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/pagevec.h>
37 #include <linux/vm_event_item.h>
38 #include <linux/smp.h>
39 #include <linux/page-flags.h>
40 #include <linux/backing-dev.h>
41 #include <linux/bit_spinlock.h>
42 #include <linux/rcupdate.h>
43 #include <linux/limits.h>
44 #include <linux/export.h>
45 #include <linux/mutex.h>
46 #include <linux/rbtree.h>
47 #include <linux/slab.h>
48 #include <linux/swap.h>
49 #include <linux/swapops.h>
50 #include <linux/spinlock.h>
51 #include <linux/eventfd.h>
52 #include <linux/poll.h>
53 #include <linux/sort.h>
55 #include <linux/seq_file.h>
56 #include <linux/vmpressure.h>
57 #include <linux/memremap.h>
58 #include <linux/mm_inline.h>
59 #include <linux/swap_cgroup.h>
60 #include <linux/cpu.h>
61 #include <linux/oom.h>
62 #include <linux/lockdep.h>
63 #include <linux/file.h>
64 #include <linux/resume_user_mode.h>
65 #include <linux/psi.h>
66 #include <linux/seq_buf.h>
67 #include <linux/sched/isolation.h>
68 #include <linux/kmemleak.h>
75 #include <linux/uaccess.h>
77 #include <trace/events/vmscan.h>
79 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
80 EXPORT_SYMBOL(memory_cgrp_subsys);
82 struct mem_cgroup *root_mem_cgroup __read_mostly;
84 /* Active memory cgroup to use from an interrupt context */
85 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
86 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
88 /* Socket memory accounting disabled? */
89 static bool cgroup_memory_nosocket __ro_after_init;
91 /* Kernel memory accounting disabled? */
92 static bool cgroup_memory_nokmem __ro_after_init;
94 /* BPF memory accounting disabled? */
95 static bool cgroup_memory_nobpf __ro_after_init;
97 #ifdef CONFIG_CGROUP_WRITEBACK
98 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
101 /* Whether legacy memory+swap accounting is active */
102 static bool do_memsw_account(void)
104 return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
107 #define THRESHOLDS_EVENTS_TARGET 128
108 #define SOFTLIMIT_EVENTS_TARGET 1024
111 * Cgroups above their limits are maintained in a RB-Tree, independent of
112 * their hierarchy representation
115 struct mem_cgroup_tree_per_node {
116 struct rb_root rb_root;
117 struct rb_node *rb_rightmost;
121 struct mem_cgroup_tree {
122 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
125 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
128 struct mem_cgroup_eventfd_list {
129 struct list_head list;
130 struct eventfd_ctx *eventfd;
134 * cgroup_event represents events which userspace want to receive.
136 struct mem_cgroup_event {
138 * memcg which the event belongs to.
140 struct mem_cgroup *memcg;
142 * eventfd to signal userspace about the event.
144 struct eventfd_ctx *eventfd;
146 * Each of these stored in a list by the cgroup.
148 struct list_head list;
150 * register_event() callback will be used to add new userspace
151 * waiter for changes related to this event. Use eventfd_signal()
152 * on eventfd to send notification to userspace.
154 int (*register_event)(struct mem_cgroup *memcg,
155 struct eventfd_ctx *eventfd, const char *args);
157 * unregister_event() callback will be called when userspace closes
158 * the eventfd or on cgroup removing. This callback must be set,
159 * if you want provide notification functionality.
161 void (*unregister_event)(struct mem_cgroup *memcg,
162 struct eventfd_ctx *eventfd);
164 * All fields below needed to unregister event when
165 * userspace closes eventfd.
168 wait_queue_head_t *wqh;
169 wait_queue_entry_t wait;
170 struct work_struct remove;
173 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
174 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
176 /* Stuffs for move charges at task migration. */
178 * Types of charges to be moved.
180 #define MOVE_ANON 0x1U
181 #define MOVE_FILE 0x2U
182 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
184 /* "mc" and its members are protected by cgroup_mutex */
185 static struct move_charge_struct {
186 spinlock_t lock; /* for from, to */
187 struct mm_struct *mm;
188 struct mem_cgroup *from;
189 struct mem_cgroup *to;
191 unsigned long precharge;
192 unsigned long moved_charge;
193 unsigned long moved_swap;
194 struct task_struct *moving_task; /* a task moving charges */
195 wait_queue_head_t waitq; /* a waitq for other context */
197 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
198 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
202 * Maximum loops in mem_cgroup_soft_reclaim(), used for soft
203 * limit reclaim to prevent infinite loops, if they ever occur.
205 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
206 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
208 /* for encoding cft->private value on file */
216 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
217 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
218 #define MEMFILE_ATTR(val) ((val) & 0xffff)
221 * Iteration constructs for visiting all cgroups (under a tree). If
222 * loops are exited prematurely (break), mem_cgroup_iter_break() must
223 * be used for reference counting.
225 #define for_each_mem_cgroup_tree(iter, root) \
226 for (iter = mem_cgroup_iter(root, NULL, NULL); \
228 iter = mem_cgroup_iter(root, iter, NULL))
230 #define for_each_mem_cgroup(iter) \
231 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
233 iter = mem_cgroup_iter(NULL, iter, NULL))
235 static inline bool task_is_dying(void)
237 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
238 (current->flags & PF_EXITING);
241 /* Some nice accessors for the vmpressure. */
242 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
245 memcg = root_mem_cgroup;
246 return &memcg->vmpressure;
249 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
251 return container_of(vmpr, struct mem_cgroup, vmpressure);
254 #define CURRENT_OBJCG_UPDATE_BIT 0
255 #define CURRENT_OBJCG_UPDATE_FLAG (1UL << CURRENT_OBJCG_UPDATE_BIT)
257 #ifdef CONFIG_MEMCG_KMEM
258 static DEFINE_SPINLOCK(objcg_lock);
260 bool mem_cgroup_kmem_disabled(void)
262 return cgroup_memory_nokmem;
265 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
266 unsigned int nr_pages);
268 static void obj_cgroup_release(struct percpu_ref *ref)
270 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
271 unsigned int nr_bytes;
272 unsigned int nr_pages;
276 * At this point all allocated objects are freed, and
277 * objcg->nr_charged_bytes can't have an arbitrary byte value.
278 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
280 * The following sequence can lead to it:
281 * 1) CPU0: objcg == stock->cached_objcg
282 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
283 * PAGE_SIZE bytes are charged
284 * 3) CPU1: a process from another memcg is allocating something,
285 * the stock if flushed,
286 * objcg->nr_charged_bytes = PAGE_SIZE - 92
287 * 5) CPU0: we do release this object,
288 * 92 bytes are added to stock->nr_bytes
289 * 6) CPU0: stock is flushed,
290 * 92 bytes are added to objcg->nr_charged_bytes
292 * In the result, nr_charged_bytes == PAGE_SIZE.
293 * This page will be uncharged in obj_cgroup_release().
295 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
296 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
297 nr_pages = nr_bytes >> PAGE_SHIFT;
300 obj_cgroup_uncharge_pages(objcg, nr_pages);
302 spin_lock_irqsave(&objcg_lock, flags);
303 list_del(&objcg->list);
304 spin_unlock_irqrestore(&objcg_lock, flags);
306 percpu_ref_exit(ref);
307 kfree_rcu(objcg, rcu);
310 static struct obj_cgroup *obj_cgroup_alloc(void)
312 struct obj_cgroup *objcg;
315 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
319 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
325 INIT_LIST_HEAD(&objcg->list);
329 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
330 struct mem_cgroup *parent)
332 struct obj_cgroup *objcg, *iter;
334 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
336 spin_lock_irq(&objcg_lock);
338 /* 1) Ready to reparent active objcg. */
339 list_add(&objcg->list, &memcg->objcg_list);
340 /* 2) Reparent active objcg and already reparented objcgs to parent. */
341 list_for_each_entry(iter, &memcg->objcg_list, list)
342 WRITE_ONCE(iter->memcg, parent);
343 /* 3) Move already reparented objcgs to the parent's list */
344 list_splice(&memcg->objcg_list, &parent->objcg_list);
346 spin_unlock_irq(&objcg_lock);
348 percpu_ref_kill(&objcg->refcnt);
352 * A lot of the calls to the cache allocation functions are expected to be
353 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
354 * conditional to this static branch, we'll have to allow modules that does
355 * kmem_cache_alloc and the such to see this symbol as well
357 DEFINE_STATIC_KEY_FALSE(memcg_kmem_online_key);
358 EXPORT_SYMBOL(memcg_kmem_online_key);
360 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key);
361 EXPORT_SYMBOL(memcg_bpf_enabled_key);
365 * mem_cgroup_css_from_folio - css of the memcg associated with a folio
366 * @folio: folio of interest
368 * If memcg is bound to the default hierarchy, css of the memcg associated
369 * with @folio is returned. The returned css remains associated with @folio
370 * until it is released.
372 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
375 struct cgroup_subsys_state *mem_cgroup_css_from_folio(struct folio *folio)
377 struct mem_cgroup *memcg = folio_memcg(folio);
379 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
380 memcg = root_mem_cgroup;
386 * page_cgroup_ino - return inode number of the memcg a page is charged to
389 * Look up the closest online ancestor of the memory cgroup @page is charged to
390 * and return its inode number or 0 if @page is not charged to any cgroup. It
391 * is safe to call this function without holding a reference to @page.
393 * Note, this function is inherently racy, because there is nothing to prevent
394 * the cgroup inode from getting torn down and potentially reallocated a moment
395 * after page_cgroup_ino() returns, so it only should be used by callers that
396 * do not care (such as procfs interfaces).
398 ino_t page_cgroup_ino(struct page *page)
400 struct mem_cgroup *memcg;
401 unsigned long ino = 0;
404 /* page_folio() is racy here, but the entire function is racy anyway */
405 memcg = folio_memcg_check(page_folio(page));
407 while (memcg && !(memcg->css.flags & CSS_ONLINE))
408 memcg = parent_mem_cgroup(memcg);
410 ino = cgroup_ino(memcg->css.cgroup);
415 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
416 struct mem_cgroup_tree_per_node *mctz,
417 unsigned long new_usage_in_excess)
419 struct rb_node **p = &mctz->rb_root.rb_node;
420 struct rb_node *parent = NULL;
421 struct mem_cgroup_per_node *mz_node;
422 bool rightmost = true;
427 mz->usage_in_excess = new_usage_in_excess;
428 if (!mz->usage_in_excess)
432 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
434 if (mz->usage_in_excess < mz_node->usage_in_excess) {
443 mctz->rb_rightmost = &mz->tree_node;
445 rb_link_node(&mz->tree_node, parent, p);
446 rb_insert_color(&mz->tree_node, &mctz->rb_root);
450 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
451 struct mem_cgroup_tree_per_node *mctz)
456 if (&mz->tree_node == mctz->rb_rightmost)
457 mctz->rb_rightmost = rb_prev(&mz->tree_node);
459 rb_erase(&mz->tree_node, &mctz->rb_root);
463 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
464 struct mem_cgroup_tree_per_node *mctz)
468 spin_lock_irqsave(&mctz->lock, flags);
469 __mem_cgroup_remove_exceeded(mz, mctz);
470 spin_unlock_irqrestore(&mctz->lock, flags);
473 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
475 unsigned long nr_pages = page_counter_read(&memcg->memory);
476 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
477 unsigned long excess = 0;
479 if (nr_pages > soft_limit)
480 excess = nr_pages - soft_limit;
485 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
487 unsigned long excess;
488 struct mem_cgroup_per_node *mz;
489 struct mem_cgroup_tree_per_node *mctz;
491 if (lru_gen_enabled()) {
492 if (soft_limit_excess(memcg))
493 lru_gen_soft_reclaim(memcg, nid);
497 mctz = soft_limit_tree.rb_tree_per_node[nid];
501 * Necessary to update all ancestors when hierarchy is used.
502 * because their event counter is not touched.
504 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
505 mz = memcg->nodeinfo[nid];
506 excess = soft_limit_excess(memcg);
508 * We have to update the tree if mz is on RB-tree or
509 * mem is over its softlimit.
511 if (excess || mz->on_tree) {
514 spin_lock_irqsave(&mctz->lock, flags);
515 /* if on-tree, remove it */
517 __mem_cgroup_remove_exceeded(mz, mctz);
519 * Insert again. mz->usage_in_excess will be updated.
520 * If excess is 0, no tree ops.
522 __mem_cgroup_insert_exceeded(mz, mctz, excess);
523 spin_unlock_irqrestore(&mctz->lock, flags);
528 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
530 struct mem_cgroup_tree_per_node *mctz;
531 struct mem_cgroup_per_node *mz;
535 mz = memcg->nodeinfo[nid];
536 mctz = soft_limit_tree.rb_tree_per_node[nid];
538 mem_cgroup_remove_exceeded(mz, mctz);
542 static struct mem_cgroup_per_node *
543 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
545 struct mem_cgroup_per_node *mz;
549 if (!mctz->rb_rightmost)
550 goto done; /* Nothing to reclaim from */
552 mz = rb_entry(mctz->rb_rightmost,
553 struct mem_cgroup_per_node, tree_node);
555 * Remove the node now but someone else can add it back,
556 * we will to add it back at the end of reclaim to its correct
557 * position in the tree.
559 __mem_cgroup_remove_exceeded(mz, mctz);
560 if (!soft_limit_excess(mz->memcg) ||
561 !css_tryget(&mz->memcg->css))
567 static struct mem_cgroup_per_node *
568 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
570 struct mem_cgroup_per_node *mz;
572 spin_lock_irq(&mctz->lock);
573 mz = __mem_cgroup_largest_soft_limit_node(mctz);
574 spin_unlock_irq(&mctz->lock);
578 /* Subset of vm_event_item to report for memcg event stats */
579 static const unsigned int memcg_vm_event_stat[] = {
595 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
600 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
608 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
609 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
611 static void init_memcg_events(void)
615 for (i = 0; i < NR_MEMCG_EVENTS; ++i)
616 mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
619 static inline int memcg_events_index(enum vm_event_item idx)
621 return mem_cgroup_events_index[idx] - 1;
624 struct memcg_vmstats_percpu {
625 /* Stats updates since the last flush */
626 unsigned int stats_updates;
628 /* Cached pointers for fast iteration in memcg_rstat_updated() */
629 struct memcg_vmstats_percpu *parent;
630 struct memcg_vmstats *vmstats;
632 /* The above should fit a single cacheline for memcg_rstat_updated() */
634 /* Local (CPU and cgroup) page state & events */
635 long state[MEMCG_NR_STAT];
636 unsigned long events[NR_MEMCG_EVENTS];
638 /* Delta calculation for lockless upward propagation */
639 long state_prev[MEMCG_NR_STAT];
640 unsigned long events_prev[NR_MEMCG_EVENTS];
642 /* Cgroup1: threshold notifications & softlimit tree updates */
643 unsigned long nr_page_events;
644 unsigned long targets[MEM_CGROUP_NTARGETS];
645 } ____cacheline_aligned;
647 struct memcg_vmstats {
648 /* Aggregated (CPU and subtree) page state & events */
649 long state[MEMCG_NR_STAT];
650 unsigned long events[NR_MEMCG_EVENTS];
652 /* Non-hierarchical (CPU aggregated) page state & events */
653 long state_local[MEMCG_NR_STAT];
654 unsigned long events_local[NR_MEMCG_EVENTS];
656 /* Pending child counts during tree propagation */
657 long state_pending[MEMCG_NR_STAT];
658 unsigned long events_pending[NR_MEMCG_EVENTS];
660 /* Stats updates since the last flush */
661 atomic64_t stats_updates;
665 * memcg and lruvec stats flushing
667 * Many codepaths leading to stats update or read are performance sensitive and
668 * adding stats flushing in such codepaths is not desirable. So, to optimize the
669 * flushing the kernel does:
671 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
672 * rstat update tree grow unbounded.
674 * 2) Flush the stats synchronously on reader side only when there are more than
675 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
676 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
677 * only for 2 seconds due to (1).
679 static void flush_memcg_stats_dwork(struct work_struct *w);
680 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
681 static u64 flush_last_time;
683 #define FLUSH_TIME (2UL*HZ)
686 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
687 * not rely on this as part of an acquired spinlock_t lock. These functions are
688 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
691 static void memcg_stats_lock(void)
693 preempt_disable_nested();
694 VM_WARN_ON_IRQS_ENABLED();
697 static void __memcg_stats_lock(void)
699 preempt_disable_nested();
702 static void memcg_stats_unlock(void)
704 preempt_enable_nested();
708 static bool memcg_vmstats_needs_flush(struct memcg_vmstats *vmstats)
710 return atomic64_read(&vmstats->stats_updates) >
711 MEMCG_CHARGE_BATCH * num_online_cpus();
714 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
716 struct memcg_vmstats_percpu *statc;
717 int cpu = smp_processor_id();
722 cgroup_rstat_updated(memcg->css.cgroup, cpu);
723 statc = this_cpu_ptr(memcg->vmstats_percpu);
724 for (; statc; statc = statc->parent) {
725 statc->stats_updates += abs(val);
726 if (statc->stats_updates < MEMCG_CHARGE_BATCH)
730 * If @memcg is already flush-able, increasing stats_updates is
731 * redundant. Avoid the overhead of the atomic update.
733 if (!memcg_vmstats_needs_flush(statc->vmstats))
734 atomic64_add(statc->stats_updates,
735 &statc->vmstats->stats_updates);
736 statc->stats_updates = 0;
740 static void do_flush_stats(struct mem_cgroup *memcg)
742 if (mem_cgroup_is_root(memcg))
743 WRITE_ONCE(flush_last_time, jiffies_64);
745 cgroup_rstat_flush(memcg->css.cgroup);
749 * mem_cgroup_flush_stats - flush the stats of a memory cgroup subtree
750 * @memcg: root of the subtree to flush
752 * Flushing is serialized by the underlying global rstat lock. There is also a
753 * minimum amount of work to be done even if there are no stat updates to flush.
754 * Hence, we only flush the stats if the updates delta exceeds a threshold. This
755 * avoids unnecessary work and contention on the underlying lock.
757 void mem_cgroup_flush_stats(struct mem_cgroup *memcg)
759 if (mem_cgroup_disabled())
763 memcg = root_mem_cgroup;
765 if (memcg_vmstats_needs_flush(memcg->vmstats))
766 do_flush_stats(memcg);
769 void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg)
771 /* Only flush if the periodic flusher is one full cycle late */
772 if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME))
773 mem_cgroup_flush_stats(memcg);
776 static void flush_memcg_stats_dwork(struct work_struct *w)
779 * Deliberately ignore memcg_vmstats_needs_flush() here so that flushing
780 * in latency-sensitive paths is as cheap as possible.
782 do_flush_stats(root_mem_cgroup);
783 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
786 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
788 long x = READ_ONCE(memcg->vmstats->state[idx]);
796 static int memcg_page_state_unit(int item);
799 * Normalize the value passed into memcg_rstat_updated() to be in pages. Round
800 * up non-zero sub-page updates to 1 page as zero page updates are ignored.
802 static int memcg_state_val_in_pages(int idx, int val)
804 int unit = memcg_page_state_unit(idx);
806 if (!val || unit == PAGE_SIZE)
809 return max(val * unit / PAGE_SIZE, 1UL);
813 * __mod_memcg_state - update cgroup memory statistics
814 * @memcg: the memory cgroup
815 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
816 * @val: delta to add to the counter, can be negative
818 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
820 if (mem_cgroup_disabled())
823 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
824 memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
827 /* idx can be of type enum memcg_stat_item or node_stat_item. */
828 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
830 long x = READ_ONCE(memcg->vmstats->state_local[idx]);
839 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
842 struct mem_cgroup_per_node *pn;
843 struct mem_cgroup *memcg;
845 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
849 * The caller from rmap relies on disabled preemption because they never
850 * update their counter from in-interrupt context. For these two
851 * counters we check that the update is never performed from an
852 * interrupt context while other caller need to have disabled interrupt.
854 __memcg_stats_lock();
855 if (IS_ENABLED(CONFIG_DEBUG_VM)) {
860 case NR_SHMEM_PMDMAPPED:
861 case NR_FILE_PMDMAPPED:
862 WARN_ON_ONCE(!in_task());
865 VM_WARN_ON_IRQS_ENABLED();
870 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
873 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
875 memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
876 memcg_stats_unlock();
880 * __mod_lruvec_state - update lruvec memory statistics
881 * @lruvec: the lruvec
882 * @idx: the stat item
883 * @val: delta to add to the counter, can be negative
885 * The lruvec is the intersection of the NUMA node and a cgroup. This
886 * function updates the all three counters that are affected by a
887 * change of state at this level: per-node, per-cgroup, per-lruvec.
889 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
893 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
895 /* Update memcg and lruvec */
896 if (!mem_cgroup_disabled())
897 __mod_memcg_lruvec_state(lruvec, idx, val);
900 void __lruvec_stat_mod_folio(struct folio *folio, enum node_stat_item idx,
903 struct mem_cgroup *memcg;
904 pg_data_t *pgdat = folio_pgdat(folio);
905 struct lruvec *lruvec;
908 memcg = folio_memcg(folio);
909 /* Untracked pages have no memcg, no lruvec. Update only the node */
912 __mod_node_page_state(pgdat, idx, val);
916 lruvec = mem_cgroup_lruvec(memcg, pgdat);
917 __mod_lruvec_state(lruvec, idx, val);
920 EXPORT_SYMBOL(__lruvec_stat_mod_folio);
922 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
924 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
925 struct mem_cgroup *memcg;
926 struct lruvec *lruvec;
929 memcg = mem_cgroup_from_slab_obj(p);
932 * Untracked pages have no memcg, no lruvec. Update only the
933 * node. If we reparent the slab objects to the root memcg,
934 * when we free the slab object, we need to update the per-memcg
935 * vmstats to keep it correct for the root memcg.
938 __mod_node_page_state(pgdat, idx, val);
940 lruvec = mem_cgroup_lruvec(memcg, pgdat);
941 __mod_lruvec_state(lruvec, idx, val);
947 * __count_memcg_events - account VM events in a cgroup
948 * @memcg: the memory cgroup
949 * @idx: the event item
950 * @count: the number of events that occurred
952 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
955 int index = memcg_events_index(idx);
957 if (mem_cgroup_disabled() || index < 0)
961 __this_cpu_add(memcg->vmstats_percpu->events[index], count);
962 memcg_rstat_updated(memcg, count);
963 memcg_stats_unlock();
966 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
968 int index = memcg_events_index(event);
972 return READ_ONCE(memcg->vmstats->events[index]);
975 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
977 int index = memcg_events_index(event);
982 return READ_ONCE(memcg->vmstats->events_local[index]);
985 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
988 /* pagein of a big page is an event. So, ignore page size */
990 __count_memcg_events(memcg, PGPGIN, 1);
992 __count_memcg_events(memcg, PGPGOUT, 1);
993 nr_pages = -nr_pages; /* for event */
996 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
999 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1000 enum mem_cgroup_events_target target)
1002 unsigned long val, next;
1004 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
1005 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
1006 /* from time_after() in jiffies.h */
1007 if ((long)(next - val) < 0) {
1009 case MEM_CGROUP_TARGET_THRESH:
1010 next = val + THRESHOLDS_EVENTS_TARGET;
1012 case MEM_CGROUP_TARGET_SOFTLIMIT:
1013 next = val + SOFTLIMIT_EVENTS_TARGET;
1018 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
1025 * Check events in order.
