mz
->memcg
->css
);
2621 return nr_reclaimed
;
2625 * Test whether @memcg has children, dead or alive. Note that this
2626 * function doesn't care whether @memcg has use_hierarchy enabled and
2627 * returns %true if there are child csses according to the cgroup
2628 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2630 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2635 ret
= css_next_child(NULL
, &memcg
->css
);
2641 * Reclaims as many pages from the given memcg as possible and moves
2642 * the rest to the parent.
2644 * Caller is responsible for holding css reference for memcg.
2646 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2648 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2650 /* we call try-to-free pages for make this cgroup empty */
2651 lru_add_drain_all();
2652 /* try to free all pages in this cgroup */
2653 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2656 if (signal_pending(current
))
2659 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2663 /* maybe some writeback is necessary */
2664 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2672 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2673 char *buf
, size_t nbytes
,
2676 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2678 if (mem_cgroup_is_root(memcg
))
2680 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2683 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2686 return mem_cgroup_from_css(css
)->use_hierarchy
;
2689 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2690 struct cftype
*cft
, u64 val
)
2693 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2694 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2696 if (memcg
->use_hierarchy
== val
)
2700 * If parent's use_hierarchy is set, we can't make any modifications
2701 * in the child subtrees. If it is unset, then the change can
2702 * occur, provided the current cgroup has no children.
2704 * For the root cgroup, parent_mem is NULL, we allow value to be
2705 * set if there are no children.
2707 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2708 (val
== 1 || val
== 0)) {
2709 if (!memcg_has_children(memcg
))
2710 memcg
->use_hierarchy
= val
;
2719 static void tree_stat(struct mem_cgroup
*memcg
, unsigned long *stat
)
2721 struct mem_cgroup
*iter
;
2724 memset(stat
, 0, sizeof(*stat
) * MEMCG_NR_STAT
);
2726 for_each_mem_cgroup_tree(iter
, memcg
) {
2727 for (i
= 0; i
< MEMCG_NR_STAT
; i
++)
2728 stat
[i
] += mem_cgroup_read_stat(iter
, i
);
2732 static void tree_events(struct mem_cgroup
*memcg
, unsigned long *events
)
2734 struct mem_cgroup
*iter
;
2737 memset(events
, 0, sizeof(*events
) * MEMCG_NR_EVENTS
);
2739 for_each_mem_cgroup_tree(iter
, memcg
) {
2740 for (i
= 0; i
< MEMCG_NR_EVENTS
; i
++)
2741 events
[i
] += mem_cgroup_read_events(iter
, i
);
2745 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2747 unsigned long val
= 0;
2749 if (mem_cgroup_is_root(memcg
)) {
2750 struct mem_cgroup
*iter
;
2752 for_each_mem_cgroup_tree(iter
, memcg
) {
2753 val
+= mem_cgroup_read_stat(iter
,
2754 MEM_CGROUP_STAT_CACHE
);
2755 val
+= mem_cgroup_read_stat(iter
,
2756 MEM_CGROUP_STAT_RSS
);
2758 val
+= mem_cgroup_read_stat(iter
,
2759 MEM_CGROUP_STAT_SWAP
);
2763 val
= page_counter_read(&memcg
->memory
);
2765 val
= page_counter_read(&memcg
->memsw
);
2778 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2781 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2782 struct page_counter
*counter
;
2784 switch (MEMFILE_TYPE(cft
->private)) {
2786 counter
= &memcg
->memory
;
2789 counter
= &memcg
->memsw
;
2792 counter
= &memcg
->kmem
;
2795 counter
= &memcg
->tcpmem
;
2801 switch (MEMFILE_ATTR(cft
->private)) {
2803 if (counter
== &memcg
->memory
)
2804 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
2805 if (counter
== &memcg
->memsw
)
2806 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
2807 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2809 return (u64
)counter
->limit
* PAGE_SIZE
;
2811 return (u64
)counter
->watermark
* PAGE_SIZE
;
2813 return counter
->failcnt
;
2814 case RES_SOFT_LIMIT
:
2815 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2822 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2826 if (cgroup_memory_nokmem
)
2829 BUG_ON(memcg
->kmemcg_id
>= 0);
2830 BUG_ON(memcg
->kmem_state
);
2832 memcg_id
= memcg_alloc_cache_id();
2836 static_branch_inc(&memcg_kmem_enabled_key
);
2838 * A memory cgroup is considered kmem-online as soon as it gets
2839 * kmemcg_id. Setting the id after enabling static branching will
2840 * guarantee no one starts accounting before all call sites are
2843 memcg
->kmemcg_id
= memcg_id
;
2844 memcg
->kmem_state
= KMEM_ONLINE
;
2849 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2851 struct cgroup_subsys_state
*css
;
2852 struct mem_cgroup
*parent
, *child
;
2855 if (memcg
->kmem_state
!= KMEM_ONLINE
)
2858 * Clear the online state before clearing memcg_caches array
2859 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2860 * guarantees that no cache will be created for this cgroup
2861 * after we are done (see memcg_create_kmem_cache()).
2863 memcg
->kmem_state
= KMEM_ALLOCATED
;
2865 memcg_deactivate_kmem_caches(memcg
);
2867 kmemcg_id
= memcg
->kmemcg_id
;
2868 BUG_ON(kmemcg_id
< 0);
2870 parent
= parent_mem_cgroup(memcg
);
2872 parent
= root_mem_cgroup
;
2875 * Change kmemcg_id of this cgroup and all its descendants to the
2876 * parent's id, and then move all entries from this cgroup's list_lrus
2877 * to ones of the parent. After we have finished, all list_lrus
2878 * corresponding to this cgroup are guaranteed to remain empty. The
2879 * ordering is imposed by list_lru_node->lock taken by
2880 * memcg_drain_all_list_lrus().
2882 css_for_each_descendant_pre(css
, &memcg
->css
) {
2883 child
= mem_cgroup_from_css(css
);
2884 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
2885 child
->kmemcg_id
= parent
->kmemcg_id
;
2886 if (!memcg
->use_hierarchy
)
2889 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
2891 memcg_free_cache_id(kmemcg_id
);
2894 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2896 /* css_alloc() failed, offlining didn't happen */
2897 if (unlikely(memcg
->kmem_state
== KMEM_ONLINE
))
2898 memcg_offline_kmem(memcg
);
2900 if (memcg
->kmem_state
== KMEM_ALLOCATED
) {
2901 memcg_destroy_kmem_caches(memcg
);
2902 static_branch_dec(&memcg_kmem_enabled_key
);
2903 WARN_ON(page_counter_read(&memcg
->kmem
));
2907 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
2911 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
2914 static void memcg_free_kmem(struct mem_cgroup
*memcg
)
2917 #endif /* !CONFIG_SLOB */
2919 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2920 unsigned long limit
)
2924 mutex_lock(&memcg_limit_mutex
);
2925 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2926 mutex_unlock(&memcg_limit_mutex
);
2930 static int memcg_update_tcp_limit(struct mem_cgroup
*memcg
, unsigned long limit
)
2934 mutex_lock(&memcg_limit_mutex
);
2936 ret
= page_counter_limit(&memcg
->tcpmem
, limit
);
2940 if (!memcg
->tcpmem_active
) {
2942 * The active flag needs to be written after the static_key
2943 * update. This is what guarantees that the socket activation
2944 * function is the last one to run. See sock_update_memcg() for
2945 * details, and note that we don't mark any socket as belonging
2946 * to this memcg until that flag is up.
2948 * We need to do this, because static_keys will span multiple
2949 * sites, but we can't control their order. If we mark a socket
2950 * as accounted, but the accounting functions are not patched in
2951 * yet, we'll lose accounting.
2953 * We never race with the readers in sock_update_memcg(),
2954 * because when this value change, the code to process it is not
2957 static_branch_inc(&memcg_sockets_enabled_key
);
2958 memcg
->tcpmem_active
= true;
2961 mutex_unlock(&memcg_limit_mutex
);
2966 * The user of this function is...
2969 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
2970 char *buf
, size_t nbytes
, loff_t off
)
2972 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2973 unsigned long nr_pages
;
2976 buf
= strstrip(buf
);
2977 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
2981 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
2983 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
2987 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
2989 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
2992 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
2995 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
2998 ret
= memcg_update_tcp_limit(memcg
, nr_pages
);
3002 case RES_SOFT_LIMIT
:
3003 memcg
->soft_limit
= nr_pages
;
3007 return ret
?: nbytes
;
3010 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3011 size_t nbytes
, loff_t off
)
3013 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3014 struct page_counter
*counter
;
3016 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3018 counter
= &memcg
->memory
;
3021 counter
= &memcg
->memsw
;
3024 counter
= &memcg
->kmem
;
3027 counter
= &memcg
->tcpmem
;
3033 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3035 page_counter_reset_watermark(counter
);
3038 counter
->failcnt
= 0;
3047 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3050 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3054 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3055 struct cftype
*cft
, u64 val
)
3057 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3059 if (val
& ~MOVE_MASK
)
3063 * No kind of locking is needed in here, because ->can_attach() will
3064 * check this value once in the beginning of the process, and then carry
3065 * on with stale data. This means that changes to this value will only
3066 * affect task migrations starting after the change.
