1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/memremap.h>
57 #include <linux/mm_inline.h>
58 #include <linux/swap_cgroup.h>
59 #include <linux/cpu.h>
60 #include <linux/oom.h>
61 #include <linux/lockdep.h>
62 #include <linux/file.h>
63 #include <linux/resume_user_mode.h>
64 #include <linux/psi.h>
65 #include <linux/seq_buf.h>
66 #include <linux/sched/isolation.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
78 EXPORT_SYMBOL(memory_cgrp_subsys
);
80 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
82 /* Active memory cgroup to use from an interrupt context */
83 DEFINE_PER_CPU(struct mem_cgroup
*, int_active_memcg
);
84 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg
);
86 /* Socket memory accounting disabled? */
87 static bool cgroup_memory_nosocket __ro_after_init
;
89 /* Kernel memory accounting disabled? */
90 static bool cgroup_memory_nokmem __ro_after_init
;
92 /* BPF memory accounting disabled? */
93 static bool cgroup_memory_nobpf __ro_after_init
;
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq
);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys
);
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node
{
114 struct rb_root rb_root
;
115 struct rb_node
*rb_rightmost
;
119 struct mem_cgroup_tree
{
120 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
126 struct mem_cgroup_eventfd_list
{
127 struct list_head list
;
128 struct eventfd_ctx
*eventfd
;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event
{
136 * memcg which the event belongs to.
138 struct mem_cgroup
*memcg
;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx
*eventfd
;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list
;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event
)(struct mem_cgroup
*memcg
,
153 struct eventfd_ctx
*eventfd
, const char *args
);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event
)(struct mem_cgroup
*memcg
,
160 struct eventfd_ctx
*eventfd
);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t
*wqh
;
167 wait_queue_entry_t wait
;
168 struct work_struct remove
;
171 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
172 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct
{
184 spinlock_t lock
; /* for from, to */
185 struct mm_struct
*mm
;
186 struct mem_cgroup
*from
;
187 struct mem_cgroup
*to
;
189 unsigned long precharge
;
190 unsigned long moved_charge
;
191 unsigned long moved_swap
;
192 struct task_struct
*moving_task
; /* a task moving charges */
193 wait_queue_head_t waitq
; /* a waitq for other context */
195 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
196 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
200 * Maximum loops in mem_cgroup_soft_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
214 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
215 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
216 #define MEMFILE_ATTR(val) ((val) & 0xffff)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool task_is_dying(void)
235 return tsk_is_oom_victim(current
) || fatal_signal_pending(current
) ||
236 (current
->flags
& PF_EXITING
);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
243 memcg
= root_mem_cgroup
;
244 return &memcg
->vmpressure
;
247 struct mem_cgroup
*vmpressure_to_memcg(struct vmpressure
*vmpr
)
249 return container_of(vmpr
, struct mem_cgroup
, vmpressure
);
252 #define CURRENT_OBJCG_UPDATE_BIT 0
253 #define CURRENT_OBJCG_UPDATE_FLAG (1UL << CURRENT_OBJCG_UPDATE_BIT)
255 #ifdef CONFIG_MEMCG_KMEM
256 static DEFINE_SPINLOCK(objcg_lock
);
258 bool mem_cgroup_kmem_disabled(void)
260 return cgroup_memory_nokmem
;
263 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
264 unsigned int nr_pages
);
266 static void obj_cgroup_release(struct percpu_ref
*ref
)
268 struct obj_cgroup
*objcg
= container_of(ref
, struct obj_cgroup
, refcnt
);
269 unsigned int nr_bytes
;
270 unsigned int nr_pages
;
274 * At this point all allocated objects are freed, and
275 * objcg->nr_charged_bytes can't have an arbitrary byte value.
276 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
278 * The following sequence can lead to it:
279 * 1) CPU0: objcg == stock->cached_objcg
280 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
281 * PAGE_SIZE bytes are charged
282 * 3) CPU1: a process from another memcg is allocating something,
283 * the stock if flushed,
284 * objcg->nr_charged_bytes = PAGE_SIZE - 92
285 * 5) CPU0: we do release this object,
286 * 92 bytes are added to stock->nr_bytes
287 * 6) CPU0: stock is flushed,
288 * 92 bytes are added to objcg->nr_charged_bytes
290 * In the result, nr_charged_bytes == PAGE_SIZE.
291 * This page will be uncharged in obj_cgroup_release().
293 nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
);
294 WARN_ON_ONCE(nr_bytes
& (PAGE_SIZE
- 1));
295 nr_pages
= nr_bytes
>> PAGE_SHIFT
;
298 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
300 spin_lock_irqsave(&objcg_lock
, flags
);
301 list_del(&objcg
->list
);
302 spin_unlock_irqrestore(&objcg_lock
, flags
);
304 percpu_ref_exit(ref
);
305 kfree_rcu(objcg
, rcu
);
308 static struct obj_cgroup
*obj_cgroup_alloc(void)
310 struct obj_cgroup
*objcg
;
313 objcg
= kzalloc(sizeof(struct obj_cgroup
), GFP_KERNEL
);
317 ret
= percpu_ref_init(&objcg
->refcnt
, obj_cgroup_release
, 0,
323 INIT_LIST_HEAD(&objcg
->list
);
327 static void memcg_reparent_objcgs(struct mem_cgroup
*memcg
,
328 struct mem_cgroup
*parent
)
330 struct obj_cgroup
*objcg
, *iter
;
332 objcg
= rcu_replace_pointer(memcg
->objcg
, NULL
, true);
334 spin_lock_irq(&objcg_lock
);
336 /* 1) Ready to reparent active objcg. */
337 list_add(&objcg
->list
, &memcg
->objcg_list
);
338 /* 2) Reparent active objcg and already reparented objcgs to parent. */
339 list_for_each_entry(iter
, &memcg
->objcg_list
, list
)
340 WRITE_ONCE(iter
->memcg
, parent
);
341 /* 3) Move already reparented objcgs to the parent's list */
342 list_splice(&memcg
->objcg_list
, &parent
->objcg_list
);
344 spin_unlock_irq(&objcg_lock
);
346 percpu_ref_kill(&objcg
->refcnt
);
350 * A lot of the calls to the cache allocation functions are expected to be
351 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
352 * conditional to this static branch, we'll have to allow modules that does
353 * kmem_cache_alloc and the such to see this symbol as well
355 DEFINE_STATIC_KEY_FALSE(memcg_kmem_online_key
);
356 EXPORT_SYMBOL(memcg_kmem_online_key
);
358 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key
);
359 EXPORT_SYMBOL(memcg_bpf_enabled_key
);
363 * mem_cgroup_css_from_folio - css of the memcg associated with a folio
364 * @folio: folio of interest
366 * If memcg is bound to the default hierarchy, css of the memcg associated
367 * with @folio is returned. The returned css remains associated with @folio
368 * until it is released.
370 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
373 struct cgroup_subsys_state
*mem_cgroup_css_from_folio(struct folio
*folio
)
375 struct mem_cgroup
*memcg
= folio_memcg(folio
);
377 if (!memcg
|| !cgroup_subsys_on_dfl(memory_cgrp_subsys
))
378 memcg
= root_mem_cgroup
;
384 * page_cgroup_ino - return inode number of the memcg a page is charged to
387 * Look up the closest online ancestor of the memory cgroup @page is charged to
388 * and return its inode number or 0 if @page is not charged to any cgroup. It
389 * is safe to call this function without holding a reference to @page.
391 * Note, this function is inherently racy, because there is nothing to prevent
392 * the cgroup inode from getting torn down and potentially reallocated a moment
393 * after page_cgroup_ino() returns, so it only should be used by callers that
394 * do not care (such as procfs interfaces).
396 ino_t
page_cgroup_ino(struct page
*page
)
398 struct mem_cgroup
*memcg
;
399 unsigned long ino
= 0;
402 /* page_folio() is racy here, but the entire function is racy anyway */
403 memcg
= folio_memcg_check(page_folio(page
));
405 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
406 memcg
= parent_mem_cgroup(memcg
);
408 ino
= cgroup_ino(memcg
->css
.cgroup
);
413 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node
*mz
,
414 struct mem_cgroup_tree_per_node
*mctz
,
415 unsigned long new_usage_in_excess
)
417 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
418 struct rb_node
*parent
= NULL
;
419 struct mem_cgroup_per_node
*mz_node
;
420 bool rightmost
= true;
425 mz
->usage_in_excess
= new_usage_in_excess
;
426 if (!mz
->usage_in_excess
)
430 mz_node
= rb_entry(parent
, struct mem_cgroup_per_node
,
432 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
) {
441 mctz
->rb_rightmost
= &mz
->tree_node
;
443 rb_link_node(&mz
->tree_node
, parent
, p
);
444 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
448 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
449 struct mem_cgroup_tree_per_node
*mctz
)
454 if (&mz
->tree_node
== mctz
->rb_rightmost
)
455 mctz
->rb_rightmost
= rb_prev(&mz
->tree_node
);
457 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
461 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node
*mz
,
462 struct mem_cgroup_tree_per_node
*mctz
)
466 spin_lock_irqsave(&mctz
->lock
, flags
);
467 __mem_cgroup_remove_exceeded(mz
, mctz
);
468 spin_unlock_irqrestore(&mctz
->lock
, flags
);
471 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
473 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
474 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
475 unsigned long excess
= 0;
477 if (nr_pages
> soft_limit
)
478 excess
= nr_pages
- soft_limit
;
483 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, int nid
)
485 unsigned long excess
;
486 struct mem_cgroup_per_node
*mz
;
487 struct mem_cgroup_tree_per_node
*mctz
;
489 if (lru_gen_enabled()) {
490 if (soft_limit_excess(memcg
))
491 lru_gen_soft_reclaim(memcg
, nid
);
495 mctz
= soft_limit_tree
.rb_tree_per_node
[nid
];
499 * Necessary to update all ancestors when hierarchy is used.
500 * because their event counter is not touched.
502 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
503 mz
= memcg
->nodeinfo
[nid
];
504 excess
= soft_limit_excess(memcg
);
506 * We have to update the tree if mz is on RB-tree or
507 * mem is over its softlimit.
509 if (excess
|| mz
->on_tree
) {
512 spin_lock_irqsave(&mctz
->lock
, flags
);
513 /* if on-tree, remove it */
515 __mem_cgroup_remove_exceeded(mz
, mctz
);
517 * Insert again. mz->usage_in_excess will be updated.
518 * If excess is 0, no tree ops.
520 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
521 spin_unlock_irqrestore(&mctz
->lock
, flags
);
526 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
528 struct mem_cgroup_tree_per_node
*mctz
;
529 struct mem_cgroup_per_node
*mz
;
533 mz
= memcg
->nodeinfo
[nid
];
534 mctz
= soft_limit_tree
.rb_tree_per_node
[nid
];
536 mem_cgroup_remove_exceeded(mz
, mctz
);
540 static struct mem_cgroup_per_node
*
541 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
543 struct mem_cgroup_per_node
*mz
;
547 if (!mctz
->rb_rightmost
)
548 goto done
; /* Nothing to reclaim from */
550 mz
= rb_entry(mctz
->rb_rightmost
,
551 struct mem_cgroup_per_node
, tree_node
);
553 * Remove the node now but someone else can add it back,
554 * we will to add it back at the end of reclaim to its correct
555 * position in the tree.
557 __mem_cgroup_remove_exceeded(mz
, mctz
);
558 if (!soft_limit_excess(mz
->memcg
) ||
559 !css_tryget(&mz
->memcg
->css
))
565 static struct mem_cgroup_per_node
*
566 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node
*mctz
)
568 struct mem_cgroup_per_node
*mz
;
570 spin_lock_irq(&mctz
->lock
);
571 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
572 spin_unlock_irq(&mctz
->lock
);
577 * memcg and lruvec stats flushing
579 * Many codepaths leading to stats update or read are performance sensitive and
580 * adding stats flushing in such codepaths is not desirable. So, to optimize the
581 * flushing the kernel does:
583 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
584 * rstat update tree grow unbounded.
586 * 2) Flush the stats synchronously on reader side only when there are more than
587 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
588 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
589 * only for 2 seconds due to (1).
591 static void flush_memcg_stats_dwork(struct work_struct
*w
);
592 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork
, flush_memcg_stats_dwork
);
593 static DEFINE_PER_CPU(unsigned int, stats_updates
);
594 static atomic_t stats_flush_ongoing
= ATOMIC_INIT(0);
595 static atomic_t stats_flush_threshold
= ATOMIC_INIT(0);
596 static u64 flush_next_time
;
598 #define FLUSH_TIME (2UL*HZ)
601 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
602 * not rely on this as part of an acquired spinlock_t lock. These functions are
603 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
606 static void memcg_stats_lock(void)
608 preempt_disable_nested();
609 VM_WARN_ON_IRQS_ENABLED();
612 static void __memcg_stats_lock(void)
614 preempt_disable_nested();
617 static void memcg_stats_unlock(void)
619 preempt_enable_nested();
622 static inline void memcg_rstat_updated(struct mem_cgroup
*memcg
, int val
)
629 cgroup_rstat_updated(memcg
->css
.cgroup
, smp_processor_id());
631 x
= __this_cpu_add_return(stats_updates
, abs(val
));
632 if (x
> MEMCG_CHARGE_BATCH
) {
634 * If stats_flush_threshold exceeds the threshold
635 * (>num_online_cpus()), cgroup stats update will be triggered
636 * in __mem_cgroup_flush_stats(). Increasing this var further
637 * is redundant and simply adds overhead in atomic update.
639 if (atomic_read(&stats_flush_threshold
) <= num_online_cpus())
640 atomic_add(x
/ MEMCG_CHARGE_BATCH
, &stats_flush_threshold
);
641 __this_cpu_write(stats_updates
, 0);
645 static void do_flush_stats(void)
648 * We always flush the entire tree, so concurrent flushers can just
649 * skip. This avoids a thundering herd problem on the rstat global lock
650 * from memcg flushers (e.g. reclaim, refault, etc).
652 if (atomic_read(&stats_flush_ongoing
) ||
653 atomic_xchg(&stats_flush_ongoing
, 1))
656 WRITE_ONCE(flush_next_time
, jiffies_64
+ 2*FLUSH_TIME
);
658 cgroup_rstat_flush(root_mem_cgroup
->css
.cgroup
);
660 atomic_set(&stats_flush_threshold
, 0);
661 atomic_set(&stats_flush_ongoing
, 0);
664 void mem_cgroup_flush_stats(void)
666 if (atomic_read(&stats_flush_threshold
) > num_online_cpus())
670 void mem_cgroup_flush_stats_ratelimited(void)
672 if (time_after64(jiffies_64
, READ_ONCE(flush_next_time
)))
673 mem_cgroup_flush_stats();
676 static void flush_memcg_stats_dwork(struct work_struct
*w
)
679 * Always flush here so that flushing in latency-sensitive paths is
680 * as cheap as possible.
683 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
, FLUSH_TIME
);
686 /* Subset of vm_event_item to report for memcg event stats */
687 static const unsigned int memcg_vm_event_stat
[] = {
703 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
707 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
715 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
716 static int mem_cgroup_events_index
[NR_VM_EVENT_ITEMS
] __read_mostly
;
718 static void init_memcg_events(void)
722 for (i
= 0; i
< NR_MEMCG_EVENTS
; ++i
)
723 mem_cgroup_events_index
[memcg_vm_event_stat
[i
]] = i
+ 1;
726 static inline int memcg_events_index(enum vm_event_item idx
)
728 return mem_cgroup_events_index
[idx
] - 1;
731 struct memcg_vmstats_percpu
{
732 /* Local (CPU and cgroup) page state & events */
733 long state
[MEMCG_NR_STAT
];
734 unsigned long events
[NR_MEMCG_EVENTS
];
736 /* Delta calculation for lockless upward propagation */
737 long state_prev
[MEMCG_NR_STAT
];
738 unsigned long events_prev
[NR_MEMCG_EVENTS
];
740 /* Cgroup1: threshold notifications & softlimit tree updates */
741 unsigned long nr_page_events
;
742 unsigned long targets
[MEM_CGROUP_NTARGETS
];
745 struct memcg_vmstats
{
746 /* Aggregated (CPU and subtree) page state & events */
747 long state
[MEMCG_NR_STAT
];
748 unsigned long events
[NR_MEMCG_EVENTS
];
750 /* Non-hierarchical (CPU aggregated) page state & events */
751 long state_local
[MEMCG_NR_STAT
];
752 unsigned long events_local
[NR_MEMCG_EVENTS
];
754 /* Pending child counts during tree propagation */
755 long state_pending
[MEMCG_NR_STAT
];
756 unsigned long events_pending
[NR_MEMCG_EVENTS
];
759 unsigned long memcg_page_state(struct mem_cgroup
*memcg
, int idx
)
761 long x
= READ_ONCE(memcg
->vmstats
->state
[idx
]);
769 static int memcg_page_state_unit(int item
);
772 * Normalize the value passed into memcg_rstat_updated() to be in pages. Round
773 * up non-zero sub-page updates to 1 page as zero page updates are ignored.
775 static int memcg_state_val_in_pages(int idx
, int val
)
777 int unit
= memcg_page_state_unit(idx
);
779 if (!val
|| unit
== PAGE_SIZE
)
782 return max(val
* unit
/ PAGE_SIZE
, 1UL);
786 * __mod_memcg_state - update cgroup memory statistics
787 * @memcg: the memory cgroup
788 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
789 * @val: delta to add to the counter, can be negative
791 void __mod_memcg_state(struct mem_cgroup
*memcg
, int idx
, int val
)
793 if (mem_cgroup_disabled())
796 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
797 memcg_rstat_updated(memcg
, memcg_state_val_in_pages(idx
, val
));
800 /* idx can be of type enum memcg_stat_item or node_stat_item. */
801 static unsigned long memcg_page_state_local(struct mem_cgroup
*memcg
, int idx
)
803 long x
= READ_ONCE(memcg
->vmstats
->state_local
[idx
]);
812 void __mod_memcg_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
815 struct mem_cgroup_per_node
*pn
;
816 struct mem_cgroup
*memcg
;
818 pn
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
822 * The caller from rmap relies on disabled preemption because they never
823 * update their counter from in-interrupt context. For these two
824 * counters we check that the update is never performed from an
825 * interrupt context while other caller need to have disabled interrupt.
827 __memcg_stats_lock();
828 if (IS_ENABLED(CONFIG_DEBUG_VM
)) {
833 case NR_SHMEM_PMDMAPPED
:
834 case NR_FILE_PMDMAPPED
:
835 WARN_ON_ONCE(!in_task());
838 VM_WARN_ON_IRQS_ENABLED();
843 __this_cpu_add(memcg
->vmstats_percpu
->state
[idx
], val
);
846 __this_cpu_add(pn
->lruvec_stats_percpu
->state
[idx
], val
);
848 memcg_rstat_updated(memcg
, memcg_state_val_in_pages(idx
, val
));
849 memcg_stats_unlock();
853 * __mod_lruvec_state - update lruvec memory statistics
854 * @lruvec: the lruvec
855 * @idx: the stat item
856 * @val: delta to add to the counter, can be negative
858 * The lruvec is the intersection of the NUMA node and a cgroup. This
859 * function updates the all three counters that are affected by a
860 * change of state at this level: per-node, per-cgroup, per-lruvec.
862 void __mod_lruvec_state(struct lruvec
*lruvec
, enum node_stat_item idx
,
866 __mod_node_page_state(lruvec_pgdat(lruvec
), idx
, val
);
868 /* Update memcg and lruvec */
869 if (!mem_cgroup_disabled())
870 __mod_memcg_lruvec_state(lruvec
, idx
, val
);
873 void __mod_lruvec_page_state(struct page
*page
, enum node_stat_item idx
,
876 struct page
*head
= compound_head(page
); /* rmap on tail pages */
877 struct mem_cgroup
*memcg
;
878 pg_data_t
*pgdat
= page_pgdat(page
);
879 struct lruvec
*lruvec
;
882 memcg
= page_memcg(head
);
883 /* Untracked pages have no memcg, no lruvec. Update only the node */
886 __mod_node_page_state(pgdat
, idx
, val
);
890 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
891 __mod_lruvec_state(lruvec
, idx
, val
);
894 EXPORT_SYMBOL(__mod_lruvec_page_state
);
896 void __mod_lruvec_kmem_state(void *p
, enum node_stat_item idx
, int val
)
898 pg_data_t
*pgdat
= page_pgdat(virt_to_page(p
));
899 struct mem_cgroup
*memcg
;
900 struct lruvec
*lruvec
;
903 memcg
= mem_cgroup_from_slab_obj(p
);
906 * Untracked pages have no memcg, no lruvec. Update only the
907 * node. If we reparent the slab objects to the root memcg,
908 * when we free the slab object, we need to update the per-memcg
909 * vmstats to keep it correct for the root memcg.
912 __mod_node_page_state(pgdat
, idx
, val
);
914 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
915 __mod_lruvec_state(lruvec
, idx
, val
);
921 * __count_memcg_events - account VM events in a cgroup
922 * @memcg: the memory cgroup
923 * @idx: the event item
924 * @count: the number of events that occurred
926 void __count_memcg_events(struct mem_cgroup
*memcg
, enum vm_event_item idx
,
929 int index
= memcg_events_index(idx
);
931 if (mem_cgroup_disabled() || index
< 0)
935 __this_cpu_add(memcg
->vmstats_percpu
->events
[index
], count
);
936 memcg_rstat_updated(memcg
, count
);
937 memcg_stats_unlock();
940 static unsigned long memcg_events(struct mem_cgroup
*memcg
, int event
)
942 int index
= memcg_events_index(event
);
946 return READ_ONCE(memcg
->vmstats
->events
[index
]);
949 static unsigned long memcg_events_local(struct mem_cgroup
*memcg
, int event
)
951 int index
= memcg_events_index(event
);
956 return READ_ONCE(memcg
->vmstats
->events_local
[index
]);
959 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
962 /* pagein of a big page is an event. So, ignore page size */
964 __count_memcg_events(memcg
, PGPGIN
, 1);
966 __count_memcg_events(memcg
, PGPGOUT
, 1);
967 nr_pages
= -nr_pages
; /* for event */
970 __this_cpu_add(memcg
->vmstats_percpu
->nr_page_events
, nr_pages
);
973 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
974 enum mem_cgroup_events_target target
)
976 unsigned long val
, next
;
978 val
= __this_cpu_read(memcg
->vmstats_percpu
->nr_page_events
);
979 next
= __this_cpu_read(memcg
->vmstats_percpu
->targets
[target
]);
980 /* from time_after() in jiffies.h */
981 if ((long)(next
- val
) < 0) {
983 case MEM_CGROUP_TARGET_THRESH
:
984 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
986 case MEM_CGROUP_TARGET_SOFTLIMIT
:
987 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
992 __this_cpu_write(memcg
->vmstats_percpu
->targets
[target
], next
);
999 * Check events in order.
1002 static void memcg_check_events(struct mem_cgroup
*memcg
, int nid
)
1004 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
1007 /* threshold event is triggered in finer grain than soft limit */
1008 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1009 MEM_CGROUP_TARGET_THRESH
))) {
1012 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1013 MEM_CGROUP_TARGET_SOFTLIMIT
);
1014 mem_cgroup_threshold(memcg
);
1015 if (unlikely(do_softlimit
))
1016 mem_cgroup_update_tree(memcg
, nid
);
1020 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1023 * mm_update_next_owner() may clear mm->owner to NULL
1024 * if it races with swapoff, page migration, etc.
1025 * So this can be called with p == NULL.
1030 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1032 EXPORT_SYMBOL(mem_cgroup_from_task
);
1034 static __always_inline
struct mem_cgroup
*active_memcg(void)
1037 return this_cpu_read(int_active_memcg
);
1039 return current
->active_memcg
;
1043 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1044 * @mm: mm from which memcg should be extracted. It can be NULL.
1046 * Obtain a reference on mm->memcg and returns it if successful. If mm
1047 * is NULL, then the memcg is chosen as follows:
1048 * 1) The active memcg, if set.
1049 * 2) current->mm->memcg, if available
1051 * If mem_cgroup is disabled, NULL is returned.
1053 struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1055 struct mem_cgroup
*memcg
;
1057 if (mem_cgroup_disabled())
1061 * Page cache insertions can happen without an
1062 * actual mm context, e.g. during disk probing
1063 * on boot, loopback IO, acct() writes etc.
1065 * No need to css_get on root memcg as the reference
1066 * counting is disabled on the root level in the
1067 * cgroup core. See CSS_NO_REF.
1069 if (unlikely(!mm
)) {
1070 memcg
= active_memcg();
1071 if (unlikely(memcg
)) {
1072 /* remote memcg must hold a ref */
1073 css_get(&memcg
->css
);
1078 return root_mem_cgroup
;
1083 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1084 if (unlikely(!memcg
))
1085 memcg
= root_mem_cgroup
;
1086 } while (!css_tryget(&memcg
->css
));
1090 EXPORT_SYMBOL(get_mem_cgroup_from_mm
);
1093 * get_mem_cgroup_from_current - Obtain a reference on current task's memcg.
1095 struct mem_cgroup
*get_mem_cgroup_from_current(void)
1097 struct mem_cgroup
*memcg
;
1099 if (mem_cgroup_disabled())
1104 memcg
= mem_cgroup_from_task(current
);
1105 if (!css_tryget(&memcg
->css
)) {
1114 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1115 * @root: hierarchy root
1116 * @prev: previously returned memcg, NULL on first invocation
1117 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1119 * Returns references to children of the hierarchy below @root, or
1120 * @root itself, or %NULL after a full round-trip.
1122 * Caller must pass the return value in @prev on subsequent
1123 * invocations for reference counting, or use mem_cgroup_iter_break()
1124 * to cancel a hierarchy walk before the round-trip is complete.
1126 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1127 * in the hierarchy among all concurrent reclaimers operating on the
1130 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1131 struct mem_cgroup
*prev
,
1132 struct mem_cgroup_reclaim_cookie
*reclaim
)
1134 struct mem_cgroup_reclaim_iter
*iter
;
1135 struct cgroup_subsys_state
*css
= NULL
;
1136 struct mem_cgroup
*memcg
= NULL
;
1137 struct mem_cgroup
*pos
= NULL
;
1139 if (mem_cgroup_disabled())
1143 root
= root_mem_cgroup
;
1148 struct mem_cgroup_per_node
*mz
;
1150 mz
= root
->nodeinfo
[reclaim
->pgdat
->node_id
];
1154 * On start, join the current reclaim iteration cycle.
1155 * Exit when a concurrent walker completes it.
