1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
3 *
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 *
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
9 *
10 * Memory thresholds
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
13 *
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
17 *
18 * Native page reclaim
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 *
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
26 */
27
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
53 #include <linux/fs.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/memremap.h>
57 #include <linux/mm_inline.h>
58 #include <linux/swap_cgroup.h>
59 #include <linux/cpu.h>
60 #include <linux/oom.h>
61 #include <linux/lockdep.h>
62 #include <linux/file.h>
63 #include <linux/resume_user_mode.h>
64 #include <linux/psi.h>
65 #include <linux/seq_buf.h>
66 #include <linux/sched/isolation.h>
67 #include "internal.h"
68 #include <net/sock.h>
69 #include <net/ip.h>
70 #include "slab.h"
71 #include "swap.h"
72
73 #include <linux/uaccess.h>
74
75 #include <trace/events/vmscan.h>
76
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
79
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
81
82 /* Active memory cgroup to use from an interrupt context */
83 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
84 EXPORT_PER_CPU_SYMBOL_GPL(int_active_memcg);
85
86 /* Socket memory accounting disabled? */
87 static bool cgroup_memory_nosocket __ro_after_init;
88
89 /* Kernel memory accounting disabled? */
90 static bool cgroup_memory_nokmem __ro_after_init;
91
92 /* BPF memory accounting disabled? */
93 static bool cgroup_memory_nobpf __ro_after_init;
94
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
97 #endif
98
99 /* Whether legacy memory+swap accounting is active */
do_memsw_account(void)100 static bool do_memsw_account(void)
101 {
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
103 }
104
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
107
108 /*
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
111 */
112
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
116 spinlock_t lock;
117 };
118
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
121 };
122
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
124
125 /* for OOM */
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
129 };
130
131 /*
132 * cgroup_event represents events which userspace want to receive.
133 */
134 struct mem_cgroup_event {
135 /*
136 * memcg which the event belongs to.
137 */
138 struct mem_cgroup *memcg;
139 /*
140 * eventfd to signal userspace about the event.
141 */
142 struct eventfd_ctx *eventfd;
143 /*
144 * Each of these stored in a list by the cgroup.
145 */
146 struct list_head list;
147 /*
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
151 */
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
154 /*
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
158 */
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
161 /*
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
164 */
165 poll_table pt;
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
169 };
170
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
173
174 /* Stuffs for move charges at task migration. */
175 /*
176 * Types of charges to be moved.
177 */
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
181
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
188 unsigned long flags;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
194 } mc = {
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 };
198
199 /*
200 * Maximum loops in mem_cgroup_soft_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
202 */
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
205
206 /* for encoding cft->private value on file */
207 enum res_type {
208 _MEM,
209 _MEMSWAP,
210 _KMEM,
211 _TCP,
212 };
213
214 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
215 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
216 #define MEMFILE_ATTR(val) ((val) & 0xffff)
217
218 /*
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
222 */
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
225 iter != NULL; \
226 iter = mem_cgroup_iter(root, iter, NULL))
227
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
230 iter != NULL; \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
232
task_is_dying(void)233 static inline bool task_is_dying(void)
234 {
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
237 }
238
239 /* Some nice accessors for the vmpressure. */
memcg_to_vmpressure(struct mem_cgroup * memcg)240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
241 {
242 if (!memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
245 }
246
vmpressure_to_memcg(struct vmpressure * vmpr)247 struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr)
248 {
249 return container_of(vmpr, struct mem_cgroup, vmpressure);
250 }
251
252 #ifdef CONFIG_MEMCG_KMEM
253 static DEFINE_SPINLOCK(objcg_lock);
254
mem_cgroup_kmem_disabled(void)255 bool mem_cgroup_kmem_disabled(void)
256 {
257 return cgroup_memory_nokmem;
258 }
259
260 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
261 unsigned int nr_pages);
262
obj_cgroup_release(struct percpu_ref * ref)263 static void obj_cgroup_release(struct percpu_ref *ref)
264 {
265 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
266 unsigned int nr_bytes;
267 unsigned int nr_pages;
268 unsigned long flags;
269
270 /*
271 * At this point all allocated objects are freed, and
272 * objcg->nr_charged_bytes can't have an arbitrary byte value.
273 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
274 *
275 * The following sequence can lead to it:
276 * 1) CPU0: objcg == stock->cached_objcg
277 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
278 * PAGE_SIZE bytes are charged
279 * 3) CPU1: a process from another memcg is allocating something,
280 * the stock if flushed,
281 * objcg->nr_charged_bytes = PAGE_SIZE - 92
282 * 5) CPU0: we do release this object,
283 * 92 bytes are added to stock->nr_bytes
284 * 6) CPU0: stock is flushed,
285 * 92 bytes are added to objcg->nr_charged_bytes
286 *
287 * In the result, nr_charged_bytes == PAGE_SIZE.
288 * This page will be uncharged in obj_cgroup_release().
289 */
290 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
291 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
292 nr_pages = nr_bytes >> PAGE_SHIFT;
293
294 if (nr_pages)
295 obj_cgroup_uncharge_pages(objcg, nr_pages);
296
297 spin_lock_irqsave(&objcg_lock, flags);
298 list_del(&objcg->list);
299 spin_unlock_irqrestore(&objcg_lock, flags);
300
301 percpu_ref_exit(ref);
302 kfree_rcu(objcg, rcu);
303 }
304
obj_cgroup_alloc(void)305 static struct obj_cgroup *obj_cgroup_alloc(void)
306 {
307 struct obj_cgroup *objcg;
308 int ret;
309
310 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
311 if (!objcg)
312 return NULL;
313
314 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
315 GFP_KERNEL);
316 if (ret) {
317 kfree(objcg);
318 return NULL;
319 }
320 INIT_LIST_HEAD(&objcg->list);
321 return objcg;
322 }
323
memcg_reparent_objcgs(struct mem_cgroup * memcg,struct mem_cgroup * parent)324 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
325 struct mem_cgroup *parent)
326 {
327 struct obj_cgroup *objcg, *iter;
328
329 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
330
331 spin_lock_irq(&objcg_lock);
332
333 /* 1) Ready to reparent active objcg. */
334 list_add(&objcg->list, &memcg->objcg_list);
335 /* 2) Reparent active objcg and already reparented objcgs to parent. */
336 list_for_each_entry(iter, &memcg->objcg_list, list)
337 WRITE_ONCE(iter->memcg, parent);
338 /* 3) Move already reparented objcgs to the parent's list */
339 list_splice(&memcg->objcg_list, &parent->objcg_list);
340
341 spin_unlock_irq(&objcg_lock);
342
343 percpu_ref_kill(&objcg->refcnt);
344 }
345
346 /*
347 * A lot of the calls to the cache allocation functions are expected to be
348 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
349 * conditional to this static branch, we'll have to allow modules that does
350 * kmem_cache_alloc and the such to see this symbol as well
351 */
352 DEFINE_STATIC_KEY_FALSE(memcg_kmem_online_key);
353 EXPORT_SYMBOL(memcg_kmem_online_key);
354
355 DEFINE_STATIC_KEY_FALSE(memcg_bpf_enabled_key);
356 EXPORT_SYMBOL(memcg_bpf_enabled_key);
357 #endif
358
359 /**
360 * mem_cgroup_css_from_folio - css of the memcg associated with a folio
361 * @folio: folio of interest
362 *
363 * If memcg is bound to the default hierarchy, css of the memcg associated
364 * with @folio is returned. The returned css remains associated with @folio
365 * until it is released.
366 *
367 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
368 * is returned.
369 */
mem_cgroup_css_from_folio(struct folio * folio)370 struct cgroup_subsys_state *mem_cgroup_css_from_folio(struct folio *folio)
371 {
372 struct mem_cgroup *memcg = folio_memcg(folio);
373
374 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
375 memcg = root_mem_cgroup;
376
377 return &memcg->css;
378 }
379
380 /**
381 * page_cgroup_ino - return inode number of the memcg a page is charged to
382 * @page: the page
383 *
384 * Look up the closest online ancestor of the memory cgroup @page is charged to
385 * and return its inode number or 0 if @page is not charged to any cgroup. It
386 * is safe to call this function without holding a reference to @page.
387 *
388 * Note, this function is inherently racy, because there is nothing to prevent
389 * the cgroup inode from getting torn down and potentially reallocated a moment
390 * after page_cgroup_ino() returns, so it only should be used by callers that
391 * do not care (such as procfs interfaces).
392 */
page_cgroup_ino(struct page * page)393 ino_t page_cgroup_ino(struct page *page)
394 {
395 struct mem_cgroup *memcg;
396 unsigned long ino = 0;
397
398 rcu_read_lock();
399 /* page_folio() is racy here, but the entire function is racy anyway */
400 memcg = folio_memcg_check(page_folio(page));
401
402 while (memcg && !(memcg->css.flags & CSS_ONLINE))
403 memcg = parent_mem_cgroup(memcg);
404 if (memcg)
405 ino = cgroup_ino(memcg->css.cgroup);
406 rcu_read_unlock();
407 return ino;
408 }
409
__mem_cgroup_insert_exceeded(struct mem_cgroup_per_node * mz,struct mem_cgroup_tree_per_node * mctz,unsigned long new_usage_in_excess)410 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
411 struct mem_cgroup_tree_per_node *mctz,
412 unsigned long new_usage_in_excess)
413 {
414 struct rb_node **p = &mctz->rb_root.rb_node;
415 struct rb_node *parent = NULL;
416 struct mem_cgroup_per_node *mz_node;
417 bool rightmost = true;
418
419 if (mz->on_tree)
420 return;
421
422 mz->usage_in_excess = new_usage_in_excess;
423 if (!mz->usage_in_excess)
424 return;
425 while (*p) {
426 parent = *p;
427 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
428 tree_node);
429 if (mz->usage_in_excess < mz_node->usage_in_excess) {
430 p = &(*p)->rb_left;
431 rightmost = false;
432 } else {
433 p = &(*p)->rb_right;
434 }
435 }
436
437 if (rightmost)
438 mctz->rb_rightmost = &mz->tree_node;
439
440 rb_link_node(&mz->tree_node, parent, p);
441 rb_insert_color(&mz->tree_node, &mctz->rb_root);
442 mz->on_tree = true;
443 }
444
__mem_cgroup_remove_exceeded(struct mem_cgroup_per_node * mz,struct mem_cgroup_tree_per_node * mctz)445 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
446 struct mem_cgroup_tree_per_node *mctz)
447 {
448 if (!mz->on_tree)
449 return;
450
451 if (&mz->tree_node == mctz->rb_rightmost)
452 mctz->rb_rightmost = rb_prev(&mz->tree_node);
453
454 rb_erase(&mz->tree_node, &mctz->rb_root);
455 mz->on_tree = false;
456 }
457
mem_cgroup_remove_exceeded(struct mem_cgroup_per_node * mz,struct mem_cgroup_tree_per_node * mctz)458 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
459 struct mem_cgroup_tree_per_node *mctz)
460 {
461 unsigned long flags;
462
463 spin_lock_irqsave(&mctz->lock, flags);
464 __mem_cgroup_remove_exceeded(mz, mctz);
465 spin_unlock_irqrestore(&mctz->lock, flags);
466 }
467
soft_limit_excess(struct mem_cgroup * memcg)468 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
469 {
470 unsigned long nr_pages = page_counter_read(&memcg->memory);
471 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
472 unsigned long excess = 0;
473
474 if (nr_pages > soft_limit)
475 excess = nr_pages - soft_limit;
476
477 return excess;
478 }
479
mem_cgroup_update_tree(struct mem_cgroup * memcg,int nid)480 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, int nid)
481 {
482 unsigned long excess;
483 struct mem_cgroup_per_node *mz;
484 struct mem_cgroup_tree_per_node *mctz;
485
486 if (lru_gen_enabled()) {
487 if (soft_limit_excess(memcg))
488 lru_gen_soft_reclaim(memcg, nid);
489 return;
490 }
491
492 mctz = soft_limit_tree.rb_tree_per_node[nid];
493 if (!mctz)
494 return;
495 /*
496 * Necessary to update all ancestors when hierarchy is used.
497 * because their event counter is not touched.
498 */
499 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
500 mz = memcg->nodeinfo[nid];
501 excess = soft_limit_excess(memcg);
502 /*
503 * We have to update the tree if mz is on RB-tree or
504 * mem is over its softlimit.
505 */
506 if (excess || mz->on_tree) {
507 unsigned long flags;
508
509 spin_lock_irqsave(&mctz->lock, flags);
510 /* if on-tree, remove it */
511 if (mz->on_tree)
512 __mem_cgroup_remove_exceeded(mz, mctz);
513 /*
514 * Insert again. mz->usage_in_excess will be updated.
515 * If excess is 0, no tree ops.
516 */
517 __mem_cgroup_insert_exceeded(mz, mctz, excess);
518 spin_unlock_irqrestore(&mctz->lock, flags);
519 }
520 }
521 }
522
mem_cgroup_remove_from_trees(struct mem_cgroup * memcg)523 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
524 {
525 struct mem_cgroup_tree_per_node *mctz;
526 struct mem_cgroup_per_node *mz;
527 int nid;
528
529 for_each_node(nid) {
530 mz = memcg->nodeinfo[nid];
531 mctz = soft_limit_tree.rb_tree_per_node[nid];
532 if (mctz)
533 mem_cgroup_remove_exceeded(mz, mctz);
534 }
535 }
536
537 static struct mem_cgroup_per_node *
__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node * mctz)538 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
539 {
540 struct mem_cgroup_per_node *mz;
541
542 retry:
543 mz = NULL;
544 if (!mctz->rb_rightmost)
545 goto done; /* Nothing to reclaim from */
546
547 mz = rb_entry(mctz->rb_rightmost,
548 struct mem_cgroup_per_node, tree_node);
549 /*
550 * Remove the node now but someone else can add it back,
551 * we will to add it back at the end of reclaim to its correct
552 * position in the tree.
553 */
554 __mem_cgroup_remove_exceeded(mz, mctz);
555 if (!soft_limit_excess(mz->memcg) ||
556 !css_tryget(&mz->memcg->css))
557 goto retry;
558 done:
559 return mz;
560 }
561
562 static struct mem_cgroup_per_node *
mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node * mctz)563 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
564 {
565 struct mem_cgroup_per_node *mz;
566
567 spin_lock_irq(&mctz->lock);
568 mz = __mem_cgroup_largest_soft_limit_node(mctz);
569 spin_unlock_irq(&mctz->lock);
570 return mz;
571 }
572
573 /*
574 * memcg and lruvec stats flushing
575 *
576 * Many codepaths leading to stats update or read are performance sensitive and
577 * adding stats flushing in such codepaths is not desirable. So, to optimize the
578 * flushing the kernel does:
579 *
580 * 1) Periodically and asynchronously flush the stats every 2 seconds to not let
581 * rstat update tree grow unbounded.
582 *
583 * 2) Flush the stats synchronously on reader side only when there are more than
584 * (MEMCG_CHARGE_BATCH * nr_cpus) update events. Though this optimization
585 * will let stats be out of sync by atmost (MEMCG_CHARGE_BATCH * nr_cpus) but
586 * only for 2 seconds due to (1).
587 */
588 static void flush_memcg_stats_dwork(struct work_struct *w);
589 static DECLARE_DEFERRABLE_WORK(stats_flush_dwork, flush_memcg_stats_dwork);
590 static DEFINE_PER_CPU(unsigned int, stats_updates);
591 static atomic_t stats_flush_ongoing = ATOMIC_INIT(0);
592 static atomic_t stats_flush_threshold = ATOMIC_INIT(0);
593 static u64 flush_next_time;
594
595 #define FLUSH_TIME (2UL*HZ)
596
597 /*
598 * Accessors to ensure that preemption is disabled on PREEMPT_RT because it can
599 * not rely on this as part of an acquired spinlock_t lock. These functions are
600 * never used in hardirq context on PREEMPT_RT and therefore disabling preemtion
601 * is sufficient.
602 */
memcg_stats_lock(void)603 static void memcg_stats_lock(void)
604 {
605 preempt_disable_nested();
606 VM_WARN_ON_IRQS_ENABLED();
607 }
608
__memcg_stats_lock(void)609 static void __memcg_stats_lock(void)
610 {
611 preempt_disable_nested();
612 }
613
memcg_stats_unlock(void)614 static void memcg_stats_unlock(void)
615 {
616 preempt_enable_nested();
617 }
618
memcg_rstat_updated(struct mem_cgroup * memcg,int val)619 static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
620 {
621 unsigned int x;
622
623 if (!val)
624 return;
625
626 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
627
628 x = __this_cpu_add_return(stats_updates, abs(val));
629 if (x > MEMCG_CHARGE_BATCH) {
630 /*
631 * If stats_flush_threshold exceeds the threshold
632 * (>num_online_cpus()), cgroup stats update will be triggered
633 * in __mem_cgroup_flush_stats(). Increasing this var further
634 * is redundant and simply adds overhead in atomic update.
635 */
636 if (atomic_read(&stats_flush_threshold) <= num_online_cpus())
637 atomic_add(x / MEMCG_CHARGE_BATCH, &stats_flush_threshold);
638 __this_cpu_write(stats_updates, 0);
639 }
640 }
641
do_flush_stats(void)642 static void do_flush_stats(void)
643 {
644 /*
645 * We always flush the entire tree, so concurrent flushers can just
646 * skip. This avoids a thundering herd problem on the rstat global lock
647 * from memcg flushers (e.g. reclaim, refault, etc).
648 */
649 if (atomic_read(&stats_flush_ongoing) ||
650 atomic_xchg(&stats_flush_ongoing, 1))
651 return;
652
653 WRITE_ONCE(flush_next_time, jiffies_64 + 2*FLUSH_TIME);
654
655 cgroup_rstat_flush(root_mem_cgroup->css.cgroup);
656
657 atomic_set(&stats_flush_threshold, 0);
658 atomic_set(&stats_flush_ongoing, 0);
659 }
660
mem_cgroup_flush_stats(void)661 void mem_cgroup_flush_stats(void)
662 {
663 if (atomic_read(&stats_flush_threshold) > num_online_cpus())
664 do_flush_stats();
665 }
666
mem_cgroup_flush_stats_ratelimited(void)667 void mem_cgroup_flush_stats_ratelimited(void)
668 {
669 if (time_after64(jiffies_64, READ_ONCE(flush_next_time)))
670 mem_cgroup_flush_stats();
671 }
672
flush_memcg_stats_dwork(struct work_struct * w)673 static void flush_memcg_stats_dwork(struct work_struct *w)
674 {
675 /*
676 * Always flush here so that flushing in latency-sensitive paths is
677 * as cheap as possible.
678 */
679 do_flush_stats();
680 queue_delayed_work(system_unbound_wq, &stats_flush_dwork, FLUSH_TIME);
681 }
682
683 /* Subset of vm_event_item to report for memcg event stats */
684 static const unsigned int memcg_vm_event_stat[] = {
685 PGPGIN,
686 PGPGOUT,
687 PGSCAN_KSWAPD,
688 PGSCAN_DIRECT,
689 PGSCAN_KHUGEPAGED,
690 PGSTEAL_KSWAPD,
691 PGSTEAL_DIRECT,
692 PGSTEAL_KHUGEPAGED,
693 PGFAULT,
694 PGMAJFAULT,
695 PGREFILL,
696 PGACTIVATE,
697 PGDEACTIVATE,
698 PGLAZYFREE,
699 PGLAZYFREED,
700 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
701 ZSWPIN,
702 ZSWPOUT,
703 #endif
704 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
705 THP_FAULT_ALLOC,
706 THP_COLLAPSE_ALLOC,
707 #endif
708 };
709
710 #define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat)
711 static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly;
712
init_memcg_events(void)713 static void init_memcg_events(void)
714 {
715 int i;
716
717 for (i = 0; i < NR_MEMCG_EVENTS; ++i)
718 mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1;
719 }
720
memcg_events_index(enum vm_event_item idx)721 static inline int memcg_events_index(enum vm_event_item idx)
722 {
723 return mem_cgroup_events_index[idx] - 1;
724 }
725
726 struct memcg_vmstats_percpu {
727 /* Local (CPU and cgroup) page state & events */
728 long state[MEMCG_NR_STAT];
729 unsigned long events[NR_MEMCG_EVENTS];
730
731 /* Delta calculation for lockless upward propagation */
732 long state_prev[MEMCG_NR_STAT];
733 unsigned long events_prev[NR_MEMCG_EVENTS];
734
735 /* Cgroup1: threshold notifications & softlimit tree updates */
736 unsigned long nr_page_events;
737 unsigned long targets[MEM_CGROUP_NTARGETS];
738 };
739
740 struct memcg_vmstats {
741 /* Aggregated (CPU and subtree) page state & events */
742 long state[MEMCG_NR_STAT];
743 unsigned long events[NR_MEMCG_EVENTS];
744
745 /* Non-hierarchical (CPU aggregated) page state & events */
746 long state_local[MEMCG_NR_STAT];
747 unsigned long events_local[NR_MEMCG_EVENTS];
748
749 /* Pending child counts during tree propagation */
750 long state_pending[MEMCG_NR_STAT];
751 unsigned long events_pending[NR_MEMCG_EVENTS];
752 };
753
memcg_page_state(struct mem_cgroup * memcg,int idx)754 unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
755 {
756 long x = READ_ONCE(memcg->vmstats->state[idx]);
757 #ifdef CONFIG_SMP
758 if (x < 0)
759 x = 0;
760 #endif
761 return x;
762 }
763
764 /**
765 * __mod_memcg_state - update cgroup memory statistics
766 * @memcg: the memory cgroup
767 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
768 * @val: delta to add to the counter, can be negative
769 */
__mod_memcg_state(struct mem_cgroup * memcg,int idx,int val)770 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
771 {
772 if (mem_cgroup_disabled())
773 return;
774
775 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
776 memcg_rstat_updated(memcg, val);
777 }
778
779 /* idx can be of type enum memcg_stat_item or node_stat_item. */
memcg_page_state_local(struct mem_cgroup * memcg,int idx)780 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
781 {
782 long x = READ_ONCE(memcg->vmstats->state_local[idx]);
783
784 #ifdef CONFIG_SMP
785 if (x < 0)
786 x = 0;
787 #endif
788 return x;
789 }
790
__mod_memcg_lruvec_state(struct lruvec * lruvec,enum node_stat_item idx,int val)791 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
792 int val)
793 {
794 struct mem_cgroup_per_node *pn;
795 struct mem_cgroup *memcg;
796
797 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
798 memcg = pn->memcg;
799
800 /*
801 * The caller from rmap relay on disabled preemption becase they never
802 * update their counter from in-interrupt context. For these two
803 * counters we check that the update is never performed from an
804 * interrupt context while other caller need to have disabled interrupt.
805 */
806 __memcg_stats_lock();
807 if (IS_ENABLED(CONFIG_DEBUG_VM)) {
808 switch (idx) {
809 case NR_ANON_MAPPED:
810 case NR_FILE_MAPPED:
811 case NR_ANON_THPS:
812 case NR_SHMEM_PMDMAPPED:
813 case NR_FILE_PMDMAPPED:
814 WARN_ON_ONCE(!in_task());
815 break;
816 default:
817 VM_WARN_ON_IRQS_ENABLED();
818 }
819 }
820
821 /* Update memcg */
822 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
823
824 /* Update lruvec */
825 __this_cpu_add(pn->lruvec_stats_percpu->state[idx], val);
826
827 memcg_rstat_updated(memcg, val);
828 memcg_stats_unlock();
829 }
830
831 /**
832 * __mod_lruvec_state - update lruvec memory statistics
833 * @lruvec: the lruvec
834 * @idx: the stat item
835 * @val: delta to add to the counter, can be negative
836 *
837 * The lruvec is the intersection of the NUMA node and a cgroup. This
838 * function updates the all three counters that are affected by a
839 * change of state at this level: per-node, per-cgroup, per-lruvec.
840 */
__mod_lruvec_state(struct lruvec * lruvec,enum node_stat_item idx,int val)841 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
842 int val)
843 {
844 /* Update node */
845 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
846
847 /* Update memcg and lruvec */
848 if (!mem_cgroup_disabled())
849 __mod_memcg_lruvec_state(lruvec, idx, val);
850 }
851
__mod_lruvec_page_state(struct page * page,enum node_stat_item idx,int val)852 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
853 int val)
854 {
855 struct page *head = compound_head(page); /* rmap on tail pages */
856 struct mem_cgroup *memcg;
857 pg_data_t *pgdat = page_pgdat(page);
858 struct lruvec *lruvec;
859
860 rcu_read_lock();
861 memcg = page_memcg(head);
862 /* Untracked pages have no memcg, no lruvec. Update only the node */
863 if (!memcg) {
864 rcu_read_unlock();
865 __mod_node_page_state(pgdat, idx, val);
866 return;
867 }
868
869 lruvec = mem_cgroup_lruvec(memcg, pgdat);
870 __mod_lruvec_state(lruvec, idx, val);
871 rcu_read_unlock();
872 }
873 EXPORT_SYMBOL(__mod_lruvec_page_state);
874
__mod_lruvec_kmem_state(void * p,enum node_stat_item idx,int val)875 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
876 {
877 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
878 struct mem_cgroup *memcg;
879 struct lruvec *lruvec;
880
881 rcu_read_lock();
882 memcg = mem_cgroup_from_slab_obj(p);
883
884 /*
885 * Untracked pages have no memcg, no lruvec. Update only the
886 * node. If we reparent the slab objects to the root memcg,
887 * when we free the slab object, we need to update the per-memcg
888 * vmstats to keep it correct for the root memcg.
889 */
890 if (!memcg) {
891 __mod_node_page_state(pgdat, idx, val);
892 } else {
893 lruvec = mem_cgroup_lruvec(memcg, pgdat);
894 __mod_lruvec_state(lruvec, idx, val);
895 }
896 rcu_read_unlock();
897 }
898
899 /**
900 * __count_memcg_events - account VM events in a cgroup
901 * @memcg: the memory cgroup
902 * @idx: the event item
903 * @count: the number of events that occurred
904 */
__count_memcg_events(struct mem_cgroup * memcg,enum vm_event_item idx,unsigned long count)905 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
906 unsigned long count)
907 {
908 int index = memcg_events_index(idx);
909
910 if (mem_cgroup_disabled() || index < 0)
911 return;
912
913 memcg_stats_lock();
914 __this_cpu_add(memcg->vmstats_percpu->events[index], count);
915 memcg_rstat_updated(memcg, count);
916 memcg_stats_unlock();
917 }
918
memcg_events(struct mem_cgroup * memcg,int event)919 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
920 {
921 int index = memcg_events_index(event);
922
923 if (index < 0)
924 return 0;
925 return READ_ONCE(memcg->vmstats->events[index]);
926 }
927
memcg_events_local(struct mem_cgroup * memcg,int event)928 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
929 {
930 int index = memcg_events_index(event);
931
932 if (index < 0)
933 return 0;
934
935 return READ_ONCE(memcg->vmstats->events_local[index]);
936 }
937
mem_cgroup_charge_statistics(struct mem_cgroup * memcg,int nr_pages)938 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
939 int nr_pages)
940 {
941 /* pagein of a big page is an event. So, ignore page size */
942 if (nr_pages > 0)
943 __count_memcg_events(memcg, PGPGIN, 1);
944 else {
945 __count_memcg_events(memcg, PGPGOUT, 1);
946 nr_pages = -nr_pages; /* for event */
947 }
948
949 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
950 }
951
mem_cgroup_event_ratelimit(struct mem_cgroup * memcg,enum mem_cgroup_events_target target)952 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
953 enum mem_cgroup_events_target target)
954 {
955 unsigned long val, next;
956
957 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
958 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
959 /* from time_after() in jiffies.h */
960 if ((long)(next - val) < 0) {
961 switch (target) {
962 case MEM_CGROUP_TARGET_THRESH:
963 next = val + THRESHOLDS_EVENTS_TARGET;
964 break;
965 case MEM_CGROUP_TARGET_SOFTLIMIT:
966 next = val + SOFTLIMIT_EVENTS_TARGET;
967 break;
968 default:
969 break;
970 }
971 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
972 return true;
973 }
974 return false;
975 }
976
977 /*
978 * Check events in order.
979 *
980 */
memcg_check_events(struct mem_cgroup * memcg,int nid)981 static void memcg_check_events(struct mem_cgroup *memcg, int nid)
982 {
983 if (IS_ENABLED(CONFIG_PREEMPT_RT))
984 return;
985
986 /* threshold event is triggered in finer grain than soft limit */
987 if (unlikely(mem_cgroup_event_ratelimit(memcg,
988 MEM_CGROUP_TARGET_THRESH))) {
989 bool do_softlimit;
990
991 do_softlimit = mem_cgroup_event_ratelimit(memcg,
992 MEM_CGROUP_TARGET_SOFTLIMIT);
993 mem_cgroup_threshold(memcg);
994 if (unlikely(do_softlimit))
995 mem_cgroup_update_tree(memcg, nid);
996 }
997 }
998
mem_cgroup_from_task(struct task_struct * p)999 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1000 {
1001 /*
1002 * mm_update_next_owner() may clear mm->owner to NULL
1003 * if it races with swapoff, page migration, etc.
1004 * So this can be called with p == NULL.
1005 */
1006 if (unlikely(!p))
1007 return NULL;
1008
1009 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1010 }
1011 EXPORT_SYMBOL(mem_cgroup_from_task);
1012
active_memcg(void)1013 static __always_inline struct mem_cgroup *active_memcg(void)
1014 {
1015 if (!in_task())
1016 return this_cpu_read(int_active_memcg);
1017 else
1018 return current->active_memcg;
1019 }
1020
1021 /**
1022 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1023 * @mm: mm from which memcg should be extracted. It can be NULL.
1024 *
1025 * Obtain a reference on mm->memcg and returns it if successful. If mm
1026 * is NULL, then the memcg is chosen as follows:
1027 * 1) The active memcg, if set.
1028 * 2) current->mm->memcg, if available
1029 * 3) root memcg
1030 * If mem_cgroup is disabled, NULL is returned.
1031 */
get_mem_cgroup_from_mm(struct mm_struct * mm)1032 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1033 {
1034 struct mem_cgroup *memcg;
1035
1036 if (mem_cgroup_disabled())
1037 return NULL;
1038
1039 /*
1040 * Page cache insertions can happen without an
1041 * actual mm context, e.g. during disk probing
1042 * on boot, loopback IO, acct() writes etc.
1043 *
1044 * No need to css_get on root memcg as the reference
1045 * counting is disabled on the root level in the
1046 * cgroup core. See CSS_NO_REF.
1047 */
1048 if (unlikely(!mm)) {
1049 memcg = active_memcg();
1050 if (unlikely(memcg)) {
1051 /* remote memcg must hold a ref */
1052 css_get(&memcg->css);
1053 return memcg;
1054 }
1055 mm = current->mm;
1056 if (unlikely(!mm))
1057 return root_mem_cgroup;
1058 }
1059
1060 rcu_read_lock();
1061 do {
1062 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1063 if (unlikely(!memcg))
1064 memcg = root_mem_cgroup;
1065 } while (!css_tryget(&memcg->css));
1066 rcu_read_unlock();
1067 return memcg;
1068 }
1069 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1070
memcg_kmem_bypass(void)1071 static __always_inline bool memcg_kmem_bypass(void)
1072 {
1073 /* Allow remote memcg charging from any context. */
1074 if (unlikely(active_memcg()))
1075 return false;
1076
1077 /* Memcg to charge can't be determined. */
1078 if (!in_task() || !current->mm || (current->flags & PF_KTHREAD))
1079 return true;
1080
1081 return false;
1082 }
1083
1084 /**
1085 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1086 * @root: hierarchy root
1087 * @prev: previously returned memcg, NULL on first invocation
1088 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1089 *
1090 * Returns references to children of the hierarchy below @root, or
1091 * @root itself, or %NULL after a full round-trip.
1092 *
1093 * Caller must pass the return value in @prev on subsequent
1094 * invocations for reference counting, or use mem_cgroup_iter_break()
1095 * to cancel a hierarchy walk before the round-trip is complete.
1096 *
1097 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1098 * in the hierarchy among all concurrent reclaimers operating on the
1099 * same node.