3068 memcg
->move_charge_at_immigrate
= val
;
3072 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3073 struct cftype
*cft
, u64 val
)
3080 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3084 unsigned int lru_mask
;
3087 static const struct numa_stat stats
[] = {
3088 { "total", LRU_ALL
},
3089 { "file", LRU_ALL_FILE
},
3090 { "anon", LRU_ALL_ANON
},
3091 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3093 const struct numa_stat
*stat
;
3096 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3098 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3099 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3100 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3101 for_each_node_state(nid
, N_MEMORY
) {
3102 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3104 seq_printf(m
, " N%d=%lu", nid
, nr
);
3109 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3110 struct mem_cgroup
*iter
;
3113 for_each_mem_cgroup_tree(iter
, memcg
)
3114 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3115 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3116 for_each_node_state(nid
, N_MEMORY
) {
3118 for_each_mem_cgroup_tree(iter
, memcg
)
3119 nr
+= mem_cgroup_node_nr_lru_pages(
3120 iter
, nid
, stat
->lru_mask
);
3121 seq_printf(m
, " N%d=%lu", nid
, nr
);
3128 #endif /* CONFIG_NUMA */
3130 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3132 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3133 unsigned long memory
, memsw
;
3134 struct mem_cgroup
*mi
;
3137 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3138 MEM_CGROUP_STAT_NSTATS
);
3139 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3140 MEM_CGROUP_EVENTS_NSTATS
);
3141 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3143 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3144 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3146 seq_printf(m
, "%s %lu\n", mem_cgroup_stat_names
[i
],
3147 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3150 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3151 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3152 mem_cgroup_read_events(memcg
, i
));
3154 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3155 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3156 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3158 /* Hierarchical information */
3159 memory
= memsw
= PAGE_COUNTER_MAX
;
3160 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3161 memory
= min(memory
, mi
->memory
.limit
);
3162 memsw
= min(memsw
, mi
->memsw
.limit
);
3164 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3165 (u64
)memory
* PAGE_SIZE
);
3166 if (do_memsw_account())
3167 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3168 (u64
)memsw
* PAGE_SIZE
);
3170 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3171 unsigned long long val
= 0;
3173 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_memsw_account())
3175 for_each_mem_cgroup_tree(mi
, memcg
)
3176 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3177 seq_printf(m
, "total_%s %llu\n", mem_cgroup_stat_names
[i
], val
);
3180 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3181 unsigned long long val
= 0;
3183 for_each_mem_cgroup_tree(mi
, memcg
)
3184 val
+= mem_cgroup_read_events(mi
, i
);
3185 seq_printf(m
, "total_%s %llu\n",
3186 mem_cgroup_events_names
[i
], val
);
3189 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3190 unsigned long long val
= 0;
3192 for_each_mem_cgroup_tree(mi
, memcg
)
3193 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3194 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3197 #ifdef CONFIG_DEBUG_VM
3200 struct mem_cgroup_per_zone
*mz
;
3201 struct zone_reclaim_stat
*rstat
;
3202 unsigned long recent_rotated
[2] = {0, 0};
3203 unsigned long recent_scanned
[2] = {0, 0};
3205 for_each_online_node(nid
)
3206 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3207 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3208 rstat
= &mz
->lruvec
.reclaim_stat
;
3210 recent_rotated
[0] += rstat
->recent_rotated
[0];
3211 recent_rotated
[1] += rstat
->recent_rotated
[1];
3212 recent_scanned
[0] += rstat
->recent_scanned
[0];
3213 recent_scanned
[1] += rstat
->recent_scanned
[1];
3215 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3216 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3217 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3218 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3225 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3228 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3230 return mem_cgroup_swappiness(memcg
);
3233 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3234 struct cftype
*cft
, u64 val
)
3236 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3242 memcg
->swappiness
= val
;
3244 vm_swappiness
= val
;
3249 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3251 struct mem_cgroup_threshold_ary
*t
;
3252 unsigned long usage
;
3257 t
= rcu_dereference(memcg
->thresholds
.primary
);
3259 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3264 usage
= mem_cgroup_usage(memcg
, swap
);
3267 * current_threshold points to threshold just below or equal to usage.
3268 * If it's not true, a threshold was crossed after last
3269 * call of __mem_cgroup_threshold().
3271 i
= t
->current_threshold
;
3274 * Iterate backward over array of thresholds starting from
3275 * current_threshold and check if a threshold is crossed.
3276 * If none of thresholds below usage is crossed, we read
3277 * only one element of the array here.
3279 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3280 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3282 /* i = current_threshold + 1 */
3286 * Iterate forward over array of thresholds starting from
3287 * current_threshold+1 and check if a threshold is crossed.
3288 * If none of thresholds above usage is crossed, we read
3289 * only one element of the array here.
3291 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3292 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3294 /* Update current_threshold */
3295 t
->current_threshold
= i
- 1;
3300 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3303 __mem_cgroup_threshold(memcg
, false);
3304 if (do_memsw_account())
3305 __mem_cgroup_threshold(memcg
, true);
3307 memcg
= parent_mem_cgroup(memcg
);
3311 static int compare_thresholds(const void *a
, const void *b
)
3313 const struct mem_cgroup_threshold
*_a
= a
;
3314 const struct mem_cgroup_threshold
*_b
= b
;
3316 if (_a
->threshold
> _b
->threshold
)
3319 if (_a
->threshold
< _b
->threshold
)
3325 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3327 struct mem_cgroup_eventfd_list
*ev
;
3329 spin_lock(&memcg_oom_lock
);
3331 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3332 eventfd_signal(ev
->eventfd
, 1);
3334 spin_unlock(&memcg_oom_lock
);
3338 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3340 struct mem_cgroup
*iter
;
3342 for_each_mem_cgroup_tree(iter
, memcg
)
3343 mem_cgroup_oom_notify_cb(iter
);
3346 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3347 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3349 struct mem_cgroup_thresholds
*thresholds
;
3350 struct mem_cgroup_threshold_ary
*new;
3351 unsigned long threshold
;
3352 unsigned long usage
;
3355 ret
= page_counter_memparse(args
, "-1", &threshold
);
3359 mutex_lock(&memcg
->thresholds_lock
);
3362 thresholds
= &memcg
->thresholds
;
3363 usage
= mem_cgroup_usage(memcg
, false);
3364 } else if (type
== _MEMSWAP
) {
3365 thresholds
= &memcg
->memsw_thresholds
;
3366 usage
= mem_cgroup_usage(memcg
, true);
3370 /* Check if a threshold crossed before adding a new one */
3371 if (thresholds
->primary
)
3372 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3374 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3376 /* Allocate memory for new array of thresholds */
3377 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3385 /* Copy thresholds (if any) to new array */
3386 if (thresholds
->primary
) {
3387 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3388 sizeof(struct mem_cgroup_threshold
));
3391 /* Add new threshold */
3392 new->entries
[size
- 1].eventfd
= eventfd
;
3393 new->entries
[size
- 1].threshold
= threshold
;
3395 /* Sort thresholds. Registering of new threshold isn't time-critical */
3396 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3397 compare_thresholds
, NULL
);
3399 /* Find current threshold */
3400 new->current_threshold
= -1;
3401 for (i
= 0; i
< size
; i
++) {
3402 if (new->entries
[i
].threshold
<= usage
) {
3404 * new->current_threshold will not be used until
3405 * rcu_assign_pointer(), so it's safe to increment
3408 ++new->current_threshold
;
3413 /* Free old spare buffer and save old primary buffer as spare */
3414 kfree(thresholds
->spare
);
3415 thresholds
->spare
= thresholds
->primary
;
3417 rcu_assign_pointer(thresholds
->primary
, new);
3419 /* To be sure that nobody uses thresholds */
3423 mutex_unlock(&memcg
->thresholds_lock
);
3428 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3429 struct eventfd_ctx
*eventfd
, const char *args
)
3431 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3434 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3435 struct eventfd_ctx
*eventfd
, const char *args
)
3437 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3440 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3441 struct eventfd_ctx
*eventfd
, enum res_type type
)
3443 struct mem_cgroup_thresholds
*thresholds
;
3444 struct mem_cgroup_threshold_ary
*new;
3445 unsigned long usage
;
3448 mutex_lock(&memcg
->thresholds_lock
);
3451 thresholds
= &memcg
->thresholds
;
3452 usage
= mem_cgroup_usage(memcg
, false);
3453 } else if (type
== _MEMSWAP
) {
3454 thresholds
= &memcg
->memsw_thresholds
;
3455 usage
= mem_cgroup_usage(memcg
, true);
3459 if (!thresholds
->primary
)
3462 /* Check if a threshold crossed before removing */
3463 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3465 /* Calculate new number of threshold */
3467 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3468 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3472 new = thresholds
->spare
;
3474 /* Set thresholds array to NULL if we don't have thresholds */
3483 /* Copy thresholds and find current threshold */
3484 new->current_threshold
= -1;
3485 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3486 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3489 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3490 if (new->entries
[j
].threshold
<= usage
) {
3492 * new->current_threshold will not be used
3493 * until rcu_assign_pointer(), so it's safe to increment
3496 ++new->current_threshold
;
3502 /* Swap primary and spare array */
3503 thresholds
->spare
= thresholds
->primary
;
3505 rcu_assign_pointer(thresholds
->primary
, new);
3507 /* To be sure that nobody uses thresholds */
3510 /* If all events are unregistered, free the spare array */
3512 kfree(thresholds
->spare
);
3513 thresholds
->spare
= NULL
;
3516 mutex_unlock(&memcg
->thresholds_lock
);
3519 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3520 struct eventfd_ctx
*eventfd
)
3522 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3525 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3526 struct eventfd_ctx
*eventfd
)
3528 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3531 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3532 struct eventfd_ctx
*eventfd
, const char *args
)
3534 struct mem_cgroup_eventfd_list
*event
;
3536 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3540 spin_lock(&memcg_oom_lock
);
3542 event
->eventfd
= eventfd
;
3543 list_add(&event
->list
, &memcg
->oom_notify
);
3545 /* already in OOM ? */
3546 if (memcg
->under_oom
)
3547 eventfd_signal(eventfd
, 1);
3548 spin_unlock(&memcg_oom_lock
);
3553 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3554 struct eventfd_ctx
*eventfd
)
3556 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3558 spin_lock(&memcg_oom_lock
);
3560 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3561 if (ev
->eventfd
== eventfd
) {
3562 list_del(&ev
->list
);
3567 spin_unlock(&memcg_oom_lock
);
3570 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3572 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3574 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3575 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3579 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3580 struct cftype
*cft
, u64 val
)
3582 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3584 /* cannot set to root cgroup and only 0 and 1 are allowed */
3585 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3588 memcg
->oom_kill_disable
= val
;
3590 memcg_oom_recover(memcg
);
3595 #ifdef CONFIG_CGROUP_WRITEBACK
3597 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3599 return &memcg
->cgwb_list
;
3602 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3604 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3607 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3609 wb_domain_exit(&memcg
->cgwb_domain
);
3612 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3614 wb_domain_size_changed(&memcg
->cgwb_domain
);
3617 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3619 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3621 if (!memcg
->css
.parent
)
3624 return &memcg
->cgwb_domain
;
3628 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3629 * @wb: bdi_writeback in question
3630 * @pfilepages: out parameter for number of file pages
3631 * @pheadroom: out parameter for number of allocatable pages according to memcg
3632 * @pdirty: out parameter for number of dirty pages
3633 * @pwriteback: out parameter for number of pages under writeback
3635 * Determine the numbers of file, headroom, dirty, and writeback pages in
3636 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3637 * is a bit more involved.