1158 reclaim
->generation
= iter
->generation
;
1159 else if (reclaim
->generation
!= iter
->generation
)
1163 pos
= READ_ONCE(iter
->position
);
1164 if (!pos
|| css_tryget(&pos
->css
))
1167 * css reference reached zero, so iter->position will
1168 * be cleared by ->css_released. However, we should not
1169 * rely on this happening soon, because ->css_released
1170 * is called from a work queue, and by busy-waiting we
1171 * might block it. So we clear iter->position right
1174 (void)cmpxchg(&iter
->position
, pos
, NULL
);
1184 css
= css_next_descendant_pre(css
, &root
->css
);
1187 * Reclaimers share the hierarchy walk, and a
1188 * new one might jump in right at the end of
1189 * the hierarchy - make sure they see at least
1190 * one group and restart from the beginning.
1198 * Verify the css and acquire a reference. The root
1199 * is provided by the caller, so we know it's alive
1200 * and kicking, and don't take an extra reference.
1202 if (css
== &root
->css
|| css_tryget(css
)) {
1203 memcg
= mem_cgroup_from_css(css
);
1210 * The position could have already been updated by a competing
1211 * thread, so check that the value hasn't changed since we read
1212 * it to avoid reclaiming from the same cgroup twice.
1214 (void)cmpxchg(&iter
->position
, pos
, memcg
);
1225 if (prev
&& prev
!= root
)
1226 css_put(&prev
->css
);
1232 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1233 * @root: hierarchy root
1234 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1236 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1237 struct mem_cgroup
*prev
)
1240 root
= root_mem_cgroup
;
1241 if (prev
&& prev
!= root
)
1242 css_put(&prev
->css
);
1245 static void __invalidate_reclaim_iterators(struct mem_cgroup
*from
,
1246 struct mem_cgroup
*dead_memcg
)
1248 struct mem_cgroup_reclaim_iter
*iter
;
1249 struct mem_cgroup_per_node
*mz
;
1252 for_each_node(nid
) {
1253 mz
= from
->nodeinfo
[nid
];
1255 cmpxchg(&iter
->position
, dead_memcg
, NULL
);
1259 static void invalidate_reclaim_iterators(struct mem_cgroup
*dead_memcg
)
1261 struct mem_cgroup
*memcg
= dead_memcg
;
1262 struct mem_cgroup
*last
;
1265 __invalidate_reclaim_iterators(memcg
, dead_memcg
);
1267 } while ((memcg
= parent_mem_cgroup(memcg
)));
1270 * When cgroup1 non-hierarchy mode is used,
1271 * parent_mem_cgroup() does not walk all the way up to the
1272 * cgroup root (root_mem_cgroup). So we have to handle
1273 * dead_memcg from cgroup root separately.
1275 if (!mem_cgroup_is_root(last
))
1276 __invalidate_reclaim_iterators(root_mem_cgroup
,
1281 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1282 * @memcg: hierarchy root
1283 * @fn: function to call for each task
1284 * @arg: argument passed to @fn
1286 * This function iterates over tasks attached to @memcg or to any of its
1287 * descendants and calls @fn for each task. If @fn returns a non-zero
1288 * value, the function breaks the iteration loop. Otherwise, it will iterate
1289 * over all tasks and return 0.
1291 * This function must not be called for the root memory cgroup.
1293 void mem_cgroup_scan_tasks(struct mem_cgroup
*memcg
,
1294 int (*fn
)(struct task_struct
*, void *), void *arg
)
1296 struct mem_cgroup
*iter
;
1299 BUG_ON(mem_cgroup_is_root(memcg
));
1301 for_each_mem_cgroup_tree(iter
, memcg
) {
1302 struct css_task_iter it
;
1303 struct task_struct
*task
;
1305 css_task_iter_start(&iter
->css
, CSS_TASK_ITER_PROCS
, &it
);
1306 while (!ret
&& (task
= css_task_iter_next(&it
)))
1307 ret
= fn(task
, arg
);
1308 css_task_iter_end(&it
);
1310 mem_cgroup_iter_break(memcg
, iter
);
1316 #ifdef CONFIG_DEBUG_VM
1317 void lruvec_memcg_debug(struct lruvec
*lruvec
, struct folio
*folio
)
1319 struct mem_cgroup
*memcg
;
1321 if (mem_cgroup_disabled())
1324 memcg
= folio_memcg(folio
);
1327 VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec
)), folio
);
1329 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec
) != memcg
, folio
);
1334 * folio_lruvec_lock - Lock the lruvec for a folio.
1335 * @folio: Pointer to the folio.
1337 * These functions are safe to use under any of the following conditions:
1339 * - folio_test_lru false
1340 * - folio_memcg_lock()
1341 * - folio frozen (refcount of 0)
1343 * Return: The lruvec this folio is on with its lock held.
1345 struct lruvec
*folio_lruvec_lock(struct folio
*folio
)
1347 struct lruvec
*lruvec
= folio_lruvec(folio
);
1349 spin_lock(&lruvec
->lru_lock
);
1350 lruvec_memcg_debug(lruvec
, folio
);
1356 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1357 * @folio: Pointer to the folio.
1359 * These functions are safe to use under any of the following conditions:
1361 * - folio_test_lru false
1362 * - folio_memcg_lock()
1363 * - folio frozen (refcount of 0)
1365 * Return: The lruvec this folio is on with its lock held and interrupts
1368 struct lruvec
*folio_lruvec_lock_irq(struct folio
*folio
)
1370 struct lruvec
*lruvec
= folio_lruvec(folio
);
1372 spin_lock_irq(&lruvec
->lru_lock
);
1373 lruvec_memcg_debug(lruvec
, folio
);
1379 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1380 * @folio: Pointer to the folio.
1381 * @flags: Pointer to irqsave flags.
1383 * These functions are safe to use under any of the following conditions:
1385 * - folio_test_lru false
1386 * - folio_memcg_lock()
1387 * - folio frozen (refcount of 0)
1389 * Return: The lruvec this folio is on with its lock held and interrupts
1392 struct lruvec
*folio_lruvec_lock_irqsave(struct folio
*folio
,
1393 unsigned long *flags
)
1395 struct lruvec
*lruvec
= folio_lruvec(folio
);
1397 spin_lock_irqsave(&lruvec
->lru_lock
, *flags
);
1398 lruvec_memcg_debug(lruvec
, folio
);
1404 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1405 * @lruvec: mem_cgroup per zone lru vector
1406 * @lru: index of lru list the page is sitting on
1407 * @zid: zone id of the accounted pages
1408 * @nr_pages: positive when adding or negative when removing
1410 * This function must be called under lru_lock, just before a page is added
1411 * to or just after a page is removed from an lru list.
1413 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1414 int zid
, int nr_pages
)
1416 struct mem_cgroup_per_node
*mz
;
1417 unsigned long *lru_size
;
1420 if (mem_cgroup_disabled())
1423 mz
= container_of(lruvec
, struct mem_cgroup_per_node
, lruvec
);
1424 lru_size
= &mz
->lru_zone_size
[zid
][lru
];
1427 *lru_size
+= nr_pages
;
1430 if (WARN_ONCE(size
< 0,
1431 "%s(%p, %d, %d): lru_size %ld\n",
1432 __func__
, lruvec
, lru
, nr_pages
, size
)) {
1438 *lru_size
+= nr_pages
;
1442 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1443 * @memcg: the memory cgroup
1445 * Returns the maximum amount of memory @mem can be charged with, in
1448 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1450 unsigned long margin
= 0;
1451 unsigned long count
;
1452 unsigned long limit
;
1454 count
= page_counter_read(&memcg
->memory
);
1455 limit
= READ_ONCE(memcg
->memory
.max
);
1457 margin
= limit
- count
;
1459 if (do_memsw_account()) {
1460 count
= page_counter_read(&memcg
->memsw
);
1461 limit
= READ_ONCE(memcg
->memsw
.max
);
1463 margin
= min(margin
, limit
- count
);
1472 * A routine for checking "mem" is under move_account() or not.
1474 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1475 * moving cgroups. This is for waiting at high-memory pressure
1478 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1480 struct mem_cgroup
*from
;
1481 struct mem_cgroup
*to
;
1484 * Unlike task_move routines, we access mc.to, mc.from not under
1485 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1487 spin_lock(&mc
.lock
);
1493 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1494 mem_cgroup_is_descendant(to
, memcg
);
1496 spin_unlock(&mc
.lock
);
1500 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1502 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1503 if (mem_cgroup_under_move(memcg
)) {
1505 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1506 /* moving charge context might have finished. */
1509 finish_wait(&mc
.waitq
, &wait
);
1516 struct memory_stat
{
1521 static const struct memory_stat memory_stats
[] = {
1522 { "anon", NR_ANON_MAPPED
},
1523 { "file", NR_FILE_PAGES
},
1524 { "kernel", MEMCG_KMEM
},
1525 { "kernel_stack", NR_KERNEL_STACK_KB
},
1526 { "pagetables", NR_PAGETABLE
},
1527 { "sec_pagetables", NR_SECONDARY_PAGETABLE
},
1528 { "percpu", MEMCG_PERCPU_B
},
1529 { "sock", MEMCG_SOCK
},
1530 { "vmalloc", MEMCG_VMALLOC
},
1531 { "shmem", NR_SHMEM
},
1532 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1533 { "zswap", MEMCG_ZSWAP_B
},
1534 { "zswapped", MEMCG_ZSWAPPED
},
1536 { "file_mapped", NR_FILE_MAPPED
},
1537 { "file_dirty", NR_FILE_DIRTY
},
1538 { "file_writeback", NR_WRITEBACK
},
1540 { "swapcached", NR_SWAPCACHE
},
1542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1543 { "anon_thp", NR_ANON_THPS
},
1544 { "file_thp", NR_FILE_THPS
},
1545 { "shmem_thp", NR_SHMEM_THPS
},
1547 { "inactive_anon", NR_INACTIVE_ANON
},
1548 { "active_anon", NR_ACTIVE_ANON
},
1549 { "inactive_file", NR_INACTIVE_FILE
},
1550 { "active_file", NR_ACTIVE_FILE
},
1551 { "unevictable", NR_UNEVICTABLE
},
1552 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B
},
1553 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B
},
1555 /* The memory events */
1556 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON
},
1557 { "workingset_refault_file", WORKINGSET_REFAULT_FILE
},
1558 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON
},
1559 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE
},
1560 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON
},
1561 { "workingset_restore_file", WORKINGSET_RESTORE_FILE
},
1562 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM
},
1565 /* The actual unit of the state item, not the same as the output unit */
1566 static int memcg_page_state_unit(int item
)
1569 case MEMCG_PERCPU_B
:
1571 case NR_SLAB_RECLAIMABLE_B
:
1572 case NR_SLAB_UNRECLAIMABLE_B
:
1574 case NR_KERNEL_STACK_KB
:
1581 /* Translate stat items to the correct unit for memory.stat output */
1582 static int memcg_page_state_output_unit(int item
)
1585 * Workingset state is actually in pages, but we export it to userspace
1586 * as a scalar count of events, so special case it here.
1589 case WORKINGSET_REFAULT_ANON
:
1590 case WORKINGSET_REFAULT_FILE
:
1591 case WORKINGSET_ACTIVATE_ANON
:
1592 case WORKINGSET_ACTIVATE_FILE
:
1593 case WORKINGSET_RESTORE_ANON
:
1594 case WORKINGSET_RESTORE_FILE
:
1595 case WORKINGSET_NODERECLAIM
:
1598 return memcg_page_state_unit(item
);
1602 static inline unsigned long memcg_page_state_output(struct mem_cgroup
*memcg
,
1605 return memcg_page_state(memcg
, item
) *
1606 memcg_page_state_output_unit(item
);
1609 static inline unsigned long memcg_page_state_local_output(
1610 struct mem_cgroup
*memcg
, int item
)
1612 return memcg_page_state_local(memcg
, item
) *
1613 memcg_page_state_output_unit(item
);
1616 static void memcg_stat_format(struct mem_cgroup
*memcg
, struct seq_buf
*s
)
1621 * Provide statistics on the state of the memory subsystem as
1622 * well as cumulative event counters that show past behavior.
1624 * This list is ordered following a combination of these gradients:
1625 * 1) generic big picture -> specifics and details
1626 * 2) reflecting userspace activity -> reflecting kernel heuristics
1628 * Current memory state:
1630 mem_cgroup_flush_stats();
1632 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
1635 size
= memcg_page_state_output(memcg
, memory_stats
[i
].idx
);
1636 seq_buf_printf(s
, "%s %llu\n", memory_stats
[i
].name
, size
);
1638 if (unlikely(memory_stats
[i
].idx
== NR_SLAB_UNRECLAIMABLE_B
)) {
1639 size
+= memcg_page_state_output(memcg
,
1640 NR_SLAB_RECLAIMABLE_B
);
1641 seq_buf_printf(s
, "slab %llu\n", size
);
1645 /* Accumulated memory events */
1646 seq_buf_printf(s
, "pgscan %lu\n",
1647 memcg_events(memcg
, PGSCAN_KSWAPD
) +
1648 memcg_events(memcg
, PGSCAN_DIRECT
) +
1649 memcg_events(memcg
, PGSCAN_KHUGEPAGED
));
1650 seq_buf_printf(s
, "pgsteal %lu\n",
1651 memcg_events(memcg
, PGSTEAL_KSWAPD
) +
1652 memcg_events(memcg
, PGSTEAL_DIRECT
) +
1653 memcg_events(memcg
, PGSTEAL_KHUGEPAGED
));
1655 for (i
= 0; i
< ARRAY_SIZE(memcg_vm_event_stat
); i
++) {
1656 if (memcg_vm_event_stat
[i
] == PGPGIN
||
1657 memcg_vm_event_stat
[i
] == PGPGOUT
)
1660 seq_buf_printf(s
, "%s %lu\n",
1661 vm_event_name(memcg_vm_event_stat
[i
]),
1662 memcg_events(memcg
, memcg_vm_event_stat
[i
]));
1665 /* The above should easily fit into one page */
1666 WARN_ON_ONCE(seq_buf_has_overflowed(s
));
1669 static void memcg1_stat_format(struct mem_cgroup
*memcg
, struct seq_buf
*s
);
1671 static void memory_stat_format(struct mem_cgroup
*memcg
, struct seq_buf
*s
)
1673 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1674 memcg_stat_format(memcg
, s
);
1676 memcg1_stat_format(memcg
, s
);
1677 WARN_ON_ONCE(seq_buf_has_overflowed(s
));
1681 * mem_cgroup_print_oom_context: Print OOM information relevant to
1682 * memory controller.
1683 * @memcg: The memory cgroup that went over limit
1684 * @p: Task that is going to be killed
1686 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1689 void mem_cgroup_print_oom_context(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1694 pr_cont(",oom_memcg=");
1695 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1697 pr_cont(",global_oom");
1699 pr_cont(",task_memcg=");
1700 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1706 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1707 * memory controller.
1708 * @memcg: The memory cgroup that went over limit
1710 void mem_cgroup_print_oom_meminfo(struct mem_cgroup
*memcg
)
1712 /* Use static buffer, for the caller is holding oom_lock. */
1713 static char buf
[PAGE_SIZE
];
1716 lockdep_assert_held(&oom_lock
);
1718 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1719 K((u64
)page_counter_read(&memcg
->memory
)),
1720 K((u64
)READ_ONCE(memcg
->memory
.max
)), memcg
->memory
.failcnt
);
1721 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
1722 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1723 K((u64
)page_counter_read(&memcg
->swap
)),
1724 K((u64
)READ_ONCE(memcg
->swap
.max
)), memcg
->swap
.failcnt
);
1726 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1727 K((u64
)page_counter_read(&memcg
->memsw
)),
1728 K((u64
)memcg
->memsw
.max
), memcg
->memsw
.failcnt
);
1729 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1730 K((u64
)page_counter_read(&memcg
->kmem
)),
1731 K((u64
)memcg
->kmem
.max
), memcg
->kmem
.failcnt
);
1734 pr_info("Memory cgroup stats for ");
1735 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1737 seq_buf_init(&s
, buf
, sizeof(buf
));
1738 memory_stat_format(memcg
, &s
);
1739 seq_buf_do_printk(&s
, KERN_INFO
);
1743 * Return the memory (and swap, if configured) limit for a memcg.
1745 unsigned long mem_cgroup_get_max(struct mem_cgroup
*memcg
)
1747 unsigned long max
= READ_ONCE(memcg
->memory
.max
);
1749 if (do_memsw_account()) {
1750 if (mem_cgroup_swappiness(memcg
)) {
1751 /* Calculate swap excess capacity from memsw limit */
1752 unsigned long swap
= READ_ONCE(memcg
->memsw
.max
) - max
;
1754 max
+= min(swap
, (unsigned long)total_swap_pages
);
1757 if (mem_cgroup_swappiness(memcg
))
1758 max
+= min(READ_ONCE(memcg
->swap
.max
),
1759 (unsigned long)total_swap_pages
);
1764 unsigned long mem_cgroup_size(struct mem_cgroup
*memcg
)
1766 return page_counter_read(&memcg
->memory
);
1769 static bool mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1772 struct oom_control oc
= {
1776 .gfp_mask
= gfp_mask
,
1781 if (mutex_lock_killable(&oom_lock
))
1784 if (mem_cgroup_margin(memcg
) >= (1 << order
))
1788 * A few threads which were not waiting at mutex_lock_killable() can
1789 * fail to bail out. Therefore, check again after holding oom_lock.
1791 ret
= task_is_dying() || out_of_memory(&oc
);
1794 mutex_unlock(&oom_lock
);
1798 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1801 unsigned long *total_scanned
)
1803 struct mem_cgroup
*victim
= NULL
;
1806 unsigned long excess
;
1807 unsigned long nr_scanned
;
1808 struct mem_cgroup_reclaim_cookie reclaim
= {
1812 excess
= soft_limit_excess(root_memcg
);
1815 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1820 * If we have not been able to reclaim
1821 * anything, it might because there are
1822 * no reclaimable pages under this hierarchy
1827 * We want to do more targeted reclaim.
1828 * excess >> 2 is not to excessive so as to
1829 * reclaim too much, nor too less that we keep
1830 * coming back to reclaim from this cgroup
1832 if (total
>= (excess
>> 2) ||
1833 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1838 total
+= mem_cgroup_shrink_node(victim
, gfp_mask
, false,
1839 pgdat
, &nr_scanned
);
1840 *total_scanned
+= nr_scanned
;
1841 if (!soft_limit_excess(root_memcg
))
1844 mem_cgroup_iter_break(root_memcg
, victim
);
1848 #ifdef CONFIG_LOCKDEP
1849 static struct lockdep_map memcg_oom_lock_dep_map
= {
1850 .name
= "memcg_oom_lock",
1854 static DEFINE_SPINLOCK(memcg_oom_lock
);
1857 * Check OOM-Killer is already running under our hierarchy.
1858 * If someone is running, return false.
1860 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1862 struct mem_cgroup
*iter
, *failed
= NULL
;
1864 spin_lock(&memcg_oom_lock
);
1866 for_each_mem_cgroup_tree(iter
, memcg
) {
1867 if (iter
->oom_lock
) {
1869 * this subtree of our hierarchy is already locked
1870 * so we cannot give a lock.
1873 mem_cgroup_iter_break(memcg
, iter
);
1876 iter
->oom_lock
= true;
1881 * OK, we failed to lock the whole subtree so we have
1882 * to clean up what we set up to the failing subtree
1884 for_each_mem_cgroup_tree(iter
, memcg
) {
1885 if (iter
== failed
) {
1886 mem_cgroup_iter_break(memcg
, iter
);
1889 iter
->oom_lock
= false;
1892 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1894 spin_unlock(&memcg_oom_lock
);
1899 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1901 struct mem_cgroup
*iter
;
1903 spin_lock(&memcg_oom_lock
);
1904 mutex_release(&memcg_oom_lock_dep_map
, _RET_IP_
);
1905 for_each_mem_cgroup_tree(iter
, memcg
)
1906 iter
->oom_lock
= false;
1907 spin_unlock(&memcg_oom_lock
);
1910 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1912 struct mem_cgroup
*iter
;
1914 spin_lock(&memcg_oom_lock
);
1915 for_each_mem_cgroup_tree(iter
, memcg
)
1917 spin_unlock(&memcg_oom_lock
);
1920 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1922 struct mem_cgroup
*iter
;
1925 * Be careful about under_oom underflows because a child memcg
1926 * could have been added after mem_cgroup_mark_under_oom.
1928 spin_lock(&memcg_oom_lock
);
1929 for_each_mem_cgroup_tree(iter
, memcg
)
1930 if (iter
->under_oom
> 0)
1932 spin_unlock(&memcg_oom_lock
);
1935 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1937 struct oom_wait_info
{
1938 struct mem_cgroup
*memcg
;
1939 wait_queue_entry_t wait
;
1942 static int memcg_oom_wake_function(wait_queue_entry_t
*wait
,
1943 unsigned mode
, int sync
, void *arg
)
1945 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1946 struct mem_cgroup
*oom_wait_memcg
;
1947 struct oom_wait_info
*oom_wait_info
;
1949 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1950 oom_wait_memcg
= oom_wait_info
->memcg
;
1952 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1953 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1955 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1958 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1961 * For the following lockless ->under_oom test, the only required
1962 * guarantee is that it must see the state asserted by an OOM when
1963 * this function is called as a result of userland actions
1964 * triggered by the notification of the OOM. This is trivially
1965 * achieved by invoking mem_cgroup_mark_under_oom() before
1966 * triggering notification.
1968 if (memcg
&& memcg
->under_oom
)
1969 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1973 * Returns true if successfully killed one or more processes. Though in some
1974 * corner cases it can return true even without killing any process.
1976 static bool mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1980 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1983 memcg_memory_event(memcg
, MEMCG_OOM
);
1986 * We are in the middle of the charge context here, so we
1987 * don't want to block when potentially sitting on a callstack
1988 * that holds all kinds of filesystem and mm locks.
1990 * cgroup1 allows disabling the OOM killer and waiting for outside
1991 * handling until the charge can succeed; remember the context and put
1992 * the task to sleep at the end of the page fault when all locks are
1995 * On the other hand, in-kernel OOM killer allows for an async victim
1996 * memory reclaim (oom_reaper) and that means that we are not solely
1997 * relying on the oom victim to make a forward progress and we can
1998 * invoke the oom killer here.
2000 * Please note that mem_cgroup_out_of_memory might fail to find a
2001 * victim and then we have to bail out from the charge path.
2003 if (READ_ONCE(memcg
->oom_kill_disable
)) {
2004 if (current
->in_user_fault
) {
2005 css_get(&memcg
->css
);
2006 current
->memcg_in_oom
= memcg
;
2007 current
->memcg_oom_gfp_mask
= mask
;
2008 current
->memcg_oom_order
= order
;
2013 mem_cgroup_mark_under_oom(memcg
);
2015 locked
= mem_cgroup_oom_trylock(memcg
);
2018 mem_cgroup_oom_notify(memcg
);
2020 mem_cgroup_unmark_under_oom(memcg
);
2021 ret
= mem_cgroup_out_of_memory(memcg
, mask
, order
);
2024 mem_cgroup_oom_unlock(memcg
);
2030 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2031 * @handle: actually kill/wait or just clean up the OOM state
2033 * This has to be called at the end of a page fault if the memcg OOM
2034 * handler was enabled.
2036 * Memcg supports userspace OOM handling where failed allocations must
2037 * sleep on a waitqueue until the userspace task resolves the
2038 * situation. Sleeping directly in the charge context with all kinds
2039 * of locks held is not a good idea, instead we remember an OOM state
2040 * in the task and mem_cgroup_oom_synchronize() has to be called at
2041 * the end of the page fault to complete the OOM handling.
2043 * Returns %true if an ongoing memcg OOM situation was detected and
2044 * completed, %false otherwise.
2046 bool mem_cgroup_oom_synchronize(bool handle
)
2048 struct mem_cgroup
*memcg
= current
->memcg_in_oom
;
2049 struct oom_wait_info owait
;
2052 /* OOM is global, do not handle */
2059 owait
.memcg
= memcg
;
2060 owait
.wait
.flags
= 0;
2061 owait
.wait
.func
= memcg_oom_wake_function
;
2062 owait
.wait
.private = current
;
2063 INIT_LIST_HEAD(&owait
.wait
.entry
);
2065 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2066 mem_cgroup_mark_under_oom(memcg
);
2068 locked
= mem_cgroup_oom_trylock(memcg
);
2071 mem_cgroup_oom_notify(memcg
);
2074 mem_cgroup_unmark_under_oom(memcg
);
2075 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2078 mem_cgroup_oom_unlock(memcg
);
2080 current
->memcg_in_oom
= NULL
;
2081 css_put(&memcg
->css
);
2086 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2087 * @victim: task to be killed by the OOM killer
2088 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2090 * Returns a pointer to a memory cgroup, which has to be cleaned up
2091 * by killing all belonging OOM-killable tasks.
2093 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2095 struct mem_cgroup
*mem_cgroup_get_oom_group(struct task_struct
*victim
,
2096 struct mem_cgroup
*oom_domain
)
2098 struct mem_cgroup
*oom_group
= NULL
;
2099 struct mem_cgroup
*memcg
;
2101 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
2105 oom_domain
= root_mem_cgroup
;
2109 memcg
= mem_cgroup_from_task(victim
);
2110 if (mem_cgroup_is_root(memcg
))
2114 * If the victim task has been asynchronously moved to a different
2115 * memory cgroup, we might end up killing tasks outside oom_domain.
2116 * In this case it's better to ignore memory.group.oom.
2118 if (unlikely(!mem_cgroup_is_descendant(memcg
, oom_domain
)))
2122 * Traverse the memory cgroup hierarchy from the victim task's
2123 * cgroup up to the OOMing cgroup (or root) to find the
2124 * highest-level memory cgroup with oom.group set.
2126 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
2127 if (READ_ONCE(memcg
->oom_group
))
2130 if (memcg
== oom_domain
)
2135 css_get(&oom_group
->css
);
2142 void mem_cgroup_print_oom_group(struct mem_cgroup
*memcg
)
2144 pr_info("Tasks in ");
2145 pr_cont_cgroup_path(memcg
->css
.cgroup
);
2146 pr_cont(" are going to be killed due to memory.oom.group set\n");
2150 * folio_memcg_lock - Bind a folio to its memcg.
2151 * @folio: The folio.
2153 * This function prevents unlocked LRU folios from being moved to
2156 * It ensures lifetime of the bound memcg. The caller is responsible
2157 * for the lifetime of the folio.
2159 void folio_memcg_lock(struct folio
*folio
)
2161 struct mem_cgroup
*memcg
;
2162 unsigned long flags
;
2165 * The RCU lock is held throughout the transaction. The fast
2166 * path can get away without acquiring the memcg->move_lock
2167 * because page moving starts with an RCU grace period.
2171 if (mem_cgroup_disabled())
2174 memcg
= folio_memcg(folio
);
2175 if (unlikely(!memcg
))
2178 #ifdef CONFIG_PROVE_LOCKING
2179 local_irq_save(flags
);
2180 might_lock(&memcg
->move_lock
);
2181 local_irq_restore(flags
);
2184 if (atomic_read(&memcg
->moving_account
) <= 0)
2187 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2188 if (memcg
!= folio_memcg(folio
)) {
2189 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2194 * When charge migration first begins, we can have multiple
2195 * critical sections holding the fast-path RCU lock and one
2196 * holding the slowpath move_lock. Track the task who has the
2197 * move_lock for folio_memcg_unlock().
