]> git.proxmox.com Git - mirror_ubuntu-focal-kernel.git/blob - mm/memcontrol.c
Merge branch 'x86/cleanups' into x86/cpu, to pick up dependent changes
[mirror_ubuntu-focal-kernel.git] / mm / memcontrol.c
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
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/mm.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
50 #include <linux/fs.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/seq_buf.h>
61 #include "internal.h"
62 #include <net/sock.h>
63 #include <net/ip.h>
64 #include "slab.h"
65
66 #include <linux/uaccess.h>
67
68 #include <trace/events/vmscan.h>
69
70 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
71 EXPORT_SYMBOL(memory_cgrp_subsys);
72
73 struct mem_cgroup *root_mem_cgroup __read_mostly;
74
75 #define MEM_CGROUP_RECLAIM_RETRIES 5
76
77 /* Socket memory accounting disabled? */
78 static bool cgroup_memory_nosocket;
79
80 /* Kernel memory accounting disabled? */
81 static bool cgroup_memory_nokmem;
82
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
86 #else
87 #define do_swap_account 0
88 #endif
89
90 /* Whether legacy memory+swap accounting is active */
91 static bool do_memsw_account(void)
92 {
93 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
94 }
95
96 static const char *const mem_cgroup_lru_names[] = {
97 "inactive_anon",
98 "active_anon",
99 "inactive_file",
100 "active_file",
101 "unevictable",
102 };
103
104 #define THRESHOLDS_EVENTS_TARGET 128
105 #define SOFTLIMIT_EVENTS_TARGET 1024
106 #define NUMAINFO_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_hierarchical_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 enum charge_type {
207 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
208 MEM_CGROUP_CHARGE_TYPE_ANON,
209 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
210 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
211 NR_CHARGE_TYPE,
212 };
213
214 /* for encoding cft->private value on file */
215 enum res_type {
216 _MEM,
217 _MEMSWAP,
218 _OOM_TYPE,
219 _KMEM,
220 _TCP,
221 };
222
223 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
224 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
225 #define MEMFILE_ATTR(val) ((val) & 0xffff)
226 /* Used for OOM nofiier */
227 #define OOM_CONTROL (0)
228
229 /*
230 * Iteration constructs for visiting all cgroups (under a tree). If
231 * loops are exited prematurely (break), mem_cgroup_iter_break() must
232 * be used for reference counting.
233 */
234 #define for_each_mem_cgroup_tree(iter, root) \
235 for (iter = mem_cgroup_iter(root, NULL, NULL); \
236 iter != NULL; \
237 iter = mem_cgroup_iter(root, iter, NULL))
238
239 #define for_each_mem_cgroup(iter) \
240 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
241 iter != NULL; \
242 iter = mem_cgroup_iter(NULL, iter, NULL))
243
244 static inline bool should_force_charge(void)
245 {
246 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
247 (current->flags & PF_EXITING);
248 }
249
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
252 {
253 if (!memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
256 }
257
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
259 {
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
261 }
262
263 #ifdef CONFIG_MEMCG_KMEM
264 /*
265 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
266 * The main reason for not using cgroup id for this:
267 * this works better in sparse environments, where we have a lot of memcgs,
268 * but only a few kmem-limited. Or also, if we have, for instance, 200
269 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
270 * 200 entry array for that.
271 *
272 * The current size of the caches array is stored in memcg_nr_cache_ids. It
273 * will double each time we have to increase it.
274 */
275 static DEFINE_IDA(memcg_cache_ida);
276 int memcg_nr_cache_ids;
277
278 /* Protects memcg_nr_cache_ids */
279 static DECLARE_RWSEM(memcg_cache_ids_sem);
280
281 void memcg_get_cache_ids(void)
282 {
283 down_read(&memcg_cache_ids_sem);
284 }
285
286 void memcg_put_cache_ids(void)
287 {
288 up_read(&memcg_cache_ids_sem);
289 }
290
291 /*
292 * MIN_SIZE is different than 1, because we would like to avoid going through
293 * the alloc/free process all the time. In a small machine, 4 kmem-limited
294 * cgroups is a reasonable guess. In the future, it could be a parameter or
295 * tunable, but that is strictly not necessary.
296 *
297 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
298 * this constant directly from cgroup, but it is understandable that this is
299 * better kept as an internal representation in cgroup.c. In any case, the
300 * cgrp_id space is not getting any smaller, and we don't have to necessarily
301 * increase ours as well if it increases.
302 */
303 #define MEMCG_CACHES_MIN_SIZE 4
304 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
305
306 /*
307 * A lot of the calls to the cache allocation functions are expected to be
308 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
309 * conditional to this static branch, we'll have to allow modules that does
310 * kmem_cache_alloc and the such to see this symbol as well
311 */
312 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
313 EXPORT_SYMBOL(memcg_kmem_enabled_key);
314
315 struct workqueue_struct *memcg_kmem_cache_wq;
316
317 static int memcg_shrinker_map_size;
318 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
319
320 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
321 {
322 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
323 }
324
325 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
326 int size, int old_size)
327 {
328 struct memcg_shrinker_map *new, *old;
329 int nid;
330
331 lockdep_assert_held(&memcg_shrinker_map_mutex);
332
333 for_each_node(nid) {
334 old = rcu_dereference_protected(
335 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
336 /* Not yet online memcg */
337 if (!old)
338 return 0;
339
340 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
341 if (!new)
342 return -ENOMEM;
343
344 /* Set all old bits, clear all new bits */
345 memset(new->map, (int)0xff, old_size);
346 memset((void *)new->map + old_size, 0, size - old_size);
347
348 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
349 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
350 }
351
352 return 0;
353 }
354
355 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
356 {
357 struct mem_cgroup_per_node *pn;
358 struct memcg_shrinker_map *map;
359 int nid;
360
361 if (mem_cgroup_is_root(memcg))
362 return;
363
364 for_each_node(nid) {
365 pn = mem_cgroup_nodeinfo(memcg, nid);
366 map = rcu_dereference_protected(pn->shrinker_map, true);
367 if (map)
368 kvfree(map);
369 rcu_assign_pointer(pn->shrinker_map, NULL);
370 }
371 }
372
373 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
374 {
375 struct memcg_shrinker_map *map;
376 int nid, size, ret = 0;
377
378 if (mem_cgroup_is_root(memcg))
379 return 0;
380
381 mutex_lock(&memcg_shrinker_map_mutex);
382 size = memcg_shrinker_map_size;
383 for_each_node(nid) {
384 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
385 if (!map) {
386 memcg_free_shrinker_maps(memcg);
387 ret = -ENOMEM;
388 break;
389 }
390 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
391 }
392 mutex_unlock(&memcg_shrinker_map_mutex);
393
394 return ret;
395 }
396
397 int memcg_expand_shrinker_maps(int new_id)
398 {
399 int size, old_size, ret = 0;
400 struct mem_cgroup *memcg;
401
402 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
403 old_size = memcg_shrinker_map_size;
404 if (size <= old_size)
405 return 0;
406
407 mutex_lock(&memcg_shrinker_map_mutex);
408 if (!root_mem_cgroup)
409 goto unlock;
410
411 for_each_mem_cgroup(memcg) {
412 if (mem_cgroup_is_root(memcg))
413 continue;
414 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
415 if (ret)
416 goto unlock;
417 }
418 unlock:
419 if (!ret)
420 memcg_shrinker_map_size = size;
421 mutex_unlock(&memcg_shrinker_map_mutex);
422 return ret;
423 }
424
425 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 {
427 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
428 struct memcg_shrinker_map *map;
429
430 rcu_read_lock();
431 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
432 /* Pairs with smp mb in shrink_slab() */
433 smp_mb__before_atomic();
434 set_bit(shrinker_id, map->map);
435 rcu_read_unlock();
436 }
437 }
438
439 #else /* CONFIG_MEMCG_KMEM */
440 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
441 {
442 return 0;
443 }
444 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) { }
445 #endif /* CONFIG_MEMCG_KMEM */
446
447 /**
448 * mem_cgroup_css_from_page - css of the memcg associated with a page
449 * @page: page of interest
450 *
451 * If memcg is bound to the default hierarchy, css of the memcg associated
452 * with @page is returned. The returned css remains associated with @page
453 * until it is released.
454 *
455 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
456 * is returned.
457 */
458 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
459 {
460 struct mem_cgroup *memcg;
461
462 memcg = page->mem_cgroup;
463
464 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
465 memcg = root_mem_cgroup;
466
467 return &memcg->css;
468 }
469
470 /**
471 * page_cgroup_ino - return inode number of the memcg a page is charged to
472 * @page: the page
473 *
474 * Look up the closest online ancestor of the memory cgroup @page is charged to
475 * and return its inode number or 0 if @page is not charged to any cgroup. It
476 * is safe to call this function without holding a reference to @page.
477 *
478 * Note, this function is inherently racy, because there is nothing to prevent
479 * the cgroup inode from getting torn down and potentially reallocated a moment
480 * after page_cgroup_ino() returns, so it only should be used by callers that
481 * do not care (such as procfs interfaces).
482 */
483 ino_t page_cgroup_ino(struct page *page)
484 {
485 struct mem_cgroup *memcg;
486 unsigned long ino = 0;
487
488 rcu_read_lock();
489 if (PageHead(page) && PageSlab(page))
490 memcg = memcg_from_slab_page(page);
491 else
492 memcg = READ_ONCE(page->mem_cgroup);
493 while (memcg && !(memcg->css.flags & CSS_ONLINE))
494 memcg = parent_mem_cgroup(memcg);
495 if (memcg)
496 ino = cgroup_ino(memcg->css.cgroup);
497 rcu_read_unlock();
498 return ino;
499 }
500
501 static struct mem_cgroup_per_node *
502 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
503 {
504 int nid = page_to_nid(page);
505
506 return memcg->nodeinfo[nid];
507 }
508
509 static struct mem_cgroup_tree_per_node *
510 soft_limit_tree_node(int nid)
511 {
512 return soft_limit_tree.rb_tree_per_node[nid];
513 }
514
515 static struct mem_cgroup_tree_per_node *
516 soft_limit_tree_from_page(struct page *page)
517 {
518 int nid = page_to_nid(page);
519
520 return soft_limit_tree.rb_tree_per_node[nid];
521 }
522
523 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
524 struct mem_cgroup_tree_per_node *mctz,
525 unsigned long new_usage_in_excess)
526 {
527 struct rb_node **p = &mctz->rb_root.rb_node;
528 struct rb_node *parent = NULL;
529 struct mem_cgroup_per_node *mz_node;
530 bool rightmost = true;
531
532 if (mz->on_tree)
533 return;
534
535 mz->usage_in_excess = new_usage_in_excess;
536 if (!mz->usage_in_excess)
537 return;
538 while (*p) {
539 parent = *p;
540 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
541 tree_node);
542 if (mz->usage_in_excess < mz_node->usage_in_excess) {
543 p = &(*p)->rb_left;
544 rightmost = false;
545 }
546
547 /*
548 * We can't avoid mem cgroups that are over their soft
549 * limit by the same amount
550 */
551 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
552 p = &(*p)->rb_right;
553 }
554
555 if (rightmost)
556 mctz->rb_rightmost = &mz->tree_node;
557
558 rb_link_node(&mz->tree_node, parent, p);
559 rb_insert_color(&mz->tree_node, &mctz->rb_root);
560 mz->on_tree = true;
561 }
562
563 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
564 struct mem_cgroup_tree_per_node *mctz)
565 {
566 if (!mz->on_tree)
567 return;
568
569 if (&mz->tree_node == mctz->rb_rightmost)
570 mctz->rb_rightmost = rb_prev(&mz->tree_node);
571
572 rb_erase(&mz->tree_node, &mctz->rb_root);
573 mz->on_tree = false;
574 }
575
576 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
577 struct mem_cgroup_tree_per_node *mctz)
578 {
579 unsigned long flags;
580
581 spin_lock_irqsave(&mctz->lock, flags);
582 __mem_cgroup_remove_exceeded(mz, mctz);
583 spin_unlock_irqrestore(&mctz->lock, flags);
584 }
585
586 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
587 {
588 unsigned long nr_pages = page_counter_read(&memcg->memory);
589 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
590 unsigned long excess = 0;
591
592 if (nr_pages > soft_limit)
593 excess = nr_pages - soft_limit;
594
595 return excess;
596 }
597
598 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
599 {
600 unsigned long excess;
601 struct mem_cgroup_per_node *mz;
602 struct mem_cgroup_tree_per_node *mctz;
603
604 mctz = soft_limit_tree_from_page(page);
605 if (!mctz)
606 return;
607 /*
608 * Necessary to update all ancestors when hierarchy is used.
609 * because their event counter is not touched.
610 */
611 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
612 mz = mem_cgroup_page_nodeinfo(memcg, page);
613 excess = soft_limit_excess(memcg);
614 /*
615 * We have to update the tree if mz is on RB-tree or
616 * mem is over its softlimit.
617 */
618 if (excess || mz->on_tree) {
619 unsigned long flags;
620
621 spin_lock_irqsave(&mctz->lock, flags);
622 /* if on-tree, remove it */
623 if (mz->on_tree)
624 __mem_cgroup_remove_exceeded(mz, mctz);
625 /*
626 * Insert again. mz->usage_in_excess will be updated.
627 * If excess is 0, no tree ops.
628 */
629 __mem_cgroup_insert_exceeded(mz, mctz, excess);
630 spin_unlock_irqrestore(&mctz->lock, flags);
631 }
632 }
633 }
634
635 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
636 {
637 struct mem_cgroup_tree_per_node *mctz;
638 struct mem_cgroup_per_node *mz;
639 int nid;
640
641 for_each_node(nid) {
642 mz = mem_cgroup_nodeinfo(memcg, nid);
643 mctz = soft_limit_tree_node(nid);
644 if (mctz)
645 mem_cgroup_remove_exceeded(mz, mctz);
646 }
647 }
648
649 static struct mem_cgroup_per_node *
650 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
651 {
652 struct mem_cgroup_per_node *mz;
653
654 retry:
655 mz = NULL;
656 if (!mctz->rb_rightmost)
657 goto done; /* Nothing to reclaim from */
658
659 mz = rb_entry(mctz->rb_rightmost,
660 struct mem_cgroup_per_node, tree_node);
661 /*
662 * Remove the node now but someone else can add it back,
663 * we will to add it back at the end of reclaim to its correct
664 * position in the tree.
665 */
666 __mem_cgroup_remove_exceeded(mz, mctz);
667 if (!soft_limit_excess(mz->memcg) ||
668 !css_tryget_online(&mz->memcg->css))
669 goto retry;
670 done:
671 return mz;
672 }
673
674 static struct mem_cgroup_per_node *
675 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
676 {
677 struct mem_cgroup_per_node *mz;
678
679 spin_lock_irq(&mctz->lock);
680 mz = __mem_cgroup_largest_soft_limit_node(mctz);
681 spin_unlock_irq(&mctz->lock);
682 return mz;
683 }
684
685 /**
686 * __mod_memcg_state - update cgroup memory statistics
687 * @memcg: the memory cgroup
688 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
689 * @val: delta to add to the counter, can be negative
690 */
691 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
692 {
693 long x;
694
695 if (mem_cgroup_disabled())
696 return;
697
698 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
699 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
700 struct mem_cgroup *mi;
701
702 /*
703 * Batch local counters to keep them in sync with
704 * the hierarchical ones.