3639 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3640 * headroom is calculated as the lowest headroom of itself and the
3641 * ancestors. Note that this doesn't consider the actual amount of
3642 * available memory in the system. The caller should further cap
3643 * *@pheadroom accordingly.
3645 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
3646 unsigned long *pheadroom
, unsigned long *pdirty
,
3647 unsigned long *pwriteback
)
3649 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3650 struct mem_cgroup
*parent
;
3652 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3654 /* this should eventually include NR_UNSTABLE_NFS */
3655 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3656 *pfilepages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3657 (1 << LRU_ACTIVE_FILE
));
3658 *pheadroom
= PAGE_COUNTER_MAX
;
3660 while ((parent
= parent_mem_cgroup(memcg
))) {
3661 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3662 unsigned long used
= page_counter_read(&memcg
->memory
);
3664 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
3669 #else /* CONFIG_CGROUP_WRITEBACK */
3671 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3676 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3680 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3684 #endif /* CONFIG_CGROUP_WRITEBACK */
3687 * DO NOT USE IN NEW FILES.
3689 * "cgroup.event_control" implementation.
3691 * This is way over-engineered. It tries to support fully configurable
3692 * events for each user. Such level of flexibility is completely
3693 * unnecessary especially in the light of the planned unified hierarchy.
3695 * Please deprecate this and replace with something simpler if at all
3700 * Unregister event and free resources.
3702 * Gets called from workqueue.
3704 static void memcg_event_remove(struct work_struct
*work
)
3706 struct mem_cgroup_event
*event
=
3707 container_of(work
, struct mem_cgroup_event
, remove
);
3708 struct mem_cgroup
*memcg
= event
->memcg
;
3710 remove_wait_queue(event
->wqh
, &event
->wait
);
3712 event
->unregister_event(memcg
, event
->eventfd
);
3714 /* Notify userspace the event is going away. */
3715 eventfd_signal(event
->eventfd
, 1);
3717 eventfd_ctx_put(event
->eventfd
);
3719 css_put(&memcg
->css
);
3723 * Gets called on POLLHUP on eventfd when user closes it.
3725 * Called with wqh->lock held and interrupts disabled.
3727 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3728 int sync
, void *key
)
3730 struct mem_cgroup_event
*event
=
3731 container_of(wait
, struct mem_cgroup_event
, wait
);
3732 struct mem_cgroup
*memcg
= event
->memcg
;
3733 unsigned long flags
= (unsigned long)key
;
3735 if (flags
& POLLHUP
) {
3737 * If the event has been detached at cgroup removal, we
3738 * can simply return knowing the other side will cleanup
3741 * We can't race against event freeing since the other
3742 * side will require wqh->lock via remove_wait_queue(),
3745 spin_lock(&memcg
->event_list_lock
);
3746 if (!list_empty(&event
->list
)) {
3747 list_del_init(&event
->list
);
3749 * We are in atomic context, but cgroup_event_remove()
3750 * may sleep, so we have to call it in workqueue.
3752 schedule_work(&event
->remove
);
3754 spin_unlock(&memcg
->event_list_lock
);
3760 static void memcg_event_ptable_queue_proc(struct file
*file
,
3761 wait_queue_head_t
*wqh
, poll_table
*pt
)
3763 struct mem_cgroup_event
*event
=
3764 container_of(pt
, struct mem_cgroup_event
, pt
);
3767 add_wait_queue(wqh
, &event
->wait
);
3771 * DO NOT USE IN NEW FILES.
3773 * Parse input and register new cgroup event handler.
3775 * Input must be in format '<event_fd> <control_fd> <args>'.
3776 * Interpretation of args is defined by control file implementation.
3778 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3779 char *buf
, size_t nbytes
, loff_t off
)
3781 struct cgroup_subsys_state
*css
= of_css(of
);
3782 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3783 struct mem_cgroup_event
*event
;
3784 struct cgroup_subsys_state
*cfile_css
;
3785 unsigned int efd
, cfd
;
3792 buf
= strstrip(buf
);
3794 efd
= simple_strtoul(buf
, &endp
, 10);
3799 cfd
= simple_strtoul(buf
, &endp
, 10);
3800 if ((*endp
!= ' ') && (*endp
!= '\0'))
3804 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3808 event
->memcg
= memcg
;
3809 INIT_LIST_HEAD(&event
->list
);
3810 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3811 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3812 INIT_WORK(&event
->remove
, memcg_event_remove
);
3820 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3821 if (IS_ERR(event
->eventfd
)) {
3822 ret
= PTR_ERR(event
->eventfd
);
3829 goto out_put_eventfd
;
3832 /* the process need read permission on control file */
3833 /* AV: shouldn't we check that it's been opened for read instead? */
3834 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3839 * Determine the event callbacks and set them in @event. This used
3840 * to be done via struct cftype but cgroup core no longer knows
3841 * about these events. The following is crude but the whole thing
3842 * is for compatibility anyway.
3844 * DO NOT ADD NEW FILES.
3846 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3848 if (!strcmp(name
, "memory.usage_in_bytes")) {
3849 event
->register_event
= mem_cgroup_usage_register_event
;
3850 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3851 } else if (!strcmp(name
, "memory.oom_control")) {
3852 event
->register_event
= mem_cgroup_oom_register_event
;
3853 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3854 } else if (!strcmp(name
, "memory.pressure_level")) {
3855 event
->register_event
= vmpressure_register_event
;
3856 event
->unregister_event
= vmpressure_unregister_event
;
3857 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3858 event
->register_event
= memsw_cgroup_usage_register_event
;
3859 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3866 * Verify @cfile should belong to @css. Also, remaining events are
3867 * automatically removed on cgroup destruction but the removal is
3868 * asynchronous, so take an extra ref on @css.