1028 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
1030 if (IS_ENABLED(CONFIG_PREEMPT_RT))
1033 /* threshold event is triggered in finer grain than soft limit */
1034 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1035 MEM_CGROUP_TARGET_THRESH))) {
1038 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1039 MEM_CGROUP_TARGET_SOFTLIMIT);
1040 mem_cgroup_threshold(memcg);
1041 if (unlikely(do_softlimit))
1042 mem_cgroup_update_tree(memcg, nid);
1046 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1049 * mm_update_next_owner() may clear mm->owner to NULL
1050 * if it races with swapoff, page migration, etc.
1051 * So this can be called with p == NULL.
1056 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1058 EXPORT_SYMBOL(mem_cgroup_from_task);
1060 static __always_inline struct mem_cgroup *active_memcg(void)
1063 return this_cpu_read(int_active_memcg);
1065 return current->active_memcg;
1069 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1070 * @mm: mm from which memcg should be extracted. It can be NULL.
1072 * Obtain a reference on mm->memcg and returns it if successful. If mm
1073 * is NULL, then the memcg is chosen as follows:
1074 * 1) The active memcg, if set.
1075 * 2) current->mm->memcg, if available
1077 * If mem_cgroup is disabled, NULL is returned.
1079 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1081 struct mem_cgroup *memcg;
1083 if (mem_cgroup_disabled())
1087 * Page cache insertions can happen without an
1088 * actual mm context, e.g. during disk probing
1089 * on boot, loopback IO, acct() writes etc.
1091 * No need to css_get on root memcg as the reference
1092 * counting is disabled on the root level in the
1093 * cgroup core. See CSS_NO_REF.
1095 if (unlikely(!mm)) {
1096 memcg = active_memcg();
1097 if (unlikely(memcg)) {
1098 /* remote memcg must hold a ref */
1099 css_get(&memcg->css);
1104 return root_mem_cgroup;
1109 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1110 if (unlikely(!memcg))
1111 memcg = root_mem_cgroup;
1112 } while (!css_tryget(&memcg->css));
1116 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1119 * get_mem_cgroup_from_current - Obtain a reference on current task's memcg.
1121 struct mem_cgroup *get_mem_cgroup_from_current(void)
1123 struct mem_cgroup *memcg;
1125 if (mem_cgroup_disabled())
1130 memcg = mem_cgroup_from_task(current);
1131 if (!css_tryget(&memcg->css)) {
1140 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1141 * @root: hierarchy root
1142 * @prev: previously returned memcg, NULL on first invocation
1143 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1145 * Returns references to children of the hierarchy below @root, or
1146 * @root itself, or %NULL after a full round-trip.
1148 * Caller must pass the return value in @prev on subsequent
1149 * invocations for reference counting, or use mem_cgroup_iter_break()
1150 * to cancel a hierarchy walk before the round-trip is complete.
1152 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1153 * in the hierarchy among all concurrent reclaimers operating on the
1156 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1157 struct mem_cgroup *prev,
1158 struct mem_cgroup_reclaim_cookie *reclaim)
1160 struct mem_cgroup_reclaim_iter *iter;
1161 struct cgroup_subsys_state *css = NULL;
1162 struct mem_cgroup *memcg = NULL;
1163 struct mem_cgroup *pos = NULL;
1165 if (mem_cgroup_disabled())
1169 root = root_mem_cgroup;
1174 struct mem_cgroup_per_node *mz;
1176 mz = root->nodeinfo[reclaim->pgdat->node_id];
1180 * On start, join the current reclaim iteration cycle.
1181 * Exit when a concurrent walker completes it.
1184 reclaim->generation = iter->generation;
1185 else if (reclaim->generation != iter->generation)
1189 pos = READ_ONCE(iter->position);
1190 if (!pos || css_tryget(&pos->css))
1193 * css reference reached zero, so iter->position will
1194 * be cleared by ->css_released. However, we should not
1195 * rely on this happening soon, because ->css_released
1196 * is called from a work queue, and by busy-waiting we
1197 * might block it. So we clear iter->position right
1200 (void)cmpxchg(&iter->position, pos, NULL);
1210 css = css_next_descendant_pre(css, &root->css);
1213 * Reclaimers share the hierarchy walk, and a
1214 * new one might jump in right at the end of
1215 * the hierarchy - make sure they see at least
1216 * one group and restart from the beginning.
1224 * Verify the css and acquire a reference. The root
1225 * is provided by the caller, so we know it's alive
1226 * and kicking, and don't take an extra reference.
1228 if (css == &root->css || css_tryget(css)) {
1229 memcg = mem_cgroup_from_css(css);
1236 * The position could have already been updated by a competing
1237 * thread, so check that the value hasn't changed since we read
1238 * it to avoid reclaiming from the same cgroup twice.
1240 (void)cmpxchg(&iter->position, pos, memcg);
1251 if (prev && prev != root)
1252 css_put(&prev->css);
1258 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1259 * @root: hierarchy root
1260 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1262 void mem_cgroup_iter_break(struct mem_cgroup *root,
1263 struct mem_cgroup *prev)
1266 root = root_mem_cgroup;
1267 if (prev && prev != root)
1268 css_put(&prev->css);
1271 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1272 struct mem_cgroup *dead_memcg)
1274 struct mem_cgroup_reclaim_iter *iter;
1275 struct mem_cgroup_per_node *mz;
1278 for_each_node(nid) {
1279 mz = from->nodeinfo[nid];
1281 cmpxchg(&iter->position, dead_memcg, NULL);
1285 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1287 struct mem_cgroup *memcg = dead_memcg;
1288 struct mem_cgroup *last;
1291 __invalidate_reclaim_iterators(memcg, dead_memcg);
1293 } while ((memcg = parent_mem_cgroup(memcg)));
1296 * When cgroup1 non-hierarchy mode is used,
1297 * parent_mem_cgroup() does not walk all the way up to the
1298 * cgroup root (root_mem_cgroup). So we have to handle
1299 * dead_memcg from cgroup root separately.
1301 if (!mem_cgroup_is_root(last))
1302 __invalidate_reclaim_iterators(root_mem_cgroup,
1307 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1308 * @memcg: hierarchy root
1309 * @fn: function to call for each task
1310 * @arg: argument passed to @fn
1312 * This function iterates over tasks attached to @memcg or to any of its
1313 * descendants and calls @fn for each task. If @fn returns a non-zero
1314 * value, the function breaks the iteration loop. Otherwise, it will iterate
1315 * over all tasks and return 0.
1317 * This function must not be called for the root memory cgroup.
1319 void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1320 int (*fn)(struct task_struct *, void *), void *arg)
1322 struct mem_cgroup *iter;
1325 BUG_ON(mem_cgroup_is_root(memcg));
1327 for_each_mem_cgroup_tree(iter, memcg) {
1328 struct css_task_iter it;
1329 struct task_struct *task;
1331 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1332 while (!ret && (task = css_task_iter_next(&it)))
1333 ret = fn(task, arg);
1334 css_task_iter_end(&it);
1336 mem_cgroup_iter_break(memcg, iter);
1342 #ifdef CONFIG_DEBUG_VM
1343 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1345 struct mem_cgroup *memcg;
1347 if (mem_cgroup_disabled())
1350 memcg = folio_memcg(folio);
1353 VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec)), folio);
1355 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1360 * folio_lruvec_lock - Lock the lruvec for a folio.
1361 * @folio: Pointer to the folio.
1363 * These functions are safe to use under any of the following conditions:
1365 * - folio_test_lru false
1366 * - folio_memcg_lock()
1367 * - folio frozen (refcount of 0)
1369 * Return: The lruvec this folio is on with its lock held.
1371 struct lruvec *folio_lruvec_lock(struct folio *folio)
1373 struct lruvec *lruvec = folio_lruvec(folio);
1375 spin_lock(&lruvec->lru_lock);
1376 lruvec_memcg_debug(lruvec, folio);
1382 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1383 * @folio: Pointer to the folio.
1385 * These functions are safe to use under any of the following conditions:
1387 * - folio_test_lru false
1388 * - folio_memcg_lock()
1389 * - folio frozen (refcount of 0)
1391 * Return: The lruvec this folio is on with its lock held and interrupts
1394 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1396 struct lruvec *lruvec = folio_lruvec(folio);
1398 spin_lock_irq(&lruvec->lru_lock);
1399 lruvec_memcg_debug(lruvec, folio);
1405 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1406 * @folio: Pointer to the folio.
1407 * @flags: Pointer to irqsave flags.
1409 * These functions are safe to use under any of the following conditions:
1411 * - folio_test_lru false
1412 * - folio_memcg_lock()
1413 * - folio frozen (refcount of 0)
1415 * Return: The lruvec this folio is on with its lock held and interrupts
1418 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1419 unsigned long *flags)
1421 struct lruvec *lruvec = folio_lruvec(folio);
1423 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1424 lruvec_memcg_debug(lruvec, folio);
1430 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1431 * @lruvec: mem_cgroup per zone lru vector
1432 * @lru: index of lru list the page is sitting on
1433 * @zid: zone id of the accounted pages
1434 * @nr_pages: positive when adding or negative when removing
1436 * This function must be called under lru_lock, just before a page is added
1437 * to or just after a page is removed from an lru list.
1439 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1440 int zid, int nr_pages)
1442 struct mem_cgroup_per_node *mz;
1443 unsigned long *lru_size;
1446 if (mem_cgroup_disabled())
1449 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1450 lru_size = &mz->lru_zone_size[zid][lru];
1453 *lru_size += nr_pages;
1456 if (WARN_ONCE(size < 0,
1457 "%s(%p, %d, %d): lru_size %ld\n",
1458 __func__, lruvec, lru, nr_pages, size)) {
1464 *lru_size += nr_pages;
1468 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1469 * @memcg: the memory cgroup
1471 * Returns the maximum amount of memory @mem can be charged with, in
1474 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1476 unsigned long margin = 0;
1477 unsigned long count;
1478 unsigned long limit;
1480 count = page_counter_read(&memcg->memory);
1481 limit = READ_ONCE(memcg->memory.max);
1483 margin = limit - count;
1485 if (do_memsw_account()) {
1486 count = page_counter_read(&memcg->memsw);
1487 limit = READ_ONCE(memcg->memsw.max);
1489 margin = min(margin, limit - count);
1498 * A routine for checking "mem" is under move_account() or not.
1500 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1501 * moving cgroups. This is for waiting at high-memory pressure
1504 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1506 struct mem_cgroup *from;
1507 struct mem_cgroup *to;
1510 * Unlike task_move routines, we access mc.to, mc.from not under
1511 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1513 spin_lock(&mc.lock);
1519 ret = mem_cgroup_is_descendant(from, memcg) ||
1520 mem_cgroup_is_descendant(to, memcg);
1522 spin_unlock(&mc.lock);
1526 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1528 if (mc.moving_task && current != mc.moving_task) {
1529 if (mem_cgroup_under_move(memcg)) {
1531 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1532 /* moving charge context might have finished. */
1535 finish_wait(&mc.waitq, &wait);
1542 struct memory_stat {
1547 static const struct memory_stat memory_stats[] = {
1548 { "anon", NR_ANON_MAPPED },
1549 { "file", NR_FILE_PAGES },
1550 { "kernel", MEMCG_KMEM },
1551 { "kernel_stack", NR_KERNEL_STACK_KB },
1552 { "pagetables", NR_PAGETABLE },
1553 { "sec_pagetables", NR_SECONDARY_PAGETABLE },
1554 { "percpu", MEMCG_PERCPU_B },
1555 { "sock", MEMCG_SOCK },
1556 { "vmalloc", MEMCG_VMALLOC },
1557 { "shmem", NR_SHMEM },
1558 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1559 { "zswap", MEMCG_ZSWAP_B },
1560 { "zswapped", MEMCG_ZSWAPPED },
1562 { "file_mapped", NR_FILE_MAPPED },
1563 { "file_dirty", NR_FILE_DIRTY },
1564 { "file_writeback", NR_WRITEBACK },
1566 { "swapcached", NR_SWAPCACHE },
1568 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1569 { "anon_thp", NR_ANON_THPS },
1570 { "file_thp", NR_FILE_THPS },
1571 { "shmem_thp", NR_SHMEM_THPS },
1573 { "inactive_anon", NR_INACTIVE_ANON },
1574 { "active_anon", NR_ACTIVE_ANON },
1575 { "inactive_file", NR_INACTIVE_FILE },
1576 { "active_file", NR_ACTIVE_FILE },
1577 { "unevictable", NR_UNEVICTABLE },
1578 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1579 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1581 /* The memory events */
1582 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1583 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1584 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1585 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1586 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1587 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1588 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1591 /* The actual unit of the state item, not the same as the output unit */
1592 static int memcg_page_state_unit(int item)
1595 case MEMCG_PERCPU_B:
1597 case NR_SLAB_RECLAIMABLE_B:
1598 case NR_SLAB_UNRECLAIMABLE_B:
1600 case NR_KERNEL_STACK_KB:
1607 /* Translate stat items to the correct unit for memory.stat output */
1608 static int memcg_page_state_output_unit(int item)
1611 * Workingset state is actually in pages, but we export it to userspace
1612 * as a scalar count of events, so special case it here.
1615 case WORKINGSET_REFAULT_ANON:
1616 case WORKINGSET_REFAULT_FILE:
1617 case WORKINGSET_ACTIVATE_ANON:
1618 case WORKINGSET_ACTIVATE_FILE:
1619 case WORKINGSET_RESTORE_ANON:
1620 case WORKINGSET_RESTORE_FILE:
1621 case WORKINGSET_NODERECLAIM:
1624 return memcg_page_state_unit(item);
1628 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1631 return memcg_page_state(memcg, item) *
1632 memcg_page_state_output_unit(item);
1635 static inline unsigned long memcg_page_state_local_output(
1636 struct mem_cgroup *memcg, int item)
1638 return memcg_page_state_local(memcg, item) *
1639 memcg_page_state_output_unit(item);
1642 static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1647 * Provide statistics on the state of the memory subsystem as
1648 * well as cumulative event counters that show past behavior.
1650 * This list is ordered following a combination of these gradients:
1651 * 1) generic big picture -> specifics and details
1652 * 2) reflecting userspace activity -> reflecting kernel heuristics
1654 * Current memory state:
1656 mem_cgroup_flush_stats(memcg);
1658 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1661 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1662 seq_buf_printf(s, "%s %llu\n", memory_stats[i].name, size);
1664 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1665 size += memcg_page_state_output(memcg,
1666 NR_SLAB_RECLAIMABLE_B);
1667 seq_buf_printf(s, "slab %llu\n", size);
1671 /* Accumulated memory events */
1672 seq_buf_printf(s, "pgscan %lu\n",
1673 memcg_events(memcg, PGSCAN_KSWAPD) +
1674 memcg_events(memcg, PGSCAN_DIRECT) +
1675 memcg_events(memcg, PGSCAN_KHUGEPAGED));
1676 seq_buf_printf(s, "pgsteal %lu\n",
1677 memcg_events(memcg, PGSTEAL_KSWAPD) +
1678 memcg_events(memcg, PGSTEAL_DIRECT) +
1679 memcg_events(memcg, PGSTEAL_KHUGEPAGED));
1681 for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1682 if (memcg_vm_event_stat[i] == PGPGIN ||
1683 memcg_vm_event_stat[i] == PGPGOUT)
1686 seq_buf_printf(s, "%s %lu\n",
1687 vm_event_name(memcg_vm_event_stat[i]),
1688 memcg_events(memcg, memcg_vm_event_stat[i]));
1691 /* The above should easily fit into one page */
1692 WARN_ON_ONCE(seq_buf_has_overflowed(s));
1695 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s);
1697 static void memory_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1699 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1700 memcg_stat_format(memcg, s);
1702 memcg1_stat_format(memcg, s);
1703 WARN_ON_ONCE(seq_buf_has_overflowed(s));
1707 * mem_cgroup_print_oom_context: Print OOM information relevant to
1708 * memory controller.
1709 * @memcg: The memory cgroup that went over limit
1710 * @p: Task that is going to be killed
1712 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1715 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1720 pr_cont(",oom_memcg=");
1721 pr_cont_cgroup_path(memcg->css.cgroup);
1723 pr_cont(",global_oom");
1725 pr_cont(",task_memcg=");
1726 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1732 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1733 * memory controller.
1734 * @memcg: The memory cgroup that went over limit
1736 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1738 /* Use static buffer, for the caller is holding oom_lock. */
1739 static char buf[PAGE_SIZE];
1742 lockdep_assert_held(&oom_lock);
1744 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1745 K((u64)page_counter_read(&memcg->memory)),
1746 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1747 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1748 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1749 K((u64)page_counter_read(&memcg->swap)),
1750 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1752 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1753 K((u64)page_counter_read(&memcg->memsw)),
1754 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1755 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1756 K((u64)page_counter_read(&memcg->kmem)),
1757 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1760 pr_info("Memory cgroup stats for ");
1761 pr_cont_cgroup_path(memcg->css.cgroup);
1763 seq_buf_init(&s, buf, sizeof(buf));
1764 memory_stat_format(memcg, &s);
1765 seq_buf_do_printk(&s, KERN_INFO);
1769 * Return the memory (and swap, if configured) limit for a memcg.
1771 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1773 unsigned long max = READ_ONCE(memcg->memory.max);
1775 if (do_memsw_account()) {
1776 if (mem_cgroup_swappiness(memcg)) {
1777 /* Calculate swap excess capacity from memsw limit */
1778 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1780 max += min(swap, (unsigned long)total_swap_pages);
1783 if (mem_cgroup_swappiness(memcg))
1784 max += min(READ_ONCE(memcg->swap.max),
1785 (unsigned long)total_swap_pages);
1790 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1792 return page_counter_read(&memcg->memory);
1795 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1798 struct oom_control oc = {
1802 .gfp_mask = gfp_mask,
1807 if (mutex_lock_killable(&oom_lock))
1810 if (mem_cgroup_margin(memcg) >= (1 << order))
1814 * A few threads which were not waiting at mutex_lock_killable() can
1815 * fail to bail out. Therefore, check again after holding oom_lock.
1817 ret = task_is_dying() || out_of_memory(&oc);
1820 mutex_unlock(&oom_lock);
1824 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1827 unsigned long *total_scanned)
1829 struct mem_cgroup *victim = NULL;
1832 unsigned long excess;
1833 unsigned long nr_scanned;
1834 struct mem_cgroup_reclaim_cookie reclaim = {
1838 excess = soft_limit_excess(root_memcg);
1841 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1846 * If we have not been able to reclaim
1847 * anything, it might because there are
1848 * no reclaimable pages under this hierarchy
1853 * We want to do more targeted reclaim.
1854 * excess >> 2 is not to excessive so as to
1855 * reclaim too much, nor too less that we keep
1856 * coming back to reclaim from this cgroup
1858 if (total >= (excess >> 2) ||
1859 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1864 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1865 pgdat, &nr_scanned);
1866 *total_scanned += nr_scanned;
1867 if (!soft_limit_excess(root_memcg))
1870 mem_cgroup_iter_break(root_memcg, victim);
1874 #ifdef CONFIG_LOCKDEP
1875 static struct lockdep_map memcg_oom_lock_dep_map = {
1876 .name = "memcg_oom_lock",
1880 static DEFINE_SPINLOCK(memcg_oom_lock);
1883 * Check OOM-Killer is already running under our hierarchy.
1884 * If someone is running, return false.
1886 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1888 struct mem_cgroup *iter, *failed = NULL;
1890 spin_lock(&memcg_oom_lock);
1892 for_each_mem_cgroup_tree(iter, memcg) {
1893 if (iter->oom_lock) {
1895 * this subtree of our hierarchy is already locked
1896 * so we cannot give a lock.
1899 mem_cgroup_iter_break(memcg, iter);
1902 iter->oom_lock = true;
1907 * OK, we failed to lock the whole subtree so we have
1908 * to clean up what we set up to the failing subtree
1910 for_each_mem_cgroup_tree(iter, memcg) {
1911 if (iter == failed) {
1912 mem_cgroup_iter_break(memcg, iter);
1915 iter->oom_lock = false;
1918 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1920 spin_unlock(&memcg_oom_lock);
1925 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1927 struct mem_cgroup *iter;
1929 spin_lock(&memcg_oom_lock);
1930 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1931 for_each_mem_cgroup_tree(iter, memcg)
1932 iter->oom_lock = false;
1933 spin_unlock(&memcg_oom_lock);
1936 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1938 struct mem_cgroup *iter;
1940 spin_lock(&memcg_oom_lock);
1941 for_each_mem_cgroup_tree(iter, memcg)
1943 spin_unlock(&memcg_oom_lock);
1946 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1948 struct mem_cgroup *iter;
1951 * Be careful about under_oom underflows because a child memcg
1952 * could have been added after mem_cgroup_mark_under_oom.
1954 spin_lock(&memcg_oom_lock);
1955 for_each_mem_cgroup_tree(iter, memcg)
1956 if (iter->under_oom > 0)
1958 spin_unlock(&memcg_oom_lock);
1961 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1963 struct oom_wait_info {
1964 struct mem_cgroup *memcg;
1965 wait_queue_entry_t wait;
1968 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1969 unsigned mode, int sync, void *arg)
1971 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1972 struct mem_cgroup *oom_wait_memcg;
1973 struct oom_wait_info *oom_wait_info;
1975 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1976 oom_wait_memcg = oom_wait_info->memcg;
1978 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1979 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1981 return autoremove_wake_function(wait, mode, sync, arg);
1984 static void memcg_oom_recover(struct mem_cgroup *memcg)
1987 * For the following lockless ->under_oom test, the only required
1988 * guarantee is that it must see the state asserted by an OOM when
1989 * this function is called as a result of userland actions
1990 * triggered by the notification of the OOM. This is trivially
1991 * achieved by invoking mem_cgroup_mark_under_oom() before
1992 * triggering notification.
1994 if (memcg && memcg->under_oom)
1995 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1999 * Returns true if successfully killed one or more processes. Though in some
2000 * corner cases it can return true even without killing any process.
2002 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2006 if (order > PAGE_ALLOC_COSTLY_ORDER)
2009 memcg_memory_event(memcg, MEMCG_OOM);
2012 * We are in the middle of the charge context here, so we
2013 * don't want to block when potentially sitting on a callstack
2014 * that holds all kinds of filesystem and mm locks.
2016 * cgroup1 allows disabling the OOM killer and waiting for outside
2017 * handling until the charge can succeed; remember the context and put
2018 * the task to sleep at the end of the page fault when all locks are
2021 * On the other hand, in-kernel OOM killer allows for an async victim
2022 * memory reclaim (oom_reaper) and that means that we are not solely
2023 * relying on the oom victim to make a forward progress and we can
2024 * invoke the oom killer here.
2026 * Please note that mem_cgroup_out_of_memory might fail to find a
2027 * victim and then we have to bail out from the charge path.
2029 if (READ_ONCE(memcg->oom_kill_disable)) {
2030 if (current->in_user_fault) {
2031 css_get(&memcg->css);
2032 current->memcg_in_oom = memcg;
2033 current->memcg_oom_gfp_mask = mask;
2034 current->memcg_oom_order = order;
2039 mem_cgroup_mark_under_oom(memcg);
2041 locked = mem_cgroup_oom_trylock(memcg);
2044 mem_cgroup_oom_notify(memcg);
2046 mem_cgroup_unmark_under_oom(memcg);
2047 ret = mem_cgroup_out_of_memory(memcg, mask, order);
2050 mem_cgroup_oom_unlock(memcg);
2056 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2057 * @handle: actually kill/wait or just clean up the OOM state
2059 * This has to be called at the end of a page fault if the memcg OOM
2060 * handler was enabled.