705 */
706 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
707 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
708 atomic_long_add(x, &mi->vmstats[idx]);
709 x = 0;
710 }
711 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
712 }
713
714 static struct mem_cgroup_per_node *
715 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
716 {
717 struct mem_cgroup *parent;
718
719 parent = parent_mem_cgroup(pn->memcg);
720 if (!parent)
721 return NULL;
722 return mem_cgroup_nodeinfo(parent, nid);
723 }
724
725 /**
726 * __mod_lruvec_state - update lruvec memory statistics
727 * @lruvec: the lruvec
728 * @idx: the stat item
729 * @val: delta to add to the counter, can be negative
730 *
731 * The lruvec is the intersection of the NUMA node and a cgroup. This
732 * function updates the all three counters that are affected by a
733 * change of state at this level: per-node, per-cgroup, per-lruvec.
734 */
735 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
736 int val)
737 {
738 pg_data_t *pgdat = lruvec_pgdat(lruvec);
739 struct mem_cgroup_per_node *pn;
740 struct mem_cgroup *memcg;
741 long x;
742
743 /* Update node */
744 __mod_node_page_state(pgdat, idx, val);
745
746 if (mem_cgroup_disabled())
747 return;
748
749 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
750 memcg = pn->memcg;
751
752 /* Update memcg */
753 __mod_memcg_state(memcg, idx, val);
754
755 /* Update lruvec */
756 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
757
758 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
759 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
760 struct mem_cgroup_per_node *pi;
761
762 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
763 atomic_long_add(x, &pi->lruvec_stat[idx]);
764 x = 0;
765 }
766 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
767 }
768
769 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
770 {
771 struct page *page = virt_to_head_page(p);
772 pg_data_t *pgdat = page_pgdat(page);
773 struct mem_cgroup *memcg;
774 struct lruvec *lruvec;
775
776 rcu_read_lock();
777 memcg = memcg_from_slab_page(page);
778
779 /* Untracked pages have no memcg, no lruvec. Update only the node */
780 if (!memcg || memcg == root_mem_cgroup) {
781 __mod_node_page_state(pgdat, idx, val);
782 } else {
783 lruvec = mem_cgroup_lruvec(pgdat, memcg);
784 __mod_lruvec_state(lruvec, idx, val);
785 }
786 rcu_read_unlock();
787 }
788
789 /**
790 * __count_memcg_events - account VM events in a cgroup
791 * @memcg: the memory cgroup
792 * @idx: the event item
793 * @count: the number of events that occured
794 */
795 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
796 unsigned long count)
797 {
798 unsigned long x;
799
800 if (mem_cgroup_disabled())
801 return;
802
803 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
804 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
805 struct mem_cgroup *mi;
806
807 /*
808 * Batch local counters to keep them in sync with
809 * the hierarchical ones.
810 */
811 __this_cpu_add(memcg->vmstats_local->events[idx], x);
812 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
813 atomic_long_add(x, &mi->vmevents[idx]);
814 x = 0;
815 }
816 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
817 }
818
819 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
820 {
821 return atomic_long_read(&memcg->vmevents[event]);
822 }
823
824 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
825 {
826 long x = 0;
827 int cpu;
828
829 for_each_possible_cpu(cpu)
830 x += per_cpu(memcg->vmstats_local->events[event], cpu);
831 return x;
832 }
833
834 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
835 struct page *page,
836 bool compound, int nr_pages)
837 {
838 /*
839 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
840 * counted as CACHE even if it's on ANON LRU.
841 */
842 if (PageAnon(page))
843 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
844 else {
845 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
846 if (PageSwapBacked(page))
847 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
848 }
849
850 if (compound) {
851 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
852 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
853 }
854
855 /* pagein of a big page is an event. So, ignore page size */
856 if (nr_pages > 0)
857 __count_memcg_events(memcg, PGPGIN, 1);
858 else {
859 __count_memcg_events(memcg, PGPGOUT, 1);
860 nr_pages = -nr_pages; /* for event */
861 }
862
863 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
864 }
865
866 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
867 enum mem_cgroup_events_target target)
868 {
869 unsigned long val, next;
870
871 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
872 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
873 /* from time_after() in jiffies.h */
874 if ((long)(next - val) < 0) {
875 switch (target) {
876 case MEM_CGROUP_TARGET_THRESH:
877 next = val + THRESHOLDS_EVENTS_TARGET;
878 break;
879 case MEM_CGROUP_TARGET_SOFTLIMIT:
880 next = val + SOFTLIMIT_EVENTS_TARGET;
881 break;
882 case MEM_CGROUP_TARGET_NUMAINFO:
883 next = val + NUMAINFO_EVENTS_TARGET;
884 break;
885 default:
886 break;
887 }
888 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
889 return true;
890 }
891 return false;
892 }
893
894 /*
895 * Check events in order.
896 *
897 */
898 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
899 {
900 /* threshold event is triggered in finer grain than soft limit */
901 if (unlikely(mem_cgroup_event_ratelimit(memcg,
902 MEM_CGROUP_TARGET_THRESH))) {
903 bool do_softlimit;
904 bool do_numainfo __maybe_unused;
905
906 do_softlimit = mem_cgroup_event_ratelimit(memcg,
907 MEM_CGROUP_TARGET_SOFTLIMIT);
908 #if MAX_NUMNODES > 1
909 do_numainfo = mem_cgroup_event_ratelimit(memcg,
910 MEM_CGROUP_TARGET_NUMAINFO);
911 #endif
912 mem_cgroup_threshold(memcg);
913 if (unlikely(do_softlimit))
914 mem_cgroup_update_tree(memcg, page);
915 #if MAX_NUMNODES > 1
916 if (unlikely(do_numainfo))
917 atomic_inc(&memcg->numainfo_events);
918 #endif
919 }
920 }
921
922 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
923 {
924 /*
925 * mm_update_next_owner() may clear mm->owner to NULL
926 * if it races with swapoff, page migration, etc.
927 * So this can be called with p == NULL.
928 */
929 if (unlikely(!p))
930 return NULL;
931
932 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
933 }
934 EXPORT_SYMBOL(mem_cgroup_from_task);
935
936 /**
937 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
938 * @mm: mm from which memcg should be extracted. It can be NULL.
939 *
940 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
941 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
942 * returned.
943 */
944 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
945 {
946 struct mem_cgroup *memcg;
947
948 if (mem_cgroup_disabled())
949 return NULL;
950
951 rcu_read_lock();
952 do {
953 /*
954 * Page cache insertions can happen withou an
955 * actual mm context, e.g. during disk probing
956 * on boot, loopback IO, acct() writes etc.
957 */
958 if (unlikely(!mm))
959 memcg = root_mem_cgroup;
960 else {
961 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
962 if (unlikely(!memcg))
963 memcg = root_mem_cgroup;
964 }
965 } while (!css_tryget_online(&memcg->css));
966 rcu_read_unlock();
967 return memcg;
968 }
969 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
970
971 /**
972 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
973 * @page: page from which memcg should be extracted.
974 *
975 * Obtain a reference on page->memcg and returns it if successful. Otherwise
976 * root_mem_cgroup is returned.
977 */
978 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
979 {
980 struct mem_cgroup *memcg = page->mem_cgroup;
981
982 if (mem_cgroup_disabled())
983 return NULL;
984
985 rcu_read_lock();
986 if (!memcg || !css_tryget_online(&memcg->css))
987 memcg = root_mem_cgroup;
988 rcu_read_unlock();
989 return memcg;
990 }
991 EXPORT_SYMBOL(get_mem_cgroup_from_page);
992
993 /**
994 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
995 */
996 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
997 {
998 if (unlikely(current->active_memcg)) {
999 struct mem_cgroup *memcg = root_mem_cgroup;
1000
1001 rcu_read_lock();
1002 if (css_tryget_online(&current->active_memcg->css))
1003 memcg = current->active_memcg;
1004 rcu_read_unlock();
1005 return memcg;
1006 }
1007 return get_mem_cgroup_from_mm(current->mm);
1008 }
1009
1010 /**
1011 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1012 * @root: hierarchy root
1013 * @prev: previously returned memcg, NULL on first invocation
1014 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1015 *
1016 * Returns references to children of the hierarchy below @root, or
1017 * @root itself, or %NULL after a full round-trip.
1018 *
1019 * Caller must pass the return value in @prev on subsequent
1020 * invocations for reference counting, or use mem_cgroup_iter_break()
1021 * to cancel a hierarchy walk before the round-trip is complete.
1022 *
1023 * Reclaimers can specify a node and a priority level in @reclaim to
1024 * divide up the memcgs in the hierarchy among all concurrent
1025 * reclaimers operating on the same node and priority.
1026 */
1027 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1028 struct mem_cgroup *prev,
1029 struct mem_cgroup_reclaim_cookie *reclaim)
1030 {
1031 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1032 struct cgroup_subsys_state *css = NULL;
1033 struct mem_cgroup *memcg = NULL;
1034 struct mem_cgroup *pos = NULL;
1035
1036 if (mem_cgroup_disabled())
1037 return NULL;
1038
1039 if (!root)
1040 root = root_mem_cgroup;
1041
1042 if (prev && !reclaim)
1043 pos = prev;
1044
1045 if (!root->use_hierarchy && root != root_mem_cgroup) {
1046 if (prev)
1047 goto out;
1048 return root;
1049 }
1050
1051 rcu_read_lock();
1052
1053 if (reclaim) {
1054 struct mem_cgroup_per_node *mz;
1055
1056 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1057 iter = &mz->iter[reclaim->priority];
1058
1059 if (prev && reclaim->generation != iter->generation)
1060 goto out_unlock;
1061
1062 while (1) {
1063 pos = READ_ONCE(iter->position);
1064 if (!pos || css_tryget(&pos->css))
1065 break;
1066 /*
1067 * css reference reached zero, so iter->position will
1068 * be cleared by ->css_released. However, we should not
1069 * rely on this happening soon, because ->css_released
1070 * is called from a work queue, and by busy-waiting we
1071 * might block it. So we clear iter->position right
1072 * away.
1073 */
1074 (void)cmpxchg(&iter->position, pos, NULL);
1075 }
1076 }
1077
1078 if (pos)
1079 css = &pos->css;
1080
1081 for (;;) {
1082 css = css_next_descendant_pre(css, &root->css);
1083 if (!css) {
1084 /*
1085 * Reclaimers share the hierarchy walk, and a
1086 * new one might jump in right at the end of
1087 * the hierarchy - make sure they see at least
1088 * one group and restart from the beginning.
1089 */
1090 if (!prev)
1091 continue;
1092 break;
1093 }
1094
1095 /*
1096 * Verify the css and acquire a reference. The root
1097 * is provided by the caller, so we know it's alive
1098 * and kicking, and don't take an extra reference.
1099 */
1100 memcg = mem_cgroup_from_css(css);
1101
1102 if (css == &root->css)
1103 break;
1104
1105 if (css_tryget(css))
1106 break;
1107
1108 memcg = NULL;
1109 }
1110
1111 if (reclaim) {
1112 /*
1113 * The position could have already been updated by a competing
1114 * thread, so check that the value hasn't changed since we read
1115 * it to avoid reclaiming from the same cgroup twice.
1116 */
1117 (void)cmpxchg(&iter->position, pos, memcg);
1118
1119 if (pos)
1120 css_put(&pos->css);
1121
1122 if (!memcg)
1123 iter->generation++;
1124 else if (!prev)
1125 reclaim->generation = iter->generation;
1126 }
1127
1128 out_unlock:
1129 rcu_read_unlock();
1130 out:
1131 if (prev && prev != root)
1132 css_put(&prev->css);
1133
1134 return memcg;
1135 }
1136
1137 /**
1138 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1139 * @root: hierarchy root
1140 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1141 */
1142 void mem_cgroup_iter_break(struct mem_cgroup *root,
1143 struct mem_cgroup *prev)
1144 {
1145 if (!root)
1146 root = root_mem_cgroup;
1147 if (prev && prev != root)
1148 css_put(&prev->css);
1149 }
1150
1151 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1152 struct mem_cgroup *dead_memcg)
1153 {
1154 struct mem_cgroup_reclaim_iter *iter;
1155 struct mem_cgroup_per_node *mz;
1156 int nid;
1157 int i;
1158
1159 for_each_node(nid) {
1160 mz = mem_cgroup_nodeinfo(from, nid);
1161 for (i = 0; i <= DEF_PRIORITY; i++) {
1162 iter = &mz->iter[i];
1163 cmpxchg(&iter->position,
1164 dead_memcg, NULL);
1165 }
1166 }
1167 }
1168
1169 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1170 {
1171 struct mem_cgroup *memcg = dead_memcg;
1172 struct mem_cgroup *last;
1173
1174 do {
1175 __invalidate_reclaim_iterators(memcg, dead_memcg);
1176 last = memcg;
1177 } while ((memcg = parent_mem_cgroup(memcg)));
1178
1179 /*
1180 * When cgruop1 non-hierarchy mode is used,
1181 * parent_mem_cgroup() does not walk all the way up to the
1182 * cgroup root (root_mem_cgroup). So we have to handle
1183 * dead_memcg from cgroup root separately.
1184 */
1185 if (last != root_mem_cgroup)
1186 __invalidate_reclaim_iterators(root_mem_cgroup,
1187 dead_memcg);
1188 }
1189
1190 /**
1191 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1192 * @memcg: hierarchy root
1193 * @fn: function to call for each task
1194 * @arg: argument passed to @fn
1195 *
1196 * This function iterates over tasks attached to @memcg or to any of its
1197 * descendants and calls @fn for each task. If @fn returns a non-zero
1198 * value, the function breaks the iteration loop and returns the value.
1199 * Otherwise, it will iterate over all tasks and return 0.
1200 *
1201 * This function must not be called for the root memory cgroup.
1202 */
1203 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1204 int (*fn)(struct task_struct *, void *), void *arg)
1205 {
1206 struct mem_cgroup *iter;
1207 int ret = 0;
1208
1209 BUG_ON(memcg == root_mem_cgroup);
1210
1211 for_each_mem_cgroup_tree(iter, memcg) {
1212 struct css_task_iter it;
1213 struct task_struct *task;
1214
1215 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1216 while (!ret && (task = css_task_iter_next(&it)))
1217 ret = fn(task, arg);
1218 css_task_iter_end(&it);
1219 if (ret) {
1220 mem_cgroup_iter_break(memcg, iter);
1221 break;
1222 }
1223 }
1224 return ret;
1225 }
1226
1227 /**
1228 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1229 * @page: the page
1230 * @pgdat: pgdat of the page
1231 *
1232 * This function is only safe when following the LRU page isolation
1233 * and putback protocol: the LRU lock must be held, and the page must
1234 * either be PageLRU() or the caller must have isolated/allocated it.
1235 */
1236 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1237 {
1238 struct mem_cgroup_per_node *mz;
1239 struct mem_cgroup *memcg;
1240 struct lruvec *lruvec;
1241
1242 if (mem_cgroup_disabled()) {
1243 lruvec = &pgdat->lruvec;
1244 goto out;
1245 }
1246
1247 memcg = page->mem_cgroup;
1248 /*
1249 * Swapcache readahead pages are added to the LRU - and
1250 * possibly migrated - before they are charged.
1251 */
1252 if (!memcg)
1253 memcg = root_mem_cgroup;
1254
1255 mz = mem_cgroup_page_nodeinfo(memcg, page);
1256 lruvec = &mz->lruvec;
1257 out:
1258 /*
1259 * Since a node can be onlined after the mem_cgroup was created,
1260 * we have to be prepared to initialize lruvec->zone here;
1261 * and if offlined then reonlined, we need to reinitialize it.