1100 */
mem_cgroup_iter(struct mem_cgroup * root,struct mem_cgroup * prev,struct mem_cgroup_reclaim_cookie * reclaim)1101 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1102 struct mem_cgroup *prev,
1103 struct mem_cgroup_reclaim_cookie *reclaim)
1104 {
1105 struct mem_cgroup_reclaim_iter *iter;
1106 struct cgroup_subsys_state *css = NULL;
1107 struct mem_cgroup *memcg = NULL;
1108 struct mem_cgroup *pos = NULL;
1109
1110 if (mem_cgroup_disabled())
1111 return NULL;
1112
1113 if (!root)
1114 root = root_mem_cgroup;
1115
1116 rcu_read_lock();
1117
1118 if (reclaim) {
1119 struct mem_cgroup_per_node *mz;
1120
1121 mz = root->nodeinfo[reclaim->pgdat->node_id];
1122 iter = &mz->iter;
1123
1124 /*
1125 * On start, join the current reclaim iteration cycle.
1126 * Exit when a concurrent walker completes it.
1127 */
1128 if (!prev)
1129 reclaim->generation = iter->generation;
1130 else if (reclaim->generation != iter->generation)
1131 goto out_unlock;
1132
1133 while (1) {
1134 pos = READ_ONCE(iter->position);
1135 if (!pos || css_tryget(&pos->css))
1136 break;
1137 /*
1138 * css reference reached zero, so iter->position will
1139 * be cleared by ->css_released. However, we should not
1140 * rely on this happening soon, because ->css_released
1141 * is called from a work queue, and by busy-waiting we
1142 * might block it. So we clear iter->position right
1143 * away.
1144 */
1145 (void)cmpxchg(&iter->position, pos, NULL);
1146 }
1147 } else if (prev) {
1148 pos = prev;
1149 }
1150
1151 if (pos)
1152 css = &pos->css;
1153
1154 for (;;) {
1155 css = css_next_descendant_pre(css, &root->css);
1156 if (!css) {
1157 /*
1158 * Reclaimers share the hierarchy walk, and a
1159 * new one might jump in right at the end of
1160 * the hierarchy - make sure they see at least
1161 * one group and restart from the beginning.
1162 */
1163 if (!prev)
1164 continue;
1165 break;
1166 }
1167
1168 /*
1169 * Verify the css and acquire a reference. The root
1170 * is provided by the caller, so we know it's alive
1171 * and kicking, and don't take an extra reference.
1172 */
1173 if (css == &root->css || css_tryget(css)) {
1174 memcg = mem_cgroup_from_css(css);
1175 break;
1176 }
1177 }
1178
1179 if (reclaim) {
1180 /*
1181 * The position could have already been updated by a competing
1182 * thread, so check that the value hasn't changed since we read
1183 * it to avoid reclaiming from the same cgroup twice.
1184 */
1185 (void)cmpxchg(&iter->position, pos, memcg);
1186
1187 if (pos)
1188 css_put(&pos->css);
1189
1190 if (!memcg)
1191 iter->generation++;
1192 }
1193
1194 out_unlock:
1195 rcu_read_unlock();
1196 if (prev && prev != root)
1197 css_put(&prev->css);
1198
1199 return memcg;
1200 }
1201
1202 /**
1203 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1204 * @root: hierarchy root
1205 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1206 */
mem_cgroup_iter_break(struct mem_cgroup * root,struct mem_cgroup * prev)1207 void mem_cgroup_iter_break(struct mem_cgroup *root,
1208 struct mem_cgroup *prev)
1209 {
1210 if (!root)
1211 root = root_mem_cgroup;
1212 if (prev && prev != root)
1213 css_put(&prev->css);
1214 }
1215
__invalidate_reclaim_iterators(struct mem_cgroup * from,struct mem_cgroup * dead_memcg)1216 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1217 struct mem_cgroup *dead_memcg)
1218 {
1219 struct mem_cgroup_reclaim_iter *iter;
1220 struct mem_cgroup_per_node *mz;
1221 int nid;
1222
1223 for_each_node(nid) {
1224 mz = from->nodeinfo[nid];
1225 iter = &mz->iter;
1226 cmpxchg(&iter->position, dead_memcg, NULL);
1227 }
1228 }
1229
invalidate_reclaim_iterators(struct mem_cgroup * dead_memcg)1230 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1231 {
1232 struct mem_cgroup *memcg = dead_memcg;
1233 struct mem_cgroup *last;
1234
1235 do {
1236 __invalidate_reclaim_iterators(memcg, dead_memcg);
1237 last = memcg;
1238 } while ((memcg = parent_mem_cgroup(memcg)));
1239
1240 /*
1241 * When cgroup1 non-hierarchy mode is used,
1242 * parent_mem_cgroup() does not walk all the way up to the
1243 * cgroup root (root_mem_cgroup). So we have to handle
1244 * dead_memcg from cgroup root separately.
1245 */
1246 if (!mem_cgroup_is_root(last))
1247 __invalidate_reclaim_iterators(root_mem_cgroup,
1248 dead_memcg);
1249 }
1250
1251 /**
1252 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1253 * @memcg: hierarchy root
1254 * @fn: function to call for each task
1255 * @arg: argument passed to @fn
1256 *
1257 * This function iterates over tasks attached to @memcg or to any of its
1258 * descendants and calls @fn for each task. If @fn returns a non-zero
1259 * value, the function breaks the iteration loop. Otherwise, it will iterate
1260 * over all tasks and return 0.
1261 *
1262 * This function must not be called for the root memory cgroup.
1263 */
mem_cgroup_scan_tasks(struct mem_cgroup * memcg,int (* fn)(struct task_struct *,void *),void * arg)1264 void mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1265 int (*fn)(struct task_struct *, void *), void *arg)
1266 {
1267 struct mem_cgroup *iter;
1268 int ret = 0;
1269 int i = 0;
1270
1271 BUG_ON(mem_cgroup_is_root(memcg));
1272
1273 for_each_mem_cgroup_tree(iter, memcg) {
1274 struct css_task_iter it;
1275 struct task_struct *task;
1276
1277 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1278 while (!ret && (task = css_task_iter_next(&it))) {
1279 /* Avoid potential softlockup warning */
1280 if ((++i & 1023) == 0)
1281 cond_resched();
1282 ret = fn(task, arg);
1283 }
1284 css_task_iter_end(&it);
1285 if (ret) {
1286 mem_cgroup_iter_break(memcg, iter);
1287 break;
1288 }
1289 }
1290 }
1291
1292 #ifdef CONFIG_DEBUG_VM
lruvec_memcg_debug(struct lruvec * lruvec,struct folio * folio)1293 void lruvec_memcg_debug(struct lruvec *lruvec, struct folio *folio)
1294 {
1295 struct mem_cgroup *memcg;
1296
1297 if (mem_cgroup_disabled())
1298 return;
1299
1300 memcg = folio_memcg(folio);
1301
1302 if (!memcg)
1303 VM_BUG_ON_FOLIO(!mem_cgroup_is_root(lruvec_memcg(lruvec)), folio);
1304 else
1305 VM_BUG_ON_FOLIO(lruvec_memcg(lruvec) != memcg, folio);
1306 }
1307 #endif
1308
1309 /**
1310 * folio_lruvec_lock - Lock the lruvec for a folio.
1311 * @folio: Pointer to the folio.
1312 *
1313 * These functions are safe to use under any of the following conditions:
1314 * - folio locked
1315 * - folio_test_lru false
1316 * - folio_memcg_lock()
1317 * - folio frozen (refcount of 0)
1318 *
1319 * Return: The lruvec this folio is on with its lock held.
1320 */
folio_lruvec_lock(struct folio * folio)1321 struct lruvec *folio_lruvec_lock(struct folio *folio)
1322 {
1323 struct lruvec *lruvec = folio_lruvec(folio);
1324
1325 spin_lock(&lruvec->lru_lock);
1326 lruvec_memcg_debug(lruvec, folio);
1327
1328 return lruvec;
1329 }
1330
1331 /**
1332 * folio_lruvec_lock_irq - Lock the lruvec for a folio.
1333 * @folio: Pointer to the folio.
1334 *
1335 * These functions are safe to use under any of the following conditions:
1336 * - folio locked
1337 * - folio_test_lru false
1338 * - folio_memcg_lock()
1339 * - folio frozen (refcount of 0)
1340 *
1341 * Return: The lruvec this folio is on with its lock held and interrupts
1342 * disabled.
1343 */
folio_lruvec_lock_irq(struct folio * folio)1344 struct lruvec *folio_lruvec_lock_irq(struct folio *folio)
1345 {
1346 struct lruvec *lruvec = folio_lruvec(folio);
1347
1348 spin_lock_irq(&lruvec->lru_lock);
1349 lruvec_memcg_debug(lruvec, folio);
1350
1351 return lruvec;
1352 }
1353
1354 /**
1355 * folio_lruvec_lock_irqsave - Lock the lruvec for a folio.
1356 * @folio: Pointer to the folio.
1357 * @flags: Pointer to irqsave flags.
1358 *
1359 * These functions are safe to use under any of the following conditions:
1360 * - folio locked
1361 * - folio_test_lru false
1362 * - folio_memcg_lock()
1363 * - folio frozen (refcount of 0)
1364 *
1365 * Return: The lruvec this folio is on with its lock held and interrupts
1366 * disabled.
1367 */
folio_lruvec_lock_irqsave(struct folio * folio,unsigned long * flags)1368 struct lruvec *folio_lruvec_lock_irqsave(struct folio *folio,
1369 unsigned long *flags)
1370 {
1371 struct lruvec *lruvec = folio_lruvec(folio);
1372
1373 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1374 lruvec_memcg_debug(lruvec, folio);
1375
1376 return lruvec;
1377 }
1378
1379 /**
1380 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1381 * @lruvec: mem_cgroup per zone lru vector
1382 * @lru: index of lru list the page is sitting on
1383 * @zid: zone id of the accounted pages
1384 * @nr_pages: positive when adding or negative when removing
1385 *
1386 * This function must be called under lru_lock, just before a page is added
1387 * to or just after a page is removed from an lru list.
1388 */
mem_cgroup_update_lru_size(struct lruvec * lruvec,enum lru_list lru,int zid,int nr_pages)1389 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1390 int zid, int nr_pages)
1391 {
1392 struct mem_cgroup_per_node *mz;
1393 unsigned long *lru_size;
1394 long size;
1395
1396 if (mem_cgroup_disabled())
1397 return;
1398
1399 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1400 lru_size = &mz->lru_zone_size[zid][lru];
1401
1402 if (nr_pages < 0)
1403 *lru_size += nr_pages;
1404
1405 size = *lru_size;
1406 if (WARN_ONCE(size < 0,
1407 "%s(%p, %d, %d): lru_size %ld\n",
1408 __func__, lruvec, lru, nr_pages, size)) {
1409 VM_BUG_ON(1);
1410 *lru_size = 0;
1411 }
1412
1413 if (nr_pages > 0)
1414 *lru_size += nr_pages;
1415 }
1416
1417 /**
1418 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1419 * @memcg: the memory cgroup
1420 *
1421 * Returns the maximum amount of memory @mem can be charged with, in
1422 * pages.
1423 */
mem_cgroup_margin(struct mem_cgroup * memcg)1424 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1425 {
1426 unsigned long margin = 0;
1427 unsigned long count;
1428 unsigned long limit;
1429
1430 count = page_counter_read(&memcg->memory);
1431 limit = READ_ONCE(memcg->memory.max);
1432 if (count < limit)
1433 margin = limit - count;
1434
1435 if (do_memsw_account()) {
1436 count = page_counter_read(&memcg->memsw);
1437 limit = READ_ONCE(memcg->memsw.max);
1438 if (count < limit)
1439 margin = min(margin, limit - count);
1440 else
1441 margin = 0;
1442 }
1443
1444 return margin;
1445 }
1446
1447 /*
1448 * A routine for checking "mem" is under move_account() or not.
1449 *
1450 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1451 * moving cgroups. This is for waiting at high-memory pressure
1452 * caused by "move".
1453 */
mem_cgroup_under_move(struct mem_cgroup * memcg)1454 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1455 {
1456 struct mem_cgroup *from;
1457 struct mem_cgroup *to;
1458 bool ret = false;
1459 /*
1460 * Unlike task_move routines, we access mc.to, mc.from not under
1461 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1462 */
1463 spin_lock(&mc.lock);
1464 from = mc.from;
1465 to = mc.to;
1466 if (!from)
1467 goto unlock;
1468
1469 ret = mem_cgroup_is_descendant(from, memcg) ||
1470 mem_cgroup_is_descendant(to, memcg);
1471 unlock:
1472 spin_unlock(&mc.lock);
1473 return ret;
1474 }
1475
mem_cgroup_wait_acct_move(struct mem_cgroup * memcg)1476 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1477 {
1478 if (mc.moving_task && current != mc.moving_task) {
1479 if (mem_cgroup_under_move(memcg)) {
1480 DEFINE_WAIT(wait);
1481 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1482 /* moving charge context might have finished. */
1483 if (mc.moving_task)
1484 schedule();
1485 finish_wait(&mc.waitq, &wait);
1486 return true;
1487 }
1488 }
1489 return false;
1490 }
1491
1492 struct memory_stat {
1493 const char *name;
1494 unsigned int idx;
1495 };
1496
1497 static const struct memory_stat memory_stats[] = {
1498 { "anon", NR_ANON_MAPPED },
1499 { "file", NR_FILE_PAGES },
1500 { "kernel", MEMCG_KMEM },
1501 { "kernel_stack", NR_KERNEL_STACK_KB },
1502 { "pagetables", NR_PAGETABLE },
1503 { "sec_pagetables", NR_SECONDARY_PAGETABLE },
1504 { "percpu", MEMCG_PERCPU_B },
1505 { "sock", MEMCG_SOCK },
1506 { "vmalloc", MEMCG_VMALLOC },
1507 { "shmem", NR_SHMEM },
1508 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
1509 { "zswap", MEMCG_ZSWAP_B },
1510 { "zswapped", MEMCG_ZSWAPPED },
1511 #endif
1512 { "file_mapped", NR_FILE_MAPPED },
1513 { "file_dirty", NR_FILE_DIRTY },
1514 { "file_writeback", NR_WRITEBACK },
1515 #ifdef CONFIG_SWAP
1516 { "swapcached", NR_SWAPCACHE },
1517 #endif
1518 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1519 { "anon_thp", NR_ANON_THPS },
1520 { "file_thp", NR_FILE_THPS },
1521 { "shmem_thp", NR_SHMEM_THPS },
1522 #endif
1523 { "inactive_anon", NR_INACTIVE_ANON },
1524 { "active_anon", NR_ACTIVE_ANON },
1525 { "inactive_file", NR_INACTIVE_FILE },
1526 { "active_file", NR_ACTIVE_FILE },
1527 { "unevictable", NR_UNEVICTABLE },
1528 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1529 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1530
1531 /* The memory events */
1532 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1533 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1534 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1535 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1536 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1537 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1538 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1539 };
1540
1541 /* Translate stat items to the correct unit for memory.stat output */
memcg_page_state_unit(int item)1542 static int memcg_page_state_unit(int item)
1543 {
1544 switch (item) {
1545 case MEMCG_PERCPU_B:
1546 case MEMCG_ZSWAP_B:
1547 case NR_SLAB_RECLAIMABLE_B:
1548 case NR_SLAB_UNRECLAIMABLE_B:
1549 case WORKINGSET_REFAULT_ANON:
1550 case WORKINGSET_REFAULT_FILE:
1551 case WORKINGSET_ACTIVATE_ANON:
1552 case WORKINGSET_ACTIVATE_FILE:
1553 case WORKINGSET_RESTORE_ANON:
1554 case WORKINGSET_RESTORE_FILE:
1555 case WORKINGSET_NODERECLAIM:
1556 return 1;
1557 case NR_KERNEL_STACK_KB:
1558 return SZ_1K;
1559 default:
1560 return PAGE_SIZE;
1561 }
1562 }
1563
memcg_page_state_output(struct mem_cgroup * memcg,int item)1564 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1565 int item)
1566 {
1567 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1568 }
1569
memcg_stat_format(struct mem_cgroup * memcg,struct seq_buf * s)1570 static void memcg_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1571 {
1572 int i;
1573
1574 /*
1575 * Provide statistics on the state of the memory subsystem as
1576 * well as cumulative event counters that show past behavior.
1577 *
1578 * This list is ordered following a combination of these gradients:
1579 * 1) generic big picture -> specifics and details
1580 * 2) reflecting userspace activity -> reflecting kernel heuristics
1581 *
1582 * Current memory state:
1583 */
1584 mem_cgroup_flush_stats();
1585
1586 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1587 u64 size;
1588
1589 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1590 seq_buf_printf(s, "%s %llu\n", memory_stats[i].name, size);
1591
1592 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1593 size += memcg_page_state_output(memcg,
1594 NR_SLAB_RECLAIMABLE_B);
1595 seq_buf_printf(s, "slab %llu\n", size);
1596 }
1597 }
1598
1599 /* Accumulated memory events */
1600 seq_buf_printf(s, "pgscan %lu\n",
1601 memcg_events(memcg, PGSCAN_KSWAPD) +
1602 memcg_events(memcg, PGSCAN_DIRECT) +
1603 memcg_events(memcg, PGSCAN_KHUGEPAGED));
1604 seq_buf_printf(s, "pgsteal %lu\n",
1605 memcg_events(memcg, PGSTEAL_KSWAPD) +
1606 memcg_events(memcg, PGSTEAL_DIRECT) +
1607 memcg_events(memcg, PGSTEAL_KHUGEPAGED));
1608
1609 for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) {
1610 if (memcg_vm_event_stat[i] == PGPGIN ||
1611 memcg_vm_event_stat[i] == PGPGOUT)
1612 continue;
1613
1614 seq_buf_printf(s, "%s %lu\n",
1615 vm_event_name(memcg_vm_event_stat[i]),
1616 memcg_events(memcg, memcg_vm_event_stat[i]));
1617 }
1618
1619 /* The above should easily fit into one page */
1620 WARN_ON_ONCE(seq_buf_has_overflowed(s));
1621 }
1622
1623 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s);
1624
memory_stat_format(struct mem_cgroup * memcg,struct seq_buf * s)1625 static void memory_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
1626 {
1627 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1628 memcg_stat_format(memcg, s);
1629 else
1630 memcg1_stat_format(memcg, s);
1631 WARN_ON_ONCE(seq_buf_has_overflowed(s));
1632 }
1633
1634 /**
1635 * mem_cgroup_print_oom_context: Print OOM information relevant to
1636 * memory controller.
1637 * @memcg: The memory cgroup that went over limit
1638 * @p: Task that is going to be killed
1639 *
1640 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1641 * enabled
1642 */
mem_cgroup_print_oom_context(struct mem_cgroup * memcg,struct task_struct * p)1643 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1644 {
1645 rcu_read_lock();
1646
1647 if (memcg) {
1648 pr_cont(",oom_memcg=");
1649 pr_cont_cgroup_path(memcg->css.cgroup);
1650 } else
1651 pr_cont(",global_oom");
1652 if (p) {
1653 pr_cont(",task_memcg=");
1654 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1655 }
1656 rcu_read_unlock();
1657 }
1658
1659 /**
1660 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1661 * memory controller.
1662 * @memcg: The memory cgroup that went over limit
1663 */
mem_cgroup_print_oom_meminfo(struct mem_cgroup * memcg)1664 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1665 {
1666 /* Use static buffer, for the caller is holding oom_lock. */
1667 static char buf[PAGE_SIZE];
1668 struct seq_buf s;
1669
1670 lockdep_assert_held(&oom_lock);
1671
1672 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1673 K((u64)page_counter_read(&memcg->memory)),
1674 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1675 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1676 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1677 K((u64)page_counter_read(&memcg->swap)),
1678 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1679 else {
1680 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1681 K((u64)page_counter_read(&memcg->memsw)),
1682 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1683 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1684 K((u64)page_counter_read(&memcg->kmem)),
1685 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1686 }
1687
1688 pr_info("Memory cgroup stats for ");
1689 pr_cont_cgroup_path(memcg->css.cgroup);
1690 pr_cont(":");
1691 seq_buf_init(&s, buf, sizeof(buf));
1692 memory_stat_format(memcg, &s);
1693 seq_buf_do_printk(&s, KERN_INFO);
1694 }
1695
1696 /*
1697 * Return the memory (and swap, if configured) limit for a memcg.
1698 */
mem_cgroup_get_max(struct mem_cgroup * memcg)1699 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1700 {
1701 unsigned long max = READ_ONCE(memcg->memory.max);
1702
1703 if (do_memsw_account()) {
1704 if (mem_cgroup_swappiness(memcg)) {
1705 /* Calculate swap excess capacity from memsw limit */
1706 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1707
1708 max += min(swap, (unsigned long)total_swap_pages);
1709 }
1710 } else {
1711 if (mem_cgroup_swappiness(memcg))
1712 max += min(READ_ONCE(memcg->swap.max),
1713 (unsigned long)total_swap_pages);
1714 }
1715 return max;
1716 }
1717
mem_cgroup_size(struct mem_cgroup * memcg)1718 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1719 {
1720 return page_counter_read(&memcg->memory);
1721 }
1722
mem_cgroup_out_of_memory(struct mem_cgroup * memcg,gfp_t gfp_mask,int order)1723 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1724 int order)
1725 {
1726 struct oom_control oc = {
1727 .zonelist = NULL,
1728 .nodemask = NULL,
1729 .memcg = memcg,
1730 .gfp_mask = gfp_mask,
1731 .order = order,
1732 };
1733 bool ret = true;
1734
1735 if (mutex_lock_killable(&oom_lock))
1736 return true;
1737
1738 if (mem_cgroup_margin(memcg) >= (1 << order))
1739 goto unlock;
1740
1741 /*
1742 * A few threads which were not waiting at mutex_lock_killable() can
1743 * fail to bail out. Therefore, check again after holding oom_lock.
1744 */
1745 ret = task_is_dying() || out_of_memory(&oc);
1746
1747 unlock:
1748 mutex_unlock(&oom_lock);
1749 return ret;
1750 }
1751
mem_cgroup_soft_reclaim(struct mem_cgroup * root_memcg,pg_data_t * pgdat,gfp_t gfp_mask,unsigned long * total_scanned)1752 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1753 pg_data_t *pgdat,
1754 gfp_t gfp_mask,
1755 unsigned long *total_scanned)
1756 {
1757 struct mem_cgroup *victim = NULL;
1758 int total = 0;
1759 int loop = 0;
1760 unsigned long excess;
1761 unsigned long nr_scanned;
1762 struct mem_cgroup_reclaim_cookie reclaim = {
1763 .pgdat = pgdat,
1764 };
1765
1766 excess = soft_limit_excess(root_memcg);
1767
1768 while (1) {
1769 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1770 if (!victim) {
1771 loop++;
1772 if (loop >= 2) {
1773 /*
1774 * If we have not been able to reclaim
1775 * anything, it might because there are
1776 * no reclaimable pages under this hierarchy
1777 */
1778 if (!total)
1779 break;
1780 /*
1781 * We want to do more targeted reclaim.
1782 * excess >> 2 is not to excessive so as to
1783 * reclaim too much, nor too less that we keep
1784 * coming back to reclaim from this cgroup
1785 */
1786 if (total >= (excess >> 2) ||
1787 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1788 break;
1789 }
1790 continue;
1791 }
1792 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1793 pgdat, &nr_scanned);
1794 *total_scanned += nr_scanned;
1795 if (!soft_limit_excess(root_memcg))
1796 break;
1797 }
1798 mem_cgroup_iter_break(root_memcg, victim);
1799 return total;
1800 }
1801
1802 #ifdef CONFIG_LOCKDEP
1803 static struct lockdep_map memcg_oom_lock_dep_map = {
1804 .name = "memcg_oom_lock",
1805 };
1806 #endif
1807
1808 static DEFINE_SPINLOCK(memcg_oom_lock);
1809
1810 /*
1811 * Check OOM-Killer is already running under our hierarchy.
1812 * If someone is running, return false.
1813 */
mem_cgroup_oom_trylock(struct mem_cgroup * memcg)1814 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1815 {
1816 struct mem_cgroup *iter, *failed = NULL;
1817
1818 spin_lock(&memcg_oom_lock);
1819
1820 for_each_mem_cgroup_tree(iter, memcg) {
1821 if (iter->oom_lock) {
1822 /*
1823 * this subtree of our hierarchy is already locked
1824 * so we cannot give a lock.
1825 */
1826 failed = iter;
1827 mem_cgroup_iter_break(memcg, iter);
1828 break;
1829 } else
1830 iter->oom_lock = true;
1831 }
1832
1833 if (failed) {
1834 /*
1835 * OK, we failed to lock the whole subtree so we have
1836 * to clean up what we set up to the failing subtree
1837 */
1838 for_each_mem_cgroup_tree(iter, memcg) {
1839 if (iter == failed) {
1840 mem_cgroup_iter_break(memcg, iter);
1841 break;
1842 }
1843 iter->oom_lock = false;
1844 }
1845 } else
1846 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1847
1848 spin_unlock(&memcg_oom_lock);
1849
1850 return !failed;
1851 }
1852
mem_cgroup_oom_unlock(struct mem_cgroup * memcg)1853 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1854 {
1855 struct mem_cgroup *iter;
1856
1857 spin_lock(&memcg_oom_lock);
1858 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1859 for_each_mem_cgroup_tree(iter, memcg)
1860 iter->oom_lock = false;
1861 spin_unlock(&memcg_oom_lock);
1862 }
1863
mem_cgroup_mark_under_oom(struct mem_cgroup * memcg)1864 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1865 {
1866 struct mem_cgroup *iter;
1867
1868 spin_lock(&memcg_oom_lock);
1869 for_each_mem_cgroup_tree(iter, memcg)
1870 iter->under_oom++;
1871 spin_unlock(&memcg_oom_lock);
1872 }
1873
mem_cgroup_unmark_under_oom(struct mem_cgroup * memcg)1874 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1875 {
1876 struct mem_cgroup *iter;
1877
1878 /*
1879 * Be careful about under_oom underflows because a child memcg
1880 * could have been added after mem_cgroup_mark_under_oom.
1881 */
1882 spin_lock(&memcg_oom_lock);
1883 for_each_mem_cgroup_tree(iter, memcg)
1884 if (iter->under_oom > 0)
1885 iter->under_oom--;
1886 spin_unlock(&memcg_oom_lock);
1887 }
1888
1889 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1890
1891 struct oom_wait_info {
1892 struct mem_cgroup *memcg;
1893 wait_queue_entry_t wait;
1894 };
1895
memcg_oom_wake_function(wait_queue_entry_t * wait,unsigned mode,int sync,void * arg)1896 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1897 unsigned mode, int sync, void *arg)
1898 {
1899 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1900 struct mem_cgroup *oom_wait_memcg;
1901 struct oom_wait_info *oom_wait_info;
1902
1903 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1904 oom_wait_memcg = oom_wait_info->memcg;
1905
1906 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1907 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1908 return 0;
1909 return autoremove_wake_function(wait, mode, sync, arg);
1910 }
1911
memcg_oom_recover(struct mem_cgroup * memcg)1912 static void memcg_oom_recover(struct mem_cgroup *memcg)
1913 {
1914 /*
1915 * For the following lockless ->under_oom test, the only required
1916 * guarantee is that it must see the state asserted by an OOM when
1917 * this function is called as a result of userland actions
1918 * triggered by the notification of the OOM. This is trivially
1919 * achieved by invoking mem_cgroup_mark_under_oom() before
1920 * triggering notification.
1921 */
1922 if (memcg && memcg->under_oom)
1923 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1924 }
1925
1926 /*
1927 * Returns true if successfully killed one or more processes. Though in some
1928 * corner cases it can return true even without killing any process.
1929 */
mem_cgroup_oom(struct mem_cgroup * memcg,gfp_t mask,int order)1930 static bool mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1931 {
1932 bool locked, ret;
1933
1934 if (order > PAGE_ALLOC_COSTLY_ORDER)
1935 return false;
1936
1937 memcg_memory_event(memcg, MEMCG_OOM);
1938
1939 /*
1940 * We are in the middle of the charge context here, so we
1941 * don't want to block when potentially sitting on a callstack
1942 * that holds all kinds of filesystem and mm locks.
1943 *
1944 * cgroup1 allows disabling the OOM killer and waiting for outside
1945 * handling until the charge can succeed; remember the context and put
1946 * the task to sleep at the end of the page fault when all locks are
1947 * released.
1948 *
1949 * On the other hand, in-kernel OOM killer allows for an async victim
1950 * memory reclaim (oom_reaper) and that means that we are not solely
1951 * relying on the oom victim to make a forward progress and we can
1952 * invoke the oom killer here.
1953 *
1954 * Please note that mem_cgroup_out_of_memory might fail to find a
1955 * victim and then we have to bail out from the charge path.
1956 */
1957 if (READ_ONCE(memcg->oom_kill_disable)) {
1958 if (current->in_user_fault) {
1959 css_get(&memcg->css);
1960 current->memcg_in_oom = memcg;
1961 current->memcg_oom_gfp_mask = mask;
1962 current->memcg_oom_order = order;
1963 }
1964 return false;
1965 }
1966
1967 mem_cgroup_mark_under_oom(memcg);
1968
1969 locked = mem_cgroup_oom_trylock(memcg);
1970
1971 if (locked)
1972 mem_cgroup_oom_notify(memcg);
1973
1974 mem_cgroup_unmark_under_oom(memcg);
1975 ret = mem_cgroup_out_of_memory(memcg, mask, order);
1976
1977 if (locked)
1978 mem_cgroup_oom_unlock(memcg);
1979
1980 return ret;
1981 }
1982
1983 /**
1984 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1985 * @handle: actually kill/wait or just clean up the OOM state
1986 *
1987 * This has to be called at the end of a page fault if the memcg OOM
1988 * handler was enabled.
1989 *
1990 * Memcg supports userspace OOM handling where failed allocations must
1991 * sleep on a waitqueue until the userspace task resolves the
1992 * situation. Sleeping directly in the charge context with all kinds
1993 * of locks held is not a good idea, instead we remember an OOM state
1994 * in the task and mem_cgroup_oom_synchronize() has to be called at
1995 * the end of the page fault to complete the OOM handling.
1996 *
1997 * Returns %true if an ongoing memcg OOM situation was detected and
1998 * completed, %false otherwise.
1999 */
mem_cgroup_oom_synchronize(bool handle)2000 bool mem_cgroup_oom_synchronize(bool handle)
2001 {
2002 struct mem_cgroup *memcg = current->memcg_in_oom;
2003 struct oom_wait_info owait;
2004 bool locked;
2005
2006 /* OOM is global, do not handle */
2007 if (!memcg)
2008 return false;
2009
2010 if (!handle)
2011 goto cleanup;
2012
2013 owait.memcg = memcg;
2014 owait.wait.flags = 0;
2015 owait.wait.func = memcg_oom_wake_function;
2016 owait.wait.private = current;
2017 INIT_LIST_HEAD(&owait.wait.entry);
2018
2019 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2020 mem_cgroup_mark_under_oom(memcg);
2021
2022 locked = mem_cgroup_oom_trylock(memcg);
2023
2024 if (locked)
2025 mem_cgroup_oom_notify(memcg);
2026
2027 schedule();
2028 mem_cgroup_unmark_under_oom(memcg);
2029 finish_wait(&memcg_oom_waitq, &owait.wait);
2030
2031 if (locked)
2032 mem_cgroup_oom_unlock(memcg);
2033 cleanup:
2034 current->memcg_in_oom = NULL;
2035 css_put(&memcg->css);
2036 return true;
2037 }
2038
2039 /**
2040 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2041 * @victim: task to be killed by the OOM killer
2042 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2043 *
2044 * Returns a pointer to a memory cgroup, which has to be cleaned up
2045 * by killing all belonging OOM-killable tasks.
2046 *
2047 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2048 */
mem_cgroup_get_oom_group(struct task_struct * victim,struct mem_cgroup * oom_domain)2049 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2050 struct mem_cgroup *oom_domain)
2051 {
2052 struct mem_cgroup *oom_group = NULL;
2053 struct mem_cgroup *memcg;
2054
2055 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2056 return NULL;
2057
2058 if (!oom_domain)
2059 oom_domain = root_mem_cgroup;
2060
2061 rcu_read_lock();
2062
2063 memcg = mem_cgroup_from_task(victim);
2064 if (mem_cgroup_is_root(memcg))
2065 goto out;
2066
2067 /*
2068 * If the victim task has been asynchronously moved to a different
2069 * memory cgroup, we might end up killing tasks outside oom_domain.
2070 * In this case it's better to ignore memory.group.oom.
2071 */
2072 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2073 goto out;
2074
2075 /*
2076 * Traverse the memory cgroup hierarchy from the victim task's
2077 * cgroup up to the OOMing cgroup (or root) to find the
2078 * highest-level memory cgroup with oom.group set.