3870 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3871 &memory_cgrp_subsys
);
3873 if (IS_ERR(cfile_css
))
3875 if (cfile_css
!= css
) {
3880 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3884 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3886 spin_lock(&memcg
->event_list_lock
);
3887 list_add(&event
->list
, &memcg
->event_list
);
3888 spin_unlock(&memcg
->event_list_lock
);
3900 eventfd_ctx_put(event
->eventfd
);
3909 static struct cftype mem_cgroup_legacy_files
[] = {
3911 .name
= "usage_in_bytes",
3912 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
3913 .read_u64
= mem_cgroup_read_u64
,
3916 .name
= "max_usage_in_bytes",
3917 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
3918 .write
= mem_cgroup_reset
,
3919 .read_u64
= mem_cgroup_read_u64
,
3922 .name
= "limit_in_bytes",
3923 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
3924 .write
= mem_cgroup_write
,
3925 .read_u64
= mem_cgroup_read_u64
,
3928 .name
= "soft_limit_in_bytes",
3929 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
3930 .write
= mem_cgroup_write
,
3931 .read_u64
= mem_cgroup_read_u64
,
3935 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
3936 .write
= mem_cgroup_reset
,
3937 .read_u64
= mem_cgroup_read_u64
,
3941 .seq_show
= memcg_stat_show
,
3944 .name
= "force_empty",
3945 .write
= mem_cgroup_force_empty_write
,
3948 .name
= "use_hierarchy",
3949 .write_u64
= mem_cgroup_hierarchy_write
,
3950 .read_u64
= mem_cgroup_hierarchy_read
,
3953 .name
= "cgroup.event_control", /* XXX: for compat */
3954 .write
= memcg_write_event_control
,
3955 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
3958 .name
= "swappiness",
3959 .read_u64
= mem_cgroup_swappiness_read
,
3960 .write_u64
= mem_cgroup_swappiness_write
,
3963 .name
= "move_charge_at_immigrate",
3964 .read_u64
= mem_cgroup_move_charge_read
,
3965 .write_u64
= mem_cgroup_move_charge_write
,
3968 .name
= "oom_control",
3969 .seq_show
= mem_cgroup_oom_control_read
,
3970 .write_u64
= mem_cgroup_oom_control_write
,
3971 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
3974 .name
= "pressure_level",
3978 .name
= "numa_stat",
3979 .seq_show
= memcg_numa_stat_show
,
3983 .name
= "kmem.limit_in_bytes",
3984 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
3985 .write
= mem_cgroup_write
,
3986 .read_u64
= mem_cgroup_read_u64
,
3989 .name
= "kmem.usage_in_bytes",
3990 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
3991 .read_u64
= mem_cgroup_read_u64
,
3994 .name
= "kmem.failcnt",
3995 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
3996 .write
= mem_cgroup_reset
,
3997 .read_u64
= mem_cgroup_read_u64
,
4000 .name
= "kmem.max_usage_in_bytes",
4001 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4002 .write
= mem_cgroup_reset
,
4003 .read_u64
= mem_cgroup_read_u64
,
4005 #ifdef CONFIG_SLABINFO
4007 .name
= "kmem.slabinfo",
4008 .seq_start
= slab_start
,
4009 .seq_next
= slab_next
,
4010 .seq_stop
= slab_stop
,
4011 .seq_show
= memcg_slab_show
,
4015 .name
= "kmem.tcp.limit_in_bytes",
4016 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
4017 .write
= mem_cgroup_write
,
4018 .read_u64
= mem_cgroup_read_u64
,
4021 .name
= "kmem.tcp.usage_in_bytes",
4022 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
4023 .read_u64
= mem_cgroup_read_u64
,
4026 .name
= "kmem.tcp.failcnt",
4027 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
4028 .write
= mem_cgroup_reset
,
4029 .read_u64
= mem_cgroup_read_u64
,
4032 .name
= "kmem.tcp.max_usage_in_bytes",
4033 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
4034 .write
= mem_cgroup_reset
,
4035 .read_u64
= mem_cgroup_read_u64
,
4037 { }, /* terminate */
4040 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4042 struct mem_cgroup_per_node
*pn
;
4043 struct mem_cgroup_per_zone
*mz
;
4044 int zone
, tmp
= node
;
4046 * This routine is called against possible nodes.
4047 * But it's BUG to call kmalloc() against offline node.
4049 * TODO: this routine can waste much memory for nodes which will
4050 * never be onlined. It's better to use memory hotplug callback
4053 if (!node_state(node
, N_NORMAL_MEMORY
))
4055 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4059 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4060 mz
= &pn
->zoneinfo
[zone
];
4061 lruvec_init(&mz
->lruvec
);
4062 mz
->usage_in_excess
= 0;
4063 mz
->on_tree
= false;
4066 memcg
->nodeinfo
[node
] = pn
;
4070 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4072 kfree(memcg
->nodeinfo
[node
]);
4075 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
4079 memcg_wb_domain_exit(memcg
);
4081 free_mem_cgroup_per_zone_info(memcg
, node
);
4082 free_percpu(memcg
->stat
);
4086 static struct mem_cgroup
*mem_cgroup_alloc(void)
4088 struct mem_cgroup
*memcg
;
4092 size
= sizeof(struct mem_cgroup
);
4093 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4095 memcg
= kzalloc(size
, GFP_KERNEL
);
4099 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4104 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4107 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4110 INIT_WORK(&memcg
->high_work
, high_work_func
);
4111 memcg
->last_scanned_node
= MAX_NUMNODES
;
4112 INIT_LIST_HEAD(&memcg
->oom_notify
);
4113 mutex_init(&memcg
->thresholds_lock
);
4114 spin_lock_init(&memcg
->move_lock
);
4115 vmpressure_init(&memcg
->vmpressure
);
4116 INIT_LIST_HEAD(&memcg
->event_list
);
4117 spin_lock_init(&memcg
->event_list_lock
);
4118 memcg
->socket_pressure
= jiffies
;
4120 memcg
->kmemcg_id
= -1;
4122 #ifdef CONFIG_CGROUP_WRITEBACK
4123 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4127 mem_cgroup_free(memcg
);
4131 static struct cgroup_subsys_state
* __ref
4132 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4134 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
4135 struct mem_cgroup
*memcg
;
4136 long error
= -ENOMEM
;
4138 memcg
= mem_cgroup_alloc();
4140 return ERR_PTR(error
);
4142 memcg
->high
= PAGE_COUNTER_MAX
;
4143 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4145 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4146 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4148 if (parent
&& parent
->use_hierarchy
) {
4149 memcg
->use_hierarchy
= true;
4150 page_counter_init(&memcg
->memory
, &parent
->memory
);
4151 page_counter_init(&memcg
->swap
, &parent
->swap
);
4152 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4153 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4154 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
4156 page_counter_init(&memcg
->memory
, NULL
);
4157 page_counter_init(&memcg
->swap
, NULL
);
4158 page_counter_init(&memcg
->memsw
, NULL
);
4159 page_counter_init(&memcg
->kmem
, NULL
);
4160 page_counter_init(&memcg
->tcpmem
, NULL
);
4162 * Deeper hierachy with use_hierarchy == false doesn't make
4163 * much sense so let cgroup subsystem know about this
4164 * unfortunate state in our controller.
4166 if (parent
!= root_mem_cgroup
)
4167 memory_cgrp_subsys
.broken_hierarchy
= true;
4170 /* The following stuff does not apply to the root */
4172 root_mem_cgroup
= memcg
;
4176 error
= memcg_online_kmem(memcg
);
4180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4181 static_branch_inc(&memcg_sockets_enabled_key
);
4185 mem_cgroup_free(memcg
);
4190 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4192 if (css
->id
> MEM_CGROUP_ID_MAX
)
4198 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4200 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4201 struct mem_cgroup_event
*event
, *tmp
;
4204 * Unregister events and notify userspace.
4205 * Notify userspace about cgroup removing only after rmdir of cgroup
4206 * directory to avoid race between userspace and kernelspace.
4208 spin_lock(&memcg
->event_list_lock
);
4209 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4210 list_del_init(&event
->list
);
4211 schedule_work(&event
->remove
);
4213 spin_unlock(&memcg
->event_list_lock
);
4215 memcg_offline_kmem(memcg
);
4216 wb_memcg_offline(memcg
);
4219 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
4221 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4223 invalidate_reclaim_iterators(memcg
);
4226 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4228 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4230 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
4231 static_branch_dec(&memcg_sockets_enabled_key
);
4233 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
4234 static_branch_dec(&memcg_sockets_enabled_key
);
4236 vmpressure_cleanup(&memcg
->vmpressure
);
4237 cancel_work_sync(&memcg
->high_work
);
4238 mem_cgroup_remove_from_trees(memcg
);
4239 memcg_free_kmem(memcg
);
4240 mem_cgroup_free(memcg
);
4244 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4245 * @css: the target css
4247 * Reset the states of the mem_cgroup associated with @css. This is
4248 * invoked when the userland requests disabling on the default hierarchy
4249 * but the memcg is pinned through dependency. The memcg should stop
4250 * applying policies and should revert to the vanilla state as it may be
4251 * made visible again.
4253 * The current implementation only resets the essential configurations.
4254 * This needs to be expanded to cover all the visible parts.
4256 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4258 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4260 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4261 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4262 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4264 memcg
->high
= PAGE_COUNTER_MAX
;
4265 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4266 memcg_wb_domain_size_changed(memcg
);
4270 /* Handlers for move charge at task migration. */
4271 static int mem_cgroup_do_precharge(unsigned long count
)
4275 /* Try a single bulk charge without reclaim first, kswapd may wake */
4276 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
4278 mc
.precharge
+= count
;
4282 /* Try charges one by one with reclaim */
4284 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4294 * get_mctgt_type - get target type of moving charge
4295 * @vma: the vma the pte to be checked belongs
4296 * @addr: the address corresponding to the pte to be checked
4297 * @ptent: the pte to be checked
4298 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4301 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4302 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4303 * move charge. if @target is not NULL, the page is stored in target->page
4304 * with extra refcnt got(Callers should handle it).
4305 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4306 * target for charge migration. if @target is not NULL, the entry is stored
4309 * Called with pte lock held.