1262 */
1263 if (unlikely(lruvec->pgdat != pgdat))
1264 lruvec->pgdat = pgdat;
1265 return lruvec;
1266 }
1267
1268 /**
1269 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1270 * @lruvec: mem_cgroup per zone lru vector
1271 * @lru: index of lru list the page is sitting on
1272 * @zid: zone id of the accounted pages
1273 * @nr_pages: positive when adding or negative when removing
1274 *
1275 * This function must be called under lru_lock, just before a page is added
1276 * to or just after a page is removed from an lru list (that ordering being
1277 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1278 */
1279 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1280 int zid, int nr_pages)
1281 {
1282 struct mem_cgroup_per_node *mz;
1283 unsigned long *lru_size;
1284 long size;
1285
1286 if (mem_cgroup_disabled())
1287 return;
1288
1289 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1290 lru_size = &mz->lru_zone_size[zid][lru];
1291
1292 if (nr_pages < 0)
1293 *lru_size += nr_pages;
1294
1295 size = *lru_size;
1296 if (WARN_ONCE(size < 0,
1297 "%s(%p, %d, %d): lru_size %ld\n",
1298 __func__, lruvec, lru, nr_pages, size)) {
1299 VM_BUG_ON(1);
1300 *lru_size = 0;
1301 }
1302
1303 if (nr_pages > 0)
1304 *lru_size += nr_pages;
1305 }
1306
1307 /**
1308 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1309 * @memcg: the memory cgroup
1310 *
1311 * Returns the maximum amount of memory @mem can be charged with, in
1312 * pages.
1313 */
1314 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1315 {
1316 unsigned long margin = 0;
1317 unsigned long count;
1318 unsigned long limit;
1319
1320 count = page_counter_read(&memcg->memory);
1321 limit = READ_ONCE(memcg->memory.max);
1322 if (count < limit)
1323 margin = limit - count;
1324
1325 if (do_memsw_account()) {
1326 count = page_counter_read(&memcg->memsw);
1327 limit = READ_ONCE(memcg->memsw.max);
1328 if (count <= limit)
1329 margin = min(margin, limit - count);
1330 else
1331 margin = 0;
1332 }
1333
1334 return margin;
1335 }
1336
1337 /*
1338 * A routine for checking "mem" is under move_account() or not.
1339 *
1340 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1341 * moving cgroups. This is for waiting at high-memory pressure
1342 * caused by "move".
1343 */
1344 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1345 {
1346 struct mem_cgroup *from;
1347 struct mem_cgroup *to;
1348 bool ret = false;
1349 /*
1350 * Unlike task_move routines, we access mc.to, mc.from not under
1351 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1352 */
1353 spin_lock(&mc.lock);
1354 from = mc.from;
1355 to = mc.to;
1356 if (!from)
1357 goto unlock;
1358
1359 ret = mem_cgroup_is_descendant(from, memcg) ||
1360 mem_cgroup_is_descendant(to, memcg);
1361 unlock:
1362 spin_unlock(&mc.lock);
1363 return ret;
1364 }
1365
1366 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1367 {
1368 if (mc.moving_task && current != mc.moving_task) {
1369 if (mem_cgroup_under_move(memcg)) {
1370 DEFINE_WAIT(wait);
1371 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1372 /* moving charge context might have finished. */
1373 if (mc.moving_task)
1374 schedule();
1375 finish_wait(&mc.waitq, &wait);
1376 return true;
1377 }
1378 }
1379 return false;
1380 }
1381
1382 static char *memory_stat_format(struct mem_cgroup *memcg)
1383 {
1384 struct seq_buf s;
1385 int i;
1386
1387 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1388 if (!s.buffer)
1389 return NULL;
1390
1391 /*
1392 * Provide statistics on the state of the memory subsystem as
1393 * well as cumulative event counters that show past behavior.
1394 *
1395 * This list is ordered following a combination of these gradients:
1396 * 1) generic big picture -> specifics and details
1397 * 2) reflecting userspace activity -> reflecting kernel heuristics
1398 *
1399 * Current memory state:
1400 */
1401
1402 seq_buf_printf(&s, "anon %llu\n",
1403 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1404 PAGE_SIZE);
1405 seq_buf_printf(&s, "file %llu\n",
1406 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1407 PAGE_SIZE);
1408 seq_buf_printf(&s, "kernel_stack %llu\n",
1409 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1410 1024);
1411 seq_buf_printf(&s, "slab %llu\n",
1412 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1413 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1414 PAGE_SIZE);
1415 seq_buf_printf(&s, "sock %llu\n",
1416 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1417 PAGE_SIZE);
1418
1419 seq_buf_printf(&s, "shmem %llu\n",
1420 (u64)memcg_page_state(memcg, NR_SHMEM) *
1421 PAGE_SIZE);
1422 seq_buf_printf(&s, "file_mapped %llu\n",
1423 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1424 PAGE_SIZE);
1425 seq_buf_printf(&s, "file_dirty %llu\n",
1426 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1427 PAGE_SIZE);
1428 seq_buf_printf(&s, "file_writeback %llu\n",
1429 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1430 PAGE_SIZE);
1431
1432 /*
1433 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1434 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1435 * arse because it requires migrating the work out of rmap to a place
1436 * where the page->mem_cgroup is set up and stable.
1437 */
1438 seq_buf_printf(&s, "anon_thp %llu\n",
1439 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1440 PAGE_SIZE);
1441
1442 for (i = 0; i < NR_LRU_LISTS; i++)
1443 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1444 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1445 PAGE_SIZE);
1446
1447 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1448 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1449 PAGE_SIZE);
1450 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1451 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1452 PAGE_SIZE);
1453
1454 /* Accumulated memory events */
1455
1456 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1457 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1458
1459 seq_buf_printf(&s, "workingset_refault %lu\n",
1460 memcg_page_state(memcg, WORKINGSET_REFAULT));
1461 seq_buf_printf(&s, "workingset_activate %lu\n",
1462 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1463 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1464 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1465
1466 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1467 seq_buf_printf(&s, "pgscan %lu\n",
1468 memcg_events(memcg, PGSCAN_KSWAPD) +
1469 memcg_events(memcg, PGSCAN_DIRECT));
1470 seq_buf_printf(&s, "pgsteal %lu\n",
1471 memcg_events(memcg, PGSTEAL_KSWAPD) +
1472 memcg_events(memcg, PGSTEAL_DIRECT));
1473 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1474 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1475 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1476 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1477
1478 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1479 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1480 memcg_events(memcg, THP_FAULT_ALLOC));
1481 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1482 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1483 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1484
1485 /* The above should easily fit into one page */
1486 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1487
1488 return s.buffer;
1489 }
1490
1491 #define K(x) ((x) << (PAGE_SHIFT-10))
1492 /**
1493 * mem_cgroup_print_oom_context: Print OOM information relevant to
1494 * memory controller.
1495 * @memcg: The memory cgroup that went over limit
1496 * @p: Task that is going to be killed
1497 *
1498 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1499 * enabled
1500 */
1501 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1502 {
1503 rcu_read_lock();
1504
1505 if (memcg) {
1506 pr_cont(",oom_memcg=");
1507 pr_cont_cgroup_path(memcg->css.cgroup);
1508 } else
1509 pr_cont(",global_oom");
1510 if (p) {
1511 pr_cont(",task_memcg=");
1512 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1513 }
1514 rcu_read_unlock();
1515 }
1516
1517 /**
1518 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1519 * memory controller.
1520 * @memcg: The memory cgroup that went over limit
1521 */
1522 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1523 {
1524 char *buf;
1525
1526 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1527 K((u64)page_counter_read(&memcg->memory)),
1528 K((u64)memcg->memory.max), memcg->memory.failcnt);
1529 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1530 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64)page_counter_read(&memcg->swap)),
1532 K((u64)memcg->swap.max), memcg->swap.failcnt);
1533 else {
1534 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1535 K((u64)page_counter_read(&memcg->memsw)),
1536 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1537 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1538 K((u64)page_counter_read(&memcg->kmem)),
1539 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1540 }
1541
1542 pr_info("Memory cgroup stats for ");
1543 pr_cont_cgroup_path(memcg->css.cgroup);
1544 pr_cont(":");
1545 buf = memory_stat_format(memcg);
1546 if (!buf)
1547 return;
1548 pr_info("%s", buf);
1549 kfree(buf);
1550 }
1551
1552 /*
1553 * Return the memory (and swap, if configured) limit for a memcg.
1554 */
1555 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1556 {
1557 unsigned long max;
1558
1559 max = memcg->memory.max;
1560 if (mem_cgroup_swappiness(memcg)) {
1561 unsigned long memsw_max;
1562 unsigned long swap_max;
1563
1564 memsw_max = memcg->memsw.max;
1565 swap_max = memcg->swap.max;
1566 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1567 max = min(max + swap_max, memsw_max);
1568 }
1569 return max;
1570 }
1571
1572 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1573 int order)
1574 {
1575 struct oom_control oc = {
1576 .zonelist = NULL,
1577 .nodemask = NULL,
1578 .memcg = memcg,
1579 .gfp_mask = gfp_mask,
1580 .order = order,
1581 };
1582 bool ret;
1583
1584 if (mutex_lock_killable(&oom_lock))
1585 return true;
1586 /*
1587 * A few threads which were not waiting at mutex_lock_killable() can
1588 * fail to bail out. Therefore, check again after holding oom_lock.
1589 */
1590 ret = should_force_charge() || out_of_memory(&oc);
1591 mutex_unlock(&oom_lock);
1592 return ret;
1593 }
1594
1595 #if MAX_NUMNODES > 1
1596
1597 /**
1598 * test_mem_cgroup_node_reclaimable
1599 * @memcg: the target memcg
1600 * @nid: the node ID to be checked.
1601 * @noswap : specify true here if the user wants flle only information.
1602 *
1603 * This function returns whether the specified memcg contains any
1604 * reclaimable pages on a node. Returns true if there are any reclaimable
1605 * pages in the node.
1606 */
1607 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1608 int nid, bool noswap)
1609 {
1610 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
1611
1612 if (lruvec_page_state(lruvec, NR_INACTIVE_FILE) ||
1613 lruvec_page_state(lruvec, NR_ACTIVE_FILE))
1614 return true;
1615 if (noswap || !total_swap_pages)
1616 return false;
1617 if (lruvec_page_state(lruvec, NR_INACTIVE_ANON) ||
1618 lruvec_page_state(lruvec, NR_ACTIVE_ANON))
1619 return true;
1620 return false;
1621
1622 }
1623
1624 /*
1625 * Always updating the nodemask is not very good - even if we have an empty
1626 * list or the wrong list here, we can start from some node and traverse all
1627 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1628 *
1629 */
1630 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1631 {
1632 int nid;
1633 /*
1634 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1635 * pagein/pageout changes since the last update.
1636 */
1637 if (!atomic_read(&memcg->numainfo_events))
1638 return;
1639 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1640 return;
1641
1642 /* make a nodemask where this memcg uses memory from */
1643 memcg->scan_nodes = node_states[N_MEMORY];
1644
1645 for_each_node_mask(nid, node_states[N_MEMORY]) {
1646
1647 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1648 node_clear(nid, memcg->scan_nodes);
1649 }
1650
1651 atomic_set(&memcg->numainfo_events, 0);
1652 atomic_set(&memcg->numainfo_updating, 0);
1653 }
1654
1655 /*
1656 * Selecting a node where we start reclaim from. Because what we need is just
1657 * reducing usage counter, start from anywhere is O,K. Considering
1658 * memory reclaim from current node, there are pros. and cons.
1659 *
1660 * Freeing memory from current node means freeing memory from a node which
1661 * we'll use or we've used. So, it may make LRU bad. And if several threads
1662 * hit limits, it will see a contention on a node. But freeing from remote
1663 * node means more costs for memory reclaim because of memory latency.
1664 *
1665 * Now, we use round-robin. Better algorithm is welcomed.
1666 */
1667 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1668 {
1669 int node;
1670
1671 mem_cgroup_may_update_nodemask(memcg);
1672 node = memcg->last_scanned_node;
1673
1674 node = next_node_in(node, memcg->scan_nodes);
1675 /*
1676 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1677 * last time it really checked all the LRUs due to rate limiting.
1678 * Fallback to the current node in that case for simplicity.
1679 */
1680 if (unlikely(node == MAX_NUMNODES))
1681 node = numa_node_id();
1682
1683 memcg->last_scanned_node = node;
1684 return node;
1685 }
1686 #else
1687 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1688 {
1689 return 0;
1690 }
1691 #endif
1692
1693 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1694 pg_data_t *pgdat,
1695 gfp_t gfp_mask,
1696 unsigned long *total_scanned)
1697 {
1698 struct mem_cgroup *victim = NULL;
1699 int total = 0;
1700 int loop = 0;
1701 unsigned long excess;
1702 unsigned long nr_scanned;
1703 struct mem_cgroup_reclaim_cookie reclaim = {
1704 .pgdat = pgdat,
1705 .priority = 0,
1706 };
1707
1708 excess = soft_limit_excess(root_memcg);
1709
1710 while (1) {
1711 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1712 if (!victim) {
1713 loop++;
1714 if (loop >= 2) {
1715 /*
1716 * If we have not been able to reclaim
1717 * anything, it might because there are
1718 * no reclaimable pages under this hierarchy
1719 */
1720 if (!total)
1721 break;
1722 /*
1723 * We want to do more targeted reclaim.
1724 * excess >> 2 is not to excessive so as to
1725 * reclaim too much, nor too less that we keep
1726 * coming back to reclaim from this cgroup
1727 */
1728 if (total >= (excess >> 2) ||
1729 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1730 break;
1731 }
1732 continue;
1733 }
1734 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1735 pgdat, &nr_scanned);
1736 *total_scanned += nr_scanned;
1737 if (!soft_limit_excess(root_memcg))
1738 break;
1739 }
1740 mem_cgroup_iter_break(root_memcg, victim);
1741 return total;
1742 }
1743
1744 #ifdef CONFIG_LOCKDEP
1745 static struct lockdep_map memcg_oom_lock_dep_map = {
1746 .name = "memcg_oom_lock",
1747 };
1748 #endif
1749
1750 static DEFINE_SPINLOCK(memcg_oom_lock);
1751
1752 /*
1753 * Check OOM-Killer is already running under our hierarchy.
1754 * If someone is running, return false.
1755 */
1756 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1757 {
1758 struct mem_cgroup *iter, *failed = NULL;
1759
1760 spin_lock(&memcg_oom_lock);
1761
1762 for_each_mem_cgroup_tree(iter, memcg) {
1763 if (iter->oom_lock) {
1764 /*
1765 * this subtree of our hierarchy is already locked
1766 * so we cannot give a lock.