2062 * Memcg supports userspace OOM handling where failed allocations must
2063 * sleep on a waitqueue until the userspace task resolves the
2064 * situation. Sleeping directly in the charge context with all kinds
2065 * of locks held is not a good idea, instead we remember an OOM state
2066 * in the task and mem_cgroup_oom_synchronize() has to be called at
2067 * the end of the page fault to complete the OOM handling.
2069 * Returns %true if an ongoing memcg OOM situation was detected and
2070 * completed, %false otherwise.
2072 bool mem_cgroup_oom_synchronize(bool handle)
2074 struct mem_cgroup *memcg = current->memcg_in_oom;
2075 struct oom_wait_info owait;
2078 /* OOM is global, do not handle */
2085 owait.memcg = memcg;
2086 owait.wait.flags = 0;
2087 owait.wait.func = memcg_oom_wake_function;
2088 owait.wait.private = current;
2089 INIT_LIST_HEAD(&owait.wait.entry);
2091 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2092 mem_cgroup_mark_under_oom(memcg);
2094 locked = mem_cgroup_oom_trylock(memcg);
2097 mem_cgroup_oom_notify(memcg);
2100 mem_cgroup_unmark_under_oom(memcg);
2101 finish_wait(&memcg_oom_waitq, &owait.wait);
2104 mem_cgroup_oom_unlock(memcg);
2106 current->memcg_in_oom = NULL;
2107 css_put(&memcg->css);
2112 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2113 * @victim: task to be killed by the OOM killer
2114 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2116 * Returns a pointer to a memory cgroup, which has to be cleaned up
2117 * by killing all belonging OOM-killable tasks.
2119 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2121 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2122 struct mem_cgroup *oom_domain)
2124 struct mem_cgroup *oom_group = NULL;
2125 struct mem_cgroup *memcg;
2127 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2131 oom_domain = root_mem_cgroup;
2135 memcg = mem_cgroup_from_task(victim);
2136 if (mem_cgroup_is_root(memcg))
2140 * If the victim task has been asynchronously moved to a different
2141 * memory cgroup, we might end up killing tasks outside oom_domain.
2142 * In this case it's better to ignore memory.group.oom.
2144 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2148 * Traverse the memory cgroup hierarchy from the victim task's
2149 * cgroup up to the OOMing cgroup (or root) to find the
2150 * highest-level memory cgroup with oom.group set.
2152 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2153 if (READ_ONCE(memcg->oom_group))
2156 if (memcg == oom_domain)
2161 css_get(&oom_group->css);
2168 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2170 pr_info("Tasks in ");
2171 pr_cont_cgroup_path(memcg->css.cgroup);
2172 pr_cont(" are going to be killed due to memory.oom.group set\n");
2176 * folio_memcg_lock - Bind a folio to its memcg.
2177 * @folio: The folio.
2179 * This function prevents unlocked LRU folios from being moved to
2182 * It ensures lifetime of the bound memcg. The caller is responsible
2183 * for the lifetime of the folio.
2185 void folio_memcg_lock(struct folio *folio)
2187 struct mem_cgroup *memcg;
2188 unsigned long flags;
2191 * The RCU lock is held throughout the transaction. The fast
2192 * path can get away without acquiring the memcg->move_lock
2193 * because page moving starts with an RCU grace period.
2197 if (mem_cgroup_disabled())
2200 memcg = folio_memcg(folio);
2201 if (unlikely(!memcg))
2204 #ifdef CONFIG_PROVE_LOCKING
2205 local_irq_save(flags);
2206 might_lock(&memcg->move_lock);
2207 local_irq_restore(flags);
2210 if (atomic_read(&memcg->moving_account) <= 0)
2213 spin_lock_irqsave(&memcg->move_lock, flags);
2214 if (memcg != folio_memcg(folio)) {
2215 spin_unlock_irqrestore(&memcg->move_lock, flags);
2220 * When charge migration first begins, we can have multiple
2221 * critical sections holding the fast-path RCU lock and one
2222 * holding the slowpath move_lock. Track the task who has the
2223 * move_lock for folio_memcg_unlock().
2225 memcg->move_lock_task = current;
2226 memcg->move_lock_flags = flags;
2229 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2231 if (memcg && memcg->move_lock_task == current) {
2232 unsigned long flags = memcg->move_lock_flags;
2234 memcg->move_lock_task = NULL;
2235 memcg->move_lock_flags = 0;
2237 spin_unlock_irqrestore(&memcg->move_lock, flags);
2244 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2245 * @folio: The folio.
2247 * This releases the binding created by folio_memcg_lock(). This does
2248 * not change the accounting of this folio to its memcg, but it does
2249 * permit others to change it.
2251 void folio_memcg_unlock(struct folio *folio)
2253 __folio_memcg_unlock(folio_memcg(folio));
2256 struct memcg_stock_pcp {
2257 local_lock_t stock_lock;
2258 struct mem_cgroup *cached; /* this never be root cgroup */
2259 unsigned int nr_pages;
2261 #ifdef CONFIG_MEMCG_KMEM
2262 struct obj_cgroup *cached_objcg;
2263 struct pglist_data *cached_pgdat;
2264 unsigned int nr_bytes;
2265 int nr_slab_reclaimable_b;
2266 int nr_slab_unreclaimable_b;
2269 struct work_struct work;
2270 unsigned long flags;
2271 #define FLUSHING_CACHED_CHARGE 0
2273 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2274 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2276 static DEFINE_MUTEX(percpu_charge_mutex);
2278 #ifdef CONFIG_MEMCG_KMEM
2279 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2280 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2281 struct mem_cgroup *root_memcg);
2282 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2285 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2289 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2290 struct mem_cgroup *root_memcg)
2294 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2300 * consume_stock: Try to consume stocked charge on this cpu.
2301 * @memcg: memcg to consume from.
2302 * @nr_pages: how many pages to charge.
2304 * The charges will only happen if @memcg matches the current cpu's memcg
2305 * stock, and at least @nr_pages are available in that stock. Failure to
2306 * service an allocation will refill the stock.
2308 * returns true if successful, false otherwise.
2310 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2312 struct memcg_stock_pcp *stock;
2313 unsigned long flags;
2316 if (nr_pages > MEMCG_CHARGE_BATCH)
2319 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2321 stock = this_cpu_ptr(&memcg_stock);
2322 if (memcg == READ_ONCE(stock->cached) && stock->nr_pages >= nr_pages) {
2323 stock->nr_pages -= nr_pages;
2327 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2333 * Returns stocks cached in percpu and reset cached information.
2335 static void drain_stock(struct memcg_stock_pcp *stock)
2337 struct mem_cgroup *old = READ_ONCE(stock->cached);
2342 if (stock->nr_pages) {
2343 page_counter_uncharge(&old->memory, stock->nr_pages);
2344 if (do_memsw_account())
2345 page_counter_uncharge(&old->memsw, stock->nr_pages);
2346 stock->nr_pages = 0;
2350 WRITE_ONCE(stock->cached, NULL);
2353 static void drain_local_stock(struct work_struct *dummy)
2355 struct memcg_stock_pcp *stock;
2356 struct obj_cgroup *old = NULL;
2357 unsigned long flags;
2360 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2361 * drain_stock races is that we always operate on local CPU stock
2362 * here with IRQ disabled
2364 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2366 stock = this_cpu_ptr(&memcg_stock);
2367 old = drain_obj_stock(stock);
2369 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2371 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2373 obj_cgroup_put(old);
2377 * Cache charges(val) to local per_cpu area.
2378 * This will be consumed by consume_stock() function, later.
2380 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2382 struct memcg_stock_pcp *stock;
2384 stock = this_cpu_ptr(&memcg_stock);
2385 if (READ_ONCE(stock->cached) != memcg) { /* reset if necessary */
2387 css_get(&memcg->css);
2388 WRITE_ONCE(stock->cached, memcg);
2390 stock->nr_pages += nr_pages;
2392 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2396 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2398 unsigned long flags;
2400 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2401 __refill_stock(memcg, nr_pages);
2402 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2406 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2407 * of the hierarchy under it.
2409 static void drain_all_stock(struct mem_cgroup *root_memcg)
2413 /* If someone's already draining, avoid adding running more workers. */
2414 if (!mutex_trylock(&percpu_charge_mutex))
2417 * Notify other cpus that system-wide "drain" is running
2418 * We do not care about races with the cpu hotplug because cpu down
2419 * as well as workers from this path always operate on the local
2420 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2423 curcpu = smp_processor_id();
2424 for_each_online_cpu(cpu) {
2425 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2426 struct mem_cgroup *memcg;
2430 memcg = READ_ONCE(stock->cached);
2431 if (memcg && stock->nr_pages &&
2432 mem_cgroup_is_descendant(memcg, root_memcg))
2434 else if (obj_stock_flush_required(stock, root_memcg))
2439 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2441 drain_local_stock(&stock->work);
2442 else if (!cpu_is_isolated(cpu))
2443 schedule_work_on(cpu, &stock->work);
2447 mutex_unlock(&percpu_charge_mutex);
2450 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2452 struct memcg_stock_pcp *stock;
2454 stock = &per_cpu(memcg_stock, cpu);
2460 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2461 unsigned int nr_pages,
2464 unsigned long nr_reclaimed = 0;
2467 unsigned long pflags;
2469 if (page_counter_read(&memcg->memory) <=
2470 READ_ONCE(memcg->memory.high))
2473 memcg_memory_event(memcg, MEMCG_HIGH);
2475 psi_memstall_enter(&pflags);
2476 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2478 MEMCG_RECLAIM_MAY_SWAP);
2479 psi_memstall_leave(&pflags);
2480 } while ((memcg = parent_mem_cgroup(memcg)) &&
2481 !mem_cgroup_is_root(memcg));
2483 return nr_reclaimed;
2486 static void high_work_func(struct work_struct *work)
2488 struct mem_cgroup *memcg;
2490 memcg = container_of(work, struct mem_cgroup, high_work);
2491 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2495 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2496 * enough to still cause a significant slowdown in most cases, while still
2497 * allowing diagnostics and tracing to proceed without becoming stuck.
2499 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2502 * When calculating the delay, we use these either side of the exponentiation to
2503 * maintain precision and scale to a reasonable number of jiffies (see the table
2506 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2507 * overage ratio to a delay.
2508 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2509 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2510 * to produce a reasonable delay curve.
2512 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2513 * reasonable delay curve compared to precision-adjusted overage, not
2514 * penalising heavily at first, but still making sure that growth beyond the
2515 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2516 * example, with a high of 100 megabytes:
2518 * +-------+------------------------+
2519 * | usage | time to allocate in ms |
2520 * +-------+------------------------+
2542 * +-------+------------------------+
2544 #define MEMCG_DELAY_PRECISION_SHIFT 20
2545 #define MEMCG_DELAY_SCALING_SHIFT 14
2547 static u64 calculate_overage(unsigned long usage, unsigned long high)
2555 * Prevent division by 0 in overage calculation by acting as if
2556 * it was a threshold of 1 page
2558 high = max(high, 1UL);
2560 overage = usage - high;
2561 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2562 return div64_u64(overage, high);
2565 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2567 u64 overage, max_overage = 0;
2570 overage = calculate_overage(page_counter_read(&memcg->memory),
2571 READ_ONCE(memcg->memory.high));
2572 max_overage = max(overage, max_overage);
2573 } while ((memcg = parent_mem_cgroup(memcg)) &&
2574 !mem_cgroup_is_root(memcg));
2579 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2581 u64 overage, max_overage = 0;
2584 overage = calculate_overage(page_counter_read(&memcg->swap),
2585 READ_ONCE(memcg->swap.high));
2587 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2588 max_overage = max(overage, max_overage);
2589 } while ((memcg = parent_mem_cgroup(memcg)) &&
2590 !mem_cgroup_is_root(memcg));
2596 * Get the number of jiffies that we should penalise a mischievous cgroup which
2597 * is exceeding its memory.high by checking both it and its ancestors.
2599 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2600 unsigned int nr_pages,
2603 unsigned long penalty_jiffies;
2609 * We use overage compared to memory.high to calculate the number of
2610 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2611 * fairly lenient on small overages, and increasingly harsh when the
2612 * memcg in question makes it clear that it has no intention of stopping
2613 * its crazy behaviour, so we exponentially increase the delay based on
2616 penalty_jiffies = max_overage * max_overage * HZ;
2617 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2618 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2621 * Factor in the task's own contribution to the overage, such that four
2622 * N-sized allocations are throttled approximately the same as one
2623 * 4N-sized allocation.
2625 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2626 * larger the current charge patch is than that.
2628 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2632 * Reclaims memory over the high limit. Called directly from
2633 * try_charge() (context permitting), as well as from the userland
2634 * return path where reclaim is always able to block.
2636 void mem_cgroup_handle_over_high(gfp_t gfp_mask)
2638 unsigned long penalty_jiffies;
2639 unsigned long pflags;
2640 unsigned long nr_reclaimed;
2641 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2642 int nr_retries = MAX_RECLAIM_RETRIES;
2643 struct mem_cgroup *memcg;
2644 bool in_retry = false;
2646 if (likely(!nr_pages))
2649 memcg = get_mem_cgroup_from_mm(current->mm);
2650 current->memcg_nr_pages_over_high = 0;
2654 * Bail if the task is already exiting. Unlike memory.max,
2655 * memory.high enforcement isn't as strict, and there is no
2656 * OOM killer involved, which means the excess could already
2657 * be much bigger (and still growing) than it could for
2658 * memory.max; the dying task could get stuck in fruitless
2659 * reclaim for a long time, which isn't desirable.
2661 if (task_is_dying())
2665 * The allocating task should reclaim at least the batch size, but for
2666 * subsequent retries we only want to do what's necessary to prevent oom
2667 * or breaching resource isolation.
2669 * This is distinct from memory.max or page allocator behaviour because
2670 * memory.high is currently batched, whereas memory.max and the page
2671 * allocator run every time an allocation is made.
2673 nr_reclaimed = reclaim_high(memcg,
2674 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2678 * memory.high is breached and reclaim is unable to keep up. Throttle
2679 * allocators proactively to slow down excessive growth.
2681 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2682 mem_find_max_overage(memcg));
2684 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2685 swap_find_max_overage(memcg));
2688 * Clamp the max delay per usermode return so as to still keep the
2689 * application moving forwards and also permit diagnostics, albeit
2692 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2695 * Don't sleep if the amount of jiffies this memcg owes us is so low
2696 * that it's not even worth doing, in an attempt to be nice to those who
2697 * go only a small amount over their memory.high value and maybe haven't
2698 * been aggressively reclaimed enough yet.
2700 if (penalty_jiffies <= HZ / 100)
2704 * If reclaim is making forward progress but we're still over
2705 * memory.high, we want to encourage that rather than doing allocator
2708 if (nr_reclaimed || nr_retries--) {
2714 * Reclaim didn't manage to push usage below the limit, slow
2715 * this allocating task down.
2717 * If we exit early, we're guaranteed to die (since
2718 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2719 * need to account for any ill-begotten jiffies to pay them off later.
2721 psi_memstall_enter(&pflags);
2722 schedule_timeout_killable(penalty_jiffies);
2723 psi_memstall_leave(&pflags);
2726 css_put(&memcg->css);
2729 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2730 unsigned int nr_pages)
2732 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2733 int nr_retries = MAX_RECLAIM_RETRIES;
2734 struct mem_cgroup *mem_over_limit;
2735 struct page_counter *counter;
2736 unsigned long nr_reclaimed;
2737 bool passed_oom = false;
2738 unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2739 bool drained = false;
2740 bool raised_max_event = false;
2741 unsigned long pflags;
2744 if (consume_stock(memcg, nr_pages))
2747 if (!do_memsw_account() ||
2748 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2749 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2751 if (do_memsw_account())
2752 page_counter_uncharge(&memcg->memsw, batch);
2753 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2755 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2756 reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2759 if (batch > nr_pages) {
2765 * Prevent unbounded recursion when reclaim operations need to
2766 * allocate memory. This might exceed the limits temporarily,
2767 * but we prefer facilitating memory reclaim and getting back
2768 * under the limit over triggering OOM kills in these cases.
2770 if (unlikely(current->flags & PF_MEMALLOC))
2773 if (unlikely(task_in_memcg_oom(current)))
2776 if (!gfpflags_allow_blocking(gfp_mask))
2779 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2780 raised_max_event = true;
2782 psi_memstall_enter(&pflags);
2783 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2784 gfp_mask, reclaim_options);
2785 psi_memstall_leave(&pflags);
2787 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2791 drain_all_stock(mem_over_limit);
2796 if (gfp_mask & __GFP_NORETRY)
2799 * Even though the limit is exceeded at this point, reclaim
2800 * may have been able to free some pages. Retry the charge
2801 * before killing the task.
2803 * Only for regular pages, though: huge pages are rather
2804 * unlikely to succeed so close to the limit, and we fall back
2805 * to regular pages anyway in case of failure.
2807 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2810 * At task move, charge accounts can be doubly counted. So, it's
2811 * better to wait until the end of task_move if something is going on.
2813 if (mem_cgroup_wait_acct_move(mem_over_limit))
2819 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2822 /* Avoid endless loop for tasks bypassed by the oom killer */
2823 if (passed_oom && task_is_dying())
2827 * keep retrying as long as the memcg oom killer is able to make
2828 * a forward progress or bypass the charge if the oom killer
2829 * couldn't make any progress.
2831 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2832 get_order(nr_pages * PAGE_SIZE))) {
2834 nr_retries = MAX_RECLAIM_RETRIES;
2839 * Memcg doesn't have a dedicated reserve for atomic
2840 * allocations. But like the global atomic pool, we need to
2841 * put the burden of reclaim on regular allocation requests
2842 * and let these go through as privileged allocations.
2844 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2848 * If the allocation has to be enforced, don't forget to raise
2849 * a MEMCG_MAX event.
2851 if (!raised_max_event)
2852 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2855 * The allocation either can't fail or will lead to more memory
2856 * being freed very soon. Allow memory usage go over the limit
2857 * temporarily by force charging it.
2859 page_counter_charge(&memcg->memory, nr_pages);
2860 if (do_memsw_account())
2861 page_counter_charge(&memcg->memsw, nr_pages);
2866 if (batch > nr_pages)
2867 refill_stock(memcg, batch - nr_pages);
2870 * If the hierarchy is above the normal consumption range, schedule
2871 * reclaim on returning to userland. We can perform reclaim here
2872 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2873 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2874 * not recorded as it most likely matches current's and won't
2875 * change in the meantime. As high limit is checked again before
2876 * reclaim, the cost of mismatch is negligible.
2879 bool mem_high, swap_high;
2881 mem_high = page_counter_read(&memcg->memory) >
2882 READ_ONCE(memcg->memory.high);
2883 swap_high = page_counter_read(&memcg->swap) >
2884 READ_ONCE(memcg->swap.high);
2886 /* Don't bother a random interrupted task */
2889 schedule_work(&memcg->high_work);
2895 if (mem_high || swap_high) {
2897 * The allocating tasks in this cgroup will need to do
2898 * reclaim or be throttled to prevent further growth
2899 * of the memory or swap footprints.
2901 * Target some best-effort fairness between the tasks,
2902 * and distribute reclaim work and delay penalties
2903 * based on how much each task is actually allocating.
2905 current->memcg_nr_pages_over_high += batch;
2906 set_notify_resume(current);
2909 } while ((memcg = parent_mem_cgroup(memcg)));
2912 * Reclaim is set up above to be called from the userland
2913 * return path. But also attempt synchronous reclaim to avoid
2914 * excessive overrun while the task is still inside the
2915 * kernel. If this is successful, the return path will see it
2916 * when it rechecks the overage and simply bail out.
2918 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2919 !(current->flags & PF_MEMALLOC) &&
2920 gfpflags_allow_blocking(gfp_mask))
2921 mem_cgroup_handle_over_high(gfp_mask);
2925 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2926 unsigned int nr_pages)
2928 if (mem_cgroup_is_root(memcg))
2931 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2935 * mem_cgroup_cancel_charge() - cancel an uncommitted try_charge() call.
2936 * @memcg: memcg previously charged.
2937 * @nr_pages: number of pages previously charged.
2939 void mem_cgroup_cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2941 if (mem_cgroup_is_root(memcg))
2944 page_counter_uncharge(&memcg->memory, nr_pages);
2945 if (do_memsw_account())
2946 page_counter_uncharge(&memcg->memsw, nr_pages);
2949 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2951 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2953 * Any of the following ensures page's memcg stability:
2957 * - folio_memcg_lock()
2958 * - exclusive reference
2959 * - mem_cgroup_trylock_pages()
2961 folio->memcg_data = (unsigned long)memcg;
2965 * mem_cgroup_commit_charge - commit a previously successful try_charge().
2966 * @folio: folio to commit the charge to.
2967 * @memcg: memcg previously charged.
2969 void mem_cgroup_commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2971 css_get(&memcg->css);
2972 commit_charge(folio, memcg);
2974 local_irq_disable();
2975 mem_cgroup_charge_statistics(memcg, folio_nr_pages(folio));
2976 memcg_check_events(memcg, folio_nid(folio));
2980 #ifdef CONFIG_MEMCG_KMEM
2982 * The allocated objcg pointers array is not accounted directly.
2983 * Moreover, it should not come from DMA buffer and is not readily
2984 * reclaimable. So those GFP bits should be masked off.
2986 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
2987 __GFP_ACCOUNT | __GFP_NOFAIL)
2990 * mod_objcg_mlstate() may be called with irq enabled, so
2991 * mod_memcg_lruvec_state() should be used.
2993 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2994 struct pglist_data *pgdat,
2995 enum node_stat_item idx, int nr)
2997 struct mem_cgroup *memcg;
2998 struct lruvec *lruvec;
3001 memcg = obj_cgroup_memcg(objcg);
3002 lruvec = mem_cgroup_lruvec(memcg, pgdat);
3003 mod_memcg_lruvec_state(lruvec, idx, nr);
3007 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
3008 gfp_t gfp, bool new_slab)
3010 unsigned int objects = objs_per_slab(s, slab);
3011 unsigned long memcg_data;
3014 gfp &= ~OBJCGS_CLEAR_MASK;
3015 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
3020 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
3023 * If the slab is brand new and nobody can yet access its
3024 * memcg_data, no synchronization is required and memcg_data can
3025 * be simply assigned.
3027 slab->memcg_data = memcg_data;
3028 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
3030 * If the slab is already in use, somebody can allocate and
3031 * assign obj_cgroups in parallel. In this case the existing
3032 * objcg vector should be reused.
3038 kmemleak_not_leak(vec);
3042 static __always_inline
3043 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
3046 * Slab objects are accounted individually, not per-page.
3047 * Memcg membership data for each individual object is saved in
3050 if (folio_test_slab(folio)) {
3051 struct obj_cgroup **objcgs;
3055 slab = folio_slab(folio);
3056 objcgs = slab_objcgs(slab);
3060 off = obj_to_index(slab->slab_cache, slab, p);
3062 return obj_cgroup_memcg(objcgs[off]);
3068 * folio_memcg_check() is used here, because in theory we can encounter
3069 * a folio where the slab flag has been cleared already, but
3070 * slab->memcg_data has not been freed yet
3071 * folio_memcg_check() will guarantee that a proper memory
3072 * cgroup pointer or NULL will be returned.
3074 return folio_memcg_check(folio);
3078 * Returns a pointer to the memory cgroup to which the kernel object is charged.
3080 * A passed kernel object can be a slab object, vmalloc object or a generic
3081 * kernel page, so different mechanisms for getting the memory cgroup pointer
3084 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
3085 * can not know for sure how the kernel object is implemented.