4316 enum mc_target_type
{
4322 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4323 unsigned long addr
, pte_t ptent
)
4325 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4327 if (!page
|| !page_mapped(page
))
4329 if (PageAnon(page
)) {
4330 if (!(mc
.flags
& MOVE_ANON
))
4333 if (!(mc
.flags
& MOVE_FILE
))
4336 if (!get_page_unless_zero(page
))
4343 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4344 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4346 struct page
*page
= NULL
;
4347 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4349 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4352 * Because lookup_swap_cache() updates some statistics counter,
4353 * we call find_get_page() with swapper_space directly.
4355 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4356 if (do_memsw_account())
4357 entry
->val
= ent
.val
;
4362 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4363 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4369 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4370 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4372 struct page
*page
= NULL
;
4373 struct address_space
*mapping
;
4376 if (!vma
->vm_file
) /* anonymous vma */
4378 if (!(mc
.flags
& MOVE_FILE
))
4381 mapping
= vma
->vm_file
->f_mapping
;
4382 pgoff
= linear_page_index(vma
, addr
);
4384 /* page is moved even if it's not RSS of this task(page-faulted). */
4386 /* shmem/tmpfs may report page out on swap: account for that too. */
4387 if (shmem_mapping(mapping
)) {
4388 page
= find_get_entry(mapping
, pgoff
);
4389 if (radix_tree_exceptional_entry(page
)) {
4390 swp_entry_t swp
= radix_to_swp_entry(page
);
4391 if (do_memsw_account())
4393 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4396 page
= find_get_page(mapping
, pgoff
);
4398 page
= find_get_page(mapping
, pgoff
);
4404 * mem_cgroup_move_account - move account of the page
4406 * @nr_pages: number of regular pages (>1 for huge pages)
4407 * @from: mem_cgroup which the page is moved from.
4408 * @to: mem_cgroup which the page is moved to. @from != @to.
4410 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4412 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4415 static int mem_cgroup_move_account(struct page
*page
,
4417 struct mem_cgroup
*from
,
4418 struct mem_cgroup
*to
)
4420 unsigned long flags
;
4421 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
4425 VM_BUG_ON(from
== to
);
4426 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4427 VM_BUG_ON(compound
&& !PageTransHuge(page
));
4430 * Prevent mem_cgroup_migrate() from looking at
4431 * page->mem_cgroup of its source page while we change it.
4434 if (!trylock_page(page
))
4438 if (page
->mem_cgroup
!= from
)
4441 anon
= PageAnon(page
);
4443 spin_lock_irqsave(&from
->move_lock
, flags
);
4445 if (!anon
&& page_mapped(page
)) {
4446 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4448 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4453 * move_lock grabbed above and caller set from->moving_account, so
4454 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4455 * So mapping should be stable for dirty pages.
4457 if (!anon
&& PageDirty(page
)) {
4458 struct address_space
*mapping
= page_mapping(page
);
4460 if (mapping_cap_account_dirty(mapping
)) {
4461 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4463 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4468 if (PageWriteback(page
)) {
4469 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4471 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4476 * It is safe to change page->mem_cgroup here because the page
4477 * is referenced, charged, and isolated - we can't race with
4478 * uncharging, charging, migration, or LRU putback.
4481 /* caller should have done css_get */
4482 page
->mem_cgroup
= to
;
4483 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4487 local_irq_disable();
4488 mem_cgroup_charge_statistics(to
, page
, compound
, nr_pages
);
4489 memcg_check_events(to
, page
);
4490 mem_cgroup_charge_statistics(from
, page
, compound
, -nr_pages
);
4491 memcg_check_events(from
, page
);
4499 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4500 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4502 struct page
*page
= NULL
;
4503 enum mc_target_type ret
= MC_TARGET_NONE
;
4504 swp_entry_t ent
= { .val
= 0 };
4506 if (pte_present(ptent
))
4507 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4508 else if (is_swap_pte(ptent
))
4509 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4510 else if (pte_none(ptent
))
4511 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4513 if (!page
&& !ent
.val
)
4517 * Do only loose check w/o serialization.
4518 * mem_cgroup_move_account() checks the page is valid or
4519 * not under LRU exclusion.
4521 if (page
->mem_cgroup
== mc
.from
) {
4522 ret
= MC_TARGET_PAGE
;
4524 target
->page
= page
;
4526 if (!ret
|| !target
)
4529 /* There is a swap entry and a page doesn't exist or isn't charged */
4530 if (ent
.val
&& !ret
&&
4531 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4532 ret
= MC_TARGET_SWAP
;
4539 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4541 * We don't consider swapping or file mapped pages because THP does not
4542 * support them for now.
4543 * Caller should make sure that pmd_trans_huge(pmd) is true.
4545 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4546 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4548 struct page
*page
= NULL
;
4549 enum mc_target_type ret
= MC_TARGET_NONE
;
4551 page
= pmd_page(pmd
);
4552 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4553 if (!(mc
.flags
& MOVE_ANON
))
4555 if (page
->mem_cgroup
== mc
.from
) {
4556 ret
= MC_TARGET_PAGE
;
4559 target
->page
= page
;
4565 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4566 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4568 return MC_TARGET_NONE
;
4572 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4573 unsigned long addr
, unsigned long end
,
4574 struct mm_walk
*walk
)
4576 struct vm_area_struct
*vma
= walk
->vma
;
4580 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4582 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4583 mc
.precharge
+= HPAGE_PMD_NR
;
4588 if (pmd_trans_unstable(pmd
))
4590 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4591 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4592 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4593 mc
.precharge
++; /* increment precharge temporarily */
4594 pte_unmap_unlock(pte
- 1, ptl
);
4600 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4602 unsigned long precharge
;
4604 struct mm_walk mem_cgroup_count_precharge_walk
= {
4605 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4608 down_read(&mm
->mmap_sem
);
4609 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4610 up_read(&mm
->mmap_sem
);
4612 precharge
= mc
.precharge
;
4618 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4620 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4622 VM_BUG_ON(mc
.moving_task
);
4623 mc
.moving_task
= current
;
4624 return mem_cgroup_do_precharge(precharge
);
4627 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4628 static void __mem_cgroup_clear_mc(void)
4630 struct mem_cgroup
*from
= mc
.from
;
4631 struct mem_cgroup
*to
= mc
.to
;
4633 /* we must uncharge all the leftover precharges from mc.to */
4635 cancel_charge(mc
.to
, mc
.precharge
);
4639 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4640 * we must uncharge here.
4642 if (mc
.moved_charge
) {
4643 cancel_charge(mc
.from
, mc
.moved_charge
);
4644 mc
.moved_charge
= 0;
4646 /* we must fixup refcnts and charges */
4647 if (mc
.moved_swap
) {
4648 /* uncharge swap account from the old cgroup */
4649 if (!mem_cgroup_is_root(mc
.from
))
4650 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4653 * we charged both to->memory and to->memsw, so we
4654 * should uncharge to->memory.
4656 if (!mem_cgroup_is_root(mc
.to
))
4657 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4659 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4661 /* we've already done css_get(mc.to) */
4664 memcg_oom_recover(from
);
4665 memcg_oom_recover(to
);
4666 wake_up_all(&mc
.waitq
);
4669 static void mem_cgroup_clear_mc(void)
4672 * we must clear moving_task before waking up waiters at the end of
4675 mc
.moving_task
= NULL
;
4676 __mem_cgroup_clear_mc();
4677 spin_lock(&mc
.lock
);
4680 spin_unlock(&mc
.lock
);
4683 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4685 struct cgroup_subsys_state
*css
;
4686 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
4687 struct mem_cgroup
*from
;
4688 struct task_struct
*leader
, *p
;
4689 struct mm_struct
*mm
;
4690 unsigned long move_flags
;
4693 /* charge immigration isn't supported on the default hierarchy */
4694 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4698 * Multi-process migrations only happen on the default hierarchy
4699 * where charge immigration is not used. Perform charge
4700 * immigration if @tset contains a leader and whine if there are
4704 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
4707 memcg
= mem_cgroup_from_css(css
);
4713 * We are now commited to this value whatever it is. Changes in this
4714 * tunable will only affect upcoming migrations, not the current one.
4715 * So we need to save it, and keep it going.
4717 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4721 from
= mem_cgroup_from_task(p
);
4723 VM_BUG_ON(from
== memcg
);
4725 mm
= get_task_mm(p
);
4728 /* We move charges only when we move a owner of the mm */
4729 if (mm
->owner
== p
) {
4732 VM_BUG_ON(mc
.precharge
);
4733 VM_BUG_ON(mc
.moved_charge
);
4734 VM_BUG_ON(mc
.moved_swap
);
4736 spin_lock(&mc
.lock
);
4739 mc
.flags
= move_flags
;
4740 spin_unlock(&mc
.lock
);
4741 /* We set mc.moving_task later */
4743 ret
= mem_cgroup_precharge_mc(mm
);
4745 mem_cgroup_clear_mc();
4751 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4754 mem_cgroup_clear_mc();
4757 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4758 unsigned long addr
, unsigned long end
,
4759 struct mm_walk
*walk
)
4762 struct vm_area_struct
*vma
= walk
->vma
;
4765 enum mc_target_type target_type
;
4766 union mc_target target
;
4769 ptl
= pmd_trans_huge_lock(pmd
, vma
);
4771 if (mc
.precharge
< HPAGE_PMD_NR
) {
4775 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4776 if (target_type
== MC_TARGET_PAGE
) {
4778 if (!isolate_lru_page(page
)) {
4779 if (!mem_cgroup_move_account(page
, true,
4781 mc
.precharge
-= HPAGE_PMD_NR
;
4782 mc
.moved_charge
+= HPAGE_PMD_NR
;
4784 putback_lru_page(page
);
4792 if (pmd_trans_unstable(pmd
))
4795 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4796 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4797 pte_t ptent
= *(pte
++);
4803 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4804 case MC_TARGET_PAGE
:
4807 * We can have a part of the split pmd here. Moving it
4808 * can be done but it would be too convoluted so simply
4809 * ignore such a partial THP and keep it in original
4810 * memcg. There should be somebody mapping the head.