2079 */
2080 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2081 if (READ_ONCE(memcg->oom_group))
2082 oom_group = memcg;
2083
2084 if (memcg == oom_domain)
2085 break;
2086 }
2087
2088 if (oom_group)
2089 css_get(&oom_group->css);
2090 out:
2091 rcu_read_unlock();
2092
2093 return oom_group;
2094 }
2095
mem_cgroup_print_oom_group(struct mem_cgroup * memcg)2096 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2097 {
2098 pr_info("Tasks in ");
2099 pr_cont_cgroup_path(memcg->css.cgroup);
2100 pr_cont(" are going to be killed due to memory.oom.group set\n");
2101 }
2102
2103 /**
2104 * folio_memcg_lock - Bind a folio to its memcg.
2105 * @folio: The folio.
2106 *
2107 * This function prevents unlocked LRU folios from being moved to
2108 * another cgroup.
2109 *
2110 * It ensures lifetime of the bound memcg. The caller is responsible
2111 * for the lifetime of the folio.
2112 */
folio_memcg_lock(struct folio * folio)2113 void folio_memcg_lock(struct folio *folio)
2114 {
2115 struct mem_cgroup *memcg;
2116 unsigned long flags;
2117
2118 /*
2119 * The RCU lock is held throughout the transaction. The fast
2120 * path can get away without acquiring the memcg->move_lock
2121 * because page moving starts with an RCU grace period.
2122 */
2123 rcu_read_lock();
2124
2125 if (mem_cgroup_disabled())
2126 return;
2127 again:
2128 memcg = folio_memcg(folio);
2129 if (unlikely(!memcg))
2130 return;
2131
2132 #ifdef CONFIG_PROVE_LOCKING
2133 local_irq_save(flags);
2134 might_lock(&memcg->move_lock);
2135 local_irq_restore(flags);
2136 #endif
2137
2138 if (atomic_read(&memcg->moving_account) <= 0)
2139 return;
2140
2141 spin_lock_irqsave(&memcg->move_lock, flags);
2142 if (memcg != folio_memcg(folio)) {
2143 spin_unlock_irqrestore(&memcg->move_lock, flags);
2144 goto again;
2145 }
2146
2147 /*
2148 * When charge migration first begins, we can have multiple
2149 * critical sections holding the fast-path RCU lock and one
2150 * holding the slowpath move_lock. Track the task who has the
2151 * move_lock for folio_memcg_unlock().
2152 */
2153 memcg->move_lock_task = current;
2154 memcg->move_lock_flags = flags;
2155 }
2156
__folio_memcg_unlock(struct mem_cgroup * memcg)2157 static void __folio_memcg_unlock(struct mem_cgroup *memcg)
2158 {
2159 if (memcg && memcg->move_lock_task == current) {
2160 unsigned long flags = memcg->move_lock_flags;
2161
2162 memcg->move_lock_task = NULL;
2163 memcg->move_lock_flags = 0;
2164
2165 spin_unlock_irqrestore(&memcg->move_lock, flags);
2166 }
2167
2168 rcu_read_unlock();
2169 }
2170
2171 /**
2172 * folio_memcg_unlock - Release the binding between a folio and its memcg.
2173 * @folio: The folio.
2174 *
2175 * This releases the binding created by folio_memcg_lock(). This does
2176 * not change the accounting of this folio to its memcg, but it does
2177 * permit others to change it.
2178 */
folio_memcg_unlock(struct folio * folio)2179 void folio_memcg_unlock(struct folio *folio)
2180 {
2181 __folio_memcg_unlock(folio_memcg(folio));
2182 }
2183
2184 struct memcg_stock_pcp {
2185 local_lock_t stock_lock;
2186 struct mem_cgroup *cached; /* this never be root cgroup */
2187 unsigned int nr_pages;
2188
2189 #ifdef CONFIG_MEMCG_KMEM
2190 struct obj_cgroup *cached_objcg;
2191 struct pglist_data *cached_pgdat;
2192 unsigned int nr_bytes;
2193 int nr_slab_reclaimable_b;
2194 int nr_slab_unreclaimable_b;
2195 #endif
2196
2197 struct work_struct work;
2198 unsigned long flags;
2199 #define FLUSHING_CACHED_CHARGE 0
2200 };
2201 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock) = {
2202 .stock_lock = INIT_LOCAL_LOCK(stock_lock),
2203 };
2204 static DEFINE_MUTEX(percpu_charge_mutex);
2205
2206 #ifdef CONFIG_MEMCG_KMEM
2207 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock);
2208 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2209 struct mem_cgroup *root_memcg);
2210 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages);
2211
2212 #else
drain_obj_stock(struct memcg_stock_pcp * stock)2213 static inline struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
2214 {
2215 return NULL;
2216 }
obj_stock_flush_required(struct memcg_stock_pcp * stock,struct mem_cgroup * root_memcg)2217 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2218 struct mem_cgroup *root_memcg)
2219 {
2220 return false;
2221 }
memcg_account_kmem(struct mem_cgroup * memcg,int nr_pages)2222 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
2223 {
2224 }
2225 #endif
2226
2227 /**
2228 * consume_stock: Try to consume stocked charge on this cpu.
2229 * @memcg: memcg to consume from.
2230 * @nr_pages: how many pages to charge.
2231 *
2232 * The charges will only happen if @memcg matches the current cpu's memcg
2233 * stock, and at least @nr_pages are available in that stock. Failure to
2234 * service an allocation will refill the stock.
2235 *
2236 * returns true if successful, false otherwise.
2237 */
consume_stock(struct mem_cgroup * memcg,unsigned int nr_pages)2238 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2239 {
2240 struct memcg_stock_pcp *stock;
2241 unsigned long flags;
2242 bool ret = false;
2243
2244 if (nr_pages > MEMCG_CHARGE_BATCH)
2245 return ret;
2246
2247 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2248
2249 stock = this_cpu_ptr(&memcg_stock);
2250 if (memcg == READ_ONCE(stock->cached) && stock->nr_pages >= nr_pages) {
2251 stock->nr_pages -= nr_pages;
2252 ret = true;
2253 }
2254
2255 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2256
2257 return ret;
2258 }
2259
2260 /*
2261 * Returns stocks cached in percpu and reset cached information.
2262 */
drain_stock(struct memcg_stock_pcp * stock)2263 static void drain_stock(struct memcg_stock_pcp *stock)
2264 {
2265 struct mem_cgroup *old = READ_ONCE(stock->cached);
2266
2267 if (!old)
2268 return;
2269
2270 if (stock->nr_pages) {
2271 page_counter_uncharge(&old->memory, stock->nr_pages);
2272 if (do_memsw_account())
2273 page_counter_uncharge(&old->memsw, stock->nr_pages);
2274 stock->nr_pages = 0;
2275 }
2276
2277 css_put(&old->css);
2278 WRITE_ONCE(stock->cached, NULL);
2279 }
2280
drain_local_stock(struct work_struct * dummy)2281 static void drain_local_stock(struct work_struct *dummy)
2282 {
2283 struct memcg_stock_pcp *stock;
2284 struct obj_cgroup *old = NULL;
2285 unsigned long flags;
2286
2287 /*
2288 * The only protection from cpu hotplug (memcg_hotplug_cpu_dead) vs.
2289 * drain_stock races is that we always operate on local CPU stock
2290 * here with IRQ disabled
2291 */
2292 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2293
2294 stock = this_cpu_ptr(&memcg_stock);
2295 old = drain_obj_stock(stock);
2296 drain_stock(stock);
2297 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2298
2299 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2300 if (old)
2301 obj_cgroup_put(old);
2302 }
2303
2304 /*
2305 * Cache charges(val) to local per_cpu area.
2306 * This will be consumed by consume_stock() function, later.
2307 */
__refill_stock(struct mem_cgroup * memcg,unsigned int nr_pages)2308 static void __refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2309 {
2310 struct memcg_stock_pcp *stock;
2311
2312 stock = this_cpu_ptr(&memcg_stock);
2313 if (READ_ONCE(stock->cached) != memcg) { /* reset if necessary */
2314 drain_stock(stock);
2315 css_get(&memcg->css);
2316 WRITE_ONCE(stock->cached, memcg);
2317 }
2318 stock->nr_pages += nr_pages;
2319
2320 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2321 drain_stock(stock);
2322 }
2323
refill_stock(struct mem_cgroup * memcg,unsigned int nr_pages)2324 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2325 {
2326 unsigned long flags;
2327
2328 local_lock_irqsave(&memcg_stock.stock_lock, flags);
2329 __refill_stock(memcg, nr_pages);
2330 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
2331 }
2332
2333 /*
2334 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2335 * of the hierarchy under it.
2336 */
drain_all_stock(struct mem_cgroup * root_memcg)2337 static void drain_all_stock(struct mem_cgroup *root_memcg)
2338 {
2339 int cpu, curcpu;
2340
2341 /* If someone's already draining, avoid adding running more workers. */
2342 if (!mutex_trylock(&percpu_charge_mutex))
2343 return;
2344 /*
2345 * Notify other cpus that system-wide "drain" is running
2346 * We do not care about races with the cpu hotplug because cpu down
2347 * as well as workers from this path always operate on the local
2348 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2349 */
2350 migrate_disable();
2351 curcpu = smp_processor_id();
2352 for_each_online_cpu(cpu) {
2353 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2354 struct mem_cgroup *memcg;
2355 bool flush = false;
2356
2357 rcu_read_lock();
2358 memcg = READ_ONCE(stock->cached);
2359 if (memcg && stock->nr_pages &&
2360 mem_cgroup_is_descendant(memcg, root_memcg))
2361 flush = true;
2362 else if (obj_stock_flush_required(stock, root_memcg))
2363 flush = true;
2364 rcu_read_unlock();
2365
2366 if (flush &&
2367 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2368 if (cpu == curcpu)
2369 drain_local_stock(&stock->work);
2370 else if (!cpu_is_isolated(cpu))
2371 schedule_work_on(cpu, &stock->work);
2372 }
2373 }
2374 migrate_enable();
2375 mutex_unlock(&percpu_charge_mutex);
2376 }
2377
memcg_hotplug_cpu_dead(unsigned int cpu)2378 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2379 {
2380 struct memcg_stock_pcp *stock;
2381
2382 stock = &per_cpu(memcg_stock, cpu);
2383 drain_stock(stock);
2384
2385 return 0;
2386 }
2387
reclaim_high(struct mem_cgroup * memcg,unsigned int nr_pages,gfp_t gfp_mask)2388 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2389 unsigned int nr_pages,
2390 gfp_t gfp_mask)
2391 {
2392 unsigned long nr_reclaimed = 0;
2393
2394 do {
2395 unsigned long pflags;
2396
2397 if (page_counter_read(&memcg->memory) <=
2398 READ_ONCE(memcg->memory.high))
2399 continue;
2400
2401 memcg_memory_event(memcg, MEMCG_HIGH);
2402
2403 psi_memstall_enter(&pflags);
2404 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2405 gfp_mask,
2406 MEMCG_RECLAIM_MAY_SWAP);
2407 psi_memstall_leave(&pflags);
2408 } while ((memcg = parent_mem_cgroup(memcg)) &&
2409 !mem_cgroup_is_root(memcg));
2410
2411 return nr_reclaimed;
2412 }
2413
high_work_func(struct work_struct * work)2414 static void high_work_func(struct work_struct *work)
2415 {
2416 struct mem_cgroup *memcg;
2417
2418 memcg = container_of(work, struct mem_cgroup, high_work);
2419 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2420 }
2421
2422 /*
2423 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2424 * enough to still cause a significant slowdown in most cases, while still
2425 * allowing diagnostics and tracing to proceed without becoming stuck.
2426 */
2427 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2428
2429 /*
2430 * When calculating the delay, we use these either side of the exponentiation to
2431 * maintain precision and scale to a reasonable number of jiffies (see the table
2432 * below.
2433 *
2434 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2435 * overage ratio to a delay.
2436 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2437 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2438 * to produce a reasonable delay curve.
2439 *
2440 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2441 * reasonable delay curve compared to precision-adjusted overage, not
2442 * penalising heavily at first, but still making sure that growth beyond the
2443 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2444 * example, with a high of 100 megabytes:
2445 *
2446 * +-------+------------------------+
2447 * | usage | time to allocate in ms |
2448 * +-------+------------------------+
2449 * | 100M | 0 |
2450 * | 101M | 6 |
2451 * | 102M | 25 |
2452 * | 103M | 57 |
2453 * | 104M | 102 |
2454 * | 105M | 159 |
2455 * | 106M | 230 |
2456 * | 107M | 313 |
2457 * | 108M | 409 |
2458 * | 109M | 518 |
2459 * | 110M | 639 |
2460 * | 111M | 774 |
2461 * | 112M | 921 |
2462 * | 113M | 1081 |
2463 * | 114M | 1254 |
2464 * | 115M | 1439 |
2465 * | 116M | 1638 |
2466 * | 117M | 1849 |
2467 * | 118M | 2000 |
2468 * | 119M | 2000 |
2469 * | 120M | 2000 |
2470 * +-------+------------------------+
2471 */
2472 #define MEMCG_DELAY_PRECISION_SHIFT 20
2473 #define MEMCG_DELAY_SCALING_SHIFT 14
2474
calculate_overage(unsigned long usage,unsigned long high)2475 static u64 calculate_overage(unsigned long usage, unsigned long high)
2476 {
2477 u64 overage;
2478
2479 if (usage <= high)
2480 return 0;
2481
2482 /*
2483 * Prevent division by 0 in overage calculation by acting as if
2484 * it was a threshold of 1 page
2485 */
2486 high = max(high, 1UL);
2487
2488 overage = usage - high;
2489 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2490 return div64_u64(overage, high);
2491 }
2492
mem_find_max_overage(struct mem_cgroup * memcg)2493 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2494 {
2495 u64 overage, max_overage = 0;
2496
2497 do {
2498 overage = calculate_overage(page_counter_read(&memcg->memory),
2499 READ_ONCE(memcg->memory.high));
2500 max_overage = max(overage, max_overage);
2501 } while ((memcg = parent_mem_cgroup(memcg)) &&
2502 !mem_cgroup_is_root(memcg));
2503
2504 return max_overage;
2505 }
2506
swap_find_max_overage(struct mem_cgroup * memcg)2507 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2508 {
2509 u64 overage, max_overage = 0;
2510
2511 do {
2512 overage = calculate_overage(page_counter_read(&memcg->swap),
2513 READ_ONCE(memcg->swap.high));
2514 if (overage)
2515 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2516 max_overage = max(overage, max_overage);
2517 } while ((memcg = parent_mem_cgroup(memcg)) &&
2518 !mem_cgroup_is_root(memcg));
2519
2520 return max_overage;
2521 }
2522
2523 /*
2524 * Get the number of jiffies that we should penalise a mischievous cgroup which
2525 * is exceeding its memory.high by checking both it and its ancestors.
2526 */
calculate_high_delay(struct mem_cgroup * memcg,unsigned int nr_pages,u64 max_overage)2527 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2528 unsigned int nr_pages,
2529 u64 max_overage)
2530 {
2531 unsigned long penalty_jiffies;
2532
2533 if (!max_overage)
2534 return 0;
2535
2536 /*
2537 * We use overage compared to memory.high to calculate the number of
2538 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2539 * fairly lenient on small overages, and increasingly harsh when the
2540 * memcg in question makes it clear that it has no intention of stopping
2541 * its crazy behaviour, so we exponentially increase the delay based on
2542 * overage amount.
2543 */
2544 penalty_jiffies = max_overage * max_overage * HZ;
2545 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2546 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2547
2548 /*
2549 * Factor in the task's own contribution to the overage, such that four
2550 * N-sized allocations are throttled approximately the same as one
2551 * 4N-sized allocation.
2552 *
2553 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2554 * larger the current charge patch is than that.
2555 */
2556 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2557 }
2558
2559 /*
2560 * Scheduled by try_charge() to be executed from the userland return path
2561 * and reclaims memory over the high limit.
2562 */
mem_cgroup_handle_over_high(gfp_t gfp_mask)2563 void mem_cgroup_handle_over_high(gfp_t gfp_mask)
2564 {
2565 unsigned long penalty_jiffies;
2566 unsigned long pflags;
2567 unsigned long nr_reclaimed;
2568 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2569 int nr_retries = MAX_RECLAIM_RETRIES;
2570 struct mem_cgroup *memcg;
2571 bool in_retry = false;
2572
2573 if (likely(!nr_pages))
2574 return;
2575
2576 memcg = get_mem_cgroup_from_mm(current->mm);
2577 current->memcg_nr_pages_over_high = 0;
2578
2579 retry_reclaim:
2580 /*
2581 * The allocating task should reclaim at least the batch size, but for
2582 * subsequent retries we only want to do what's necessary to prevent oom
2583 * or breaching resource isolation.
2584 *
2585 * This is distinct from memory.max or page allocator behaviour because
2586 * memory.high is currently batched, whereas memory.max and the page
2587 * allocator run every time an allocation is made.
2588 */
2589 nr_reclaimed = reclaim_high(memcg,
2590 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2591 gfp_mask);
2592
2593 /*
2594 * memory.high is breached and reclaim is unable to keep up. Throttle
2595 * allocators proactively to slow down excessive growth.
2596 */
2597 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2598 mem_find_max_overage(memcg));
2599
2600 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2601 swap_find_max_overage(memcg));
2602
2603 /*
2604 * Clamp the max delay per usermode return so as to still keep the
2605 * application moving forwards and also permit diagnostics, albeit
2606 * extremely slowly.
2607 */
2608 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2609
2610 /*
2611 * Don't sleep if the amount of jiffies this memcg owes us is so low
2612 * that it's not even worth doing, in an attempt to be nice to those who
2613 * go only a small amount over their memory.high value and maybe haven't
2614 * been aggressively reclaimed enough yet.
2615 */
2616 if (penalty_jiffies <= HZ / 100)
2617 goto out;
2618
2619 /*
2620 * If reclaim is making forward progress but we're still over
2621 * memory.high, we want to encourage that rather than doing allocator
2622 * throttling.
2623 */
2624 if (nr_reclaimed || nr_retries--) {
2625 in_retry = true;
2626 goto retry_reclaim;
2627 }
2628
2629 /*
2630 * If we exit early, we're guaranteed to die (since
2631 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2632 * need to account for any ill-begotten jiffies to pay them off later.
2633 */
2634 psi_memstall_enter(&pflags);
2635 schedule_timeout_killable(penalty_jiffies);
2636 psi_memstall_leave(&pflags);
2637
2638 out:
2639 css_put(&memcg->css);
2640 }
2641
try_charge_memcg(struct mem_cgroup * memcg,gfp_t gfp_mask,unsigned int nr_pages)2642 static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
2643 unsigned int nr_pages)
2644 {
2645 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2646 int nr_retries = MAX_RECLAIM_RETRIES;
2647 struct mem_cgroup *mem_over_limit;
2648 struct page_counter *counter;
2649 unsigned long nr_reclaimed;
2650 bool passed_oom = false;
2651 unsigned int reclaim_options = MEMCG_RECLAIM_MAY_SWAP;
2652 bool drained = false;
2653 bool raised_max_event = false;
2654 unsigned long pflags;
2655
2656 retry:
2657 if (consume_stock(memcg, nr_pages))
2658 return 0;
2659
2660 if (!do_memsw_account() ||
2661 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2662 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2663 goto done_restock;
2664 if (do_memsw_account())
2665 page_counter_uncharge(&memcg->memsw, batch);
2666 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2667 } else {
2668 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2669 reclaim_options &= ~MEMCG_RECLAIM_MAY_SWAP;
2670 }
2671
2672 if (batch > nr_pages) {
2673 batch = nr_pages;
2674 goto retry;
2675 }
2676
2677 /*
2678 * Prevent unbounded recursion when reclaim operations need to
2679 * allocate memory. This might exceed the limits temporarily,
2680 * but we prefer facilitating memory reclaim and getting back
2681 * under the limit over triggering OOM kills in these cases.
2682 */
2683 if (unlikely(current->flags & PF_MEMALLOC))
2684 goto force;
2685
2686 if (unlikely(task_in_memcg_oom(current)))
2687 goto nomem;
2688
2689 if (!gfpflags_allow_blocking(gfp_mask))
2690 goto nomem;
2691
2692 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2693 raised_max_event = true;
2694
2695 psi_memstall_enter(&pflags);
2696 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2697 gfp_mask, reclaim_options);
2698 psi_memstall_leave(&pflags);
2699
2700 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2701 goto retry;
2702
2703 if (!drained) {
2704 drain_all_stock(mem_over_limit);
2705 drained = true;
2706 goto retry;
2707 }
2708
2709 if (gfp_mask & __GFP_NORETRY)
2710 goto nomem;
2711 /*
2712 * Even though the limit is exceeded at this point, reclaim
2713 * may have been able to free some pages. Retry the charge
2714 * before killing the task.
2715 *
2716 * Only for regular pages, though: huge pages are rather
2717 * unlikely to succeed so close to the limit, and we fall back
2718 * to regular pages anyway in case of failure.
2719 */
2720 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2721 goto retry;
2722 /*
2723 * At task move, charge accounts can be doubly counted. So, it's
2724 * better to wait until the end of task_move if something is going on.
2725 */
2726 if (mem_cgroup_wait_acct_move(mem_over_limit))
2727 goto retry;
2728
2729 if (nr_retries--)
2730 goto retry;
2731
2732 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2733 goto nomem;
2734
2735 /* Avoid endless loop for tasks bypassed by the oom killer */
2736 if (passed_oom && task_is_dying())
2737 goto nomem;
2738
2739 /*
2740 * keep retrying as long as the memcg oom killer is able to make
2741 * a forward progress or bypass the charge if the oom killer
2742 * couldn't make any progress.
2743 */
2744 if (mem_cgroup_oom(mem_over_limit, gfp_mask,
2745 get_order(nr_pages * PAGE_SIZE))) {
2746 passed_oom = true;
2747 nr_retries = MAX_RECLAIM_RETRIES;
2748 goto retry;
2749 }
2750 nomem:
2751 /*
2752 * Memcg doesn't have a dedicated reserve for atomic
2753 * allocations. But like the global atomic pool, we need to
2754 * put the burden of reclaim on regular allocation requests
2755 * and let these go through as privileged allocations.
2756 */
2757 if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
2758 return -ENOMEM;
2759 force:
2760 /*
2761 * If the allocation has to be enforced, don't forget to raise
2762 * a MEMCG_MAX event.
2763 */
2764 if (!raised_max_event)
2765 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2766
2767 /*
2768 * The allocation either can't fail or will lead to more memory
2769 * being freed very soon. Allow memory usage go over the limit
2770 * temporarily by force charging it.
2771 */
2772 page_counter_charge(&memcg->memory, nr_pages);
2773 if (do_memsw_account())
2774 page_counter_charge(&memcg->memsw, nr_pages);
2775
2776 return 0;
2777
2778 done_restock:
2779 if (batch > nr_pages)
2780 refill_stock(memcg, batch - nr_pages);
2781
2782 /*
2783 * If the hierarchy is above the normal consumption range, schedule
2784 * reclaim on returning to userland. We can perform reclaim here
2785 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2786 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2787 * not recorded as it most likely matches current's and won't
2788 * change in the meantime. As high limit is checked again before
2789 * reclaim, the cost of mismatch is negligible.
2790 */
2791 do {
2792 bool mem_high, swap_high;
2793
2794 mem_high = page_counter_read(&memcg->memory) >
2795 READ_ONCE(memcg->memory.high);
2796 swap_high = page_counter_read(&memcg->swap) >
2797 READ_ONCE(memcg->swap.high);
2798
2799 /* Don't bother a random interrupted task */
2800 if (!in_task()) {
2801 if (mem_high) {
2802 schedule_work(&memcg->high_work);
2803 break;
2804 }
2805 continue;
2806 }
2807
2808 if (mem_high || swap_high) {
2809 /*
2810 * The allocating tasks in this cgroup will need to do
2811 * reclaim or be throttled to prevent further growth
2812 * of the memory or swap footprints.
2813 *
2814 * Target some best-effort fairness between the tasks,
2815 * and distribute reclaim work and delay penalties
2816 * based on how much each task is actually allocating.
2817 */
2818 current->memcg_nr_pages_over_high += batch;
2819 set_notify_resume(current);
2820 break;
2821 }
2822 } while ((memcg = parent_mem_cgroup(memcg)));
2823
2824 if (current->memcg_nr_pages_over_high > MEMCG_CHARGE_BATCH &&
2825 !(current->flags & PF_MEMALLOC) &&
2826 gfpflags_allow_blocking(gfp_mask)) {
2827 mem_cgroup_handle_over_high(gfp_mask);
2828 }
2829 return 0;
2830 }
2831
try_charge(struct mem_cgroup * memcg,gfp_t gfp_mask,unsigned int nr_pages)2832 static inline int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2833 unsigned int nr_pages)
2834 {
2835 if (mem_cgroup_is_root(memcg))
2836 return 0;
2837
2838 return try_charge_memcg(memcg, gfp_mask, nr_pages);
2839 }
2840
cancel_charge(struct mem_cgroup * memcg,unsigned int nr_pages)2841 static inline void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2842 {
2843 if (mem_cgroup_is_root(memcg))
2844 return;
2845
2846 page_counter_uncharge(&memcg->memory, nr_pages);
2847 if (do_memsw_account())
2848 page_counter_uncharge(&memcg->memsw, nr_pages);
2849 }
2850
commit_charge(struct folio * folio,struct mem_cgroup * memcg)2851 static void commit_charge(struct folio *folio, struct mem_cgroup *memcg)
2852 {
2853 VM_BUG_ON_FOLIO(folio_memcg(folio), folio);
2854 /*
2855 * Any of the following ensures page's memcg stability:
2856 *
2857 * - the page lock
2858 * - LRU isolation
2859 * - folio_memcg_lock()
2860 * - exclusive reference
2861 * - mem_cgroup_trylock_pages()
2862 */
2863 folio->memcg_data = (unsigned long)memcg;
2864 }
2865
2866 #ifdef CONFIG_MEMCG_KMEM
2867 /*
2868 * The allocated objcg pointers array is not accounted directly.
2869 * Moreover, it should not come from DMA buffer and is not readily
2870 * reclaimable. So those GFP bits should be masked off.
2871 */
2872 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
2873 __GFP_ACCOUNT | __GFP_NOFAIL)
2874
2875 /*
2876 * mod_objcg_mlstate() may be called with irq enabled, so
2877 * mod_memcg_lruvec_state() should be used.
2878 */
mod_objcg_mlstate(struct obj_cgroup * objcg,struct pglist_data * pgdat,enum node_stat_item idx,int nr)2879 static inline void mod_objcg_mlstate(struct obj_cgroup *objcg,
2880 struct pglist_data *pgdat,
2881 enum node_stat_item idx, int nr)
2882 {
2883 struct mem_cgroup *memcg;
2884 struct lruvec *lruvec;
2885
2886 rcu_read_lock();
2887 memcg = obj_cgroup_memcg(objcg);
2888 lruvec = mem_cgroup_lruvec(memcg, pgdat);
2889 mod_memcg_lruvec_state(lruvec, idx, nr);
2890 rcu_read_unlock();
2891 }
2892
memcg_alloc_slab_cgroups(struct slab * slab,struct kmem_cache * s,gfp_t gfp,bool new_slab)2893 int memcg_alloc_slab_cgroups(struct slab *slab, struct kmem_cache *s,
2894 gfp_t gfp, bool new_slab)
2895 {
2896 unsigned int objects = objs_per_slab(s, slab);
2897 unsigned long memcg_data;
2898 void *vec;
2899
2900 gfp &= ~OBJCGS_CLEAR_MASK;
2901 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2902 slab_nid(slab));
2903 if (!vec)
2904 return -ENOMEM;
2905
2906 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2907 if (new_slab) {
2908 /*
2909 * If the slab is brand new and nobody can yet access its
2910 * memcg_data, no synchronization is required and memcg_data can
2911 * be simply assigned.
2912 */
2913 slab->memcg_data = memcg_data;
2914 } else if (cmpxchg(&slab->memcg_data, 0, memcg_data)) {
2915 /*
2916 * If the slab is already in use, somebody can allocate and
2917 * assign obj_cgroups in parallel. In this case the existing
2918 * objcg vector should be reused.
2919 */
2920 kfree(vec);
2921 return 0;
2922 }
2923
2924 kmemleak_not_leak(vec);
2925 return 0;
2926 }
2927
2928 static __always_inline
mem_cgroup_from_obj_folio(struct folio * folio,void * p)2929 struct mem_cgroup *mem_cgroup_from_obj_folio(struct folio *folio, void *p)
2930 {
2931 /*
2932 * Slab objects are accounted individually, not per-page.
2933 * Memcg membership data for each individual object is saved in
2934 * slab->memcg_data.
2935 */
2936 if (folio_test_slab(folio)) {
2937 struct obj_cgroup **objcgs;
2938 struct slab *slab;
2939 unsigned int off;
2940
2941 slab = folio_slab(folio);
2942 objcgs = slab_objcgs(slab);
2943 if (!objcgs)
2944 return NULL;
2945
2946 off = obj_to_index(slab->slab_cache, slab, p);
2947 if (objcgs[off])
2948 return obj_cgroup_memcg(objcgs[off]);
2949
2950 return NULL;
2951 }
2952
2953 /*
2954 * folio_memcg_check() is used here, because in theory we can encounter
2955 * a folio where the slab flag has been cleared already, but
2956 * slab->memcg_data has not been freed yet
2957 * folio_memcg_check() will guarantee that a proper memory
2958 * cgroup pointer or NULL will be returned.
2959 */
2960 return folio_memcg_check(folio);
2961 }
2962
2963 /*
2964 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2965 *
2966 * A passed kernel object can be a slab object, vmalloc object or a generic
2967 * kernel page, so different mechanisms for getting the memory cgroup pointer
2968 * should be used.
2969 *
2970 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2971 * can not know for sure how the kernel object is implemented.
2972 * mem_cgroup_from_obj() can be safely used in such cases.
2973 *
2974 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2975 * cgroup_mutex, etc.
2976 */
mem_cgroup_from_obj(void * p)2977 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2978 {
2979 struct folio *folio;
2980
2981 if (mem_cgroup_disabled())
2982 return NULL;
2983
2984 if (unlikely(is_vmalloc_addr(p)))
2985 folio = page_folio(vmalloc_to_page(p));
2986 else
2987 folio = virt_to_folio(p);
2988
2989 return mem_cgroup_from_obj_folio(folio, p);
2990 }
2991
2992 /*
2993 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2994 * Similar to mem_cgroup_from_obj(), but faster and not suitable for objects,
2995 * allocated using vmalloc().
2996 *
2997 * A passed kernel object must be a slab object or a generic kernel page.
2998 *
2999 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
3000 * cgroup_mutex, etc.
3001 */
mem_cgroup_from_slab_obj(void * p)3002 struct mem_cgroup *mem_cgroup_from_slab_obj(void *p)
3003 {
3004 if (mem_cgroup_disabled())
3005 return NULL;
3006
3007 return mem_cgroup_from_obj_folio(virt_to_folio(p), p);
3008 }
3009
__get_obj_cgroup_from_memcg(struct mem_cgroup * memcg)3010 static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg)
3011 {
3012 struct obj_cgroup *objcg = NULL;
3013
3014 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
3015 objcg = rcu_dereference(memcg->objcg);
3016 if (objcg && obj_cgroup_tryget(objcg))
3017 break;
3018 objcg = NULL;
3019 }
3020 return objcg;
3021 }
3022
get_obj_cgroup_from_current(void)3023 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
3024 {
3025 struct obj_cgroup *objcg = NULL;
3026 struct mem_cgroup *memcg;
3027
3028 if (memcg_kmem_bypass())
3029 return NULL;
3030
3031 rcu_read_lock();
3032 if (unlikely(active_memcg()))
3033 memcg = active_memcg();
3034 else
3035 memcg = mem_cgroup_from_task(current);
3036 objcg = __get_obj_cgroup_from_memcg(memcg);
3037 rcu_read_unlock();
3038 return objcg;
3039 }
3040
get_obj_cgroup_from_folio(struct folio * folio)3041 struct obj_cgroup *get_obj_cgroup_from_folio(struct folio *folio)
3042 {
3043 struct obj_cgroup *objcg;
3044
3045 if (!memcg_kmem_online())
3046 return NULL;
3047
3048 if (folio_memcg_kmem(folio)) {
3049 objcg = __folio_objcg(folio);
3050 obj_cgroup_get(objcg);
3051 } else {
3052 struct mem_cgroup *memcg;
3053
3054 rcu_read_lock();
3055 memcg = __folio_memcg(folio);
3056 if (memcg)
3057 objcg = __get_obj_cgroup_from_memcg(memcg);
3058 else
3059 objcg = NULL;
3060 rcu_read_unlock();
3061 }
3062 return objcg;
3063 }
3064
memcg_account_kmem(struct mem_cgroup * memcg,int nr_pages)3065 static void memcg_account_kmem(struct mem_cgroup *memcg, int nr_pages)
3066 {
3067 mod_memcg_state(memcg, MEMCG_KMEM, nr_pages);
3068 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
3069 if (nr_pages > 0)
3070 page_counter_charge(&memcg->kmem, nr_pages);
3071 else
3072 page_counter_uncharge(&memcg->kmem, -nr_pages);
3073 }
3074 }
3075
3076
3077 /*
3078 * obj_cgroup_uncharge_pages: uncharge a number of kernel pages from a objcg
3079 * @objcg: object cgroup to uncharge
3080 * @nr_pages: number of pages to uncharge
3081 */
obj_cgroup_uncharge_pages(struct obj_cgroup * objcg,unsigned int nr_pages)3082 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3083 unsigned int nr_pages)
3084 {
3085 struct mem_cgroup *memcg;
3086
3087 memcg = get_mem_cgroup_from_objcg(objcg);
3088
3089 memcg_account_kmem(memcg, -nr_pages);
3090 refill_stock(memcg, nr_pages);
3091
3092 css_put(&memcg->css);
3093 }
3094
3095 /*
3096 * obj_cgroup_charge_pages: charge a number of kernel pages to a objcg
3097 * @objcg: object cgroup to charge
3098 * @gfp: reclaim mode
3099 * @nr_pages: number of pages to charge
3100 *
3101 * Returns 0 on success, an error code on failure.