1767 */
1768 failed = iter;
1769 mem_cgroup_iter_break(memcg, iter);
1770 break;
1771 } else
1772 iter->oom_lock = true;
1773 }
1774
1775 if (failed) {
1776 /*
1777 * OK, we failed to lock the whole subtree so we have
1778 * to clean up what we set up to the failing subtree
1779 */
1780 for_each_mem_cgroup_tree(iter, memcg) {
1781 if (iter == failed) {
1782 mem_cgroup_iter_break(memcg, iter);
1783 break;
1784 }
1785 iter->oom_lock = false;
1786 }
1787 } else
1788 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1789
1790 spin_unlock(&memcg_oom_lock);
1791
1792 return !failed;
1793 }
1794
1795 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1796 {
1797 struct mem_cgroup *iter;
1798
1799 spin_lock(&memcg_oom_lock);
1800 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1801 for_each_mem_cgroup_tree(iter, memcg)
1802 iter->oom_lock = false;
1803 spin_unlock(&memcg_oom_lock);
1804 }
1805
1806 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1807 {
1808 struct mem_cgroup *iter;
1809
1810 spin_lock(&memcg_oom_lock);
1811 for_each_mem_cgroup_tree(iter, memcg)
1812 iter->under_oom++;
1813 spin_unlock(&memcg_oom_lock);
1814 }
1815
1816 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1817 {
1818 struct mem_cgroup *iter;
1819
1820 /*
1821 * When a new child is created while the hierarchy is under oom,
1822 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1823 */
1824 spin_lock(&memcg_oom_lock);
1825 for_each_mem_cgroup_tree(iter, memcg)
1826 if (iter->under_oom > 0)
1827 iter->under_oom--;
1828 spin_unlock(&memcg_oom_lock);
1829 }
1830
1831 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1832
1833 struct oom_wait_info {
1834 struct mem_cgroup *memcg;
1835 wait_queue_entry_t wait;
1836 };
1837
1838 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1839 unsigned mode, int sync, void *arg)
1840 {
1841 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1842 struct mem_cgroup *oom_wait_memcg;
1843 struct oom_wait_info *oom_wait_info;
1844
1845 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1846 oom_wait_memcg = oom_wait_info->memcg;
1847
1848 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1849 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1850 return 0;
1851 return autoremove_wake_function(wait, mode, sync, arg);
1852 }
1853
1854 static void memcg_oom_recover(struct mem_cgroup *memcg)
1855 {
1856 /*
1857 * For the following lockless ->under_oom test, the only required
1858 * guarantee is that it must see the state asserted by an OOM when
1859 * this function is called as a result of userland actions
1860 * triggered by the notification of the OOM. This is trivially
1861 * achieved by invoking mem_cgroup_mark_under_oom() before
1862 * triggering notification.
1863 */
1864 if (memcg && memcg->under_oom)
1865 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1866 }
1867
1868 enum oom_status {
1869 OOM_SUCCESS,
1870 OOM_FAILED,
1871 OOM_ASYNC,
1872 OOM_SKIPPED
1873 };
1874
1875 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1876 {
1877 enum oom_status ret;
1878 bool locked;
1879
1880 if (order > PAGE_ALLOC_COSTLY_ORDER)
1881 return OOM_SKIPPED;
1882
1883 memcg_memory_event(memcg, MEMCG_OOM);
1884
1885 /*
1886 * We are in the middle of the charge context here, so we
1887 * don't want to block when potentially sitting on a callstack
1888 * that holds all kinds of filesystem and mm locks.
1889 *
1890 * cgroup1 allows disabling the OOM killer and waiting for outside
1891 * handling until the charge can succeed; remember the context and put
1892 * the task to sleep at the end of the page fault when all locks are
1893 * released.
1894 *
1895 * On the other hand, in-kernel OOM killer allows for an async victim
1896 * memory reclaim (oom_reaper) and that means that we are not solely
1897 * relying on the oom victim to make a forward progress and we can
1898 * invoke the oom killer here.
1899 *
1900 * Please note that mem_cgroup_out_of_memory might fail to find a
1901 * victim and then we have to bail out from the charge path.
1902 */
1903 if (memcg->oom_kill_disable) {
1904 if (!current->in_user_fault)
1905 return OOM_SKIPPED;
1906 css_get(&memcg->css);
1907 current->memcg_in_oom = memcg;
1908 current->memcg_oom_gfp_mask = mask;
1909 current->memcg_oom_order = order;
1910
1911 return OOM_ASYNC;
1912 }
1913
1914 mem_cgroup_mark_under_oom(memcg);
1915
1916 locked = mem_cgroup_oom_trylock(memcg);
1917
1918 if (locked)
1919 mem_cgroup_oom_notify(memcg);
1920
1921 mem_cgroup_unmark_under_oom(memcg);
1922 if (mem_cgroup_out_of_memory(memcg, mask, order))
1923 ret = OOM_SUCCESS;
1924 else
1925 ret = OOM_FAILED;
1926
1927 if (locked)
1928 mem_cgroup_oom_unlock(memcg);
1929
1930 return ret;
1931 }
1932
1933 /**
1934 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1935 * @handle: actually kill/wait or just clean up the OOM state
1936 *
1937 * This has to be called at the end of a page fault if the memcg OOM
1938 * handler was enabled.
1939 *
1940 * Memcg supports userspace OOM handling where failed allocations must
1941 * sleep on a waitqueue until the userspace task resolves the
1942 * situation. Sleeping directly in the charge context with all kinds
1943 * of locks held is not a good idea, instead we remember an OOM state
1944 * in the task and mem_cgroup_oom_synchronize() has to be called at
1945 * the end of the page fault to complete the OOM handling.
1946 *
1947 * Returns %true if an ongoing memcg OOM situation was detected and
1948 * completed, %false otherwise.
1949 */
1950 bool mem_cgroup_oom_synchronize(bool handle)
1951 {
1952 struct mem_cgroup *memcg = current->memcg_in_oom;
1953 struct oom_wait_info owait;
1954 bool locked;
1955
1956 /* OOM is global, do not handle */
1957 if (!memcg)
1958 return false;
1959
1960 if (!handle)
1961 goto cleanup;
1962
1963 owait.memcg = memcg;
1964 owait.wait.flags = 0;
1965 owait.wait.func = memcg_oom_wake_function;
1966 owait.wait.private = current;
1967 INIT_LIST_HEAD(&owait.wait.entry);
1968
1969 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1970 mem_cgroup_mark_under_oom(memcg);
1971
1972 locked = mem_cgroup_oom_trylock(memcg);
1973
1974 if (locked)
1975 mem_cgroup_oom_notify(memcg);
1976
1977 if (locked && !memcg->oom_kill_disable) {
1978 mem_cgroup_unmark_under_oom(memcg);
1979 finish_wait(&memcg_oom_waitq, &owait.wait);
1980 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1981 current->memcg_oom_order);
1982 } else {
1983 schedule();
1984 mem_cgroup_unmark_under_oom(memcg);
1985 finish_wait(&memcg_oom_waitq, &owait.wait);
1986 }
1987
1988 if (locked) {
1989 mem_cgroup_oom_unlock(memcg);
1990 /*
1991 * There is no guarantee that an OOM-lock contender
1992 * sees the wakeups triggered by the OOM kill
1993 * uncharges. Wake any sleepers explicitely.
1994 */
1995 memcg_oom_recover(memcg);
1996 }
1997 cleanup:
1998 current->memcg_in_oom = NULL;
1999 css_put(&memcg->css);
2000 return true;
2001 }
2002
2003 /**
2004 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2005 * @victim: task to be killed by the OOM killer
2006 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2007 *
2008 * Returns a pointer to a memory cgroup, which has to be cleaned up
2009 * by killing all belonging OOM-killable tasks.
2010 *
2011 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2012 */
2013 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2014 struct mem_cgroup *oom_domain)
2015 {
2016 struct mem_cgroup *oom_group = NULL;
2017 struct mem_cgroup *memcg;
2018
2019 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2020 return NULL;
2021
2022 if (!oom_domain)
2023 oom_domain = root_mem_cgroup;
2024
2025 rcu_read_lock();
2026
2027 memcg = mem_cgroup_from_task(victim);
2028 if (memcg == root_mem_cgroup)
2029 goto out;
2030
2031 /*
2032 * Traverse the memory cgroup hierarchy from the victim task's
2033 * cgroup up to the OOMing cgroup (or root) to find the
2034 * highest-level memory cgroup with oom.group set.
2035 */
2036 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2037 if (memcg->oom_group)
2038 oom_group = memcg;
2039
2040 if (memcg == oom_domain)
2041 break;
2042 }
2043
2044 if (oom_group)
2045 css_get(&oom_group->css);
2046 out:
2047 rcu_read_unlock();
2048
2049 return oom_group;
2050 }
2051
2052 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2053 {
2054 pr_info("Tasks in ");
2055 pr_cont_cgroup_path(memcg->css.cgroup);
2056 pr_cont(" are going to be killed due to memory.oom.group set\n");
2057 }
2058
2059 /**
2060 * lock_page_memcg - lock a page->mem_cgroup binding
2061 * @page: the page
2062 *
2063 * This function protects unlocked LRU pages from being moved to
2064 * another cgroup.
2065 *
2066 * It ensures lifetime of the returned memcg. Caller is responsible
2067 * for the lifetime of the page; __unlock_page_memcg() is available
2068 * when @page might get freed inside the locked section.
2069 */
2070 struct mem_cgroup *lock_page_memcg(struct page *page)
2071 {
2072 struct mem_cgroup *memcg;
2073 unsigned long flags;
2074
2075 /*
2076 * The RCU lock is held throughout the transaction. The fast
2077 * path can get away without acquiring the memcg->move_lock
2078 * because page moving starts with an RCU grace period.
2079 *
2080 * The RCU lock also protects the memcg from being freed when
2081 * the page state that is going to change is the only thing
2082 * preventing the page itself from being freed. E.g. writeback
2083 * doesn't hold a page reference and relies on PG_writeback to
2084 * keep off truncation, migration and so forth.
2085 */
2086 rcu_read_lock();
2087
2088 if (mem_cgroup_disabled())
2089 return NULL;
2090 again:
2091 memcg = page->mem_cgroup;
2092 if (unlikely(!memcg))
2093 return NULL;
2094
2095 if (atomic_read(&memcg->moving_account) <= 0)
2096 return memcg;
2097
2098 spin_lock_irqsave(&memcg->move_lock, flags);
2099 if (memcg != page->mem_cgroup) {
2100 spin_unlock_irqrestore(&memcg->move_lock, flags);
2101 goto again;
2102 }
2103
2104 /*
2105 * When charge migration first begins, we can have locked and
2106 * unlocked page stat updates happening concurrently. Track
2107 * the task who has the lock for unlock_page_memcg().
2108 */
2109 memcg->move_lock_task = current;
2110 memcg->move_lock_flags = flags;
2111
2112 return memcg;
2113 }
2114 EXPORT_SYMBOL(lock_page_memcg);
2115
2116 /**
2117 * __unlock_page_memcg - unlock and unpin a memcg
2118 * @memcg: the memcg
2119 *
2120 * Unlock and unpin a memcg returned by lock_page_memcg().
2121 */
2122 void __unlock_page_memcg(struct mem_cgroup *memcg)
2123 {
2124 if (memcg && memcg->move_lock_task == current) {
2125 unsigned long flags = memcg->move_lock_flags;
2126
2127 memcg->move_lock_task = NULL;
2128 memcg->move_lock_flags = 0;
2129
2130 spin_unlock_irqrestore(&memcg->move_lock, flags);
2131 }
2132
2133 rcu_read_unlock();
2134 }
2135
2136 /**
2137 * unlock_page_memcg - unlock a page->mem_cgroup binding
2138 * @page: the page
2139 */
2140 void unlock_page_memcg(struct page *page)
2141 {
2142 __unlock_page_memcg(page->mem_cgroup);
2143 }
2144 EXPORT_SYMBOL(unlock_page_memcg);
2145
2146 struct memcg_stock_pcp {
2147 struct mem_cgroup *cached; /* this never be root cgroup */
2148 unsigned int nr_pages;
2149 struct work_struct work;
2150 unsigned long flags;
2151 #define FLUSHING_CACHED_CHARGE 0
2152 };
2153 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2154 static DEFINE_MUTEX(percpu_charge_mutex);
2155
2156 /**
2157 * consume_stock: Try to consume stocked charge on this cpu.
2158 * @memcg: memcg to consume from.
2159 * @nr_pages: how many pages to charge.
2160 *
2161 * The charges will only happen if @memcg matches the current cpu's memcg
2162 * stock, and at least @nr_pages are available in that stock. Failure to
2163 * service an allocation will refill the stock.
2164 *
2165 * returns true if successful, false otherwise.
2166 */
2167 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2168 {
2169 struct memcg_stock_pcp *stock;
2170 unsigned long flags;
2171 bool ret = false;
2172
2173 if (nr_pages > MEMCG_CHARGE_BATCH)
2174 return ret;
2175
2176 local_irq_save(flags);
2177
2178 stock = this_cpu_ptr(&memcg_stock);
2179 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2180 stock->nr_pages -= nr_pages;
2181 ret = true;
2182 }
2183
2184 local_irq_restore(flags);
2185
2186 return ret;
2187 }
2188
2189 /*
2190 * Returns stocks cached in percpu and reset cached information.
2191 */
2192 static void drain_stock(struct memcg_stock_pcp *stock)
2193 {
2194 struct mem_cgroup *old = stock->cached;
2195
2196 if (stock->nr_pages) {
2197 page_counter_uncharge(&old->memory, stock->nr_pages);
2198 if (do_memsw_account())
2199 page_counter_uncharge(&old->memsw, stock->nr_pages);
2200 css_put_many(&old->css, stock->nr_pages);
2201 stock->nr_pages = 0;
2202 }
2203 stock->cached = NULL;
2204 }
2205
2206 static void drain_local_stock(struct work_struct *dummy)
2207 {
2208 struct memcg_stock_pcp *stock;
2209 unsigned long flags;
2210
2211 /*
2212 * The only protection from memory hotplug vs. drain_stock races is
2213 * that we always operate on local CPU stock here with IRQ disabled
2214 */
2215 local_irq_save(flags);
2216
2217 stock = this_cpu_ptr(&memcg_stock);
2218 drain_stock(stock);
2219 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2220
2221 local_irq_restore(flags);
2222 }
2223
2224 /*
2225 * Cache charges(val) to local per_cpu area.
2226 * This will be consumed by consume_stock() function, later.
2227 */
2228 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2229 {
2230 struct memcg_stock_pcp *stock;
2231 unsigned long flags;
2232
2233 local_irq_save(flags);
2234
2235 stock = this_cpu_ptr(&memcg_stock);
2236 if (stock->cached != memcg) { /* reset if necessary */
2237 drain_stock(stock);
2238 stock->cached = memcg;
2239 }
2240 stock->nr_pages += nr_pages;
2241
2242 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2243 drain_stock(stock);
2244
2245 local_irq_restore(flags);
2246 }
2247
2248 /*
2249 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2250 * of the hierarchy under it.