4812 if (PageTransCompound(page
))
4814 if (isolate_lru_page(page
))
4816 if (!mem_cgroup_move_account(page
, false,
4819 /* we uncharge from mc.from later. */
4822 putback_lru_page(page
);
4823 put
: /* get_mctgt_type() gets the page */
4826 case MC_TARGET_SWAP
:
4828 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4830 /* we fixup refcnts and charges later. */
4838 pte_unmap_unlock(pte
- 1, ptl
);
4843 * We have consumed all precharges we got in can_attach().
4844 * We try charge one by one, but don't do any additional
4845 * charges to mc.to if we have failed in charge once in attach()
4848 ret
= mem_cgroup_do_precharge(1);
4856 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4858 struct mm_walk mem_cgroup_move_charge_walk
= {
4859 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4863 lru_add_drain_all();
4865 * Signal lock_page_memcg() to take the memcg's move_lock
4866 * while we're moving its pages to another memcg. Then wait
4867 * for already started RCU-only updates to finish.
4869 atomic_inc(&mc
.from
->moving_account
);
4872 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4874 * Someone who are holding the mmap_sem might be waiting in
4875 * waitq. So we cancel all extra charges, wake up all waiters,
4876 * and retry. Because we cancel precharges, we might not be able
4877 * to move enough charges, but moving charge is a best-effort
4878 * feature anyway, so it wouldn't be a big problem.
4880 __mem_cgroup_clear_mc();
4885 * When we have consumed all precharges and failed in doing
4886 * additional charge, the page walk just aborts.
4888 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
4889 up_read(&mm
->mmap_sem
);
4890 atomic_dec(&mc
.from
->moving_account
);
4893 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4895 struct cgroup_subsys_state
*css
;
4896 struct task_struct
*p
= cgroup_taskset_first(tset
, &css
);
4897 struct mm_struct
*mm
= get_task_mm(p
);
4901 mem_cgroup_move_charge(mm
);
4905 mem_cgroup_clear_mc();
4907 #else /* !CONFIG_MMU */
4908 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
4912 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
4915 static void mem_cgroup_move_task(struct cgroup_taskset
*tset
)
4921 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4922 * to verify whether we're attached to the default hierarchy on each mount
4925 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
4928 * use_hierarchy is forced on the default hierarchy. cgroup core
4929 * guarantees that @root doesn't have any children, so turning it
4930 * on for the root memcg is enough.
4932 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
4933 root_mem_cgroup
->use_hierarchy
= true;
4935 root_mem_cgroup
->use_hierarchy
= false;
4938 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
4941 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4943 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
4946 static int memory_low_show(struct seq_file
*m
, void *v
)
4948 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4949 unsigned long low
= READ_ONCE(memcg
->low
);
4951 if (low
== PAGE_COUNTER_MAX
)
4952 seq_puts(m
, "max\n");
4954 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
4959 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
4960 char *buf
, size_t nbytes
, loff_t off
)
4962 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4966 buf
= strstrip(buf
);
4967 err
= page_counter_memparse(buf
, "max", &low
);
4976 static int memory_high_show(struct seq_file
*m
, void *v
)
4978 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4979 unsigned long high
= READ_ONCE(memcg
->high
);
4981 if (high
== PAGE_COUNTER_MAX
)
4982 seq_puts(m
, "max\n");
4984 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
4989 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
4990 char *buf
, size_t nbytes
, loff_t off
)
4992 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4996 buf
= strstrip(buf
);
4997 err
= page_counter_memparse(buf
, "max", &high
);
5003 memcg_wb_domain_size_changed(memcg
);
5007 static int memory_max_show(struct seq_file
*m
, void *v
)
5009 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5010 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5012 if (max
== PAGE_COUNTER_MAX
)
5013 seq_puts(m
, "max\n");
5015 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5020 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5021 char *buf
, size_t nbytes
, loff_t off
)
5023 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5027 buf
= strstrip(buf
);
5028 err
= page_counter_memparse(buf
, "max", &max
);
5032 err
= mem_cgroup_resize_limit(memcg
, max
);
5036 memcg_wb_domain_size_changed(memcg
);
5040 static int memory_events_show(struct seq_file
*m
, void *v
)
5042 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5044 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5045 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5046 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5047 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5052 static int memory_stat_show(struct seq_file
*m
, void *v
)
5054 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5055 unsigned long stat
[MEMCG_NR_STAT
];
5056 unsigned long events
[MEMCG_NR_EVENTS
];
5060 * Provide statistics on the state of the memory subsystem as
5061 * well as cumulative event counters that show past behavior.
5063 * This list is ordered following a combination of these gradients:
5064 * 1) generic big picture -> specifics and details
5065 * 2) reflecting userspace activity -> reflecting kernel heuristics
5067 * Current memory state:
5070 tree_stat(memcg
, stat
);
5071 tree_events(memcg
, events
);
5073 seq_printf(m
, "anon %llu\n",
5074 (u64
)stat
[MEM_CGROUP_STAT_RSS
] * PAGE_SIZE
);
5075 seq_printf(m
, "file %llu\n",
5076 (u64
)stat
[MEM_CGROUP_STAT_CACHE
] * PAGE_SIZE
);
5077 seq_printf(m
, "kernel_stack %llu\n",
5078 (u64
)stat
[MEMCG_KERNEL_STACK
] * PAGE_SIZE
);
5079 seq_printf(m
, "slab %llu\n",
5080 (u64
)(stat
[MEMCG_SLAB_RECLAIMABLE
] +
5081 stat
[MEMCG_SLAB_UNRECLAIMABLE
]) * PAGE_SIZE
);
5082 seq_printf(m
, "sock %llu\n",
5083 (u64
)stat
[MEMCG_SOCK
] * PAGE_SIZE
);
5085 seq_printf(m
, "file_mapped %llu\n",
5086 (u64
)stat
[MEM_CGROUP_STAT_FILE_MAPPED
] * PAGE_SIZE
);
5087 seq_printf(m
, "file_dirty %llu\n",
5088 (u64
)stat
[MEM_CGROUP_STAT_DIRTY
] * PAGE_SIZE
);
5089 seq_printf(m
, "file_writeback %llu\n",
5090 (u64
)stat
[MEM_CGROUP_STAT_WRITEBACK
] * PAGE_SIZE
);
5092 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5093 struct mem_cgroup
*mi
;
5094 unsigned long val
= 0;
5096 for_each_mem_cgroup_tree(mi
, memcg
)
5097 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
));
5098 seq_printf(m
, "%s %llu\n",
5099 mem_cgroup_lru_names
[i
], (u64
)val
* PAGE_SIZE
);
5102 seq_printf(m
, "slab_reclaimable %llu\n",
5103 (u64
)stat
[MEMCG_SLAB_RECLAIMABLE
] * PAGE_SIZE
);
5104 seq_printf(m
, "slab_unreclaimable %llu\n",
5105 (u64
)stat
[MEMCG_SLAB_UNRECLAIMABLE
] * PAGE_SIZE
);
5107 /* Accumulated memory events */
5109 seq_printf(m
, "pgfault %lu\n",
5110 events
[MEM_CGROUP_EVENTS_PGFAULT
]);
5111 seq_printf(m
, "pgmajfault %lu\n",
5112 events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
5117 static struct cftype memory_files
[] = {
5120 .flags
= CFTYPE_NOT_ON_ROOT
,
5121 .read_u64
= memory_current_read
,
5125 .flags
= CFTYPE_NOT_ON_ROOT
,
5126 .seq_show
= memory_low_show
,
5127 .write
= memory_low_write
,
5131 .flags
= CFTYPE_NOT_ON_ROOT
,
5132 .seq_show
= memory_high_show
,
5133 .write
= memory_high_write
,
5137 .flags
= CFTYPE_NOT_ON_ROOT
,
5138 .seq_show
= memory_max_show
,
5139 .write
= memory_max_write
,
5143 .flags
= CFTYPE_NOT_ON_ROOT
,
5144 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
5145 .seq_show
= memory_events_show
,
5149 .flags
= CFTYPE_NOT_ON_ROOT
,
5150 .seq_show
= memory_stat_show
,
5155 struct cgroup_subsys memory_cgrp_subsys
= {
5156 .css_alloc
= mem_cgroup_css_alloc
,
5157 .css_online
= mem_cgroup_css_online
,
5158 .css_offline
= mem_cgroup_css_offline
,
5159 .css_released
= mem_cgroup_css_released
,
5160 .css_free
= mem_cgroup_css_free
,
5161 .css_reset
= mem_cgroup_css_reset
,
5162 .can_attach
= mem_cgroup_can_attach
,
5163 .cancel_attach
= mem_cgroup_cancel_attach
,
5164 .attach
= mem_cgroup_move_task
,
5165 .bind
= mem_cgroup_bind
,
5166 .dfl_cftypes
= memory_files
,
5167 .legacy_cftypes
= mem_cgroup_legacy_files
,
5172 * mem_cgroup_low - check if memory consumption is below the normal range
5173 * @root: the highest ancestor to consider
5174 * @memcg: the memory cgroup to check
5176 * Returns %true if memory consumption of @memcg, and that of all
5177 * configurable ancestors up to @root, is below the normal range.