3086 * mem_cgroup_from_obj() can be safely used in such cases.
3088 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3089 * cgroup_mutex, etc.
3091 struct mem_cgroup *mem_cgroup_from_obj(void *p)
3093 struct folio *folio;
3095 if (mem_cgroup_disabled())
3098 if (unlikely(is_vmalloc_addr(p)))
3099 folio = page_folio(vmalloc_to_page(p));
3101 folio = virt_to_folio(p);
3103 return mem_cgroup_from_obj_folio(folio, p);
3107 * Returns a pointer to the memory cgroup to which the kernel object is charged.
3108 * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
3109 * allocated using vmalloc().
3111 * A passed kernel object must be a slab object or a generic kernel page.
3113 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3114 * cgroup_mutex, etc.
3116 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
3118 if (mem_cgroup_disabled())
3121 return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
3124 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
3126 struct obj_cgroup *objcg = NULL;
3128 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3129 objcg = rcu_dereference(memcg->objcg);
3130 if (likely(objcg && obj_cgroup_tryget(objcg)))
3137 static struct obj_cgroup *current_objcg_update(void)
3139 struct mem_cgroup *memcg;
3140 struct obj_cgroup *old, *objcg = NULL;
3143 /* Atomically drop the update bit. */
3144 old = xchg(¤t->objcg, NULL);
3146 old = (struct obj_cgroup *)
3147 ((unsigned long)old & ~CURRENT_OBJCG_UPDATE_FLAG);
3149 obj_cgroup_put(old);
3154 /* If new objcg is NULL, no reason for the second atomic update. */
3155 if (!current->mm || (current->flags & PF_KTHREAD))
3159 * Release the objcg pointer from the previous iteration,
3160 * if try_cmpxcg() below fails.
3162 if (unlikely(objcg)) {
3163 obj_cgroup_put(objcg);
3168 * Obtain the new objcg pointer. The current task can be
3169 * asynchronously moved to another memcg and the previous
3170 * memcg can be offlined. So let's get the memcg pointer
3171 * and try get a reference to objcg under a rcu read lock.
3175 memcg = mem_cgroup_from_task(current);
3176 objcg = __get_obj_cgroup_from_memcg(memcg);
3180 * Try set up a new objcg pointer atomically. If it
3181 * fails, it means the update flag was set concurrently, so
3182 * the whole procedure should be repeated.
3184 } while (!try_cmpxchg(¤t->objcg, &old, objcg));
3189 __always_inline struct obj_cgroup *current_obj_cgroup(void)
3191 struct mem_cgroup *memcg;
3192 struct obj_cgroup *objcg;
3195 memcg = current->active_memcg;
3196 if (unlikely(memcg))
3199 objcg = READ_ONCE(current->objcg);
3200 if (unlikely((unsigned long)objcg & CURRENT_OBJCG_UPDATE_FLAG))
3201 objcg = current_objcg_update();
3203 * Objcg reference is kept by the task, so it's safe
3204 * to use the objcg by the current task.
3209 memcg = this_cpu_read(int_active_memcg);
3210 if (unlikely(memcg))
3217 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3219 * Memcg pointer is protected by scope (see set_active_memcg())
3220 * and is pinning the corresponding objcg, so objcg can't go
3221 * away and can be used within the scope without any additional
3224 objcg = rcu_dereference_check(memcg->objcg, 1);
3232 struct obj_cgroup *get_obj_cgroup_from_folio(struct folio *folio)
3234 struct obj_cgroup *objcg;
3236 if (!memcg_kmem_online())
3239 if (folio_memcg_kmem(folio)) {
3240 objcg = __folio_objcg(folio);
3241 obj_cgroup_get(objcg);
3243 struct mem_cgroup *memcg;
3246 memcg = __folio_memcg(folio);
3248 objcg = __get_obj_cgroup_from_memcg(memcg);
3256 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3258 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3259 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3261 page_counter_charge(&memcg->kmem, nr_pages);
3263 page_counter_uncharge(&memcg->kmem, -nr_pages);
3269 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3270 * @objcg: object cgroup to uncharge
3271 * @nr_pages: number of pages to uncharge
3273 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3274 unsigned int nr_pages)
3276 struct mem_cgroup *memcg;
3278 memcg = get_mem_cgroup_from_objcg(objcg);
3280 memcg_account_kmem(memcg, -nr_pages);
3281 refill_stock(memcg, nr_pages);
3283 css_put(&memcg->css);
3287 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3288 * @objcg: object cgroup to charge
3289 * @gfp: reclaim mode
3290 * @nr_pages: number of pages to charge
3292 * Returns 0 on success, an error code on failure.
3294 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3295 unsigned int nr_pages)
3297 struct mem_cgroup *memcg;
3300 memcg = get_mem_cgroup_from_objcg(objcg);
3302 ret = try_charge_memcg(memcg, gfp, nr_pages);
3306 memcg_account_kmem(memcg, nr_pages);
3308 css_put(&memcg->css);
3314 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3315 * @page: page to charge
3316 * @gfp: reclaim mode
3317 * @order: allocation order
3319 * Returns 0 on success, an error code on failure.
3321 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3323 struct obj_cgroup *objcg;
3326 objcg = current_obj_cgroup();
3328 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3330 obj_cgroup_get(objcg);
3331 page->memcg_data = (unsigned long)objcg |
3340 * __memcg_kmem_uncharge_page: uncharge a kmem page
3341 * @page: page to uncharge
3342 * @order: allocation order
3344 void __memcg_kmem_uncharge_page(struct page *page, int order)
3346 struct folio *folio = page_folio(page);
3347 struct obj_cgroup *objcg;
3348 unsigned int nr_pages = 1 << order;
3350 if (!folio_memcg_kmem(folio))
3353 objcg = __folio_objcg(folio);
3354 obj_cgroup_uncharge_pages(objcg, nr_pages);
3355 folio->memcg_data = 0;
3356 obj_cgroup_put(objcg);
3359 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3360 enum node_stat_item idx, int nr)
3362 struct memcg_stock_pcp *stock;
3363 struct obj_cgroup *old = NULL;
3364 unsigned long flags;
3367 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3368 stock = this_cpu_ptr(&memcg_stock);
3371 * Save vmstat data in stock and skip vmstat array update unless
3372 * accumulating over a page of vmstat data or when pgdat or idx
3375 if (READ_ONCE(stock->cached_objcg) != objcg) {
3376 old = drain_obj_stock(stock);
3377 obj_cgroup_get(objcg);
3378 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3379 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3380 WRITE_ONCE(stock->cached_objcg, objcg);
3381 stock->cached_pgdat = pgdat;
3382 } else if (stock->cached_pgdat != pgdat) {
3383 /* Flush the existing cached vmstat data */
3384 struct pglist_data *oldpg = stock->cached_pgdat;
3386 if (stock->nr_slab_reclaimable_b) {
3387 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3388 stock->nr_slab_reclaimable_b);
3389 stock->nr_slab_reclaimable_b = 0;
3391 if (stock->nr_slab_unreclaimable_b) {
3392 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3393 stock->nr_slab_unreclaimable_b);
3394 stock->nr_slab_unreclaimable_b = 0;
3396 stock->cached_pgdat = pgdat;
3399 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3400 : &stock->nr_slab_unreclaimable_b;
3402 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3403 * cached locally at least once before pushing it out.
3410 if (abs(*bytes) > PAGE_SIZE) {
3418 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3420 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3422 obj_cgroup_put(old);
3425 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3427 struct memcg_stock_pcp *stock;
3428 unsigned long flags;
3431 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3433 stock = this_cpu_ptr(&memcg_stock);
3434 if (objcg == READ_ONCE(stock->cached_objcg) && stock->nr_bytes >= nr_bytes) {
3435 stock->nr_bytes -= nr_bytes;
3439 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3444 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3446 struct obj_cgroup *old = READ_ONCE(stock->cached_objcg);
3451 if (stock->nr_bytes) {
3452 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3453 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3456 struct mem_cgroup *memcg;
3458 memcg = get_mem_cgroup_from_objcg(old);
3460 memcg_account_kmem(memcg, -nr_pages);
3461 __refill_stock(memcg, nr_pages);
3463 css_put(&memcg->css);
3467 * The leftover is flushed to the centralized per-memcg value.
3468 * On the next attempt to refill obj stock it will be moved
3469 * to a per-cpu stock (probably, on an other CPU), see
3470 * refill_obj_stock().
3472 * How often it's flushed is a trade-off between the memory
3473 * limit enforcement accuracy and potential CPU contention,
3474 * so it might be changed in the future.
3476 atomic_add(nr_bytes, &old->nr_charged_bytes);
3477 stock->nr_bytes = 0;
3481 * Flush the vmstat data in current stock
3483 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3484 if (stock->nr_slab_reclaimable_b) {
3485 mod_objcg_mlstate(old, stock->cached_pgdat,
3486 NR_SLAB_RECLAIMABLE_B,
3487 stock->nr_slab_reclaimable_b);
3488 stock->nr_slab_reclaimable_b = 0;
3490 if (stock->nr_slab_unreclaimable_b) {
3491 mod_objcg_mlstate(old, stock->cached_pgdat,
3492 NR_SLAB_UNRECLAIMABLE_B,
3493 stock->nr_slab_unreclaimable_b);
3494 stock->nr_slab_unreclaimable_b = 0;
3496 stock->cached_pgdat = NULL;
3499 WRITE_ONCE(stock->cached_objcg, NULL);
3501 * The `old' objects needs to be released by the caller via
3502 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3507 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3508 struct mem_cgroup *root_memcg)
3510 struct obj_cgroup *objcg = READ_ONCE(stock->cached_objcg);
3511 struct mem_cgroup *memcg;
3514 memcg = obj_cgroup_memcg(objcg);
3515 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3522 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3523 bool allow_uncharge)
3525 struct memcg_stock_pcp *stock;
3526 struct obj_cgroup *old = NULL;
3527 unsigned long flags;
3528 unsigned int nr_pages = 0;
3530 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3532 stock = this_cpu_ptr(&memcg_stock);
3533 if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
3534 old = drain_obj_stock(stock);
3535 obj_cgroup_get(objcg);
3536 WRITE_ONCE(stock->cached_objcg, objcg);
3537 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3538 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3539 allow_uncharge = true; /* Allow uncharge when objcg changes */
3541 stock->nr_bytes += nr_bytes;
3543 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3544 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3545 stock->nr_bytes &= (PAGE_SIZE - 1);
3548 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3550 obj_cgroup_put(old);
3553 obj_cgroup_uncharge_pages(objcg, nr_pages);
3556 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3558 unsigned int nr_pages, nr_bytes;
3561 if (consume_obj_stock(objcg, size))
3565 * In theory, objcg->nr_charged_bytes can have enough
3566 * pre-charged bytes to satisfy the allocation. However,
3567 * flushing objcg->nr_charged_bytes requires two atomic
3568 * operations, and objcg->nr_charged_bytes can't be big.
3569 * The shared objcg->nr_charged_bytes can also become a
3570 * performance bottleneck if all tasks of the same memcg are
3571 * trying to update it. So it's better to ignore it and try
3572 * grab some new pages. The stock's nr_bytes will be flushed to
3573 * objcg->nr_charged_bytes later on when objcg changes.
3575 * The stock's nr_bytes may contain enough pre-charged bytes
3576 * to allow one less page from being charged, but we can't rely
3577 * on the pre-charged bytes not being changed outside of
3578 * consume_obj_stock() or refill_obj_stock(). So ignore those
3579 * pre-charged bytes as well when charging pages. To avoid a
3580 * page uncharge right after a page charge, we set the
3581 * allow_uncharge flag to false when calling refill_obj_stock()
3582 * to temporarily allow the pre-charged bytes to exceed the page
3583 * size limit. The maximum reachable value of the pre-charged
3584 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3587 nr_pages = size >> PAGE_SHIFT;
3588 nr_bytes = size & (PAGE_SIZE - 1);
3593 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3594 if (!ret && nr_bytes)
3595 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3600 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3602 refill_obj_stock(objcg, size, true);
3605 #endif /* CONFIG_MEMCG_KMEM */
3608 * Because page_memcg(head) is not set on tails, set it now.
3610 void split_page_memcg(struct page *head, int old_order, int new_order)
3612 struct folio *folio = page_folio(head);
3613 struct mem_cgroup *memcg = folio_memcg(folio);
3615 unsigned int old_nr = 1 << old_order;
3616 unsigned int new_nr = 1 << new_order;
3618 if (mem_cgroup_disabled() || !memcg)
3621 for (i = new_nr; i < old_nr; i += new_nr)
3622 folio_page(folio, i)->memcg_data = folio->memcg_data;
3624 if (folio_memcg_kmem(folio))
3625 obj_cgroup_get_many(__folio_objcg(folio), old_nr / new_nr - 1);
3627 css_get_many(&memcg->css, old_nr / new_nr - 1);
3632 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3633 * @entry: swap entry to be moved
3634 * @from: mem_cgroup which the entry is moved from
3635 * @to: mem_cgroup which the entry is moved to
3637 * It succeeds only when the swap_cgroup's record for this entry is the same
3638 * as the mem_cgroup's id of @from.
3640 * Returns 0 on success, -EINVAL on failure.
3642 * The caller must have charged to @to, IOW, called page_counter_charge() about
3643 * both res and memsw, and called css_get().
3645 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3646 struct mem_cgroup *from, struct mem_cgroup *to)
3648 unsigned short old_id, new_id;
3650 old_id = mem_cgroup_id(from);
3651 new_id = mem_cgroup_id(to);
3653 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3654 mod_memcg_state(from, MEMCG_SWAP, -1);
3655 mod_memcg_state(to, MEMCG_SWAP, 1);
3661 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3662 struct mem_cgroup *from, struct mem_cgroup *to)
3668 static DEFINE_MUTEX(memcg_max_mutex);
3670 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3671 unsigned long max, bool memsw)
3673 bool enlarge = false;
3674 bool drained = false;
3676 bool limits_invariant;
3677 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3680 if (signal_pending(current)) {
3685 mutex_lock(&memcg_max_mutex);
3687 * Make sure that the new limit (memsw or memory limit) doesn't
3688 * break our basic invariant rule memory.max <= memsw.max.
3690 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3691 max <= memcg->memsw.max;
3692 if (!limits_invariant) {
3693 mutex_unlock(&memcg_max_mutex);
3697 if (max > counter->max)
3699 ret = page_counter_set_max(counter, max);
3700 mutex_unlock(&memcg_max_mutex);
3706 drain_all_stock(memcg);
3711 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3712 memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3718 if (!ret && enlarge)
3719 memcg_oom_recover(memcg);
3724 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3726 unsigned long *total_scanned)
3728 unsigned long nr_reclaimed = 0;
3729 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3730 unsigned long reclaimed;
3732 struct mem_cgroup_tree_per_node *mctz;
3733 unsigned long excess;
3735 if (lru_gen_enabled())
3741 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3744 * Do not even bother to check the largest node if the root
3745 * is empty. Do it lockless to prevent lock bouncing. Races
3746 * are acceptable as soft limit is best effort anyway.
3748 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3752 * This loop can run a while, specially if mem_cgroup's continuously
3753 * keep exceeding their soft limit and putting the system under
3760 mz = mem_cgroup_largest_soft_limit_node(mctz);
3764 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3765 gfp_mask, total_scanned);
3766 nr_reclaimed += reclaimed;
3767 spin_lock_irq(&mctz->lock);
3770 * If we failed to reclaim anything from this memory cgroup
3771 * it is time to move on to the next cgroup
3775 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3777 excess = soft_limit_excess(mz->memcg);
3779 * One school of thought says that we should not add
3780 * back the node to the tree if reclaim returns 0.
3781 * But our reclaim could return 0, simply because due
3782 * to priority we are exposing a smaller subset of
3783 * memory to reclaim from. Consider this as a longer
3786 /* If excess == 0, no tree ops */
3787 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3788 spin_unlock_irq(&mctz->lock);
3789 css_put(&mz->memcg->css);
3792 * Could not reclaim anything and there are no more
3793 * mem cgroups to try or we seem to be looping without
3794 * reclaiming anything.
3796 if (!nr_reclaimed &&
3798 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3800 } while (!nr_reclaimed);
3802 css_put(&next_mz->memcg->css);
3803 return nr_reclaimed;
3807 * Reclaims as many pages from the given memcg as possible.
3809 * Caller is responsible for holding css reference for memcg.
3811 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3813 int nr_retries = MAX_RECLAIM_RETRIES;
3815 /* we call try-to-free pages for make this cgroup empty */
3816 lru_add_drain_all();
3818 drain_all_stock(memcg);
3820 /* try to free all pages in this cgroup */
3821 while (nr_retries && page_counter_read(&memcg->memory)) {
3822 if (signal_pending(current))
3825 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3826 MEMCG_RECLAIM_MAY_SWAP))
3833 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3834 char *buf, size_t nbytes,
3837 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3839 if (mem_cgroup_is_root(memcg))
3841 return mem_cgroup_force_empty(memcg) ?: nbytes;
3844 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3850 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3851 struct cftype *cft, u64 val)
3856 pr_warn_once("Non-hierarchical mode is deprecated. "
3857 "Please report your usecase to linux-mm@kvack.org if you "
3858 "depend on this functionality.\n");
3863 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3867 if (mem_cgroup_is_root(memcg)) {
3869 * Approximate root's usage from global state. This isn't
3870 * perfect, but the root usage was always an approximation.
3872 val = global_node_page_state(NR_FILE_PAGES) +
3873 global_node_page_state(NR_ANON_MAPPED);
3875 val += total_swap_pages - get_nr_swap_pages();
3878 val = page_counter_read(&memcg->memory);
3880 val = page_counter_read(&memcg->memsw);
3893 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3897 struct page_counter *counter;
3899 switch (MEMFILE_TYPE(cft->private)) {
3901 counter = &memcg->memory;
3904 counter = &memcg->memsw;
3907 counter = &memcg->kmem;
3910 counter = &memcg->tcpmem;
3916 switch (MEMFILE_ATTR(cft->private)) {
3918 if (counter == &memcg->memory)
3919 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3920 if (counter == &memcg->memsw)
3921 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3922 return (u64)page_counter_read(counter) * PAGE_SIZE;
3924 return (u64)counter->max * PAGE_SIZE;
3926 return (u64)counter->watermark * PAGE_SIZE;
3928 return counter->failcnt;
3929 case RES_SOFT_LIMIT:
3930 return (u64)READ_ONCE(memcg->soft_limit) * PAGE_SIZE;
3937 * This function doesn't do anything useful. Its only job is to provide a read
3938 * handler for a file so that cgroup_file_mode() will add read permissions.
3940 static int mem_cgroup_dummy_seq_show(__always_unused struct seq_file *m,
3941 __always_unused void *v)
3946 #ifdef CONFIG_MEMCG_KMEM
3947 static int memcg_online_kmem(struct mem_cgroup *memcg)
3949 struct obj_cgroup *objcg;
3951 if (mem_cgroup_kmem_disabled())
3954 if (unlikely(mem_cgroup_is_root(memcg)))
3957 objcg = obj_cgroup_alloc();
3961 objcg->memcg = memcg;
3962 rcu_assign_pointer(memcg->objcg, objcg);
3963 obj_cgroup_get(objcg);
3964 memcg->orig_objcg = objcg;
3966 static_branch_enable(&memcg_kmem_online_key);
3968 memcg->kmemcg_id = memcg->id.id;
3973 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3975 struct mem_cgroup *parent;
3977 if (mem_cgroup_kmem_disabled())
3980 if (unlikely(mem_cgroup_is_root(memcg)))
3983 parent = parent_mem_cgroup(memcg);
3985 parent = root_mem_cgroup;
3987 memcg_reparent_objcgs(memcg, parent);
3990 * After we have finished memcg_reparent_objcgs(), all list_lrus
3991 * corresponding to this cgroup are guaranteed to remain empty.
3992 * The ordering is imposed by list_lru_node->lock taken by
3993 * memcg_reparent_list_lrus().
3995 memcg_reparent_list_lrus(memcg, parent);
3998 static int memcg_online_kmem(struct mem_cgroup *memcg)
4002 static void memcg_offline_kmem(struct mem_cgroup *memcg)
4005 #endif /* CONFIG_MEMCG_KMEM */
4007 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
4011 mutex_lock(&memcg_max_mutex);
4013 ret = page_counter_set_max(&memcg->tcpmem, max);
4017 if (!memcg->tcpmem_active) {
4019 * The active flag needs to be written after the static_key
4020 * update. This is what guarantees that the socket activation
4021 * function is the last one to run. See mem_cgroup_sk_alloc()
4022 * for details, and note that we don't mark any socket as
4023 * belonging to this memcg until that flag is up.
4025 * We need to do this, because static_keys will span multiple
4026 * sites, but we can't control their order. If we mark a socket
4027 * as accounted, but the accounting functions are not patched in
4028 * yet, we'll lose accounting.
4030 * We never race with the readers in mem_cgroup_sk_alloc(),
4031 * because when this value change, the code to process it is not
4034 static_branch_inc(&memcg_sockets_enabled_key);
4035 memcg->tcpmem_active = true;
4038 mutex_unlock(&memcg_max_mutex);
4043 * The user of this function is...
4046 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4047 char *buf, size_t nbytes, loff_t off)
4049 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4050 unsigned long nr_pages;
4053 buf = strstrip(buf);
4054 ret = page_counter_memparse(buf, "-1", &nr_pages);
4058 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4060 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4064 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4066 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
4069 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
4072 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
4073 "Writing any value to this file has no effect. "
4074 "Please report your usecase to linux-mm@kvack.org if you "
4075 "depend on this functionality.\n");
4079 ret = memcg_update_tcp_max(memcg, nr_pages);
4083 case RES_SOFT_LIMIT:
4084 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4087 WRITE_ONCE(memcg->soft_limit, nr_pages);
4092 return ret ?: nbytes;
4095 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4096 size_t nbytes, loff_t off)
4098 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4099 struct page_counter *counter;
4101 switch (MEMFILE_TYPE(of_cft(of)->private)) {
4103 counter = &memcg->memory;
4106 counter = &memcg->memsw;
4109 counter = &memcg->kmem;
4112 counter = &memcg->tcpmem;
4118 switch (MEMFILE_ATTR(of_cft(of)->private)) {
4120 page_counter_reset_watermark(counter);
4123 counter->failcnt = 0;
4132 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4135 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4139 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4140 struct cftype *cft, u64 val)
4142 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4144 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
4145 "Please report your usecase to linux-mm@kvack.org if you "
4146 "depend on this functionality.\n");
4148 if (val & ~MOVE_MASK)
4152 * No kind of locking is needed in here, because ->can_attach() will
4153 * check this value once in the beginning of the process, and then carry
4154 * on with stale data. This means that changes to this value will only
4155 * affect task migrations starting after the change.