2251 */
2252 static void drain_all_stock(struct mem_cgroup *root_memcg)
2253 {
2254 int cpu, curcpu;
2255
2256 /* If someone's already draining, avoid adding running more workers. */
2257 if (!mutex_trylock(&percpu_charge_mutex))
2258 return;
2259 /*
2260 * Notify other cpus that system-wide "drain" is running
2261 * We do not care about races with the cpu hotplug because cpu down
2262 * as well as workers from this path always operate on the local
2263 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2264 */
2265 curcpu = get_cpu();
2266 for_each_online_cpu(cpu) {
2267 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2268 struct mem_cgroup *memcg;
2269
2270 memcg = stock->cached;
2271 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
2272 continue;
2273 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
2274 css_put(&memcg->css);
2275 continue;
2276 }
2277 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2278 if (cpu == curcpu)
2279 drain_local_stock(&stock->work);
2280 else
2281 schedule_work_on(cpu, &stock->work);
2282 }
2283 css_put(&memcg->css);
2284 }
2285 put_cpu();
2286 mutex_unlock(&percpu_charge_mutex);
2287 }
2288
2289 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2290 {
2291 struct memcg_stock_pcp *stock;
2292 struct mem_cgroup *memcg, *mi;
2293
2294 stock = &per_cpu(memcg_stock, cpu);
2295 drain_stock(stock);
2296
2297 for_each_mem_cgroup(memcg) {
2298 int i;
2299
2300 for (i = 0; i < MEMCG_NR_STAT; i++) {
2301 int nid;
2302 long x;
2303
2304 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2305 if (x)
2306 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2307 atomic_long_add(x, &memcg->vmstats[i]);
2308
2309 if (i >= NR_VM_NODE_STAT_ITEMS)
2310 continue;
2311
2312 for_each_node(nid) {
2313 struct mem_cgroup_per_node *pn;
2314
2315 pn = mem_cgroup_nodeinfo(memcg, nid);
2316 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2317 if (x)
2318 do {
2319 atomic_long_add(x, &pn->lruvec_stat[i]);
2320 } while ((pn = parent_nodeinfo(pn, nid)));
2321 }
2322 }
2323
2324 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2325 long x;
2326
2327 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2328 if (x)
2329 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2330 atomic_long_add(x, &memcg->vmevents[i]);
2331 }
2332 }
2333
2334 return 0;
2335 }
2336
2337 static void reclaim_high(struct mem_cgroup *memcg,
2338 unsigned int nr_pages,
2339 gfp_t gfp_mask)
2340 {
2341 do {
2342 if (page_counter_read(&memcg->memory) <= memcg->high)
2343 continue;
2344 memcg_memory_event(memcg, MEMCG_HIGH);
2345 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2346 } while ((memcg = parent_mem_cgroup(memcg)));
2347 }
2348
2349 static void high_work_func(struct work_struct *work)
2350 {
2351 struct mem_cgroup *memcg;
2352
2353 memcg = container_of(work, struct mem_cgroup, high_work);
2354 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2355 }
2356
2357 /*
2358 * Scheduled by try_charge() to be executed from the userland return path
2359 * and reclaims memory over the high limit.
2360 */
2361 void mem_cgroup_handle_over_high(void)
2362 {
2363 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2364 struct mem_cgroup *memcg;
2365
2366 if (likely(!nr_pages))
2367 return;
2368
2369 memcg = get_mem_cgroup_from_mm(current->mm);
2370 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2371 css_put(&memcg->css);
2372 current->memcg_nr_pages_over_high = 0;
2373 }
2374
2375 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2376 unsigned int nr_pages)
2377 {
2378 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2379 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2380 struct mem_cgroup *mem_over_limit;
2381 struct page_counter *counter;
2382 unsigned long nr_reclaimed;
2383 bool may_swap = true;
2384 bool drained = false;
2385 enum oom_status oom_status;
2386
2387 if (mem_cgroup_is_root(memcg))
2388 return 0;
2389 retry:
2390 if (consume_stock(memcg, nr_pages))
2391 return 0;
2392
2393 if (!do_memsw_account() ||
2394 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2395 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2396 goto done_restock;
2397 if (do_memsw_account())
2398 page_counter_uncharge(&memcg->memsw, batch);
2399 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2400 } else {
2401 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2402 may_swap = false;
2403 }
2404
2405 if (batch > nr_pages) {
2406 batch = nr_pages;
2407 goto retry;
2408 }
2409
2410 /*
2411 * Unlike in global OOM situations, memcg is not in a physical
2412 * memory shortage. Allow dying and OOM-killed tasks to
2413 * bypass the last charges so that they can exit quickly and
2414 * free their memory.
2415 */
2416 if (unlikely(should_force_charge()))
2417 goto force;
2418
2419 /*
2420 * Prevent unbounded recursion when reclaim operations need to
2421 * allocate memory. This might exceed the limits temporarily,
2422 * but we prefer facilitating memory reclaim and getting back
2423 * under the limit over triggering OOM kills in these cases.
2424 */
2425 if (unlikely(current->flags & PF_MEMALLOC))
2426 goto force;
2427
2428 if (unlikely(task_in_memcg_oom(current)))
2429 goto nomem;
2430
2431 if (!gfpflags_allow_blocking(gfp_mask))
2432 goto nomem;
2433
2434 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2435
2436 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2437 gfp_mask, may_swap);
2438
2439 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2440 goto retry;
2441
2442 if (!drained) {
2443 drain_all_stock(mem_over_limit);
2444 drained = true;
2445 goto retry;
2446 }
2447
2448 if (gfp_mask & __GFP_NORETRY)
2449 goto nomem;
2450 /*
2451 * Even though the limit is exceeded at this point, reclaim
2452 * may have been able to free some pages. Retry the charge
2453 * before killing the task.
2454 *
2455 * Only for regular pages, though: huge pages are rather
2456 * unlikely to succeed so close to the limit, and we fall back
2457 * to regular pages anyway in case of failure.
2458 */
2459 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2460 goto retry;
2461 /*
2462 * At task move, charge accounts can be doubly counted. So, it's
2463 * better to wait until the end of task_move if something is going on.
2464 */
2465 if (mem_cgroup_wait_acct_move(mem_over_limit))
2466 goto retry;
2467
2468 if (nr_retries--)
2469 goto retry;
2470
2471 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2472 goto nomem;
2473
2474 if (gfp_mask & __GFP_NOFAIL)
2475 goto force;
2476
2477 if (fatal_signal_pending(current))
2478 goto force;
2479
2480 /*
2481 * keep retrying as long as the memcg oom killer is able to make
2482 * a forward progress or bypass the charge if the oom killer
2483 * couldn't make any progress.
2484 */
2485 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2486 get_order(nr_pages * PAGE_SIZE));
2487 switch (oom_status) {
2488 case OOM_SUCCESS:
2489 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2490 goto retry;
2491 case OOM_FAILED:
2492 goto force;
2493 default:
2494 goto nomem;
2495 }
2496 nomem:
2497 if (!(gfp_mask & __GFP_NOFAIL))
2498 return -ENOMEM;
2499 force:
2500 /*
2501 * The allocation either can't fail or will lead to more memory
2502 * being freed very soon. Allow memory usage go over the limit
2503 * temporarily by force charging it.
2504 */
2505 page_counter_charge(&memcg->memory, nr_pages);
2506 if (do_memsw_account())
2507 page_counter_charge(&memcg->memsw, nr_pages);
2508 css_get_many(&memcg->css, nr_pages);
2509
2510 return 0;
2511
2512 done_restock:
2513 css_get_many(&memcg->css, batch);
2514 if (batch > nr_pages)
2515 refill_stock(memcg, batch - nr_pages);
2516
2517 /*
2518 * If the hierarchy is above the normal consumption range, schedule
2519 * reclaim on returning to userland. We can perform reclaim here
2520 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2521 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2522 * not recorded as it most likely matches current's and won't
2523 * change in the meantime. As high limit is checked again before
2524 * reclaim, the cost of mismatch is negligible.
2525 */
2526 do {
2527 if (page_counter_read(&memcg->memory) > memcg->high) {
2528 /* Don't bother a random interrupted task */
2529 if (in_interrupt()) {
2530 schedule_work(&memcg->high_work);
2531 break;
2532 }
2533 current->memcg_nr_pages_over_high += batch;
2534 set_notify_resume(current);
2535 break;
2536 }
2537 } while ((memcg = parent_mem_cgroup(memcg)));
2538
2539 return 0;
2540 }
2541
2542 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2543 {
2544 if (mem_cgroup_is_root(memcg))
2545 return;
2546
2547 page_counter_uncharge(&memcg->memory, nr_pages);
2548 if (do_memsw_account())
2549 page_counter_uncharge(&memcg->memsw, nr_pages);
2550
2551 css_put_many(&memcg->css, nr_pages);
2552 }
2553
2554 static void lock_page_lru(struct page *page, int *isolated)
2555 {
2556 pg_data_t *pgdat = page_pgdat(page);
2557
2558 spin_lock_irq(&pgdat->lru_lock);
2559 if (PageLRU(page)) {
2560 struct lruvec *lruvec;
2561
2562 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2563 ClearPageLRU(page);
2564 del_page_from_lru_list(page, lruvec, page_lru(page));
2565 *isolated = 1;
2566 } else
2567 *isolated = 0;
2568 }
2569
2570 static void unlock_page_lru(struct page *page, int isolated)
2571 {
2572 pg_data_t *pgdat = page_pgdat(page);
2573
2574 if (isolated) {
2575 struct lruvec *lruvec;
2576
2577 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2578 VM_BUG_ON_PAGE(PageLRU(page), page);
2579 SetPageLRU(page);
2580 add_page_to_lru_list(page, lruvec, page_lru(page));
2581 }
2582 spin_unlock_irq(&pgdat->lru_lock);
2583 }
2584
2585 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2586 bool lrucare)
2587 {
2588 int isolated;
2589
2590 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2591
2592 /*
2593 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2594 * may already be on some other mem_cgroup's LRU. Take care of it.
2595 */
2596 if (lrucare)
2597 lock_page_lru(page, &isolated);
2598
2599 /*
2600 * Nobody should be changing or seriously looking at
2601 * page->mem_cgroup at this point:
2602 *
2603 * - the page is uncharged
2604 *
2605 * - the page is off-LRU
2606 *
2607 * - an anonymous fault has exclusive page access, except for
2608 * a locked page table
2609 *
2610 * - a page cache insertion, a swapin fault, or a migration
2611 * have the page locked
2612 */
2613 page->mem_cgroup = memcg;
2614
2615 if (lrucare)
2616 unlock_page_lru(page, isolated);
2617 }
2618
2619 #ifdef CONFIG_MEMCG_KMEM
2620 static int memcg_alloc_cache_id(void)
2621 {
2622 int id, size;
2623 int err;
2624
2625 id = ida_simple_get(&memcg_cache_ida,
2626 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2627 if (id < 0)
2628 return id;
2629
2630 if (id < memcg_nr_cache_ids)
2631 return id;
2632
2633 /*
2634 * There's no space for the new id in memcg_caches arrays,
2635 * so we have to grow them.
2636 */
2637 down_write(&memcg_cache_ids_sem);
2638
2639 size = 2 * (id + 1);
2640 if (size < MEMCG_CACHES_MIN_SIZE)
2641 size = MEMCG_CACHES_MIN_SIZE;
2642 else if (size > MEMCG_CACHES_MAX_SIZE)
2643 size = MEMCG_CACHES_MAX_SIZE;
2644
2645 err = memcg_update_all_caches(size);
2646 if (!err)
2647 err = memcg_update_all_list_lrus(size);
2648 if (!err)
2649 memcg_nr_cache_ids = size;
2650
2651 up_write(&memcg_cache_ids_sem);
2652
2653 if (err) {
2654 ida_simple_remove(&memcg_cache_ida, id);
2655 return err;
2656 }
2657 return id;
2658 }
2659
2660 static void memcg_free_cache_id(int id)
2661 {
2662 ida_simple_remove(&memcg_cache_ida, id);
2663 }
2664
2665 struct memcg_kmem_cache_create_work {
2666 struct mem_cgroup *memcg;
2667 struct kmem_cache *cachep;
2668 struct work_struct work;
2669 };
2670
2671 static void memcg_kmem_cache_create_func(struct work_struct *w)
2672 {
2673 struct memcg_kmem_cache_create_work *cw =
2674 container_of(w, struct memcg_kmem_cache_create_work, work);
2675 struct mem_cgroup *memcg = cw->memcg;
2676 struct kmem_cache *cachep = cw->cachep;
2677
2678 memcg_create_kmem_cache(memcg, cachep);
2679
2680 css_put(&memcg->css);
2681 kfree(cw);
2682 }
2683
2684 /*
2685 * Enqueue the creation of a per-memcg kmem_cache.
2686 */
2687 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2688 struct kmem_cache *cachep)
2689 {
2690 struct memcg_kmem_cache_create_work *cw;
2691
2692 if (!css_tryget_online(&memcg->css))
2693 return;
2694
2695 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2696 if (!cw)
2697 return;
2698
2699 cw->memcg = memcg;
2700 cw->cachep = cachep;
2701 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2702
2703 queue_work(memcg_kmem_cache_wq, &cw->work);
2704 }
2705
2706 static inline bool memcg_kmem_bypass(void)
2707 {
2708 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2709 return true;
2710 return false;
2711 }
2712
2713 /**
2714 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2715 * @cachep: the original global kmem cache
2716 *
2717 * Return the kmem_cache we're supposed to use for a slab allocation.
2718 * We try to use the current memcg's version of the cache.
2719 *
2720 * If the cache does not exist yet, if we are the first user of it, we
2721 * create it asynchronously in a workqueue and let the current allocation
2722 * go through with the original cache.
2723 *
2724 * This function takes a reference to the cache it returns to assure it
2725 * won't get destroyed while we are working with it. Once the caller is
2726 * done with it, memcg_kmem_put_cache() must be called to release the
2727 * reference.
2728 */
2729 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2730 {
2731 struct mem_cgroup *memcg;
2732 struct kmem_cache *memcg_cachep;
2733 struct memcg_cache_array *arr;
2734 int kmemcg_id;
2735
2736 VM_BUG_ON(!is_root_cache(cachep));
2737
2738 if (memcg_kmem_bypass())
2739 return cachep;
2740
2741 rcu_read_lock();
2742
2743 if (unlikely(current->active_memcg))
2744 memcg = current->active_memcg;
2745 else
2746 memcg = mem_cgroup_from_task(current);
2747
2748 if (!memcg || memcg == root_mem_cgroup)
2749 goto out_unlock;
2750
2751 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2752 if (kmemcg_id < 0)
2753 goto out_unlock;
2754
2755 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2756
2757 /*
2758 * Make sure we will access the up-to-date value. The code updating
2759 * memcg_caches issues a write barrier to match the data dependency
2760 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2761 */
2762 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2763
2764 /*
2765 * If we are in a safe context (can wait, and not in interrupt
2766 * context), we could be be predictable and return right away.
2767 * This would guarantee that the allocation being performed
2768 * already belongs in the new cache.
2769 *
2770 * However, there are some clashes that can arrive from locking.
2771 * For instance, because we acquire the slab_mutex while doing
2772 * memcg_create_kmem_cache, this means no further allocation
2773 * could happen with the slab_mutex held. So it's better to
2774 * defer everything.
2775 *
2776 * If the memcg is dying or memcg_cache is about to be released,
2777 * don't bother creating new kmem_caches. Because memcg_cachep
2778 * is ZEROed as the fist step of kmem offlining, we don't need
2779 * percpu_ref_tryget_live() here. css_tryget_online() check in
2780 * memcg_schedule_kmem_cache_create() will prevent us from
2781 * creation of a new kmem_cache.
2782 */
2783 if (unlikely(!memcg_cachep))
2784 memcg_schedule_kmem_cache_create(memcg, cachep);
2785 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2786 cachep = memcg_cachep;
2787 out_unlock:
2788 rcu_read_unlock();
2789 return cachep;
2790 }
2791
2792 /**
2793 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2794 * @cachep: the cache returned by memcg_kmem_get_cache
2795 */
2796 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2797 {
2798 if (!is_root_cache(cachep))
2799 percpu_ref_put(&cachep->memcg_params.refcnt);
2800 }
2801
2802 /**
2803 * __memcg_kmem_charge_memcg: charge a kmem page
2804 * @page: page to charge
2805 * @gfp: reclaim mode
2806 * @order: allocation order
2807 * @memcg: memory cgroup to charge
2808 *
2809 * Returns 0 on success, an error code on failure.
2810 */
2811 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2812 struct mem_cgroup *memcg)
2813 {
2814 unsigned int nr_pages = 1 << order;
2815 struct page_counter *counter;
2816 int ret;
2817
2818 ret = try_charge(memcg, gfp, nr_pages);
2819 if (ret)
2820 return ret;
2821
2822 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2823 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2824 cancel_charge(memcg, nr_pages);
2825 return -ENOMEM;
2826 }
2827 return 0;
2828 }
2829
2830 /**
2831 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2832 * @page: page to charge
2833 * @gfp: reclaim mode
2834 * @order: allocation order
2835 *
2836 * Returns 0 on success, an error code on failure.