3102 */
obj_cgroup_charge_pages(struct obj_cgroup * objcg,gfp_t gfp,unsigned int nr_pages)3103 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3104 unsigned int nr_pages)
3105 {
3106 struct mem_cgroup *memcg;
3107 int ret;
3108
3109 memcg = get_mem_cgroup_from_objcg(objcg);
3110
3111 ret = try_charge_memcg(memcg, gfp, nr_pages);
3112 if (ret)
3113 goto out;
3114
3115 memcg_account_kmem(memcg, nr_pages);
3116 out:
3117 css_put(&memcg->css);
3118
3119 return ret;
3120 }
3121
3122 /**
3123 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3124 * @page: page to charge
3125 * @gfp: reclaim mode
3126 * @order: allocation order
3127 *
3128 * Returns 0 on success, an error code on failure.
3129 */
__memcg_kmem_charge_page(struct page * page,gfp_t gfp,int order)3130 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3131 {
3132 struct obj_cgroup *objcg;
3133 int ret = 0;
3134
3135 objcg = get_obj_cgroup_from_current();
3136 if (objcg) {
3137 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3138 if (!ret) {
3139 page->memcg_data = (unsigned long)objcg |
3140 MEMCG_DATA_KMEM;
3141 return 0;
3142 }
3143 obj_cgroup_put(objcg);
3144 }
3145 return ret;
3146 }
3147
3148 /**
3149 * __memcg_kmem_uncharge_page: uncharge a kmem page
3150 * @page: page to uncharge
3151 * @order: allocation order
3152 */
__memcg_kmem_uncharge_page(struct page * page,int order)3153 void __memcg_kmem_uncharge_page(struct page *page, int order)
3154 {
3155 struct folio *folio = page_folio(page);
3156 struct obj_cgroup *objcg;
3157 unsigned int nr_pages = 1 << order;
3158
3159 if (!folio_memcg_kmem(folio))
3160 return;
3161
3162 objcg = __folio_objcg(folio);
3163 obj_cgroup_uncharge_pages(objcg, nr_pages);
3164 folio->memcg_data = 0;
3165 obj_cgroup_put(objcg);
3166 }
3167
mod_objcg_state(struct obj_cgroup * objcg,struct pglist_data * pgdat,enum node_stat_item idx,int nr)3168 void mod_objcg_state(struct obj_cgroup *objcg, struct pglist_data *pgdat,
3169 enum node_stat_item idx, int nr)
3170 {
3171 struct memcg_stock_pcp *stock;
3172 struct obj_cgroup *old = NULL;
3173 unsigned long flags;
3174 int *bytes;
3175
3176 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3177 stock = this_cpu_ptr(&memcg_stock);
3178
3179 /*
3180 * Save vmstat data in stock and skip vmstat array update unless
3181 * accumulating over a page of vmstat data or when pgdat or idx
3182 * changes.
3183 */
3184 if (READ_ONCE(stock->cached_objcg) != objcg) {
3185 old = drain_obj_stock(stock);
3186 obj_cgroup_get(objcg);
3187 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3188 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3189 WRITE_ONCE(stock->cached_objcg, objcg);
3190 stock->cached_pgdat = pgdat;
3191 } else if (stock->cached_pgdat != pgdat) {
3192 /* Flush the existing cached vmstat data */
3193 struct pglist_data *oldpg = stock->cached_pgdat;
3194
3195 if (stock->nr_slab_reclaimable_b) {
3196 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_RECLAIMABLE_B,
3197 stock->nr_slab_reclaimable_b);
3198 stock->nr_slab_reclaimable_b = 0;
3199 }
3200 if (stock->nr_slab_unreclaimable_b) {
3201 mod_objcg_mlstate(objcg, oldpg, NR_SLAB_UNRECLAIMABLE_B,
3202 stock->nr_slab_unreclaimable_b);
3203 stock->nr_slab_unreclaimable_b = 0;
3204 }
3205 stock->cached_pgdat = pgdat;
3206 }
3207
3208 bytes = (idx == NR_SLAB_RECLAIMABLE_B) ? &stock->nr_slab_reclaimable_b
3209 : &stock->nr_slab_unreclaimable_b;
3210 /*
3211 * Even for large object >= PAGE_SIZE, the vmstat data will still be
3212 * cached locally at least once before pushing it out.
3213 */
3214 if (!*bytes) {
3215 *bytes = nr;
3216 nr = 0;
3217 } else {
3218 *bytes += nr;
3219 if (abs(*bytes) > PAGE_SIZE) {
3220 nr = *bytes;
3221 *bytes = 0;
3222 } else {
3223 nr = 0;
3224 }
3225 }
3226 if (nr)
3227 mod_objcg_mlstate(objcg, pgdat, idx, nr);
3228
3229 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3230 if (old)
3231 obj_cgroup_put(old);
3232 }
3233
consume_obj_stock(struct obj_cgroup * objcg,unsigned int nr_bytes)3234 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3235 {
3236 struct memcg_stock_pcp *stock;
3237 unsigned long flags;
3238 bool ret = false;
3239
3240 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3241
3242 stock = this_cpu_ptr(&memcg_stock);
3243 if (objcg == READ_ONCE(stock->cached_objcg) && stock->nr_bytes >= nr_bytes) {
3244 stock->nr_bytes -= nr_bytes;
3245 ret = true;
3246 }
3247
3248 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3249
3250 return ret;
3251 }
3252
drain_obj_stock(struct memcg_stock_pcp * stock)3253 static struct obj_cgroup *drain_obj_stock(struct memcg_stock_pcp *stock)
3254 {
3255 struct obj_cgroup *old = READ_ONCE(stock->cached_objcg);
3256
3257 if (!old)
3258 return NULL;
3259
3260 if (stock->nr_bytes) {
3261 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3262 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3263
3264 if (nr_pages) {
3265 struct mem_cgroup *memcg;
3266
3267 memcg = get_mem_cgroup_from_objcg(old);
3268
3269 memcg_account_kmem(memcg, -nr_pages);
3270 __refill_stock(memcg, nr_pages);
3271
3272 css_put(&memcg->css);
3273 }
3274
3275 /*
3276 * The leftover is flushed to the centralized per-memcg value.
3277 * On the next attempt to refill obj stock it will be moved
3278 * to a per-cpu stock (probably, on an other CPU), see
3279 * refill_obj_stock().
3280 *
3281 * How often it's flushed is a trade-off between the memory
3282 * limit enforcement accuracy and potential CPU contention,
3283 * so it might be changed in the future.
3284 */
3285 atomic_add(nr_bytes, &old->nr_charged_bytes);
3286 stock->nr_bytes = 0;
3287 }
3288
3289 /*
3290 * Flush the vmstat data in current stock
3291 */
3292 if (stock->nr_slab_reclaimable_b || stock->nr_slab_unreclaimable_b) {
3293 if (stock->nr_slab_reclaimable_b) {
3294 mod_objcg_mlstate(old, stock->cached_pgdat,
3295 NR_SLAB_RECLAIMABLE_B,
3296 stock->nr_slab_reclaimable_b);
3297 stock->nr_slab_reclaimable_b = 0;
3298 }
3299 if (stock->nr_slab_unreclaimable_b) {
3300 mod_objcg_mlstate(old, stock->cached_pgdat,
3301 NR_SLAB_UNRECLAIMABLE_B,
3302 stock->nr_slab_unreclaimable_b);
3303 stock->nr_slab_unreclaimable_b = 0;
3304 }
3305 stock->cached_pgdat = NULL;
3306 }
3307
3308 WRITE_ONCE(stock->cached_objcg, NULL);
3309 /*
3310 * The `old' objects needs to be released by the caller via
3311 * obj_cgroup_put() outside of memcg_stock_pcp::stock_lock.
3312 */
3313 return old;
3314 }
3315
obj_stock_flush_required(struct memcg_stock_pcp * stock,struct mem_cgroup * root_memcg)3316 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3317 struct mem_cgroup *root_memcg)
3318 {
3319 struct obj_cgroup *objcg = READ_ONCE(stock->cached_objcg);
3320 struct mem_cgroup *memcg;
3321
3322 if (objcg) {
3323 memcg = obj_cgroup_memcg(objcg);
3324 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3325 return true;
3326 }
3327
3328 return false;
3329 }
3330
refill_obj_stock(struct obj_cgroup * objcg,unsigned int nr_bytes,bool allow_uncharge)3331 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes,
3332 bool allow_uncharge)
3333 {
3334 struct memcg_stock_pcp *stock;
3335 struct obj_cgroup *old = NULL;
3336 unsigned long flags;
3337 unsigned int nr_pages = 0;
3338
3339 local_lock_irqsave(&memcg_stock.stock_lock, flags);
3340
3341 stock = this_cpu_ptr(&memcg_stock);
3342 if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
3343 old = drain_obj_stock(stock);
3344 obj_cgroup_get(objcg);
3345 WRITE_ONCE(stock->cached_objcg, objcg);
3346 stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
3347 ? atomic_xchg(&objcg->nr_charged_bytes, 0) : 0;
3348 allow_uncharge = true; /* Allow uncharge when objcg changes */
3349 }
3350 stock->nr_bytes += nr_bytes;
3351
3352 if (allow_uncharge && (stock->nr_bytes > PAGE_SIZE)) {
3353 nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3354 stock->nr_bytes &= (PAGE_SIZE - 1);
3355 }
3356
3357 local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
3358 if (old)
3359 obj_cgroup_put(old);
3360
3361 if (nr_pages)
3362 obj_cgroup_uncharge_pages(objcg, nr_pages);
3363 }
3364
obj_cgroup_charge(struct obj_cgroup * objcg,gfp_t gfp,size_t size)3365 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3366 {
3367 unsigned int nr_pages, nr_bytes;
3368 int ret;
3369
3370 if (consume_obj_stock(objcg, size))
3371 return 0;
3372
3373 /*
3374 * In theory, objcg->nr_charged_bytes can have enough
3375 * pre-charged bytes to satisfy the allocation. However,
3376 * flushing objcg->nr_charged_bytes requires two atomic
3377 * operations, and objcg->nr_charged_bytes can't be big.
3378 * The shared objcg->nr_charged_bytes can also become a
3379 * performance bottleneck if all tasks of the same memcg are
3380 * trying to update it. So it's better to ignore it and try
3381 * grab some new pages. The stock's nr_bytes will be flushed to
3382 * objcg->nr_charged_bytes later on when objcg changes.
3383 *
3384 * The stock's nr_bytes may contain enough pre-charged bytes
3385 * to allow one less page from being charged, but we can't rely
3386 * on the pre-charged bytes not being changed outside of
3387 * consume_obj_stock() or refill_obj_stock(). So ignore those
3388 * pre-charged bytes as well when charging pages. To avoid a
3389 * page uncharge right after a page charge, we set the
3390 * allow_uncharge flag to false when calling refill_obj_stock()
3391 * to temporarily allow the pre-charged bytes to exceed the page
3392 * size limit. The maximum reachable value of the pre-charged
3393 * bytes is (sizeof(object) + PAGE_SIZE - 2) if there is no data
3394 * race.
3395 */
3396 nr_pages = size >> PAGE_SHIFT;
3397 nr_bytes = size & (PAGE_SIZE - 1);
3398
3399 if (nr_bytes)
3400 nr_pages += 1;
3401
3402 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3403 if (!ret && nr_bytes)
3404 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes, false);
3405
3406 return ret;
3407 }
3408
obj_cgroup_uncharge(struct obj_cgroup * objcg,size_t size)3409 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3410 {
3411 refill_obj_stock(objcg, size, true);
3412 }
3413
3414 #endif /* CONFIG_MEMCG_KMEM */
3415
3416 /*
3417 * Because page_memcg(head) is not set on tails, set it now.
3418 */
split_page_memcg(struct page * head,unsigned int nr)3419 void split_page_memcg(struct page *head, unsigned int nr)
3420 {
3421 struct folio *folio = page_folio(head);
3422 struct mem_cgroup *memcg = folio_memcg(folio);
3423 int i;
3424
3425 if (mem_cgroup_disabled() || !memcg)
3426 return;
3427
3428 for (i = 1; i < nr; i++)
3429 folio_page(folio, i)->memcg_data = folio->memcg_data;
3430
3431 if (folio_memcg_kmem(folio))
3432 obj_cgroup_get_many(__folio_objcg(folio), nr - 1);
3433 else
3434 css_get_many(&memcg->css, nr - 1);
3435 }
3436
3437 #ifdef CONFIG_SWAP
3438 /**
3439 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3440 * @entry: swap entry to be moved
3441 * @from: mem_cgroup which the entry is moved from
3442 * @to: mem_cgroup which the entry is moved to
3443 *
3444 * It succeeds only when the swap_cgroup's record for this entry is the same
3445 * as the mem_cgroup's id of @from.
3446 *
3447 * Returns 0 on success, -EINVAL on failure.
3448 *
3449 * The caller must have charged to @to, IOW, called page_counter_charge() about
3450 * both res and memsw, and called css_get().
3451 */
mem_cgroup_move_swap_account(swp_entry_t entry,struct mem_cgroup * from,struct mem_cgroup * to)3452 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3453 struct mem_cgroup *from, struct mem_cgroup *to)
3454 {
3455 unsigned short old_id, new_id;
3456
3457 old_id = mem_cgroup_id(from);
3458 new_id = mem_cgroup_id(to);
3459
3460 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3461 mod_memcg_state(from, MEMCG_SWAP, -1);
3462 mod_memcg_state(to, MEMCG_SWAP, 1);
3463 return 0;
3464 }
3465 return -EINVAL;
3466 }
3467 #else
mem_cgroup_move_swap_account(swp_entry_t entry,struct mem_cgroup * from,struct mem_cgroup * to)3468 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3469 struct mem_cgroup *from, struct mem_cgroup *to)
3470 {
3471 return -EINVAL;
3472 }
3473 #endif
3474
3475 static DEFINE_MUTEX(memcg_max_mutex);
3476
mem_cgroup_resize_max(struct mem_cgroup * memcg,unsigned long max,bool memsw)3477 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3478 unsigned long max, bool memsw)
3479 {
3480 bool enlarge = false;
3481 bool drained = false;
3482 int ret;
3483 bool limits_invariant;
3484 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3485
3486 do {
3487 if (signal_pending(current)) {
3488 ret = -EINTR;
3489 break;
3490 }
3491
3492 mutex_lock(&memcg_max_mutex);
3493 /*
3494 * Make sure that the new limit (memsw or memory limit) doesn't
3495 * break our basic invariant rule memory.max <= memsw.max.
3496 */
3497 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3498 max <= memcg->memsw.max;
3499 if (!limits_invariant) {
3500 mutex_unlock(&memcg_max_mutex);
3501 ret = -EINVAL;
3502 break;
3503 }
3504 if (max > counter->max)
3505 enlarge = true;
3506 ret = page_counter_set_max(counter, max);
3507 mutex_unlock(&memcg_max_mutex);
3508
3509 if (!ret)
3510 break;
3511
3512 if (!drained) {
3513 drain_all_stock(memcg);
3514 drained = true;
3515 continue;
3516 }
3517
3518 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3519 memsw ? 0 : MEMCG_RECLAIM_MAY_SWAP)) {
3520 ret = -EBUSY;
3521 break;
3522 }
3523 } while (true);
3524
3525 if (!ret && enlarge)
3526 memcg_oom_recover(memcg);
3527
3528 return ret;
3529 }
3530
mem_cgroup_soft_limit_reclaim(pg_data_t * pgdat,int order,gfp_t gfp_mask,unsigned long * total_scanned)3531 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3532 gfp_t gfp_mask,
3533 unsigned long *total_scanned)
3534 {
3535 unsigned long nr_reclaimed = 0;
3536 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3537 unsigned long reclaimed;
3538 int loop = 0;
3539 struct mem_cgroup_tree_per_node *mctz;
3540 unsigned long excess;
3541
3542 if (lru_gen_enabled())
3543 return 0;
3544
3545 if (order > 0)
3546 return 0;
3547
3548 mctz = soft_limit_tree.rb_tree_per_node[pgdat->node_id];
3549
3550 /*
3551 * Do not even bother to check the largest node if the root
3552 * is empty. Do it lockless to prevent lock bouncing. Races
3553 * are acceptable as soft limit is best effort anyway.
3554 */
3555 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3556 return 0;
3557
3558 /*
3559 * This loop can run a while, specially if mem_cgroup's continuously
3560 * keep exceeding their soft limit and putting the system under
3561 * pressure
3562 */
3563 do {
3564 if (next_mz)
3565 mz = next_mz;
3566 else
3567 mz = mem_cgroup_largest_soft_limit_node(mctz);
3568 if (!mz)
3569 break;
3570
3571 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3572 gfp_mask, total_scanned);
3573 nr_reclaimed += reclaimed;
3574 spin_lock_irq(&mctz->lock);
3575
3576 /*
3577 * If we failed to reclaim anything from this memory cgroup
3578 * it is time to move on to the next cgroup
3579 */
3580 next_mz = NULL;
3581 if (!reclaimed)
3582 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3583
3584 excess = soft_limit_excess(mz->memcg);
3585 /*
3586 * One school of thought says that we should not add
3587 * back the node to the tree if reclaim returns 0.
3588 * But our reclaim could return 0, simply because due
3589 * to priority we are exposing a smaller subset of
3590 * memory to reclaim from. Consider this as a longer
3591 * term TODO.
3592 */
3593 /* If excess == 0, no tree ops */
3594 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3595 spin_unlock_irq(&mctz->lock);
3596 css_put(&mz->memcg->css);
3597 loop++;
3598 /*
3599 * Could not reclaim anything and there are no more
3600 * mem cgroups to try or we seem to be looping without
3601 * reclaiming anything.
3602 */
3603 if (!nr_reclaimed &&
3604 (next_mz == NULL ||
3605 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3606 break;
3607 } while (!nr_reclaimed);
3608 if (next_mz)
3609 css_put(&next_mz->memcg->css);
3610 return nr_reclaimed;
3611 }
3612
3613 /*
3614 * Reclaims as many pages from the given memcg as possible.
3615 *
3616 * Caller is responsible for holding css reference for memcg.
3617 */
mem_cgroup_force_empty(struct mem_cgroup * memcg)3618 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3619 {
3620 int nr_retries = MAX_RECLAIM_RETRIES;
3621
3622 /* we call try-to-free pages for make this cgroup empty */
3623 lru_add_drain_all();
3624
3625 drain_all_stock(memcg);
3626
3627 /* try to free all pages in this cgroup */
3628 while (nr_retries && page_counter_read(&memcg->memory)) {
3629 if (signal_pending(current))
3630 return -EINTR;
3631
3632 if (!try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL,
3633 MEMCG_RECLAIM_MAY_SWAP))
3634 nr_retries--;
3635 }
3636
3637 return 0;
3638 }
3639
mem_cgroup_force_empty_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3640 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3641 char *buf, size_t nbytes,
3642 loff_t off)
3643 {
3644 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3645
3646 if (mem_cgroup_is_root(memcg))
3647 return -EINVAL;
3648 return mem_cgroup_force_empty(memcg) ?: nbytes;
3649 }
3650
mem_cgroup_hierarchy_read(struct cgroup_subsys_state * css,struct cftype * cft)3651 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3652 struct cftype *cft)
3653 {
3654 return 1;
3655 }
3656
mem_cgroup_hierarchy_write(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)3657 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3658 struct cftype *cft, u64 val)
3659 {
3660 if (val == 1)
3661 return 0;
3662
3663 pr_warn_once("Non-hierarchical mode is deprecated. "
3664 "Please report your usecase to linux-mm@kvack.org if you "
3665 "depend on this functionality.\n");
3666
3667 return -EINVAL;
3668 }
3669
mem_cgroup_usage(struct mem_cgroup * memcg,bool swap)3670 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3671 {
3672 unsigned long val;
3673
3674 if (mem_cgroup_is_root(memcg)) {
3675 /*
3676 * Approximate root's usage from global state. This isn't
3677 * perfect, but the root usage was always an approximation.
3678 */
3679 val = global_node_page_state(NR_FILE_PAGES) +
3680 global_node_page_state(NR_ANON_MAPPED);
3681 if (swap)
3682 val += total_swap_pages - get_nr_swap_pages();
3683 } else {
3684 if (!swap)
3685 val = page_counter_read(&memcg->memory);
3686 else
3687 val = page_counter_read(&memcg->memsw);
3688 }
3689 return val;
3690 }
3691
3692 enum {
3693 RES_USAGE,
3694 RES_LIMIT,
3695 RES_MAX_USAGE,
3696 RES_FAILCNT,
3697 RES_SOFT_LIMIT,
3698 };
3699
mem_cgroup_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)3700 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3701 struct cftype *cft)
3702 {
3703 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3704 struct page_counter *counter;
3705
3706 switch (MEMFILE_TYPE(cft->private)) {
3707 case _MEM:
3708 counter = &memcg->memory;
3709 break;
3710 case _MEMSWAP:
3711 counter = &memcg->memsw;
3712 break;
3713 case _KMEM:
3714 counter = &memcg->kmem;
3715 break;
3716 case _TCP:
3717 counter = &memcg->tcpmem;
3718 break;
3719 default:
3720 BUG();
3721 }
3722
3723 switch (MEMFILE_ATTR(cft->private)) {
3724 case RES_USAGE:
3725 if (counter == &memcg->memory)
3726 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3727 if (counter == &memcg->memsw)
3728 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3729 return (u64)page_counter_read(counter) * PAGE_SIZE;
3730 case RES_LIMIT:
3731 return (u64)counter->max * PAGE_SIZE;
3732 case RES_MAX_USAGE:
3733 return (u64)counter->watermark * PAGE_SIZE;
3734 case RES_FAILCNT:
3735 return counter->failcnt;
3736 case RES_SOFT_LIMIT:
3737 return (u64)READ_ONCE(memcg->soft_limit) * PAGE_SIZE;
3738 default:
3739 BUG();
3740 }
3741 }
3742
3743 /*
3744 * This function doesn't do anything useful. Its only job is to provide a read
3745 * handler for a file so that cgroup_file_mode() will add read permissions.
3746 */
mem_cgroup_dummy_seq_show(__always_unused struct seq_file * m,__always_unused void * v)3747 static int mem_cgroup_dummy_seq_show(__always_unused struct seq_file *m,
3748 __always_unused void *v)
3749 {
3750 return -EINVAL;
3751 }
3752
3753 #ifdef CONFIG_MEMCG_KMEM
memcg_online_kmem(struct mem_cgroup * memcg)3754 static int memcg_online_kmem(struct mem_cgroup *memcg)
3755 {
3756 struct obj_cgroup *objcg;
3757
3758 if (mem_cgroup_kmem_disabled())
3759 return 0;
3760
3761 if (unlikely(mem_cgroup_is_root(memcg)))
3762 return 0;
3763
3764 objcg = obj_cgroup_alloc();
3765 if (!objcg)
3766 return -ENOMEM;
3767
3768 objcg->memcg = memcg;
3769 rcu_assign_pointer(memcg->objcg, objcg);
3770
3771 static_branch_enable(&memcg_kmem_online_key);
3772
3773 memcg->kmemcg_id = memcg->id.id;
3774
3775 return 0;
3776 }
3777
memcg_offline_kmem(struct mem_cgroup * memcg)3778 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3779 {
3780 struct mem_cgroup *parent;
3781
3782 if (mem_cgroup_kmem_disabled())
3783 return;
3784
3785 if (unlikely(mem_cgroup_is_root(memcg)))
3786 return;
3787
3788 parent = parent_mem_cgroup(memcg);
3789 if (!parent)
3790 parent = root_mem_cgroup;
3791
3792 memcg_reparent_objcgs(memcg, parent);
3793
3794 /*
3795 * After we have finished memcg_reparent_objcgs(), all list_lrus
3796 * corresponding to this cgroup are guaranteed to remain empty.
3797 * The ordering is imposed by list_lru_node->lock taken by
3798 * memcg_reparent_list_lrus().
3799 */
3800 memcg_reparent_list_lrus(memcg, parent);
3801 }
3802 #else
memcg_online_kmem(struct mem_cgroup * memcg)3803 static int memcg_online_kmem(struct mem_cgroup *memcg)
3804 {
3805 return 0;
3806 }
memcg_offline_kmem(struct mem_cgroup * memcg)3807 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3808 {
3809 }
3810 #endif /* CONFIG_MEMCG_KMEM */
3811
memcg_update_tcp_max(struct mem_cgroup * memcg,unsigned long max)3812 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3813 {
3814 int ret;
3815
3816 mutex_lock(&memcg_max_mutex);
3817
3818 ret = page_counter_set_max(&memcg->tcpmem, max);
3819 if (ret)
3820 goto out;
3821
3822 if (!memcg->tcpmem_active) {
3823 /*
3824 * The active flag needs to be written after the static_key
3825 * update. This is what guarantees that the socket activation
3826 * function is the last one to run. See mem_cgroup_sk_alloc()
3827 * for details, and note that we don't mark any socket as
3828 * belonging to this memcg until that flag is up.
3829 *
3830 * We need to do this, because static_keys will span multiple
3831 * sites, but we can't control their order. If we mark a socket
3832 * as accounted, but the accounting functions are not patched in
3833 * yet, we'll lose accounting.
3834 *
3835 * We never race with the readers in mem_cgroup_sk_alloc(),
3836 * because when this value change, the code to process it is not
3837 * patched in yet.
3838 */
3839 static_branch_inc(&memcg_sockets_enabled_key);
3840 memcg->tcpmem_active = true;
3841 }
3842 out:
3843 mutex_unlock(&memcg_max_mutex);
3844 return ret;
3845 }
3846
3847 /*
3848 * The user of this function is...
3849 * RES_LIMIT.
3850 */
mem_cgroup_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3851 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3852 char *buf, size_t nbytes, loff_t off)
3853 {
3854 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3855 unsigned long nr_pages;
3856 int ret;
3857
3858 buf = strstrip(buf);
3859 ret = page_counter_memparse(buf, "-1", &nr_pages);
3860 if (ret)
3861 return ret;
3862
3863 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3864 case RES_LIMIT:
3865 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3866 ret = -EINVAL;
3867 break;
3868 }
3869 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3870 case _MEM:
3871 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3872 break;
3873 case _MEMSWAP:
3874 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3875 break;
3876 case _KMEM:
3877 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3878 "Writing any value to this file has no effect. "
3879 "Please report your usecase to linux-mm@kvack.org if you "
3880 "depend on this functionality.\n");
3881 ret = 0;
3882 break;
3883 case _TCP:
3884 ret = memcg_update_tcp_max(memcg, nr_pages);
3885 break;
3886 }
3887 break;
3888 case RES_SOFT_LIMIT:
3889 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
3890 ret = -EOPNOTSUPP;
3891 } else {
3892 WRITE_ONCE(memcg->soft_limit, nr_pages);
3893 ret = 0;
3894 }
3895 break;
3896 }
3897 return ret ?: nbytes;
3898 }
3899
mem_cgroup_reset(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3900 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3901 size_t nbytes, loff_t off)
3902 {
3903 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3904 struct page_counter *counter;
3905
3906 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3907 case _MEM:
3908 counter = &memcg->memory;
3909 break;
3910 case _MEMSWAP:
3911 counter = &memcg->memsw;
3912 break;
3913 case _KMEM:
3914 counter = &memcg->kmem;
3915 break;
3916 case _TCP:
3917 counter = &memcg->tcpmem;
3918 break;
3919 default:
3920 BUG();
3921 }
3922
3923 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3924 case RES_MAX_USAGE:
3925 page_counter_reset_watermark(counter);
3926 break;
3927 case RES_FAILCNT:
3928 counter->failcnt = 0;
3929 break;
3930 default:
3931 BUG();
3932 }
3933
3934 return nbytes;
3935 }
3936
mem_cgroup_move_charge_read(struct cgroup_subsys_state * css,struct cftype * cft)3937 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3938 struct cftype *cft)
3939 {
3940 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3941 }
3942
3943 #ifdef CONFIG_MMU
mem_cgroup_move_charge_write(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)3944 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3945 struct cftype *cft, u64 val)
3946 {
3947 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3948
3949 pr_warn_once("Cgroup memory moving (move_charge_at_immigrate) is deprecated. "
3950 "Please report your usecase to linux-mm@kvack.org if you "
3951 "depend on this functionality.\n");
3952
3953 if (val & ~MOVE_MASK)
3954 return -EINVAL;
3955
3956 /*
3957 * No kind of locking is needed in here, because ->can_attach() will
3958 * check this value once in the beginning of the process, and then carry
3959 * on with stale data. This means that changes to this value will only
3960 * affect task migrations starting after the change.