4157 memcg->move_charge_at_immigrate = val;
4161 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4162 struct cftype *cft, u64 val)
4170 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
4171 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
4172 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
4174 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
4175 int nid, unsigned int lru_mask, bool tree)
4177 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
4178 unsigned long nr = 0;
4181 VM_BUG_ON((unsigned)nid >= nr_node_ids);
4184 if (!(BIT(lru) & lru_mask))
4187 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
4189 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
4194 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4195 unsigned int lru_mask,
4198 unsigned long nr = 0;
4202 if (!(BIT(lru) & lru_mask))
4205 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4207 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4212 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4216 unsigned int lru_mask;
4219 static const struct numa_stat stats[] = {
4220 { "total", LRU_ALL },
4221 { "file", LRU_ALL_FILE },
4222 { "anon", LRU_ALL_ANON },
4223 { "unevictable", BIT(LRU_UNEVICTABLE) },
4225 const struct numa_stat *stat;
4227 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4229 mem_cgroup_flush_stats(memcg);
4231 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4232 seq_printf(m, "%s=%lu", stat->name,
4233 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4235 for_each_node_state(nid, N_MEMORY)
4236 seq_printf(m, " N%d=%lu", nid,
4237 mem_cgroup_node_nr_lru_pages(memcg, nid,
4238 stat->lru_mask, false));
4242 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4244 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4245 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4247 for_each_node_state(nid, N_MEMORY)
4248 seq_printf(m, " N%d=%lu", nid,
4249 mem_cgroup_node_nr_lru_pages(memcg, nid,
4250 stat->lru_mask, true));
4256 #endif /* CONFIG_NUMA */
4258 static const unsigned int memcg1_stats[] = {
4261 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4268 WORKINGSET_REFAULT_ANON,
4269 WORKINGSET_REFAULT_FILE,
4276 static const char *const memcg1_stat_names[] = {
4279 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4286 "workingset_refault_anon",
4287 "workingset_refault_file",
4294 /* Universal VM events cgroup1 shows, original sort order */
4295 static const unsigned int memcg1_events[] = {
4302 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
4304 unsigned long memory, memsw;
4305 struct mem_cgroup *mi;
4308 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4310 mem_cgroup_flush_stats(memcg);
4312 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4315 nr = memcg_page_state_local_output(memcg, memcg1_stats[i]);
4316 seq_buf_printf(s, "%s %lu\n", memcg1_stat_names[i], nr);
4319 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4320 seq_buf_printf(s, "%s %lu\n", vm_event_name(memcg1_events[i]),
4321 memcg_events_local(memcg, memcg1_events[i]));
4323 for (i = 0; i < NR_LRU_LISTS; i++)
4324 seq_buf_printf(s, "%s %lu\n", lru_list_name(i),
4325 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4328 /* Hierarchical information */
4329 memory = memsw = PAGE_COUNTER_MAX;
4330 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4331 memory = min(memory, READ_ONCE(mi->memory.max));
4332 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4334 seq_buf_printf(s, "hierarchical_memory_limit %llu\n",
4335 (u64)memory * PAGE_SIZE);
4336 seq_buf_printf(s, "hierarchical_memsw_limit %llu\n",
4337 (u64)memsw * PAGE_SIZE);
4339 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4342 nr = memcg_page_state_output(memcg, memcg1_stats[i]);
4343 seq_buf_printf(s, "total_%s %llu\n", memcg1_stat_names[i],
4347 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4348 seq_buf_printf(s, "total_%s %llu\n",
4349 vm_event_name(memcg1_events[i]),
4350 (u64)memcg_events(memcg, memcg1_events[i]));
4352 for (i = 0; i < NR_LRU_LISTS; i++)
4353 seq_buf_printf(s, "total_%s %llu\n", lru_list_name(i),
4354 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4357 #ifdef CONFIG_DEBUG_VM
4360 struct mem_cgroup_per_node *mz;
4361 unsigned long anon_cost = 0;
4362 unsigned long file_cost = 0;
4364 for_each_online_pgdat(pgdat) {
4365 mz = memcg->nodeinfo[pgdat->node_id];
4367 anon_cost += mz->lruvec.anon_cost;
4368 file_cost += mz->lruvec.file_cost;
4370 seq_buf_printf(s, "anon_cost %lu\n", anon_cost);
4371 seq_buf_printf(s, "file_cost %lu\n", file_cost);
4376 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4381 return mem_cgroup_swappiness(memcg);
4384 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4385 struct cftype *cft, u64 val)
4387 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4392 if (!mem_cgroup_is_root(memcg))
4393 WRITE_ONCE(memcg->swappiness, val);
4395 WRITE_ONCE(vm_swappiness, val);
4400 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4402 struct mem_cgroup_threshold_ary *t;
4403 unsigned long usage;
4408 t = rcu_dereference(memcg->thresholds.primary);
4410 t = rcu_dereference(memcg->memsw_thresholds.primary);
4415 usage = mem_cgroup_usage(memcg, swap);
4418 * current_threshold points to threshold just below or equal to usage.
4419 * If it's not true, a threshold was crossed after last
4420 * call of __mem_cgroup_threshold().
4422 i = t->current_threshold;
4425 * Iterate backward over array of thresholds starting from
4426 * current_threshold and check if a threshold is crossed.
4427 * If none of thresholds below usage is crossed, we read
4428 * only one element of the array here.
4430 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4431 eventfd_signal(t->entries[i].eventfd);
4433 /* i = current_threshold + 1 */
4437 * Iterate forward over array of thresholds starting from
4438 * current_threshold+1 and check if a threshold is crossed.
4439 * If none of thresholds above usage is crossed, we read
4440 * only one element of the array here.
4442 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4443 eventfd_signal(t->entries[i].eventfd);
4445 /* Update current_threshold */
4446 t->current_threshold = i - 1;
4451 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4454 __mem_cgroup_threshold(memcg, false);
4455 if (do_memsw_account())
4456 __mem_cgroup_threshold(memcg, true);
4458 memcg = parent_mem_cgroup(memcg);
4462 static int compare_thresholds(const void *a, const void *b)
4464 const struct mem_cgroup_threshold *_a = a;
4465 const struct mem_cgroup_threshold *_b = b;
4467 if (_a->threshold > _b->threshold)
4470 if (_a->threshold < _b->threshold)
4476 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4478 struct mem_cgroup_eventfd_list *ev;
4480 spin_lock(&memcg_oom_lock);
4482 list_for_each_entry(ev, &memcg->oom_notify, list)
4483 eventfd_signal(ev->eventfd);
4485 spin_unlock(&memcg_oom_lock);
4489 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4491 struct mem_cgroup *iter;
4493 for_each_mem_cgroup_tree(iter, memcg)
4494 mem_cgroup_oom_notify_cb(iter);
4497 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4498 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4500 struct mem_cgroup_thresholds *thresholds;
4501 struct mem_cgroup_threshold_ary *new;
4502 unsigned long threshold;
4503 unsigned long usage;
4506 ret = page_counter_memparse(args, "-1", &threshold);
4510 mutex_lock(&memcg->thresholds_lock);
4513 thresholds = &memcg->thresholds;
4514 usage = mem_cgroup_usage(memcg, false);
4515 } else if (type == _MEMSWAP) {
4516 thresholds = &memcg->memsw_thresholds;
4517 usage = mem_cgroup_usage(memcg, true);
4521 /* Check if a threshold crossed before adding a new one */
4522 if (thresholds->primary)
4523 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4525 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4527 /* Allocate memory for new array of thresholds */
4528 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4535 /* Copy thresholds (if any) to new array */
4536 if (thresholds->primary)
4537 memcpy(new->entries, thresholds->primary->entries,
4538 flex_array_size(new, entries, size - 1));
4540 /* Add new threshold */
4541 new->entries[size - 1].eventfd = eventfd;
4542 new->entries[size - 1].threshold = threshold;
4544 /* Sort thresholds. Registering of new threshold isn't time-critical */
4545 sort(new->entries, size, sizeof(*new->entries),
4546 compare_thresholds, NULL);
4548 /* Find current threshold */
4549 new->current_threshold = -1;
4550 for (i = 0; i < size; i++) {
4551 if (new->entries[i].threshold <= usage) {
4553 * new->current_threshold will not be used until
4554 * rcu_assign_pointer(), so it's safe to increment
4557 ++new->current_threshold;
4562 /* Free old spare buffer and save old primary buffer as spare */
4563 kfree(thresholds->spare);
4564 thresholds->spare = thresholds->primary;
4566 rcu_assign_pointer(thresholds->primary, new);
4568 /* To be sure that nobody uses thresholds */
4572 mutex_unlock(&memcg->thresholds_lock);
4577 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4578 struct eventfd_ctx *eventfd, const char *args)
4580 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4583 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4584 struct eventfd_ctx *eventfd, const char *args)
4586 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4589 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4590 struct eventfd_ctx *eventfd, enum res_type type)
4592 struct mem_cgroup_thresholds *thresholds;
4593 struct mem_cgroup_threshold_ary *new;
4594 unsigned long usage;
4595 int i, j, size, entries;
4597 mutex_lock(&memcg->thresholds_lock);
4600 thresholds = &memcg->thresholds;
4601 usage = mem_cgroup_usage(memcg, false);
4602 } else if (type == _MEMSWAP) {
4603 thresholds = &memcg->memsw_thresholds;
4604 usage = mem_cgroup_usage(memcg, true);
4608 if (!thresholds->primary)
4611 /* Check if a threshold crossed before removing */
4612 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4614 /* Calculate new number of threshold */
4616 for (i = 0; i < thresholds->primary->size; i++) {
4617 if (thresholds->primary->entries[i].eventfd != eventfd)
4623 new = thresholds->spare;
4625 /* If no items related to eventfd have been cleared, nothing to do */
4629 /* Set thresholds array to NULL if we don't have thresholds */
4638 /* Copy thresholds and find current threshold */
4639 new->current_threshold = -1;
4640 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4641 if (thresholds->primary->entries[i].eventfd == eventfd)
4644 new->entries[j] = thresholds->primary->entries[i];
4645 if (new->entries[j].threshold <= usage) {
4647 * new->current_threshold will not be used
4648 * until rcu_assign_pointer(), so it's safe to increment
4651 ++new->current_threshold;
4657 /* Swap primary and spare array */
4658 thresholds->spare = thresholds->primary;
4660 rcu_assign_pointer(thresholds->primary, new);
4662 /* To be sure that nobody uses thresholds */
4665 /* If all events are unregistered, free the spare array */
4667 kfree(thresholds->spare);
4668 thresholds->spare = NULL;
4671 mutex_unlock(&memcg->thresholds_lock);
4674 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4675 struct eventfd_ctx *eventfd)
4677 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4680 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4681 struct eventfd_ctx *eventfd)
4683 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4686 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4687 struct eventfd_ctx *eventfd, const char *args)
4689 struct mem_cgroup_eventfd_list *event;
4691 event = kmalloc(sizeof(*event), GFP_KERNEL);
4695 spin_lock(&memcg_oom_lock);
4697 event->eventfd = eventfd;
4698 list_add(&event->list, &memcg->oom_notify);
4700 /* already in OOM ? */
4701 if (memcg->under_oom)
4702 eventfd_signal(eventfd);
4703 spin_unlock(&memcg_oom_lock);
4708 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4709 struct eventfd_ctx *eventfd)
4711 struct mem_cgroup_eventfd_list *ev, *tmp;
4713 spin_lock(&memcg_oom_lock);
4715 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4716 if (ev->eventfd == eventfd) {
4717 list_del(&ev->list);
4722 spin_unlock(&memcg_oom_lock);
4725 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4727 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4729 seq_printf(sf, "oom_kill_disable %d\n", READ_ONCE(memcg->oom_kill_disable));
4730 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4731 seq_printf(sf, "oom_kill %lu\n",
4732 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4736 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4737 struct cftype *cft, u64 val)
4739 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4741 /* cannot set to root cgroup and only 0 and 1 are allowed */
4742 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4745 WRITE_ONCE(memcg->oom_kill_disable, val);
4747 memcg_oom_recover(memcg);
4752 #ifdef CONFIG_CGROUP_WRITEBACK
4754 #include <trace/events/writeback.h>
4756 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4758 return wb_domain_init(&memcg->cgwb_domain, gfp);
4761 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4763 wb_domain_exit(&memcg->cgwb_domain);
4766 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4768 wb_domain_size_changed(&memcg->cgwb_domain);
4771 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4773 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4775 if (!memcg->css.parent)
4778 return &memcg->cgwb_domain;
4782 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4783 * @wb: bdi_writeback in question
4784 * @pfilepages: out parameter for number of file pages
4785 * @pheadroom: out parameter for number of allocatable pages according to memcg
4786 * @pdirty: out parameter for number of dirty pages
4787 * @pwriteback: out parameter for number of pages under writeback
4789 * Determine the numbers of file, headroom, dirty, and writeback pages in
4790 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4791 * is a bit more involved.
4793 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4794 * headroom is calculated as the lowest headroom of itself and the
4795 * ancestors. Note that this doesn't consider the actual amount of
4796 * available memory in the system. The caller should further cap
4797 * *@pheadroom accordingly.
4799 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4800 unsigned long *pheadroom, unsigned long *pdirty,
4801 unsigned long *pwriteback)
4803 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4804 struct mem_cgroup *parent;
4806 mem_cgroup_flush_stats_ratelimited(memcg);
4808 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4809 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4810 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4811 memcg_page_state(memcg, NR_ACTIVE_FILE);
4813 *pheadroom = PAGE_COUNTER_MAX;
4814 while ((parent = parent_mem_cgroup(memcg))) {
4815 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4816 READ_ONCE(memcg->memory.high));
4817 unsigned long used = page_counter_read(&memcg->memory);
4819 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4825 * Foreign dirty flushing
4827 * There's an inherent mismatch between memcg and writeback. The former
4828 * tracks ownership per-page while the latter per-inode. This was a
4829 * deliberate design decision because honoring per-page ownership in the
4830 * writeback path is complicated, may lead to higher CPU and IO overheads
4831 * and deemed unnecessary given that write-sharing an inode across
4832 * different cgroups isn't a common use-case.
4834 * Combined with inode majority-writer ownership switching, this works well
4835 * enough in most cases but there are some pathological cases. For
4836 * example, let's say there are two cgroups A and B which keep writing to
4837 * different but confined parts of the same inode. B owns the inode and
4838 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4839 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4840 * triggering background writeback. A will be slowed down without a way to
4841 * make writeback of the dirty pages happen.
4843 * Conditions like the above can lead to a cgroup getting repeatedly and
4844 * severely throttled after making some progress after each
4845 * dirty_expire_interval while the underlying IO device is almost
4848 * Solving this problem completely requires matching the ownership tracking
4849 * granularities between memcg and writeback in either direction. However,
4850 * the more egregious behaviors can be avoided by simply remembering the
4851 * most recent foreign dirtying events and initiating remote flushes on
4852 * them when local writeback isn't enough to keep the memory clean enough.
4854 * The following two functions implement such mechanism. When a foreign
4855 * page - a page whose memcg and writeback ownerships don't match - is
4856 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4857 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4858 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4859 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4860 * foreign bdi_writebacks which haven't expired. Both the numbers of
4861 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4862 * limited to MEMCG_CGWB_FRN_CNT.
4864 * The mechanism only remembers IDs and doesn't hold any object references.
4865 * As being wrong occasionally doesn't matter, updates and accesses to the
4866 * records are lockless and racy.
4868 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4869 struct bdi_writeback *wb)
4871 struct mem_cgroup *memcg = folio_memcg(folio);
4872 struct memcg_cgwb_frn *frn;
4873 u64 now = get_jiffies_64();
4874 u64 oldest_at = now;
4878 trace_track_foreign_dirty(folio, wb);
4881 * Pick the slot to use. If there is already a slot for @wb, keep
4882 * using it. If not replace the oldest one which isn't being
4885 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4886 frn = &memcg->cgwb_frn[i];
4887 if (frn->bdi_id == wb->bdi->id &&
4888 frn->memcg_id == wb->memcg_css->id)
4890 if (time_before64(frn->at, oldest_at) &&
4891 atomic_read(&frn->done.cnt) == 1) {
4893 oldest_at = frn->at;
4897 if (i < MEMCG_CGWB_FRN_CNT) {
4899 * Re-using an existing one. Update timestamp lazily to
4900 * avoid making the cacheline hot. We want them to be
4901 * reasonably up-to-date and significantly shorter than
4902 * dirty_expire_interval as that's what expires the record.
4903 * Use the shorter of 1s and dirty_expire_interval / 8.
4905 unsigned long update_intv =
4906 min_t(unsigned long, HZ,
4907 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4909 if (time_before64(frn->at, now - update_intv))
4911 } else if (oldest >= 0) {
4912 /* replace the oldest free one */
4913 frn = &memcg->cgwb_frn[oldest];
4914 frn->bdi_id = wb->bdi->id;
4915 frn->memcg_id = wb->memcg_css->id;
4920 /* issue foreign writeback flushes for recorded foreign dirtying events */
4921 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4923 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4924 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4925 u64 now = jiffies_64;
4928 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4929 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4932 * If the record is older than dirty_expire_interval,
4933 * writeback on it has already started. No need to kick it
4934 * off again. Also, don't start a new one if there's
4935 * already one in flight.
4937 if (time_after64(frn->at, now - intv) &&
4938 atomic_read(&frn->done.cnt) == 1) {
4940 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4941 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4942 WB_REASON_FOREIGN_FLUSH,
4948 #else /* CONFIG_CGROUP_WRITEBACK */
4950 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4955 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4959 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4963 #endif /* CONFIG_CGROUP_WRITEBACK */
4966 * DO NOT USE IN NEW FILES.
4968 * "cgroup.event_control" implementation.
4970 * This is way over-engineered. It tries to support fully configurable
4971 * events for each user. Such level of flexibility is completely
4972 * unnecessary especially in the light of the planned unified hierarchy.
4974 * Please deprecate this and replace with something simpler if at all
4979 * Unregister event and free resources.
4981 * Gets called from workqueue.
4983 static void memcg_event_remove(struct work_struct *work)
4985 struct mem_cgroup_event *event =
4986 container_of(work, struct mem_cgroup_event, remove);
4987 struct mem_cgroup *memcg = event->memcg;
4989 remove_wait_queue(event->wqh, &event->wait);
4991 event->unregister_event(memcg, event->eventfd);
4993 /* Notify userspace the event is going away. */
4994 eventfd_signal(event->eventfd);
4996 eventfd_ctx_put(event->eventfd);
4998 css_put(&memcg->css);
5002 * Gets called on EPOLLHUP on eventfd when user closes it.
5004 * Called with wqh->lock held and interrupts disabled.
5006 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
5007 int sync, void *key)
5009 struct mem_cgroup_event *event =
5010 container_of(wait, struct mem_cgroup_event, wait);
5011 struct mem_cgroup *memcg = event->memcg;
5012 __poll_t flags = key_to_poll(key);
5014 if (flags & EPOLLHUP) {
5016 * If the event has been detached at cgroup removal, we
5017 * can simply return knowing the other side will cleanup
5020 * We can't race against event freeing since the other
5021 * side will require wqh->lock via remove_wait_queue(),
5024 spin_lock(&memcg->event_list_lock);
5025 if (!list_empty(&event->list)) {
5026 list_del_init(&event->list);
5028 * We are in atomic context, but cgroup_event_remove()
5029 * may sleep, so we have to call it in workqueue.
5031 schedule_work(&event->remove);
5033 spin_unlock(&memcg->event_list_lock);
5039 static void memcg_event_ptable_queue_proc(struct file *file,
5040 wait_queue_head_t *wqh, poll_table *pt)
5042 struct mem_cgroup_event *event =
5043 container_of(pt, struct mem_cgroup_event, pt);
5046 add_wait_queue(wqh, &event->wait);
5050 * DO NOT USE IN NEW FILES.
5052 * Parse input and register new cgroup event handler.
5054 * Input must be in format '<event_fd> <control_fd> <args>'.
5055 * Interpretation of args is defined by control file implementation.
5057 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5058 char *buf, size_t nbytes, loff_t off)
5060 struct cgroup_subsys_state *css = of_css(of);
5061 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5062 struct mem_cgroup_event *event;
5063 struct cgroup_subsys_state *cfile_css;
5064 unsigned int efd, cfd;
5067 struct dentry *cdentry;
5072 if (IS_ENABLED(CONFIG_PREEMPT_RT))
5075 buf = strstrip(buf);
5077 efd = simple_strtoul(buf, &endp, 10);
5082 cfd = simple_strtoul(buf, &endp, 10);
5083 if ((*endp != ' ') && (*endp != '\0'))
5087 event = kzalloc(sizeof(*event), GFP_KERNEL);
5091 event->memcg = memcg;
5092 INIT_LIST_HEAD(&event->list);
5093 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5094 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5095 INIT_WORK(&event->remove, memcg_event_remove);
5103 event->eventfd = eventfd_ctx_fileget(efile.file);
5104 if (IS_ERR(event->eventfd)) {
5105 ret = PTR_ERR(event->eventfd);
5112 goto out_put_eventfd;
5115 /* the process need read permission on control file */
5116 /* AV: shouldn't we check that it's been opened for read instead? */
5117 ret = file_permission(cfile.file, MAY_READ);
5122 * The control file must be a regular cgroup1 file. As a regular cgroup
5123 * file can't be renamed, it's safe to access its name afterwards.
5125 cdentry = cfile.file->f_path.dentry;
5126 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
5132 * Determine the event callbacks and set them in @event. This used
5133 * to be done via struct cftype but cgroup core no longer knows
5134 * about these events. The following is crude but the whole thing
5135 * is for compatibility anyway.
5137 * DO NOT ADD NEW FILES.
5139 name = cdentry->d_name.name;
5141 if (!strcmp(name, "memory.usage_in_bytes")) {
5142 event->register_event = mem_cgroup_usage_register_event;
5143 event->unregister_event = mem_cgroup_usage_unregister_event;
5144 } else if (!strcmp(name, "memory.oom_control")) {
5145 event->register_event = mem_cgroup_oom_register_event;
5146 event->unregister_event = mem_cgroup_oom_unregister_event;
5147 } else if (!strcmp(name, "memory.pressure_level")) {
5148 event->register_event = vmpressure_register_event;
5149 event->unregister_event = vmpressure_unregister_event;
5150 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5151 event->register_event = memsw_cgroup_usage_register_event;
5152 event->unregister_event = memsw_cgroup_usage_unregister_event;
5159 * Verify @cfile should belong to @css. Also, remaining events are
5160 * automatically removed on cgroup destruction but the removal is
5161 * asynchronous, so take an extra ref on @css.