2837 */
2838 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2839 {
2840 struct mem_cgroup *memcg;
2841 int ret = 0;
2842
2843 if (memcg_kmem_bypass())
2844 return 0;
2845
2846 memcg = get_mem_cgroup_from_current();
2847 if (!mem_cgroup_is_root(memcg)) {
2848 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2849 if (!ret) {
2850 page->mem_cgroup = memcg;
2851 __SetPageKmemcg(page);
2852 }
2853 }
2854 css_put(&memcg->css);
2855 return ret;
2856 }
2857
2858 /**
2859 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2860 * @memcg: memcg to uncharge
2861 * @nr_pages: number of pages to uncharge
2862 */
2863 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2864 unsigned int nr_pages)
2865 {
2866 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2867 page_counter_uncharge(&memcg->kmem, nr_pages);
2868
2869 page_counter_uncharge(&memcg->memory, nr_pages);
2870 if (do_memsw_account())
2871 page_counter_uncharge(&memcg->memsw, nr_pages);
2872 }
2873 /**
2874 * __memcg_kmem_uncharge: uncharge a kmem page
2875 * @page: page to uncharge
2876 * @order: allocation order
2877 */
2878 void __memcg_kmem_uncharge(struct page *page, int order)
2879 {
2880 struct mem_cgroup *memcg = page->mem_cgroup;
2881 unsigned int nr_pages = 1 << order;
2882
2883 if (!memcg)
2884 return;
2885
2886 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2887 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2888 page->mem_cgroup = NULL;
2889
2890 /* slab pages do not have PageKmemcg flag set */
2891 if (PageKmemcg(page))
2892 __ClearPageKmemcg(page);
2893
2894 css_put_many(&memcg->css, nr_pages);
2895 }
2896 #endif /* CONFIG_MEMCG_KMEM */
2897
2898 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2899
2900 /*
2901 * Because tail pages are not marked as "used", set it. We're under
2902 * pgdat->lru_lock and migration entries setup in all page mappings.
2903 */
2904 void mem_cgroup_split_huge_fixup(struct page *head)
2905 {
2906 int i;
2907
2908 if (mem_cgroup_disabled())
2909 return;
2910
2911 for (i = 1; i < HPAGE_PMD_NR; i++)
2912 head[i].mem_cgroup = head->mem_cgroup;
2913
2914 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2915 }
2916 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2917
2918 #ifdef CONFIG_MEMCG_SWAP
2919 /**
2920 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2921 * @entry: swap entry to be moved
2922 * @from: mem_cgroup which the entry is moved from
2923 * @to: mem_cgroup which the entry is moved to
2924 *
2925 * It succeeds only when the swap_cgroup's record for this entry is the same
2926 * as the mem_cgroup's id of @from.
2927 *
2928 * Returns 0 on success, -EINVAL on failure.
2929 *
2930 * The caller must have charged to @to, IOW, called page_counter_charge() about
2931 * both res and memsw, and called css_get().
2932 */
2933 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2934 struct mem_cgroup *from, struct mem_cgroup *to)
2935 {
2936 unsigned short old_id, new_id;
2937
2938 old_id = mem_cgroup_id(from);
2939 new_id = mem_cgroup_id(to);
2940
2941 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2942 mod_memcg_state(from, MEMCG_SWAP, -1);
2943 mod_memcg_state(to, MEMCG_SWAP, 1);
2944 return 0;
2945 }
2946 return -EINVAL;
2947 }
2948 #else
2949 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2950 struct mem_cgroup *from, struct mem_cgroup *to)
2951 {
2952 return -EINVAL;
2953 }
2954 #endif
2955
2956 static DEFINE_MUTEX(memcg_max_mutex);
2957
2958 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2959 unsigned long max, bool memsw)
2960 {
2961 bool enlarge = false;
2962 bool drained = false;
2963 int ret;
2964 bool limits_invariant;
2965 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2966
2967 do {
2968 if (signal_pending(current)) {
2969 ret = -EINTR;
2970 break;
2971 }
2972
2973 mutex_lock(&memcg_max_mutex);
2974 /*
2975 * Make sure that the new limit (memsw or memory limit) doesn't
2976 * break our basic invariant rule memory.max <= memsw.max.
2977 */
2978 limits_invariant = memsw ? max >= memcg->memory.max :
2979 max <= memcg->memsw.max;
2980 if (!limits_invariant) {
2981 mutex_unlock(&memcg_max_mutex);
2982 ret = -EINVAL;
2983 break;
2984 }
2985 if (max > counter->max)
2986 enlarge = true;
2987 ret = page_counter_set_max(counter, max);
2988 mutex_unlock(&memcg_max_mutex);
2989
2990 if (!ret)
2991 break;
2992
2993 if (!drained) {
2994 drain_all_stock(memcg);
2995 drained = true;
2996 continue;
2997 }
2998
2999 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3000 GFP_KERNEL, !memsw)) {
3001 ret = -EBUSY;
3002 break;
3003 }
3004 } while (true);
3005
3006 if (!ret && enlarge)
3007 memcg_oom_recover(memcg);
3008
3009 return ret;
3010 }
3011
3012 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3013 gfp_t gfp_mask,
3014 unsigned long *total_scanned)
3015 {
3016 unsigned long nr_reclaimed = 0;
3017 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3018 unsigned long reclaimed;
3019 int loop = 0;
3020 struct mem_cgroup_tree_per_node *mctz;
3021 unsigned long excess;
3022 unsigned long nr_scanned;
3023
3024 if (order > 0)
3025 return 0;
3026
3027 mctz = soft_limit_tree_node(pgdat->node_id);
3028
3029 /*
3030 * Do not even bother to check the largest node if the root
3031 * is empty. Do it lockless to prevent lock bouncing. Races
3032 * are acceptable as soft limit is best effort anyway.
3033 */
3034 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3035 return 0;
3036
3037 /*
3038 * This loop can run a while, specially if mem_cgroup's continuously
3039 * keep exceeding their soft limit and putting the system under
3040 * pressure
3041 */
3042 do {
3043 if (next_mz)
3044 mz = next_mz;
3045 else
3046 mz = mem_cgroup_largest_soft_limit_node(mctz);
3047 if (!mz)
3048 break;
3049
3050 nr_scanned = 0;
3051 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3052 gfp_mask, &nr_scanned);
3053 nr_reclaimed += reclaimed;
3054 *total_scanned += nr_scanned;
3055 spin_lock_irq(&mctz->lock);
3056 __mem_cgroup_remove_exceeded(mz, mctz);
3057
3058 /*
3059 * If we failed to reclaim anything from this memory cgroup
3060 * it is time to move on to the next cgroup
3061 */
3062 next_mz = NULL;
3063 if (!reclaimed)
3064 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3065
3066 excess = soft_limit_excess(mz->memcg);
3067 /*
3068 * One school of thought says that we should not add
3069 * back the node to the tree if reclaim returns 0.
3070 * But our reclaim could return 0, simply because due
3071 * to priority we are exposing a smaller subset of
3072 * memory to reclaim from. Consider this as a longer
3073 * term TODO.
3074 */
3075 /* If excess == 0, no tree ops */
3076 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3077 spin_unlock_irq(&mctz->lock);
3078 css_put(&mz->memcg->css);
3079 loop++;
3080 /*
3081 * Could not reclaim anything and there are no more
3082 * mem cgroups to try or we seem to be looping without
3083 * reclaiming anything.
3084 */
3085 if (!nr_reclaimed &&
3086 (next_mz == NULL ||
3087 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3088 break;
3089 } while (!nr_reclaimed);
3090 if (next_mz)
3091 css_put(&next_mz->memcg->css);
3092 return nr_reclaimed;
3093 }
3094
3095 /*
3096 * Test whether @memcg has children, dead or alive. Note that this
3097 * function doesn't care whether @memcg has use_hierarchy enabled and
3098 * returns %true if there are child csses according to the cgroup
3099 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3100 */
3101 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3102 {
3103 bool ret;
3104
3105 rcu_read_lock();
3106 ret = css_next_child(NULL, &memcg->css);
3107 rcu_read_unlock();
3108 return ret;
3109 }
3110
3111 /*
3112 * Reclaims as many pages from the given memcg as possible.
3113 *
3114 * Caller is responsible for holding css reference for memcg.
3115 */
3116 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3117 {
3118 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3119
3120 /* we call try-to-free pages for make this cgroup empty */
3121 lru_add_drain_all();
3122
3123 drain_all_stock(memcg);
3124
3125 /* try to free all pages in this cgroup */
3126 while (nr_retries && page_counter_read(&memcg->memory)) {
3127 int progress;
3128
3129 if (signal_pending(current))
3130 return -EINTR;
3131
3132 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3133 GFP_KERNEL, true);
3134 if (!progress) {
3135 nr_retries--;
3136 /* maybe some writeback is necessary */
3137 congestion_wait(BLK_RW_ASYNC, HZ/10);
3138 }
3139
3140 }
3141
3142 return 0;
3143 }
3144
3145 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3146 char *buf, size_t nbytes,
3147 loff_t off)
3148 {
3149 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3150
3151 if (mem_cgroup_is_root(memcg))
3152 return -EINVAL;
3153 return mem_cgroup_force_empty(memcg) ?: nbytes;
3154 }
3155
3156 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3157 struct cftype *cft)
3158 {
3159 return mem_cgroup_from_css(css)->use_hierarchy;
3160 }
3161
3162 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3163 struct cftype *cft, u64 val)
3164 {
3165 int retval = 0;
3166 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3167 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3168
3169 if (memcg->use_hierarchy == val)
3170 return 0;
3171
3172 /*
3173 * If parent's use_hierarchy is set, we can't make any modifications
3174 * in the child subtrees. If it is unset, then the change can
3175 * occur, provided the current cgroup has no children.
3176 *
3177 * For the root cgroup, parent_mem is NULL, we allow value to be
3178 * set if there are no children.
3179 */
3180 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3181 (val == 1 || val == 0)) {
3182 if (!memcg_has_children(memcg))
3183 memcg->use_hierarchy = val;
3184 else
3185 retval = -EBUSY;
3186 } else
3187 retval = -EINVAL;
3188
3189 return retval;
3190 }
3191
3192 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3193 {
3194 unsigned long val;
3195
3196 if (mem_cgroup_is_root(memcg)) {
3197 val = memcg_page_state(memcg, MEMCG_CACHE) +
3198 memcg_page_state(memcg, MEMCG_RSS);
3199 if (swap)
3200 val += memcg_page_state(memcg, MEMCG_SWAP);
3201 } else {
3202 if (!swap)
3203 val = page_counter_read(&memcg->memory);
3204 else
3205 val = page_counter_read(&memcg->memsw);
3206 }
3207 return val;
3208 }
3209
3210 enum {
3211 RES_USAGE,
3212 RES_LIMIT,
3213 RES_MAX_USAGE,
3214 RES_FAILCNT,
3215 RES_SOFT_LIMIT,
3216 };
3217
3218 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3219 struct cftype *cft)
3220 {
3221 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3222 struct page_counter *counter;
3223
3224 switch (MEMFILE_TYPE(cft->private)) {
3225 case _MEM:
3226 counter = &memcg->memory;
3227 break;
3228 case _MEMSWAP:
3229 counter = &memcg->memsw;
3230 break;
3231 case _KMEM:
3232 counter = &memcg->kmem;
3233 break;
3234 case _TCP:
3235 counter = &memcg->tcpmem;
3236 break;
3237 default:
3238 BUG();
3239 }
3240
3241 switch (MEMFILE_ATTR(cft->private)) {
3242 case RES_USAGE:
3243 if (counter == &memcg->memory)
3244 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3245 if (counter == &memcg->memsw)
3246 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3247 return (u64)page_counter_read(counter) * PAGE_SIZE;
3248 case RES_LIMIT:
3249 return (u64)counter->max * PAGE_SIZE;
3250 case RES_MAX_USAGE:
3251 return (u64)counter->watermark * PAGE_SIZE;
3252 case RES_FAILCNT:
3253 return counter->failcnt;
3254 case RES_SOFT_LIMIT:
3255 return (u64)memcg->soft_limit * PAGE_SIZE;
3256 default:
3257 BUG();
3258 }
3259 }
3260
3261 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
3262 {
3263 unsigned long stat[MEMCG_NR_STAT];
3264 struct mem_cgroup *mi;
3265 int node, cpu, i;
3266 int min_idx, max_idx;
3267
3268 if (slab_only) {
3269 min_idx = NR_SLAB_RECLAIMABLE;
3270 max_idx = NR_SLAB_UNRECLAIMABLE;
3271 } else {
3272 min_idx = 0;
3273 max_idx = MEMCG_NR_STAT;
3274 }
3275
3276 for (i = min_idx; i < max_idx; i++)
3277 stat[i] = 0;
3278
3279 for_each_online_cpu(cpu)
3280 for (i = min_idx; i < max_idx; i++)
3281 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3282
3283 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3284 for (i = min_idx; i < max_idx; i++)
3285 atomic_long_add(stat[i], &mi->vmstats[i]);
3286
3287 if (!slab_only)
3288 max_idx = NR_VM_NODE_STAT_ITEMS;
3289
3290 for_each_node(node) {
3291 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3292 struct mem_cgroup_per_node *pi;
3293
3294 for (i = min_idx; i < max_idx; i++)
3295 stat[i] = 0;
3296
3297 for_each_online_cpu(cpu)
3298 for (i = min_idx; i < max_idx; i++)
3299 stat[i] += per_cpu(
3300 pn->lruvec_stat_cpu->count[i], cpu);
3301
3302 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3303 for (i = min_idx; i < max_idx; i++)
3304 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3305 }
3306 }
3307
3308 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3309 {
3310 unsigned long events[NR_VM_EVENT_ITEMS];
3311 struct mem_cgroup *mi;
3312 int cpu, i;
3313
3314 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3315 events[i] = 0;
3316
3317 for_each_online_cpu(cpu)
3318 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3319 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3320 cpu);
3321
3322 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3323 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3324 atomic_long_add(events[i], &mi->vmevents[i]);
3325 }
3326
3327 #ifdef CONFIG_MEMCG_KMEM
3328 static int memcg_online_kmem(struct mem_cgroup *memcg)
3329 {
3330 int memcg_id;
3331
3332 if (cgroup_memory_nokmem)
3333 return 0;
3334
3335 BUG_ON(memcg->kmemcg_id >= 0);
3336 BUG_ON(memcg->kmem_state);
3337
3338 memcg_id = memcg_alloc_cache_id();
3339 if (memcg_id < 0)
3340 return memcg_id;
3341
3342 static_branch_inc(&memcg_kmem_enabled_key);
3343 /*
3344 * A memory cgroup is considered kmem-online as soon as it gets
3345 * kmemcg_id. Setting the id after enabling static branching will
3346 * guarantee no one starts accounting before all call sites are
3347 * patched.
3348 */
3349 memcg->kmemcg_id = memcg_id;
3350 memcg->kmem_state = KMEM_ONLINE;
3351 INIT_LIST_HEAD(&memcg->kmem_caches);
3352
3353 return 0;
3354 }
3355
3356 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3357 {
3358 struct cgroup_subsys_state *css;
3359 struct mem_cgroup *parent, *child;
3360 int kmemcg_id;
3361
3362 if (memcg->kmem_state != KMEM_ONLINE)
3363 return;
3364 /*
3365 * Clear the online state before clearing memcg_caches array
3366 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3367 * guarantees that no cache will be created for this cgroup
3368 * after we are done (see memcg_create_kmem_cache()).