2199 memcg
->move_lock_task
= current
;
2200 memcg
->move_lock_flags
= flags
;
2203 static void __folio_memcg_unlock(struct mem_cgroup
*memcg
)
2205 if (memcg
&& memcg
->move_lock_task
== current
) {
2206 unsigned long flags
= memcg
->move_lock_flags
;
2208 memcg
->move_lock_task
= NULL
;
2209 memcg
->move_lock_flags
= 0;
2211 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2218 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2219 * @folio: The folio.
2221 * This releases the binding created by folio_memcg_lock(). This does
2222 * not change the accounting of this folio to its memcg, but it does
2223 * permit others to change it.
2225 void folio_memcg_unlock(struct folio
*folio
)
2227 __folio_memcg_unlock(folio_memcg(folio
));
2230 struct memcg_stock_pcp
{
2231 local_lock_t stock_lock
;
2232 struct mem_cgroup
*cached
; /* this never be root cgroup */
2233 unsigned int nr_pages
;
2235 #ifdef CONFIG_MEMCG_KMEM
2236 struct obj_cgroup
*cached_objcg
;
2237 struct pglist_data
*cached_pgdat
;
2238 unsigned int nr_bytes
;
2239 int nr_slab_reclaimable_b
;
2240 int nr_slab_unreclaimable_b
;
2243 struct work_struct work
;
2244 unsigned long flags
;
2245 #define FLUSHING_CACHED_CHARGE 0
2247 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
) = {
2248 .stock_lock
= INIT_LOCAL_LOCK(stock_lock
),
2250 static DEFINE_MUTEX(percpu_charge_mutex
);
2252 #ifdef CONFIG_MEMCG_KMEM
2253 static struct obj_cgroup
*drain_obj_stock(struct memcg_stock_pcp
*stock
);
2254 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2255 struct mem_cgroup
*root_memcg
);
2256 static void memcg_account_kmem(struct mem_cgroup
*memcg
, int nr_pages
);
2259 static inline struct obj_cgroup
*drain_obj_stock(struct memcg_stock_pcp
*stock
)
2263 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
2264 struct mem_cgroup
*root_memcg
)
2268 static void memcg_account_kmem(struct mem_cgroup
*memcg
, int nr_pages
)
2274 * consume_stock: Try to consume stocked charge on this cpu.
2275 * @memcg: memcg to consume from.
2276 * @nr_pages: how many pages to charge.
2278 * The charges will only happen if @memcg matches the current cpu's memcg
2279 * stock, and at least @nr_pages are available in that stock. Failure to
2280 * service an allocation will refill the stock.
2282 * returns true if successful, false otherwise.
2284 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2286 struct memcg_stock_pcp
*stock
;
2287 unsigned long flags
;
2290 if (nr_pages
> MEMCG_CHARGE_BATCH
)
2293 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
2295 stock
= this_cpu_ptr(&memcg_stock
);
2296 if (memcg
== READ_ONCE(stock
->cached
) && stock
->nr_pages
>= nr_pages
) {
2297 stock
->nr_pages
-= nr_pages
;
2301 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
2307 * Returns stocks cached in percpu and reset cached information.
2309 static void drain_stock(struct memcg_stock_pcp
*stock
)
2311 struct mem_cgroup
*old
= READ_ONCE(stock
->cached
);
2316 if (stock
->nr_pages
) {
2317 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2318 if (do_memsw_account())
2319 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2320 stock
->nr_pages
= 0;
2324 WRITE_ONCE(stock
->cached
, NULL
);
2327 static void drain_local_stock(struct work_struct
*dummy
)
2329 struct memcg_stock_pcp
*stock
;
2330 struct obj_cgroup
*old
= NULL
;
2331 unsigned long flags
;
2334 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2335 * drain_stock races is that we always operate on local CPU stock
2336 * here with IRQ disabled
2338 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
2340 stock
= this_cpu_ptr(&memcg_stock
);
2341 old
= drain_obj_stock(stock
);
2343 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2345 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
2347 obj_cgroup_put(old
);
2351 * Cache charges(val) to local per_cpu area.
2352 * This will be consumed by consume_stock() function, later.
2354 static void __refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2356 struct memcg_stock_pcp
*stock
;
2358 stock
= this_cpu_ptr(&memcg_stock
);
2359 if (READ_ONCE(stock
->cached
) != memcg
) { /* reset if necessary */
2361 css_get(&memcg
->css
);
2362 WRITE_ONCE(stock
->cached
, memcg
);
2364 stock
->nr_pages
+= nr_pages
;
2366 if (stock
->nr_pages
> MEMCG_CHARGE_BATCH
)
2370 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2372 unsigned long flags
;
2374 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
2375 __refill_stock(memcg
, nr_pages
);
2376 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
2380 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2381 * of the hierarchy under it.
2383 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2387 /* If someone's already draining, avoid adding running more workers. */
2388 if (!mutex_trylock(&percpu_charge_mutex
))
2391 * Notify other cpus that system-wide "drain" is running
2392 * We do not care about races with the cpu hotplug because cpu down
2393 * as well as workers from this path always operate on the local
2394 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2397 curcpu
= smp_processor_id();
2398 for_each_online_cpu(cpu
) {
2399 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2400 struct mem_cgroup
*memcg
;
2404 memcg
= READ_ONCE(stock
->cached
);
2405 if (memcg
&& stock
->nr_pages
&&
2406 mem_cgroup_is_descendant(memcg
, root_memcg
))
2408 else if (obj_stock_flush_required(stock
, root_memcg
))
2413 !test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2415 drain_local_stock(&stock
->work
);
2416 else if (!cpu_is_isolated(cpu
))
2417 schedule_work_on(cpu
, &stock
->work
);
2421 mutex_unlock(&percpu_charge_mutex
);
2424 static int memcg_hotplug_cpu_dead(unsigned int cpu
)
2426 struct memcg_stock_pcp
*stock
;
2428 stock
= &per_cpu(memcg_stock
, cpu
);
2434 static unsigned long reclaim_high(struct mem_cgroup
*memcg
,
2435 unsigned int nr_pages
,
2438 unsigned long nr_reclaimed
= 0;
2441 unsigned long pflags
;
2443 if (page_counter_read(&memcg
->memory
) <=
2444 READ_ONCE(memcg
->memory
.high
))
2447 memcg_memory_event(memcg
, MEMCG_HIGH
);
2449 psi_memstall_enter(&pflags
);
2450 nr_reclaimed
+= try_to_free_mem_cgroup_pages(memcg
, nr_pages
,
2452 MEMCG_RECLAIM_MAY_SWAP
);
2453 psi_memstall_leave(&pflags
);
2454 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2455 !mem_cgroup_is_root(memcg
));
2457 return nr_reclaimed
;
2460 static void high_work_func(struct work_struct
*work
)
2462 struct mem_cgroup
*memcg
;
2464 memcg
= container_of(work
, struct mem_cgroup
, high_work
);
2465 reclaim_high(memcg
, MEMCG_CHARGE_BATCH
, GFP_KERNEL
);
2469 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2470 * enough to still cause a significant slowdown in most cases, while still
2471 * allowing diagnostics and tracing to proceed without becoming stuck.
2473 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2476 * When calculating the delay, we use these either side of the exponentiation to
2477 * maintain precision and scale to a reasonable number of jiffies (see the table
2480 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2481 * overage ratio to a delay.
2482 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2483 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2484 * to produce a reasonable delay curve.
2486 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2487 * reasonable delay curve compared to precision-adjusted overage, not
2488 * penalising heavily at first, but still making sure that growth beyond the
2489 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2490 * example, with a high of 100 megabytes:
2492 * +-------+------------------------+
2493 * | usage | time to allocate in ms |
2494 * +-------+------------------------+
2516 * +-------+------------------------+
2518 #define MEMCG_DELAY_PRECISION_SHIFT 20
2519 #define MEMCG_DELAY_SCALING_SHIFT 14
2521 static u64
calculate_overage(unsigned long usage
, unsigned long high
)
2529 * Prevent division by 0 in overage calculation by acting as if
2530 * it was a threshold of 1 page
2532 high
= max(high
, 1UL);
2534 overage
= usage
- high
;
2535 overage
<<= MEMCG_DELAY_PRECISION_SHIFT
;
2536 return div64_u64(overage
, high
);
2539 static u64
mem_find_max_overage(struct mem_cgroup
*memcg
)
2541 u64 overage
, max_overage
= 0;
2544 overage
= calculate_overage(page_counter_read(&memcg
->memory
),
2545 READ_ONCE(memcg
->memory
.high
));
2546 max_overage
= max(overage
, max_overage
);
2547 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2548 !mem_cgroup_is_root(memcg
));
2553 static u64
swap_find_max_overage(struct mem_cgroup
*memcg
)
2555 u64 overage
, max_overage
= 0;
2558 overage
= calculate_overage(page_counter_read(&memcg
->swap
),
2559 READ_ONCE(memcg
->swap
.high
));
2561 memcg_memory_event(memcg
, MEMCG_SWAP_HIGH
);
2562 max_overage
= max(overage
, max_overage
);
2563 } while ((memcg
= parent_mem_cgroup(memcg
)) &&
2564 !mem_cgroup_is_root(memcg
));
2570 * Get the number of jiffies that we should penalise a mischievous cgroup which
2571 * is exceeding its memory.high by checking both it and its ancestors.
2573 static unsigned long calculate_high_delay(struct mem_cgroup
*memcg
,
2574 unsigned int nr_pages
,
2577 unsigned long penalty_jiffies
;
2583 * We use overage compared to memory.high to calculate the number of
2584 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2585 * fairly lenient on small overages, and increasingly harsh when the
2586 * memcg in question makes it clear that it has no intention of stopping
2587 * its crazy behaviour, so we exponentially increase the delay based on
2590 penalty_jiffies
= max_overage
* max_overage
* HZ
;
2591 penalty_jiffies
>>= MEMCG_DELAY_PRECISION_SHIFT
;
2592 penalty_jiffies
>>= MEMCG_DELAY_SCALING_SHIFT
;
2595 * Factor in the task's own contribution to the overage, such that four
2596 * N-sized allocations are throttled approximately the same as one
2597 * 4N-sized allocation.
2599 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2600 * larger the current charge patch is than that.
2602 return penalty_jiffies
* nr_pages
/ MEMCG_CHARGE_BATCH
;
2606 * Scheduled by try_charge() to be executed from the userland return path
2607 * and reclaims memory over the high limit.
2609 void mem_cgroup_handle_over_high(gfp_t gfp_mask
)
2611 unsigned long penalty_jiffies
;
2612 unsigned long pflags
;
2613 unsigned long nr_reclaimed
;
2614 unsigned int nr_pages
= current
->memcg_nr_pages_over_high
;
2615 int nr_retries
= MAX_RECLAIM_RETRIES
;
2616 struct mem_cgroup
*memcg
;
2617 bool in_retry
= false;
2619 if (likely(!nr_pages
))
2622 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2623 current
->memcg_nr_pages_over_high
= 0;
2627 * The allocating task should reclaim at least the batch size, but for
2628 * subsequent retries we only want to do what's necessary to prevent oom
2629 * or breaching resource isolation.
2631 * This is distinct from memory.max or page allocator behaviour because
2632 * memory.high is currently batched, whereas memory.max and the page
2633 * allocator run every time an allocation is made.
2635 nr_reclaimed
= reclaim_high(memcg
,
2636 in_retry
? SWAP_CLUSTER_MAX
: nr_pages
,
2640 * memory.high is breached and reclaim is unable to keep up. Throttle
2641 * allocators proactively to slow down excessive growth.
2643 penalty_jiffies
= calculate_high_delay(memcg
, nr_pages
,
2644 mem_find_max_overage(memcg
));
2646 penalty_jiffies
+= calculate_high_delay(memcg
, nr_pages
,
2647 swap_find_max_overage(memcg
));
2650 * Clamp the max delay per usermode return so as to still keep the
2651 * application moving forwards and also permit diagnostics, albeit
2654 penalty_jiffies
= min(penalty_jiffies
, MEMCG_MAX_HIGH_DELAY_JIFFIES
);
2657 * Don't sleep if the amount of jiffies this memcg owes us is so low
2658 * that it's not even worth doing, in an attempt to be nice to those who
2659 * go only a small amount over their memory.high value and maybe haven't
2660 * been aggressively reclaimed enough yet.
2662 if (penalty_jiffies
<= HZ
/ 100)
2666 * If reclaim is making forward progress but we're still over
2667 * memory.high, we want to encourage that rather than doing allocator
2670 if (nr_reclaimed
|| nr_retries
--) {
2676 * If we exit early, we're guaranteed to die (since
2677 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2678 * need to account for any ill-begotten jiffies to pay them off later.
2680 psi_memstall_enter(&pflags
);
2681 schedule_timeout_killable(penalty_jiffies
);
2682 psi_memstall_leave(&pflags
);
2685 css_put(&memcg
->css
);
2688 static int try_charge_memcg(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2689 unsigned int nr_pages
)
2691 unsigned int batch
= max(MEMCG_CHARGE_BATCH
, nr_pages
);
2692 int nr_retries
= MAX_RECLAIM_RETRIES
;
2693 struct mem_cgroup
*mem_over_limit
;
2694 struct page_counter
*counter
;
2695 unsigned long nr_reclaimed
;
2696 bool passed_oom
= false;
2697 unsigned int reclaim_options
= MEMCG_RECLAIM_MAY_SWAP
;
2698 bool drained
= false;
2699 bool raised_max_event
= false;
2700 unsigned long pflags
;
2703 if (consume_stock(memcg
, nr_pages
))
2706 if (!do_memsw_account() ||
2707 page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2708 if (page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2710 if (do_memsw_account())
2711 page_counter_uncharge(&memcg
->memsw
, batch
);
2712 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2714 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2715 reclaim_options
&= ~MEMCG_RECLAIM_MAY_SWAP
;
2718 if (batch
> nr_pages
) {
2724 * Prevent unbounded recursion when reclaim operations need to
2725 * allocate memory. This might exceed the limits temporarily,
2726 * but we prefer facilitating memory reclaim and getting back
2727 * under the limit over triggering OOM kills in these cases.
2729 if (unlikely(current
->flags
& PF_MEMALLOC
))
2732 if (unlikely(task_in_memcg_oom(current
)))
2735 if (!gfpflags_allow_blocking(gfp_mask
))
2738 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2739 raised_max_event
= true;
2741 psi_memstall_enter(&pflags
);
2742 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2743 gfp_mask
, reclaim_options
);
2744 psi_memstall_leave(&pflags
);
2746 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2750 drain_all_stock(mem_over_limit
);
2755 if (gfp_mask
& __GFP_NORETRY
)
2758 * Even though the limit is exceeded at this point, reclaim
2759 * may have been able to free some pages. Retry the charge
2760 * before killing the task.
2762 * Only for regular pages, though: huge pages are rather
2763 * unlikely to succeed so close to the limit, and we fall back
2764 * to regular pages anyway in case of failure.
2766 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2769 * At task move, charge accounts can be doubly counted. So, it's
2770 * better to wait until the end of task_move if something is going on.
2772 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2778 if (gfp_mask
& __GFP_RETRY_MAYFAIL
)
2781 /* Avoid endless loop for tasks bypassed by the oom killer */
2782 if (passed_oom
&& task_is_dying())
2786 * keep retrying as long as the memcg oom killer is able to make
2787 * a forward progress or bypass the charge if the oom killer
2788 * couldn't make any progress.
2790 if (mem_cgroup_oom(mem_over_limit
, gfp_mask
,
2791 get_order(nr_pages
* PAGE_SIZE
))) {
2793 nr_retries
= MAX_RECLAIM_RETRIES
;
2798 * Memcg doesn't have a dedicated reserve for atomic
2799 * allocations. But like the global atomic pool, we need to
2800 * put the burden of reclaim on regular allocation requests
2801 * and let these go through as privileged allocations.
2803 if (!(gfp_mask
& (__GFP_NOFAIL
| __GFP_HIGH
)))
2807 * If the allocation has to be enforced, don't forget to raise
2808 * a MEMCG_MAX event.
2810 if (!raised_max_event
)
2811 memcg_memory_event(mem_over_limit
, MEMCG_MAX
);
2814 * The allocation either can't fail or will lead to more memory
2815 * being freed very soon. Allow memory usage go over the limit
2816 * temporarily by force charging it.
2818 page_counter_charge(&memcg
->memory
, nr_pages
);
2819 if (do_memsw_account())
2820 page_counter_charge(&memcg
->memsw
, nr_pages
);
2825 if (batch
> nr_pages
)
2826 refill_stock(memcg
, batch
- nr_pages
);
2829 * If the hierarchy is above the normal consumption range, schedule
2830 * reclaim on returning to userland. We can perform reclaim here
2831 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2832 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2833 * not recorded as it most likely matches current's and won't
2834 * change in the meantime. As high limit is checked again before
2835 * reclaim, the cost of mismatch is negligible.
2838 bool mem_high
, swap_high
;
2840 mem_high
= page_counter_read(&memcg
->memory
) >
2841 READ_ONCE(memcg
->memory
.high
);
2842 swap_high
= page_counter_read(&memcg
->swap
) >
2843 READ_ONCE(memcg
->swap
.high
);
2845 /* Don't bother a random interrupted task */
2848 schedule_work(&memcg
->high_work
);
2854 if (mem_high
|| swap_high
) {
2856 * The allocating tasks in this cgroup will need to do
2857 * reclaim or be throttled to prevent further growth
2858 * of the memory or swap footprints.
2860 * Target some best-effort fairness between the tasks,
2861 * and distribute reclaim work and delay penalties
2862 * based on how much each task is actually allocating.
2864 current
->memcg_nr_pages_over_high
+= batch
;
2865 set_notify_resume(current
);
2868 } while ((memcg
= parent_mem_cgroup(memcg
)));
2870 if (current
->memcg_nr_pages_over_high
> MEMCG_CHARGE_BATCH
&&
2871 !(current
->flags
& PF_MEMALLOC
) &&
2872 gfpflags_allow_blocking(gfp_mask
)) {
2873 mem_cgroup_handle_over_high(gfp_mask
);
2878 static inline int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2879 unsigned int nr_pages
)
2881 if (mem_cgroup_is_root(memcg
))
2884 return try_charge_memcg(memcg
, gfp_mask
, nr_pages
);
2888 * mem_cgroup_cancel_charge() - cancel an uncommitted try_charge() call.
2889 * @memcg: memcg previously charged.
2890 * @nr_pages: number of pages previously charged.
2892 void mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2894 if (mem_cgroup_is_root(memcg
))
2897 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2898 if (do_memsw_account())
2899 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2902 static void commit_charge(struct folio
*folio
, struct mem_cgroup
*memcg
)
2904 VM_BUG_ON_FOLIO(folio_memcg(folio
), folio
);
2906 * Any of the following ensures page's memcg stability:
2910 * - folio_memcg_lock()
2911 * - exclusive reference
2912 * - mem_cgroup_trylock_pages()
2914 folio
->memcg_data
= (unsigned long)memcg
;
2918 * mem_cgroup_commit_charge - commit a previously successful try_charge().
2919 * @folio: folio to commit the charge to.
2920 * @memcg: memcg previously charged.
2922 void mem_cgroup_commit_charge(struct folio
*folio
, struct mem_cgroup
*memcg
)
2924 css_get(&memcg
->css
);
2925 commit_charge(folio
, memcg
);
2927 local_irq_disable();
2928 mem_cgroup_charge_statistics(memcg
, folio_nr_pages(folio
));
2929 memcg_check_events(memcg
, folio_nid(folio
));
2933 #ifdef CONFIG_MEMCG_KMEM
2935 * The allocated objcg pointers array is not accounted directly.
2936 * Moreover, it should not come from DMA buffer and is not readily
2937 * reclaimable. So those GFP bits should be masked off.
2939 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | __GFP_ACCOUNT)
2942 * mod_objcg_mlstate() may be called with irq enabled, so
2943 * mod_memcg_lruvec_state() should be used.
2945 static inline void mod_objcg_mlstate(struct obj_cgroup
*objcg
,
2946 struct pglist_data
*pgdat
,
2947 enum node_stat_item idx
, int nr
)
2949 struct mem_cgroup
*memcg
;
2950 struct lruvec
*lruvec
;
2953 memcg
= obj_cgroup_memcg(objcg
);
2954 lruvec
= mem_cgroup_lruvec(memcg
, pgdat
);
2955 mod_memcg_lruvec_state(lruvec
, idx
, nr
);
2959 int memcg_alloc_slab_cgroups(struct slab
*slab
, struct kmem_cache
*s
,
2960 gfp_t gfp
, bool new_slab
)
2962 unsigned int objects
= objs_per_slab(s
, slab
);
2963 unsigned long memcg_data
;
2966 gfp
&= ~OBJCGS_CLEAR_MASK
;
2967 vec
= kcalloc_node(objects
, sizeof(struct obj_cgroup
*), gfp
,
2972 memcg_data
= (unsigned long) vec
| MEMCG_DATA_OBJCGS
;
2975 * If the slab is brand new and nobody can yet access its
2976 * memcg_data, no synchronization is required and memcg_data can
2977 * be simply assigned.
2979 slab
->memcg_data
= memcg_data
;
2980 } else if (cmpxchg(&slab
->memcg_data
, 0, memcg_data
)) {
2982 * If the slab is already in use, somebody can allocate and
2983 * assign obj_cgroups in parallel. In this case the existing
2984 * objcg vector should be reused.
2990 kmemleak_not_leak(vec
);
2994 static __always_inline
2995 struct mem_cgroup
*mem_cgroup_from_obj_folio(struct folio
*folio
, void *p
)
2998 * Slab objects are accounted individually, not per-page.
2999 * Memcg membership data for each individual object is saved in
3002 if (folio_test_slab(folio
)) {
3003 struct obj_cgroup
**objcgs
;
3007 slab
= folio_slab(folio
);
3008 objcgs
= slab_objcgs(slab
);
3012 off
= obj_to_index(slab
->slab_cache
, slab
, p
);
3014 return obj_cgroup_memcg(objcgs
[off
]);
3020 * folio_memcg_check() is used here, because in theory we can encounter
3021 * a folio where the slab flag has been cleared already, but
3022 * slab->memcg_data has not been freed yet
3023 * folio_memcg_check() will guarantee that a proper memory
3024 * cgroup pointer or NULL will be returned.
3026 return folio_memcg_check(folio
);
3030 * Returns a pointer to the memory cgroup to which the kernel object is charged.
3032 * A passed kernel object can be a slab object, vmalloc object or a generic
3033 * kernel page, so different mechanisms for getting the memory cgroup pointer
3036 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
3037 * can not know for sure how the kernel object is implemented.
3038 * mem_cgroup_from_obj() can be safely used in such cases.
3040 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3041 * cgroup_mutex, etc.
3043 struct mem_cgroup
*mem_cgroup_from_obj(void *p
)
3045 struct folio
*folio
;
3047 if (mem_cgroup_disabled())
3050 if (unlikely(is_vmalloc_addr(p
)))
3051 folio
= page_folio(vmalloc_to_page(p
));
3053 folio
= virt_to_folio(p
);
3055 return mem_cgroup_from_obj_folio(folio
, p
);
3059 * Returns a pointer to the memory cgroup to which the kernel object is charged.
3060 * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
3061 * allocated using vmalloc().
3063 * A passed kernel object must be a slab object or a generic kernel page.
3065 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3066 * cgroup_mutex, etc.
3068 struct mem_cgroup
*mem_cgroup_from_slab_obj(void *p
)
3070 if (mem_cgroup_disabled())
3073 return mem_cgroup_from_obj_folio(virt_to_folio(p
), p
);
3076 static struct obj_cgroup
*__get_obj_cgroup_from_memcg(struct mem_cgroup
*memcg
)
3078 struct obj_cgroup
*objcg
= NULL
;
3080 for (; !mem_cgroup_is_root(memcg
); memcg
= parent_mem_cgroup(memcg
)) {
3081 objcg
= rcu_dereference(memcg
->objcg
);
3082 if (likely(objcg
&& obj_cgroup_tryget(objcg
)))
3089 static struct obj_cgroup
*current_objcg_update(void)
3091 struct mem_cgroup
*memcg
;
3092 struct obj_cgroup
*old
, *objcg
= NULL
;
3095 /* Atomically drop the update bit. */
3096 old
= xchg(¤t
->objcg
, NULL
);
3098 old
= (struct obj_cgroup
*)
3099 ((unsigned long)old
& ~CURRENT_OBJCG_UPDATE_FLAG
);
3101 obj_cgroup_put(old
);
3106 /* If new objcg is NULL, no reason for the second atomic update. */
3107 if (!current
->mm
|| (current
->flags
& PF_KTHREAD
))
3111 * Release the objcg pointer from the previous iteration,
3112 * if try_cmpxcg() below fails.
3114 if (unlikely(objcg
)) {
3115 obj_cgroup_put(objcg
);
3120 * Obtain the new objcg pointer. The current task can be
3121 * asynchronously moved to another memcg and the previous
3122 * memcg can be offlined. So let's get the memcg pointer
3123 * and try get a reference to objcg under a rcu read lock.
3127 memcg
= mem_cgroup_from_task(current
);
3128 objcg
= __get_obj_cgroup_from_memcg(memcg
);
3132 * Try set up a new objcg pointer atomically. If it
3133 * fails, it means the update flag was set concurrently, so
3134 * the whole procedure should be repeated.
3136 } while (!try_cmpxchg(¤t
->objcg
, &old
, objcg
));
3141 __always_inline
struct obj_cgroup
*current_obj_cgroup(void)
3143 struct mem_cgroup
*memcg
;
3144 struct obj_cgroup
*objcg
;
3147 memcg
= current
->active_memcg
;
3148 if (unlikely(memcg
))
3151 objcg
= READ_ONCE(current
->objcg
);
3152 if (unlikely((unsigned long)objcg
& CURRENT_OBJCG_UPDATE_FLAG
))
3153 objcg
= current_objcg_update();
3155 * Objcg reference is kept by the task, so it's safe
3156 * to use the objcg by the current task.