5163 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
5164 &memory_cgrp_subsys);
5166 if (IS_ERR(cfile_css))
5168 if (cfile_css != css) {
5173 ret = event->register_event(memcg, event->eventfd, buf);
5177 vfs_poll(efile.file, &event->pt);
5179 spin_lock_irq(&memcg->event_list_lock);
5180 list_add(&event->list, &memcg->event_list);
5181 spin_unlock_irq(&memcg->event_list_lock);
5193 eventfd_ctx_put(event->eventfd);
5202 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_SLUB_DEBUG)
5203 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
5207 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
5213 static int memory_stat_show(struct seq_file *m, void *v);
5215 static struct cftype mem_cgroup_legacy_files[] = {
5217 .name = "usage_in_bytes",
5218 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5219 .read_u64 = mem_cgroup_read_u64,
5222 .name = "max_usage_in_bytes",
5223 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5224 .write = mem_cgroup_reset,
5225 .read_u64 = mem_cgroup_read_u64,
5228 .name = "limit_in_bytes",
5229 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5230 .write = mem_cgroup_write,
5231 .read_u64 = mem_cgroup_read_u64,
5234 .name = "soft_limit_in_bytes",
5235 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5236 .write = mem_cgroup_write,
5237 .read_u64 = mem_cgroup_read_u64,
5241 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5242 .write = mem_cgroup_reset,
5243 .read_u64 = mem_cgroup_read_u64,
5247 .seq_show = memory_stat_show,
5250 .name = "force_empty",
5251 .write = mem_cgroup_force_empty_write,
5254 .name = "use_hierarchy",
5255 .write_u64 = mem_cgroup_hierarchy_write,
5256 .read_u64 = mem_cgroup_hierarchy_read,
5259 .name = "cgroup.event_control", /* XXX: for compat */
5260 .write = memcg_write_event_control,
5261 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5264 .name = "swappiness",
5265 .read_u64 = mem_cgroup_swappiness_read,
5266 .write_u64 = mem_cgroup_swappiness_write,
5269 .name = "move_charge_at_immigrate",
5270 .read_u64 = mem_cgroup_move_charge_read,
5271 .write_u64 = mem_cgroup_move_charge_write,
5274 .name = "oom_control",
5275 .seq_show = mem_cgroup_oom_control_read,
5276 .write_u64 = mem_cgroup_oom_control_write,
5279 .name = "pressure_level",
5280 .seq_show = mem_cgroup_dummy_seq_show,
5284 .name = "numa_stat",
5285 .seq_show = memcg_numa_stat_show,
5289 .name = "kmem.limit_in_bytes",
5290 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5291 .write = mem_cgroup_write,
5292 .read_u64 = mem_cgroup_read_u64,
5295 .name = "kmem.usage_in_bytes",
5296 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5297 .read_u64 = mem_cgroup_read_u64,
5300 .name = "kmem.failcnt",
5301 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5302 .write = mem_cgroup_reset,
5303 .read_u64 = mem_cgroup_read_u64,
5306 .name = "kmem.max_usage_in_bytes",
5307 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5308 .write = mem_cgroup_reset,
5309 .read_u64 = mem_cgroup_read_u64,
5311 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_SLUB_DEBUG)
5313 .name = "kmem.slabinfo",
5314 .seq_show = mem_cgroup_slab_show,
5318 .name = "kmem.tcp.limit_in_bytes",
5319 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5320 .write = mem_cgroup_write,
5321 .read_u64 = mem_cgroup_read_u64,
5324 .name = "kmem.tcp.usage_in_bytes",
5325 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5326 .read_u64 = mem_cgroup_read_u64,
5329 .name = "kmem.tcp.failcnt",
5330 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5331 .write = mem_cgroup_reset,
5332 .read_u64 = mem_cgroup_read_u64,
5335 .name = "kmem.tcp.max_usage_in_bytes",
5336 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5337 .write = mem_cgroup_reset,
5338 .read_u64 = mem_cgroup_read_u64,
5340 { }, /* terminate */
5344 * Private memory cgroup IDR
5346 * Swap-out records and page cache shadow entries need to store memcg
5347 * references in constrained space, so we maintain an ID space that is
5348 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5349 * memory-controlled cgroups to 64k.
5351 * However, there usually are many references to the offline CSS after
5352 * the cgroup has been destroyed, such as page cache or reclaimable
5353 * slab objects, that don't need to hang on to the ID. We want to keep
5354 * those dead CSS from occupying IDs, or we might quickly exhaust the
5355 * relatively small ID space and prevent the creation of new cgroups
5356 * even when there are much fewer than 64k cgroups - possibly none.
5358 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5359 * be freed and recycled when it's no longer needed, which is usually
5360 * when the CSS is offlined.
5362 * The only exception to that are records of swapped out tmpfs/shmem
5363 * pages that need to be attributed to live ancestors on swapin. But
5364 * those references are manageable from userspace.
5367 #define MEM_CGROUP_ID_MAX ((1UL << MEM_CGROUP_ID_SHIFT) - 1)
5368 static DEFINE_IDR(mem_cgroup_idr);
5370 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5372 if (memcg->id.id > 0) {
5373 idr_remove(&mem_cgroup_idr, memcg->id.id);
5378 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5381 refcount_add(n, &memcg->id.ref);
5384 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5386 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5387 mem_cgroup_id_remove(memcg);
5389 /* Memcg ID pins CSS */
5390 css_put(&memcg->css);
5394 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5396 mem_cgroup_id_put_many(memcg, 1);
5400 * mem_cgroup_from_id - look up a memcg from a memcg id
5401 * @id: the memcg id to look up
5403 * Caller must hold rcu_read_lock().
5405 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5407 WARN_ON_ONCE(!rcu_read_lock_held());
5408 return idr_find(&mem_cgroup_idr, id);
5411 #ifdef CONFIG_SHRINKER_DEBUG
5412 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5414 struct cgroup *cgrp;
5415 struct cgroup_subsys_state *css;
5416 struct mem_cgroup *memcg;
5418 cgrp = cgroup_get_from_id(ino);
5420 return ERR_CAST(cgrp);
5422 css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5424 memcg = container_of(css, struct mem_cgroup, css);
5426 memcg = ERR_PTR(-ENOENT);
5434 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5436 struct mem_cgroup_per_node *pn;
5438 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5442 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5443 GFP_KERNEL_ACCOUNT);
5444 if (!pn->lruvec_stats_percpu) {
5449 lruvec_init(&pn->lruvec);
5452 memcg->nodeinfo[node] = pn;
5456 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5458 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5463 free_percpu(pn->lruvec_stats_percpu);
5467 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5471 if (memcg->orig_objcg)
5472 obj_cgroup_put(memcg->orig_objcg);
5475 free_mem_cgroup_per_node_info(memcg, node);
5476 kfree(memcg->vmstats);
5477 free_percpu(memcg->vmstats_percpu);
5481 static void mem_cgroup_free(struct mem_cgroup *memcg)
5483 lru_gen_exit_memcg(memcg);
5484 memcg_wb_domain_exit(memcg);
5485 __mem_cgroup_free(memcg);
5488 static struct mem_cgroup *mem_cgroup_alloc(struct mem_cgroup *parent)
5490 struct memcg_vmstats_percpu *statc, *pstatc;
5491 struct mem_cgroup *memcg;
5493 int __maybe_unused i;
5494 long error = -ENOMEM;
5496 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5498 return ERR_PTR(error);
5500 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5501 1, MEM_CGROUP_ID_MAX + 1, GFP_KERNEL);
5502 if (memcg->id.id < 0) {
5503 error = memcg->id.id;
5507 memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5508 if (!memcg->vmstats)
5511 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5512 GFP_KERNEL_ACCOUNT);
5513 if (!memcg->vmstats_percpu)
5516 for_each_possible_cpu(cpu) {
5518 pstatc = per_cpu_ptr(parent->vmstats_percpu, cpu);
5519 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5520 statc->parent = parent ? pstatc : NULL;
5521 statc->vmstats = memcg->vmstats;
5525 if (alloc_mem_cgroup_per_node_info(memcg, node))
5528 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5531 INIT_WORK(&memcg->high_work, high_work_func);
5532 INIT_LIST_HEAD(&memcg->oom_notify);
5533 mutex_init(&memcg->thresholds_lock);
5534 spin_lock_init(&memcg->move_lock);
5535 vmpressure_init(&memcg->vmpressure);
5536 INIT_LIST_HEAD(&memcg->event_list);
5537 spin_lock_init(&memcg->event_list_lock);
5538 memcg->socket_pressure = jiffies;
5539 #ifdef CONFIG_MEMCG_KMEM
5540 memcg->kmemcg_id = -1;
5541 INIT_LIST_HEAD(&memcg->objcg_list);
5543 #ifdef CONFIG_CGROUP_WRITEBACK
5544 INIT_LIST_HEAD(&memcg->cgwb_list);
5545 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5546 memcg->cgwb_frn[i].done =
5547 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5549 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5550 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5551 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5552 memcg->deferred_split_queue.split_queue_len = 0;
5554 lru_gen_init_memcg(memcg);
5557 mem_cgroup_id_remove(memcg);
5558 __mem_cgroup_free(memcg);
5559 return ERR_PTR(error);
5562 static struct cgroup_subsys_state * __ref
5563 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5565 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5566 struct mem_cgroup *memcg, *old_memcg;
5568 old_memcg = set_active_memcg(parent);
5569 memcg = mem_cgroup_alloc(parent);
5570 set_active_memcg(old_memcg);
5572 return ERR_CAST(memcg);
5574 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5575 WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5576 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5577 memcg->zswap_max = PAGE_COUNTER_MAX;
5578 WRITE_ONCE(memcg->zswap_writeback,
5579 !parent || READ_ONCE(parent->zswap_writeback));
5581 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5583 WRITE_ONCE(memcg->swappiness, mem_cgroup_swappiness(parent));
5584 WRITE_ONCE(memcg->oom_kill_disable, READ_ONCE(parent->oom_kill_disable));
5586 page_counter_init(&memcg->memory, &parent->memory);
5587 page_counter_init(&memcg->swap, &parent->swap);
5588 page_counter_init(&memcg->kmem, &parent->kmem);
5589 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5591 init_memcg_events();
5592 page_counter_init(&memcg->memory, NULL);
5593 page_counter_init(&memcg->swap, NULL);
5594 page_counter_init(&memcg->kmem, NULL);
5595 page_counter_init(&memcg->tcpmem, NULL);
5597 root_mem_cgroup = memcg;
5601 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5602 static_branch_inc(&memcg_sockets_enabled_key);
5604 #if defined(CONFIG_MEMCG_KMEM)
5605 if (!cgroup_memory_nobpf)
5606 static_branch_inc(&memcg_bpf_enabled_key);
5612 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5614 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5616 if (memcg_online_kmem(memcg))
5620 * A memcg must be visible for expand_shrinker_info()
5621 * by the time the maps are allocated. So, we allocate maps
5622 * here, when for_each_mem_cgroup() can't skip it.
5624 if (alloc_shrinker_info(memcg))
5627 if (unlikely(mem_cgroup_is_root(memcg)) && !mem_cgroup_disabled())
5628 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5630 lru_gen_online_memcg(memcg);
5632 /* Online state pins memcg ID, memcg ID pins CSS */
5633 refcount_set(&memcg->id.ref, 1);
5637 * Ensure mem_cgroup_from_id() works once we're fully online.
5639 * We could do this earlier and require callers to filter with
5640 * css_tryget_online(). But right now there are no users that
5641 * need earlier access, and the workingset code relies on the
5642 * cgroup tree linkage (mem_cgroup_get_nr_swap_pages()). So
5643 * publish it here at the end of onlining. This matches the
5644 * regular ID destruction during offlining.
5646 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5650 memcg_offline_kmem(memcg);
5652 mem_cgroup_id_remove(memcg);
5656 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5658 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5659 struct mem_cgroup_event *event, *tmp;
5662 * Unregister events and notify userspace.
5663 * Notify userspace about cgroup removing only after rmdir of cgroup
5664 * directory to avoid race between userspace and kernelspace.
5666 spin_lock_irq(&memcg->event_list_lock);
5667 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5668 list_del_init(&event->list);
5669 schedule_work(&event->remove);
5671 spin_unlock_irq(&memcg->event_list_lock);
5673 page_counter_set_min(&memcg->memory, 0);
5674 page_counter_set_low(&memcg->memory, 0);
5676 zswap_memcg_offline_cleanup(memcg);
5678 memcg_offline_kmem(memcg);
5679 reparent_shrinker_deferred(memcg);
5680 wb_memcg_offline(memcg);
5681 lru_gen_offline_memcg(memcg);
5683 drain_all_stock(memcg);
5685 mem_cgroup_id_put(memcg);
5688 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5690 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5692 invalidate_reclaim_iterators(memcg);
5693 lru_gen_release_memcg(memcg);
5696 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5698 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5699 int __maybe_unused i;
5701 #ifdef CONFIG_CGROUP_WRITEBACK
5702 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5703 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5705 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5706 static_branch_dec(&memcg_sockets_enabled_key);
5708 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5709 static_branch_dec(&memcg_sockets_enabled_key);
5711 #if defined(CONFIG_MEMCG_KMEM)
5712 if (!cgroup_memory_nobpf)
5713 static_branch_dec(&memcg_bpf_enabled_key);
5716 vmpressure_cleanup(&memcg->vmpressure);
5717 cancel_work_sync(&memcg->high_work);
5718 mem_cgroup_remove_from_trees(memcg);
5719 free_shrinker_info(memcg);
5720 mem_cgroup_free(memcg);
5724 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5725 * @css: the target css
5727 * Reset the states of the mem_cgroup associated with @css. This is
5728 * invoked when the userland requests disabling on the default hierarchy
5729 * but the memcg is pinned through dependency. The memcg should stop
5730 * applying policies and should revert to the vanilla state as it may be
5731 * made visible again.
5733 * The current implementation only resets the essential configurations.
5734 * This needs to be expanded to cover all the visible parts.
5736 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5738 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5740 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5741 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5742 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5743 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5744 page_counter_set_min(&memcg->memory, 0);
5745 page_counter_set_low(&memcg->memory, 0);
5746 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5747 WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5748 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5749 memcg_wb_domain_size_changed(memcg);
5752 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5754 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5755 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5756 struct memcg_vmstats_percpu *statc;
5757 long delta, delta_cpu, v;
5760 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5762 for (i = 0; i < MEMCG_NR_STAT; i++) {
5764 * Collect the aggregated propagation counts of groups
5765 * below us. We're in a per-cpu loop here and this is
5766 * a global counter, so the first cycle will get them.
5768 delta = memcg->vmstats->state_pending[i];
5770 memcg->vmstats->state_pending[i] = 0;
5772 /* Add CPU changes on this level since the last flush */
5774 v = READ_ONCE(statc->state[i]);
5775 if (v != statc->state_prev[i]) {
5776 delta_cpu = v - statc->state_prev[i];
5778 statc->state_prev[i] = v;
5781 /* Aggregate counts on this level and propagate upwards */
5783 memcg->vmstats->state_local[i] += delta_cpu;
5786 memcg->vmstats->state[i] += delta;
5788 parent->vmstats->state_pending[i] += delta;
5792 for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5793 delta = memcg->vmstats->events_pending[i];
5795 memcg->vmstats->events_pending[i] = 0;
5798 v = READ_ONCE(statc->events[i]);
5799 if (v != statc->events_prev[i]) {
5800 delta_cpu = v - statc->events_prev[i];
5802 statc->events_prev[i] = v;
5806 memcg->vmstats->events_local[i] += delta_cpu;
5809 memcg->vmstats->events[i] += delta;
5811 parent->vmstats->events_pending[i] += delta;
5815 for_each_node_state(nid, N_MEMORY) {
5816 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5817 struct mem_cgroup_per_node *ppn = NULL;
5818 struct lruvec_stats_percpu *lstatc;
5821 ppn = parent->nodeinfo[nid];
5823 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5825 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5826 delta = pn->lruvec_stats.state_pending[i];
5828 pn->lruvec_stats.state_pending[i] = 0;
5831 v = READ_ONCE(lstatc->state[i]);
5832 if (v != lstatc->state_prev[i]) {
5833 delta_cpu = v - lstatc->state_prev[i];
5835 lstatc->state_prev[i] = v;
5839 pn->lruvec_stats.state_local[i] += delta_cpu;
5842 pn->lruvec_stats.state[i] += delta;
5844 ppn->lruvec_stats.state_pending[i] += delta;
5848 statc->stats_updates = 0;
5849 /* We are in a per-cpu loop here, only do the atomic write once */
5850 if (atomic64_read(&memcg->vmstats->stats_updates))
5851 atomic64_set(&memcg->vmstats->stats_updates, 0);
5855 /* Handlers for move charge at task migration. */
5856 static int mem_cgroup_do_precharge(unsigned long count)
5860 /* Try a single bulk charge without reclaim first, kswapd may wake */
5861 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5863 mc.precharge += count;
5867 /* Try charges one by one with reclaim, but do not retry */
5869 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5879 struct folio *folio;
5883 enum mc_target_type {
5890 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5891 unsigned long addr, pte_t ptent)
5893 struct page *page = vm_normal_page(vma, addr, ptent);
5897 if (PageAnon(page)) {
5898 if (!(mc.flags & MOVE_ANON))
5901 if (!(mc.flags & MOVE_FILE))
5909 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5910 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5911 pte_t ptent, swp_entry_t *entry)
5913 struct page *page = NULL;
5914 swp_entry_t ent = pte_to_swp_entry(ptent);
5916 if (!(mc.flags & MOVE_ANON))
5920 * Handle device private pages that are not accessible by the CPU, but
5921 * stored as special swap entries in the page table.
5923 if (is_device_private_entry(ent)) {
5924 page = pfn_swap_entry_to_page(ent);
5925 if (!get_page_unless_zero(page))
5930 if (non_swap_entry(ent))
5934 * Because swap_cache_get_folio() updates some statistics counter,
5935 * we call find_get_page() with swapper_space directly.
5937 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5938 entry->val = ent.val;
5943 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5944 pte_t ptent, swp_entry_t *entry)
5950 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5951 unsigned long addr, pte_t ptent)
5953 unsigned long index;
5954 struct folio *folio;
5956 if (!vma->vm_file) /* anonymous vma */
5958 if (!(mc.flags & MOVE_FILE))
5961 /* folio is moved even if it's not RSS of this task(page-faulted). */
5962 /* shmem/tmpfs may report page out on swap: account for that too. */
5963 index = linear_page_index(vma, addr);
5964 folio = filemap_get_incore_folio(vma->vm_file->f_mapping, index);
5967 return folio_file_page(folio, index);
5971 * mem_cgroup_move_account - move account of the folio
5972 * @folio: The folio.
5973 * @compound: charge the page as compound or small page
5974 * @from: mem_cgroup which the folio is moved from.
5975 * @to: mem_cgroup which the folio is moved to. @from != @to.
5977 * The folio must be locked and not on the LRU.
5979 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5982 static int mem_cgroup_move_account(struct folio *folio,
5984 struct mem_cgroup *from,
5985 struct mem_cgroup *to)
5987 struct lruvec *from_vec, *to_vec;
5988 struct pglist_data *pgdat;
5989 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5992 VM_BUG_ON(from == to);
5993 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
5994 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5995 VM_BUG_ON(compound && !folio_test_large(folio));
5998 if (folio_memcg(folio) != from)
6001 pgdat = folio_pgdat(folio);
6002 from_vec = mem_cgroup_lruvec(from, pgdat);
6003 to_vec = mem_cgroup_lruvec(to, pgdat);
6005 folio_memcg_lock(folio);
6007 if (folio_test_anon(folio)) {
6008 if (folio_mapped(folio)) {
6009 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
6010 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
6011 if (folio_test_pmd_mappable(folio)) {
6012 __mod_lruvec_state(from_vec, NR_ANON_THPS,
6014 __mod_lruvec_state(to_vec, NR_ANON_THPS,
6019 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
6020 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
6022 if (folio_test_swapbacked(folio)) {
6023 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
6024 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
6027 if (folio_mapped(folio)) {
6028 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
6029 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
6032 if (folio_test_dirty(folio)) {
6033 struct address_space *mapping = folio_mapping(folio);
6035 if (mapping_can_writeback(mapping)) {
6036 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
6038 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
6045 if (folio_test_swapcache(folio)) {
6046 __mod_lruvec_state(from_vec, NR_SWAPCACHE, -nr_pages);
6047 __mod_lruvec_state(to_vec, NR_SWAPCACHE, nr_pages);
6050 if (folio_test_writeback(folio)) {
6051 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
6052 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
6056 * All state has been migrated, let's switch to the new memcg.
6058 * It is safe to change page's memcg here because the page
6059 * is referenced, charged, isolated, and locked: we can't race
6060 * with (un)charging, migration, LRU putback, or anything else
6061 * that would rely on a stable page's memory cgroup.
6063 * Note that folio_memcg_lock is a memcg lock, not a page lock,
6064 * to save space. As soon as we switch page's memory cgroup to a
6065 * new memcg that isn't locked, the above state can change
6066 * concurrently again. Make sure we're truly done with it.
6071 css_put(&from->css);
6073 folio->memcg_data = (unsigned long)to;
6075 __folio_memcg_unlock(from);
6078 nid = folio_nid(folio);
6080 local_irq_disable();
6081 mem_cgroup_charge_statistics(to, nr_pages);
6082 memcg_check_events(to, nid);
6083 mem_cgroup_charge_statistics(from, -nr_pages);
6084 memcg_check_events(from, nid);
6091 * get_mctgt_type - get target type of moving charge
6092 * @vma: the vma the pte to be checked belongs
6093 * @addr: the address corresponding to the pte to be checked
6094 * @ptent: the pte to be checked
6095 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6097 * Context: Called with pte lock held.
6099 * * MC_TARGET_NONE - If the pte is not a target for move charge.
6100 * * MC_TARGET_PAGE - If the page corresponding to this pte is a target for
6101 * move charge. If @target is not NULL, the folio is stored in target->folio
6102 * with extra refcnt taken (Caller should release it).
6103 * * MC_TARGET_SWAP - If the swap entry corresponding to this pte is a
6104 * target for charge migration. If @target is not NULL, the entry is
6105 * stored in target->ent.
6106 * * MC_TARGET_DEVICE - Like MC_TARGET_PAGE but page is device memory and
6107 * thus not on the lru. For now such page is charged like a regular page
6108 * would be as it is just special memory taking the place of a regular page.
6109 * See Documentations/vm/hmm.txt and include/linux/hmm.h
6111 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6112 unsigned long addr, pte_t ptent, union mc_target *target)
6114 struct page *page = NULL;
6115 struct folio *folio;
6116 enum mc_target_type ret = MC_TARGET_NONE;
6117 swp_entry_t ent = { .val = 0 };
6119 if (pte_present(ptent))
6120 page = mc_handle_present_pte(vma, addr, ptent);
6121 else if (pte_none_mostly(ptent))
6123 * PTE markers should be treated as a none pte here, separated
6124 * from other swap handling below.
6126 page = mc_handle_file_pte(vma, addr, ptent);
6127 else if (is_swap_pte(ptent))
6128 page = mc_handle_swap_pte(vma, ptent, &ent);
6131 folio = page_folio(page);
6132 if (target && page) {
6133 if (!folio_trylock(folio)) {
6138 * page_mapped() must be stable during the move. This
6139 * pte is locked, so if it's present, the page cannot
6140 * become unmapped. If it isn't, we have only partial
6141 * control over the mapped state: the page lock will
6142 * prevent new faults against pagecache and swapcache,
6143 * so an unmapped page cannot become mapped. However,
6144 * if the page is already mapped elsewhere, it can
6145 * unmap, and there is nothing we can do about it.
6146 * Alas, skip moving the page in this case.
6148 if (!pte_present(ptent) && page_mapped(page)) {
6149 folio_unlock(folio);
6155 if (!page && !ent.val)
6159 * Do only loose check w/o serialization.
6160 * mem_cgroup_move_account() checks the page is valid or
6161 * not under LRU exclusion.
6163 if (folio_memcg(folio) == mc.from) {
6164 ret = MC_TARGET_PAGE;
6165 if (folio_is_device_private(folio) ||
6166 folio_is_device_coherent(folio))
6167 ret = MC_TARGET_DEVICE;
6169 target->folio = folio;
6171 if (!ret || !target) {
6173 folio_unlock(folio);
6178 * There is a swap entry and a page doesn't exist or isn't charged.
6179 * But we cannot move a tail-page in a THP.
6181 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
6182 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6183 ret = MC_TARGET_SWAP;
6190 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6192 * We don't consider PMD mapped swapping or file mapped pages because THP does
6193 * not support them for now.
6194 * Caller should make sure that pmd_trans_huge(pmd) is true.