3961 */
3962 memcg->move_charge_at_immigrate = val;
3963 return 0;
3964 }
3965 #else
mem_cgroup_move_charge_write(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)3966 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3967 struct cftype *cft, u64 val)
3968 {
3969 return -ENOSYS;
3970 }
3971 #endif
3972
3973 #ifdef CONFIG_NUMA
3974
3975 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3976 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3977 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3978
mem_cgroup_node_nr_lru_pages(struct mem_cgroup * memcg,int nid,unsigned int lru_mask,bool tree)3979 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3980 int nid, unsigned int lru_mask, bool tree)
3981 {
3982 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3983 unsigned long nr = 0;
3984 enum lru_list lru;
3985
3986 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3987
3988 for_each_lru(lru) {
3989 if (!(BIT(lru) & lru_mask))
3990 continue;
3991 if (tree)
3992 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3993 else
3994 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3995 }
3996 return nr;
3997 }
3998
mem_cgroup_nr_lru_pages(struct mem_cgroup * memcg,unsigned int lru_mask,bool tree)3999 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
4000 unsigned int lru_mask,
4001 bool tree)
4002 {
4003 unsigned long nr = 0;
4004 enum lru_list lru;
4005
4006 for_each_lru(lru) {
4007 if (!(BIT(lru) & lru_mask))
4008 continue;
4009 if (tree)
4010 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
4011 else
4012 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
4013 }
4014 return nr;
4015 }
4016
memcg_numa_stat_show(struct seq_file * m,void * v)4017 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4018 {
4019 struct numa_stat {
4020 const char *name;
4021 unsigned int lru_mask;
4022 };
4023
4024 static const struct numa_stat stats[] = {
4025 { "total", LRU_ALL },
4026 { "file", LRU_ALL_FILE },
4027 { "anon", LRU_ALL_ANON },
4028 { "unevictable", BIT(LRU_UNEVICTABLE) },
4029 };
4030 const struct numa_stat *stat;
4031 int nid;
4032 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4033
4034 mem_cgroup_flush_stats();
4035
4036 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4037 seq_printf(m, "%s=%lu", stat->name,
4038 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4039 false));
4040 for_each_node_state(nid, N_MEMORY)
4041 seq_printf(m, " N%d=%lu", nid,
4042 mem_cgroup_node_nr_lru_pages(memcg, nid,
4043 stat->lru_mask, false));
4044 seq_putc(m, '\n');
4045 }
4046
4047 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4048
4049 seq_printf(m, "hierarchical_%s=%lu", stat->name,
4050 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
4051 true));
4052 for_each_node_state(nid, N_MEMORY)
4053 seq_printf(m, " N%d=%lu", nid,
4054 mem_cgroup_node_nr_lru_pages(memcg, nid,
4055 stat->lru_mask, true));
4056 seq_putc(m, '\n');
4057 }
4058
4059 return 0;
4060 }
4061 #endif /* CONFIG_NUMA */
4062
4063 static const unsigned int memcg1_stats[] = {
4064 NR_FILE_PAGES,
4065 NR_ANON_MAPPED,
4066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4067 NR_ANON_THPS,
4068 #endif
4069 NR_SHMEM,
4070 NR_FILE_MAPPED,
4071 NR_FILE_DIRTY,
4072 NR_WRITEBACK,
4073 WORKINGSET_REFAULT_ANON,
4074 WORKINGSET_REFAULT_FILE,
4075 MEMCG_SWAP,
4076 };
4077
4078 static const char *const memcg1_stat_names[] = {
4079 "cache",
4080 "rss",
4081 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4082 "rss_huge",
4083 #endif
4084 "shmem",
4085 "mapped_file",
4086 "dirty",
4087 "writeback",
4088 "workingset_refault_anon",
4089 "workingset_refault_file",
4090 "swap",
4091 };
4092
4093 /* Universal VM events cgroup1 shows, original sort order */
4094 static const unsigned int memcg1_events[] = {
4095 PGPGIN,
4096 PGPGOUT,
4097 PGFAULT,
4098 PGMAJFAULT,
4099 };
4100
memcg1_stat_format(struct mem_cgroup * memcg,struct seq_buf * s)4101 static void memcg1_stat_format(struct mem_cgroup *memcg, struct seq_buf *s)
4102 {
4103 unsigned long memory, memsw;
4104 struct mem_cgroup *mi;
4105 unsigned int i;
4106
4107 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4108
4109 mem_cgroup_flush_stats();
4110
4111 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4112 unsigned long nr;
4113
4114 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4115 continue;
4116 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4117 seq_buf_printf(s, "%s %lu\n", memcg1_stat_names[i],
4118 nr * memcg_page_state_unit(memcg1_stats[i]));
4119 }
4120
4121 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4122 seq_buf_printf(s, "%s %lu\n", vm_event_name(memcg1_events[i]),
4123 memcg_events_local(memcg, memcg1_events[i]));
4124
4125 for (i = 0; i < NR_LRU_LISTS; i++)
4126 seq_buf_printf(s, "%s %lu\n", lru_list_name(i),
4127 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4128 PAGE_SIZE);
4129
4130 /* Hierarchical information */
4131 memory = memsw = PAGE_COUNTER_MAX;
4132 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4133 memory = min(memory, READ_ONCE(mi->memory.max));
4134 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4135 }
4136 seq_buf_printf(s, "hierarchical_memory_limit %llu\n",
4137 (u64)memory * PAGE_SIZE);
4138 if (do_memsw_account())
4139 seq_buf_printf(s, "hierarchical_memsw_limit %llu\n",
4140 (u64)memsw * PAGE_SIZE);
4141
4142 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4143 unsigned long nr;
4144
4145 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4146 continue;
4147 nr = memcg_page_state(memcg, memcg1_stats[i]);
4148 seq_buf_printf(s, "total_%s %llu\n", memcg1_stat_names[i],
4149 (u64)nr * memcg_page_state_unit(memcg1_stats[i]));
4150 }
4151
4152 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4153 seq_buf_printf(s, "total_%s %llu\n",
4154 vm_event_name(memcg1_events[i]),
4155 (u64)memcg_events(memcg, memcg1_events[i]));
4156
4157 for (i = 0; i < NR_LRU_LISTS; i++)
4158 seq_buf_printf(s, "total_%s %llu\n", lru_list_name(i),
4159 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4160 PAGE_SIZE);
4161
4162 #ifdef CONFIG_DEBUG_VM
4163 {
4164 pg_data_t *pgdat;
4165 struct mem_cgroup_per_node *mz;
4166 unsigned long anon_cost = 0;
4167 unsigned long file_cost = 0;
4168
4169 for_each_online_pgdat(pgdat) {
4170 mz = memcg->nodeinfo[pgdat->node_id];
4171
4172 anon_cost += mz->lruvec.anon_cost;
4173 file_cost += mz->lruvec.file_cost;
4174 }
4175 seq_buf_printf(s, "anon_cost %lu\n", anon_cost);
4176 seq_buf_printf(s, "file_cost %lu\n", file_cost);
4177 }
4178 #endif
4179 }
4180
mem_cgroup_swappiness_read(struct cgroup_subsys_state * css,struct cftype * cft)4181 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4182 struct cftype *cft)
4183 {
4184 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4185
4186 return mem_cgroup_swappiness(memcg);
4187 }
4188
mem_cgroup_swappiness_write(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)4189 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4190 struct cftype *cft, u64 val)
4191 {
4192 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4193
4194 if (val > 200)
4195 return -EINVAL;
4196
4197 if (!mem_cgroup_is_root(memcg))
4198 WRITE_ONCE(memcg->swappiness, val);
4199 else
4200 WRITE_ONCE(vm_swappiness, val);
4201
4202 return 0;
4203 }
4204
__mem_cgroup_threshold(struct mem_cgroup * memcg,bool swap)4205 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4206 {
4207 struct mem_cgroup_threshold_ary *t;
4208 unsigned long usage;
4209 int i;
4210
4211 rcu_read_lock();
4212 if (!swap)
4213 t = rcu_dereference(memcg->thresholds.primary);
4214 else
4215 t = rcu_dereference(memcg->memsw_thresholds.primary);
4216
4217 if (!t)
4218 goto unlock;
4219
4220 usage = mem_cgroup_usage(memcg, swap);
4221
4222 /*
4223 * current_threshold points to threshold just below or equal to usage.
4224 * If it's not true, a threshold was crossed after last
4225 * call of __mem_cgroup_threshold().
4226 */
4227 i = t->current_threshold;
4228
4229 /*
4230 * Iterate backward over array of thresholds starting from
4231 * current_threshold and check if a threshold is crossed.
4232 * If none of thresholds below usage is crossed, we read
4233 * only one element of the array here.
4234 */
4235 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4236 eventfd_signal(t->entries[i].eventfd, 1);
4237
4238 /* i = current_threshold + 1 */
4239 i++;
4240
4241 /*
4242 * Iterate forward over array of thresholds starting from
4243 * current_threshold+1 and check if a threshold is crossed.
4244 * If none of thresholds above usage is crossed, we read
4245 * only one element of the array here.
4246 */
4247 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4248 eventfd_signal(t->entries[i].eventfd, 1);
4249
4250 /* Update current_threshold */
4251 t->current_threshold = i - 1;
4252 unlock:
4253 rcu_read_unlock();
4254 }
4255
mem_cgroup_threshold(struct mem_cgroup * memcg)4256 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4257 {
4258 while (memcg) {
4259 __mem_cgroup_threshold(memcg, false);
4260 if (do_memsw_account())
4261 __mem_cgroup_threshold(memcg, true);
4262
4263 memcg = parent_mem_cgroup(memcg);
4264 }
4265 }
4266
compare_thresholds(const void * a,const void * b)4267 static int compare_thresholds(const void *a, const void *b)
4268 {
4269 const struct mem_cgroup_threshold *_a = a;
4270 const struct mem_cgroup_threshold *_b = b;
4271
4272 if (_a->threshold > _b->threshold)
4273 return 1;
4274
4275 if (_a->threshold < _b->threshold)
4276 return -1;
4277
4278 return 0;
4279 }
4280
mem_cgroup_oom_notify_cb(struct mem_cgroup * memcg)4281 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4282 {
4283 struct mem_cgroup_eventfd_list *ev;
4284
4285 spin_lock(&memcg_oom_lock);
4286
4287 list_for_each_entry(ev, &memcg->oom_notify, list)
4288 eventfd_signal(ev->eventfd, 1);
4289
4290 spin_unlock(&memcg_oom_lock);
4291 return 0;
4292 }
4293
mem_cgroup_oom_notify(struct mem_cgroup * memcg)4294 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4295 {
4296 struct mem_cgroup *iter;
4297
4298 for_each_mem_cgroup_tree(iter, memcg)
4299 mem_cgroup_oom_notify_cb(iter);
4300 }
4301
__mem_cgroup_usage_register_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd,const char * args,enum res_type type)4302 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4303 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4304 {
4305 struct mem_cgroup_thresholds *thresholds;
4306 struct mem_cgroup_threshold_ary *new;
4307 unsigned long threshold;
4308 unsigned long usage;
4309 int i, size, ret;
4310
4311 ret = page_counter_memparse(args, "-1", &threshold);
4312 if (ret)
4313 return ret;
4314
4315 mutex_lock(&memcg->thresholds_lock);
4316
4317 if (type == _MEM) {
4318 thresholds = &memcg->thresholds;
4319 usage = mem_cgroup_usage(memcg, false);
4320 } else if (type == _MEMSWAP) {
4321 thresholds = &memcg->memsw_thresholds;
4322 usage = mem_cgroup_usage(memcg, true);
4323 } else
4324 BUG();
4325
4326 /* Check if a threshold crossed before adding a new one */
4327 if (thresholds->primary)
4328 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4329
4330 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4331
4332 /* Allocate memory for new array of thresholds */
4333 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4334 if (!new) {
4335 ret = -ENOMEM;
4336 goto unlock;
4337 }
4338 new->size = size;
4339
4340 /* Copy thresholds (if any) to new array */
4341 if (thresholds->primary)
4342 memcpy(new->entries, thresholds->primary->entries,
4343 flex_array_size(new, entries, size - 1));
4344
4345 /* Add new threshold */
4346 new->entries[size - 1].eventfd = eventfd;
4347 new->entries[size - 1].threshold = threshold;
4348
4349 /* Sort thresholds. Registering of new threshold isn't time-critical */
4350 sort(new->entries, size, sizeof(*new->entries),
4351 compare_thresholds, NULL);
4352
4353 /* Find current threshold */
4354 new->current_threshold = -1;
4355 for (i = 0; i < size; i++) {
4356 if (new->entries[i].threshold <= usage) {
4357 /*
4358 * new->current_threshold will not be used until
4359 * rcu_assign_pointer(), so it's safe to increment
4360 * it here.
4361 */
4362 ++new->current_threshold;
4363 } else
4364 break;
4365 }
4366
4367 /* Free old spare buffer and save old primary buffer as spare */
4368 kfree(thresholds->spare);
4369 thresholds->spare = thresholds->primary;
4370
4371 rcu_assign_pointer(thresholds->primary, new);
4372
4373 /* To be sure that nobody uses thresholds */
4374 synchronize_rcu();
4375
4376 unlock:
4377 mutex_unlock(&memcg->thresholds_lock);
4378
4379 return ret;
4380 }
4381
mem_cgroup_usage_register_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd,const char * args)4382 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4383 struct eventfd_ctx *eventfd, const char *args)
4384 {
4385 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4386 }
4387
memsw_cgroup_usage_register_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd,const char * args)4388 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4389 struct eventfd_ctx *eventfd, const char *args)
4390 {
4391 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4392 }
4393
__mem_cgroup_usage_unregister_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd,enum res_type type)4394 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4395 struct eventfd_ctx *eventfd, enum res_type type)
4396 {
4397 struct mem_cgroup_thresholds *thresholds;
4398 struct mem_cgroup_threshold_ary *new;
4399 unsigned long usage;
4400 int i, j, size, entries;
4401
4402 mutex_lock(&memcg->thresholds_lock);
4403
4404 if (type == _MEM) {
4405 thresholds = &memcg->thresholds;
4406 usage = mem_cgroup_usage(memcg, false);
4407 } else if (type == _MEMSWAP) {
4408 thresholds = &memcg->memsw_thresholds;
4409 usage = mem_cgroup_usage(memcg, true);
4410 } else
4411 BUG();
4412
4413 if (!thresholds->primary)
4414 goto unlock;
4415
4416 /* Check if a threshold crossed before removing */
4417 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4418
4419 /* Calculate new number of threshold */
4420 size = entries = 0;
4421 for (i = 0; i < thresholds->primary->size; i++) {
4422 if (thresholds->primary->entries[i].eventfd != eventfd)
4423 size++;
4424 else
4425 entries++;
4426 }
4427
4428 new = thresholds->spare;
4429
4430 /* If no items related to eventfd have been cleared, nothing to do */
4431 if (!entries)
4432 goto unlock;
4433
4434 /* Set thresholds array to NULL if we don't have thresholds */
4435 if (!size) {
4436 kfree(new);
4437 new = NULL;
4438 goto swap_buffers;
4439 }
4440
4441 new->size = size;
4442
4443 /* Copy thresholds and find current threshold */
4444 new->current_threshold = -1;
4445 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4446 if (thresholds->primary->entries[i].eventfd == eventfd)
4447 continue;
4448
4449 new->entries[j] = thresholds->primary->entries[i];
4450 if (new->entries[j].threshold <= usage) {
4451 /*
4452 * new->current_threshold will not be used
4453 * until rcu_assign_pointer(), so it's safe to increment
4454 * it here.
4455 */
4456 ++new->current_threshold;
4457 }
4458 j++;
4459 }
4460
4461 swap_buffers:
4462 /* Swap primary and spare array */
4463 thresholds->spare = thresholds->primary;
4464
4465 rcu_assign_pointer(thresholds->primary, new);
4466
4467 /* To be sure that nobody uses thresholds */
4468 synchronize_rcu();
4469
4470 /* If all events are unregistered, free the spare array */
4471 if (!new) {
4472 kfree(thresholds->spare);
4473 thresholds->spare = NULL;
4474 }
4475 unlock:
4476 mutex_unlock(&memcg->thresholds_lock);
4477 }
4478
mem_cgroup_usage_unregister_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd)4479 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4480 struct eventfd_ctx *eventfd)
4481 {
4482 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4483 }
4484
memsw_cgroup_usage_unregister_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd)4485 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4486 struct eventfd_ctx *eventfd)
4487 {
4488 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4489 }
4490
mem_cgroup_oom_register_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd,const char * args)4491 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4492 struct eventfd_ctx *eventfd, const char *args)
4493 {
4494 struct mem_cgroup_eventfd_list *event;
4495
4496 event = kmalloc(sizeof(*event), GFP_KERNEL);
4497 if (!event)
4498 return -ENOMEM;
4499
4500 spin_lock(&memcg_oom_lock);
4501
4502 event->eventfd = eventfd;
4503 list_add(&event->list, &memcg->oom_notify);
4504
4505 /* already in OOM ? */
4506 if (memcg->under_oom)
4507 eventfd_signal(eventfd, 1);
4508 spin_unlock(&memcg_oom_lock);
4509
4510 return 0;
4511 }
4512
mem_cgroup_oom_unregister_event(struct mem_cgroup * memcg,struct eventfd_ctx * eventfd)4513 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4514 struct eventfd_ctx *eventfd)
4515 {
4516 struct mem_cgroup_eventfd_list *ev, *tmp;
4517
4518 spin_lock(&memcg_oom_lock);
4519
4520 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4521 if (ev->eventfd == eventfd) {
4522 list_del(&ev->list);
4523 kfree(ev);
4524 }
4525 }
4526
4527 spin_unlock(&memcg_oom_lock);
4528 }
4529
mem_cgroup_oom_control_read(struct seq_file * sf,void * v)4530 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4531 {
4532 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4533
4534 seq_printf(sf, "oom_kill_disable %d\n", READ_ONCE(memcg->oom_kill_disable));
4535 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4536 seq_printf(sf, "oom_kill %lu\n",
4537 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4538 return 0;
4539 }
4540
mem_cgroup_oom_control_write(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)4541 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4542 struct cftype *cft, u64 val)
4543 {
4544 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4545
4546 /* cannot set to root cgroup and only 0 and 1 are allowed */
4547 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4548 return -EINVAL;
4549
4550 WRITE_ONCE(memcg->oom_kill_disable, val);
4551 if (!val)
4552 memcg_oom_recover(memcg);
4553
4554 return 0;
4555 }
4556
4557 #ifdef CONFIG_CGROUP_WRITEBACK
4558
4559 #include <trace/events/writeback.h>
4560
memcg_wb_domain_init(struct mem_cgroup * memcg,gfp_t gfp)4561 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4562 {
4563 return wb_domain_init(&memcg->cgwb_domain, gfp);
4564 }
4565
memcg_wb_domain_exit(struct mem_cgroup * memcg)4566 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4567 {
4568 wb_domain_exit(&memcg->cgwb_domain);
4569 }
4570
memcg_wb_domain_size_changed(struct mem_cgroup * memcg)4571 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4572 {
4573 wb_domain_size_changed(&memcg->cgwb_domain);
4574 }
4575
mem_cgroup_wb_domain(struct bdi_writeback * wb)4576 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4577 {
4578 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4579
4580 if (!memcg->css.parent)
4581 return NULL;
4582
4583 return &memcg->cgwb_domain;
4584 }
4585
4586 /**
4587 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4588 * @wb: bdi_writeback in question
4589 * @pfilepages: out parameter for number of file pages
4590 * @pheadroom: out parameter for number of allocatable pages according to memcg
4591 * @pdirty: out parameter for number of dirty pages
4592 * @pwriteback: out parameter for number of pages under writeback
4593 *
4594 * Determine the numbers of file, headroom, dirty, and writeback pages in
4595 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4596 * is a bit more involved.
4597 *
4598 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4599 * headroom is calculated as the lowest headroom of itself and the
4600 * ancestors. Note that this doesn't consider the actual amount of
4601 * available memory in the system. The caller should further cap
4602 * *@pheadroom accordingly.
4603 */
mem_cgroup_wb_stats(struct bdi_writeback * wb,unsigned long * pfilepages,unsigned long * pheadroom,unsigned long * pdirty,unsigned long * pwriteback)4604 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4605 unsigned long *pheadroom, unsigned long *pdirty,
4606 unsigned long *pwriteback)
4607 {
4608 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4609 struct mem_cgroup *parent;
4610
4611 mem_cgroup_flush_stats();
4612
4613 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4614 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4615 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4616 memcg_page_state(memcg, NR_ACTIVE_FILE);
4617
4618 *pheadroom = PAGE_COUNTER_MAX;
4619 while ((parent = parent_mem_cgroup(memcg))) {
4620 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4621 READ_ONCE(memcg->memory.high));
4622 unsigned long used = page_counter_read(&memcg->memory);
4623
4624 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4625 memcg = parent;
4626 }
4627 }
4628
4629 /*
4630 * Foreign dirty flushing
4631 *
4632 * There's an inherent mismatch between memcg and writeback. The former
4633 * tracks ownership per-page while the latter per-inode. This was a
4634 * deliberate design decision because honoring per-page ownership in the
4635 * writeback path is complicated, may lead to higher CPU and IO overheads
4636 * and deemed unnecessary given that write-sharing an inode across
4637 * different cgroups isn't a common use-case.
4638 *
4639 * Combined with inode majority-writer ownership switching, this works well
4640 * enough in most cases but there are some pathological cases. For
4641 * example, let's say there are two cgroups A and B which keep writing to
4642 * different but confined parts of the same inode. B owns the inode and
4643 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4644 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4645 * triggering background writeback. A will be slowed down without a way to
4646 * make writeback of the dirty pages happen.
4647 *
4648 * Conditions like the above can lead to a cgroup getting repeatedly and
4649 * severely throttled after making some progress after each
4650 * dirty_expire_interval while the underlying IO device is almost
4651 * completely idle.
4652 *
4653 * Solving this problem completely requires matching the ownership tracking
4654 * granularities between memcg and writeback in either direction. However,
4655 * the more egregious behaviors can be avoided by simply remembering the
4656 * most recent foreign dirtying events and initiating remote flushes on
4657 * them when local writeback isn't enough to keep the memory clean enough.
4658 *
4659 * The following two functions implement such mechanism. When a foreign
4660 * page - a page whose memcg and writeback ownerships don't match - is
4661 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4662 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4663 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4664 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4665 * foreign bdi_writebacks which haven't expired. Both the numbers of
4666 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4667 * limited to MEMCG_CGWB_FRN_CNT.
4668 *
4669 * The mechanism only remembers IDs and doesn't hold any object references.
4670 * As being wrong occasionally doesn't matter, updates and accesses to the
4671 * records are lockless and racy.
4672 */
mem_cgroup_track_foreign_dirty_slowpath(struct folio * folio,struct bdi_writeback * wb)4673 void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
4674 struct bdi_writeback *wb)
4675 {
4676 struct mem_cgroup *memcg = folio_memcg(folio);
4677 struct memcg_cgwb_frn *frn;
4678 u64 now = get_jiffies_64();
4679 u64 oldest_at = now;
4680 int oldest = -1;
4681 int i;
4682
4683 trace_track_foreign_dirty(folio, wb);
4684
4685 /*
4686 * Pick the slot to use. If there is already a slot for @wb, keep
4687 * using it. If not replace the oldest one which isn't being
4688 * written out.
4689 */
4690 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4691 frn = &memcg->cgwb_frn[i];
4692 if (frn->bdi_id == wb->bdi->id &&
4693 frn->memcg_id == wb->memcg_css->id)
4694 break;
4695 if (time_before64(frn->at, oldest_at) &&
4696 atomic_read(&frn->done.cnt) == 1) {
4697 oldest = i;
4698 oldest_at = frn->at;
4699 }
4700 }
4701
4702 if (i < MEMCG_CGWB_FRN_CNT) {
4703 /*
4704 * Re-using an existing one. Update timestamp lazily to
4705 * avoid making the cacheline hot. We want them to be
4706 * reasonably up-to-date and significantly shorter than
4707 * dirty_expire_interval as that's what expires the record.
4708 * Use the shorter of 1s and dirty_expire_interval / 8.
4709 */
4710 unsigned long update_intv =
4711 min_t(unsigned long, HZ,
4712 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4713
4714 if (time_before64(frn->at, now - update_intv))
4715 frn->at = now;
4716 } else if (oldest >= 0) {
4717 /* replace the oldest free one */
4718 frn = &memcg->cgwb_frn[oldest];
4719 frn->bdi_id = wb->bdi->id;
4720 frn->memcg_id = wb->memcg_css->id;
4721 frn->at = now;
4722 }
4723 }
4724
4725 /* issue foreign writeback flushes for recorded foreign dirtying events */
mem_cgroup_flush_foreign(struct bdi_writeback * wb)4726 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4727 {
4728 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4729 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4730 u64 now = jiffies_64;
4731 int i;
4732
4733 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4734 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4735
4736 /*
4737 * If the record is older than dirty_expire_interval,
4738 * writeback on it has already started. No need to kick it
4739 * off again. Also, don't start a new one if there's
4740 * already one in flight.
4741 */
4742 if (time_after64(frn->at, now - intv) &&
4743 atomic_read(&frn->done.cnt) == 1) {
4744 frn->at = 0;
4745 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4746 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id,
4747 WB_REASON_FOREIGN_FLUSH,
4748 &frn->done);
4749 }
4750 }
4751 }
4752
4753 #else /* CONFIG_CGROUP_WRITEBACK */
4754
memcg_wb_domain_init(struct mem_cgroup * memcg,gfp_t gfp)4755 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4756 {
4757 return 0;
4758 }
4759
memcg_wb_domain_exit(struct mem_cgroup * memcg)4760 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4761 {
4762 }
4763
memcg_wb_domain_size_changed(struct mem_cgroup * memcg)4764 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4765 {
4766 }
4767
4768 #endif /* CONFIG_CGROUP_WRITEBACK */
4769
4770 /*
4771 * DO NOT USE IN NEW FILES.
4772 *
4773 * "cgroup.event_control" implementation.
4774 *
4775 * This is way over-engineered. It tries to support fully configurable
4776 * events for each user. Such level of flexibility is completely
4777 * unnecessary especially in the light of the planned unified hierarchy.
4778 *
4779 * Please deprecate this and replace with something simpler if at all
4780 * possible.
4781 */
4782
4783 /*
4784 * Unregister event and free resources.
4785 *
4786 * Gets called from workqueue.
4787 */
memcg_event_remove(struct work_struct * work)4788 static void memcg_event_remove(struct work_struct *work)
4789 {
4790 struct mem_cgroup_event *event =
4791 container_of(work, struct mem_cgroup_event, remove);
4792 struct mem_cgroup *memcg = event->memcg;
4793
4794 remove_wait_queue(event->wqh, &event->wait);
4795
4796 event->unregister_event(memcg, event->eventfd);
4797
4798 /* Notify userspace the event is going away. */
4799 eventfd_signal(event->eventfd, 1);
4800
4801 eventfd_ctx_put(event->eventfd);
4802 kfree(event);
4803 css_put(&memcg->css);
4804 }
4805
4806 /*
4807 * Gets called on EPOLLHUP on eventfd when user closes it.
4808 *
4809 * Called with wqh->lock held and interrupts disabled.
4810 */
memcg_event_wake(wait_queue_entry_t * wait,unsigned mode,int sync,void * key)4811 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4812 int sync, void *key)
4813 {
4814 struct mem_cgroup_event *event =
4815 container_of(wait, struct mem_cgroup_event, wait);
4816 struct mem_cgroup *memcg = event->memcg;
4817 __poll_t flags = key_to_poll(key);
4818
4819 if (flags & EPOLLHUP) {
4820 /*
4821 * If the event has been detached at cgroup removal, we
4822 * can simply return knowing the other side will cleanup
4823 * for us.
4824 *
4825 * We can't race against event freeing since the other
4826 * side will require wqh->lock via remove_wait_queue(),
4827 * which we hold.
4828 */
4829 spin_lock(&memcg->event_list_lock);
4830 if (!list_empty(&event->list)) {
4831 list_del_init(&event->list);
4832 /*
4833 * We are in atomic context, but cgroup_event_remove()
4834 * may sleep, so we have to call it in workqueue.
4835 */
4836 schedule_work(&event->remove);
4837 }
4838 spin_unlock(&memcg->event_list_lock);
4839 }
4840
4841 return 0;
4842 }
4843
memcg_event_ptable_queue_proc(struct file * file,wait_queue_head_t * wqh,poll_table * pt)4844 static void memcg_event_ptable_queue_proc(struct file *file,
4845 wait_queue_head_t *wqh, poll_table *pt)
4846 {
4847 struct mem_cgroup_event *event =
4848 container_of(pt, struct mem_cgroup_event, pt);
4849
4850 event->wqh = wqh;
4851 add_wait_queue(wqh, &event->wait);
4852 }
4853
4854 /*
4855 * DO NOT USE IN NEW FILES.
4856 *
4857 * Parse input and register new cgroup event handler.
4858 *
4859 * Input must be in format '<event_fd> <control_fd> <args>'.
4860 * Interpretation of args is defined by control file implementation.
4861 */
memcg_write_event_control(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)4862 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4863 char *buf, size_t nbytes, loff_t off)
4864 {
4865 struct cgroup_subsys_state *css = of_css(of);
4866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4867 struct mem_cgroup_event *event;
4868 struct cgroup_subsys_state *cfile_css;
4869 unsigned int efd, cfd;
4870 struct fd efile;
4871 struct fd cfile;
4872 struct dentry *cdentry;
4873 const char *name;
4874 char *endp;
4875 int ret;
4876
4877 if (IS_ENABLED(CONFIG_PREEMPT_RT))
4878 return -EOPNOTSUPP;
4879
4880 buf = strstrip(buf);
4881
4882 efd = simple_strtoul(buf, &endp, 10);
4883 if (*endp != ' ')
4884 return -EINVAL;
4885 buf = endp + 1;
4886
4887 cfd = simple_strtoul(buf, &endp, 10);
4888 if (*endp == '\0')
4889 buf = endp;
4890 else if (*endp == ' ')
4891 buf = endp + 1;
4892 else
4893 return -EINVAL;
4894
4895 event = kzalloc(sizeof(*event), GFP_KERNEL);
4896 if (!event)
4897 return -ENOMEM;
4898
4899 event->memcg = memcg;
4900 INIT_LIST_HEAD(&event->list);
4901 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4902 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4903 INIT_WORK(&event->remove, memcg_event_remove);
4904
4905 efile = fdget(efd);
4906 if (!efile.file) {
4907 ret = -EBADF;
4908 goto out_kfree;
4909 }
4910
4911 event->eventfd = eventfd_ctx_fileget(efile.file);
4912 if (IS_ERR(event->eventfd)) {
4913 ret = PTR_ERR(event->eventfd);
4914 goto out_put_efile;
4915 }
4916
4917 cfile = fdget(cfd);
4918 if (!cfile.file) {
4919 ret = -EBADF;
4920 goto out_put_eventfd;
4921 }
4922
4923 /* the process need read permission on control file */
4924 /* AV: shouldn't we check that it's been opened for read instead? */
4925 ret = file_permission(cfile.file, MAY_READ);
4926 if (ret < 0)
4927 goto out_put_cfile;
4928
4929 /*
4930 * The control file must be a regular cgroup1 file. As a regular cgroup
4931 * file can't be renamed, it's safe to access its name afterwards.
4932 */
4933 cdentry = cfile.file->f_path.dentry;
4934 if (cdentry->d_sb->s_type != &cgroup_fs_type || !d_is_reg(cdentry)) {
4935 ret = -EINVAL;
4936 goto out_put_cfile;
4937 }
4938
4939 /*
4940 * Determine the event callbacks and set them in @event. This used
4941 * to be done via struct cftype but cgroup core no longer knows
4942 * about these events. The following is crude but the whole thing
4943 * is for compatibility anyway.
4944 *
4945 * DO NOT ADD NEW FILES.
4946 */
4947 name = cdentry->d_name.name;
4948
4949 if (!strcmp(name, "memory.usage_in_bytes")) {
4950 event->register_event = mem_cgroup_usage_register_event;
4951 event->unregister_event = mem_cgroup_usage_unregister_event;
4952 } else if (!strcmp(name, "memory.oom_control")) {
4953 event->register_event = mem_cgroup_oom_register_event;
4954 event->unregister_event = mem_cgroup_oom_unregister_event;
4955 } else if (!strcmp(name, "memory.pressure_level")) {
4956 event->register_event = vmpressure_register_event;
4957 event->unregister_event = vmpressure_unregister_event;
4958 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4959 event->register_event = memsw_cgroup_usage_register_event;
4960 event->unregister_event = memsw_cgroup_usage_unregister_event;
4961 } else {
4962 ret = -EINVAL;
4963 goto out_put_cfile;
4964 }
4965
4966 /*
4967 * Verify @cfile should belong to @css. Also, remaining events are
4968 * automatically removed on cgroup destruction but the removal is
4969 * asynchronous, so take an extra ref on @css.
4970 */
4971 cfile_css = css_tryget_online_from_dir(cdentry->d_parent,
4972 &memory_cgrp_subsys);
4973 ret = -EINVAL;
4974 if (IS_ERR(cfile_css))
4975 goto out_put_cfile;
4976 if (cfile_css != css) {
4977 css_put(cfile_css);
4978 goto out_put_cfile;
4979 }
4980
4981 ret = event->register_event(memcg, event->eventfd, buf);
4982 if (ret)
4983 goto out_put_css;
4984
4985 vfs_poll(efile.file, &event->pt);
4986
4987 spin_lock_irq(&memcg->event_list_lock);
4988 list_add(&event->list, &memcg->event_list);
4989 spin_unlock_irq(&memcg->event_list_lock);
4990
4991 fdput(cfile);
4992 fdput(efile);
4993
4994 return nbytes;
4995
4996 out_put_css:
4997 css_put(css);
4998 out_put_cfile:
4999 fdput(cfile);
5000 out_put_eventfd:
5001 eventfd_ctx_put(event->eventfd);
5002 out_put_efile:
5003 fdput(efile);
5004 out_kfree:
5005 kfree(event);
5006
5007 return ret;
5008 }
5009
5010 #if defined(CONFIG_MEMCG_KMEM) && (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
mem_cgroup_slab_show(struct seq_file * m,void * p)5011 static int mem_cgroup_slab_show(struct seq_file *m, void *p)
5012 {
5013 /*
5014 * Deprecated.
5015 * Please, take a look at tools/cgroup/memcg_slabinfo.py .