3161 memcg
= this_cpu_read(int_active_memcg
);
3162 if (unlikely(memcg
))
3168 for (; !mem_cgroup_is_root(memcg
); memcg
= parent_mem_cgroup(memcg
)) {
3170 * Memcg pointer is protected by scope (see set_active_memcg())
3171 * and is pinning the corresponding objcg, so objcg can't go
3172 * away and can be used within the scope without any additional
3175 objcg
= rcu_dereference_check(memcg
->objcg
, 1);
3184 struct obj_cgroup
*get_obj_cgroup_from_folio(struct folio
*folio
)
3186 struct obj_cgroup
*objcg
;
3188 if (!memcg_kmem_online())
3191 if (folio_memcg_kmem(folio
)) {
3192 objcg
= __folio_objcg(folio
);
3193 obj_cgroup_get(objcg
);
3195 struct mem_cgroup
*memcg
;
3198 memcg
= __folio_memcg(folio
);
3200 objcg
= __get_obj_cgroup_from_memcg(memcg
);
3208 static void memcg_account_kmem(struct mem_cgroup
*memcg
, int nr_pages
)
3210 mod_memcg_state(memcg
, MEMCG_KMEM
, nr_pages
);
3211 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
3213 page_counter_charge(&memcg
->kmem
, nr_pages
);
3215 page_counter_uncharge(&memcg
->kmem
, -nr_pages
);
3221 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3222 * @objcg: object cgroup to uncharge
3223 * @nr_pages: number of pages to uncharge
3225 static void obj_cgroup_uncharge_pages(struct obj_cgroup
*objcg
,
3226 unsigned int nr_pages
)
3228 struct mem_cgroup
*memcg
;
3230 memcg
= get_mem_cgroup_from_objcg(objcg
);
3232 memcg_account_kmem(memcg
, -nr_pages
);
3233 refill_stock(memcg
, nr_pages
);
3235 css_put(&memcg
->css
);
3239 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3240 * @objcg: object cgroup to charge
3241 * @gfp: reclaim mode
3242 * @nr_pages: number of pages to charge
3244 * Returns 0 on success, an error code on failure.
3246 static int obj_cgroup_charge_pages(struct obj_cgroup
*objcg
, gfp_t gfp
,
3247 unsigned int nr_pages
)
3249 struct mem_cgroup
*memcg
;
3252 memcg
= get_mem_cgroup_from_objcg(objcg
);
3254 ret
= try_charge_memcg(memcg
, gfp
, nr_pages
);
3258 memcg_account_kmem(memcg
, nr_pages
);
3260 css_put(&memcg
->css
);
3266 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3267 * @page: page to charge
3268 * @gfp: reclaim mode
3269 * @order: allocation order
3271 * Returns 0 on success, an error code on failure.
3273 int __memcg_kmem_charge_page(struct page
*page
, gfp_t gfp
, int order
)
3275 struct obj_cgroup
*objcg
;
3278 objcg
= current_obj_cgroup();
3280 ret
= obj_cgroup_charge_pages(objcg
, gfp
, 1 << order
);
3282 obj_cgroup_get(objcg
);
3283 page
->memcg_data
= (unsigned long)objcg
|
3292 * __memcg_kmem_uncharge_page: uncharge a kmem page
3293 * @page: page to uncharge
3294 * @order: allocation order
3296 void __memcg_kmem_uncharge_page(struct page
*page
, int order
)
3298 struct folio
*folio
= page_folio(page
);
3299 struct obj_cgroup
*objcg
;
3300 unsigned int nr_pages
= 1 << order
;
3302 if (!folio_memcg_kmem(folio
))
3305 objcg
= __folio_objcg(folio
);
3306 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3307 folio
->memcg_data
= 0;
3308 obj_cgroup_put(objcg
);
3311 void mod_objcg_state(struct obj_cgroup
*objcg
, struct pglist_data
*pgdat
,
3312 enum node_stat_item idx
, int nr
)
3314 struct memcg_stock_pcp
*stock
;
3315 struct obj_cgroup
*old
= NULL
;
3316 unsigned long flags
;
3319 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
3320 stock
= this_cpu_ptr(&memcg_stock
);
3323 * Save vmstat data in stock and skip vmstat array update unless
3324 * accumulating over a page of vmstat data or when pgdat or idx
3327 if (READ_ONCE(stock
->cached_objcg
) != objcg
) {
3328 old
= drain_obj_stock(stock
);
3329 obj_cgroup_get(objcg
);
3330 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3331 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3332 WRITE_ONCE(stock
->cached_objcg
, objcg
);
3333 stock
->cached_pgdat
= pgdat
;
3334 } else if (stock
->cached_pgdat
!= pgdat
) {
3335 /* Flush the existing cached vmstat data */
3336 struct pglist_data
*oldpg
= stock
->cached_pgdat
;
3338 if (stock
->nr_slab_reclaimable_b
) {
3339 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_RECLAIMABLE_B
,
3340 stock
->nr_slab_reclaimable_b
);
3341 stock
->nr_slab_reclaimable_b
= 0;
3343 if (stock
->nr_slab_unreclaimable_b
) {
3344 mod_objcg_mlstate(objcg
, oldpg
, NR_SLAB_UNRECLAIMABLE_B
,
3345 stock
->nr_slab_unreclaimable_b
);
3346 stock
->nr_slab_unreclaimable_b
= 0;
3348 stock
->cached_pgdat
= pgdat
;
3351 bytes
= (idx
== NR_SLAB_RECLAIMABLE_B
) ? &stock
->nr_slab_reclaimable_b
3352 : &stock
->nr_slab_unreclaimable_b
;
3354 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3355 * cached locally at least once before pushing it out.
3362 if (abs(*bytes
) > PAGE_SIZE
) {
3370 mod_objcg_mlstate(objcg
, pgdat
, idx
, nr
);
3372 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
3374 obj_cgroup_put(old
);
3377 static bool consume_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
)
3379 struct memcg_stock_pcp
*stock
;
3380 unsigned long flags
;
3383 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
3385 stock
= this_cpu_ptr(&memcg_stock
);
3386 if (objcg
== READ_ONCE(stock
->cached_objcg
) && stock
->nr_bytes
>= nr_bytes
) {
3387 stock
->nr_bytes
-= nr_bytes
;
3391 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
3396 static struct obj_cgroup
*drain_obj_stock(struct memcg_stock_pcp
*stock
)
3398 struct obj_cgroup
*old
= READ_ONCE(stock
->cached_objcg
);
3403 if (stock
->nr_bytes
) {
3404 unsigned int nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3405 unsigned int nr_bytes
= stock
->nr_bytes
& (PAGE_SIZE
- 1);
3408 struct mem_cgroup
*memcg
;
3410 memcg
= get_mem_cgroup_from_objcg(old
);
3412 memcg_account_kmem(memcg
, -nr_pages
);
3413 __refill_stock(memcg
, nr_pages
);
3415 css_put(&memcg
->css
);
3419 * The leftover is flushed to the centralized per-memcg value.
3420 * On the next attempt to refill obj stock it will be moved
3421 * to a per-cpu stock (probably, on an other CPU), see
3422 * refill_obj_stock().
3424 * How often it's flushed is a trade-off between the memory
3425 * limit enforcement accuracy and potential CPU contention,
3426 * so it might be changed in the future.
3428 atomic_add(nr_bytes
, &old
->nr_charged_bytes
);
3429 stock
->nr_bytes
= 0;
3433 * Flush the vmstat data in current stock
3435 if (stock
->nr_slab_reclaimable_b
|| stock
->nr_slab_unreclaimable_b
) {
3436 if (stock
->nr_slab_reclaimable_b
) {
3437 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3438 NR_SLAB_RECLAIMABLE_B
,
3439 stock
->nr_slab_reclaimable_b
);
3440 stock
->nr_slab_reclaimable_b
= 0;
3442 if (stock
->nr_slab_unreclaimable_b
) {
3443 mod_objcg_mlstate(old
, stock
->cached_pgdat
,
3444 NR_SLAB_UNRECLAIMABLE_B
,
3445 stock
->nr_slab_unreclaimable_b
);
3446 stock
->nr_slab_unreclaimable_b
= 0;
3448 stock
->cached_pgdat
= NULL
;
3451 WRITE_ONCE(stock
->cached_objcg
, NULL
);
3453 * The `old' objects needs to be released by the caller via
3454 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3459 static bool obj_stock_flush_required(struct memcg_stock_pcp
*stock
,
3460 struct mem_cgroup
*root_memcg
)
3462 struct obj_cgroup
*objcg
= READ_ONCE(stock
->cached_objcg
);
3463 struct mem_cgroup
*memcg
;
3466 memcg
= obj_cgroup_memcg(objcg
);
3467 if (memcg
&& mem_cgroup_is_descendant(memcg
, root_memcg
))
3474 static void refill_obj_stock(struct obj_cgroup
*objcg
, unsigned int nr_bytes
,
3475 bool allow_uncharge
)
3477 struct memcg_stock_pcp
*stock
;
3478 struct obj_cgroup
*old
= NULL
;
3479 unsigned long flags
;
3480 unsigned int nr_pages
= 0;
3482 local_lock_irqsave(&memcg_stock
.stock_lock
, flags
);
3484 stock
= this_cpu_ptr(&memcg_stock
);
3485 if (READ_ONCE(stock
->cached_objcg
) != objcg
) { /* reset if necessary */
3486 old
= drain_obj_stock(stock
);
3487 obj_cgroup_get(objcg
);
3488 WRITE_ONCE(stock
->cached_objcg
, objcg
);
3489 stock
->nr_bytes
= atomic_read(&objcg
->nr_charged_bytes
)
3490 ? atomic_xchg(&objcg
->nr_charged_bytes
, 0) : 0;
3491 allow_uncharge
= true; /* Allow uncharge when objcg changes */
3493 stock
->nr_bytes
+= nr_bytes
;
3495 if (allow_uncharge
&& (stock
->nr_bytes
> PAGE_SIZE
)) {
3496 nr_pages
= stock
->nr_bytes
>> PAGE_SHIFT
;
3497 stock
->nr_bytes
&= (PAGE_SIZE
- 1);
3500 local_unlock_irqrestore(&memcg_stock
.stock_lock
, flags
);
3502 obj_cgroup_put(old
);
3505 obj_cgroup_uncharge_pages(objcg
, nr_pages
);
3508 int obj_cgroup_charge(struct obj_cgroup
*objcg
, gfp_t gfp
, size_t size
)
3510 unsigned int nr_pages
, nr_bytes
;
3513 if (consume_obj_stock(objcg
, size
))
3517 * In theory, objcg->nr_charged_bytes can have enough
3518 * pre-charged bytes to satisfy the allocation. However,
3519 * flushing objcg->nr_charged_bytes requires two atomic
3520 * operations, and objcg->nr_charged_bytes can't be big.
3521 * The shared objcg->nr_charged_bytes can also become a
3522 * performance bottleneck if all tasks of the same memcg are
3523 * trying to update it. So it's better to ignore it and try
3524 * grab some new pages. The stock's nr_bytes will be flushed to
3525 * objcg->nr_charged_bytes later on when objcg changes.
3527 * The stock's nr_bytes may contain enough pre-charged bytes
3528 * to allow one less page from being charged, but we can't rely
3529 * on the pre-charged bytes not being changed outside of
3530 * consume_obj_stock() or refill_obj_stock(). So ignore those
3531 * pre-charged bytes as well when charging pages. To avoid a
3532 * page uncharge right after a page charge, we set the
3533 * allow_uncharge flag to false when calling refill_obj_stock()
3534 * to temporarily allow the pre-charged bytes to exceed the page
3535 * size limit. The maximum reachable value of the pre-charged
3536 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3539 nr_pages
= size
>> PAGE_SHIFT
;
3540 nr_bytes
= size
& (PAGE_SIZE
- 1);
3545 ret
= obj_cgroup_charge_pages(objcg
, gfp
, nr_pages
);
3546 if (!ret
&& nr_bytes
)
3547 refill_obj_stock(objcg
, PAGE_SIZE
- nr_bytes
, false);
3552 void obj_cgroup_uncharge(struct obj_cgroup
*objcg
, size_t size
)
3554 refill_obj_stock(objcg
, size
, true);
3557 #endif /* CONFIG_MEMCG_KMEM */
3560 * Because page_memcg(head) is not set on tails, set it now.
3562 void split_page_memcg(struct page
*head
, unsigned int nr
)
3564 struct folio
*folio
= page_folio(head
);
3565 struct mem_cgroup
*memcg
= folio_memcg(folio
);
3568 if (mem_cgroup_disabled() || !memcg
)
3571 for (i
= 1; i
< nr
; i
++)
3572 folio_page(folio
, i
)->memcg_data
= folio
->memcg_data
;
3574 if (folio_memcg_kmem(folio
))
3575 obj_cgroup_get_many(__folio_objcg(folio
), nr
- 1);
3577 css_get_many(&memcg
->css
, nr
- 1);
3582 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3583 * @entry: swap entry to be moved
3584 * @from: mem_cgroup which the entry is moved from
3585 * @to: mem_cgroup which the entry is moved to
3587 * It succeeds only when the swap_cgroup's record for this entry is the same
3588 * as the mem_cgroup's id of @from.
3590 * Returns 0 on success, -EINVAL on failure.
3592 * The caller must have charged to @to, IOW, called page_counter_charge() about
3593 * both res and memsw, and called css_get().
3595 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3596 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3598 unsigned short old_id
, new_id
;
3600 old_id
= mem_cgroup_id(from
);
3601 new_id
= mem_cgroup_id(to
);
3603 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3604 mod_memcg_state(from
, MEMCG_SWAP
, -1);
3605 mod_memcg_state(to
, MEMCG_SWAP
, 1);
3611 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3612 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3618 static DEFINE_MUTEX(memcg_max_mutex
);
3620 static int mem_cgroup_resize_max(struct mem_cgroup
*memcg
,
3621 unsigned long max
, bool memsw
)
3623 bool enlarge
= false;
3624 bool drained
= false;
3626 bool limits_invariant
;
3627 struct page_counter
*counter
= memsw
? &memcg
->memsw
: &memcg
->memory
;
3630 if (signal_pending(current
)) {
3635 mutex_lock(&memcg_max_mutex
);
3637 * Make sure that the new limit (memsw or memory limit) doesn't
3638 * break our basic invariant rule memory.max <= memsw.max.
3640 limits_invariant
= memsw
? max
>= READ_ONCE(memcg
->memory
.max
) :
3641 max
<= memcg
->memsw
.max
;
3642 if (!limits_invariant
) {
3643 mutex_unlock(&memcg_max_mutex
);
3647 if (max
> counter
->max
)
3649 ret
= page_counter_set_max(counter
, max
);
3650 mutex_unlock(&memcg_max_mutex
);
3656 drain_all_stock(memcg
);
3661 if (!try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
,
3662 memsw
? 0 : MEMCG_RECLAIM_MAY_SWAP
)) {
3668 if (!ret
&& enlarge
)
3669 memcg_oom_recover(memcg
);
3674 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t
*pgdat
, int order
,
3676 unsigned long *total_scanned
)
3678 unsigned long nr_reclaimed
= 0;
3679 struct mem_cgroup_per_node
*mz
, *next_mz
= NULL
;
3680 unsigned long reclaimed
;
3682 struct mem_cgroup_tree_per_node
*mctz
;
3683 unsigned long excess
;
3685 if (lru_gen_enabled())
3691 mctz
= soft_limit_tree
.rb_tree_per_node
[pgdat
->node_id
];
3694 * Do not even bother to check the largest node if the root
3695 * is empty. Do it lockless to prevent lock bouncing. Races
3696 * are acceptable as soft limit is best effort anyway.
3698 if (!mctz
|| RB_EMPTY_ROOT(&mctz
->rb_root
))
3702 * This loop can run a while, specially if mem_cgroup's continuously
3703 * keep exceeding their soft limit and putting the system under
3710 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3714 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, pgdat
,
3715 gfp_mask
, total_scanned
);
3716 nr_reclaimed
+= reclaimed
;
3717 spin_lock_irq(&mctz
->lock
);
3720 * If we failed to reclaim anything from this memory cgroup
3721 * it is time to move on to the next cgroup
3725 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3727 excess
= soft_limit_excess(mz
->memcg
);
3729 * One school of thought says that we should not add
3730 * back the node to the tree if reclaim returns 0.
3731 * But our reclaim could return 0, simply because due
3732 * to priority we are exposing a smaller subset of
3733 * memory to reclaim from. Consider this as a longer
3736 /* If excess == 0, no tree ops */
3737 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3738 spin_unlock_irq(&mctz
->lock
);
3739 css_put(&mz
->memcg
->css
);
3742 * Could not reclaim anything and there are no more
3743 * mem cgroups to try or we seem to be looping without
3744 * reclaiming anything.
3746 if (!nr_reclaimed
&&
3748 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3750 } while (!nr_reclaimed
);
3752 css_put(&next_mz
->memcg
->css
);
3753 return nr_reclaimed
;
3757 * Reclaims as many pages from the given memcg as possible.
3759 * Caller is responsible for holding css reference for memcg.
3761 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3763 int nr_retries
= MAX_RECLAIM_RETRIES
;
3765 /* we call try-to-free pages for make this cgroup empty */
3766 lru_add_drain_all();
3768 drain_all_stock(memcg
);
3770 /* try to free all pages in this cgroup */
3771 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3772 if (signal_pending(current
))
3775 if (!try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
,
3776 MEMCG_RECLAIM_MAY_SWAP
))
3783 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3784 char *buf
, size_t nbytes
,
3787 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3789 if (mem_cgroup_is_root(memcg
))
3791 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3794 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3800 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3801 struct cftype
*cft
, u64 val
)
3806 pr_warn_once("Non-hierarchical mode is deprecated. "
3807 "Please report your usecase to linux-mm@kvack.org if you "
3808 "depend on this functionality.\n");
3813 static unsigned long mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3817 if (mem_cgroup_is_root(memcg
)) {
3819 * Approximate root's usage from global state. This isn't
3820 * perfect, but the root usage was always an approximation.
3822 val
= global_node_page_state(NR_FILE_PAGES
) +
3823 global_node_page_state(NR_ANON_MAPPED
);
3825 val
+= total_swap_pages
- get_nr_swap_pages();
3828 val
= page_counter_read(&memcg
->memory
);
3830 val
= page_counter_read(&memcg
->memsw
);
3843 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3846 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3847 struct page_counter
*counter
;
3849 switch (MEMFILE_TYPE(cft
->private)) {
3851 counter
= &memcg
->memory
;
3854 counter
= &memcg
->memsw
;
3857 counter
= &memcg
->kmem
;
3860 counter
= &memcg
->tcpmem
;
3866 switch (MEMFILE_ATTR(cft
->private)) {
3868 if (counter
== &memcg
->memory
)
3869 return (u64
)mem_cgroup_usage(memcg
, false) * PAGE_SIZE
;
3870 if (counter
== &memcg
->memsw
)
3871 return (u64
)mem_cgroup_usage(memcg
, true) * PAGE_SIZE
;
3872 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3874 return (u64
)counter
->max
* PAGE_SIZE
;
3876 return (u64
)counter
->watermark
* PAGE_SIZE
;
3878 return counter
->failcnt
;
3879 case RES_SOFT_LIMIT
:
3880 return (u64
)READ_ONCE(memcg
->soft_limit
) * PAGE_SIZE
;
3887 * This function doesn't do anything useful. Its only job is to provide a read
3888 * handler for a file so that cgroup_file_mode() will add read permissions.
3890 static int mem_cgroup_dummy_seq_show(__always_unused
struct seq_file
*m
,
3891 __always_unused
void *v
)
3896 #ifdef CONFIG_MEMCG_KMEM
3897 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3899 struct obj_cgroup
*objcg
;
3901 if (mem_cgroup_kmem_disabled())
3904 if (unlikely(mem_cgroup_is_root(memcg
)))
3907 objcg
= obj_cgroup_alloc();
3911 objcg
->memcg
= memcg
;
3912 rcu_assign_pointer(memcg
->objcg
, objcg
);
3913 obj_cgroup_get(objcg
);
3914 memcg
->orig_objcg
= objcg
;
3916 static_branch_enable(&memcg_kmem_online_key
);
3918 memcg
->kmemcg_id
= memcg
->id
.id
;
3923 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3925 struct mem_cgroup
*parent
;
3927 if (mem_cgroup_kmem_disabled())
3930 if (unlikely(mem_cgroup_is_root(memcg
)))
3933 parent
= parent_mem_cgroup(memcg
);
3935 parent
= root_mem_cgroup
;
3937 memcg_reparent_objcgs(memcg
, parent
);
3940 * After we have finished memcg_reparent_objcgs(), all list_lrus
3941 * corresponding to this cgroup are guaranteed to remain empty.
3942 * The ordering is imposed by list_lru_node->lock taken by
3943 * memcg_reparent_list_lrus().
3945 memcg_reparent_list_lrus(memcg
, parent
);
3948 static int memcg_online_kmem(struct mem_cgroup
*memcg
)
3952 static void memcg_offline_kmem(struct mem_cgroup
*memcg
)
3955 #endif /* CONFIG_MEMCG_KMEM */
3957 static int memcg_update_tcp_max(struct mem_cgroup
*memcg
, unsigned long max
)
3961 mutex_lock(&memcg_max_mutex
);
3963 ret
= page_counter_set_max(&memcg
->tcpmem
, max
);
3967 if (!memcg
->tcpmem_active
) {
3969 * The active flag needs to be written after the static_key
3970 * update. This is what guarantees that the socket activation
3971 * function is the last one to run. See mem_cgroup_sk_alloc()
3972 * for details, and note that we don't mark any socket as
3973 * belonging to this memcg until that flag is up.
3975 * We need to do this, because static_keys will span multiple
3976 * sites, but we can't control their order. If we mark a socket
3977 * as accounted, but the accounting functions are not patched in
3978 * yet, we'll lose accounting.
3980 * We never race with the readers in mem_cgroup_sk_alloc(),
3981 * because when this value change, the code to process it is not
3984 static_branch_inc(&memcg_sockets_enabled_key
);
3985 memcg
->tcpmem_active
= true;
3988 mutex_unlock(&memcg_max_mutex
);
3993 * The user of this function is...
3996 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3997 char *buf
, size_t nbytes
, loff_t off
)
3999 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4000 unsigned long nr_pages
;
4003 buf
= strstrip(buf
);
4004 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
4008 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
4010 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4014 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
4016 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, false);
4019 ret
= mem_cgroup_resize_max(memcg
, nr_pages
, true);
4022 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
4023 "Writing any value to this file has no effect. "
4024 "Please report your usecase to linux-mm@kvack.org if you "
4025 "depend on this functionality.\n");
4029 ret
= memcg_update_tcp_max(memcg
, nr_pages
);
4033 case RES_SOFT_LIMIT
:
4034 if (IS_ENABLED(CONFIG_PREEMPT_RT
)) {
4037 WRITE_ONCE(memcg
->soft_limit
, nr_pages
);
4042 return ret
?: nbytes
;
4045 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
4046 size_t nbytes
, loff_t off
)
4048 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4049 struct page_counter
*counter
;
4051 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
4053 counter
= &memcg
->memory
;
4056 counter
= &memcg
->memsw
;
4059 counter
= &memcg
->kmem
;
4062 counter
= &memcg
->tcpmem
;
4068 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
4070 page_counter_reset_watermark(counter
);
4073 counter
->failcnt
= 0;
4082 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
4085 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
4089 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4090 struct cftype
*cft
, u64 val
)
4092 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4094 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
4095 "Please report your usecase to linux-mm@kvack.org if you "
4096 "depend on this functionality.\n");
4098 if (val
& ~MOVE_MASK
)
4102 * No kind of locking is needed in here, because ->can_attach() will
4103 * check this value once in the beginning of the process, and then carry
4104 * on with stale data. This means that changes to this value will only
4105 * affect task migrations starting after the change.
4107 memcg
->move_charge_at_immigrate
= val
;
4111 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4112 struct cftype
*cft
, u64 val
)
4120 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
4121 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
4122 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
4124 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
4125 int nid
, unsigned int lru_mask
, bool tree
)
4127 struct lruvec
*lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
4128 unsigned long nr
= 0;
4131 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
4134 if (!(BIT(lru
) & lru_mask
))
4137 nr
+= lruvec_page_state(lruvec
, NR_LRU_BASE
+ lru
);
4139 nr
+= lruvec_page_state_local(lruvec
, NR_LRU_BASE
+ lru
);
4144 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
4145 unsigned int lru_mask
,
4148 unsigned long nr
= 0;
4152 if (!(BIT(lru
) & lru_mask
))
4155 nr
+= memcg_page_state(memcg
, NR_LRU_BASE
+ lru
);
4157 nr
+= memcg_page_state_local(memcg
, NR_LRU_BASE
+ lru
);
4162 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
4166 unsigned int lru_mask
;
4169 static const struct numa_stat stats
[] = {
4170 { "total", LRU_ALL
},
4171 { "file", LRU_ALL_FILE
},
4172 { "anon", LRU_ALL_ANON
},
4173 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4175 const struct numa_stat
*stat
;
4177 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
4179 mem_cgroup_flush_stats();
4181 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4182 seq_printf(m
, "%s=%lu", stat
->name
,
4183 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4185 for_each_node_state(nid
, N_MEMORY
)
4186 seq_printf(m
, " N%d=%lu", nid
,
4187 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4188 stat
->lru_mask
, false));
4192 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4194 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
,
4195 mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
,
4197 for_each_node_state(nid
, N_MEMORY
)
4198 seq_printf(m
, " N%d=%lu", nid
,
4199 mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4200 stat
->lru_mask
, true));
4206 #endif /* CONFIG_NUMA */
4208 static const unsigned int memcg1_stats
[] = {
4211 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4218 WORKINGSET_REFAULT_ANON
,
4219 WORKINGSET_REFAULT_FILE
,
4226 static const char *const memcg1_stat_names
[] = {
4229 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4236 "workingset_refault_anon",
4237 "workingset_refault_file",
4244 /* Universal VM events cgroup1 shows, original sort order */
4245 static const unsigned int memcg1_events
[] = {
4252 static void memcg1_stat_format(struct mem_cgroup
*memcg
, struct seq_buf
*s
)
4254 unsigned long memory
, memsw
;
4255 struct mem_cgroup
*mi
;
4258 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names
) != ARRAY_SIZE(memcg1_stats
));
4260 mem_cgroup_flush_stats();
4262 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4265 nr
= memcg_page_state_local_output(memcg
, memcg1_stats
[i
]);
4266 seq_buf_printf(s
, "%s %lu\n", memcg1_stat_names
[i
], nr
);
4269 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4270 seq_buf_printf(s
, "%s %lu\n", vm_event_name(memcg1_events
[i
]),
4271 memcg_events_local(memcg
, memcg1_events
[i
]));
4273 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4274 seq_buf_printf(s
, "%s %lu\n", lru_list_name(i
),
4275 memcg_page_state_local(memcg
, NR_LRU_BASE
+ i
) *
4278 /* Hierarchical information */
4279 memory
= memsw
= PAGE_COUNTER_MAX
;
4280 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
4281 memory
= min(memory
, READ_ONCE(mi
->memory
.max
));
4282 memsw
= min(memsw
, READ_ONCE(mi
->memsw
.max
));
4284 seq_buf_printf(s
, "hierarchical_memory_limit %llu\n",
4285 (u64
)memory
* PAGE_SIZE
);
4286 seq_buf_printf(s
, "hierarchical_memsw_limit %llu\n",
4287 (u64
)memsw
* PAGE_SIZE
);
4289 for (i
= 0; i
< ARRAY_SIZE(memcg1_stats
); i
++) {
4292 nr
= memcg_page_state_output(memcg
, memcg1_stats
[i
]);
4293 seq_buf_printf(s
, "total_%s %llu\n", memcg1_stat_names
[i
],
4297 for (i
= 0; i
< ARRAY_SIZE(memcg1_events
); i
++)
4298 seq_buf_printf(s
, "total_%s %llu\n",
4299 vm_event_name(memcg1_events
[i
]),
4300 (u64
)memcg_events(memcg
, memcg1_events
[i
]));
4302 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4303 seq_buf_printf(s
, "total_%s %llu\n", lru_list_name(i
),
4304 (u64
)memcg_page_state(memcg
, NR_LRU_BASE
+ i
) *
4307 #ifdef CONFIG_DEBUG_VM
4310 struct mem_cgroup_per_node
*mz
;
4311 unsigned long anon_cost
= 0;
4312 unsigned long file_cost
= 0;
4314 for_each_online_pgdat(pgdat
) {
4315 mz
= memcg
->nodeinfo
[pgdat
->node_id
];
4317 anon_cost
+= mz
->lruvec
.anon_cost
;
4318 file_cost
+= mz
->lruvec
.file_cost
;
4320 seq_buf_printf(s
, "anon_cost %lu\n", anon_cost
);
4321 seq_buf_printf(s
, "file_cost %lu\n", file_cost
);
4326 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4329 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4331 return mem_cgroup_swappiness(memcg
);
4334 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4335 struct cftype
*cft
, u64 val
)
4337 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4342 if (!mem_cgroup_is_root(memcg
))
4343 WRITE_ONCE(memcg
->swappiness
, val
);
4345 WRITE_ONCE(vm_swappiness
, val
);
4350 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4352 struct mem_cgroup_threshold_ary
*t
;
4353 unsigned long usage
;
4358 t
= rcu_dereference(memcg
->thresholds
.primary
);
4360 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4365 usage
= mem_cgroup_usage(memcg
, swap
);
4368 * current_threshold points to threshold just below or equal to usage.