6196 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6197 unsigned long addr, pmd_t pmd, union mc_target *target)
6199 struct page *page = NULL;
6200 struct folio *folio;
6201 enum mc_target_type ret = MC_TARGET_NONE;
6203 if (unlikely(is_swap_pmd(pmd))) {
6204 VM_BUG_ON(thp_migration_supported() &&
6205 !is_pmd_migration_entry(pmd));
6208 page = pmd_page(pmd);
6209 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6210 folio = page_folio(page);
6211 if (!(mc.flags & MOVE_ANON))
6213 if (folio_memcg(folio) == mc.from) {
6214 ret = MC_TARGET_PAGE;
6217 if (!folio_trylock(folio)) {
6219 return MC_TARGET_NONE;
6221 target->folio = folio;
6227 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6228 unsigned long addr, pmd_t pmd, union mc_target *target)
6230 return MC_TARGET_NONE;
6234 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6235 unsigned long addr, unsigned long end,
6236 struct mm_walk *walk)
6238 struct vm_area_struct *vma = walk->vma;
6242 ptl = pmd_trans_huge_lock(pmd, vma);
6245 * Note their can not be MC_TARGET_DEVICE for now as we do not
6246 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
6247 * this might change.
6249 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6250 mc.precharge += HPAGE_PMD_NR;
6255 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6258 for (; addr != end; pte++, addr += PAGE_SIZE)
6259 if (get_mctgt_type(vma, addr, ptep_get(pte), NULL))
6260 mc.precharge++; /* increment precharge temporarily */
6261 pte_unmap_unlock(pte - 1, ptl);
6267 static const struct mm_walk_ops precharge_walk_ops = {
6268 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6269 .walk_lock = PGWALK_RDLOCK,
6272 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6274 unsigned long precharge;
6277 walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
6278 mmap_read_unlock(mm);
6280 precharge = mc.precharge;
6286 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6288 unsigned long precharge = mem_cgroup_count_precharge(mm);
6290 VM_BUG_ON(mc.moving_task);
6291 mc.moving_task = current;
6292 return mem_cgroup_do_precharge(precharge);
6295 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6296 static void __mem_cgroup_clear_mc(void)
6298 struct mem_cgroup *from = mc.from;
6299 struct mem_cgroup *to = mc.to;
6301 /* we must uncharge all the leftover precharges from mc.to */
6303 mem_cgroup_cancel_charge(mc.to, mc.precharge);
6307 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6308 * we must uncharge here.
6310 if (mc.moved_charge) {
6311 mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6312 mc.moved_charge = 0;
6314 /* we must fixup refcnts and charges */
6315 if (mc.moved_swap) {
6316 /* uncharge swap account from the old cgroup */
6317 if (!mem_cgroup_is_root(mc.from))
6318 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6320 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6323 * we charged both to->memory and to->memsw, so we
6324 * should uncharge to->memory.
6326 if (!mem_cgroup_is_root(mc.to))
6327 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6331 memcg_oom_recover(from);
6332 memcg_oom_recover(to);
6333 wake_up_all(&mc.waitq);
6336 static void mem_cgroup_clear_mc(void)
6338 struct mm_struct *mm = mc.mm;
6341 * we must clear moving_task before waking up waiters at the end of
6344 mc.moving_task = NULL;
6345 __mem_cgroup_clear_mc();
6346 spin_lock(&mc.lock);
6350 spin_unlock(&mc.lock);
6355 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6357 struct cgroup_subsys_state *css;
6358 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6359 struct mem_cgroup *from;
6360 struct task_struct *leader, *p;
6361 struct mm_struct *mm;
6362 unsigned long move_flags;
6365 /* charge immigration isn't supported on the default hierarchy */
6366 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6370 * Multi-process migrations only happen on the default hierarchy
6371 * where charge immigration is not used. Perform charge
6372 * immigration if @tset contains a leader and whine if there are
6376 cgroup_taskset_for_each_leader(leader, css, tset) {
6379 memcg = mem_cgroup_from_css(css);
6385 * We are now committed to this value whatever it is. Changes in this
6386 * tunable will only affect upcoming migrations, not the current one.
6387 * So we need to save it, and keep it going.
6389 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6393 from = mem_cgroup_from_task(p);
6395 VM_BUG_ON(from == memcg);
6397 mm = get_task_mm(p);
6400 /* We move charges only when we move a owner of the mm */
6401 if (mm->owner == p) {
6404 VM_BUG_ON(mc.precharge);
6405 VM_BUG_ON(mc.moved_charge);
6406 VM_BUG_ON(mc.moved_swap);
6408 spin_lock(&mc.lock);
6412 mc.flags = move_flags;
6413 spin_unlock(&mc.lock);
6414 /* We set mc.moving_task later */
6416 ret = mem_cgroup_precharge_mc(mm);
6418 mem_cgroup_clear_mc();
6425 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6428 mem_cgroup_clear_mc();
6431 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6432 unsigned long addr, unsigned long end,
6433 struct mm_walk *walk)
6436 struct vm_area_struct *vma = walk->vma;
6439 enum mc_target_type target_type;
6440 union mc_target target;
6441 struct folio *folio;
6443 ptl = pmd_trans_huge_lock(pmd, vma);
6445 if (mc.precharge < HPAGE_PMD_NR) {
6449 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6450 if (target_type == MC_TARGET_PAGE) {
6451 folio = target.folio;
6452 if (folio_isolate_lru(folio)) {
6453 if (!mem_cgroup_move_account(folio, true,
6455 mc.precharge -= HPAGE_PMD_NR;
6456 mc.moved_charge += HPAGE_PMD_NR;
6458 folio_putback_lru(folio);
6460 folio_unlock(folio);
6462 } else if (target_type == MC_TARGET_DEVICE) {
6463 folio = target.folio;
6464 if (!mem_cgroup_move_account(folio, true,
6466 mc.precharge -= HPAGE_PMD_NR;
6467 mc.moved_charge += HPAGE_PMD_NR;
6469 folio_unlock(folio);
6477 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6480 for (; addr != end; addr += PAGE_SIZE) {
6481 pte_t ptent = ptep_get(pte++);
6482 bool device = false;
6488 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6489 case MC_TARGET_DEVICE:
6492 case MC_TARGET_PAGE:
6493 folio = target.folio;
6495 * We can have a part of the split pmd here. Moving it
6496 * can be done but it would be too convoluted so simply
6497 * ignore such a partial THP and keep it in original
6498 * memcg. There should be somebody mapping the head.
6500 if (folio_test_large(folio))
6502 if (!device && !folio_isolate_lru(folio))
6504 if (!mem_cgroup_move_account(folio, false,
6507 /* we uncharge from mc.from later. */
6511 folio_putback_lru(folio);
6512 put: /* get_mctgt_type() gets & locks the page */
6513 folio_unlock(folio);
6516 case MC_TARGET_SWAP:
6518 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6520 mem_cgroup_id_get_many(mc.to, 1);
6521 /* we fixup other refcnts and charges later. */
6529 pte_unmap_unlock(pte - 1, ptl);
6534 * We have consumed all precharges we got in can_attach().
6535 * We try charge one by one, but don't do any additional
6536 * charges to mc.to if we have failed in charge once in attach()
6539 ret = mem_cgroup_do_precharge(1);
6547 static const struct mm_walk_ops charge_walk_ops = {
6548 .pmd_entry = mem_cgroup_move_charge_pte_range,
6549 .walk_lock = PGWALK_RDLOCK,
6552 static void mem_cgroup_move_charge(void)
6554 lru_add_drain_all();
6556 * Signal folio_memcg_lock() to take the memcg's move_lock
6557 * while we're moving its pages to another memcg. Then wait
6558 * for already started RCU-only updates to finish.
6560 atomic_inc(&mc.from->moving_account);
6563 if (unlikely(!mmap_read_trylock(mc.mm))) {
6565 * Someone who are holding the mmap_lock might be waiting in
6566 * waitq. So we cancel all extra charges, wake up all waiters,
6567 * and retry. Because we cancel precharges, we might not be able
6568 * to move enough charges, but moving charge is a best-effort
6569 * feature anyway, so it wouldn't be a big problem.
6571 __mem_cgroup_clear_mc();
6576 * When we have consumed all precharges and failed in doing
6577 * additional charge, the page walk just aborts.
6579 walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6580 mmap_read_unlock(mc.mm);
6581 atomic_dec(&mc.from->moving_account);
6584 static void mem_cgroup_move_task(void)
6587 mem_cgroup_move_charge();
6588 mem_cgroup_clear_mc();
6592 #else /* !CONFIG_MMU */
6593 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6597 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6600 static void mem_cgroup_move_task(void)
6605 #ifdef CONFIG_MEMCG_KMEM
6606 static void mem_cgroup_fork(struct task_struct *task)
6609 * Set the update flag to cause task->objcg to be initialized lazily
6610 * on the first allocation. It can be done without any synchronization
6611 * because it's always performed on the current task, so does
6612 * current_objcg_update().
6614 task->objcg = (struct obj_cgroup *)CURRENT_OBJCG_UPDATE_FLAG;
6617 static void mem_cgroup_exit(struct task_struct *task)
6619 struct obj_cgroup *objcg = task->objcg;
6621 objcg = (struct obj_cgroup *)
6622 ((unsigned long)objcg & ~CURRENT_OBJCG_UPDATE_FLAG);
6624 obj_cgroup_put(objcg);
6627 * Some kernel allocations can happen after this point,
6628 * but let's ignore them. It can be done without any synchronization
6629 * because it's always performed on the current task, so does
6630 * current_objcg_update().
6636 #ifdef CONFIG_LRU_GEN
6637 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset)
6639 struct task_struct *task;
6640 struct cgroup_subsys_state *css;
6642 /* find the first leader if there is any */
6643 cgroup_taskset_for_each_leader(task, css, tset)
6650 if (task->mm && READ_ONCE(task->mm->owner) == task)
6651 lru_gen_migrate_mm(task->mm);
6655 static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset) {}
6656 #endif /* CONFIG_LRU_GEN */
6658 #ifdef CONFIG_MEMCG_KMEM
6659 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset)
6661 struct task_struct *task;
6662 struct cgroup_subsys_state *css;
6664 cgroup_taskset_for_each(task, css, tset) {
6665 /* atomically set the update bit */
6666 set_bit(CURRENT_OBJCG_UPDATE_BIT, (unsigned long *)&task->objcg);
6670 static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset) {}
6671 #endif /* CONFIG_MEMCG_KMEM */
6673 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
6674 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6676 mem_cgroup_lru_gen_attach(tset);
6677 mem_cgroup_kmem_attach(tset);
6681 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6683 if (value == PAGE_COUNTER_MAX)
6684 seq_puts(m, "max\n");
6686 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6691 static u64 memory_current_read(struct cgroup_subsys_state *css,
6694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6696 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6699 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6702 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6704 return (u64)memcg->memory.watermark * PAGE_SIZE;
6707 static int memory_min_show(struct seq_file *m, void *v)
6709 return seq_puts_memcg_tunable(m,
6710 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6713 static ssize_t memory_min_write(struct kernfs_open_file *of,
6714 char *buf, size_t nbytes, loff_t off)
6716 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6720 buf = strstrip(buf);
6721 err = page_counter_memparse(buf, "max", &min);
6725 page_counter_set_min(&memcg->memory, min);
6730 static int memory_low_show(struct seq_file *m, void *v)
6732 return seq_puts_memcg_tunable(m,
6733 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6736 static ssize_t memory_low_write(struct kernfs_open_file *of,
6737 char *buf, size_t nbytes, loff_t off)
6739 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6743 buf = strstrip(buf);
6744 err = page_counter_memparse(buf, "max", &low);
6748 page_counter_set_low(&memcg->memory, low);
6753 static int memory_high_show(struct seq_file *m, void *v)
6755 return seq_puts_memcg_tunable(m,
6756 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6759 static ssize_t memory_high_write(struct kernfs_open_file *of,
6760 char *buf, size_t nbytes, loff_t off)
6762 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6763 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6764 bool drained = false;
6768 buf = strstrip(buf);
6769 err = page_counter_memparse(buf, "max", &high);
6773 page_counter_set_high(&memcg->memory, high);
6776 unsigned long nr_pages = page_counter_read(&memcg->memory);
6777 unsigned long reclaimed;
6779 if (nr_pages <= high)
6782 if (signal_pending(current))
6786 drain_all_stock(memcg);
6791 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6792 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6794 if (!reclaimed && !nr_retries--)
6798 memcg_wb_domain_size_changed(memcg);
6802 static int memory_max_show(struct seq_file *m, void *v)
6804 return seq_puts_memcg_tunable(m,
6805 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6808 static ssize_t memory_max_write(struct kernfs_open_file *of,
6809 char *buf, size_t nbytes, loff_t off)
6811 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6812 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6813 bool drained = false;
6817 buf = strstrip(buf);
6818 err = page_counter_memparse(buf, "max", &max);
6822 xchg(&memcg->memory.max, max);
6825 unsigned long nr_pages = page_counter_read(&memcg->memory);
6827 if (nr_pages <= max)
6830 if (signal_pending(current))
6834 drain_all_stock(memcg);
6840 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6841 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6846 memcg_memory_event(memcg, MEMCG_OOM);
6847 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6851 memcg_wb_domain_size_changed(memcg);
6856 * Note: don't forget to update the 'samples/cgroup/memcg_event_listener'
6857 * if any new events become available.
6859 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6861 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6862 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6863 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6864 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6865 seq_printf(m, "oom_kill %lu\n",
6866 atomic_long_read(&events[MEMCG_OOM_KILL]));
6867 seq_printf(m, "oom_group_kill %lu\n",
6868 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6871 static int memory_events_show(struct seq_file *m, void *v)
6873 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6875 __memory_events_show(m, memcg->memory_events);
6879 static int memory_events_local_show(struct seq_file *m, void *v)
6881 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6883 __memory_events_show(m, memcg->memory_events_local);
6887 static int memory_stat_show(struct seq_file *m, void *v)
6889 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6890 char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6895 seq_buf_init(&s, buf, PAGE_SIZE);
6896 memory_stat_format(memcg, &s);
6903 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6906 return lruvec_page_state(lruvec, item) *
6907 memcg_page_state_output_unit(item);
6910 static int memory_numa_stat_show(struct seq_file *m, void *v)
6913 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6915 mem_cgroup_flush_stats(memcg);
6917 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6920 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6923 seq_printf(m, "%s", memory_stats[i].name);
6924 for_each_node_state(nid, N_MEMORY) {
6926 struct lruvec *lruvec;
6928 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6929 size = lruvec_page_state_output(lruvec,
6930 memory_stats[i].idx);
6931 seq_printf(m, " N%d=%llu", nid, size);
6940 static int memory_oom_group_show(struct seq_file *m, void *v)
6942 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6944 seq_printf(m, "%d\n", READ_ONCE(memcg->oom_group));
6949 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6950 char *buf, size_t nbytes, loff_t off)
6952 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6955 buf = strstrip(buf);
6959 ret = kstrtoint(buf, 0, &oom_group);
6963 if (oom_group != 0 && oom_group != 1)
6966 WRITE_ONCE(memcg->oom_group, oom_group);
6971 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6972 size_t nbytes, loff_t off)
6974 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6975 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6976 unsigned long nr_to_reclaim, nr_reclaimed = 0;
6977 unsigned int reclaim_options;
6980 buf = strstrip(buf);
6981 err = page_counter_memparse(buf, "", &nr_to_reclaim);
6985 reclaim_options = MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6986 while (nr_reclaimed < nr_to_reclaim) {
6987 /* Will converge on zero, but reclaim enforces a minimum */
6988 unsigned long batch_size = (nr_to_reclaim - nr_reclaimed) / 4;
6989 unsigned long reclaimed;
6991 if (signal_pending(current))
6995 * This is the final attempt, drain percpu lru caches in the
6996 * hope of introducing more evictable pages for
6997 * try_to_free_mem_cgroup_pages().
7000 lru_add_drain_all();
7002 reclaimed = try_to_free_mem_cgroup_pages(memcg,
7003 batch_size, GFP_KERNEL, reclaim_options);
7005 if (!reclaimed && !nr_retries--)
7008 nr_reclaimed += reclaimed;
7014 static struct cftype memory_files[] = {
7017 .flags = CFTYPE_NOT_ON_ROOT,
7018 .read_u64 = memory_current_read,
7022 .flags = CFTYPE_NOT_ON_ROOT,
7023 .read_u64 = memory_peak_read,
7027 .flags = CFTYPE_NOT_ON_ROOT,
7028 .seq_show = memory_min_show,
7029 .write = memory_min_write,
7033 .flags = CFTYPE_NOT_ON_ROOT,
7034 .seq_show = memory_low_show,
7035 .write = memory_low_write,
7039 .flags = CFTYPE_NOT_ON_ROOT,
7040 .seq_show = memory_high_show,
7041 .write = memory_high_write,
7045 .flags = CFTYPE_NOT_ON_ROOT,
7046 .seq_show = memory_max_show,
7047 .write = memory_max_write,
7051 .flags = CFTYPE_NOT_ON_ROOT,
7052 .file_offset = offsetof(struct mem_cgroup, events_file),
7053 .seq_show = memory_events_show,
7056 .name = "events.local",
7057 .flags = CFTYPE_NOT_ON_ROOT,
7058 .file_offset = offsetof(struct mem_cgroup, events_local_file),
7059 .seq_show = memory_events_local_show,
7063 .seq_show = memory_stat_show,
7067 .name = "numa_stat",
7068 .seq_show = memory_numa_stat_show,
7072 .name = "oom.group",
7073 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
7074 .seq_show = memory_oom_group_show,
7075 .write = memory_oom_group_write,
7079 .flags = CFTYPE_NS_DELEGATABLE,
7080 .write = memory_reclaim,
7085 struct cgroup_subsys memory_cgrp_subsys = {
7086 .css_alloc = mem_cgroup_css_alloc,
7087 .css_online = mem_cgroup_css_online,
7088 .css_offline = mem_cgroup_css_offline,
7089 .css_released = mem_cgroup_css_released,
7090 .css_free = mem_cgroup_css_free,
7091 .css_reset = mem_cgroup_css_reset,
7092 .css_rstat_flush = mem_cgroup_css_rstat_flush,
7093 .can_attach = mem_cgroup_can_attach,
7094 #if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM)
7095 .attach = mem_cgroup_attach,
7097 .cancel_attach = mem_cgroup_cancel_attach,
7098 .post_attach = mem_cgroup_move_task,
7099 #ifdef CONFIG_MEMCG_KMEM
7100 .fork = mem_cgroup_fork,
7101 .exit = mem_cgroup_exit,
7103 .dfl_cftypes = memory_files,
7104 .legacy_cftypes = mem_cgroup_legacy_files,
7109 * This function calculates an individual cgroup's effective
7110 * protection which is derived from its own memory.min/low, its
7111 * parent's and siblings' settings, as well as the actual memory
7112 * distribution in the tree.
7114 * The following rules apply to the effective protection values:
7116 * 1. At the first level of reclaim, effective protection is equal to
7117 * the declared protection in memory.min and memory.low.
7119 * 2. To enable safe delegation of the protection configuration, at
7120 * subsequent levels the effective protection is capped to the
7121 * parent's effective protection.
7123 * 3. To make complex and dynamic subtrees easier to configure, the
7124 * user is allowed to overcommit the declared protection at a given
7125 * level. If that is the case, the parent's effective protection is
7126 * distributed to the children in proportion to how much protection
7127 * they have declared and how much of it they are utilizing.
7129 * This makes distribution proportional, but also work-conserving:
7130 * if one cgroup claims much more protection than it uses memory,
7131 * the unused remainder is available to its siblings.
7133 * 4. Conversely, when the declared protection is undercommitted at a
7134 * given level, the distribution of the larger parental protection
7135 * budget is NOT proportional. A cgroup's protection from a sibling
7136 * is capped to its own memory.min/low setting.
7138 * 5. However, to allow protecting recursive subtrees from each other
7139 * without having to declare each individual cgroup's fixed share
7140 * of the ancestor's claim to protection, any unutilized -
7141 * "floating" - protection from up the tree is distributed in
7142 * proportion to each cgroup's *usage*. This makes the protection
7143 * neutral wrt sibling cgroups and lets them compete freely over
7144 * the shared parental protection budget, but it protects the
7145 * subtree as a whole from neighboring subtrees.
7147 * Note that 4. and 5. are not in conflict: 4. is about protecting
7148 * against immediate siblings whereas 5. is about protecting against
7149 * neighboring subtrees.
7151 static unsigned long effective_protection(unsigned long usage,
7152 unsigned long parent_usage,
7153 unsigned long setting,
7154 unsigned long parent_effective,
7155 unsigned long siblings_protected)
7157 unsigned long protected;
7160 protected = min(usage, setting);
7162 * If all cgroups at this level combined claim and use more
7163 * protection than what the parent affords them, distribute
7164 * shares in proportion to utilization.
7166 * We are using actual utilization rather than the statically
7167 * claimed protection in order to be work-conserving: claimed
7168 * but unused protection is available to siblings that would
7169 * otherwise get a smaller chunk than what they claimed.
7171 if (siblings_protected > parent_effective)
7172 return protected * parent_effective / siblings_protected;
7175 * Ok, utilized protection of all children is within what the
7176 * parent affords them, so we know whatever this child claims
7177 * and utilizes is effectively protected.
7179 * If there is unprotected usage beyond this value, reclaim
7180 * will apply pressure in proportion to that amount.
7182 * If there is unutilized protection, the cgroup will be fully
7183 * shielded from reclaim, but we do return a smaller value for
7184 * protection than what the group could enjoy in theory. This
7185 * is okay. With the overcommit distribution above, effective
7186 * protection is always dependent on how memory is actually
7187 * consumed among the siblings anyway.
7192 * If the children aren't claiming (all of) the protection
7193 * afforded to them by the parent, distribute the remainder in
7194 * proportion to the (unprotected) memory of each cgroup. That
7195 * way, cgroups that aren't explicitly prioritized wrt each
7196 * other compete freely over the allowance, but they are
7197 * collectively protected from neighboring trees.
7199 * We're using unprotected memory for the weight so that if
7200 * some cgroups DO claim explicit protection, we don't protect
7201 * the same bytes twice.
7203 * Check both usage and parent_usage against the respective
7204 * protected values. One should imply the other, but they
7205 * aren't read atomically - make sure the division is sane.
7207 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
7209 if (parent_effective > siblings_protected &&
7210 parent_usage > siblings_protected &&
7211 usage > protected) {
7212 unsigned long unclaimed;
7214 unclaimed = parent_effective - siblings_protected;
7215 unclaimed *= usage - protected;
7216 unclaimed /= parent_usage - siblings_protected;
7225 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
7226 * @root: the top ancestor of the sub-tree being checked
7227 * @memcg: the memory cgroup to check
7229 * WARNING: This function is not stateless! It can only be used as part
7230 * of a top-down tree iteration, not for isolated queries.
7232 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
7233 struct mem_cgroup *memcg)
7235 unsigned long usage, parent_usage;
7236 struct mem_cgroup *parent;
7238 if (mem_cgroup_disabled())
7242 root = root_mem_cgroup;
7245 * Effective values of the reclaim targets are ignored so they
7246 * can be stale. Have a look at mem_cgroup_protection for more
7248 * TODO: calculation should be more robust so that we do not need
7249 * that special casing.
7254 usage = page_counter_read(&memcg->memory);
7258 parent = parent_mem_cgroup(memcg);
7260 if (parent == root) {
7261 memcg->memory.emin = READ_ONCE(memcg->memory.min);
7262 memcg->memory.elow = READ_ONCE(memcg->memory.low);
7266 parent_usage = page_counter_read(&parent->memory);
7268 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
7269 READ_ONCE(memcg->memory.min),
7270 READ_ONCE(parent->memory.emin),
7271 atomic_long_read(&parent->memory.children_min_usage)));
7273 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
7274 READ_ONCE(memcg->memory.low),
7275 READ_ONCE(parent->memory.elow),
7276 atomic_long_read(&parent->memory.children_low_usage)));
7279 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
7284 ret = try_charge(memcg, gfp, folio_nr_pages(folio));
7288 mem_cgroup_commit_charge(folio, memcg);
7293 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
7295 struct mem_cgroup *memcg;
7298 memcg = get_mem_cgroup_from_mm(mm);
7299 ret = charge_memcg(folio, memcg, gfp);
7300 css_put(&memcg->css);
7306 * mem_cgroup_hugetlb_try_charge - try to charge the memcg for a hugetlb folio
7307 * @memcg: memcg to charge.