5179 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5181 if (mem_cgroup_disabled())
5185 * The toplevel group doesn't have a configurable range, so
5186 * it's never low when looked at directly, and it is not
5187 * considered an ancestor when assessing the hierarchy.
5190 if (memcg
== root_mem_cgroup
)
5193 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5196 while (memcg
!= root
) {
5197 memcg
= parent_mem_cgroup(memcg
);
5199 if (memcg
== root_mem_cgroup
)
5202 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5209 * mem_cgroup_try_charge - try charging a page
5210 * @page: page to charge
5211 * @mm: mm context of the victim
5212 * @gfp_mask: reclaim mode
5213 * @memcgp: charged memcg return
5215 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5216 * pages according to @gfp_mask if necessary.
5218 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5219 * Otherwise, an error code is returned.
5221 * After page->mapping has been set up, the caller must finalize the
5222 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5223 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5225 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5226 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
,
5229 struct mem_cgroup
*memcg
= NULL
;
5230 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5233 if (mem_cgroup_disabled())
5236 if (PageSwapCache(page
)) {
5238 * Every swap fault against a single page tries to charge the
5239 * page, bail as early as possible. shmem_unuse() encounters
5240 * already charged pages, too. The USED bit is protected by
5241 * the page lock, which serializes swap cache removal, which
5242 * in turn serializes uncharging.
5244 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5245 if (page
->mem_cgroup
)
5248 if (do_swap_account
) {
5249 swp_entry_t ent
= { .val
= page_private(page
), };
5250 unsigned short id
= lookup_swap_cgroup_id(ent
);
5253 memcg
= mem_cgroup_from_id(id
);
5254 if (memcg
&& !css_tryget_online(&memcg
->css
))
5261 memcg
= get_mem_cgroup_from_mm(mm
);
5263 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5265 css_put(&memcg
->css
);
5272 * mem_cgroup_commit_charge - commit a page charge
5273 * @page: page to charge
5274 * @memcg: memcg to charge the page to
5275 * @lrucare: page might be on LRU already
5277 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5278 * after page->mapping has been set up. This must happen atomically
5279 * as part of the page instantiation, i.e. under the page table lock
5280 * for anonymous pages, under the page lock for page and swap cache.
5282 * In addition, the page must not be on the LRU during the commit, to
5283 * prevent racing with task migration. If it might be, use @lrucare.
5285 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5287 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5288 bool lrucare
, bool compound
)
5290 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5292 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5293 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5295 if (mem_cgroup_disabled())
5298 * Swap faults will attempt to charge the same page multiple
5299 * times. But reuse_swap_page() might have removed the page
5300 * from swapcache already, so we can't check PageSwapCache().
5305 commit_charge(page
, memcg
, lrucare
);
5307 local_irq_disable();
5308 mem_cgroup_charge_statistics(memcg
, page
, compound
, nr_pages
);
5309 memcg_check_events(memcg
, page
);
5312 if (do_memsw_account() && PageSwapCache(page
)) {
5313 swp_entry_t entry
= { .val
= page_private(page
) };
5315 * The swap entry might not get freed for a long time,
5316 * let's not wait for it. The page already received a
5317 * memory+swap charge, drop the swap entry duplicate.
5319 mem_cgroup_uncharge_swap(entry
);
5324 * mem_cgroup_cancel_charge - cancel a page charge
5325 * @page: page to charge
5326 * @memcg: memcg to charge the page to
5328 * Cancel a charge transaction started by mem_cgroup_try_charge().
5330 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5333 unsigned int nr_pages
= compound
? hpage_nr_pages(page
) : 1;
5335 if (mem_cgroup_disabled())
5338 * Swap faults will attempt to charge the same page multiple
5339 * times. But reuse_swap_page() might have removed the page
5340 * from swapcache already, so we can't check PageSwapCache().
5345 cancel_charge(memcg
, nr_pages
);
5348 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5349 unsigned long nr_anon
, unsigned long nr_file
,
5350 unsigned long nr_huge
, struct page
*dummy_page
)
5352 unsigned long nr_pages
= nr_anon
+ nr_file
;
5353 unsigned long flags
;
5355 if (!mem_cgroup_is_root(memcg
)) {
5356 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5357 if (do_memsw_account())
5358 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5359 memcg_oom_recover(memcg
);
5362 local_irq_save(flags
);
5363 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5364 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5365 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5366 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5367 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5368 memcg_check_events(memcg
, dummy_page
);
5369 local_irq_restore(flags
);
5371 if (!mem_cgroup_is_root(memcg
))
5372 css_put_many(&memcg
->css
, nr_pages
);
5375 static void uncharge_list(struct list_head
*page_list
)
5377 struct mem_cgroup
*memcg
= NULL
;
5378 unsigned long nr_anon
= 0;
5379 unsigned long nr_file
= 0;
5380 unsigned long nr_huge
= 0;
5381 unsigned long pgpgout
= 0;
5382 struct list_head
*next
;
5385 next
= page_list
->next
;
5387 unsigned int nr_pages
= 1;
5389 page
= list_entry(next
, struct page
, lru
);
5390 next
= page
->lru
.next
;
5392 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5393 VM_BUG_ON_PAGE(page_count(page
), page
);
5395 if (!page
->mem_cgroup
)
5399 * Nobody should be changing or seriously looking at
5400 * page->mem_cgroup at this point, we have fully
5401 * exclusive access to the page.
5404 if (memcg
!= page
->mem_cgroup
) {
5406 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5408 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5410 memcg
= page
->mem_cgroup
;
5413 if (PageTransHuge(page
)) {
5414 nr_pages
<<= compound_order(page
);
5415 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5416 nr_huge
+= nr_pages
;
5420 nr_anon
+= nr_pages
;
5422 nr_file
+= nr_pages
;
5424 page
->mem_cgroup
= NULL
;
5427 } while (next
!= page_list
);
5430 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5435 * mem_cgroup_uncharge - uncharge a page
5436 * @page: page to uncharge
5438 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5439 * mem_cgroup_commit_charge().
5441 void mem_cgroup_uncharge(struct page
*page
)
5443 if (mem_cgroup_disabled())
5446 /* Don't touch page->lru of any random page, pre-check: */
5447 if (!page
->mem_cgroup
)
5450 INIT_LIST_HEAD(&page
->lru
);
5451 uncharge_list(&page
->lru
);
5455 * mem_cgroup_uncharge_list - uncharge a list of page
5456 * @page_list: list of pages to uncharge
5458 * Uncharge a list of pages previously charged with
5459 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5461 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5463 if (mem_cgroup_disabled())
5466 if (!list_empty(page_list
))
5467 uncharge_list(page_list
);
5471 * mem_cgroup_migrate - charge a page's replacement
5472 * @oldpage: currently circulating page
5473 * @newpage: replacement page
5475 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5476 * be uncharged upon free.
5478 * Both pages must be locked, @newpage->mapping must be set up.
5480 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
)
5482 struct mem_cgroup
*memcg
;
5483 unsigned int nr_pages
;
5486 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5487 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5488 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5489 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5492 if (mem_cgroup_disabled())
5495 /* Page cache replacement: new page already charged? */
5496 if (newpage
->mem_cgroup
)
5499 /* Swapcache readahead pages can get replaced before being charged */
5500 memcg
= oldpage
->mem_cgroup
;
5504 /* Force-charge the new page. The old one will be freed soon */
5505 compound
= PageTransHuge(newpage
);
5506 nr_pages
= compound
? hpage_nr_pages(newpage
) : 1;
5508 page_counter_charge(&memcg
->memory
, nr_pages
);
5509 if (do_memsw_account())
5510 page_counter_charge(&memcg
->memsw
, nr_pages
);
5511 css_get_many(&memcg
->css
, nr_pages
);
5513 commit_charge(newpage
, memcg
, false);
5515 local_irq_disable();
5516 mem_cgroup_charge_statistics(memcg
, newpage
, compound
, nr_pages
);
5517 memcg_check_events(memcg
, newpage
);
5521 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
5522 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
5524 void sock_update_memcg(struct sock
*sk
)
5526 struct mem_cgroup
*memcg
;
5528 /* Socket cloning can throw us here with sk_cgrp already
5529 * filled. It won't however, necessarily happen from
5530 * process context. So the test for root memcg given
5531 * the current task's memcg won't help us in this case.
5533 * Respecting the original socket's memcg is a better
5534 * decision in this case.