4369 * If it's not true, a threshold was crossed after last
4370 * call of __mem_cgroup_threshold().
4372 i
= t
->current_threshold
;
4375 * Iterate backward over array of thresholds starting from
4376 * current_threshold and check if a threshold is crossed.
4377 * If none of thresholds below usage is crossed, we read
4378 * only one element of the array here.
4380 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4381 eventfd_signal(t
->entries
[i
].eventfd
);
4383 /* i = current_threshold + 1 */
4387 * Iterate forward over array of thresholds starting from
4388 * current_threshold+1 and check if a threshold is crossed.
4389 * If none of thresholds above usage is crossed, we read
4390 * only one element of the array here.
4392 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4393 eventfd_signal(t
->entries
[i
].eventfd
);
4395 /* Update current_threshold */
4396 t
->current_threshold
= i
- 1;
4401 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4404 __mem_cgroup_threshold(memcg
, false);
4405 if (do_memsw_account())
4406 __mem_cgroup_threshold(memcg
, true);
4408 memcg
= parent_mem_cgroup(memcg
);
4412 static int compare_thresholds(const void *a
, const void *b
)
4414 const struct mem_cgroup_threshold
*_a
= a
;
4415 const struct mem_cgroup_threshold
*_b
= b
;
4417 if (_a
->threshold
> _b
->threshold
)
4420 if (_a
->threshold
< _b
->threshold
)
4426 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4428 struct mem_cgroup_eventfd_list
*ev
;
4430 spin_lock(&memcg_oom_lock
);
4432 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4433 eventfd_signal(ev
->eventfd
);
4435 spin_unlock(&memcg_oom_lock
);
4439 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4441 struct mem_cgroup
*iter
;
4443 for_each_mem_cgroup_tree(iter
, memcg
)
4444 mem_cgroup_oom_notify_cb(iter
);
4447 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4448 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4450 struct mem_cgroup_thresholds
*thresholds
;
4451 struct mem_cgroup_threshold_ary
*new;
4452 unsigned long threshold
;
4453 unsigned long usage
;
4456 ret
= page_counter_memparse(args
, "-1", &threshold
);
4460 mutex_lock(&memcg
->thresholds_lock
);
4463 thresholds
= &memcg
->thresholds
;
4464 usage
= mem_cgroup_usage(memcg
, false);
4465 } else if (type
== _MEMSWAP
) {
4466 thresholds
= &memcg
->memsw_thresholds
;
4467 usage
= mem_cgroup_usage(memcg
, true);
4471 /* Check if a threshold crossed before adding a new one */
4472 if (thresholds
->primary
)
4473 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4475 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4477 /* Allocate memory for new array of thresholds */
4478 new = kmalloc(struct_size(new, entries
, size
), GFP_KERNEL
);
4485 /* Copy thresholds (if any) to new array */
4486 if (thresholds
->primary
)
4487 memcpy(new->entries
, thresholds
->primary
->entries
,
4488 flex_array_size(new, entries
, size
- 1));
4490 /* Add new threshold */
4491 new->entries
[size
- 1].eventfd
= eventfd
;
4492 new->entries
[size
- 1].threshold
= threshold
;
4494 /* Sort thresholds. Registering of new threshold isn't time-critical */
4495 sort(new->entries
, size
, sizeof(*new->entries
),
4496 compare_thresholds
, NULL
);
4498 /* Find current threshold */
4499 new->current_threshold
= -1;
4500 for (i
= 0; i
< size
; i
++) {
4501 if (new->entries
[i
].threshold
<= usage
) {
4503 * new->current_threshold will not be used until
4504 * rcu_assign_pointer(), so it's safe to increment
4507 ++new->current_threshold
;
4512 /* Free old spare buffer and save old primary buffer as spare */
4513 kfree(thresholds
->spare
);
4514 thresholds
->spare
= thresholds
->primary
;
4516 rcu_assign_pointer(thresholds
->primary
, new);
4518 /* To be sure that nobody uses thresholds */
4522 mutex_unlock(&memcg
->thresholds_lock
);
4527 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4528 struct eventfd_ctx
*eventfd
, const char *args
)
4530 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4533 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4534 struct eventfd_ctx
*eventfd
, const char *args
)
4536 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4539 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4540 struct eventfd_ctx
*eventfd
, enum res_type type
)
4542 struct mem_cgroup_thresholds
*thresholds
;
4543 struct mem_cgroup_threshold_ary
*new;
4544 unsigned long usage
;
4545 int i
, j
, size
, entries
;
4547 mutex_lock(&memcg
->thresholds_lock
);
4550 thresholds
= &memcg
->thresholds
;
4551 usage
= mem_cgroup_usage(memcg
, false);
4552 } else if (type
== _MEMSWAP
) {
4553 thresholds
= &memcg
->memsw_thresholds
;
4554 usage
= mem_cgroup_usage(memcg
, true);
4558 if (!thresholds
->primary
)
4561 /* Check if a threshold crossed before removing */
4562 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4564 /* Calculate new number of threshold */
4566 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4567 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4573 new = thresholds
->spare
;
4575 /* If no items related to eventfd have been cleared, nothing to do */
4579 /* Set thresholds array to NULL if we don't have thresholds */
4588 /* Copy thresholds and find current threshold */
4589 new->current_threshold
= -1;
4590 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4591 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4594 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4595 if (new->entries
[j
].threshold
<= usage
) {
4597 * new->current_threshold will not be used
4598 * until rcu_assign_pointer(), so it's safe to increment
4601 ++new->current_threshold
;
4607 /* Swap primary and spare array */
4608 thresholds
->spare
= thresholds
->primary
;
4610 rcu_assign_pointer(thresholds
->primary
, new);
4612 /* To be sure that nobody uses thresholds */
4615 /* If all events are unregistered, free the spare array */
4617 kfree(thresholds
->spare
);
4618 thresholds
->spare
= NULL
;
4621 mutex_unlock(&memcg
->thresholds_lock
);
4624 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4625 struct eventfd_ctx
*eventfd
)
4627 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4630 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4631 struct eventfd_ctx
*eventfd
)
4633 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4636 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4637 struct eventfd_ctx
*eventfd
, const char *args
)
4639 struct mem_cgroup_eventfd_list
*event
;
4641 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4645 spin_lock(&memcg_oom_lock
);
4647 event
->eventfd
= eventfd
;
4648 list_add(&event
->list
, &memcg
->oom_notify
);
4650 /* already in OOM ? */
4651 if (memcg
->under_oom
)
4652 eventfd_signal(eventfd
);
4653 spin_unlock(&memcg_oom_lock
);
4658 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4659 struct eventfd_ctx
*eventfd
)
4661 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4663 spin_lock(&memcg_oom_lock
);
4665 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4666 if (ev
->eventfd
== eventfd
) {
4667 list_del(&ev
->list
);
4672 spin_unlock(&memcg_oom_lock
);
4675 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4677 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(sf
);
4679 seq_printf(sf
, "oom_kill_disable %d\n", READ_ONCE(memcg
->oom_kill_disable
));
4680 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
4681 seq_printf(sf
, "oom_kill %lu\n",
4682 atomic_long_read(&memcg
->memory_events
[MEMCG_OOM_KILL
]));
4686 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4687 struct cftype
*cft
, u64 val
)
4689 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4691 /* cannot set to root cgroup and only 0 and 1 are allowed */
4692 if (mem_cgroup_is_root(memcg
) || !((val
== 0) || (val
== 1)))
4695 WRITE_ONCE(memcg
->oom_kill_disable
, val
);
4697 memcg_oom_recover(memcg
);
4702 #ifdef CONFIG_CGROUP_WRITEBACK
4704 #include <trace/events/writeback.h>
4706 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4708 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4711 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4713 wb_domain_exit(&memcg
->cgwb_domain
);
4716 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4718 wb_domain_size_changed(&memcg
->cgwb_domain
);
4721 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4723 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4725 if (!memcg
->css
.parent
)
4728 return &memcg
->cgwb_domain
;
4732 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4733 * @wb: bdi_writeback in question
4734 * @pfilepages: out parameter for number of file pages
4735 * @pheadroom: out parameter for number of allocatable pages according to memcg
4736 * @pdirty: out parameter for number of dirty pages
4737 * @pwriteback: out parameter for number of pages under writeback
4739 * Determine the numbers of file, headroom, dirty, and writeback pages in
4740 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4741 * is a bit more involved.
4743 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4744 * headroom is calculated as the lowest headroom of itself and the
4745 * ancestors. Note that this doesn't consider the actual amount of
4746 * available memory in the system. The caller should further cap
4747 * *@pheadroom accordingly.
4749 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pfilepages
,
4750 unsigned long *pheadroom
, unsigned long *pdirty
,
4751 unsigned long *pwriteback
)
4753 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4754 struct mem_cgroup
*parent
;
4756 mem_cgroup_flush_stats();
4758 *pdirty
= memcg_page_state(memcg
, NR_FILE_DIRTY
);
4759 *pwriteback
= memcg_page_state(memcg
, NR_WRITEBACK
);
4760 *pfilepages
= memcg_page_state(memcg
, NR_INACTIVE_FILE
) +
4761 memcg_page_state(memcg
, NR_ACTIVE_FILE
);
4763 *pheadroom
= PAGE_COUNTER_MAX
;
4764 while ((parent
= parent_mem_cgroup(memcg
))) {
4765 unsigned long ceiling
= min(READ_ONCE(memcg
->memory
.max
),
4766 READ_ONCE(memcg
->memory
.high
));
4767 unsigned long used
= page_counter_read(&memcg
->memory
);
4769 *pheadroom
= min(*pheadroom
, ceiling
- min(ceiling
, used
));
4775 * Foreign dirty flushing
4777 * There's an inherent mismatch between memcg and writeback. The former
4778 * tracks ownership per-page while the latter per-inode. This was a
4779 * deliberate design decision because honoring per-page ownership in the
4780 * writeback path is complicated, may lead to higher CPU and IO overheads
4781 * and deemed unnecessary given that write-sharing an inode across
4782 * different cgroups isn't a common use-case.
4784 * Combined with inode majority-writer ownership switching, this works well
4785 * enough in most cases but there are some pathological cases. For
4786 * example, let's say there are two cgroups A and B which keep writing to
4787 * different but confined parts of the same inode. B owns the inode and
4788 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4789 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4790 * triggering background writeback. A will be slowed down without a way to
4791 * make writeback of the dirty pages happen.
4793 * Conditions like the above can lead to a cgroup getting repeatedly and
4794 * severely throttled after making some progress after each
4795 * dirty_expire_interval while the underlying IO device is almost
4798 * Solving this problem completely requires matching the ownership tracking
4799 * granularities between memcg and writeback in either direction. However,
4800 * the more egregious behaviors can be avoided by simply remembering the
4801 * most recent foreign dirtying events and initiating remote flushes on
4802 * them when local writeback isn't enough to keep the memory clean enough.
4804 * The following two functions implement such mechanism. When a foreign
4805 * page - a page whose memcg and writeback ownerships don't match - is
4806 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4807 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4808 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4809 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4810 * foreign bdi_writebacks which haven't expired. Both the numbers of
4811 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4812 * limited to MEMCG_CGWB_FRN_CNT.
4814 * The mechanism only remembers IDs and doesn't hold any object references.
4815 * As being wrong occasionally doesn't matter, updates and accesses to the
4816 * records are lockless and racy.
4818 void mem_cgroup_track_foreign_dirty_slowpath(struct folio
*folio
,
4819 struct bdi_writeback
*wb
)
4821 struct mem_cgroup
*memcg
= folio_memcg(folio
);
4822 struct memcg_cgwb_frn
*frn
;
4823 u64 now
= get_jiffies_64();
4824 u64 oldest_at
= now
;
4828 trace_track_foreign_dirty(folio
, wb
);
4831 * Pick the slot to use. If there is already a slot for @wb, keep
4832 * using it. If not replace the oldest one which isn't being
4835 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4836 frn
= &memcg
->cgwb_frn
[i
];
4837 if (frn
->bdi_id
== wb
->bdi
->id
&&
4838 frn
->memcg_id
== wb
->memcg_css
->id
)
4840 if (time_before64(frn
->at
, oldest_at
) &&
4841 atomic_read(&frn
->done
.cnt
) == 1) {
4843 oldest_at
= frn
->at
;
4847 if (i
< MEMCG_CGWB_FRN_CNT
) {
4849 * Re-using an existing one. Update timestamp lazily to
4850 * avoid making the cacheline hot. We want them to be
4851 * reasonably up-to-date and significantly shorter than
4852 * dirty_expire_interval as that's what expires the record.
4853 * Use the shorter of 1s and dirty_expire_interval / 8.
4855 unsigned long update_intv
=
4856 min_t(unsigned long, HZ
,
4857 msecs_to_jiffies(dirty_expire_interval
* 10) / 8);
4859 if (time_before64(frn
->at
, now
- update_intv
))
4861 } else if (oldest
>= 0) {
4862 /* replace the oldest free one */
4863 frn
= &memcg
->cgwb_frn
[oldest
];
4864 frn
->bdi_id
= wb
->bdi
->id
;
4865 frn
->memcg_id
= wb
->memcg_css
->id
;
4870 /* issue foreign writeback flushes for recorded foreign dirtying events */
4871 void mem_cgroup_flush_foreign(struct bdi_writeback
*wb
)
4873 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4874 unsigned long intv
= msecs_to_jiffies(dirty_expire_interval
* 10);
4875 u64 now
= jiffies_64
;
4878 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++) {
4879 struct memcg_cgwb_frn
*frn
= &memcg
->cgwb_frn
[i
];
4882 * If the record is older than dirty_expire_interval,
4883 * writeback on it has already started. No need to kick it
4884 * off again. Also, don't start a new one if there's
4885 * already one in flight.
4887 if (time_after64(frn
->at
, now
- intv
) &&
4888 atomic_read(&frn
->done
.cnt
) == 1) {
4890 trace_flush_foreign(wb
, frn
->bdi_id
, frn
->memcg_id
);
4891 cgroup_writeback_by_id(frn
->bdi_id
, frn
->memcg_id
,
4892 WB_REASON_FOREIGN_FLUSH
,
4898 #else /* CONFIG_CGROUP_WRITEBACK */
4900 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4905 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4909 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4913 #endif /* CONFIG_CGROUP_WRITEBACK */
4916 * DO NOT USE IN NEW FILES.
4918 * "cgroup.event_control" implementation.
4920 * This is way over-engineered. It tries to support fully configurable
4921 * events for each user. Such level of flexibility is completely
4922 * unnecessary especially in the light of the planned unified hierarchy.
4924 * Please deprecate this and replace with something simpler if at all
4929 * Unregister event and free resources.
4931 * Gets called from workqueue.
4933 static void memcg_event_remove(struct work_struct
*work
)
4935 struct mem_cgroup_event
*event
=
4936 container_of(work
, struct mem_cgroup_event
, remove
);
4937 struct mem_cgroup
*memcg
= event
->memcg
;
4939 remove_wait_queue(event
->wqh
, &event
->wait
);
4941 event
->unregister_event(memcg
, event
->eventfd
);
4943 /* Notify userspace the event is going away. */
4944 eventfd_signal(event
->eventfd
);
4946 eventfd_ctx_put(event
->eventfd
);
4948 css_put(&memcg
->css
);
4952 * Gets called on EPOLLHUP on eventfd when user closes it.
4954 * Called with wqh->lock held and interrupts disabled.
4956 static int memcg_event_wake(wait_queue_entry_t
*wait
, unsigned mode
,
4957 int sync
, void *key
)
4959 struct mem_cgroup_event
*event
=
4960 container_of(wait
, struct mem_cgroup_event
, wait
);
4961 struct mem_cgroup
*memcg
= event
->memcg
;
4962 __poll_t flags
= key_to_poll(key
);
4964 if (flags
& EPOLLHUP
) {
4966 * If the event has been detached at cgroup removal, we
4967 * can simply return knowing the other side will cleanup
4970 * We can't race against event freeing since the other
4971 * side will require wqh->lock via remove_wait_queue(),
4974 spin_lock(&memcg
->event_list_lock
);
4975 if (!list_empty(&event
->list
)) {
4976 list_del_init(&event
->list
);
4978 * We are in atomic context, but cgroup_event_remove()
4979 * may sleep, so we have to call it in workqueue.
4981 schedule_work(&event
->remove
);
4983 spin_unlock(&memcg
->event_list_lock
);
4989 static void memcg_event_ptable_queue_proc(struct file
*file
,
4990 wait_queue_head_t
*wqh
, poll_table
*pt
)
4992 struct mem_cgroup_event
*event
=
4993 container_of(pt
, struct mem_cgroup_event
, pt
);
4996 add_wait_queue(wqh
, &event
->wait
);
5000 * DO NOT USE IN NEW FILES.
5002 * Parse input and register new cgroup event handler.
5004 * Input must be in format '<event_fd> <control_fd> <args>'.
5005 * Interpretation of args is defined by control file implementation.
5007 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
5008 char *buf
, size_t nbytes
, loff_t off
)
5010 struct cgroup_subsys_state
*css
= of_css(of
);
5011 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5012 struct mem_cgroup_event
*event
;
5013 struct cgroup_subsys_state
*cfile_css
;
5014 unsigned int efd
, cfd
;
5017 struct dentry
*cdentry
;
5022 if (IS_ENABLED(CONFIG_PREEMPT_RT
))
5025 buf
= strstrip(buf
);
5027 efd
= simple_strtoul(buf
, &endp
, 10);
5032 cfd
= simple_strtoul(buf
, &endp
, 10);
5033 if ((*endp
!= ' ') && (*endp
!= '\0'))
5037 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5041 event
->memcg
= memcg
;
5042 INIT_LIST_HEAD(&event
->list
);
5043 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5044 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5045 INIT_WORK(&event
->remove
, memcg_event_remove
);
5053 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5054 if (IS_ERR(event
->eventfd
)) {
5055 ret
= PTR_ERR(event
->eventfd
);
5062 goto out_put_eventfd
;
5065 /* the process need read permission on control file */
5066 /* AV: shouldn't we check that it's been opened for read instead? */
5067 ret
= file_permission(cfile
.file
, MAY_READ
);
5072 * The control file must be a regular cgroup1 file. As a regular cgroup
5073 * file can't be renamed, it's safe to access its name afterwards.
5075 cdentry
= cfile
.file
->f_path
.dentry
;
5076 if (cdentry
->d_sb
->s_type
!= &cgroup_fs_type
|| !d_is_reg(cdentry
)) {
5082 * Determine the event callbacks and set them in @event. This used
5083 * to be done via struct cftype but cgroup core no longer knows
5084 * about these events. The following is crude but the whole thing
5085 * is for compatibility anyway.
5087 * DO NOT ADD NEW FILES.
5089 name
= cdentry
->d_name
.name
;
5091 if (!strcmp(name
, "memory.usage_in_bytes")) {
5092 event
->register_event
= mem_cgroup_usage_register_event
;
5093 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5094 } else if (!strcmp(name
, "memory.oom_control")) {
5095 event
->register_event
= mem_cgroup_oom_register_event
;
5096 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5097 } else if (!strcmp(name
, "memory.pressure_level")) {
5098 event
->register_event
= vmpressure_register_event
;
5099 event
->unregister_event
= vmpressure_unregister_event
;
5100 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5101 event
->register_event
= memsw_cgroup_usage_register_event
;
5102 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5109 * Verify @cfile should belong to @css. Also, remaining events are
5110 * automatically removed on cgroup destruction but the removal is
5111 * asynchronous, so take an extra ref on @css.
5113 cfile_css
= css_tryget_online_from_dir(cdentry
->d_parent
,
5114 &memory_cgrp_subsys
);
5116 if (IS_ERR(cfile_css
))
5118 if (cfile_css
!= css
) {
5123 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
5127 vfs_poll(efile
.file
, &event
->pt
);
5129 spin_lock_irq(&memcg
->event_list_lock
);
5130 list_add(&event
->list
, &memcg
->event_list
);
5131 spin_unlock_irq(&memcg
->event_list_lock
);
5143 eventfd_ctx_put(event
->eventfd
);
5152 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5153 static int mem_cgroup_slab_show(struct seq_file
*m
, void *p
)
5157 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
5163 static int memory_stat_show(struct seq_file
*m
, void *v
);
5165 static struct cftype mem_cgroup_legacy_files
[] = {
5167 .name
= "usage_in_bytes",
5168 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5169 .read_u64
= mem_cgroup_read_u64
,
5172 .name
= "max_usage_in_bytes",
5173 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5174 .write
= mem_cgroup_reset
,
5175 .read_u64
= mem_cgroup_read_u64
,
5178 .name
= "limit_in_bytes",
5179 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5180 .write
= mem_cgroup_write
,
5181 .read_u64
= mem_cgroup_read_u64
,
5184 .name
= "soft_limit_in_bytes",
5185 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5186 .write
= mem_cgroup_write
,
5187 .read_u64
= mem_cgroup_read_u64
,
5191 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5192 .write
= mem_cgroup_reset
,
5193 .read_u64
= mem_cgroup_read_u64
,
5197 .seq_show
= memory_stat_show
,
5200 .name
= "force_empty",
5201 .write
= mem_cgroup_force_empty_write
,
5204 .name
= "use_hierarchy",
5205 .write_u64
= mem_cgroup_hierarchy_write
,
5206 .read_u64
= mem_cgroup_hierarchy_read
,
5209 .name
= "cgroup.event_control", /* XXX: for compat */
5210 .write
= memcg_write_event_control
,
5211 .flags
= CFTYPE_NO_PREFIX
| CFTYPE_WORLD_WRITABLE
,
5214 .name
= "swappiness",
5215 .read_u64
= mem_cgroup_swappiness_read
,
5216 .write_u64
= mem_cgroup_swappiness_write
,
5219 .name
= "move_charge_at_immigrate",
5220 .read_u64
= mem_cgroup_move_charge_read
,
5221 .write_u64
= mem_cgroup_move_charge_write
,
5224 .name
= "oom_control",
5225 .seq_show
= mem_cgroup_oom_control_read
,
5226 .write_u64
= mem_cgroup_oom_control_write
,
5229 .name
= "pressure_level",
5230 .seq_show
= mem_cgroup_dummy_seq_show
,
5234 .name
= "numa_stat",
5235 .seq_show
= memcg_numa_stat_show
,
5239 .name
= "kmem.limit_in_bytes",
5240 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5241 .write
= mem_cgroup_write
,
5242 .read_u64
= mem_cgroup_read_u64
,
5245 .name
= "kmem.usage_in_bytes",
5246 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5247 .read_u64
= mem_cgroup_read_u64
,
5250 .name
= "kmem.failcnt",
5251 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5252 .write
= mem_cgroup_reset
,
5253 .read_u64
= mem_cgroup_read_u64
,
5256 .name
= "kmem.max_usage_in_bytes",
5257 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5258 .write
= mem_cgroup_reset
,
5259 .read_u64
= mem_cgroup_read_u64
,
5261 #if defined(CONFIG_MEMCG_KMEM) && \
5262 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5264 .name
= "kmem.slabinfo",
5265 .seq_show
= mem_cgroup_slab_show
,
5269 .name
= "kmem.tcp.limit_in_bytes",
5270 .private = MEMFILE_PRIVATE(_TCP
, RES_LIMIT
),
5271 .write
= mem_cgroup_write
,
5272 .read_u64
= mem_cgroup_read_u64
,
5275 .name
= "kmem.tcp.usage_in_bytes",
5276 .private = MEMFILE_PRIVATE(_TCP
, RES_USAGE
),
5277 .read_u64
= mem_cgroup_read_u64
,
5280 .name
= "kmem.tcp.failcnt",
5281 .private = MEMFILE_PRIVATE(_TCP
, RES_FAILCNT
),
5282 .write
= mem_cgroup_reset
,
5283 .read_u64
= mem_cgroup_read_u64
,
5286 .name
= "kmem.tcp.max_usage_in_bytes",
5287 .private = MEMFILE_PRIVATE(_TCP
, RES_MAX_USAGE
),
5288 .write
= mem_cgroup_reset
,
5289 .read_u64
= mem_cgroup_read_u64
,
5291 { }, /* terminate */
5295 * Private memory cgroup IDR
5297 * Swap-out records and page cache shadow entries need to store memcg
5298 * references in constrained space, so we maintain an ID space that is
5299 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5300 * memory-controlled cgroups to 64k.
5302 * However, there usually are many references to the offline CSS after
5303 * the cgroup has been destroyed, such as page cache or reclaimable
5304 * slab objects, that don't need to hang on to the ID. We want to keep
5305 * those dead CSS from occupying IDs, or we might quickly exhaust the
5306 * relatively small ID space and prevent the creation of new cgroups
5307 * even when there are much fewer than 64k cgroups - possibly none.
5309 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5310 * be freed and recycled when it's no longer needed, which is usually
5311 * when the CSS is offlined.
5313 * The only exception to that are records of swapped out tmpfs/shmem
5314 * pages that need to be attributed to live ancestors on swapin. But
5315 * those references are manageable from userspace.