3369 */
3370 memcg->kmem_state = KMEM_ALLOCATED;
3371
3372 parent = parent_mem_cgroup(memcg);
3373 if (!parent)
3374 parent = root_mem_cgroup;
3375
3376 /*
3377 * Deactivate and reparent kmem_caches. Then flush percpu
3378 * slab statistics to have precise values at the parent and
3379 * all ancestor levels. It's required to keep slab stats
3380 * accurate after the reparenting of kmem_caches.
3381 */
3382 memcg_deactivate_kmem_caches(memcg, parent);
3383 memcg_flush_percpu_vmstats(memcg, true);
3384
3385 kmemcg_id = memcg->kmemcg_id;
3386 BUG_ON(kmemcg_id < 0);
3387
3388 /*
3389 * Change kmemcg_id of this cgroup and all its descendants to the
3390 * parent's id, and then move all entries from this cgroup's list_lrus
3391 * to ones of the parent. After we have finished, all list_lrus
3392 * corresponding to this cgroup are guaranteed to remain empty. The
3393 * ordering is imposed by list_lru_node->lock taken by
3394 * memcg_drain_all_list_lrus().
3395 */
3396 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3397 css_for_each_descendant_pre(css, &memcg->css) {
3398 child = mem_cgroup_from_css(css);
3399 BUG_ON(child->kmemcg_id != kmemcg_id);
3400 child->kmemcg_id = parent->kmemcg_id;
3401 if (!memcg->use_hierarchy)
3402 break;
3403 }
3404 rcu_read_unlock();
3405
3406 memcg_drain_all_list_lrus(kmemcg_id, parent);
3407
3408 memcg_free_cache_id(kmemcg_id);
3409 }
3410
3411 static void memcg_free_kmem(struct mem_cgroup *memcg)
3412 {
3413 /* css_alloc() failed, offlining didn't happen */
3414 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3415 memcg_offline_kmem(memcg);
3416
3417 if (memcg->kmem_state == KMEM_ALLOCATED) {
3418 WARN_ON(!list_empty(&memcg->kmem_caches));
3419 static_branch_dec(&memcg_kmem_enabled_key);
3420 }
3421 }
3422 #else
3423 static int memcg_online_kmem(struct mem_cgroup *memcg)
3424 {
3425 return 0;
3426 }
3427 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3428 {
3429 }
3430 static void memcg_free_kmem(struct mem_cgroup *memcg)
3431 {
3432 }
3433 #endif /* CONFIG_MEMCG_KMEM */
3434
3435 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3436 unsigned long max)
3437 {
3438 int ret;
3439
3440 mutex_lock(&memcg_max_mutex);
3441 ret = page_counter_set_max(&memcg->kmem, max);
3442 mutex_unlock(&memcg_max_mutex);
3443 return ret;
3444 }
3445
3446 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3447 {
3448 int ret;
3449
3450 mutex_lock(&memcg_max_mutex);
3451
3452 ret = page_counter_set_max(&memcg->tcpmem, max);
3453 if (ret)
3454 goto out;
3455
3456 if (!memcg->tcpmem_active) {
3457 /*
3458 * The active flag needs to be written after the static_key
3459 * update. This is what guarantees that the socket activation
3460 * function is the last one to run. See mem_cgroup_sk_alloc()
3461 * for details, and note that we don't mark any socket as
3462 * belonging to this memcg until that flag is up.
3463 *
3464 * We need to do this, because static_keys will span multiple
3465 * sites, but we can't control their order. If we mark a socket
3466 * as accounted, but the accounting functions are not patched in
3467 * yet, we'll lose accounting.
3468 *
3469 * We never race with the readers in mem_cgroup_sk_alloc(),
3470 * because when this value change, the code to process it is not
3471 * patched in yet.
3472 */
3473 static_branch_inc(&memcg_sockets_enabled_key);
3474 memcg->tcpmem_active = true;
3475 }
3476 out:
3477 mutex_unlock(&memcg_max_mutex);
3478 return ret;
3479 }
3480
3481 /*
3482 * The user of this function is...
3483 * RES_LIMIT.
3484 */
3485 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3486 char *buf, size_t nbytes, loff_t off)
3487 {
3488 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3489 unsigned long nr_pages;
3490 int ret;
3491
3492 buf = strstrip(buf);
3493 ret = page_counter_memparse(buf, "-1", &nr_pages);
3494 if (ret)
3495 return ret;
3496
3497 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3498 case RES_LIMIT:
3499 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3500 ret = -EINVAL;
3501 break;
3502 }
3503 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3504 case _MEM:
3505 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3506 break;
3507 case _MEMSWAP:
3508 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3509 break;
3510 case _KMEM:
3511 ret = memcg_update_kmem_max(memcg, nr_pages);
3512 break;
3513 case _TCP:
3514 ret = memcg_update_tcp_max(memcg, nr_pages);
3515 break;
3516 }
3517 break;
3518 case RES_SOFT_LIMIT:
3519 memcg->soft_limit = nr_pages;
3520 ret = 0;
3521 break;
3522 }
3523 return ret ?: nbytes;
3524 }
3525
3526 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3527 size_t nbytes, loff_t off)
3528 {
3529 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3530 struct page_counter *counter;
3531
3532 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3533 case _MEM:
3534 counter = &memcg->memory;
3535 break;
3536 case _MEMSWAP:
3537 counter = &memcg->memsw;
3538 break;
3539 case _KMEM:
3540 counter = &memcg->kmem;
3541 break;
3542 case _TCP:
3543 counter = &memcg->tcpmem;
3544 break;
3545 default:
3546 BUG();
3547 }
3548
3549 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3550 case RES_MAX_USAGE:
3551 page_counter_reset_watermark(counter);
3552 break;
3553 case RES_FAILCNT:
3554 counter->failcnt = 0;
3555 break;
3556 default:
3557 BUG();
3558 }
3559
3560 return nbytes;
3561 }
3562
3563 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3564 struct cftype *cft)
3565 {
3566 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3567 }
3568
3569 #ifdef CONFIG_MMU
3570 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3571 struct cftype *cft, u64 val)
3572 {
3573 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3574
3575 if (val & ~MOVE_MASK)
3576 return -EINVAL;
3577
3578 /*
3579 * No kind of locking is needed in here, because ->can_attach() will
3580 * check this value once in the beginning of the process, and then carry
3581 * on with stale data. This means that changes to this value will only
3582 * affect task migrations starting after the change.
3583 */
3584 memcg->move_charge_at_immigrate = val;
3585 return 0;
3586 }
3587 #else
3588 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3589 struct cftype *cft, u64 val)
3590 {
3591 return -ENOSYS;
3592 }
3593 #endif
3594
3595 #ifdef CONFIG_NUMA
3596
3597 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3598 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3599 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3600
3601 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3602 int nid, unsigned int lru_mask)
3603 {
3604 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3605 unsigned long nr = 0;
3606 enum lru_list lru;
3607
3608 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3609
3610 for_each_lru(lru) {
3611 if (!(BIT(lru) & lru_mask))
3612 continue;
3613 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3614 }
3615 return nr;
3616 }
3617
3618 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3619 unsigned int lru_mask)
3620 {
3621 unsigned long nr = 0;
3622 enum lru_list lru;
3623
3624 for_each_lru(lru) {
3625 if (!(BIT(lru) & lru_mask))
3626 continue;
3627 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3628 }
3629 return nr;
3630 }
3631
3632 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3633 {
3634 struct numa_stat {
3635 const char *name;
3636 unsigned int lru_mask;
3637 };
3638
3639 static const struct numa_stat stats[] = {
3640 { "total", LRU_ALL },
3641 { "file", LRU_ALL_FILE },
3642 { "anon", LRU_ALL_ANON },
3643 { "unevictable", BIT(LRU_UNEVICTABLE) },
3644 };
3645 const struct numa_stat *stat;
3646 int nid;
3647 unsigned long nr;
3648 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3649
3650 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3651 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3652 seq_printf(m, "%s=%lu", stat->name, nr);
3653 for_each_node_state(nid, N_MEMORY) {
3654 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3655 stat->lru_mask);
3656 seq_printf(m, " N%d=%lu", nid, nr);
3657 }
3658 seq_putc(m, '\n');
3659 }
3660
3661 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3662 struct mem_cgroup *iter;
3663
3664 nr = 0;
3665 for_each_mem_cgroup_tree(iter, memcg)
3666 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3667 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3668 for_each_node_state(nid, N_MEMORY) {
3669 nr = 0;
3670 for_each_mem_cgroup_tree(iter, memcg)
3671 nr += mem_cgroup_node_nr_lru_pages(
3672 iter, nid, stat->lru_mask);
3673 seq_printf(m, " N%d=%lu", nid, nr);
3674 }
3675 seq_putc(m, '\n');
3676 }
3677
3678 return 0;
3679 }
3680 #endif /* CONFIG_NUMA */
3681
3682 static const unsigned int memcg1_stats[] = {
3683 MEMCG_CACHE,
3684 MEMCG_RSS,
3685 MEMCG_RSS_HUGE,
3686 NR_SHMEM,
3687 NR_FILE_MAPPED,
3688 NR_FILE_DIRTY,
3689 NR_WRITEBACK,
3690 MEMCG_SWAP,
3691 };
3692
3693 static const char *const memcg1_stat_names[] = {
3694 "cache",
3695 "rss",
3696 "rss_huge",
3697 "shmem",
3698 "mapped_file",
3699 "dirty",
3700 "writeback",
3701 "swap",
3702 };
3703
3704 /* Universal VM events cgroup1 shows, original sort order */
3705 static const unsigned int memcg1_events[] = {
3706 PGPGIN,
3707 PGPGOUT,
3708 PGFAULT,
3709 PGMAJFAULT,
3710 };
3711
3712 static const char *const memcg1_event_names[] = {
3713 "pgpgin",
3714 "pgpgout",
3715 "pgfault",
3716 "pgmajfault",
3717 };
3718
3719 static int memcg_stat_show(struct seq_file *m, void *v)
3720 {
3721 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3722 unsigned long memory, memsw;
3723 struct mem_cgroup *mi;
3724 unsigned int i;
3725
3726 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3727 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3728
3729 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3730 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3731 continue;
3732 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3733 memcg_page_state_local(memcg, memcg1_stats[i]) *
3734 PAGE_SIZE);
3735 }
3736
3737 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3738 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3739 memcg_events_local(memcg, memcg1_events[i]));
3740
3741 for (i = 0; i < NR_LRU_LISTS; i++)
3742 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3743 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3744 PAGE_SIZE);
3745
3746 /* Hierarchical information */
3747 memory = memsw = PAGE_COUNTER_MAX;
3748 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3749 memory = min(memory, mi->memory.max);
3750 memsw = min(memsw, mi->memsw.max);
3751 }
3752 seq_printf(m, "hierarchical_memory_limit %llu\n",
3753 (u64)memory * PAGE_SIZE);
3754 if (do_memsw_account())
3755 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3756 (u64)memsw * PAGE_SIZE);
3757
3758 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3759 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3760 continue;
3761 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3762 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3763 PAGE_SIZE);
3764 }
3765
3766 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3767 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3768 (u64)memcg_events(memcg, memcg1_events[i]));
3769
3770 for (i = 0; i < NR_LRU_LISTS; i++)
3771 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3772 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3773 PAGE_SIZE);
3774
3775 #ifdef CONFIG_DEBUG_VM
3776 {
3777 pg_data_t *pgdat;
3778 struct mem_cgroup_per_node *mz;
3779 struct zone_reclaim_stat *rstat;
3780 unsigned long recent_rotated[2] = {0, 0};
3781 unsigned long recent_scanned[2] = {0, 0};
3782
3783 for_each_online_pgdat(pgdat) {
3784 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3785 rstat = &mz->lruvec.reclaim_stat;
3786
3787 recent_rotated[0] += rstat->recent_rotated[0];
3788 recent_rotated[1] += rstat->recent_rotated[1];
3789 recent_scanned[0] += rstat->recent_scanned[0];
3790 recent_scanned[1] += rstat->recent_scanned[1];
3791 }
3792 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3793 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3794 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3795 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3796 }
3797 #endif
3798
3799 return 0;
3800 }
3801
3802 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3803 struct cftype *cft)
3804 {
3805 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3806
3807 return mem_cgroup_swappiness(memcg);
3808 }
3809
3810 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3811 struct cftype *cft, u64 val)
3812 {
3813 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3814
3815 if (val > 100)
3816 return -EINVAL;
3817
3818 if (css->parent)
3819 memcg->swappiness = val;
3820 else
3821 vm_swappiness = val;
3822
3823 return 0;
3824 }
3825
3826 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3827 {
3828 struct mem_cgroup_threshold_ary *t;
3829 unsigned long usage;
3830 int i;
3831
3832 rcu_read_lock();
3833 if (!swap)
3834 t = rcu_dereference(memcg->thresholds.primary);
3835 else
3836 t = rcu_dereference(memcg->memsw_thresholds.primary);
3837
3838 if (!t)
3839 goto unlock;
3840
3841 usage = mem_cgroup_usage(memcg, swap);
3842
3843 /*
3844 * current_threshold points to threshold just below or equal to usage.
3845 * If it's not true, a threshold was crossed after last
3846 * call of __mem_cgroup_threshold().
3847 */
3848 i = t->current_threshold;
3849
3850 /*
3851 * Iterate backward over array of thresholds starting from
3852 * current_threshold and check if a threshold is crossed.
3853 * If none of thresholds below usage is crossed, we read
3854 * only one element of the array here.
3855 */
3856 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3857 eventfd_signal(t->entries[i].eventfd, 1);
3858
3859 /* i = current_threshold + 1 */
3860 i++;
3861
3862 /*
3863 * Iterate forward over array of thresholds starting from
3864 * current_threshold+1 and check if a threshold is crossed.
3865 * If none of thresholds above usage is crossed, we read
3866 * only one element of the array here.