5537 BUG_ON(mem_cgroup_is_root(sk
->sk_memcg
));
5538 css_get(&sk
->sk_memcg
->css
);
5543 memcg
= mem_cgroup_from_task(current
);
5544 if (memcg
== root_mem_cgroup
)
5546 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
5548 if (css_tryget_online(&memcg
->css
))
5549 sk
->sk_memcg
= memcg
;
5553 EXPORT_SYMBOL(sock_update_memcg
);
5555 void sock_release_memcg(struct sock
*sk
)
5557 WARN_ON(!sk
->sk_memcg
);
5558 css_put(&sk
->sk_memcg
->css
);
5562 * mem_cgroup_charge_skmem - charge socket memory
5563 * @memcg: memcg to charge
5564 * @nr_pages: number of pages to charge
5566 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5567 * @memcg's configured limit, %false if the charge had to be forced.
5569 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5571 gfp_t gfp_mask
= GFP_KERNEL
;
5573 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5574 struct page_counter
*fail
;
5576 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
5577 memcg
->tcpmem_pressure
= 0;
5580 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
5581 memcg
->tcpmem_pressure
= 1;
5585 /* Don't block in the packet receive path */
5587 gfp_mask
= GFP_NOWAIT
;
5589 this_cpu_add(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5591 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0)
5594 try_charge(memcg
, gfp_mask
|__GFP_NOFAIL
, nr_pages
);
5599 * mem_cgroup_uncharge_skmem - uncharge socket memory
5600 * @memcg - memcg to uncharge
5601 * @nr_pages - number of pages to uncharge
5603 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
5605 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
5606 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
5610 this_cpu_sub(memcg
->stat
->count
[MEMCG_SOCK
], nr_pages
);
5612 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5613 css_put_many(&memcg
->css
, nr_pages
);
5616 static int __init
cgroup_memory(char *s
)
5620 while ((token
= strsep(&s
, ",")) != NULL
) {
5623 if (!strcmp(token
, "nosocket"))
5624 cgroup_memory_nosocket
= true;
5625 if (!strcmp(token
, "nokmem"))
5626 cgroup_memory_nokmem
= true;
5630 __setup("cgroup.memory=", cgroup_memory
);
5633 * subsys_initcall() for memory controller.
5635 * Some parts like hotcpu_notifier() have to be initialized from this context
5636 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5637 * everything that doesn't depend on a specific mem_cgroup structure should
5638 * be initialized from here.
5640 static int __init
mem_cgroup_init(void)
5644 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5646 for_each_possible_cpu(cpu
)
5647 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5650 for_each_node(node
) {
5651 struct mem_cgroup_tree_per_node
*rtpn
;
5654 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5655 node_online(node
) ? node
: NUMA_NO_NODE
);
5657 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5658 struct mem_cgroup_tree_per_zone
*rtpz
;
5660 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5661 rtpz
->rb_root
= RB_ROOT
;
5662 spin_lock_init(&rtpz
->lock
);
5664 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5669 subsys_initcall(mem_cgroup_init
);
5671 #ifdef CONFIG_MEMCG_SWAP
5673 * mem_cgroup_swapout - transfer a memsw charge to swap
5674 * @page: page whose memsw charge to transfer
5675 * @entry: swap entry to move the charge to
5677 * Transfer the memsw charge of @page to @entry.
5679 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5681 struct mem_cgroup
*memcg
;
5682 unsigned short oldid
;
5684 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5685 VM_BUG_ON_PAGE(page_count(page
), page
);
5687 if (!do_memsw_account())
5690 memcg
= page
->mem_cgroup
;
5692 /* Readahead page, never charged */
5696 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5697 VM_BUG_ON_PAGE(oldid
, page
);
5698 mem_cgroup_swap_statistics(memcg
, true);
5700 page
->mem_cgroup
= NULL
;
5702 if (!mem_cgroup_is_root(memcg
))
5703 page_counter_uncharge(&memcg
->memory
, 1);
5706 * Interrupts should be disabled here because the caller holds the
5707 * mapping->tree_lock lock which is taken with interrupts-off. It is
5708 * important here to have the interrupts disabled because it is the
5709 * only synchronisation we have for udpating the per-CPU variables.
5711 VM_BUG_ON(!irqs_disabled());
5712 mem_cgroup_charge_statistics(memcg
, page
, false, -1);
5713 memcg_check_events(memcg
, page
);
5717 * mem_cgroup_try_charge_swap - try charging a swap entry
5718 * @page: page being added to swap
5719 * @entry: swap entry to charge
5721 * Try to charge @entry to the memcg that @page belongs to.
5723 * Returns 0 on success, -ENOMEM on failure.
5725 int mem_cgroup_try_charge_swap(struct page
*page
, swp_entry_t entry
)
5727 struct mem_cgroup
*memcg
;
5728 struct page_counter
*counter
;
5729 unsigned short oldid
;
5731 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) || !do_swap_account
)
5734 memcg
= page
->mem_cgroup
;
5736 /* Readahead page, never charged */
5740 if (!mem_cgroup_is_root(memcg
) &&
5741 !page_counter_try_charge(&memcg
->swap
, 1, &counter
))
5744 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5745 VM_BUG_ON_PAGE(oldid
, page
);
5746 mem_cgroup_swap_statistics(memcg
, true);
5748 css_get(&memcg
->css
);
5753 * mem_cgroup_uncharge_swap - uncharge a swap entry
5754 * @entry: swap entry to uncharge
5756 * Drop the swap charge associated with @entry.
5758 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5760 struct mem_cgroup
*memcg
;
5763 if (!do_swap_account
)
5766 id
= swap_cgroup_record(entry
, 0);
5768 memcg
= mem_cgroup_from_id(id
);
5770 if (!mem_cgroup_is_root(memcg
)) {
5771 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5772 page_counter_uncharge(&memcg
->swap
, 1);
5774 page_counter_uncharge(&memcg
->memsw
, 1);
5776 mem_cgroup_swap_statistics(memcg
, false);
5777 css_put(&memcg
->css
);
5782 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
5784 long nr_swap_pages
= get_nr_swap_pages();
5786 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5787 return nr_swap_pages
;
5788 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5789 nr_swap_pages
= min_t(long, nr_swap_pages
,
5790 READ_ONCE(memcg
->swap
.limit
) -
5791 page_counter_read(&memcg
->swap
));
5792 return nr_swap_pages
;
5795 bool mem_cgroup_swap_full(struct page
*page
)
5797 struct mem_cgroup
*memcg
;
5799 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5803 if (!do_swap_account
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
5806 memcg
= page
->mem_cgroup
;
5810 for (; memcg
!= root_mem_cgroup
; memcg
= parent_mem_cgroup(memcg
))
5811 if (page_counter_read(&memcg
->swap
) * 2 >= memcg
->swap
.limit
)
5817 /* for remember boot option*/
5818 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5819 static int really_do_swap_account __initdata
= 1;
5821 static int really_do_swap_account __initdata
;
5824 static int __init
enable_swap_account(char *s
)
5826 if (!strcmp(s
, "1"))
5827 really_do_swap_account
= 1;
5828 else if (!strcmp(s
, "0"))
5829 really_do_swap_account
= 0;
5832 __setup("swapaccount=", enable_swap_account
);
5834 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
5837 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5839 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
5842 static int swap_max_show(struct seq_file
*m
, void *v
)
5844 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5845 unsigned long max
= READ_ONCE(memcg
->swap
.limit
);
5847 if (max
== PAGE_COUNTER_MAX
)
5848 seq_puts(m
, "max\n");
5850 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5855 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
5856 char *buf
, size_t nbytes
, loff_t off
)
5858 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5862 buf
= strstrip(buf
);
5863 err
= page_counter_memparse(buf
, "max", &max
);
5867 mutex_lock(&memcg_limit_mutex
);
5868 err
= page_counter_limit(&memcg
->swap
, max
);
5869 mutex_unlock(&memcg_limit_mutex
);
5876 static struct cftype swap_files
[] = {
5878 .name
= "swap.current",
5879 .flags
= CFTYPE_NOT_ON_ROOT
,
5880 .read_u64
= swap_current_read
,
5884 .flags
= CFTYPE_NOT_ON_ROOT
,
5885 .seq_show
= swap_max_show
,
5886 .write
= swap_max_write
,
5891 static struct cftype memsw_cgroup_files
[] = {
5893 .name
= "memsw.usage_in_bytes",
5894 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5895 .read_u64
= mem_cgroup_read_u64
,
5898 .name
= "memsw.max_usage_in_bytes",
5899 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5900 .write
= mem_cgroup_reset
,
5901 .read_u64
= mem_cgroup_read_u64
,
5904 .name
= "memsw.limit_in_bytes",
5905 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5906 .write
= mem_cgroup_write
,
5907 .read_u64
= mem_cgroup_read_u64
,
5910 .name
= "memsw.failcnt",
5911 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5912 .write
= mem_cgroup_reset
,
5913 .read_u64
= mem_cgroup_read_u64
,
5915 { }, /* terminate */
5918 static int __init
mem_cgroup_swap_init(void)
5920 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5921 do_swap_account
= 1;
5922 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
,
5924 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5925 memsw_cgroup_files
));
5929 subsys_initcall(mem_cgroup_swap_init
);
5931 #endif /* CONFIG_MEMCG_SWAP */