5016 */
5017 return 0;
5018 }
5019 #endif
5020
5021 static int memory_stat_show(struct seq_file *m, void *v);
5022
5023 static struct cftype mem_cgroup_legacy_files[] = {
5024 {
5025 .name = "usage_in_bytes",
5026 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5027 .read_u64 = mem_cgroup_read_u64,
5028 },
5029 {
5030 .name = "max_usage_in_bytes",
5031 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5032 .write = mem_cgroup_reset,
5033 .read_u64 = mem_cgroup_read_u64,
5034 },
5035 {
5036 .name = "limit_in_bytes",
5037 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5038 .write = mem_cgroup_write,
5039 .read_u64 = mem_cgroup_read_u64,
5040 },
5041 {
5042 .name = "soft_limit_in_bytes",
5043 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5044 .write = mem_cgroup_write,
5045 .read_u64 = mem_cgroup_read_u64,
5046 },
5047 {
5048 .name = "failcnt",
5049 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5050 .write = mem_cgroup_reset,
5051 .read_u64 = mem_cgroup_read_u64,
5052 },
5053 {
5054 .name = "stat",
5055 .seq_show = memory_stat_show,
5056 },
5057 {
5058 .name = "force_empty",
5059 .write = mem_cgroup_force_empty_write,
5060 },
5061 {
5062 .name = "use_hierarchy",
5063 .write_u64 = mem_cgroup_hierarchy_write,
5064 .read_u64 = mem_cgroup_hierarchy_read,
5065 },
5066 {
5067 .name = "cgroup.event_control", /* XXX: for compat */
5068 .write = memcg_write_event_control,
5069 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5070 },
5071 {
5072 .name = "swappiness",
5073 .read_u64 = mem_cgroup_swappiness_read,
5074 .write_u64 = mem_cgroup_swappiness_write,
5075 },
5076 {
5077 .name = "move_charge_at_immigrate",
5078 .read_u64 = mem_cgroup_move_charge_read,
5079 .write_u64 = mem_cgroup_move_charge_write,
5080 },
5081 {
5082 .name = "oom_control",
5083 .seq_show = mem_cgroup_oom_control_read,
5084 .write_u64 = mem_cgroup_oom_control_write,
5085 },
5086 {
5087 .name = "pressure_level",
5088 .seq_show = mem_cgroup_dummy_seq_show,
5089 },
5090 #ifdef CONFIG_NUMA
5091 {
5092 .name = "numa_stat",
5093 .seq_show = memcg_numa_stat_show,
5094 },
5095 #endif
5096 {
5097 .name = "kmem.limit_in_bytes",
5098 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5099 .write = mem_cgroup_write,
5100 .read_u64 = mem_cgroup_read_u64,
5101 },
5102 {
5103 .name = "kmem.usage_in_bytes",
5104 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5105 .read_u64 = mem_cgroup_read_u64,
5106 },
5107 {
5108 .name = "kmem.failcnt",
5109 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5110 .write = mem_cgroup_reset,
5111 .read_u64 = mem_cgroup_read_u64,
5112 },
5113 {
5114 .name = "kmem.max_usage_in_bytes",
5115 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5116 .write = mem_cgroup_reset,
5117 .read_u64 = mem_cgroup_read_u64,
5118 },
5119 #if defined(CONFIG_MEMCG_KMEM) && \
5120 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5121 {
5122 .name = "kmem.slabinfo",
5123 .seq_show = mem_cgroup_slab_show,
5124 },
5125 #endif
5126 {
5127 .name = "kmem.tcp.limit_in_bytes",
5128 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5129 .write = mem_cgroup_write,
5130 .read_u64 = mem_cgroup_read_u64,
5131 },
5132 {
5133 .name = "kmem.tcp.usage_in_bytes",
5134 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5135 .read_u64 = mem_cgroup_read_u64,
5136 },
5137 {
5138 .name = "kmem.tcp.failcnt",
5139 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5140 .write = mem_cgroup_reset,
5141 .read_u64 = mem_cgroup_read_u64,
5142 },
5143 {
5144 .name = "kmem.tcp.max_usage_in_bytes",
5145 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5146 .write = mem_cgroup_reset,
5147 .read_u64 = mem_cgroup_read_u64,
5148 },
5149 { }, /* terminate */
5150 };
5151
5152 /*
5153 * Private memory cgroup IDR
5154 *
5155 * Swap-out records and page cache shadow entries need to store memcg
5156 * references in constrained space, so we maintain an ID space that is
5157 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5158 * memory-controlled cgroups to 64k.
5159 *
5160 * However, there usually are many references to the offline CSS after
5161 * the cgroup has been destroyed, such as page cache or reclaimable
5162 * slab objects, that don't need to hang on to the ID. We want to keep
5163 * those dead CSS from occupying IDs, or we might quickly exhaust the
5164 * relatively small ID space and prevent the creation of new cgroups
5165 * even when there are much fewer than 64k cgroups - possibly none.
5166 *
5167 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5168 * be freed and recycled when it's no longer needed, which is usually
5169 * when the CSS is offlined.
5170 *
5171 * The only exception to that are records of swapped out tmpfs/shmem
5172 * pages that need to be attributed to live ancestors on swapin. But
5173 * those references are manageable from userspace.
5174 */
5175
5176 #define MEM_CGROUP_ID_MAX ((1UL << MEM_CGROUP_ID_SHIFT) - 1)
5177 static DEFINE_IDR(mem_cgroup_idr);
5178 static DEFINE_SPINLOCK(memcg_idr_lock);
5179
mem_cgroup_alloc_id(void)5180 static int mem_cgroup_alloc_id(void)
5181 {
5182 int ret;
5183
5184 idr_preload(GFP_KERNEL);
5185 spin_lock(&memcg_idr_lock);
5186 ret = idr_alloc(&mem_cgroup_idr, NULL, 1, MEM_CGROUP_ID_MAX + 1,
5187 GFP_NOWAIT);
5188 spin_unlock(&memcg_idr_lock);
5189 idr_preload_end();
5190 return ret;
5191 }
5192
mem_cgroup_id_remove(struct mem_cgroup * memcg)5193 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5194 {
5195 if (memcg->id.id > 0) {
5196 spin_lock(&memcg_idr_lock);
5197 idr_remove(&mem_cgroup_idr, memcg->id.id);
5198 spin_unlock(&memcg_idr_lock);
5199
5200 memcg->id.id = 0;
5201 }
5202 }
5203
mem_cgroup_id_get_many(struct mem_cgroup * memcg,unsigned int n)5204 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5205 unsigned int n)
5206 {
5207 refcount_add(n, &memcg->id.ref);
5208 }
5209
mem_cgroup_id_put_many(struct mem_cgroup * memcg,unsigned int n)5210 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5211 {
5212 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5213 mem_cgroup_id_remove(memcg);
5214
5215 /* Memcg ID pins CSS */
5216 css_put(&memcg->css);
5217 }
5218 }
5219
mem_cgroup_id_put(struct mem_cgroup * memcg)5220 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5221 {
5222 mem_cgroup_id_put_many(memcg, 1);
5223 }
5224
5225 /**
5226 * mem_cgroup_from_id - look up a memcg from a memcg id
5227 * @id: the memcg id to look up
5228 *
5229 * Caller must hold rcu_read_lock().
5230 */
mem_cgroup_from_id(unsigned short id)5231 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5232 {
5233 WARN_ON_ONCE(!rcu_read_lock_held());
5234 return idr_find(&mem_cgroup_idr, id);
5235 }
5236
5237 #ifdef CONFIG_SHRINKER_DEBUG
mem_cgroup_get_from_ino(unsigned long ino)5238 struct mem_cgroup *mem_cgroup_get_from_ino(unsigned long ino)
5239 {
5240 struct cgroup *cgrp;
5241 struct cgroup_subsys_state *css;
5242 struct mem_cgroup *memcg;
5243
5244 cgrp = cgroup_get_from_id(ino);
5245 if (IS_ERR(cgrp))
5246 return ERR_CAST(cgrp);
5247
5248 css = cgroup_get_e_css(cgrp, &memory_cgrp_subsys);
5249 if (css)
5250 memcg = container_of(css, struct mem_cgroup, css);
5251 else
5252 memcg = ERR_PTR(-ENOENT);
5253
5254 cgroup_put(cgrp);
5255
5256 return memcg;
5257 }
5258 #endif
5259
alloc_mem_cgroup_per_node_info(struct mem_cgroup * memcg,int node)5260 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5261 {
5262 struct mem_cgroup_per_node *pn;
5263
5264 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, node);
5265 if (!pn)
5266 return 1;
5267
5268 pn->lruvec_stats_percpu = alloc_percpu_gfp(struct lruvec_stats_percpu,
5269 GFP_KERNEL_ACCOUNT);
5270 if (!pn->lruvec_stats_percpu) {
5271 kfree(pn);
5272 return 1;
5273 }
5274
5275 lruvec_init(&pn->lruvec);
5276 pn->memcg = memcg;
5277
5278 memcg->nodeinfo[node] = pn;
5279 return 0;
5280 }
5281
free_mem_cgroup_per_node_info(struct mem_cgroup * memcg,int node)5282 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5283 {
5284 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5285
5286 if (!pn)
5287 return;
5288
5289 free_percpu(pn->lruvec_stats_percpu);
5290 kfree(pn);
5291 }
5292
__mem_cgroup_free(struct mem_cgroup * memcg)5293 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5294 {
5295 int node;
5296
5297 for_each_node(node)
5298 free_mem_cgroup_per_node_info(memcg, node);
5299 kfree(memcg->vmstats);
5300 free_percpu(memcg->vmstats_percpu);
5301 kfree(memcg);
5302 }
5303
mem_cgroup_free(struct mem_cgroup * memcg)5304 static void mem_cgroup_free(struct mem_cgroup *memcg)
5305 {
5306 lru_gen_exit_memcg(memcg);
5307 memcg_wb_domain_exit(memcg);
5308 __mem_cgroup_free(memcg);
5309 }
5310
mem_cgroup_alloc(void)5311 static struct mem_cgroup *mem_cgroup_alloc(void)
5312 {
5313 struct mem_cgroup *memcg;
5314 int node;
5315 int __maybe_unused i;
5316 long error = -ENOMEM;
5317
5318 memcg = kzalloc(struct_size(memcg, nodeinfo, nr_node_ids), GFP_KERNEL);
5319 if (!memcg)
5320 return ERR_PTR(error);
5321
5322 memcg->id.id = mem_cgroup_alloc_id();
5323 if (memcg->id.id < 0) {
5324 error = memcg->id.id;
5325 goto fail;
5326 }
5327
5328 memcg->vmstats = kzalloc(sizeof(struct memcg_vmstats), GFP_KERNEL);
5329 if (!memcg->vmstats)
5330 goto fail;
5331
5332 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5333 GFP_KERNEL_ACCOUNT);
5334 if (!memcg->vmstats_percpu)
5335 goto fail;
5336
5337 for_each_node(node)
5338 if (alloc_mem_cgroup_per_node_info(memcg, node))
5339 goto fail;
5340
5341 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5342 goto fail;
5343
5344 INIT_WORK(&memcg->high_work, high_work_func);
5345 INIT_LIST_HEAD(&memcg->oom_notify);
5346 mutex_init(&memcg->thresholds_lock);
5347 spin_lock_init(&memcg->move_lock);
5348 vmpressure_init(&memcg->vmpressure);
5349 INIT_LIST_HEAD(&memcg->event_list);
5350 spin_lock_init(&memcg->event_list_lock);
5351 memcg->socket_pressure = jiffies;
5352 #ifdef CONFIG_MEMCG_KMEM
5353 memcg->kmemcg_id = -1;
5354 INIT_LIST_HEAD(&memcg->objcg_list);
5355 #endif
5356 #ifdef CONFIG_CGROUP_WRITEBACK
5357 INIT_LIST_HEAD(&memcg->cgwb_list);
5358 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5359 memcg->cgwb_frn[i].done =
5360 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5361 #endif
5362 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5363 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5364 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5365 memcg->deferred_split_queue.split_queue_len = 0;
5366 #endif
5367 lru_gen_init_memcg(memcg);
5368 return memcg;
5369 fail:
5370 mem_cgroup_id_remove(memcg);
5371 __mem_cgroup_free(memcg);
5372 return ERR_PTR(error);
5373 }
5374
5375 static struct cgroup_subsys_state * __ref
mem_cgroup_css_alloc(struct cgroup_subsys_state * parent_css)5376 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5377 {
5378 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5379 struct mem_cgroup *memcg, *old_memcg;
5380
5381 old_memcg = set_active_memcg(parent);
5382 memcg = mem_cgroup_alloc();
5383 set_active_memcg(old_memcg);
5384 if (IS_ERR(memcg))
5385 return ERR_CAST(memcg);
5386
5387 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5388 WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5389 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
5390 memcg->zswap_max = PAGE_COUNTER_MAX;
5391 #endif
5392 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5393 if (parent) {
5394 WRITE_ONCE(memcg->swappiness, mem_cgroup_swappiness(parent));
5395 WRITE_ONCE(memcg->oom_kill_disable, READ_ONCE(parent->oom_kill_disable));
5396
5397 page_counter_init(&memcg->memory, &parent->memory);
5398 page_counter_init(&memcg->swap, &parent->swap);
5399 page_counter_init(&memcg->kmem, &parent->kmem);
5400 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5401 } else {
5402 init_memcg_events();
5403 page_counter_init(&memcg->memory, NULL);
5404 page_counter_init(&memcg->swap, NULL);
5405 page_counter_init(&memcg->kmem, NULL);
5406 page_counter_init(&memcg->tcpmem, NULL);
5407
5408 root_mem_cgroup = memcg;
5409 return &memcg->css;
5410 }
5411
5412 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5413 static_branch_inc(&memcg_sockets_enabled_key);
5414
5415 #if defined(CONFIG_MEMCG_KMEM)
5416 if (!cgroup_memory_nobpf)
5417 static_branch_inc(&memcg_bpf_enabled_key);
5418 #endif
5419
5420 return &memcg->css;
5421 }
5422
mem_cgroup_css_online(struct cgroup_subsys_state * css)5423 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5424 {
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5426
5427 if (memcg_online_kmem(memcg))
5428 goto remove_id;
5429
5430 /*
5431 * A memcg must be visible for expand_shrinker_info()
5432 * by the time the maps are allocated. So, we allocate maps
5433 * here, when for_each_mem_cgroup() can't skip it.
5434 */
5435 if (alloc_shrinker_info(memcg))
5436 goto offline_kmem;
5437
5438 if (unlikely(mem_cgroup_is_root(memcg)))
5439 queue_delayed_work(system_unbound_wq, &stats_flush_dwork,
5440 FLUSH_TIME);
5441 lru_gen_online_memcg(memcg);
5442
5443 /* Online state pins memcg ID, memcg ID pins CSS */
5444 refcount_set(&memcg->id.ref, 1);
5445 css_get(css);
5446
5447 /*
5448 * Ensure mem_cgroup_from_id() works once we're fully online.
5449 *
5450 * We could do this earlier and require callers to filter with
5451 * css_tryget_online(). But right now there are no users that
5452 * need earlier access, and the workingset code relies on the
5453 * cgroup tree linkage (mem_cgroup_get_nr_swap_pages()). So
5454 * publish it here at the end of onlining. This matches the
5455 * regular ID destruction during offlining.
5456 */
5457 spin_lock(&memcg_idr_lock);
5458 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5459 spin_unlock(&memcg_idr_lock);
5460
5461 return 0;
5462 offline_kmem:
5463 memcg_offline_kmem(memcg);
5464 remove_id:
5465 mem_cgroup_id_remove(memcg);
5466 return -ENOMEM;
5467 }
5468
mem_cgroup_css_offline(struct cgroup_subsys_state * css)5469 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5470 {
5471 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5472 struct mem_cgroup_event *event, *tmp;
5473
5474 /*
5475 * Unregister events and notify userspace.
5476 * Notify userspace about cgroup removing only after rmdir of cgroup
5477 * directory to avoid race between userspace and kernelspace.
5478 */
5479 spin_lock_irq(&memcg->event_list_lock);
5480 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5481 list_del_init(&event->list);
5482 schedule_work(&event->remove);
5483 }
5484 spin_unlock_irq(&memcg->event_list_lock);
5485
5486 page_counter_set_min(&memcg->memory, 0);
5487 page_counter_set_low(&memcg->memory, 0);
5488
5489 memcg_offline_kmem(memcg);
5490 reparent_shrinker_deferred(memcg);
5491 wb_memcg_offline(memcg);
5492 lru_gen_offline_memcg(memcg);
5493
5494 drain_all_stock(memcg);
5495
5496 mem_cgroup_id_put(memcg);
5497 }
5498
mem_cgroup_css_released(struct cgroup_subsys_state * css)5499 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5500 {
5501 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5502
5503 invalidate_reclaim_iterators(memcg);
5504 lru_gen_release_memcg(memcg);
5505 }
5506
mem_cgroup_css_free(struct cgroup_subsys_state * css)5507 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5508 {
5509 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5510 int __maybe_unused i;
5511
5512 #ifdef CONFIG_CGROUP_WRITEBACK
5513 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5514 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5515 #endif
5516 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5517 static_branch_dec(&memcg_sockets_enabled_key);
5518
5519 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5520 static_branch_dec(&memcg_sockets_enabled_key);
5521
5522 #if defined(CONFIG_MEMCG_KMEM)
5523 if (!cgroup_memory_nobpf)
5524 static_branch_dec(&memcg_bpf_enabled_key);
5525 #endif
5526
5527 vmpressure_cleanup(&memcg->vmpressure);
5528 cancel_work_sync(&memcg->high_work);
5529 mem_cgroup_remove_from_trees(memcg);
5530 free_shrinker_info(memcg);
5531 mem_cgroup_free(memcg);
5532 }
5533
5534 /**
5535 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5536 * @css: the target css
5537 *
5538 * Reset the states of the mem_cgroup associated with @css. This is
5539 * invoked when the userland requests disabling on the default hierarchy
5540 * but the memcg is pinned through dependency. The memcg should stop
5541 * applying policies and should revert to the vanilla state as it may be
5542 * made visible again.
5543 *
5544 * The current implementation only resets the essential configurations.
5545 * This needs to be expanded to cover all the visible parts.
5546 */
mem_cgroup_css_reset(struct cgroup_subsys_state * css)5547 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5548 {
5549 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5550
5551 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5552 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5553 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5554 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5555 page_counter_set_min(&memcg->memory, 0);
5556 page_counter_set_low(&memcg->memory, 0);
5557 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5558 WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
5559 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5560 memcg_wb_domain_size_changed(memcg);
5561 }
5562
mem_cgroup_css_rstat_flush(struct cgroup_subsys_state * css,int cpu)5563 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5564 {
5565 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5566 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5567 struct memcg_vmstats_percpu *statc;
5568 long delta, delta_cpu, v;
5569 int i, nid;
5570
5571 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5572
5573 for (i = 0; i < MEMCG_NR_STAT; i++) {
5574 /*
5575 * Collect the aggregated propagation counts of groups
5576 * below us. We're in a per-cpu loop here and this is
5577 * a global counter, so the first cycle will get them.
5578 */
5579 delta = memcg->vmstats->state_pending[i];
5580 if (delta)
5581 memcg->vmstats->state_pending[i] = 0;
5582
5583 /* Add CPU changes on this level since the last flush */
5584 delta_cpu = 0;
5585 v = READ_ONCE(statc->state[i]);
5586 if (v != statc->state_prev[i]) {
5587 delta_cpu = v - statc->state_prev[i];
5588 delta += delta_cpu;
5589 statc->state_prev[i] = v;
5590 }
5591
5592 /* Aggregate counts on this level and propagate upwards */
5593 if (delta_cpu)
5594 memcg->vmstats->state_local[i] += delta_cpu;
5595
5596 if (delta) {
5597 memcg->vmstats->state[i] += delta;
5598 if (parent)
5599 parent->vmstats->state_pending[i] += delta;
5600 }
5601 }
5602
5603 for (i = 0; i < NR_MEMCG_EVENTS; i++) {
5604 delta = memcg->vmstats->events_pending[i];
5605 if (delta)
5606 memcg->vmstats->events_pending[i] = 0;
5607
5608 delta_cpu = 0;
5609 v = READ_ONCE(statc->events[i]);
5610 if (v != statc->events_prev[i]) {
5611 delta_cpu = v - statc->events_prev[i];
5612 delta += delta_cpu;
5613 statc->events_prev[i] = v;
5614 }
5615
5616 if (delta_cpu)
5617 memcg->vmstats->events_local[i] += delta_cpu;
5618
5619 if (delta) {
5620 memcg->vmstats->events[i] += delta;
5621 if (parent)
5622 parent->vmstats->events_pending[i] += delta;
5623 }
5624 }
5625
5626 for_each_node_state(nid, N_MEMORY) {
5627 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
5628 struct mem_cgroup_per_node *ppn = NULL;
5629 struct lruvec_stats_percpu *lstatc;
5630
5631 if (parent)
5632 ppn = parent->nodeinfo[nid];
5633
5634 lstatc = per_cpu_ptr(pn->lruvec_stats_percpu, cpu);
5635
5636 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
5637 delta = pn->lruvec_stats.state_pending[i];
5638 if (delta)
5639 pn->lruvec_stats.state_pending[i] = 0;
5640
5641 delta_cpu = 0;
5642 v = READ_ONCE(lstatc->state[i]);
5643 if (v != lstatc->state_prev[i]) {
5644 delta_cpu = v - lstatc->state_prev[i];
5645 delta += delta_cpu;
5646 lstatc->state_prev[i] = v;
5647 }
5648
5649 if (delta_cpu)
5650 pn->lruvec_stats.state_local[i] += delta_cpu;
5651
5652 if (delta) {
5653 pn->lruvec_stats.state[i] += delta;
5654 if (ppn)
5655 ppn->lruvec_stats.state_pending[i] += delta;
5656 }
5657 }
5658 }
5659 }
5660
5661 #ifdef CONFIG_MMU
5662 /* Handlers for move charge at task migration. */
mem_cgroup_do_precharge(unsigned long count)5663 static int mem_cgroup_do_precharge(unsigned long count)
5664 {
5665 int ret;
5666
5667 /* Try a single bulk charge without reclaim first, kswapd may wake */
5668 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5669 if (!ret) {
5670 mc.precharge += count;
5671 return ret;
5672 }
5673
5674 /* Try charges one by one with reclaim, but do not retry */
5675 while (count--) {
5676 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5677 if (ret)
5678 return ret;
5679 mc.precharge++;
5680 cond_resched();
5681 }
5682 return 0;
5683 }
5684
5685 union mc_target {
5686 struct page *page;
5687 swp_entry_t ent;
5688 };
5689
5690 enum mc_target_type {
5691 MC_TARGET_NONE = 0,
5692 MC_TARGET_PAGE,
5693 MC_TARGET_SWAP,
5694 MC_TARGET_DEVICE,
5695 };
5696
mc_handle_present_pte(struct vm_area_struct * vma,unsigned long addr,pte_t ptent)5697 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5698 unsigned long addr, pte_t ptent)
5699 {
5700 struct page *page = vm_normal_page(vma, addr, ptent);
5701
5702 if (!page)
5703 return NULL;
5704 if (PageAnon(page)) {
5705 if (!(mc.flags & MOVE_ANON))
5706 return NULL;
5707 } else {
5708 if (!(mc.flags & MOVE_FILE))
5709 return NULL;
5710 }
5711 get_page(page);
5712
5713 return page;
5714 }
5715
5716 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
mc_handle_swap_pte(struct vm_area_struct * vma,pte_t ptent,swp_entry_t * entry)5717 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5718 pte_t ptent, swp_entry_t *entry)
5719 {
5720 struct page *page = NULL;
5721 swp_entry_t ent = pte_to_swp_entry(ptent);
5722
5723 if (!(mc.flags & MOVE_ANON))
5724 return NULL;
5725
5726 /*
5727 * Handle device private pages that are not accessible by the CPU, but
5728 * stored as special swap entries in the page table.
5729 */
5730 if (is_device_private_entry(ent)) {
5731 page = pfn_swap_entry_to_page(ent);
5732 if (!get_page_unless_zero(page))
5733 return NULL;
5734 return page;
5735 }
5736
5737 if (non_swap_entry(ent))
5738 return NULL;
5739
5740 /*
5741 * Because swap_cache_get_folio() updates some statistics counter,
5742 * we call find_get_page() with swapper_space directly.
5743 */
5744 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5745 entry->val = ent.val;
5746
5747 return page;
5748 }
5749 #else
mc_handle_swap_pte(struct vm_area_struct * vma,pte_t ptent,swp_entry_t * entry)5750 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5751 pte_t ptent, swp_entry_t *entry)
5752 {
5753 return NULL;
5754 }
5755 #endif
5756
mc_handle_file_pte(struct vm_area_struct * vma,unsigned long addr,pte_t ptent)5757 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5758 unsigned long addr, pte_t ptent)
5759 {
5760 unsigned long index;
5761 struct folio *folio;
5762
5763 if (!vma->vm_file) /* anonymous vma */
5764 return NULL;
5765 if (!(mc.flags & MOVE_FILE))
5766 return NULL;
5767
5768 /* folio is moved even if it's not RSS of this task(page-faulted). */
5769 /* shmem/tmpfs may report page out on swap: account for that too. */
5770 index = linear_page_index(vma, addr);
5771 folio = filemap_get_incore_folio(vma->vm_file->f_mapping, index);
5772 if (IS_ERR(folio))
5773 return NULL;
5774 return folio_file_page(folio, index);
5775 }
5776
5777 /**
5778 * mem_cgroup_move_account - move account of the page
5779 * @page: the page
5780 * @compound: charge the page as compound or small page
5781 * @from: mem_cgroup which the page is moved from.
5782 * @to: mem_cgroup which the page is moved to. @from != @to.
5783 *
5784 * The page must be locked and not on the LRU.
5785 *
5786 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5787 * from old cgroup.
5788 */
mem_cgroup_move_account(struct page * page,bool compound,struct mem_cgroup * from,struct mem_cgroup * to)5789 static int mem_cgroup_move_account(struct page *page,
5790 bool compound,
5791 struct mem_cgroup *from,
5792 struct mem_cgroup *to)
5793 {
5794 struct folio *folio = page_folio(page);
5795 struct lruvec *from_vec, *to_vec;
5796 struct pglist_data *pgdat;
5797 unsigned int nr_pages = compound ? folio_nr_pages(folio) : 1;
5798 int nid, ret;
5799
5800 VM_BUG_ON(from == to);
5801 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
5802 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
5803 VM_BUG_ON(compound && !folio_test_large(folio));
5804
5805 ret = -EINVAL;
5806 if (folio_memcg(folio) != from)
5807 goto out;
5808
5809 pgdat = folio_pgdat(folio);
5810 from_vec = mem_cgroup_lruvec(from, pgdat);
5811 to_vec = mem_cgroup_lruvec(to, pgdat);
5812
5813 folio_memcg_lock(folio);
5814
5815 if (folio_test_anon(folio)) {
5816 if (folio_mapped(folio)) {
5817 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5818 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5819 if (folio_test_pmd_mappable(folio)) {
5820 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5821 -nr_pages);
5822 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5823 nr_pages);
5824 }
5825 }
5826 } else {
5827 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5828 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5829
5830 if (folio_test_swapbacked(folio)) {
5831 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5832 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5833 }
5834
5835 if (folio_mapped(folio)) {
5836 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5837 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5838 }
5839
5840 if (folio_test_dirty(folio)) {
5841 struct address_space *mapping = folio_mapping(folio);
5842
5843 if (mapping_can_writeback(mapping)) {
5844 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5845 -nr_pages);
5846 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5847 nr_pages);
5848 }
5849 }
5850 }
5851
5852 #ifdef CONFIG_SWAP
5853 if (folio_test_swapcache(folio)) {
5854 __mod_lruvec_state(from_vec, NR_SWAPCACHE, -nr_pages);
5855 __mod_lruvec_state(to_vec, NR_SWAPCACHE, nr_pages);
5856 }
5857 #endif
5858 if (folio_test_writeback(folio)) {
5859 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5860 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5861 }
5862
5863 /*
5864 * All state has been migrated, let's switch to the new memcg.
5865 *
5866 * It is safe to change page's memcg here because the page
5867 * is referenced, charged, isolated, and locked: we can't race
5868 * with (un)charging, migration, LRU putback, or anything else
5869 * that would rely on a stable page's memory cgroup.
5870 *
5871 * Note that folio_memcg_lock is a memcg lock, not a page lock,
5872 * to save space. As soon as we switch page's memory cgroup to a
5873 * new memcg that isn't locked, the above state can change
5874 * concurrently again. Make sure we're truly done with it.
5875 */
5876 smp_mb();
5877
5878 css_get(&to->css);
5879 css_put(&from->css);
5880
5881 /* Warning should never happen, so don't worry about refcount non-0 */
5882 WARN_ON_ONCE(folio_unqueue_deferred_split(folio));
5883 folio->memcg_data = (unsigned long)to;
5884
5885 __folio_memcg_unlock(from);
5886
5887 ret = 0;
5888 nid = folio_nid(folio);
5889
5890 local_irq_disable();
5891 mem_cgroup_charge_statistics(to, nr_pages);
5892 memcg_check_events(to, nid);
5893 mem_cgroup_charge_statistics(from, -nr_pages);
5894 memcg_check_events(from, nid);
5895 local_irq_enable();
5896 out:
5897 return ret;
5898 }
5899
5900 /**
5901 * get_mctgt_type - get target type of moving charge
5902 * @vma: the vma the pte to be checked belongs
5903 * @addr: the address corresponding to the pte to be checked
5904 * @ptent: the pte to be checked
5905 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5906 *
5907 * Context: Called with pte lock held.
5908 * Return:
5909 * * MC_TARGET_NONE - If the pte is not a target for move charge.
5910 * * MC_TARGET_PAGE - If the page corresponding to this pte is a target for
5911 * move charge. If @target is not NULL, the page is stored in target->page
5912 * with extra refcnt taken (Caller should release it).
5913 * * MC_TARGET_SWAP - If the swap entry corresponding to this pte is a
5914 * target for charge migration. If @target is not NULL, the entry is
5915 * stored in target->ent.
5916 * * MC_TARGET_DEVICE - Like MC_TARGET_PAGE but page is device memory and
5917 * thus not on the lru. For now such page is charged like a regular page
5918 * would be as it is just special memory taking the place of a regular page.
5919 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5920 */
get_mctgt_type(struct vm_area_struct * vma,unsigned long addr,pte_t ptent,union mc_target * target)5921 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5922 unsigned long addr, pte_t ptent, union mc_target *target)
5923 {
5924 struct page *page = NULL;
5925 enum mc_target_type ret = MC_TARGET_NONE;
5926 swp_entry_t ent = { .val = 0 };
5927
5928 if (pte_present(ptent))
5929 page = mc_handle_present_pte(vma, addr, ptent);
5930 else if (pte_none_mostly(ptent))
5931 /*
5932 * PTE markers should be treated as a none pte here, separated
5933 * from other swap handling below.
5934 */
5935 page = mc_handle_file_pte(vma, addr, ptent);
5936 else if (is_swap_pte(ptent))
5937 page = mc_handle_swap_pte(vma, ptent, &ent);
5938
5939 if (target && page) {
5940 if (!trylock_page(page)) {
5941 put_page(page);
5942 return ret;
5943 }
5944 /*
5945 * page_mapped() must be stable during the move. This
5946 * pte is locked, so if it's present, the page cannot
5947 * become unmapped. If it isn't, we have only partial
5948 * control over the mapped state: the page lock will
5949 * prevent new faults against pagecache and swapcache,
5950 * so an unmapped page cannot become mapped. However,
5951 * if the page is already mapped elsewhere, it can
5952 * unmap, and there is nothing we can do about it.
5953 * Alas, skip moving the page in this case.
5954 */
5955 if (!pte_present(ptent) && page_mapped(page)) {
5956 unlock_page(page);
5957 put_page(page);
5958 return ret;
5959 }
5960 }
5961
5962 if (!page && !ent.val)
5963 return ret;
5964 if (page) {
5965 /*
5966 * Do only loose check w/o serialization.
5967 * mem_cgroup_move_account() checks the page is valid or
5968 * not under LRU exclusion.
5969 */
5970 if (page_memcg(page) == mc.from) {
5971 ret = MC_TARGET_PAGE;
5972 if (is_device_private_page(page) ||
5973 is_device_coherent_page(page))
5974 ret = MC_TARGET_DEVICE;
5975 if (target)
5976 target->page = page;
5977 }
5978 if (!ret || !target) {
5979 if (target)
5980 unlock_page(page);
5981 put_page(page);
5982 }
5983 }
5984 /*
5985 * There is a swap entry and a page doesn't exist or isn't charged.
5986 * But we cannot move a tail-page in a THP.