5318 #define MEM_CGROUP_ID_MAX ((1UL << MEM_CGROUP_ID_SHIFT) - 1)
5319 static DEFINE_IDR(mem_cgroup_idr
);
5321 static void mem_cgroup_id_remove(struct mem_cgroup
*memcg
)
5323 if (memcg
->id
.id
> 0) {
5324 idr_remove(&mem_cgroup_idr
, memcg
->id
.id
);
5329 static void __maybe_unused
mem_cgroup_id_get_many(struct mem_cgroup
*memcg
,
5332 refcount_add(n
, &memcg
->id
.ref
);
5335 static void mem_cgroup_id_put_many(struct mem_cgroup
*memcg
, unsigned int n
)
5337 if (refcount_sub_and_test(n
, &memcg
->id
.ref
)) {
5338 mem_cgroup_id_remove(memcg
);
5340 /* Memcg ID pins CSS */
5341 css_put(&memcg
->css
);
5345 static inline void mem_cgroup_id_put(struct mem_cgroup
*memcg
)
5347 mem_cgroup_id_put_many(memcg
, 1);
5351 * mem_cgroup_from_id - look up a memcg from a memcg id
5352 * @id: the memcg id to look up
5354 * Caller must hold rcu_read_lock().
5356 struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
5358 WARN_ON_ONCE(!rcu_read_lock_held());
5359 return idr_find(&mem_cgroup_idr
, id
);
5362 #ifdef CONFIG_SHRINKER_DEBUG
5363 struct mem_cgroup
*mem_cgroup_get_from_ino(unsigned long ino
)
5365 struct cgroup
*cgrp
;
5366 struct cgroup_subsys_state
*css
;
5367 struct mem_cgroup
*memcg
;
5369 cgrp
= cgroup_get_from_id(ino
);
5371 return ERR_CAST(cgrp
);
5373 css
= cgroup_get_e_css(cgrp
, &memory_cgrp_subsys
);
5375 memcg
= container_of(css
, struct mem_cgroup
, css
);
5377 memcg
= ERR_PTR(-ENOENT
);
5385 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5387 struct mem_cgroup_per_node
*pn
;
5389 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, node
);
5393 pn
->lruvec_stats_percpu
= alloc_percpu_gfp(struct lruvec_stats_percpu
,
5394 GFP_KERNEL_ACCOUNT
);
5395 if (!pn
->lruvec_stats_percpu
) {
5400 lruvec_init(&pn
->lruvec
);
5403 memcg
->nodeinfo
[node
] = pn
;
5407 static void free_mem_cgroup_per_node_info(struct mem_cgroup
*memcg
, int node
)
5409 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[node
];
5414 free_percpu(pn
->lruvec_stats_percpu
);
5418 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5422 if (memcg
->orig_objcg
)
5423 obj_cgroup_put(memcg
->orig_objcg
);
5426 free_mem_cgroup_per_node_info(memcg
, node
);
5427 kfree(memcg
->vmstats
);
5428 free_percpu(memcg
->vmstats_percpu
);
5432 static void mem_cgroup_free(struct mem_cgroup
*memcg
)
5434 lru_gen_exit_memcg(memcg
);
5435 memcg_wb_domain_exit(memcg
);
5436 __mem_cgroup_free(memcg
);
5439 static struct mem_cgroup
*mem_cgroup_alloc(void)
5441 struct mem_cgroup
*memcg
;
5443 int __maybe_unused i
;
5444 long error
= -ENOMEM
;
5446 memcg
= kzalloc(struct_size(memcg
, nodeinfo
, nr_node_ids
), GFP_KERNEL
);
5448 return ERR_PTR(error
);
5450 memcg
->id
.id
= idr_alloc(&mem_cgroup_idr
, NULL
,
5451 1, MEM_CGROUP_ID_MAX
+ 1, GFP_KERNEL
);
5452 if (memcg
->id
.id
< 0) {
5453 error
= memcg
->id
.id
;
5457 memcg
->vmstats
= kzalloc(sizeof(struct memcg_vmstats
), GFP_KERNEL
);
5458 if (!memcg
->vmstats
)
5461 memcg
->vmstats_percpu
= alloc_percpu_gfp(struct memcg_vmstats_percpu
,
5462 GFP_KERNEL_ACCOUNT
);
5463 if (!memcg
->vmstats_percpu
)
5467 if (alloc_mem_cgroup_per_node_info(memcg
, node
))
5470 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
5473 INIT_WORK(&memcg
->high_work
, high_work_func
);
5474 INIT_LIST_HEAD(&memcg
->oom_notify
);
5475 mutex_init(&memcg
->thresholds_lock
);
5476 spin_lock_init(&memcg
->move_lock
);
5477 vmpressure_init(&memcg
->vmpressure
);
5478 INIT_LIST_HEAD(&memcg
->event_list
);
5479 spin_lock_init(&memcg
->event_list_lock
);
5480 memcg
->socket_pressure
= jiffies
;
5481 #ifdef CONFIG_MEMCG_KMEM
5482 memcg
->kmemcg_id
= -1;
5483 INIT_LIST_HEAD(&memcg
->objcg_list
);
5485 #ifdef CONFIG_CGROUP_WRITEBACK
5486 INIT_LIST_HEAD(&memcg
->cgwb_list
);
5487 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5488 memcg
->cgwb_frn
[i
].done
=
5489 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq
);
5491 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5492 spin_lock_init(&memcg
->deferred_split_queue
.split_queue_lock
);
5493 INIT_LIST_HEAD(&memcg
->deferred_split_queue
.split_queue
);
5494 memcg
->deferred_split_queue
.split_queue_len
= 0;
5496 lru_gen_init_memcg(memcg
);
5499 mem_cgroup_id_remove(memcg
);
5500 __mem_cgroup_free(memcg
);
5501 return ERR_PTR(error
);
5504 static struct cgroup_subsys_state
* __ref
5505 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5507 struct mem_cgroup
*parent
= mem_cgroup_from_css(parent_css
);
5508 struct mem_cgroup
*memcg
, *old_memcg
;
5510 old_memcg
= set_active_memcg(parent
);
5511 memcg
= mem_cgroup_alloc();
5512 set_active_memcg(old_memcg
);
5514 return ERR_CAST(memcg
);
5516 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5517 WRITE_ONCE(memcg
->soft_limit
, PAGE_COUNTER_MAX
);
5518 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5519 memcg
->zswap_max
= PAGE_COUNTER_MAX
;
5521 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5523 WRITE_ONCE(memcg
->swappiness
, mem_cgroup_swappiness(parent
));
5524 WRITE_ONCE(memcg
->oom_kill_disable
, READ_ONCE(parent
->oom_kill_disable
));
5526 page_counter_init(&memcg
->memory
, &parent
->memory
);
5527 page_counter_init(&memcg
->swap
, &parent
->swap
);
5528 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
5529 page_counter_init(&memcg
->tcpmem
, &parent
->tcpmem
);
5531 init_memcg_events();
5532 page_counter_init(&memcg
->memory
, NULL
);
5533 page_counter_init(&memcg
->swap
, NULL
);
5534 page_counter_init(&memcg
->kmem
, NULL
);
5535 page_counter_init(&memcg
->tcpmem
, NULL
);
5537 root_mem_cgroup
= memcg
;
5541 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5542 static_branch_inc(&memcg_sockets_enabled_key
);
5544 #if defined(CONFIG_MEMCG_KMEM)
5545 if (!cgroup_memory_nobpf
)
5546 static_branch_inc(&memcg_bpf_enabled_key
);
5552 static int mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5554 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5556 if (memcg_online_kmem(memcg
))
5560 * A memcg must be visible for expand_shrinker_info()
5561 * by the time the maps are allocated. So, we allocate maps
5562 * here, when for_each_mem_cgroup() can't skip it.
5564 if (alloc_shrinker_info(memcg
))
5567 if (unlikely(mem_cgroup_is_root(memcg
)))
5568 queue_delayed_work(system_unbound_wq
, &stats_flush_dwork
,
5570 lru_gen_online_memcg(memcg
);
5572 /* Online state pins memcg ID, memcg ID pins CSS */
5573 refcount_set(&memcg
->id
.ref
, 1);
5577 * Ensure mem_cgroup_from_id() works once we're fully online.
5579 * We could do this earlier and require callers to filter with
5580 * css_tryget_online(). But right now there are no users that
5581 * need earlier access, and the workingset code relies on the
5582 * cgroup tree linkage (mem_cgroup_get_nr_swap_pages()). So
5583 * publish it here at the end of onlining. This matches the
5584 * regular ID destruction during offlining.
5586 idr_replace(&mem_cgroup_idr
, memcg
, memcg
->id
.id
);
5590 memcg_offline_kmem(memcg
);
5592 mem_cgroup_id_remove(memcg
);
5596 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5598 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5599 struct mem_cgroup_event
*event
, *tmp
;
5602 * Unregister events and notify userspace.
5603 * Notify userspace about cgroup removing only after rmdir of cgroup
5604 * directory to avoid race between userspace and kernelspace.
5606 spin_lock_irq(&memcg
->event_list_lock
);
5607 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5608 list_del_init(&event
->list
);
5609 schedule_work(&event
->remove
);
5611 spin_unlock_irq(&memcg
->event_list_lock
);
5613 page_counter_set_min(&memcg
->memory
, 0);
5614 page_counter_set_low(&memcg
->memory
, 0);
5616 memcg_offline_kmem(memcg
);
5617 reparent_shrinker_deferred(memcg
);
5618 wb_memcg_offline(memcg
);
5619 lru_gen_offline_memcg(memcg
);
5621 drain_all_stock(memcg
);
5623 mem_cgroup_id_put(memcg
);
5626 static void mem_cgroup_css_released(struct cgroup_subsys_state
*css
)
5628 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5630 invalidate_reclaim_iterators(memcg
);
5631 lru_gen_release_memcg(memcg
);
5634 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5636 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5637 int __maybe_unused i
;
5639 #ifdef CONFIG_CGROUP_WRITEBACK
5640 for (i
= 0; i
< MEMCG_CGWB_FRN_CNT
; i
++)
5641 wb_wait_for_completion(&memcg
->cgwb_frn
[i
].done
);
5643 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !cgroup_memory_nosocket
)
5644 static_branch_dec(&memcg_sockets_enabled_key
);
5646 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && memcg
->tcpmem_active
)
5647 static_branch_dec(&memcg_sockets_enabled_key
);
5649 #if defined(CONFIG_MEMCG_KMEM)
5650 if (!cgroup_memory_nobpf
)
5651 static_branch_dec(&memcg_bpf_enabled_key
);
5654 vmpressure_cleanup(&memcg
->vmpressure
);
5655 cancel_work_sync(&memcg
->high_work
);
5656 mem_cgroup_remove_from_trees(memcg
);
5657 free_shrinker_info(memcg
);
5658 mem_cgroup_free(memcg
);
5662 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5663 * @css: the target css
5665 * Reset the states of the mem_cgroup associated with @css. This is
5666 * invoked when the userland requests disabling on the default hierarchy
5667 * but the memcg is pinned through dependency. The memcg should stop
5668 * applying policies and should revert to the vanilla state as it may be
5669 * made visible again.
5671 * The current implementation only resets the essential configurations.
5672 * This needs to be expanded to cover all the visible parts.
5674 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5676 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5678 page_counter_set_max(&memcg
->memory
, PAGE_COUNTER_MAX
);
5679 page_counter_set_max(&memcg
->swap
, PAGE_COUNTER_MAX
);
5680 page_counter_set_max(&memcg
->kmem
, PAGE_COUNTER_MAX
);
5681 page_counter_set_max(&memcg
->tcpmem
, PAGE_COUNTER_MAX
);
5682 page_counter_set_min(&memcg
->memory
, 0);
5683 page_counter_set_low(&memcg
->memory
, 0);
5684 page_counter_set_high(&memcg
->memory
, PAGE_COUNTER_MAX
);
5685 WRITE_ONCE(memcg
->soft_limit
, PAGE_COUNTER_MAX
);
5686 page_counter_set_high(&memcg
->swap
, PAGE_COUNTER_MAX
);
5687 memcg_wb_domain_size_changed(memcg
);
5690 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state
*css
, int cpu
)
5692 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5693 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5694 struct memcg_vmstats_percpu
*statc
;
5695 long delta
, delta_cpu
, v
;
5698 statc
= per_cpu_ptr(memcg
->vmstats_percpu
, cpu
);
5700 for (i
= 0; i
< MEMCG_NR_STAT
; i
++) {
5702 * Collect the aggregated propagation counts of groups
5703 * below us. We're in a per-cpu loop here and this is
5704 * a global counter, so the first cycle will get them.
5706 delta
= memcg
->vmstats
->state_pending
[i
];
5708 memcg
->vmstats
->state_pending
[i
] = 0;
5710 /* Add CPU changes on this level since the last flush */
5712 v
= READ_ONCE(statc
->state
[i
]);
5713 if (v
!= statc
->state_prev
[i
]) {
5714 delta_cpu
= v
- statc
->state_prev
[i
];
5716 statc
->state_prev
[i
] = v
;
5719 /* Aggregate counts on this level and propagate upwards */
5721 memcg
->vmstats
->state_local
[i
] += delta_cpu
;
5724 memcg
->vmstats
->state
[i
] += delta
;
5726 parent
->vmstats
->state_pending
[i
] += delta
;
5730 for (i
= 0; i
< NR_MEMCG_EVENTS
; i
++) {
5731 delta
= memcg
->vmstats
->events_pending
[i
];
5733 memcg
->vmstats
->events_pending
[i
] = 0;
5736 v
= READ_ONCE(statc
->events
[i
]);
5737 if (v
!= statc
->events_prev
[i
]) {
5738 delta_cpu
= v
- statc
->events_prev
[i
];
5740 statc
->events_prev
[i
] = v
;
5744 memcg
->vmstats
->events_local
[i
] += delta_cpu
;
5747 memcg
->vmstats
->events
[i
] += delta
;
5749 parent
->vmstats
->events_pending
[i
] += delta
;
5753 for_each_node_state(nid
, N_MEMORY
) {
5754 struct mem_cgroup_per_node
*pn
= memcg
->nodeinfo
[nid
];
5755 struct mem_cgroup_per_node
*ppn
= NULL
;
5756 struct lruvec_stats_percpu
*lstatc
;
5759 ppn
= parent
->nodeinfo
[nid
];
5761 lstatc
= per_cpu_ptr(pn
->lruvec_stats_percpu
, cpu
);
5763 for (i
= 0; i
< NR_VM_NODE_STAT_ITEMS
; i
++) {
5764 delta
= pn
->lruvec_stats
.state_pending
[i
];
5766 pn
->lruvec_stats
.state_pending
[i
] = 0;
5769 v
= READ_ONCE(lstatc
->state
[i
]);
5770 if (v
!= lstatc
->state_prev
[i
]) {
5771 delta_cpu
= v
- lstatc
->state_prev
[i
];
5773 lstatc
->state_prev
[i
] = v
;
5777 pn
->lruvec_stats
.state_local
[i
] += delta_cpu
;
5780 pn
->lruvec_stats
.state
[i
] += delta
;
5782 ppn
->lruvec_stats
.state_pending
[i
] += delta
;
5789 /* Handlers for move charge at task migration. */
5790 static int mem_cgroup_do_precharge(unsigned long count
)
5794 /* Try a single bulk charge without reclaim first, kswapd may wake */
5795 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_DIRECT_RECLAIM
, count
);
5797 mc
.precharge
+= count
;
5801 /* Try charges one by one with reclaim, but do not retry */
5803 ret
= try_charge(mc
.to
, GFP_KERNEL
| __GFP_NORETRY
, 1);
5817 enum mc_target_type
{
5824 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5825 unsigned long addr
, pte_t ptent
)
5827 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5831 if (PageAnon(page
)) {
5832 if (!(mc
.flags
& MOVE_ANON
))
5835 if (!(mc
.flags
& MOVE_FILE
))
5843 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5844 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5845 pte_t ptent
, swp_entry_t
*entry
)
5847 struct page
*page
= NULL
;
5848 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5850 if (!(mc
.flags
& MOVE_ANON
))
5854 * Handle device private pages that are not accessible by the CPU, but
5855 * stored as special swap entries in the page table.
5857 if (is_device_private_entry(ent
)) {
5858 page
= pfn_swap_entry_to_page(ent
);
5859 if (!get_page_unless_zero(page
))
5864 if (non_swap_entry(ent
))
5868 * Because swap_cache_get_folio() updates some statistics counter,
5869 * we call find_get_page() with swapper_space directly.
5871 page
= find_get_page(swap_address_space(ent
), swp_offset(ent
));
5872 entry
->val
= ent
.val
;
5877 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5878 pte_t ptent
, swp_entry_t
*entry
)
5884 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5885 unsigned long addr
, pte_t ptent
)
5887 unsigned long index
;
5888 struct folio
*folio
;
5890 if (!vma
->vm_file
) /* anonymous vma */
5892 if (!(mc
.flags
& MOVE_FILE
))
5895 /* folio is moved even if it's not RSS of this task(page-faulted). */
5896 /* shmem/tmpfs may report page out on swap: account for that too. */
5897 index
= linear_page_index(vma
, addr
);
5898 folio
= filemap_get_incore_folio(vma
->vm_file
->f_mapping
, index
);
5901 return folio_file_page(folio
, index
);
5905 * mem_cgroup_move_account - move account of the page
5907 * @compound: charge the page as compound or small page
5908 * @from: mem_cgroup which the page is moved from.
5909 * @to: mem_cgroup which the page is moved to. @from != @to.
5911 * The page must be locked and not on the LRU.
5913 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5916 static int mem_cgroup_move_account(struct page
*page
,
5918 struct mem_cgroup
*from
,
5919 struct mem_cgroup
*to
)
5921 struct folio
*folio
= page_folio(page
);
5922 struct lruvec
*from_vec
, *to_vec
;
5923 struct pglist_data
*pgdat
;
5924 unsigned int nr_pages
= compound
? folio_nr_pages(folio
) : 1;
5927 VM_BUG_ON(from
== to
);
5928 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
5929 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
5930 VM_BUG_ON(compound
&& !folio_test_large(folio
));
5933 if (folio_memcg(folio
) != from
)
5936 pgdat
= folio_pgdat(folio
);
5937 from_vec
= mem_cgroup_lruvec(from
, pgdat
);
5938 to_vec
= mem_cgroup_lruvec(to
, pgdat
);
5940 folio_memcg_lock(folio
);
5942 if (folio_test_anon(folio
)) {
5943 if (folio_mapped(folio
)) {
5944 __mod_lruvec_state(from_vec
, NR_ANON_MAPPED
, -nr_pages
);
5945 __mod_lruvec_state(to_vec
, NR_ANON_MAPPED
, nr_pages
);
5946 if (folio_test_pmd_mappable(folio
)) {
5947 __mod_lruvec_state(from_vec
, NR_ANON_THPS
,
5949 __mod_lruvec_state(to_vec
, NR_ANON_THPS
,
5954 __mod_lruvec_state(from_vec
, NR_FILE_PAGES
, -nr_pages
);
5955 __mod_lruvec_state(to_vec
, NR_FILE_PAGES
, nr_pages
);
5957 if (folio_test_swapbacked(folio
)) {
5958 __mod_lruvec_state(from_vec
, NR_SHMEM
, -nr_pages
);
5959 __mod_lruvec_state(to_vec
, NR_SHMEM
, nr_pages
);
5962 if (folio_mapped(folio
)) {
5963 __mod_lruvec_state(from_vec
, NR_FILE_MAPPED
, -nr_pages
);
5964 __mod_lruvec_state(to_vec
, NR_FILE_MAPPED
, nr_pages
);
5967 if (folio_test_dirty(folio
)) {
5968 struct address_space
*mapping
= folio_mapping(folio
);
5970 if (mapping_can_writeback(mapping
)) {
5971 __mod_lruvec_state(from_vec
, NR_FILE_DIRTY
,
5973 __mod_lruvec_state(to_vec
, NR_FILE_DIRTY
,
5980 if (folio_test_swapcache(folio
)) {
5981 __mod_lruvec_state(from_vec
, NR_SWAPCACHE
, -nr_pages
);
5982 __mod_lruvec_state(to_vec
, NR_SWAPCACHE
, nr_pages
);
5985 if (folio_test_writeback(folio
)) {
5986 __mod_lruvec_state(from_vec
, NR_WRITEBACK
, -nr_pages
);
5987 __mod_lruvec_state(to_vec
, NR_WRITEBACK
, nr_pages
);
5991 * All state has been migrated, let's switch to the new memcg.
5993 * It is safe to change page's memcg here because the page
5994 * is referenced, charged, isolated, and locked: we can't race
5995 * with (un)charging, migration, LRU putback, or anything else
5996 * that would rely on a stable page's memory cgroup.
5998 * Note that folio_memcg_lock is a memcg lock, not a page lock,
5999 * to save space. As soon as we switch page's memory cgroup to a
6000 * new memcg that isn't locked, the above state can change
6001 * concurrently again. Make sure we're truly done with it.
6006 css_put(&from
->css
);
6008 folio
->memcg_data
= (unsigned long)to
;
6010 __folio_memcg_unlock(from
);
6013 nid
= folio_nid(folio
);
6015 local_irq_disable();
6016 mem_cgroup_charge_statistics(to
, nr_pages
);
6017 memcg_check_events(to
, nid
);
6018 mem_cgroup_charge_statistics(from
, -nr_pages
);
6019 memcg_check_events(from
, nid
);
6026 * get_mctgt_type - get target type of moving charge
6027 * @vma: the vma the pte to be checked belongs
6028 * @addr: the address corresponding to the pte to be checked
6029 * @ptent: the pte to be checked
6030 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6032 * Context: Called with pte lock held.
6034 * * MC_TARGET_NONE - If the pte is not a target for move charge.
6035 * * MC_TARGET_PAGE - If the page corresponding to this pte is a target for
6036 * move charge. If @target is not NULL, the page is stored in target->page
6037 * with extra refcnt taken (Caller should release it).
6038 * * MC_TARGET_SWAP - If the swap entry corresponding to this pte is a
6039 * target for charge migration. If @target is not NULL, the entry is
6040 * stored in target->ent.
6041 * * MC_TARGET_DEVICE - Like MC_TARGET_PAGE but page is device memory and
6042 * thus not on the lru. For now such page is charged like a regular page
6043 * would be as it is just special memory taking the place of a regular page.
6044 * See Documentations/vm/hmm.txt and include/linux/hmm.h
6046 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6047 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6049 struct page
*page
= NULL
;
6050 enum mc_target_type ret
= MC_TARGET_NONE
;
6051 swp_entry_t ent
= { .val
= 0 };
6053 if (pte_present(ptent
))
6054 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6055 else if (pte_none_mostly(ptent
))
6057 * PTE markers should be treated as a none pte here, separated
6058 * from other swap handling below.
6060 page
= mc_handle_file_pte(vma
, addr
, ptent
);
6061 else if (is_swap_pte(ptent
))
6062 page
= mc_handle_swap_pte(vma
, ptent
, &ent
);
6064 if (target
&& page
) {
6065 if (!trylock_page(page
)) {
6070 * page_mapped() must be stable during the move. This
6071 * pte is locked, so if it's present, the page cannot
6072 * become unmapped. If it isn't, we have only partial
6073 * control over the mapped state: the page lock will
6074 * prevent new faults against pagecache and swapcache,
6075 * so an unmapped page cannot become mapped. However,
6076 * if the page is already mapped elsewhere, it can
6077 * unmap, and there is nothing we can do about it.
6078 * Alas, skip moving the page in this case.
6080 if (!pte_present(ptent
) && page_mapped(page
)) {
6087 if (!page
&& !ent
.val
)
6091 * Do only loose check w/o serialization.
6092 * mem_cgroup_move_account() checks the page is valid or
6093 * not under LRU exclusion.
6095 if (page_memcg(page
) == mc
.from
) {
6096 ret
= MC_TARGET_PAGE
;
6097 if (is_device_private_page(page
) ||
6098 is_device_coherent_page(page
))
6099 ret
= MC_TARGET_DEVICE
;
6101 target
->page
= page
;
6103 if (!ret
|| !target
) {
6110 * There is a swap entry and a page doesn't exist or isn't charged.
6111 * But we cannot move a tail-page in a THP.
6113 if (ent
.val
&& !ret
&& (!page
|| !PageTransCompound(page
)) &&
6114 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6115 ret
= MC_TARGET_SWAP
;
6122 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6124 * We don't consider PMD mapped swapping or file mapped pages because THP does
6125 * not support them for now.
6126 * Caller should make sure that pmd_trans_huge(pmd) is true.
6128 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6129 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6131 struct page
*page
= NULL
;
6132 enum mc_target_type ret
= MC_TARGET_NONE
;
6134 if (unlikely(is_swap_pmd(pmd
))) {
6135 VM_BUG_ON(thp_migration_supported() &&
6136 !is_pmd_migration_entry(pmd
));
6139 page
= pmd_page(pmd
);
6140 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6141 if (!(mc
.flags
& MOVE_ANON
))
6143 if (page_memcg(page
) == mc
.from
) {
6144 ret
= MC_TARGET_PAGE
;
6147 if (!trylock_page(page
)) {
6149 return MC_TARGET_NONE
;
6151 target
->page
= page
;
6157 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6158 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6160 return MC_TARGET_NONE
;
6164 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6165 unsigned long addr
, unsigned long end
,
6166 struct mm_walk
*walk
)
6168 struct vm_area_struct
*vma
= walk
->vma
;
6172 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6175 * Note their can not be MC_TARGET_DEVICE for now as we do not
6176 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
6177 * this might change.
6179 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6180 mc
.precharge
+= HPAGE_PMD_NR
;
6185 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6188 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6189 if (get_mctgt_type(vma
, addr
, ptep_get(pte
), NULL
))
6190 mc
.precharge
++; /* increment precharge temporarily */
6191 pte_unmap_unlock(pte
- 1, ptl
);
6197 static const struct mm_walk_ops precharge_walk_ops
= {
6198 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6199 .walk_lock
= PGWALK_RDLOCK
,
6202 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6204 unsigned long precharge
;
6207 walk_page_range(mm
, 0, ULONG_MAX
, &precharge_walk_ops
, NULL
);
6208 mmap_read_unlock(mm
);
6210 precharge
= mc
.precharge
;
6216 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6218 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6220 VM_BUG_ON(mc
.moving_task
);
6221 mc
.moving_task
= current
;
6222 return mem_cgroup_do_precharge(precharge
);
6225 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6226 static void __mem_cgroup_clear_mc(void)
6228 struct mem_cgroup
*from
= mc
.from
;
6229 struct mem_cgroup
*to
= mc
.to
;
6231 /* we must uncharge all the leftover precharges from mc.to */
6233 mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6237 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6238 * we must uncharge here.