7308 * @gfp: reclaim mode.
7309 * @nr_pages: number of pages to charge.
7311 * This function is called when allocating a huge page folio to determine if
7312 * the memcg has the capacity for it. It does not commit the charge yet,
7313 * as the hugetlb folio itself has not been obtained from the hugetlb pool.
7315 * Once we have obtained the hugetlb folio, we can call
7316 * mem_cgroup_commit_charge() to commit the charge. If we fail to obtain the
7317 * folio, we should instead call mem_cgroup_cancel_charge() to undo the effect
7320 * Returns 0 on success. Otherwise, an error code is returned.
7322 int mem_cgroup_hugetlb_try_charge(struct mem_cgroup *memcg, gfp_t gfp,
7326 * If hugetlb memcg charging is not enabled, do not fail hugetlb allocation,
7327 * but do not attempt to commit charge later (or cancel on error) either.
7329 if (mem_cgroup_disabled() || !memcg ||
7330 !cgroup_subsys_on_dfl(memory_cgrp_subsys) ||
7331 !(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING))
7334 if (try_charge(memcg, gfp, nr_pages))
7341 * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
7342 * @folio: folio to charge.
7343 * @mm: mm context of the victim
7344 * @gfp: reclaim mode
7345 * @entry: swap entry for which the folio is allocated
7347 * This function charges a folio allocated for swapin. Please call this before
7348 * adding the folio to the swapcache.
7350 * Returns 0 on success. Otherwise, an error code is returned.
7352 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
7353 gfp_t gfp, swp_entry_t entry)
7355 struct mem_cgroup *memcg;
7359 if (mem_cgroup_disabled())
7362 id = lookup_swap_cgroup_id(entry);
7364 memcg = mem_cgroup_from_id(id);
7365 if (!memcg || !css_tryget_online(&memcg->css))
7366 memcg = get_mem_cgroup_from_mm(mm);
7369 ret = charge_memcg(folio, memcg, gfp);
7371 css_put(&memcg->css);
7376 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
7377 * @entry: swap entry for which the page is charged
7379 * Call this function after successfully adding the charged page to swapcache.
7381 * Note: This function assumes the page for which swap slot is being uncharged
7384 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
7387 * Cgroup1's unified memory+swap counter has been charged with the
7388 * new swapcache page, finish the transfer by uncharging the swap
7389 * slot. The swap slot would also get uncharged when it dies, but
7390 * it can stick around indefinitely and we'd count the page twice
7393 * Cgroup2 has separate resource counters for memory and swap,
7394 * so this is a non-issue here. Memory and swap charge lifetimes
7395 * correspond 1:1 to page and swap slot lifetimes: we charge the
7396 * page to memory here, and uncharge swap when the slot is freed.
7398 if (!mem_cgroup_disabled() && do_memsw_account()) {
7400 * The swap entry might not get freed for a long time,
7401 * let's not wait for it. The page already received a
7402 * memory+swap charge, drop the swap entry duplicate.
7404 mem_cgroup_uncharge_swap(entry, 1);
7408 struct uncharge_gather {
7409 struct mem_cgroup *memcg;
7410 unsigned long nr_memory;
7411 unsigned long pgpgout;
7412 unsigned long nr_kmem;
7416 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7418 memset(ug, 0, sizeof(*ug));
7421 static void uncharge_batch(const struct uncharge_gather *ug)
7423 unsigned long flags;
7425 if (ug->nr_memory) {
7426 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7427 if (do_memsw_account())
7428 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7430 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7431 memcg_oom_recover(ug->memcg);
7434 local_irq_save(flags);
7435 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7436 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7437 memcg_check_events(ug->memcg, ug->nid);
7438 local_irq_restore(flags);
7440 /* drop reference from uncharge_folio */
7441 css_put(&ug->memcg->css);
7444 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7447 struct mem_cgroup *memcg;
7448 struct obj_cgroup *objcg;
7450 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7453 * Nobody should be changing or seriously looking at
7454 * folio memcg or objcg at this point, we have fully
7455 * exclusive access to the folio.
7457 if (folio_memcg_kmem(folio)) {
7458 objcg = __folio_objcg(folio);
7460 * This get matches the put at the end of the function and
7461 * kmem pages do not hold memcg references anymore.
7463 memcg = get_mem_cgroup_from_objcg(objcg);
7465 memcg = __folio_memcg(folio);
7471 if (ug->memcg != memcg) {
7474 uncharge_gather_clear(ug);
7477 ug->nid = folio_nid(folio);
7479 /* pairs with css_put in uncharge_batch */
7480 css_get(&memcg->css);
7483 nr_pages = folio_nr_pages(folio);
7485 if (folio_memcg_kmem(folio)) {
7486 ug->nr_memory += nr_pages;
7487 ug->nr_kmem += nr_pages;
7489 folio->memcg_data = 0;
7490 obj_cgroup_put(objcg);
7492 /* LRU pages aren't accounted at the root level */
7493 if (!mem_cgroup_is_root(memcg))
7494 ug->nr_memory += nr_pages;
7497 folio->memcg_data = 0;
7500 css_put(&memcg->css);
7503 void __mem_cgroup_uncharge(struct folio *folio)
7505 struct uncharge_gather ug;
7507 /* Don't touch folio->lru of any random page, pre-check: */
7508 if (!folio_memcg(folio))
7511 uncharge_gather_clear(&ug);
7512 uncharge_folio(folio, &ug);
7513 uncharge_batch(&ug);
7516 void __mem_cgroup_uncharge_folios(struct folio_batch *folios)
7518 struct uncharge_gather ug;
7521 uncharge_gather_clear(&ug);
7522 for (i = 0; i < folios->nr; i++)
7523 uncharge_folio(folios->folios[i], &ug);
7525 uncharge_batch(&ug);
7529 * mem_cgroup_replace_folio - Charge a folio's replacement.
7530 * @old: Currently circulating folio.
7531 * @new: Replacement folio.
7533 * Charge @new as a replacement folio for @old. @old will
7534 * be uncharged upon free. This is only used by the page cache
7535 * (in replace_page_cache_folio()).
7537 * Both folios must be locked, @new->mapping must be set up.
7539 void mem_cgroup_replace_folio(struct folio *old, struct folio *new)
7541 struct mem_cgroup *memcg;
7542 long nr_pages = folio_nr_pages(new);
7543 unsigned long flags;
7545 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7546 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7547 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7548 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7550 if (mem_cgroup_disabled())
7553 /* Page cache replacement: new folio already charged? */
7554 if (folio_memcg(new))
7557 memcg = folio_memcg(old);
7558 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7562 /* Force-charge the new page. The old one will be freed soon */
7563 if (!mem_cgroup_is_root(memcg)) {
7564 page_counter_charge(&memcg->memory, nr_pages);
7565 if (do_memsw_account())
7566 page_counter_charge(&memcg->memsw, nr_pages);
7569 css_get(&memcg->css);
7570 commit_charge(new, memcg);
7572 local_irq_save(flags);
7573 mem_cgroup_charge_statistics(memcg, nr_pages);
7574 memcg_check_events(memcg, folio_nid(new));
7575 local_irq_restore(flags);
7579 * mem_cgroup_migrate - Transfer the memcg data from the old to the new folio.
7580 * @old: Currently circulating folio.
7581 * @new: Replacement folio.
7583 * Transfer the memcg data from the old folio to the new folio for migration.
7584 * The old folio's data info will be cleared. Note that the memory counters
7585 * will remain unchanged throughout the process.
7587 * Both folios must be locked, @new->mapping must be set up.
7589 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7591 struct mem_cgroup *memcg;
7593 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7594 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7595 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7596 VM_BUG_ON_FOLIO(folio_nr_pages(old) != folio_nr_pages(new), new);
7598 if (mem_cgroup_disabled())
7601 memcg = folio_memcg(old);
7603 * Note that it is normal to see !memcg for a hugetlb folio.
7604 * For e.g, itt could have been allocated when memory_hugetlb_accounting
7607 VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(old) && !memcg, old);
7611 /* Transfer the charge and the css ref */
7612 commit_charge(new, memcg);
7614 * If the old folio is a large folio and is in the split queue, it needs
7615 * to be removed from the split queue now, in case getting an incorrect
7616 * split queue in destroy_large_folio() after the memcg of the old folio
7619 * In addition, the old folio is about to be freed after migration, so
7620 * removing from the split queue a bit earlier seems reasonable.
7622 if (folio_test_large(old) && folio_test_large_rmappable(old))
7623 folio_undo_large_rmappable(old);
7624 old->memcg_data = 0;
7627 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7628 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7630 void mem_cgroup_sk_alloc(struct sock *sk)
7632 struct mem_cgroup *memcg;
7634 if (!mem_cgroup_sockets_enabled)
7637 /* Do not associate the sock with unrelated interrupted task's memcg. */
7642 memcg = mem_cgroup_from_task(current);
7643 if (mem_cgroup_is_root(memcg))
7645 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7647 if (css_tryget(&memcg->css))
7648 sk->sk_memcg = memcg;
7653 void mem_cgroup_sk_free(struct sock *sk)
7656 css_put(&sk->sk_memcg->css);
7660 * mem_cgroup_charge_skmem - charge socket memory
7661 * @memcg: memcg to charge
7662 * @nr_pages: number of pages to charge
7663 * @gfp_mask: reclaim mode
7665 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7666 * @memcg's configured limit, %false if it doesn't.
7668 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7671 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7672 struct page_counter *fail;
7674 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7675 memcg->tcpmem_pressure = 0;
7678 memcg->tcpmem_pressure = 1;
7679 if (gfp_mask & __GFP_NOFAIL) {
7680 page_counter_charge(&memcg->tcpmem, nr_pages);
7686 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7687 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7695 * mem_cgroup_uncharge_skmem - uncharge socket memory
7696 * @memcg: memcg to uncharge
7697 * @nr_pages: number of pages to uncharge
7699 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7701 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7702 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7706 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7708 refill_stock(memcg, nr_pages);
7711 static int __init cgroup_memory(char *s)
7715 while ((token = strsep(&s, ",")) != NULL) {
7718 if (!strcmp(token, "nosocket"))
7719 cgroup_memory_nosocket = true;
7720 if (!strcmp(token, "nokmem"))
7721 cgroup_memory_nokmem = true;
7722 if (!strcmp(token, "nobpf"))
7723 cgroup_memory_nobpf = true;
7727 __setup("cgroup.memory=", cgroup_memory);
7730 * subsys_initcall() for memory controller.
7732 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7733 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7734 * basically everything that doesn't depend on a specific mem_cgroup structure
7735 * should be initialized from here.
7737 static int __init mem_cgroup_init(void)
7742 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7743 * used for per-memcg-per-cpu caching of per-node statistics. In order
7744 * to work fine, we should make sure that the overfill threshold can't
7745 * exceed S32_MAX / PAGE_SIZE.
7747 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7749 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7750 memcg_hotplug_cpu_dead);
7752 for_each_possible_cpu(cpu)
7753 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7756 for_each_node(node) {
7757 struct mem_cgroup_tree_per_node *rtpn;
7759 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, node);
7761 rtpn->rb_root = RB_ROOT;
7762 rtpn->rb_rightmost = NULL;
7763 spin_lock_init(&rtpn->lock);
7764 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7769 subsys_initcall(mem_cgroup_init);
7772 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7774 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7776 * The root cgroup cannot be destroyed, so it's refcount must
7779 if (WARN_ON_ONCE(mem_cgroup_is_root(memcg))) {
7783 memcg = parent_mem_cgroup(memcg);
7785 memcg = root_mem_cgroup;
7791 * mem_cgroup_swapout - transfer a memsw charge to swap
7792 * @folio: folio whose memsw charge to transfer
7793 * @entry: swap entry to move the charge to
7795 * Transfer the memsw charge of @folio to @entry.
7797 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7799 struct mem_cgroup *memcg, *swap_memcg;
7800 unsigned int nr_entries;
7801 unsigned short oldid;
7803 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7804 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7806 if (mem_cgroup_disabled())
7809 if (!do_memsw_account())
7812 memcg = folio_memcg(folio);
7814 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7819 * In case the memcg owning these pages has been offlined and doesn't
7820 * have an ID allocated to it anymore, charge the closest online
7821 * ancestor for the swap instead and transfer the memory+swap charge.
7823 swap_memcg = mem_cgroup_id_get_online(memcg);
7824 nr_entries = folio_nr_pages(folio);
7825 /* Get references for the tail pages, too */
7827 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7828 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7830 VM_BUG_ON_FOLIO(oldid, folio);
7831 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7833 folio->memcg_data = 0;
7835 if (!mem_cgroup_is_root(memcg))
7836 page_counter_uncharge(&memcg->memory, nr_entries);
7838 if (memcg != swap_memcg) {
7839 if (!mem_cgroup_is_root(swap_memcg))
7840 page_counter_charge(&swap_memcg->memsw, nr_entries);
7841 page_counter_uncharge(&memcg->memsw, nr_entries);
7845 * Interrupts should be disabled here because the caller holds the
7846 * i_pages lock which is taken with interrupts-off. It is
7847 * important here to have the interrupts disabled because it is the
7848 * only synchronisation we have for updating the per-CPU variables.
7851 mem_cgroup_charge_statistics(memcg, -nr_entries);
7852 memcg_stats_unlock();
7853 memcg_check_events(memcg, folio_nid(folio));
7855 css_put(&memcg->css);
7859 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7860 * @folio: folio being added to swap
7861 * @entry: swap entry to charge
7863 * Try to charge @folio's memcg for the swap space at @entry.
7865 * Returns 0 on success, -ENOMEM on failure.
7867 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7869 unsigned int nr_pages = folio_nr_pages(folio);
7870 struct page_counter *counter;
7871 struct mem_cgroup *memcg;
7872 unsigned short oldid;
7874 if (do_memsw_account())
7877 memcg = folio_memcg(folio);
7879 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7884 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7888 memcg = mem_cgroup_id_get_online(memcg);
7890 if (!mem_cgroup_is_root(memcg) &&
7891 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7892 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7893 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7894 mem_cgroup_id_put(memcg);
7898 /* Get references for the tail pages, too */
7900 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7901 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7902 VM_BUG_ON_FOLIO(oldid, folio);
7903 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7909 * __mem_cgroup_uncharge_swap - uncharge swap space
7910 * @entry: swap entry to uncharge
7911 * @nr_pages: the amount of swap space to uncharge
7913 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7915 struct mem_cgroup *memcg;
7918 id = swap_cgroup_record(entry, 0, nr_pages);
7920 memcg = mem_cgroup_from_id(id);
7922 if (!mem_cgroup_is_root(memcg)) {
7923 if (do_memsw_account())
7924 page_counter_uncharge(&memcg->memsw, nr_pages);
7926 page_counter_uncharge(&memcg->swap, nr_pages);
7928 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7929 mem_cgroup_id_put_many(memcg, nr_pages);
7934 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7936 long nr_swap_pages = get_nr_swap_pages();
7938 if (mem_cgroup_disabled() || do_memsw_account())
7939 return nr_swap_pages;
7940 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
7941 nr_swap_pages = min_t(long, nr_swap_pages,
7942 READ_ONCE(memcg->swap.max) -
7943 page_counter_read(&memcg->swap));
7944 return nr_swap_pages;
7947 bool mem_cgroup_swap_full(struct folio *folio)
7949 struct mem_cgroup *memcg;
7951 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7955 if (do_memsw_account())
7958 memcg = folio_memcg(folio);
7962 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
7963 unsigned long usage = page_counter_read(&memcg->swap);
7965 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7966 usage * 2 >= READ_ONCE(memcg->swap.max))
7973 static int __init setup_swap_account(char *s)
7977 if (!kstrtobool(s, &res) && !res)
7978 pr_warn_once("The swapaccount=0 commandline option is deprecated "
7979 "in favor of configuring swap control via cgroupfs. "
7980 "Please report your usecase to linux-mm@kvack.org if you "
7981 "depend on this functionality.\n");
7984 __setup("swapaccount=", setup_swap_account);
7986 static u64 swap_current_read(struct cgroup_subsys_state *css,
7989 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7991 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7994 static u64 swap_peak_read(struct cgroup_subsys_state *css,
7997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7999 return (u64)memcg->swap.watermark * PAGE_SIZE;
8002 static int swap_high_show(struct seq_file *m, void *v)
8004 return seq_puts_memcg_tunable(m,
8005 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
8008 static ssize_t swap_high_write(struct kernfs_open_file *of,
8009 char *buf, size_t nbytes, loff_t off)
8011 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8015 buf = strstrip(buf);
8016 err = page_counter_memparse(buf, "max", &high);
8020 page_counter_set_high(&memcg->swap, high);
8025 static int swap_max_show(struct seq_file *m, void *v)
8027 return seq_puts_memcg_tunable(m,
8028 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
8031 static ssize_t swap_max_write(struct kernfs_open_file *of,
8032 char *buf, size_t nbytes, loff_t off)
8034 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8038 buf = strstrip(buf);
8039 err = page_counter_memparse(buf, "max", &max);
8043 xchg(&memcg->swap.max, max);
8048 static int swap_events_show(struct seq_file *m, void *v)
8050 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
8052 seq_printf(m, "high %lu\n",
8053 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
8054 seq_printf(m, "max %lu\n",
8055 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
8056 seq_printf(m, "fail %lu\n",
8057 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
8062 static struct cftype swap_files[] = {
8064 .name = "swap.current",
8065 .flags = CFTYPE_NOT_ON_ROOT,
8066 .read_u64 = swap_current_read,
8069 .name = "swap.high",
8070 .flags = CFTYPE_NOT_ON_ROOT,
8071 .seq_show = swap_high_show,
8072 .write = swap_high_write,
8076 .flags = CFTYPE_NOT_ON_ROOT,
8077 .seq_show = swap_max_show,
8078 .write = swap_max_write,
8081 .name = "swap.peak",
8082 .flags = CFTYPE_NOT_ON_ROOT,
8083 .read_u64 = swap_peak_read,
8086 .name = "swap.events",
8087 .flags = CFTYPE_NOT_ON_ROOT,
8088 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
8089 .seq_show = swap_events_show,
8094 static struct cftype memsw_files[] = {
8096 .name = "memsw.usage_in_bytes",
8097 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
8098 .read_u64 = mem_cgroup_read_u64,
8101 .name = "memsw.max_usage_in_bytes",
8102 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
8103 .write = mem_cgroup_reset,
8104 .read_u64 = mem_cgroup_read_u64,
8107 .name = "memsw.limit_in_bytes",
8108 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
8109 .write = mem_cgroup_write,
8110 .read_u64 = mem_cgroup_read_u64,
8113 .name = "memsw.failcnt",
8114 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
8115 .write = mem_cgroup_reset,
8116 .read_u64 = mem_cgroup_read_u64,
8118 { }, /* terminate */
8121 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8123 * obj_cgroup_may_zswap - check if this cgroup can zswap
8124 * @objcg: the object cgroup
8126 * Check if the hierarchical zswap limit has been reached.
8128 * This doesn't check for specific headroom, and it is not atomic
8129 * either. But with zswap, the size of the allocation is only known
8130 * once compression has occurred, and this optimistic pre-check avoids
8131 * spending cycles on compression when there is already no room left
8132 * or zswap is disabled altogether somewhere in the hierarchy.
8134 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
8136 struct mem_cgroup *memcg, *original_memcg;
8139 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8142 original_memcg = get_mem_cgroup_from_objcg(objcg);
8143 for (memcg = original_memcg; !mem_cgroup_is_root(memcg);
8144 memcg = parent_mem_cgroup(memcg)) {
8145 unsigned long max = READ_ONCE(memcg->zswap_max);
8146 unsigned long pages;
8148 if (max == PAGE_COUNTER_MAX)
8156 * mem_cgroup_flush_stats() ignores small changes. Use
8157 * do_flush_stats() directly to get accurate stats for charging.
8159 do_flush_stats(memcg);
8160 pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
8166 mem_cgroup_put(original_memcg);
8171 * obj_cgroup_charge_zswap - charge compression backend memory
8172 * @objcg: the object cgroup
8173 * @size: size of compressed object
8175 * This forces the charge after obj_cgroup_may_zswap() allowed
8176 * compression and storage in zwap for this cgroup to go ahead.
8178 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
8180 struct mem_cgroup *memcg;
8182 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8185 VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
8187 /* PF_MEMALLOC context, charging must succeed */
8188 if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
8192 memcg = obj_cgroup_memcg(objcg);
8193 mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
8194 mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
8199 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
8200 * @objcg: the object cgroup
8201 * @size: size of compressed object
8203 * Uncharges zswap memory on page in.
8205 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
8207 struct mem_cgroup *memcg;
8209 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
8212 obj_cgroup_uncharge(objcg, size);
8215 memcg = obj_cgroup_memcg(objcg);
8216 mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
8217 mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
8221 bool mem_cgroup_zswap_writeback_enabled(struct mem_cgroup *memcg)
8223 /* if zswap is disabled, do not block pages going to the swapping device */
8224 return !is_zswap_enabled() || !memcg || READ_ONCE(memcg->zswap_writeback);
8227 static u64 zswap_current_read(struct cgroup_subsys_state *css,
8230 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
8232 mem_cgroup_flush_stats(memcg);
8233 return memcg_page_state(memcg, MEMCG_ZSWAP_B);
8236 static int zswap_max_show(struct seq_file *m, void *v)
8238 return seq_puts_memcg_tunable(m,
8239 READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
8242 static ssize_t zswap_max_write(struct kernfs_open_file *of,
8243 char *buf, size_t nbytes, loff_t off)
8245 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8249 buf = strstrip(buf);
8250 err = page_counter_memparse(buf, "max", &max);
8254 xchg(&memcg->zswap_max, max);
8259 static int zswap_writeback_show(struct seq_file *m, void *v)
8261 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
8263 seq_printf(m, "%d\n", READ_ONCE(memcg->zswap_writeback));
8267 static ssize_t zswap_writeback_write(struct kernfs_open_file *of,
8268 char *buf, size_t nbytes, loff_t off)
8270 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
8271 int zswap_writeback;
8272 ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_writeback);
8277 if (zswap_writeback != 0 && zswap_writeback != 1)
8280 WRITE_ONCE(memcg->zswap_writeback, zswap_writeback);
8284 static struct cftype zswap_files[] = {
8286 .name = "zswap.current",
8287 .flags = CFTYPE_NOT_ON_ROOT,
8288 .read_u64 = zswap_current_read,
8291 .name = "zswap.max",
8292 .flags = CFTYPE_NOT_ON_ROOT,
8293 .seq_show = zswap_max_show,
8294 .write = zswap_max_write,
8297 .name = "zswap.writeback",
8298 .seq_show = zswap_writeback_show,
8299 .write = zswap_writeback_write,
8303 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
8305 static int __init mem_cgroup_swap_init(void)
8307 if (mem_cgroup_disabled())
8310 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
8311 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
8312 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
8313 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
8317 subsys_initcall(mem_cgroup_swap_init);
8319 #endif /* CONFIG_SWAP */