5987 */
5988 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5989 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5990 ret = MC_TARGET_SWAP;
5991 if (target)
5992 target->ent = ent;
5993 }
5994 return ret;
5995 }
5996
5997 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5998 /*
5999 * We don't consider PMD mapped swapping or file mapped pages because THP does
6000 * not support them for now.
6001 * Caller should make sure that pmd_trans_huge(pmd) is true.
6002 */
get_mctgt_type_thp(struct vm_area_struct * vma,unsigned long addr,pmd_t pmd,union mc_target * target)6003 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6004 unsigned long addr, pmd_t pmd, union mc_target *target)
6005 {
6006 struct page *page = NULL;
6007 enum mc_target_type ret = MC_TARGET_NONE;
6008
6009 if (unlikely(is_swap_pmd(pmd))) {
6010 VM_BUG_ON(thp_migration_supported() &&
6011 !is_pmd_migration_entry(pmd));
6012 return ret;
6013 }
6014 page = pmd_page(pmd);
6015 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6016 if (!(mc.flags & MOVE_ANON))
6017 return ret;
6018 if (page_memcg(page) == mc.from) {
6019 ret = MC_TARGET_PAGE;
6020 if (target) {
6021 get_page(page);
6022 if (!trylock_page(page)) {
6023 put_page(page);
6024 return MC_TARGET_NONE;
6025 }
6026 target->page = page;
6027 }
6028 }
6029 return ret;
6030 }
6031 #else
get_mctgt_type_thp(struct vm_area_struct * vma,unsigned long addr,pmd_t pmd,union mc_target * target)6032 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6033 unsigned long addr, pmd_t pmd, union mc_target *target)
6034 {
6035 return MC_TARGET_NONE;
6036 }
6037 #endif
6038
mem_cgroup_count_precharge_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)6039 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6040 unsigned long addr, unsigned long end,
6041 struct mm_walk *walk)
6042 {
6043 struct vm_area_struct *vma = walk->vma;
6044 pte_t *pte;
6045 spinlock_t *ptl;
6046
6047 ptl = pmd_trans_huge_lock(pmd, vma);
6048 if (ptl) {
6049 /*
6050 * Note their can not be MC_TARGET_DEVICE for now as we do not
6051 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
6052 * this might change.
6053 */
6054 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6055 mc.precharge += HPAGE_PMD_NR;
6056 spin_unlock(ptl);
6057 return 0;
6058 }
6059
6060 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6061 if (!pte)
6062 return 0;
6063 for (; addr != end; pte++, addr += PAGE_SIZE)
6064 if (get_mctgt_type(vma, addr, ptep_get(pte), NULL))
6065 mc.precharge++; /* increment precharge temporarily */
6066 pte_unmap_unlock(pte - 1, ptl);
6067 cond_resched();
6068
6069 return 0;
6070 }
6071
6072 static const struct mm_walk_ops precharge_walk_ops = {
6073 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6074 .walk_lock = PGWALK_RDLOCK,
6075 };
6076
mem_cgroup_count_precharge(struct mm_struct * mm)6077 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6078 {
6079 unsigned long precharge;
6080
6081 mmap_read_lock(mm);
6082 walk_page_range(mm, 0, ULONG_MAX, &precharge_walk_ops, NULL);
6083 mmap_read_unlock(mm);
6084
6085 precharge = mc.precharge;
6086 mc.precharge = 0;
6087
6088 return precharge;
6089 }
6090
mem_cgroup_precharge_mc(struct mm_struct * mm)6091 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6092 {
6093 unsigned long precharge = mem_cgroup_count_precharge(mm);
6094
6095 VM_BUG_ON(mc.moving_task);
6096 mc.moving_task = current;
6097 return mem_cgroup_do_precharge(precharge);
6098 }
6099
6100 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
__mem_cgroup_clear_mc(void)6101 static void __mem_cgroup_clear_mc(void)
6102 {
6103 struct mem_cgroup *from = mc.from;
6104 struct mem_cgroup *to = mc.to;
6105
6106 /* we must uncharge all the leftover precharges from mc.to */
6107 if (mc.precharge) {
6108 cancel_charge(mc.to, mc.precharge);
6109 mc.precharge = 0;
6110 }
6111 /*
6112 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6113 * we must uncharge here.
6114 */
6115 if (mc.moved_charge) {
6116 cancel_charge(mc.from, mc.moved_charge);
6117 mc.moved_charge = 0;
6118 }
6119 /* we must fixup refcnts and charges */
6120 if (mc.moved_swap) {
6121 /* uncharge swap account from the old cgroup */
6122 if (!mem_cgroup_is_root(mc.from))
6123 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
6124
6125 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
6126
6127 /*
6128 * we charged both to->memory and to->memsw, so we
6129 * should uncharge to->memory.
6130 */
6131 if (!mem_cgroup_is_root(mc.to))
6132 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
6133
6134 mc.moved_swap = 0;
6135 }
6136 memcg_oom_recover(from);
6137 memcg_oom_recover(to);
6138 wake_up_all(&mc.waitq);
6139 }
6140
mem_cgroup_clear_mc(void)6141 static void mem_cgroup_clear_mc(void)
6142 {
6143 struct mm_struct *mm = mc.mm;
6144
6145 /*
6146 * we must clear moving_task before waking up waiters at the end of
6147 * task migration.
6148 */
6149 mc.moving_task = NULL;
6150 __mem_cgroup_clear_mc();
6151 spin_lock(&mc.lock);
6152 mc.from = NULL;
6153 mc.to = NULL;
6154 mc.mm = NULL;
6155 spin_unlock(&mc.lock);
6156
6157 mmput(mm);
6158 }
6159
mem_cgroup_can_attach(struct cgroup_taskset * tset)6160 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6161 {
6162 struct cgroup_subsys_state *css;
6163 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
6164 struct mem_cgroup *from;
6165 struct task_struct *leader, *p;
6166 struct mm_struct *mm;
6167 unsigned long move_flags;
6168 int ret = 0;
6169
6170 /* charge immigration isn't supported on the default hierarchy */
6171 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6172 return 0;
6173
6174 /*
6175 * Multi-process migrations only happen on the default hierarchy
6176 * where charge immigration is not used. Perform charge
6177 * immigration if @tset contains a leader and whine if there are
6178 * multiple.
6179 */
6180 p = NULL;
6181 cgroup_taskset_for_each_leader(leader, css, tset) {
6182 WARN_ON_ONCE(p);
6183 p = leader;
6184 memcg = mem_cgroup_from_css(css);
6185 }
6186 if (!p)
6187 return 0;
6188
6189 /*
6190 * We are now committed to this value whatever it is. Changes in this
6191 * tunable will only affect upcoming migrations, not the current one.
6192 * So we need to save it, and keep it going.
6193 */
6194 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
6195 if (!move_flags)
6196 return 0;
6197
6198 from = mem_cgroup_from_task(p);
6199
6200 VM_BUG_ON(from == memcg);
6201
6202 mm = get_task_mm(p);
6203 if (!mm)
6204 return 0;
6205 /* We move charges only when we move a owner of the mm */
6206 if (mm->owner == p) {
6207 VM_BUG_ON(mc.from);
6208 VM_BUG_ON(mc.to);
6209 VM_BUG_ON(mc.precharge);
6210 VM_BUG_ON(mc.moved_charge);
6211 VM_BUG_ON(mc.moved_swap);
6212
6213 spin_lock(&mc.lock);
6214 mc.mm = mm;
6215 mc.from = from;
6216 mc.to = memcg;
6217 mc.flags = move_flags;
6218 spin_unlock(&mc.lock);
6219 /* We set mc.moving_task later */
6220
6221 ret = mem_cgroup_precharge_mc(mm);
6222 if (ret)
6223 mem_cgroup_clear_mc();
6224 } else {
6225 mmput(mm);
6226 }
6227 return ret;
6228 }
6229
mem_cgroup_cancel_attach(struct cgroup_taskset * tset)6230 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6231 {
6232 if (mc.to)
6233 mem_cgroup_clear_mc();
6234 }
6235
mem_cgroup_move_charge_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)6236 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6237 unsigned long addr, unsigned long end,
6238 struct mm_walk *walk)
6239 {
6240 int ret = 0;
6241 struct vm_area_struct *vma = walk->vma;
6242 pte_t *pte;
6243 spinlock_t *ptl;
6244 enum mc_target_type target_type;
6245 union mc_target target;
6246 struct page *page;
6247 struct folio *folio;
6248 bool tried_split_before = false;
6249
6250 retry_pmd:
6251 ptl = pmd_trans_huge_lock(pmd, vma);
6252 if (ptl) {
6253 if (mc.precharge < HPAGE_PMD_NR) {
6254 spin_unlock(ptl);
6255 return 0;
6256 }
6257 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6258 if (target_type == MC_TARGET_PAGE) {
6259 page = target.page;
6260 folio = page_folio(page);
6261 /*
6262 * Deferred split queue locking depends on memcg,
6263 * and unqueue is unsafe unless folio refcount is 0:
6264 * split or skip if on the queue? first try to split.
6265 */
6266 if (!list_empty(&folio->_deferred_list)) {
6267 spin_unlock(ptl);
6268 if (!tried_split_before)
6269 split_folio(folio);
6270 folio_unlock(folio);
6271 folio_put(folio);
6272 if (tried_split_before)
6273 return 0;
6274 tried_split_before = true;
6275 goto retry_pmd;
6276 }
6277 /*
6278 * So long as that pmd lock is held, the folio cannot
6279 * be racily added to the _deferred_list, because
6280 * page_remove_rmap() will find it still pmdmapped.
6281 */
6282 if (isolate_lru_page(page)) {
6283 if (!mem_cgroup_move_account(page, true,
6284 mc.from, mc.to)) {
6285 mc.precharge -= HPAGE_PMD_NR;
6286 mc.moved_charge += HPAGE_PMD_NR;
6287 }
6288 putback_lru_page(page);
6289 }
6290 unlock_page(page);
6291 put_page(page);
6292 } else if (target_type == MC_TARGET_DEVICE) {
6293 page = target.page;
6294 if (!mem_cgroup_move_account(page, true,
6295 mc.from, mc.to)) {
6296 mc.precharge -= HPAGE_PMD_NR;
6297 mc.moved_charge += HPAGE_PMD_NR;
6298 }
6299 unlock_page(page);
6300 put_page(page);
6301 }
6302 spin_unlock(ptl);
6303 return 0;
6304 }
6305
6306 retry:
6307 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6308 if (!pte)
6309 return 0;
6310 for (; addr != end; addr += PAGE_SIZE) {
6311 pte_t ptent = ptep_get(pte++);
6312 bool device = false;
6313 swp_entry_t ent;
6314
6315 if (!mc.precharge)
6316 break;
6317
6318 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6319 case MC_TARGET_DEVICE:
6320 device = true;
6321 fallthrough;
6322 case MC_TARGET_PAGE:
6323 page = target.page;
6324 /*
6325 * We can have a part of the split pmd here. Moving it
6326 * can be done but it would be too convoluted so simply
6327 * ignore such a partial THP and keep it in original
6328 * memcg. There should be somebody mapping the head.
6329 */
6330 if (PageTransCompound(page))
6331 goto put;
6332 if (!device && !isolate_lru_page(page))
6333 goto put;
6334 if (!mem_cgroup_move_account(page, false,
6335 mc.from, mc.to)) {
6336 mc.precharge--;
6337 /* we uncharge from mc.from later. */
6338 mc.moved_charge++;
6339 }
6340 if (!device)
6341 putback_lru_page(page);
6342 put: /* get_mctgt_type() gets & locks the page */
6343 unlock_page(page);
6344 put_page(page);
6345 break;
6346 case MC_TARGET_SWAP:
6347 ent = target.ent;
6348 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6349 mc.precharge--;
6350 mem_cgroup_id_get_many(mc.to, 1);
6351 /* we fixup other refcnts and charges later. */
6352 mc.moved_swap++;
6353 }
6354 break;
6355 default:
6356 break;
6357 }
6358 }
6359 pte_unmap_unlock(pte - 1, ptl);
6360 cond_resched();
6361
6362 if (addr != end) {
6363 /*
6364 * We have consumed all precharges we got in can_attach().
6365 * We try charge one by one, but don't do any additional
6366 * charges to mc.to if we have failed in charge once in attach()
6367 * phase.
6368 */
6369 ret = mem_cgroup_do_precharge(1);
6370 if (!ret)
6371 goto retry;
6372 }
6373
6374 return ret;
6375 }
6376
6377 static const struct mm_walk_ops charge_walk_ops = {
6378 .pmd_entry = mem_cgroup_move_charge_pte_range,
6379 .walk_lock = PGWALK_RDLOCK,
6380 };
6381
mem_cgroup_move_charge(void)6382 static void mem_cgroup_move_charge(void)
6383 {
6384 lru_add_drain_all();
6385 /*
6386 * Signal folio_memcg_lock() to take the memcg's move_lock
6387 * while we're moving its pages to another memcg. Then wait
6388 * for already started RCU-only updates to finish.
6389 */
6390 atomic_inc(&mc.from->moving_account);
6391 synchronize_rcu();
6392 retry:
6393 if (unlikely(!mmap_read_trylock(mc.mm))) {
6394 /*
6395 * Someone who are holding the mmap_lock might be waiting in
6396 * waitq. So we cancel all extra charges, wake up all waiters,
6397 * and retry. Because we cancel precharges, we might not be able
6398 * to move enough charges, but moving charge is a best-effort
6399 * feature anyway, so it wouldn't be a big problem.
6400 */
6401 __mem_cgroup_clear_mc();
6402 cond_resched();
6403 goto retry;
6404 }
6405 /*
6406 * When we have consumed all precharges and failed in doing
6407 * additional charge, the page walk just aborts.
6408 */
6409 walk_page_range(mc.mm, 0, ULONG_MAX, &charge_walk_ops, NULL);
6410 mmap_read_unlock(mc.mm);
6411 atomic_dec(&mc.from->moving_account);
6412 }
6413
mem_cgroup_move_task(void)6414 static void mem_cgroup_move_task(void)
6415 {
6416 if (mc.to) {
6417 mem_cgroup_move_charge();
6418 mem_cgroup_clear_mc();
6419 }
6420 }
6421 #else /* !CONFIG_MMU */
mem_cgroup_can_attach(struct cgroup_taskset * tset)6422 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6423 {
6424 return 0;
6425 }
mem_cgroup_cancel_attach(struct cgroup_taskset * tset)6426 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6427 {
6428 }
mem_cgroup_move_task(void)6429 static void mem_cgroup_move_task(void)
6430 {
6431 }
6432 #endif
6433
6434 #ifdef CONFIG_LRU_GEN
mem_cgroup_attach(struct cgroup_taskset * tset)6435 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6436 {
6437 struct task_struct *task;
6438 struct cgroup_subsys_state *css;
6439
6440 /* find the first leader if there is any */
6441 cgroup_taskset_for_each_leader(task, css, tset)
6442 break;
6443
6444 if (!task)
6445 return;
6446
6447 task_lock(task);
6448 if (task->mm && READ_ONCE(task->mm->owner) == task)
6449 lru_gen_migrate_mm(task->mm);
6450 task_unlock(task);
6451 }
6452 #else
mem_cgroup_attach(struct cgroup_taskset * tset)6453 static void mem_cgroup_attach(struct cgroup_taskset *tset)
6454 {
6455 }
6456 #endif /* CONFIG_LRU_GEN */
6457
seq_puts_memcg_tunable(struct seq_file * m,unsigned long value)6458 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6459 {
6460 if (value == PAGE_COUNTER_MAX)
6461 seq_puts(m, "max\n");
6462 else
6463 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6464
6465 return 0;
6466 }
6467
memory_current_read(struct cgroup_subsys_state * css,struct cftype * cft)6468 static u64 memory_current_read(struct cgroup_subsys_state *css,
6469 struct cftype *cft)
6470 {
6471 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6472
6473 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6474 }
6475
memory_peak_read(struct cgroup_subsys_state * css,struct cftype * cft)6476 static u64 memory_peak_read(struct cgroup_subsys_state *css,
6477 struct cftype *cft)
6478 {
6479 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6480
6481 return (u64)memcg->memory.watermark * PAGE_SIZE;
6482 }
6483
memory_min_show(struct seq_file * m,void * v)6484 static int memory_min_show(struct seq_file *m, void *v)
6485 {
6486 return seq_puts_memcg_tunable(m,
6487 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6488 }
6489
memory_min_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6490 static ssize_t memory_min_write(struct kernfs_open_file *of,
6491 char *buf, size_t nbytes, loff_t off)
6492 {
6493 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6494 unsigned long min;
6495 int err;
6496
6497 buf = strstrip(buf);
6498 err = page_counter_memparse(buf, "max", &min);
6499 if (err)
6500 return err;
6501
6502 page_counter_set_min(&memcg->memory, min);
6503
6504 return nbytes;
6505 }
6506
memory_low_show(struct seq_file * m,void * v)6507 static int memory_low_show(struct seq_file *m, void *v)
6508 {
6509 return seq_puts_memcg_tunable(m,
6510 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6511 }
6512
memory_low_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6513 static ssize_t memory_low_write(struct kernfs_open_file *of,
6514 char *buf, size_t nbytes, loff_t off)
6515 {
6516 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6517 unsigned long low;
6518 int err;
6519
6520 buf = strstrip(buf);
6521 err = page_counter_memparse(buf, "max", &low);
6522 if (err)
6523 return err;
6524
6525 page_counter_set_low(&memcg->memory, low);
6526
6527 return nbytes;
6528 }
6529
memory_high_show(struct seq_file * m,void * v)6530 static int memory_high_show(struct seq_file *m, void *v)
6531 {
6532 return seq_puts_memcg_tunable(m,
6533 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6534 }
6535
memory_high_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6536 static ssize_t memory_high_write(struct kernfs_open_file *of,
6537 char *buf, size_t nbytes, loff_t off)
6538 {
6539 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6540 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6541 bool drained = false;
6542 unsigned long high;
6543 int err;
6544
6545 buf = strstrip(buf);
6546 err = page_counter_memparse(buf, "max", &high);
6547 if (err)
6548 return err;
6549
6550 page_counter_set_high(&memcg->memory, high);
6551
6552 for (;;) {
6553 unsigned long nr_pages = page_counter_read(&memcg->memory);
6554 unsigned long reclaimed;
6555
6556 if (nr_pages <= high)
6557 break;
6558
6559 if (signal_pending(current))
6560 break;
6561
6562 if (!drained) {
6563 drain_all_stock(memcg);
6564 drained = true;
6565 continue;
6566 }
6567
6568 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6569 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP);
6570
6571 if (!reclaimed && !nr_retries--)
6572 break;
6573 }
6574
6575 memcg_wb_domain_size_changed(memcg);
6576 return nbytes;
6577 }
6578
memory_max_show(struct seq_file * m,void * v)6579 static int memory_max_show(struct seq_file *m, void *v)
6580 {
6581 return seq_puts_memcg_tunable(m,
6582 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6583 }
6584
memory_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6585 static ssize_t memory_max_write(struct kernfs_open_file *of,
6586 char *buf, size_t nbytes, loff_t off)
6587 {
6588 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6589 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6590 bool drained = false;
6591 unsigned long max;
6592 int err;
6593
6594 buf = strstrip(buf);
6595 err = page_counter_memparse(buf, "max", &max);
6596 if (err)
6597 return err;
6598
6599 xchg(&memcg->memory.max, max);
6600
6601 for (;;) {
6602 unsigned long nr_pages = page_counter_read(&memcg->memory);
6603
6604 if (nr_pages <= max)
6605 break;
6606
6607 if (signal_pending(current))
6608 break;
6609
6610 if (!drained) {
6611 drain_all_stock(memcg);
6612 drained = true;
6613 continue;
6614 }
6615
6616 if (nr_reclaims) {
6617 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6618 GFP_KERNEL, MEMCG_RECLAIM_MAY_SWAP))
6619 nr_reclaims--;
6620 continue;
6621 }
6622
6623 memcg_memory_event(memcg, MEMCG_OOM);
6624 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6625 break;
6626 }
6627
6628 memcg_wb_domain_size_changed(memcg);
6629 return nbytes;
6630 }
6631
__memory_events_show(struct seq_file * m,atomic_long_t * events)6632 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6633 {
6634 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6635 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6636 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6637 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6638 seq_printf(m, "oom_kill %lu\n",
6639 atomic_long_read(&events[MEMCG_OOM_KILL]));
6640 seq_printf(m, "oom_group_kill %lu\n",
6641 atomic_long_read(&events[MEMCG_OOM_GROUP_KILL]));
6642 }
6643
memory_events_show(struct seq_file * m,void * v)6644 static int memory_events_show(struct seq_file *m, void *v)
6645 {
6646 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6647
6648 __memory_events_show(m, memcg->memory_events);
6649 return 0;
6650 }
6651
memory_events_local_show(struct seq_file * m,void * v)6652 static int memory_events_local_show(struct seq_file *m, void *v)
6653 {
6654 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6655
6656 __memory_events_show(m, memcg->memory_events_local);
6657 return 0;
6658 }
6659
memory_stat_show(struct seq_file * m,void * v)6660 static int memory_stat_show(struct seq_file *m, void *v)
6661 {
6662 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6663 char *buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
6664 struct seq_buf s;
6665
6666 if (!buf)
6667 return -ENOMEM;
6668 seq_buf_init(&s, buf, PAGE_SIZE);
6669 memory_stat_format(memcg, &s);
6670 seq_puts(m, buf);
6671 kfree(buf);
6672 return 0;
6673 }
6674
6675 #ifdef CONFIG_NUMA
lruvec_page_state_output(struct lruvec * lruvec,int item)6676 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6677 int item)
6678 {
6679 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6680 }
6681
memory_numa_stat_show(struct seq_file * m,void * v)6682 static int memory_numa_stat_show(struct seq_file *m, void *v)
6683 {
6684 int i;
6685 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6686
6687 mem_cgroup_flush_stats();
6688
6689 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6690 int nid;
6691
6692 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6693 continue;
6694
6695 seq_printf(m, "%s", memory_stats[i].name);
6696 for_each_node_state(nid, N_MEMORY) {
6697 u64 size;
6698 struct lruvec *lruvec;
6699
6700 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6701 size = lruvec_page_state_output(lruvec,
6702 memory_stats[i].idx);
6703 seq_printf(m, " N%d=%llu", nid, size);
6704 }
6705 seq_putc(m, '\n');
6706 }
6707
6708 return 0;
6709 }
6710 #endif
6711
memory_oom_group_show(struct seq_file * m,void * v)6712 static int memory_oom_group_show(struct seq_file *m, void *v)
6713 {
6714 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6715
6716 seq_printf(m, "%d\n", READ_ONCE(memcg->oom_group));
6717
6718 return 0;
6719 }
6720
memory_oom_group_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6721 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6722 char *buf, size_t nbytes, loff_t off)
6723 {
6724 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6725 int ret, oom_group;
6726
6727 buf = strstrip(buf);
6728 if (!buf)
6729 return -EINVAL;
6730
6731 ret = kstrtoint(buf, 0, &oom_group);
6732 if (ret)
6733 return ret;
6734
6735 if (oom_group != 0 && oom_group != 1)
6736 return -EINVAL;
6737
6738 WRITE_ONCE(memcg->oom_group, oom_group);
6739
6740 return nbytes;
6741 }
6742
memory_reclaim(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)6743 static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
6744 size_t nbytes, loff_t off)
6745 {
6746 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6747 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6748 unsigned long nr_to_reclaim, nr_reclaimed = 0;
6749 unsigned int reclaim_options;
6750 int err;
6751
6752 buf = strstrip(buf);
6753 err = page_counter_memparse(buf, "", &nr_to_reclaim);
6754 if (err)
6755 return err;
6756
6757 reclaim_options = MEMCG_RECLAIM_MAY_SWAP | MEMCG_RECLAIM_PROACTIVE;
6758 while (nr_reclaimed < nr_to_reclaim) {
6759 unsigned long reclaimed;
6760
6761 if (signal_pending(current))
6762 return -EINTR;
6763
6764 /*
6765 * This is the final attempt, drain percpu lru caches in the
6766 * hope of introducing more evictable pages for
6767 * try_to_free_mem_cgroup_pages().
6768 */
6769 if (!nr_retries)
6770 lru_add_drain_all();
6771
6772 reclaimed = try_to_free_mem_cgroup_pages(memcg,
6773 min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
6774 GFP_KERNEL, reclaim_options);
6775
6776 if (!reclaimed && !nr_retries--)
6777 return -EAGAIN;
6778
6779 nr_reclaimed += reclaimed;
6780 }
6781
6782 return nbytes;
6783 }
6784
6785 static struct cftype memory_files[] = {
6786 {
6787 .name = "current",
6788 .flags = CFTYPE_NOT_ON_ROOT,
6789 .read_u64 = memory_current_read,
6790 },
6791 {
6792 .name = "peak",
6793 .flags = CFTYPE_NOT_ON_ROOT,
6794 .read_u64 = memory_peak_read,
6795 },
6796 {
6797 .name = "min",
6798 .flags = CFTYPE_NOT_ON_ROOT,
6799 .seq_show = memory_min_show,
6800 .write = memory_min_write,
6801 },
6802 {
6803 .name = "low",
6804 .flags = CFTYPE_NOT_ON_ROOT,
6805 .seq_show = memory_low_show,
6806 .write = memory_low_write,
6807 },
6808 {
6809 .name = "high",
6810 .flags = CFTYPE_NOT_ON_ROOT,
6811 .seq_show = memory_high_show,
6812 .write = memory_high_write,
6813 },
6814 {
6815 .name = "max",
6816 .flags = CFTYPE_NOT_ON_ROOT,
6817 .seq_show = memory_max_show,
6818 .write = memory_max_write,
6819 },
6820 {
6821 .name = "events",
6822 .flags = CFTYPE_NOT_ON_ROOT,
6823 .file_offset = offsetof(struct mem_cgroup, events_file),
6824 .seq_show = memory_events_show,
6825 },
6826 {
6827 .name = "events.local",
6828 .flags = CFTYPE_NOT_ON_ROOT,
6829 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6830 .seq_show = memory_events_local_show,
6831 },
6832 {
6833 .name = "stat",
6834 .seq_show = memory_stat_show,
6835 },
6836 #ifdef CONFIG_NUMA
6837 {
6838 .name = "numa_stat",
6839 .seq_show = memory_numa_stat_show,
6840 },
6841 #endif
6842 {
6843 .name = "oom.group",
6844 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6845 .seq_show = memory_oom_group_show,
6846 .write = memory_oom_group_write,
6847 },
6848 {
6849 .name = "reclaim",
6850 .flags = CFTYPE_NS_DELEGATABLE,
6851 .write = memory_reclaim,
6852 },
6853 { } /* terminate */
6854 };
6855
6856 struct cgroup_subsys memory_cgrp_subsys = {
6857 .css_alloc = mem_cgroup_css_alloc,
6858 .css_online = mem_cgroup_css_online,
6859 .css_offline = mem_cgroup_css_offline,
6860 .css_released = mem_cgroup_css_released,
6861 .css_free = mem_cgroup_css_free,
6862 .css_reset = mem_cgroup_css_reset,
6863 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6864 .can_attach = mem_cgroup_can_attach,
6865 .attach = mem_cgroup_attach,
6866 .cancel_attach = mem_cgroup_cancel_attach,
6867 .post_attach = mem_cgroup_move_task,
6868 .dfl_cftypes = memory_files,
6869 .legacy_cftypes = mem_cgroup_legacy_files,
6870 .early_init = 0,
6871 };
6872
6873 /*
6874 * This function calculates an individual cgroup's effective
6875 * protection which is derived from its own memory.min/low, its
6876 * parent's and siblings' settings, as well as the actual memory
6877 * distribution in the tree.
6878 *
6879 * The following rules apply to the effective protection values:
6880 *
6881 * 1. At the first level of reclaim, effective protection is equal to
6882 * the declared protection in memory.min and memory.low.
6883 *
6884 * 2. To enable safe delegation of the protection configuration, at
6885 * subsequent levels the effective protection is capped to the
6886 * parent's effective protection.
6887 *
6888 * 3. To make complex and dynamic subtrees easier to configure, the
6889 * user is allowed to overcommit the declared protection at a given
6890 * level. If that is the case, the parent's effective protection is
6891 * distributed to the children in proportion to how much protection
6892 * they have declared and how much of it they are utilizing.
6893 *
6894 * This makes distribution proportional, but also work-conserving:
6895 * if one cgroup claims much more protection than it uses memory,
6896 * the unused remainder is available to its siblings.
6897 *
6898 * 4. Conversely, when the declared protection is undercommitted at a
6899 * given level, the distribution of the larger parental protection
6900 * budget is NOT proportional. A cgroup's protection from a sibling
6901 * is capped to its own memory.min/low setting.
6902 *
6903 * 5. However, to allow protecting recursive subtrees from each other
6904 * without having to declare each individual cgroup's fixed share
6905 * of the ancestor's claim to protection, any unutilized -
6906 * "floating" - protection from up the tree is distributed in
6907 * proportion to each cgroup's *usage*. This makes the protection
6908 * neutral wrt sibling cgroups and lets them compete freely over
6909 * the shared parental protection budget, but it protects the
6910 * subtree as a whole from neighboring subtrees.
6911 *
6912 * Note that 4. and 5. are not in conflict: 4. is about protecting
6913 * against immediate siblings whereas 5. is about protecting against
6914 * neighboring subtrees.
6915 */
effective_protection(unsigned long usage,unsigned long parent_usage,unsigned long setting,unsigned long parent_effective,unsigned long siblings_protected)6916 static unsigned long effective_protection(unsigned long usage,
6917 unsigned long parent_usage,
6918 unsigned long setting,
6919 unsigned long parent_effective,
6920 unsigned long siblings_protected)
6921 {
6922 unsigned long protected;
6923 unsigned long ep;
6924
6925 protected = min(usage, setting);
6926 /*
6927 * If all cgroups at this level combined claim and use more
6928 * protection than what the parent affords them, distribute
6929 * shares in proportion to utilization.
6930 *
6931 * We are using actual utilization rather than the statically
6932 * claimed protection in order to be work-conserving: claimed
6933 * but unused protection is available to siblings that would
6934 * otherwise get a smaller chunk than what they claimed.
6935 */
6936 if (siblings_protected > parent_effective)
6937 return protected * parent_effective / siblings_protected;
6938
6939 /*
6940 * Ok, utilized protection of all children is within what the
6941 * parent affords them, so we know whatever this child claims
6942 * and utilizes is effectively protected.
6943 *
6944 * If there is unprotected usage beyond this value, reclaim
6945 * will apply pressure in proportion to that amount.
6946 *
6947 * If there is unutilized protection, the cgroup will be fully
6948 * shielded from reclaim, but we do return a smaller value for
6949 * protection than what the group could enjoy in theory. This
6950 * is okay. With the overcommit distribution above, effective
6951 * protection is always dependent on how memory is actually
6952 * consumed among the siblings anyway.
6953 */
6954 ep = protected;
6955
6956 /*
6957 * If the children aren't claiming (all of) the protection
6958 * afforded to them by the parent, distribute the remainder in
6959 * proportion to the (unprotected) memory of each cgroup. That
6960 * way, cgroups that aren't explicitly prioritized wrt each
6961 * other compete freely over the allowance, but they are
6962 * collectively protected from neighboring trees.
6963 *
6964 * We're using unprotected memory for the weight so that if
6965 * some cgroups DO claim explicit protection, we don't protect
6966 * the same bytes twice.
6967 *
6968 * Check both usage and parent_usage against the respective
6969 * protected values. One should imply the other, but they
6970 * aren't read atomically - make sure the division is sane.
6971 */
6972 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6973 return ep;
6974 if (parent_effective > siblings_protected &&
6975 parent_usage > siblings_protected &&
6976 usage > protected) {
6977 unsigned long unclaimed;
6978
6979 unclaimed = parent_effective - siblings_protected;
6980 unclaimed *= usage - protected;
6981 unclaimed /= parent_usage - siblings_protected;
6982
6983 ep += unclaimed;
6984 }
6985
6986 return ep;
6987 }
6988
6989 /**
6990 * mem_cgroup_calculate_protection - check if memory consumption is in the normal range
6991 * @root: the top ancestor of the sub-tree being checked
6992 * @memcg: the memory cgroup to check
6993 *
6994 * WARNING: This function is not stateless! It can only be used as part
6995 * of a top-down tree iteration, not for isolated queries.