6240 if (mc
.moved_charge
) {
6241 mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6242 mc
.moved_charge
= 0;
6244 /* we must fixup refcnts and charges */
6245 if (mc
.moved_swap
) {
6246 /* uncharge swap account from the old cgroup */
6247 if (!mem_cgroup_is_root(mc
.from
))
6248 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
6250 mem_cgroup_id_put_many(mc
.from
, mc
.moved_swap
);
6253 * we charged both to->memory and to->memsw, so we
6254 * should uncharge to->memory.
6256 if (!mem_cgroup_is_root(mc
.to
))
6257 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
6261 memcg_oom_recover(from
);
6262 memcg_oom_recover(to
);
6263 wake_up_all(&mc
.waitq
);
6266 static void mem_cgroup_clear_mc(void)
6268 struct mm_struct
*mm
= mc
.mm
;
6271 * we must clear moving_task before waking up waiters at the end of
6274 mc
.moving_task
= NULL
;
6275 __mem_cgroup_clear_mc();
6276 spin_lock(&mc
.lock
);
6280 spin_unlock(&mc
.lock
);
6285 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6287 struct cgroup_subsys_state
*css
;
6288 struct mem_cgroup
*memcg
= NULL
; /* unneeded init to make gcc happy */
6289 struct mem_cgroup
*from
;
6290 struct task_struct
*leader
, *p
;
6291 struct mm_struct
*mm
;
6292 unsigned long move_flags
;
6295 /* charge immigration isn't supported on the default hierarchy */
6296 if (cgroup_subsys_on_dfl(memory_cgrp_subsys
))
6300 * Multi-process migrations only happen on the default hierarchy
6301 * where charge immigration is not used. Perform charge
6302 * immigration if @tset contains a leader and whine if there are
6306 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
6309 memcg
= mem_cgroup_from_css(css
);
6315 * We are now committed to this value whatever it is. Changes in this
6316 * tunable will only affect upcoming migrations, not the current one.
6317 * So we need to save it, and keep it going.
6319 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
6323 from
= mem_cgroup_from_task(p
);
6325 VM_BUG_ON(from
== memcg
);
6327 mm
= get_task_mm(p
);
6330 /* We move charges only when we move a owner of the mm */
6331 if (mm
->owner
== p
) {
6334 VM_BUG_ON(mc
.precharge
);
6335 VM_BUG_ON(mc
.moved_charge
);
6336 VM_BUG_ON(mc
.moved_swap
);
6338 spin_lock(&mc
.lock
);
6342 mc
.flags
= move_flags
;
6343 spin_unlock(&mc
.lock
);
6344 /* We set mc.moving_task later */
6346 ret
= mem_cgroup_precharge_mc(mm
);
6348 mem_cgroup_clear_mc();
6355 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6358 mem_cgroup_clear_mc();
6361 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6362 unsigned long addr
, unsigned long end
,
6363 struct mm_walk
*walk
)
6366 struct vm_area_struct
*vma
= walk
->vma
;
6369 enum mc_target_type target_type
;
6370 union mc_target target
;
6373 ptl
= pmd_trans_huge_lock(pmd
, vma
);
6375 if (mc
.precharge
< HPAGE_PMD_NR
) {
6379 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6380 if (target_type
== MC_TARGET_PAGE
) {
6382 if (isolate_lru_page(page
)) {
6383 if (!mem_cgroup_move_account(page
, true,
6385 mc
.precharge
-= HPAGE_PMD_NR
;
6386 mc
.moved_charge
+= HPAGE_PMD_NR
;
6388 putback_lru_page(page
);
6392 } else if (target_type
== MC_TARGET_DEVICE
) {
6394 if (!mem_cgroup_move_account(page
, true,
6396 mc
.precharge
-= HPAGE_PMD_NR
;
6397 mc
.moved_charge
+= HPAGE_PMD_NR
;
6407 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6410 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6411 pte_t ptent
= ptep_get(pte
++);
6412 bool device
= false;
6418 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6419 case MC_TARGET_DEVICE
:
6422 case MC_TARGET_PAGE
:
6425 * We can have a part of the split pmd here. Moving it
6426 * can be done but it would be too convoluted so simply
6427 * ignore such a partial THP and keep it in original
6428 * memcg. There should be somebody mapping the head.
6430 if (PageTransCompound(page
))
6432 if (!device
&& !isolate_lru_page(page
))
6434 if (!mem_cgroup_move_account(page
, false,
6437 /* we uncharge from mc.from later. */
6441 putback_lru_page(page
);
6442 put
: /* get_mctgt_type() gets & locks the page */
6446 case MC_TARGET_SWAP
:
6448 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6450 mem_cgroup_id_get_many(mc
.to
, 1);
6451 /* we fixup other refcnts and charges later. */
6459 pte_unmap_unlock(pte
- 1, ptl
);
6464 * We have consumed all precharges we got in can_attach().
6465 * We try charge one by one, but don't do any additional
6466 * charges to mc.to if we have failed in charge once in attach()
6469 ret
= mem_cgroup_do_precharge(1);
6477 static const struct mm_walk_ops charge_walk_ops
= {
6478 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6479 .walk_lock
= PGWALK_RDLOCK
,
6482 static void mem_cgroup_move_charge(void)
6484 lru_add_drain_all();
6486 * Signal folio_memcg_lock() to take the memcg's move_lock
6487 * while we're moving its pages to another memcg. Then wait
6488 * for already started RCU-only updates to finish.
6490 atomic_inc(&mc
.from
->moving_account
);
6493 if (unlikely(!mmap_read_trylock(mc
.mm
))) {
6495 * Someone who are holding the mmap_lock might be waiting in
6496 * waitq. So we cancel all extra charges, wake up all waiters,
6497 * and retry. Because we cancel precharges, we might not be able
6498 * to move enough charges, but moving charge is a best-effort
6499 * feature anyway, so it wouldn't be a big problem.
6501 __mem_cgroup_clear_mc();
6506 * When we have consumed all precharges and failed in doing
6507 * additional charge, the page walk just aborts.
6509 walk_page_range(mc
.mm
, 0, ULONG_MAX
, &charge_walk_ops
, NULL
);
6510 mmap_read_unlock(mc
.mm
);
6511 atomic_dec(&mc
.from
->moving_account
);
6514 static void mem_cgroup_move_task(void)
6517 mem_cgroup_move_charge();
6518 mem_cgroup_clear_mc();
6522 #else /* !CONFIG_MMU */
6523 static int mem_cgroup_can_attach(struct cgroup_taskset
*tset
)
6527 static void mem_cgroup_cancel_attach(struct cgroup_taskset
*tset
)
6530 static void mem_cgroup_move_task(void)
6535 #ifdef CONFIG_MEMCG_KMEM
6536 static void mem_cgroup_fork(struct task_struct
*task
)
6539 * Set the update flag to cause task->objcg to be initialized lazily
6540 * on the first allocation. It can be done without any synchronization
6541 * because it's always performed on the current task, so does
6542 * current_objcg_update().
6544 task
->objcg
= (struct obj_cgroup
*)CURRENT_OBJCG_UPDATE_FLAG
;
6547 static void mem_cgroup_exit(struct task_struct
*task
)
6549 struct obj_cgroup
*objcg
= task
->objcg
;
6551 objcg
= (struct obj_cgroup
*)
6552 ((unsigned long)objcg
& ~CURRENT_OBJCG_UPDATE_FLAG
);
6554 obj_cgroup_put(objcg
);
6557 * Some kernel allocations can happen after this point,
6558 * but let's ignore them. It can be done without any synchronization
6559 * because it's always performed on the current task, so does
6560 * current_objcg_update().
6566 #ifdef CONFIG_LRU_GEN
6567 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset
*tset
)
6569 struct task_struct
*task
;
6570 struct cgroup_subsys_state
*css
;
6572 /* find the first leader if there is any */
6573 cgroup_taskset_for_each_leader(task
, css
, tset
)
6580 if (task
->mm
&& READ_ONCE(task
->mm
->owner
) == task
)
6581 lru_gen_migrate_mm(task
->mm
);
6585 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset
*tset
) {}
6586 #endif /* CONFIG_LRU_GEN */
6588 #ifdef CONFIG_MEMCG_KMEM
6589 static void mem_cgroup_kmem_attach(struct cgroup_taskset
*tset
)
6591 struct task_struct
*task
;
6592 struct cgroup_subsys_state
*css
;
6594 cgroup_taskset_for_each(task
, css
, tset
) {
6595 /* atomically set the update bit */
6596 set_bit(CURRENT_OBJCG_UPDATE_BIT
, (unsigned long *)&task
->objcg
);
6600 static void mem_cgroup_kmem_attach(struct cgroup_taskset
*tset
) {}
6601 #endif /* CONFIG_MEMCG_KMEM */
6603 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
6604 static void mem_cgroup_attach(struct cgroup_taskset
*tset
)
6606 mem_cgroup_lru_gen_attach(tset
);
6607 mem_cgroup_kmem_attach(tset
);
6611 static int seq_puts_memcg_tunable(struct seq_file
*m
, unsigned long value
)
6613 if (value
== PAGE_COUNTER_MAX
)
6614 seq_puts(m
, "max\n");
6616 seq_printf(m
, "%llu\n", (u64
)value
* PAGE_SIZE
);
6621 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
6624 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6626 return (u64
)page_counter_read(&memcg
->memory
) * PAGE_SIZE
;
6629 static u64
memory_peak_read(struct cgroup_subsys_state
*css
,
6632 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6634 return (u64
)memcg
->memory
.watermark
* PAGE_SIZE
;
6637 static int memory_min_show(struct seq_file
*m
, void *v
)
6639 return seq_puts_memcg_tunable(m
,
6640 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.min
));
6643 static ssize_t
memory_min_write(struct kernfs_open_file
*of
,
6644 char *buf
, size_t nbytes
, loff_t off
)
6646 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6650 buf
= strstrip(buf
);
6651 err
= page_counter_memparse(buf
, "max", &min
);
6655 page_counter_set_min(&memcg
->memory
, min
);
6660 static int memory_low_show(struct seq_file
*m
, void *v
)
6662 return seq_puts_memcg_tunable(m
,
6663 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.low
));
6666 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
6667 char *buf
, size_t nbytes
, loff_t off
)
6669 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6673 buf
= strstrip(buf
);
6674 err
= page_counter_memparse(buf
, "max", &low
);
6678 page_counter_set_low(&memcg
->memory
, low
);
6683 static int memory_high_show(struct seq_file
*m
, void *v
)
6685 return seq_puts_memcg_tunable(m
,
6686 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.high
));
6689 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
6690 char *buf
, size_t nbytes
, loff_t off
)
6692 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6693 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6694 bool drained
= false;
6698 buf
= strstrip(buf
);
6699 err
= page_counter_memparse(buf
, "max", &high
);
6703 page_counter_set_high(&memcg
->memory
, high
);
6706 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6707 unsigned long reclaimed
;
6709 if (nr_pages
<= high
)
6712 if (signal_pending(current
))
6716 drain_all_stock(memcg
);
6721 reclaimed
= try_to_free_mem_cgroup_pages(memcg
, nr_pages
- high
,
6722 GFP_KERNEL
, MEMCG_RECLAIM_MAY_SWAP
);
6724 if (!reclaimed
&& !nr_retries
--)
6728 memcg_wb_domain_size_changed(memcg
);
6732 static int memory_max_show(struct seq_file
*m
, void *v
)
6734 return seq_puts_memcg_tunable(m
,
6735 READ_ONCE(mem_cgroup_from_seq(m
)->memory
.max
));
6738 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
6739 char *buf
, size_t nbytes
, loff_t off
)
6741 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6742 unsigned int nr_reclaims
= MAX_RECLAIM_RETRIES
;
6743 bool drained
= false;
6747 buf
= strstrip(buf
);
6748 err
= page_counter_memparse(buf
, "max", &max
);
6752 xchg(&memcg
->memory
.max
, max
);
6755 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
6757 if (nr_pages
<= max
)
6760 if (signal_pending(current
))
6764 drain_all_stock(memcg
);
6770 if (!try_to_free_mem_cgroup_pages(memcg
, nr_pages
- max
,
6771 GFP_KERNEL
, MEMCG_RECLAIM_MAY_SWAP
))
6776 memcg_memory_event(memcg
, MEMCG_OOM
);
6777 if (!mem_cgroup_out_of_memory(memcg
, GFP_KERNEL
, 0))
6781 memcg_wb_domain_size_changed(memcg
);
6785 static void __memory_events_show(struct seq_file
*m
, atomic_long_t
*events
)
6787 seq_printf(m
, "low %lu\n", atomic_long_read(&events
[MEMCG_LOW
]));
6788 seq_printf(m
, "high %lu\n", atomic_long_read(&events
[MEMCG_HIGH
]));
6789 seq_printf(m
, "max %lu\n", atomic_long_read(&events
[MEMCG_MAX
]));
6790 seq_printf(m
, "oom %lu\n", atomic_long_read(&events
[MEMCG_OOM
]));
6791 seq_printf(m
, "oom_kill %lu\n",
6792 atomic_long_read(&events
[MEMCG_OOM_KILL
]));
6793 seq_printf(m
, "oom_group_kill %lu\n",
6794 atomic_long_read(&events
[MEMCG_OOM_GROUP_KILL
]));
6797 static int memory_events_show(struct seq_file
*m
, void *v
)
6799 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6801 __memory_events_show(m
, memcg
->memory_events
);
6805 static int memory_events_local_show(struct seq_file
*m
, void *v
)
6807 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6809 __memory_events_show(m
, memcg
->memory_events_local
);
6813 static int memory_stat_show(struct seq_file
*m
, void *v
)
6815 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6816 char *buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
6821 seq_buf_init(&s
, buf
, PAGE_SIZE
);
6822 memory_stat_format(memcg
, &s
);
6829 static inline unsigned long lruvec_page_state_output(struct lruvec
*lruvec
,
6832 return lruvec_page_state(lruvec
, item
) *
6833 memcg_page_state_output_unit(item
);
6836 static int memory_numa_stat_show(struct seq_file
*m
, void *v
)
6839 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6841 mem_cgroup_flush_stats();
6843 for (i
= 0; i
< ARRAY_SIZE(memory_stats
); i
++) {
6846 if (memory_stats
[i
].idx
>= NR_VM_NODE_STAT_ITEMS
)
6849 seq_printf(m
, "%s", memory_stats
[i
].name
);
6850 for_each_node_state(nid
, N_MEMORY
) {
6852 struct lruvec
*lruvec
;
6854 lruvec
= mem_cgroup_lruvec(memcg
, NODE_DATA(nid
));
6855 size
= lruvec_page_state_output(lruvec
,
6856 memory_stats
[i
].idx
);
6857 seq_printf(m
, " N%d=%llu", nid
, size
);
6866 static int memory_oom_group_show(struct seq_file
*m
, void *v
)
6868 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
6870 seq_printf(m
, "%d\n", READ_ONCE(memcg
->oom_group
));
6875 static ssize_t
memory_oom_group_write(struct kernfs_open_file
*of
,
6876 char *buf
, size_t nbytes
, loff_t off
)
6878 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6881 buf
= strstrip(buf
);
6885 ret
= kstrtoint(buf
, 0, &oom_group
);
6889 if (oom_group
!= 0 && oom_group
!= 1)
6892 WRITE_ONCE(memcg
->oom_group
, oom_group
);
6897 static ssize_t
memory_reclaim(struct kernfs_open_file
*of
, char *buf
,
6898 size_t nbytes
, loff_t off
)
6900 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
6901 unsigned int nr_retries
= MAX_RECLAIM_RETRIES
;
6902 unsigned long nr_to_reclaim
, nr_reclaimed
= 0;
6903 unsigned int reclaim_options
;
6906 buf
= strstrip(buf
);
6907 err
= page_counter_memparse(buf
, "", &nr_to_reclaim
);
6911 reclaim_options
= MEMCG_RECLAIM_MAY_SWAP
| MEMCG_RECLAIM_PROACTIVE
;
6912 while (nr_reclaimed
< nr_to_reclaim
) {
6913 unsigned long reclaimed
;
6915 if (signal_pending(current
))
6919 * This is the final attempt, drain percpu lru caches in the
6920 * hope of introducing more evictable pages for
6921 * try_to_free_mem_cgroup_pages().
6924 lru_add_drain_all();
6926 reclaimed
= try_to_free_mem_cgroup_pages(memcg
,
6927 min(nr_to_reclaim
- nr_reclaimed
, SWAP_CLUSTER_MAX
),
6928 GFP_KERNEL
, reclaim_options
);
6930 if (!reclaimed
&& !nr_retries
--)
6933 nr_reclaimed
+= reclaimed
;
6939 static struct cftype memory_files
[] = {
6942 .flags
= CFTYPE_NOT_ON_ROOT
,
6943 .read_u64
= memory_current_read
,
6947 .flags
= CFTYPE_NOT_ON_ROOT
,
6948 .read_u64
= memory_peak_read
,
6952 .flags
= CFTYPE_NOT_ON_ROOT
,
6953 .seq_show
= memory_min_show
,
6954 .write
= memory_min_write
,
6958 .flags
= CFTYPE_NOT_ON_ROOT
,
6959 .seq_show
= memory_low_show
,
6960 .write
= memory_low_write
,
6964 .flags
= CFTYPE_NOT_ON_ROOT
,
6965 .seq_show
= memory_high_show
,
6966 .write
= memory_high_write
,
6970 .flags
= CFTYPE_NOT_ON_ROOT
,
6971 .seq_show
= memory_max_show
,
6972 .write
= memory_max_write
,
6976 .flags
= CFTYPE_NOT_ON_ROOT
,
6977 .file_offset
= offsetof(struct mem_cgroup
, events_file
),
6978 .seq_show
= memory_events_show
,
6981 .name
= "events.local",
6982 .flags
= CFTYPE_NOT_ON_ROOT
,
6983 .file_offset
= offsetof(struct mem_cgroup
, events_local_file
),
6984 .seq_show
= memory_events_local_show
,
6988 .seq_show
= memory_stat_show
,
6992 .name
= "numa_stat",
6993 .seq_show
= memory_numa_stat_show
,
6997 .name
= "oom.group",
6998 .flags
= CFTYPE_NOT_ON_ROOT
| CFTYPE_NS_DELEGATABLE
,
6999 .seq_show
= memory_oom_group_show
,
7000 .write
= memory_oom_group_write
,
7004 .flags
= CFTYPE_NS_DELEGATABLE
,
7005 .write
= memory_reclaim
,
7010 struct cgroup_subsys memory_cgrp_subsys
= {
7011 .css_alloc
= mem_cgroup_css_alloc
,
7012 .css_online
= mem_cgroup_css_online
,
7013 .css_offline
= mem_cgroup_css_offline
,
7014 .css_released
= mem_cgroup_css_released
,
7015 .css_free
= mem_cgroup_css_free
,
7016 .css_reset
= mem_cgroup_css_reset
,
7017 .css_rstat_flush
= mem_cgroup_css_rstat_flush
,
7018 .can_attach
= mem_cgroup_can_attach
,
7019 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
7020 .attach
= mem_cgroup_attach
,
7022 .cancel_attach
= mem_cgroup_cancel_attach
,
7023 .post_attach
= mem_cgroup_move_task
,
7024 #ifdef CONFIG_MEMCG_KMEM
7025 .fork
= mem_cgroup_fork
,
7026 .exit
= mem_cgroup_exit
,
7028 .dfl_cftypes
= memory_files
,
7029 .legacy_cftypes
= mem_cgroup_legacy_files
,
7034 * This function calculates an individual cgroup's effective
7035 * protection which is derived from its own memory.min/low, its
7036 * parent's and siblings' settings, as well as the actual memory
7037 * distribution in the tree.
7039 * The following rules apply to the effective protection values:
7041 * 1. At the first level of reclaim, effective protection is equal to
7042 * the declared protection in memory.min and memory.low.
7044 * 2. To enable safe delegation of the protection configuration, at
7045 * subsequent levels the effective protection is capped to the
7046 * parent's effective protection.
7048 * 3. To make complex and dynamic subtrees easier to configure, the
7049 * user is allowed to overcommit the declared protection at a given
7050 * level. If that is the case, the parent's effective protection is
7051 * distributed to the children in proportion to how much protection
7052 * they have declared and how much of it they are utilizing.
7054 * This makes distribution proportional, but also work-conserving:
7055 * if one cgroup claims much more protection than it uses memory,
7056 * the unused remainder is available to its siblings.
7058 * 4. Conversely, when the declared protection is undercommitted at a
7059 * given level, the distribution of the larger parental protection
7060 * budget is NOT proportional. A cgroup's protection from a sibling
7061 * is capped to its own memory.min/low setting.
7063 * 5. However, to allow protecting recursive subtrees from each other
7064 * without having to declare each individual cgroup's fixed share
7065 * of the ancestor's claim to protection, any unutilized -
7066 * "floating" - protection from up the tree is distributed in
7067 * proportion to each cgroup's *usage*. This makes the protection
7068 * neutral wrt sibling cgroups and lets them compete freely over
7069 * the shared parental protection budget, but it protects the
7070 * subtree as a whole from neighboring subtrees.
7072 * Note that 4. and 5. are not in conflict: 4. is about protecting
7073 * against immediate siblings whereas 5. is about protecting against
7074 * neighboring subtrees.
7076 static unsigned long effective_protection(unsigned long usage
,
7077 unsigned long parent_usage
,
7078 unsigned long setting
,
7079 unsigned long parent_effective
,
7080 unsigned long siblings_protected
)
7082 unsigned long protected;
7085 protected = min(usage
, setting
);
7087 * If all cgroups at this level combined claim and use more
7088 * protection than what the parent affords them, distribute
7089 * shares in proportion to utilization.
7091 * We are using actual utilization rather than the statically
7092 * claimed protection in order to be work-conserving: claimed
7093 * but unused protection is available to siblings that would
7094 * otherwise get a smaller chunk than what they claimed.
7096 if (siblings_protected
> parent_effective
)
7097 return protected * parent_effective
/ siblings_protected
;
7100 * Ok, utilized protection of all children is within what the
7101 * parent affords them, so we know whatever this child claims
7102 * and utilizes is effectively protected.
7104 * If there is unprotected usage beyond this value, reclaim
7105 * will apply pressure in proportion to that amount.
7107 * If there is unutilized protection, the cgroup will be fully
7108 * shielded from reclaim, but we do return a smaller value for
7109 * protection than what the group could enjoy in theory. This
7110 * is okay. With the overcommit distribution above, effective
7111 * protection is always dependent on how memory is actually
7112 * consumed among the siblings anyway.
7117 * If the children aren't claiming (all of) the protection
7118 * afforded to them by the parent, distribute the remainder in
7119 * proportion to the (unprotected) memory of each cgroup. That
7120 * way, cgroups that aren't explicitly prioritized wrt each
7121 * other compete freely over the allowance, but they are
7122 * collectively protected from neighboring trees.
7124 * We're using unprotected memory for the weight so that if
7125 * some cgroups DO claim explicit protection, we don't protect
7126 * the same bytes twice.
7128 * Check both usage and parent_usage against the respective
7129 * protected values. One should imply the other, but they
7130 * aren't read atomically - make sure the division is sane.
7132 if (!(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_RECURSIVE_PROT
))
7134 if (parent_effective
> siblings_protected
&&
7135 parent_usage
> siblings_protected
&&
7136 usage
> protected) {
7137 unsigned long unclaimed
;
7139 unclaimed
= parent_effective
- siblings_protected
;
7140 unclaimed
*= usage
- protected;
7141 unclaimed
/= parent_usage
- siblings_protected
;
7150 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
7151 * @root: the top ancestor of the sub-tree being checked
7152 * @memcg: the memory cgroup to check
7154 * WARNING: This function is not stateless! It can only be used as part
7155 * of a top-down tree iteration, not for isolated queries.
7157 void mem_cgroup_calculate_protection(struct mem_cgroup
*root
,
7158 struct mem_cgroup
*memcg
)
7160 unsigned long usage
, parent_usage
;
7161 struct mem_cgroup
*parent
;
7163 if (mem_cgroup_disabled())
7167 root
= root_mem_cgroup
;
7170 * Effective values of the reclaim targets are ignored so they
7171 * can be stale. Have a look at mem_cgroup_protection for more
7173 * TODO: calculation should be more robust so that we do not need
7174 * that special casing.
7179 usage
= page_counter_read(&memcg
->memory
);
7183 parent
= parent_mem_cgroup(memcg
);
7185 if (parent
== root
) {
7186 memcg
->memory
.emin
= READ_ONCE(memcg
->memory
.min
);
7187 memcg
->memory
.elow
= READ_ONCE(memcg
->memory
.low
);
7191 parent_usage
= page_counter_read(&parent
->memory
);
7193 WRITE_ONCE(memcg
->memory
.emin
, effective_protection(usage
, parent_usage
,
7194 READ_ONCE(memcg
->memory
.min
),
7195 READ_ONCE(parent
->memory
.emin
),
7196 atomic_long_read(&parent
->memory
.children_min_usage
)));
7198 WRITE_ONCE(memcg
->memory
.elow
, effective_protection(usage
, parent_usage
,
7199 READ_ONCE(memcg
->memory
.low
),
7200 READ_ONCE(parent
->memory
.elow
),
7201 atomic_long_read(&parent
->memory
.children_low_usage
)));
7204 static int charge_memcg(struct folio
*folio
, struct mem_cgroup
*memcg
,
7209 ret
= try_charge(memcg
, gfp
, folio_nr_pages(folio
));
7213 mem_cgroup_commit_charge(folio
, memcg
);
7218 int __mem_cgroup_charge(struct folio
*folio
, struct mm_struct
*mm
, gfp_t gfp
)
7220 struct mem_cgroup
*memcg
;
7223 memcg
= get_mem_cgroup_from_mm(mm
);
7224 ret
= charge_memcg(folio
, memcg
, gfp
);
7225 css_put(&memcg
->css
);
7231 * mem_cgroup_hugetlb_try_charge - try to charge the memcg for a hugetlb folio
7232 * @memcg: memcg to charge.
7233 * @gfp: reclaim mode.
7234 * @nr_pages: number of pages to charge.
7236 * This function is called when allocating a huge page folio to determine if
7237 * the memcg has the capacity for it. It does not commit the charge yet,
7238 * as the hugetlb folio itself has not been obtained from the hugetlb pool.
7240 * Once we have obtained the hugetlb folio, we can call
7241 * mem_cgroup_commit_charge() to commit the charge. If we fail to obtain the
7242 * folio, we should instead call mem_cgroup_cancel_charge() to undo the effect
7245 * Returns 0 on success. Otherwise, an error code is returned.
7247 int mem_cgroup_hugetlb_try_charge(struct mem_cgroup
*memcg
, gfp_t gfp
,
7251 * If hugetlb memcg charging is not enabled, do not fail hugetlb allocation,
7252 * but do not attempt to commit charge later (or cancel on error) either.
7254 if (mem_cgroup_disabled() || !memcg
||
7255 !cgroup_subsys_on_dfl(memory_cgrp_subsys
) ||
7256 !(cgrp_dfl_root
.flags
& CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING
))
7259 if (try_charge(memcg
, gfp
, nr_pages
))
7266 * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
7267 * @folio: folio to charge.