6996 */
mem_cgroup_calculate_protection(struct mem_cgroup * root,struct mem_cgroup * memcg)6997 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6998 struct mem_cgroup *memcg)
6999 {
7000 unsigned long usage, parent_usage;
7001 struct mem_cgroup *parent;
7002
7003 if (mem_cgroup_disabled())
7004 return;
7005
7006 if (!root)
7007 root = root_mem_cgroup;
7008
7009 /*
7010 * Effective values of the reclaim targets are ignored so they
7011 * can be stale. Have a look at mem_cgroup_protection for more
7012 * details.
7013 * TODO: calculation should be more robust so that we do not need
7014 * that special casing.
7015 */
7016 if (memcg == root)
7017 return;
7018
7019 usage = page_counter_read(&memcg->memory);
7020 if (!usage)
7021 return;
7022
7023 parent = parent_mem_cgroup(memcg);
7024
7025 if (parent == root) {
7026 memcg->memory.emin = READ_ONCE(memcg->memory.min);
7027 memcg->memory.elow = READ_ONCE(memcg->memory.low);
7028 return;
7029 }
7030
7031 parent_usage = page_counter_read(&parent->memory);
7032
7033 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
7034 READ_ONCE(memcg->memory.min),
7035 READ_ONCE(parent->memory.emin),
7036 atomic_long_read(&parent->memory.children_min_usage)));
7037
7038 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
7039 READ_ONCE(memcg->memory.low),
7040 READ_ONCE(parent->memory.elow),
7041 atomic_long_read(&parent->memory.children_low_usage)));
7042 }
7043
charge_memcg(struct folio * folio,struct mem_cgroup * memcg,gfp_t gfp)7044 static int charge_memcg(struct folio *folio, struct mem_cgroup *memcg,
7045 gfp_t gfp)
7046 {
7047 long nr_pages = folio_nr_pages(folio);
7048 int ret;
7049
7050 ret = try_charge(memcg, gfp, nr_pages);
7051 if (ret)
7052 goto out;
7053
7054 css_get(&memcg->css);
7055 commit_charge(folio, memcg);
7056
7057 local_irq_disable();
7058 mem_cgroup_charge_statistics(memcg, nr_pages);
7059 memcg_check_events(memcg, folio_nid(folio));
7060 local_irq_enable();
7061 out:
7062 return ret;
7063 }
7064
__mem_cgroup_charge(struct folio * folio,struct mm_struct * mm,gfp_t gfp)7065 int __mem_cgroup_charge(struct folio *folio, struct mm_struct *mm, gfp_t gfp)
7066 {
7067 struct mem_cgroup *memcg;
7068 int ret;
7069
7070 memcg = get_mem_cgroup_from_mm(mm);
7071 ret = charge_memcg(folio, memcg, gfp);
7072 css_put(&memcg->css);
7073
7074 return ret;
7075 }
7076
7077 /**
7078 * mem_cgroup_swapin_charge_folio - Charge a newly allocated folio for swapin.
7079 * @folio: folio to charge.
7080 * @mm: mm context of the victim
7081 * @gfp: reclaim mode
7082 * @entry: swap entry for which the folio is allocated
7083 *
7084 * This function charges a folio allocated for swapin. Please call this before
7085 * adding the folio to the swapcache.
7086 *
7087 * Returns 0 on success. Otherwise, an error code is returned.
7088 */
mem_cgroup_swapin_charge_folio(struct folio * folio,struct mm_struct * mm,gfp_t gfp,swp_entry_t entry)7089 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
7090 gfp_t gfp, swp_entry_t entry)
7091 {
7092 struct mem_cgroup *memcg;
7093 unsigned short id;
7094 int ret;
7095
7096 if (mem_cgroup_disabled())
7097 return 0;
7098
7099 id = lookup_swap_cgroup_id(entry);
7100 rcu_read_lock();
7101 memcg = mem_cgroup_from_id(id);
7102 if (!memcg || !css_tryget_online(&memcg->css))
7103 memcg = get_mem_cgroup_from_mm(mm);
7104 rcu_read_unlock();
7105
7106 ret = charge_memcg(folio, memcg, gfp);
7107
7108 css_put(&memcg->css);
7109 return ret;
7110 }
7111
7112 /*
7113 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
7114 * @entry: swap entry for which the page is charged
7115 *
7116 * Call this function after successfully adding the charged page to swapcache.
7117 *
7118 * Note: This function assumes the page for which swap slot is being uncharged
7119 * is order 0 page.
7120 */
mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)7121 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
7122 {
7123 /*
7124 * Cgroup1's unified memory+swap counter has been charged with the
7125 * new swapcache page, finish the transfer by uncharging the swap
7126 * slot. The swap slot would also get uncharged when it dies, but
7127 * it can stick around indefinitely and we'd count the page twice
7128 * the entire time.
7129 *
7130 * Cgroup2 has separate resource counters for memory and swap,
7131 * so this is a non-issue here. Memory and swap charge lifetimes
7132 * correspond 1:1 to page and swap slot lifetimes: we charge the
7133 * page to memory here, and uncharge swap when the slot is freed.
7134 */
7135 if (!mem_cgroup_disabled() && do_memsw_account()) {
7136 /*
7137 * The swap entry might not get freed for a long time,
7138 * let's not wait for it. The page already received a
7139 * memory+swap charge, drop the swap entry duplicate.
7140 */
7141 mem_cgroup_uncharge_swap(entry, 1);
7142 }
7143 }
7144
7145 struct uncharge_gather {
7146 struct mem_cgroup *memcg;
7147 unsigned long nr_memory;
7148 unsigned long pgpgout;
7149 unsigned long nr_kmem;
7150 int nid;
7151 };
7152
uncharge_gather_clear(struct uncharge_gather * ug)7153 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
7154 {
7155 memset(ug, 0, sizeof(*ug));
7156 }
7157
uncharge_batch(const struct uncharge_gather * ug)7158 static void uncharge_batch(const struct uncharge_gather *ug)
7159 {
7160 unsigned long flags;
7161
7162 if (ug->nr_memory) {
7163 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
7164 if (do_memsw_account())
7165 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
7166 if (ug->nr_kmem)
7167 memcg_account_kmem(ug->memcg, -ug->nr_kmem);
7168 memcg_oom_recover(ug->memcg);
7169 }
7170
7171 local_irq_save(flags);
7172 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
7173 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
7174 memcg_check_events(ug->memcg, ug->nid);
7175 local_irq_restore(flags);
7176
7177 /* drop reference from uncharge_folio */
7178 css_put(&ug->memcg->css);
7179 }
7180
uncharge_folio(struct folio * folio,struct uncharge_gather * ug)7181 static void uncharge_folio(struct folio *folio, struct uncharge_gather *ug)
7182 {
7183 long nr_pages;
7184 struct mem_cgroup *memcg;
7185 struct obj_cgroup *objcg;
7186
7187 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7188
7189 /*
7190 * Nobody should be changing or seriously looking at
7191 * folio memcg or objcg at this point, we have fully
7192 * exclusive access to the folio.
7193 */
7194 if (folio_memcg_kmem(folio)) {
7195 objcg = __folio_objcg(folio);
7196 /*
7197 * This get matches the put at the end of the function and
7198 * kmem pages do not hold memcg references anymore.
7199 */
7200 memcg = get_mem_cgroup_from_objcg(objcg);
7201 } else {
7202 memcg = __folio_memcg(folio);
7203 }
7204
7205 if (!memcg)
7206 return;
7207
7208 if (ug->memcg != memcg) {
7209 if (ug->memcg) {
7210 uncharge_batch(ug);
7211 uncharge_gather_clear(ug);
7212 }
7213 ug->memcg = memcg;
7214 ug->nid = folio_nid(folio);
7215
7216 /* pairs with css_put in uncharge_batch */
7217 css_get(&memcg->css);
7218 }
7219
7220 nr_pages = folio_nr_pages(folio);
7221
7222 if (folio_memcg_kmem(folio)) {
7223 ug->nr_memory += nr_pages;
7224 ug->nr_kmem += nr_pages;
7225
7226 folio->memcg_data = 0;
7227 obj_cgroup_put(objcg);
7228 } else {
7229 /* LRU pages aren't accounted at the root level */
7230 if (!mem_cgroup_is_root(memcg))
7231 ug->nr_memory += nr_pages;
7232 ug->pgpgout++;
7233
7234 WARN_ON_ONCE(folio_unqueue_deferred_split(folio));
7235 folio->memcg_data = 0;
7236 }
7237
7238 css_put(&memcg->css);
7239 }
7240
__mem_cgroup_uncharge(struct folio * folio)7241 void __mem_cgroup_uncharge(struct folio *folio)
7242 {
7243 struct uncharge_gather ug;
7244
7245 /* Don't touch folio->lru of any random page, pre-check: */
7246 if (!folio_memcg(folio))
7247 return;
7248
7249 uncharge_gather_clear(&ug);
7250 uncharge_folio(folio, &ug);
7251 uncharge_batch(&ug);
7252 }
7253
7254 /**
7255 * __mem_cgroup_uncharge_list - uncharge a list of page
7256 * @page_list: list of pages to uncharge
7257 *
7258 * Uncharge a list of pages previously charged with
7259 * __mem_cgroup_charge().
7260 */
__mem_cgroup_uncharge_list(struct list_head * page_list)7261 void __mem_cgroup_uncharge_list(struct list_head *page_list)
7262 {
7263 struct uncharge_gather ug;
7264 struct folio *folio;
7265
7266 uncharge_gather_clear(&ug);
7267 list_for_each_entry(folio, page_list, lru)
7268 uncharge_folio(folio, &ug);
7269 if (ug.memcg)
7270 uncharge_batch(&ug);
7271 }
7272
7273 /**
7274 * mem_cgroup_migrate - Charge a folio's replacement.
7275 * @old: Currently circulating folio.
7276 * @new: Replacement folio.
7277 *
7278 * Charge @new as a replacement folio for @old. @old will
7279 * be uncharged upon free.
7280 *
7281 * Both folios must be locked, @new->mapping must be set up.
7282 */
mem_cgroup_migrate(struct folio * old,struct folio * new)7283 void mem_cgroup_migrate(struct folio *old, struct folio *new)
7284 {
7285 struct mem_cgroup *memcg;
7286 long nr_pages = folio_nr_pages(new);
7287 unsigned long flags;
7288
7289 VM_BUG_ON_FOLIO(!folio_test_locked(old), old);
7290 VM_BUG_ON_FOLIO(!folio_test_locked(new), new);
7291 VM_BUG_ON_FOLIO(folio_test_anon(old) != folio_test_anon(new), new);
7292 VM_BUG_ON_FOLIO(folio_nr_pages(old) != nr_pages, new);
7293
7294 if (mem_cgroup_disabled())
7295 return;
7296
7297 /* Page cache replacement: new folio already charged? */
7298 if (folio_memcg(new))
7299 return;
7300
7301 memcg = folio_memcg(old);
7302 VM_WARN_ON_ONCE_FOLIO(!memcg, old);
7303 if (!memcg)
7304 return;
7305
7306 /* Force-charge the new page. The old one will be freed soon */
7307 if (!mem_cgroup_is_root(memcg)) {
7308 page_counter_charge(&memcg->memory, nr_pages);
7309 if (do_memsw_account())
7310 page_counter_charge(&memcg->memsw, nr_pages);
7311 }
7312
7313 css_get(&memcg->css);
7314 commit_charge(new, memcg);
7315
7316 local_irq_save(flags);
7317 mem_cgroup_charge_statistics(memcg, nr_pages);
7318 memcg_check_events(memcg, folio_nid(new));
7319 local_irq_restore(flags);
7320 }
7321
7322 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
7323 EXPORT_SYMBOL(memcg_sockets_enabled_key);
7324
mem_cgroup_sk_alloc(struct sock * sk)7325 void mem_cgroup_sk_alloc(struct sock *sk)
7326 {
7327 struct mem_cgroup *memcg;
7328
7329 if (!mem_cgroup_sockets_enabled)
7330 return;
7331
7332 /* Do not associate the sock with unrelated interrupted task's memcg. */
7333 if (!in_task())
7334 return;
7335
7336 rcu_read_lock();
7337 memcg = mem_cgroup_from_task(current);
7338 if (mem_cgroup_is_root(memcg))
7339 goto out;
7340 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
7341 goto out;
7342 if (css_tryget(&memcg->css))
7343 sk->sk_memcg = memcg;
7344 out:
7345 rcu_read_unlock();
7346 }
7347
mem_cgroup_sk_free(struct sock * sk)7348 void mem_cgroup_sk_free(struct sock *sk)
7349 {
7350 if (sk->sk_memcg)
7351 css_put(&sk->sk_memcg->css);
7352 }
7353
7354 /**
7355 * mem_cgroup_charge_skmem - charge socket memory
7356 * @memcg: memcg to charge
7357 * @nr_pages: number of pages to charge
7358 * @gfp_mask: reclaim mode
7359 *
7360 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
7361 * @memcg's configured limit, %false if it doesn't.
7362 */
mem_cgroup_charge_skmem(struct mem_cgroup * memcg,unsigned int nr_pages,gfp_t gfp_mask)7363 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages,
7364 gfp_t gfp_mask)
7365 {
7366 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7367 struct page_counter *fail;
7368
7369 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7370 memcg->tcpmem_pressure = 0;
7371 return true;
7372 }
7373 memcg->tcpmem_pressure = 1;
7374 if (gfp_mask & __GFP_NOFAIL) {
7375 page_counter_charge(&memcg->tcpmem, nr_pages);
7376 return true;
7377 }
7378 return false;
7379 }
7380
7381 if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
7382 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7383 return true;
7384 }
7385
7386 return false;
7387 }
7388
7389 /**
7390 * mem_cgroup_uncharge_skmem - uncharge socket memory
7391 * @memcg: memcg to uncharge
7392 * @nr_pages: number of pages to uncharge
7393 */
mem_cgroup_uncharge_skmem(struct mem_cgroup * memcg,unsigned int nr_pages)7394 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7395 {
7396 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7397 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7398 return;
7399 }
7400
7401 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7402
7403 refill_stock(memcg, nr_pages);
7404 }
7405
cgroup_memory(char * s)7406 static int __init cgroup_memory(char *s)
7407 {
7408 char *token;
7409
7410 while ((token = strsep(&s, ",")) != NULL) {
7411 if (!*token)
7412 continue;
7413 if (!strcmp(token, "nosocket"))
7414 cgroup_memory_nosocket = true;
7415 if (!strcmp(token, "nokmem"))
7416 cgroup_memory_nokmem = true;
7417 if (!strcmp(token, "nobpf"))
7418 cgroup_memory_nobpf = true;
7419 }
7420 return 1;
7421 }
7422 __setup("cgroup.memory=", cgroup_memory);
7423
7424 /*
7425 * subsys_initcall() for memory controller.
7426 *
7427 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7428 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7429 * basically everything that doesn't depend on a specific mem_cgroup structure
7430 * should be initialized from here.
7431 */
mem_cgroup_init(void)7432 static int __init mem_cgroup_init(void)
7433 {
7434 int cpu, node;
7435
7436 /*
7437 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7438 * used for per-memcg-per-cpu caching of per-node statistics. In order
7439 * to work fine, we should make sure that the overfill threshold can't
7440 * exceed S32_MAX / PAGE_SIZE.
7441 */
7442 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7443
7444 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7445 memcg_hotplug_cpu_dead);
7446
7447 for_each_possible_cpu(cpu)
7448 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7449 drain_local_stock);
7450
7451 for_each_node(node) {
7452 struct mem_cgroup_tree_per_node *rtpn;
7453
7454 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, node);
7455
7456 rtpn->rb_root = RB_ROOT;
7457 rtpn->rb_rightmost = NULL;
7458 spin_lock_init(&rtpn->lock);
7459 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7460 }
7461
7462 return 0;
7463 }
7464 subsys_initcall(mem_cgroup_init);
7465
7466 #ifdef CONFIG_SWAP
mem_cgroup_id_get_online(struct mem_cgroup * memcg)7467 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7468 {
7469 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7470 /*
7471 * The root cgroup cannot be destroyed, so it's refcount must
7472 * always be >= 1.
7473 */
7474 if (WARN_ON_ONCE(mem_cgroup_is_root(memcg))) {
7475 VM_BUG_ON(1);
7476 break;
7477 }
7478 memcg = parent_mem_cgroup(memcg);
7479 if (!memcg)
7480 memcg = root_mem_cgroup;
7481 }
7482 return memcg;
7483 }
7484
7485 /**
7486 * mem_cgroup_swapout - transfer a memsw charge to swap
7487 * @folio: folio whose memsw charge to transfer
7488 * @entry: swap entry to move the charge to
7489 *
7490 * Transfer the memsw charge of @folio to @entry.
7491 */
mem_cgroup_swapout(struct folio * folio,swp_entry_t entry)7492 void mem_cgroup_swapout(struct folio *folio, swp_entry_t entry)
7493 {
7494 struct mem_cgroup *memcg, *swap_memcg;
7495 unsigned int nr_entries;
7496 unsigned short oldid;
7497
7498 VM_BUG_ON_FOLIO(folio_test_lru(folio), folio);
7499 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
7500
7501 if (mem_cgroup_disabled())
7502 return;
7503
7504 if (!do_memsw_account())
7505 return;
7506
7507 memcg = folio_memcg(folio);
7508
7509 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7510 if (!memcg)
7511 return;
7512
7513 /*
7514 * In case the memcg owning these pages has been offlined and doesn't
7515 * have an ID allocated to it anymore, charge the closest online
7516 * ancestor for the swap instead and transfer the memory+swap charge.
7517 */
7518 swap_memcg = mem_cgroup_id_get_online(memcg);
7519 nr_entries = folio_nr_pages(folio);
7520 /* Get references for the tail pages, too */
7521 if (nr_entries > 1)
7522 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7523 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7524 nr_entries);
7525 VM_BUG_ON_FOLIO(oldid, folio);
7526 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7527
7528 folio_unqueue_deferred_split(folio);
7529 folio->memcg_data = 0;
7530
7531 if (!mem_cgroup_is_root(memcg))
7532 page_counter_uncharge(&memcg->memory, nr_entries);
7533
7534 if (memcg != swap_memcg) {
7535 if (!mem_cgroup_is_root(swap_memcg))
7536 page_counter_charge(&swap_memcg->memsw, nr_entries);
7537 page_counter_uncharge(&memcg->memsw, nr_entries);
7538 }
7539
7540 /*
7541 * Interrupts should be disabled here because the caller holds the
7542 * i_pages lock which is taken with interrupts-off. It is
7543 * important here to have the interrupts disabled because it is the
7544 * only synchronisation we have for updating the per-CPU variables.
7545 */
7546 memcg_stats_lock();
7547 mem_cgroup_charge_statistics(memcg, -nr_entries);
7548 memcg_stats_unlock();
7549 memcg_check_events(memcg, folio_nid(folio));
7550
7551 css_put(&memcg->css);
7552 }
7553
7554 /**
7555 * __mem_cgroup_try_charge_swap - try charging swap space for a folio
7556 * @folio: folio being added to swap
7557 * @entry: swap entry to charge
7558 *
7559 * Try to charge @folio's memcg for the swap space at @entry.
7560 *
7561 * Returns 0 on success, -ENOMEM on failure.
7562 */
__mem_cgroup_try_charge_swap(struct folio * folio,swp_entry_t entry)7563 int __mem_cgroup_try_charge_swap(struct folio *folio, swp_entry_t entry)
7564 {
7565 unsigned int nr_pages = folio_nr_pages(folio);
7566 struct page_counter *counter;
7567 struct mem_cgroup *memcg;
7568 unsigned short oldid;
7569
7570 if (do_memsw_account())
7571 return 0;
7572
7573 memcg = folio_memcg(folio);
7574
7575 VM_WARN_ON_ONCE_FOLIO(!memcg, folio);
7576 if (!memcg)
7577 return 0;
7578
7579 if (!entry.val) {
7580 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7581 return 0;
7582 }
7583
7584 memcg = mem_cgroup_id_get_online(memcg);
7585
7586 if (!mem_cgroup_is_root(memcg) &&
7587 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7588 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7589 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7590 mem_cgroup_id_put(memcg);
7591 return -ENOMEM;
7592 }
7593
7594 /* Get references for the tail pages, too */
7595 if (nr_pages > 1)
7596 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7597 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7598 VM_BUG_ON_FOLIO(oldid, folio);
7599 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7600
7601 return 0;
7602 }
7603
7604 /**
7605 * __mem_cgroup_uncharge_swap - uncharge swap space
7606 * @entry: swap entry to uncharge
7607 * @nr_pages: the amount of swap space to uncharge
7608 */
__mem_cgroup_uncharge_swap(swp_entry_t entry,unsigned int nr_pages)7609 void __mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7610 {
7611 struct mem_cgroup *memcg;
7612 unsigned short id;
7613
7614 id = swap_cgroup_record(entry, 0, nr_pages);
7615 rcu_read_lock();
7616 memcg = mem_cgroup_from_id(id);
7617 if (memcg) {
7618 if (!mem_cgroup_is_root(memcg)) {
7619 if (do_memsw_account())
7620 page_counter_uncharge(&memcg->memsw, nr_pages);
7621 else
7622 page_counter_uncharge(&memcg->swap, nr_pages);
7623 }
7624 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7625 mem_cgroup_id_put_many(memcg, nr_pages);
7626 }
7627 rcu_read_unlock();
7628 }
7629
mem_cgroup_get_nr_swap_pages(struct mem_cgroup * memcg)7630 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7631 {
7632 long nr_swap_pages = get_nr_swap_pages();
7633
7634 if (mem_cgroup_disabled() || do_memsw_account())
7635 return nr_swap_pages;
7636 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
7637 nr_swap_pages = min_t(long, nr_swap_pages,
7638 READ_ONCE(memcg->swap.max) -
7639 page_counter_read(&memcg->swap));
7640 return nr_swap_pages;
7641 }
7642
mem_cgroup_swap_full(struct folio * folio)7643 bool mem_cgroup_swap_full(struct folio *folio)
7644 {
7645 struct mem_cgroup *memcg;
7646
7647 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
7648
7649 if (vm_swap_full())
7650 return true;
7651 if (do_memsw_account())
7652 return false;
7653
7654 memcg = folio_memcg(folio);
7655 if (!memcg)
7656 return false;
7657
7658 for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg)) {
7659 unsigned long usage = page_counter_read(&memcg->swap);
7660
7661 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7662 usage * 2 >= READ_ONCE(memcg->swap.max))
7663 return true;
7664 }
7665
7666 return false;
7667 }
7668
setup_swap_account(char * s)7669 static int __init setup_swap_account(char *s)
7670 {
7671 bool res;
7672
7673 if (!kstrtobool(s, &res) && !res)
7674 pr_warn_once("The swapaccount=0 commandline option is deprecated "
7675 "in favor of configuring swap control via cgroupfs. "
7676 "Please report your usecase to linux-mm@kvack.org if you "
7677 "depend on this functionality.\n");
7678 return 1;
7679 }
7680 __setup("swapaccount=", setup_swap_account);
7681
swap_current_read(struct cgroup_subsys_state * css,struct cftype * cft)7682 static u64 swap_current_read(struct cgroup_subsys_state *css,
7683 struct cftype *cft)
7684 {
7685 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7686
7687 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7688 }
7689
swap_peak_read(struct cgroup_subsys_state * css,struct cftype * cft)7690 static u64 swap_peak_read(struct cgroup_subsys_state *css,
7691 struct cftype *cft)
7692 {
7693 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7694
7695 return (u64)memcg->swap.watermark * PAGE_SIZE;
7696 }
7697
swap_high_show(struct seq_file * m,void * v)7698 static int swap_high_show(struct seq_file *m, void *v)
7699 {
7700 return seq_puts_memcg_tunable(m,
7701 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7702 }
7703
swap_high_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)7704 static ssize_t swap_high_write(struct kernfs_open_file *of,
7705 char *buf, size_t nbytes, loff_t off)
7706 {
7707 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7708 unsigned long high;
7709 int err;
7710
7711 buf = strstrip(buf);
7712 err = page_counter_memparse(buf, "max", &high);
7713 if (err)
7714 return err;
7715
7716 page_counter_set_high(&memcg->swap, high);
7717
7718 return nbytes;
7719 }
7720
swap_max_show(struct seq_file * m,void * v)7721 static int swap_max_show(struct seq_file *m, void *v)
7722 {
7723 return seq_puts_memcg_tunable(m,
7724 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7725 }
7726
swap_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)7727 static ssize_t swap_max_write(struct kernfs_open_file *of,
7728 char *buf, size_t nbytes, loff_t off)
7729 {
7730 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7731 unsigned long max;
7732 int err;
7733
7734 buf = strstrip(buf);
7735 err = page_counter_memparse(buf, "max", &max);
7736 if (err)
7737 return err;
7738
7739 xchg(&memcg->swap.max, max);
7740
7741 return nbytes;
7742 }
7743
swap_events_show(struct seq_file * m,void * v)7744 static int swap_events_show(struct seq_file *m, void *v)
7745 {
7746 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7747
7748 seq_printf(m, "high %lu\n",
7749 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7750 seq_printf(m, "max %lu\n",
7751 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7752 seq_printf(m, "fail %lu\n",
7753 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7754
7755 return 0;
7756 }
7757
7758 static struct cftype swap_files[] = {
7759 {
7760 .name = "swap.current",
7761 .flags = CFTYPE_NOT_ON_ROOT,
7762 .read_u64 = swap_current_read,
7763 },
7764 {
7765 .name = "swap.high",
7766 .flags = CFTYPE_NOT_ON_ROOT,
7767 .seq_show = swap_high_show,
7768 .write = swap_high_write,
7769 },
7770 {
7771 .name = "swap.max",
7772 .flags = CFTYPE_NOT_ON_ROOT,
7773 .seq_show = swap_max_show,
7774 .write = swap_max_write,
7775 },
7776 {
7777 .name = "swap.peak",
7778 .flags = CFTYPE_NOT_ON_ROOT,
7779 .read_u64 = swap_peak_read,
7780 },
7781 {
7782 .name = "swap.events",
7783 .flags = CFTYPE_NOT_ON_ROOT,
7784 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7785 .seq_show = swap_events_show,
7786 },
7787 { } /* terminate */
7788 };
7789
7790 static struct cftype memsw_files[] = {
7791 {
7792 .name = "memsw.usage_in_bytes",
7793 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7794 .read_u64 = mem_cgroup_read_u64,
7795 },
7796 {
7797 .name = "memsw.max_usage_in_bytes",
7798 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7799 .write = mem_cgroup_reset,
7800 .read_u64 = mem_cgroup_read_u64,
7801 },
7802 {
7803 .name = "memsw.limit_in_bytes",
7804 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7805 .write = mem_cgroup_write,
7806 .read_u64 = mem_cgroup_read_u64,
7807 },
7808 {
7809 .name = "memsw.failcnt",
7810 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7811 .write = mem_cgroup_reset,
7812 .read_u64 = mem_cgroup_read_u64,
7813 },
7814 { }, /* terminate */
7815 };
7816
7817 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7818 /**
7819 * obj_cgroup_may_zswap - check if this cgroup can zswap
7820 * @objcg: the object cgroup
7821 *
7822 * Check if the hierarchical zswap limit has been reached.
7823 *
7824 * This doesn't check for specific headroom, and it is not atomic
7825 * either. But with zswap, the size of the allocation is only known
7826 * once compression has occured, and this optimistic pre-check avoids
7827 * spending cycles on compression when there is already no room left
7828 * or zswap is disabled altogether somewhere in the hierarchy.
7829 */
obj_cgroup_may_zswap(struct obj_cgroup * objcg)7830 bool obj_cgroup_may_zswap(struct obj_cgroup *objcg)
7831 {
7832 struct mem_cgroup *memcg, *original_memcg;
7833 bool ret = true;
7834
7835 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7836 return true;
7837
7838 original_memcg = get_mem_cgroup_from_objcg(objcg);
7839 for (memcg = original_memcg; !mem_cgroup_is_root(memcg);
7840 memcg = parent_mem_cgroup(memcg)) {
7841 unsigned long max = READ_ONCE(memcg->zswap_max);
7842 unsigned long pages;
7843
7844 if (max == PAGE_COUNTER_MAX)
7845 continue;
7846 if (max == 0) {
7847 ret = false;
7848 break;
7849 }
7850
7851 cgroup_rstat_flush(memcg->css.cgroup);
7852 pages = memcg_page_state(memcg, MEMCG_ZSWAP_B) / PAGE_SIZE;
7853 if (pages < max)
7854 continue;
7855 ret = false;
7856 break;
7857 }
7858 mem_cgroup_put(original_memcg);
7859 return ret;
7860 }
7861
7862 /**
7863 * obj_cgroup_charge_zswap - charge compression backend memory
7864 * @objcg: the object cgroup
7865 * @size: size of compressed object
7866 *
7867 * This forces the charge after obj_cgroup_may_zswap() allowed
7868 * compression and storage in zwap for this cgroup to go ahead.
7869 */
obj_cgroup_charge_zswap(struct obj_cgroup * objcg,size_t size)7870 void obj_cgroup_charge_zswap(struct obj_cgroup *objcg, size_t size)
7871 {
7872 struct mem_cgroup *memcg;
7873
7874 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7875 return;
7876
7877 VM_WARN_ON_ONCE(!(current->flags & PF_MEMALLOC));
7878
7879 /* PF_MEMALLOC context, charging must succeed */
7880 if (obj_cgroup_charge(objcg, GFP_KERNEL, size))
7881 VM_WARN_ON_ONCE(1);
7882
7883 rcu_read_lock();
7884 memcg = obj_cgroup_memcg(objcg);
7885 mod_memcg_state(memcg, MEMCG_ZSWAP_B, size);
7886 mod_memcg_state(memcg, MEMCG_ZSWAPPED, 1);
7887 rcu_read_unlock();
7888 }
7889
7890 /**
7891 * obj_cgroup_uncharge_zswap - uncharge compression backend memory
7892 * @objcg: the object cgroup
7893 * @size: size of compressed object
7894 *
7895 * Uncharges zswap memory on page in.
7896 */
obj_cgroup_uncharge_zswap(struct obj_cgroup * objcg,size_t size)7897 void obj_cgroup_uncharge_zswap(struct obj_cgroup *objcg, size_t size)
7898 {
7899 struct mem_cgroup *memcg;
7900
7901 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7902 return;
7903
7904 obj_cgroup_uncharge(objcg, size);
7905
7906 rcu_read_lock();
7907 memcg = obj_cgroup_memcg(objcg);
7908 mod_memcg_state(memcg, MEMCG_ZSWAP_B, -size);
7909 mod_memcg_state(memcg, MEMCG_ZSWAPPED, -1);
7910 rcu_read_unlock();
7911 }
7912
zswap_current_read(struct cgroup_subsys_state * css,struct cftype * cft)7913 static u64 zswap_current_read(struct cgroup_subsys_state *css,
7914 struct cftype *cft)
7915 {
7916 cgroup_rstat_flush(css->cgroup);
7917 return memcg_page_state(mem_cgroup_from_css(css), MEMCG_ZSWAP_B);
7918 }
7919
zswap_max_show(struct seq_file * m,void * v)7920 static int zswap_max_show(struct seq_file *m, void *v)
7921 {
7922 return seq_puts_memcg_tunable(m,
7923 READ_ONCE(mem_cgroup_from_seq(m)->zswap_max));
7924 }
7925
zswap_max_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)7926 static ssize_t zswap_max_write(struct kernfs_open_file *of,
7927 char *buf, size_t nbytes, loff_t off)
7928 {
7929 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7930 unsigned long max;
7931 int err;
7932
7933 buf = strstrip(buf);
7934 err = page_counter_memparse(buf, "max", &max);
7935 if (err)
7936 return err;
7937
7938 xchg(&memcg->zswap_max, max);
7939
7940 return nbytes;
7941 }
7942
7943 static struct cftype zswap_files[] = {
7944 {
7945 .name = "zswap.current",
7946 .flags = CFTYPE_NOT_ON_ROOT,
7947 .read_u64 = zswap_current_read,
7948 },
7949 {
7950 .name = "zswap.max",
7951 .flags = CFTYPE_NOT_ON_ROOT,
7952 .seq_show = zswap_max_show,
7953 .write = zswap_max_write,
7954 },
7955 { } /* terminate */
7956 };
7957 #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
7958
mem_cgroup_swap_init(void)7959 static int __init mem_cgroup_swap_init(void)
7960 {
7961 if (mem_cgroup_disabled())
7962 return 0;
7963
7964 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7965 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7966 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
7967 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, zswap_files));
7968 #endif
7969 return 0;
7970 }
7971 subsys_initcall(mem_cgroup_swap_init);
7972
7973 #endif /* CONFIG_SWAP */
7974