7268 * @mm: mm context of the victim
7269 * @gfp: reclaim mode
7270 * @entry: swap entry for which the folio is allocated
7272 * This function charges a folio allocated for swapin. Please call this before
7273 * adding the folio to the swapcache.
7275 * Returns 0 on success. Otherwise, an error code is returned.
7277 int mem_cgroup_swapin_charge_folio(struct folio
*folio
, struct mm_struct
*mm
,
7278 gfp_t gfp
, swp_entry_t entry
)
7280 struct mem_cgroup
*memcg
;
7284 if (mem_cgroup_disabled())
7287 id
= lookup_swap_cgroup_id(entry
);
7289 memcg
= mem_cgroup_from_id(id
);
7290 if (!memcg
|| !css_tryget_online(&memcg
->css
))
7291 memcg
= get_mem_cgroup_from_mm(mm
);
7294 ret
= charge_memcg(folio
, memcg
, gfp
);
7296 css_put(&memcg
->css
);
7301 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
7302 * @entry: swap entry for which the page is charged
7304 * Call this function after successfully adding the charged page to swapcache.
7306 * Note: This function assumes the page for which swap slot is being uncharged
7309 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry
)
7312 * Cgroup1's unified memory+swap counter has been charged with the
7313 * new swapcache page, finish the transfer by uncharging the swap
7314 * slot. The swap slot would also get uncharged when it dies, but
7315 * it can stick around indefinitely and we'd count the page twice
7318 * Cgroup2 has separate resource counters for memory and swap,
7319 * so this is a non-issue here. Memory and swap charge lifetimes
7320 * correspond 1:1 to page and swap slot lifetimes: we charge the
7321 * page to memory here, and uncharge swap when the slot is freed.
7323 if (!mem_cgroup_disabled() && do_memsw_account()) {
7325 * The swap entry might not get freed for a long time,
7326 * let's not wait for it. The page already received a
7327 * memory+swap charge, drop the swap entry duplicate.
7329 mem_cgroup_uncharge_swap(entry
, 1);
7333 struct uncharge_gather
{
7334 struct mem_cgroup
*memcg
;
7335 unsigned long nr_memory
;
7336 unsigned long pgpgout
;
7337 unsigned long nr_kmem
;
7341 static inline void uncharge_gather_clear(struct uncharge_gather
*ug
)
7343 memset(ug
, 0, sizeof(*ug
));
7346 static void uncharge_batch(const struct uncharge_gather
*ug
)
7348 unsigned long flags
;
7350 if (ug
->nr_memory
) {
7351 page_counter_uncharge(&ug
->memcg
->memory
, ug
->nr_memory
);
7352 if (do_memsw_account())
7353 page_counter_uncharge(&ug
->memcg
->memsw
, ug
->nr_memory
);
7355 memcg_account_kmem(ug
->memcg
, -ug
->nr_kmem
);
7356 memcg_oom_recover(ug
->memcg
);
7359 local_irq_save(flags
);
7360 __count_memcg_events(ug
->memcg
, PGPGOUT
, ug
->pgpgout
);
7361 __this_cpu_add(ug
->memcg
->vmstats_percpu
->nr_page_events
, ug
->nr_memory
);
7362 memcg_check_events(ug
->memcg
, ug
->nid
);
7363 local_irq_restore(flags
);
7365 /* drop reference from uncharge_folio */
7366 css_put(&ug
->memcg
->css
);
7369 static void uncharge_folio(struct folio
*folio
, struct uncharge_gather
*ug
)
7372 struct mem_cgroup
*memcg
;
7373 struct obj_cgroup
*objcg
;
7375 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
7378 * Nobody should be changing or seriously looking at
7379 * folio memcg or objcg at this point, we have fully
7380 * exclusive access to the folio.
7382 if (folio_memcg_kmem(folio
)) {
7383 objcg
= __folio_objcg(folio
);
7385 * This get matches the put at the end of the function and
7386 * kmem pages do not hold memcg references anymore.
7388 memcg
= get_mem_cgroup_from_objcg(objcg
);
7390 memcg
= __folio_memcg(folio
);
7396 if (ug
->memcg
!= memcg
) {
7399 uncharge_gather_clear(ug
);
7402 ug
->nid
= folio_nid(folio
);
7404 /* pairs with css_put in uncharge_batch */
7405 css_get(&memcg
->css
);
7408 nr_pages
= folio_nr_pages(folio
);
7410 if (folio_memcg_kmem(folio
)) {
7411 ug
->nr_memory
+= nr_pages
;
7412 ug
->nr_kmem
+= nr_pages
;
7414 folio
->memcg_data
= 0;
7415 obj_cgroup_put(objcg
);
7417 /* LRU pages aren't accounted at the root level */
7418 if (!mem_cgroup_is_root(memcg
))
7419 ug
->nr_memory
+= nr_pages
;
7422 folio
->memcg_data
= 0;
7425 css_put(&memcg
->css
);
7428 void __mem_cgroup_uncharge(struct folio
*folio
)
7430 struct uncharge_gather ug
;
7432 /* Don't touch folio->lru of any random page, pre-check: */
7433 if (!folio_memcg(folio
))
7436 uncharge_gather_clear(&ug
);
7437 uncharge_folio(folio
, &ug
);
7438 uncharge_batch(&ug
);
7442 * __mem_cgroup_uncharge_list - uncharge a list of page
7443 * @page_list: list of pages to uncharge
7445 * Uncharge a list of pages previously charged with
7446 * __mem_cgroup_charge().
7448 void __mem_cgroup_uncharge_list(struct list_head
*page_list
)
7450 struct uncharge_gather ug
;
7451 struct folio
*folio
;
7453 uncharge_gather_clear(&ug
);
7454 list_for_each_entry(folio
, page_list
, lru
)
7455 uncharge_folio(folio
, &ug
);
7457 uncharge_batch(&ug
);
7461 * mem_cgroup_replace_folio - Charge a folio's replacement.
7462 * @old: Currently circulating folio.
7463 * @new: Replacement folio.
7465 * Charge @new as a replacement folio for @old. @old will
7466 * be uncharged upon free. This is only used by the page cache
7467 * (in replace_page_cache_folio()).
7469 * Both folios must be locked, @new->mapping must be set up.
7471 void mem_cgroup_replace_folio(struct folio
*old
, struct folio
*new)
7473 struct mem_cgroup
*memcg
;
7474 long nr_pages
= folio_nr_pages(new);
7475 unsigned long flags
;
7477 VM_BUG_ON_FOLIO(!folio_test_locked(old
), old
);
7478 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7479 VM_BUG_ON_FOLIO(folio_test_anon(old
) != folio_test_anon(new), new);
7480 VM_BUG_ON_FOLIO(folio_nr_pages(old
) != nr_pages
, new);
7482 if (mem_cgroup_disabled())
7485 /* Page cache replacement: new folio already charged? */
7486 if (folio_memcg(new))
7489 memcg
= folio_memcg(old
);
7490 VM_WARN_ON_ONCE_FOLIO(!memcg
, old
);
7494 /* Force-charge the new page. The old one will be freed soon */
7495 if (!mem_cgroup_is_root(memcg
)) {
7496 page_counter_charge(&memcg
->memory
, nr_pages
);
7497 if (do_memsw_account())
7498 page_counter_charge(&memcg
->memsw
, nr_pages
);
7501 css_get(&memcg
->css
);
7502 commit_charge(new, memcg
);
7504 local_irq_save(flags
);
7505 mem_cgroup_charge_statistics(memcg
, nr_pages
);
7506 memcg_check_events(memcg
, folio_nid(new));
7507 local_irq_restore(flags
);
7511 * mem_cgroup_migrate - Transfer the memcg data from the old to the new folio.
7512 * @old: Currently circulating folio.
7513 * @new: Replacement folio.
7515 * Transfer the memcg data from the old folio to the new folio for migration.
7516 * The old folio's data info will be cleared. Note that the memory counters
7517 * will remain unchanged throughout the process.
7519 * Both folios must be locked, @new->mapping must be set up.
7521 void mem_cgroup_migrate(struct folio
*old
, struct folio
*new)
7523 struct mem_cgroup
*memcg
;
7525 VM_BUG_ON_FOLIO(!folio_test_locked(old
), old
);
7526 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7527 VM_BUG_ON_FOLIO(folio_test_anon(old
) != folio_test_anon(new), new);
7528 VM_BUG_ON_FOLIO(folio_nr_pages(old
) != folio_nr_pages(new), new);
7530 if (mem_cgroup_disabled())
7533 memcg
= folio_memcg(old
);
7535 * Note that it is normal to see !memcg for a hugetlb folio.
7536 * For e.g, itt could have been allocated when memory_hugetlb_accounting
7539 VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(old
) && !memcg
, old
);
7543 /* Transfer the charge and the css ref */
7544 commit_charge(new, memcg
);
7545 old
->memcg_data
= 0;
7548 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key
);
7549 EXPORT_SYMBOL(memcg_sockets_enabled_key
);
7551 void mem_cgroup_sk_alloc(struct sock
*sk
)
7553 struct mem_cgroup
*memcg
;
7555 if (!mem_cgroup_sockets_enabled
)
7558 /* Do not associate the sock with unrelated interrupted task's memcg. */
7563 memcg
= mem_cgroup_from_task(current
);
7564 if (mem_cgroup_is_root(memcg
))
7566 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
) && !memcg
->tcpmem_active
)
7568 if (css_tryget(&memcg
->css
))
7569 sk
->sk_memcg
= memcg
;
7574 void mem_cgroup_sk_free(struct sock
*sk
)
7577 css_put(&sk
->sk_memcg
->css
);
7581 * mem_cgroup_charge_skmem - charge socket memory
7582 * @memcg: memcg to charge
7583 * @nr_pages: number of pages to charge
7584 * @gfp_mask: reclaim mode
7586 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7587 * @memcg's configured limit, %false if it doesn't.
7589 bool mem_cgroup_charge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
,
7592 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7593 struct page_counter
*fail
;
7595 if (page_counter_try_charge(&memcg
->tcpmem
, nr_pages
, &fail
)) {
7596 memcg
->tcpmem_pressure
= 0;
7599 memcg
->tcpmem_pressure
= 1;
7600 if (gfp_mask
& __GFP_NOFAIL
) {
7601 page_counter_charge(&memcg
->tcpmem
, nr_pages
);
7607 if (try_charge(memcg
, gfp_mask
, nr_pages
) == 0) {
7608 mod_memcg_state(memcg
, MEMCG_SOCK
, nr_pages
);
7616 * mem_cgroup_uncharge_skmem - uncharge socket memory
7617 * @memcg: memcg to uncharge
7618 * @nr_pages: number of pages to uncharge
7620 void mem_cgroup_uncharge_skmem(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
7622 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
)) {
7623 page_counter_uncharge(&memcg
->tcpmem
, nr_pages
);
7627 mod_memcg_state(memcg
, MEMCG_SOCK
, -nr_pages
);
7629 refill_stock(memcg
, nr_pages
);
7632 static int __init
cgroup_memory(char *s
)
7636 while ((token
= strsep(&s
, ",")) != NULL
) {
7639 if (!strcmp(token
, "nosocket"))
7640 cgroup_memory_nosocket
= true;
7641 if (!strcmp(token
, "nokmem"))
7642 cgroup_memory_nokmem
= true;
7643 if (!strcmp(token
, "nobpf"))
7644 cgroup_memory_nobpf
= true;
7648 __setup("cgroup.memory=", cgroup_memory
);
7651 * subsys_initcall() for memory controller.
7653 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7654 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7655 * basically everything that doesn't depend on a specific mem_cgroup structure
7656 * should be initialized from here.
7658 static int __init
mem_cgroup_init(void)
7663 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7664 * used for per-memcg-per-cpu caching of per-node statistics. In order
7665 * to work fine, we should make sure that the overfill threshold can't
7666 * exceed S32_MAX / PAGE_SIZE.
7668 BUILD_BUG_ON(MEMCG_CHARGE_BATCH
> S32_MAX
/ PAGE_SIZE
);
7670 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD
, "mm/memctrl:dead", NULL
,
7671 memcg_hotplug_cpu_dead
);
7673 for_each_possible_cpu(cpu
)
7674 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
7677 for_each_node(node
) {
7678 struct mem_cgroup_tree_per_node
*rtpn
;
7680 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, node
);
7682 rtpn
->rb_root
= RB_ROOT
;
7683 rtpn
->rb_rightmost
= NULL
;
7684 spin_lock_init(&rtpn
->lock
);
7685 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
7690 subsys_initcall(mem_cgroup_init
);
7693 static struct mem_cgroup
*mem_cgroup_id_get_online(struct mem_cgroup
*memcg
)
7695 while (!refcount_inc_not_zero(&memcg
->id
.ref
)) {
7697 * The root cgroup cannot be destroyed, so it's refcount must
7700 if (WARN_ON_ONCE(mem_cgroup_is_root(memcg
))) {
7704 memcg
= parent_mem_cgroup(memcg
);
7706 memcg
= root_mem_cgroup
;
7712 * mem_cgroup_swapout - transfer a memsw charge to swap
7713 * @folio: folio whose memsw charge to transfer
7714 * @entry: swap entry to move the charge to
7716 * Transfer the memsw charge of @folio to @entry.
7718 void mem_cgroup_swapout(struct folio
*folio
, swp_entry_t entry
)
7720 struct mem_cgroup
*memcg
, *swap_memcg
;
7721 unsigned int nr_entries
;
7722 unsigned short oldid
;
7724 VM_BUG_ON_FOLIO(folio_test_lru(folio
), folio
);
7725 VM_BUG_ON_FOLIO(folio_ref_count(folio
), folio
);
7727 if (mem_cgroup_disabled())
7730 if (!do_memsw_account())
7733 memcg
= folio_memcg(folio
);
7735 VM_WARN_ON_ONCE_FOLIO(!memcg
, folio
);
7740 * In case the memcg owning these pages has been offlined and doesn't
7741 * have an ID allocated to it anymore, charge the closest online
7742 * ancestor for the swap instead and transfer the memory+swap charge.
7744 swap_memcg
= mem_cgroup_id_get_online(memcg
);
7745 nr_entries
= folio_nr_pages(folio
);
7746 /* Get references for the tail pages, too */
7748 mem_cgroup_id_get_many(swap_memcg
, nr_entries
- 1);
7749 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(swap_memcg
),
7751 VM_BUG_ON_FOLIO(oldid
, folio
);
7752 mod_memcg_state(swap_memcg
, MEMCG_SWAP
, nr_entries
);
7754 folio
->memcg_data
= 0;
7756 if (!mem_cgroup_is_root(memcg
))
7757 page_counter_uncharge(&memcg
->memory
, nr_entries
);
7759 if (memcg
!= swap_memcg
) {
7760 if (!mem_cgroup_is_root(swap_memcg
))
7761 page_counter_charge(&swap_memcg
->memsw
, nr_entries
);
7762 page_counter_uncharge(&memcg
->memsw
, nr_entries
);
7766 * Interrupts should be disabled here because the caller holds the
7767 * i_pages lock which is taken with interrupts-off. It is
7768 * important here to have the interrupts disabled because it is the
7769 * only synchronisation we have for updating the per-CPU variables.
7772 mem_cgroup_charge_statistics(memcg
, -nr_entries
);
7773 memcg_stats_unlock();
7774 memcg_check_events(memcg
, folio_nid(folio
));
7776 css_put(&memcg
->css
);
7780 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7781 * @folio: folio being added to swap
7782 * @entry: swap entry to charge
7784 * Try to charge @folio's memcg for the swap space at @entry.
7786 * Returns 0 on success, -ENOMEM on failure.
7788 int __mem_cgroup_try_charge_swap(struct folio
*folio
, swp_entry_t entry
)
7790 unsigned int nr_pages
= folio_nr_pages(folio
);
7791 struct page_counter
*counter
;
7792 struct mem_cgroup
*memcg
;
7793 unsigned short oldid
;
7795 if (do_memsw_account())
7798 memcg
= folio_memcg(folio
);
7800 VM_WARN_ON_ONCE_FOLIO(!memcg
, folio
);
7805 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7809 memcg
= mem_cgroup_id_get_online(memcg
);
7811 if (!mem_cgroup_is_root(memcg
) &&
7812 !page_counter_try_charge(&memcg
->swap
, nr_pages
, &counter
)) {
7813 memcg_memory_event(memcg
, MEMCG_SWAP_MAX
);
7814 memcg_memory_event(memcg
, MEMCG_SWAP_FAIL
);
7815 mem_cgroup_id_put(memcg
);
7819 /* Get references for the tail pages, too */
7821 mem_cgroup_id_get_many(memcg
, nr_pages
- 1);
7822 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
), nr_pages
);
7823 VM_BUG_ON_FOLIO(oldid
, folio
);
7824 mod_memcg_state(memcg
, MEMCG_SWAP
, nr_pages
);
7830 * __mem_cgroup_uncharge_swap - uncharge swap space
7831 * @entry: swap entry to uncharge
7832 * @nr_pages: the amount of swap space to uncharge
7834 void __mem_cgroup_uncharge_swap(swp_entry_t entry
, unsigned int nr_pages
)
7836 struct mem_cgroup
*memcg
;
7839 id
= swap_cgroup_record(entry
, 0, nr_pages
);
7841 memcg
= mem_cgroup_from_id(id
);
7843 if (!mem_cgroup_is_root(memcg
)) {
7844 if (do_memsw_account())
7845 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
7847 page_counter_uncharge(&memcg
->swap
, nr_pages
);
7849 mod_memcg_state(memcg
, MEMCG_SWAP
, -nr_pages
);
7850 mem_cgroup_id_put_many(memcg
, nr_pages
);
7855 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup
*memcg
)
7857 long nr_swap_pages
= get_nr_swap_pages();
7859 if (mem_cgroup_disabled() || do_memsw_account())
7860 return nr_swap_pages
;
7861 for (; !mem_cgroup_is_root(memcg
); memcg
= parent_mem_cgroup(memcg
))
7862 nr_swap_pages
= min_t(long, nr_swap_pages
,
7863 READ_ONCE(memcg
->swap
.max
) -
7864 page_counter_read(&memcg
->swap
));
7865 return nr_swap_pages
;
7868 bool mem_cgroup_swap_full(struct folio
*folio
)
7870 struct mem_cgroup
*memcg
;
7872 VM_BUG_ON_FOLIO(!folio_test_locked(folio
), folio
);
7876 if (do_memsw_account())
7879 memcg
= folio_memcg(folio
);
7883 for (; !mem_cgroup_is_root(memcg
); memcg
= parent_mem_cgroup(memcg
)) {
7884 unsigned long usage
= page_counter_read(&memcg
->swap
);
7886 if (usage
* 2 >= READ_ONCE(memcg
->swap
.high
) ||
7887 usage
* 2 >= READ_ONCE(memcg
->swap
.max
))
7894 static int __init
setup_swap_account(char *s
)
7896 pr_warn_once("The swapaccount= commandline option is deprecated. "
7897 "Please report your usecase to linux-mm@kvack.org if you "
7898 "depend on this functionality.\n");
7901 __setup("swapaccount=", setup_swap_account
);
7903 static u64
swap_current_read(struct cgroup_subsys_state
*css
,
7906 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7908 return (u64
)page_counter_read(&memcg
->swap
) * PAGE_SIZE
;
7911 static u64
swap_peak_read(struct cgroup_subsys_state
*css
,
7914 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7916 return (u64
)memcg
->swap
.watermark
* PAGE_SIZE
;
7919 static int swap_high_show(struct seq_file
*m
, void *v
)
7921 return seq_puts_memcg_tunable(m
,
7922 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.high
));
7925 static ssize_t
swap_high_write(struct kernfs_open_file
*of
,
7926 char *buf
, size_t nbytes
, loff_t off
)
7928 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7932 buf
= strstrip(buf
);
7933 err
= page_counter_memparse(buf
, "max", &high
);
7937 page_counter_set_high(&memcg
->swap
, high
);
7942 static int swap_max_show(struct seq_file
*m
, void *v
)
7944 return seq_puts_memcg_tunable(m
,
7945 READ_ONCE(mem_cgroup_from_seq(m
)->swap
.max
));
7948 static ssize_t
swap_max_write(struct kernfs_open_file
*of
,
7949 char *buf
, size_t nbytes
, loff_t off
)
7951 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
7955 buf
= strstrip(buf
);
7956 err
= page_counter_memparse(buf
, "max", &max
);
7960 xchg(&memcg
->swap
.max
, max
);
7965 static int swap_events_show(struct seq_file
*m
, void *v
)
7967 struct mem_cgroup
*memcg
= mem_cgroup_from_seq(m
);
7969 seq_printf(m
, "high %lu\n",
7970 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_HIGH
]));
7971 seq_printf(m
, "max %lu\n",
7972 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_MAX
]));
7973 seq_printf(m
, "fail %lu\n",
7974 atomic_long_read(&memcg
->memory_events
[MEMCG_SWAP_FAIL
]));
7979 static struct cftype swap_files
[] = {
7981 .name
= "swap.current",
7982 .flags
= CFTYPE_NOT_ON_ROOT
,
7983 .read_u64
= swap_current_read
,
7986 .name
= "swap.high",
7987 .flags
= CFTYPE_NOT_ON_ROOT
,
7988 .seq_show
= swap_high_show
,
7989 .write
= swap_high_write
,
7993 .flags
= CFTYPE_NOT_ON_ROOT
,
7994 .seq_show
= swap_max_show
,
7995 .write
= swap_max_write
,
7998 .name
= "swap.peak",
7999 .flags
= CFTYPE_NOT_ON_ROOT
,
8000 .read_u64
= swap_peak_read
,
8003 .name
= "swap.events",
8004 .flags
= CFTYPE_NOT_ON_ROOT
,
8005 .file_offset
= offsetof(struct mem_cgroup
, swap_events_file
),
8006 .seq_show
= swap_events_show
,
8011 static struct cftype memsw_files
[] = {
8013 .name
= "memsw.usage_in_bytes",
8014 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
8015 .read_u64
= mem_cgroup_read_u64
,
8018 .name
= "memsw.max_usage_in_bytes",
8019 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
8020 .write
= mem_cgroup_reset
,
8021 .read_u64
= mem_cgroup_read_u64
,
8024 .name
= "memsw.limit_in_bytes",
8025 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
8026 .write
= mem_cgroup_write
,
8027 .read_u64
= mem_cgroup_read_u64
,
8030 .name
= "memsw.failcnt",
8031 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
8032 .write
= mem_cgroup_reset
,
8033 .read_u64
= mem_cgroup_read_u64
,
8035 { }, /* terminate */
8038 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8040 * obj_cgroup_may_zswap - check if this cgroup can zswap
8041 * @objcg: the object cgroup
8043 * Check if the hierarchical zswap limit has been reached.
8045 * This doesn't check for specific headroom, and it is not atomic
8046 * either. But with zswap, the size of the allocation is only known
8047 * once compression has occurred, and this optimistic pre-check avoids
8048 * spending cycles on compression when there is already no room left
8049 * or zswap is disabled altogether somewhere in the hierarchy.
8051 bool obj_cgroup_may_zswap(struct obj_cgroup
*objcg
)
8053 struct mem_cgroup
*memcg
, *original_memcg
;
8056 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
8059 original_memcg
= get_mem_cgroup_from_objcg(objcg
);
8060 for (memcg
= original_memcg
; !mem_cgroup_is_root(memcg
);
8061 memcg
= parent_mem_cgroup(memcg
)) {
8062 unsigned long max
= READ_ONCE(memcg
->zswap_max
);
8063 unsigned long pages
;
8065 if (max
== PAGE_COUNTER_MAX
)
8072 cgroup_rstat_flush(memcg
->css
.cgroup
);
8073 pages
= memcg_page_state(memcg
, MEMCG_ZSWAP_B
) / PAGE_SIZE
;
8079 mem_cgroup_put(original_memcg
);
8084 * obj_cgroup_charge_zswap - charge compression backend memory
8085 * @objcg: the object cgroup
8086 * @size: size of compressed object
8088 * This forces the charge after obj_cgroup_may_zswap() allowed
8089 * compression and storage in zwap for this cgroup to go ahead.
8091 void obj_cgroup_charge_zswap(struct obj_cgroup
*objcg
, size_t size
)
8093 struct mem_cgroup
*memcg
;
8095 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
8098 VM_WARN_ON_ONCE(!(current
->flags
& PF_MEMALLOC
));
8100 /* PF_MEMALLOC context, charging must succeed */
8101 if (obj_cgroup_charge(objcg
, GFP_KERNEL
, size
))
8105 memcg
= obj_cgroup_memcg(objcg
);
8106 mod_memcg_state(memcg
, MEMCG_ZSWAP_B
, size
);
8107 mod_memcg_state(memcg
, MEMCG_ZSWAPPED
, 1);
8112 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
8113 * @objcg: the object cgroup
8114 * @size: size of compressed object
8116 * Uncharges zswap memory on page in.
8118 void obj_cgroup_uncharge_zswap(struct obj_cgroup
*objcg
, size_t size
)
8120 struct mem_cgroup
*memcg
;
8122 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys
))
8125 obj_cgroup_uncharge(objcg
, size
);
8128 memcg
= obj_cgroup_memcg(objcg
);
8129 mod_memcg_state(memcg
, MEMCG_ZSWAP_B
, -size
);
8130 mod_memcg_state(memcg
, MEMCG_ZSWAPPED
, -1);
8134 static u64
zswap_current_read(struct cgroup_subsys_state
*css
,
8137 cgroup_rstat_flush(css
->cgroup
);
8138 return memcg_page_state(mem_cgroup_from_css(css
), MEMCG_ZSWAP_B
);
8141 static int zswap_max_show(struct seq_file
*m
, void *v
)
8143 return seq_puts_memcg_tunable(m
,
8144 READ_ONCE(mem_cgroup_from_seq(m
)->zswap_max
));
8147 static ssize_t
zswap_max_write(struct kernfs_open_file
*of
,
8148 char *buf
, size_t nbytes
, loff_t off
)
8150 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
8154 buf
= strstrip(buf
);
8155 err
= page_counter_memparse(buf
, "max", &max
);
8159 xchg(&memcg
->zswap_max
, max
);
8164 static struct cftype zswap_files
[] = {
8166 .name
= "zswap.current",
8167 .flags
= CFTYPE_NOT_ON_ROOT
,
8168 .read_u64
= zswap_current_read
,
8171 .name
= "zswap.max",
8172 .flags
= CFTYPE_NOT_ON_ROOT
,
8173 .seq_show
= zswap_max_show
,
8174 .write
= zswap_max_write
,
8178 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
8180 static int __init
mem_cgroup_swap_init(void)
8182 if (mem_cgroup_disabled())
8185 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, swap_files
));
8186 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
, memsw_files
));
8187 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8188 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys
, zswap_files
));
8192 subsys_initcall(mem_cgroup_swap_init
);
8194 #endif /* CONFIG_SWAP */