3867 */
3868 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3869 eventfd_signal(t->entries[i].eventfd, 1);
3870
3871 /* Update current_threshold */
3872 t->current_threshold = i - 1;
3873 unlock:
3874 rcu_read_unlock();
3875 }
3876
3877 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3878 {
3879 while (memcg) {
3880 __mem_cgroup_threshold(memcg, false);
3881 if (do_memsw_account())
3882 __mem_cgroup_threshold(memcg, true);
3883
3884 memcg = parent_mem_cgroup(memcg);
3885 }
3886 }
3887
3888 static int compare_thresholds(const void *a, const void *b)
3889 {
3890 const struct mem_cgroup_threshold *_a = a;
3891 const struct mem_cgroup_threshold *_b = b;
3892
3893 if (_a->threshold > _b->threshold)
3894 return 1;
3895
3896 if (_a->threshold < _b->threshold)
3897 return -1;
3898
3899 return 0;
3900 }
3901
3902 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3903 {
3904 struct mem_cgroup_eventfd_list *ev;
3905
3906 spin_lock(&memcg_oom_lock);
3907
3908 list_for_each_entry(ev, &memcg->oom_notify, list)
3909 eventfd_signal(ev->eventfd, 1);
3910
3911 spin_unlock(&memcg_oom_lock);
3912 return 0;
3913 }
3914
3915 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3916 {
3917 struct mem_cgroup *iter;
3918
3919 for_each_mem_cgroup_tree(iter, memcg)
3920 mem_cgroup_oom_notify_cb(iter);
3921 }
3922
3923 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3924 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3925 {
3926 struct mem_cgroup_thresholds *thresholds;
3927 struct mem_cgroup_threshold_ary *new;
3928 unsigned long threshold;
3929 unsigned long usage;
3930 int i, size, ret;
3931
3932 ret = page_counter_memparse(args, "-1", &threshold);
3933 if (ret)
3934 return ret;
3935
3936 mutex_lock(&memcg->thresholds_lock);
3937
3938 if (type == _MEM) {
3939 thresholds = &memcg->thresholds;
3940 usage = mem_cgroup_usage(memcg, false);
3941 } else if (type == _MEMSWAP) {
3942 thresholds = &memcg->memsw_thresholds;
3943 usage = mem_cgroup_usage(memcg, true);
3944 } else
3945 BUG();
3946
3947 /* Check if a threshold crossed before adding a new one */
3948 if (thresholds->primary)
3949 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3950
3951 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3952
3953 /* Allocate memory for new array of thresholds */
3954 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3955 if (!new) {
3956 ret = -ENOMEM;
3957 goto unlock;
3958 }
3959 new->size = size;
3960
3961 /* Copy thresholds (if any) to new array */
3962 if (thresholds->primary) {
3963 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3964 sizeof(struct mem_cgroup_threshold));
3965 }
3966
3967 /* Add new threshold */
3968 new->entries[size - 1].eventfd = eventfd;
3969 new->entries[size - 1].threshold = threshold;
3970
3971 /* Sort thresholds. Registering of new threshold isn't time-critical */
3972 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3973 compare_thresholds, NULL);
3974
3975 /* Find current threshold */
3976 new->current_threshold = -1;
3977 for (i = 0; i < size; i++) {
3978 if (new->entries[i].threshold <= usage) {
3979 /*
3980 * new->current_threshold will not be used until
3981 * rcu_assign_pointer(), so it's safe to increment
3982 * it here.
3983 */
3984 ++new->current_threshold;
3985 } else
3986 break;
3987 }
3988
3989 /* Free old spare buffer and save old primary buffer as spare */
3990 kfree(thresholds->spare);
3991 thresholds->spare = thresholds->primary;
3992
3993 rcu_assign_pointer(thresholds->primary, new);
3994
3995 /* To be sure that nobody uses thresholds */
3996 synchronize_rcu();
3997
3998 unlock:
3999 mutex_unlock(&memcg->thresholds_lock);
4000
4001 return ret;
4002 }
4003
4004 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4005 struct eventfd_ctx *eventfd, const char *args)
4006 {
4007 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4008 }
4009
4010 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4011 struct eventfd_ctx *eventfd, const char *args)
4012 {
4013 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4014 }
4015
4016 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4017 struct eventfd_ctx *eventfd, enum res_type type)
4018 {
4019 struct mem_cgroup_thresholds *thresholds;
4020 struct mem_cgroup_threshold_ary *new;
4021 unsigned long usage;
4022 int i, j, size;
4023
4024 mutex_lock(&memcg->thresholds_lock);
4025
4026 if (type == _MEM) {
4027 thresholds = &memcg->thresholds;
4028 usage = mem_cgroup_usage(memcg, false);
4029 } else if (type == _MEMSWAP) {
4030 thresholds = &memcg->memsw_thresholds;
4031 usage = mem_cgroup_usage(memcg, true);
4032 } else
4033 BUG();
4034
4035 if (!thresholds->primary)
4036 goto unlock;
4037
4038 /* Check if a threshold crossed before removing */
4039 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4040
4041 /* Calculate new number of threshold */
4042 size = 0;
4043 for (i = 0; i < thresholds->primary->size; i++) {
4044 if (thresholds->primary->entries[i].eventfd != eventfd)
4045 size++;
4046 }
4047
4048 new = thresholds->spare;
4049
4050 /* Set thresholds array to NULL if we don't have thresholds */
4051 if (!size) {
4052 kfree(new);
4053 new = NULL;
4054 goto swap_buffers;
4055 }
4056
4057 new->size = size;
4058
4059 /* Copy thresholds and find current threshold */
4060 new->current_threshold = -1;
4061 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4062 if (thresholds->primary->entries[i].eventfd == eventfd)
4063 continue;
4064
4065 new->entries[j] = thresholds->primary->entries[i];
4066 if (new->entries[j].threshold <= usage) {
4067 /*
4068 * new->current_threshold will not be used
4069 * until rcu_assign_pointer(), so it's safe to increment
4070 * it here.
4071 */
4072 ++new->current_threshold;
4073 }
4074 j++;
4075 }
4076
4077 swap_buffers:
4078 /* Swap primary and spare array */
4079 thresholds->spare = thresholds->primary;
4080
4081 rcu_assign_pointer(thresholds->primary, new);
4082
4083 /* To be sure that nobody uses thresholds */
4084 synchronize_rcu();
4085
4086 /* If all events are unregistered, free the spare array */
4087 if (!new) {
4088 kfree(thresholds->spare);
4089 thresholds->spare = NULL;
4090 }
4091 unlock:
4092 mutex_unlock(&memcg->thresholds_lock);
4093 }
4094
4095 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4096 struct eventfd_ctx *eventfd)
4097 {
4098 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4099 }
4100
4101 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4102 struct eventfd_ctx *eventfd)
4103 {
4104 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4105 }
4106
4107 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4108 struct eventfd_ctx *eventfd, const char *args)
4109 {
4110 struct mem_cgroup_eventfd_list *event;
4111
4112 event = kmalloc(sizeof(*event), GFP_KERNEL);
4113 if (!event)
4114 return -ENOMEM;
4115
4116 spin_lock(&memcg_oom_lock);
4117
4118 event->eventfd = eventfd;
4119 list_add(&event->list, &memcg->oom_notify);
4120
4121 /* already in OOM ? */
4122 if (memcg->under_oom)
4123 eventfd_signal(eventfd, 1);
4124 spin_unlock(&memcg_oom_lock);
4125
4126 return 0;
4127 }
4128
4129 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4130 struct eventfd_ctx *eventfd)
4131 {
4132 struct mem_cgroup_eventfd_list *ev, *tmp;
4133
4134 spin_lock(&memcg_oom_lock);
4135
4136 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4137 if (ev->eventfd == eventfd) {
4138 list_del(&ev->list);
4139 kfree(ev);
4140 }
4141 }
4142
4143 spin_unlock(&memcg_oom_lock);
4144 }
4145
4146 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4147 {
4148 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4149
4150 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4151 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4152 seq_printf(sf, "oom_kill %lu\n",
4153 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4154 return 0;
4155 }
4156
4157 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4158 struct cftype *cft, u64 val)
4159 {
4160 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4161
4162 /* cannot set to root cgroup and only 0 and 1 are allowed */
4163 if (!css->parent || !((val == 0) || (val == 1)))
4164 return -EINVAL;
4165
4166 memcg->oom_kill_disable = val;
4167 if (!val)
4168 memcg_oom_recover(memcg);
4169
4170 return 0;
4171 }
4172
4173 #ifdef CONFIG_CGROUP_WRITEBACK
4174
4175 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4176 {
4177 return wb_domain_init(&memcg->cgwb_domain, gfp);
4178 }
4179
4180 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4181 {
4182 wb_domain_exit(&memcg->cgwb_domain);
4183 }
4184
4185 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4186 {
4187 wb_domain_size_changed(&memcg->cgwb_domain);
4188 }
4189
4190 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4191 {
4192 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4193
4194 if (!memcg->css.parent)
4195 return NULL;
4196
4197 return &memcg->cgwb_domain;
4198 }
4199
4200 /*
4201 * idx can be of type enum memcg_stat_item or node_stat_item.
4202 * Keep in sync with memcg_exact_page().
4203 */
4204 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4205 {
4206 long x = atomic_long_read(&memcg->vmstats[idx]);
4207 int cpu;
4208
4209 for_each_online_cpu(cpu)
4210 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4211 if (x < 0)
4212 x = 0;
4213 return x;
4214 }
4215
4216 /**
4217 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4218 * @wb: bdi_writeback in question
4219 * @pfilepages: out parameter for number of file pages
4220 * @pheadroom: out parameter for number of allocatable pages according to memcg
4221 * @pdirty: out parameter for number of dirty pages
4222 * @pwriteback: out parameter for number of pages under writeback
4223 *
4224 * Determine the numbers of file, headroom, dirty, and writeback pages in
4225 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4226 * is a bit more involved.
4227 *
4228 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4229 * headroom is calculated as the lowest headroom of itself and the
4230 * ancestors. Note that this doesn't consider the actual amount of
4231 * available memory in the system. The caller should further cap
4232 * *@pheadroom accordingly.
4233 */
4234 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4235 unsigned long *pheadroom, unsigned long *pdirty,
4236 unsigned long *pwriteback)
4237 {
4238 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4239 struct mem_cgroup *parent;
4240
4241 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4242
4243 /* this should eventually include NR_UNSTABLE_NFS */
4244 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4245 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4246 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4247 *pheadroom = PAGE_COUNTER_MAX;
4248
4249 while ((parent = parent_mem_cgroup(memcg))) {
4250 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4251 unsigned long used = page_counter_read(&memcg->memory);
4252
4253 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4254 memcg = parent;
4255 }
4256 }
4257
4258 #else /* CONFIG_CGROUP_WRITEBACK */
4259
4260 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4261 {
4262 return 0;
4263 }
4264
4265 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4266 {
4267 }
4268
4269 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4270 {
4271 }
4272
4273 #endif /* CONFIG_CGROUP_WRITEBACK */
4274
4275 /*
4276 * DO NOT USE IN NEW FILES.
4277 *
4278 * "cgroup.event_control" implementation.
4279 *
4280 * This is way over-engineered. It tries to support fully configurable
4281 * events for each user. Such level of flexibility is completely
4282 * unnecessary especially in the light of the planned unified hierarchy.
4283 *
4284 * Please deprecate this and replace with something simpler if at all
4285 * possible.
4286 */
4287
4288 /*
4289 * Unregister event and free resources.
4290 *
4291 * Gets called from workqueue.
4292 */
4293 static void memcg_event_remove(struct work_struct *work)
4294 {
4295 struct mem_cgroup_event *event =
4296 container_of(work, struct mem_cgroup_event, remove);
4297 struct mem_cgroup *memcg = event->memcg;
4298
4299 remove_wait_queue(event->wqh, &event->wait);
4300
4301 event->unregister_event(memcg, event->eventfd);
4302
4303 /* Notify userspace the event is going away. */
4304 eventfd_signal(event->eventfd, 1);
4305
4306 eventfd_ctx_put(event->eventfd);
4307 kfree(event);
4308 css_put(&memcg->css);
4309 }
4310
4311 /*
4312 * Gets called on EPOLLHUP on eventfd when user closes it.
4313 *
4314 * Called with wqh->lock held and interrupts disabled.
4315 */
4316 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4317 int sync, void *key)
4318 {
4319 struct mem_cgroup_event *event =
4320 container_of(wait, struct mem_cgroup_event, wait);
4321 struct mem_cgroup *memcg = event->memcg;
4322 __poll_t flags = key_to_poll(key);
4323
4324 if (flags & EPOLLHUP) {
4325 /*
4326 * If the event has been detached at cgroup removal, we
4327 * can simply return knowing the other side will cleanup
4328 * for us.
4329 *
4330 * We can't race against event freeing since the other
4331 * side will require wqh->lock via remove_wait_queue(),
4332 * which we hold.
4333 */
4334 spin_lock(&memcg->event_list_lock);
4335 if (!list_empty(&event->list)) {
4336 list_del_init(&event->list);
4337 /*
4338 * We are in atomic context, but cgroup_event_remove()
4339 * may sleep, so we have to call it in workqueue.
4340 */
4341 schedule_work(&event->remove);
4342 }
4343 spin_unlock(&memcg->event_list_lock);
4344 }
4345
4346 return 0;
4347 }
4348
4349 static void memcg_event_ptable_queue_proc(struct file *file,
4350 wait_queue_head_t *wqh, poll_table *pt)
4351 {
4352 struct mem_cgroup_event *event =
4353 container_of(pt, struct mem_cgroup_event, pt);
4354
4355 event->wqh = wqh;
4356 add_wait_queue(wqh, &event->wait);
4357 }
4358
4359 /*
4360 * DO NOT USE IN NEW FILES.
4361 *
4362 * Parse input and register new cgroup event handler.
4363 *
4364 * Input must be in format '<event_fd> <control_fd> <args>'.
4365 * Interpretation of args is defined by control file implementation.
4366 */
4367 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4368 char *buf, size_t nbytes, loff_t off)
4369 {
4370 struct cgroup_subsys_state *css = of_css(of);
4371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4372 struct mem_cgroup_event *event;
4373 struct cgroup_subsys_state *cfile_css;
4374 unsigned int efd, cfd;
4375 struct fd efile;
4376 struct fd cfile;
4377 const char *name;
4378 char *endp;
4379 int ret;
4380
4381 buf = strstrip(buf);
4382
4383 efd = simple_strtoul(buf, &endp, 10);
4384 if (*endp != ' ')
4385 return -EINVAL;
4386 buf = endp + 1;
4387
4388 cfd = simple_strtoul(buf, &endp, 10);
4389 if ((*endp != ' ') && (*endp != '\0'))
4390 return -EINVAL;
4391 buf = endp + 1;
4392
4393 event = kzalloc(sizeof(*event), GFP_KERNEL);
4394 if (!event)
4395 return -ENOMEM;
4396
4397 event->memcg = memcg;
4398 INIT_LIST_HEAD(&event->list);
4399 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4400 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4401 INIT_WORK(&event->remove, memcg_event_remove);
4402
4403 efile = fdget(efd);
4404 if (!efile.file) {
4405 ret = -EBADF;
4406 goto out_kfree;
4407 }
4408
4409 event->eventfd = eventfd_ctx_fileget(efile.file);
4410 if (IS_ERR(event->eventfd)) {
4411 ret = PTR_ERR(event->eventfd);
4412 goto out_put_efile;
4413 }
4414
4415 cfile = fdget(cfd);
4416 if (!cfile.file) {
4417 ret = -EBADF;
4418 goto out_put_eventfd;
4419 }
4420
4421 /* the process need read permission on control file */
4422 /* AV: shouldn't we check that it's been opened for read instead? */
4423 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4424 if (ret < 0)
4425 goto out_put_cfile;
4426
4427 /*
4428 * Determine the event callbacks and set them in @event. This used
4429 * to be done via struct cftype but cgroup core no longer knows
4430 * about these events. The following is crude but the whole thing
4431 * is for compatibility anyway.
4432 *
4433 * DO NOT ADD NEW FILES.
4434 */
4435 name = cfile.file->f_path.dentry->d_name.name;
4436
4437 if (!strcmp(name, "memory.usage_in_bytes")) {
4438 event->register_event = mem_cgroup_usage_register_event;
4439 event->unregister_event = mem_cgroup_usage_unregister_event;
4440 } else if (!strcmp(name, "memory.oom_control")) {
4441 event->register_event = mem_cgroup_oom_register_event;
4442 event->unregister_event = mem_cgroup_oom_unregister_event;
4443 } else if (!strcmp(name, "memory.pressure_level")) {
4444 event->register_event = vmpressure_register_event;
4445 event->unregister_event = vmpressure_unregister_event;
4446 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4447 event->register_event = memsw_cgroup_usage_register_event;
4448 event->unregister_event = memsw_cgroup_usage_unregister_event;
4449 } else {
4450 ret = -EINVAL;
4451 goto out_put_cfile;
4452 }
4453
4454 /*
4455 * Verify @cfile should belong to @css. Also, remaining events are
4456 * automatically removed on cgroup destruction but the removal is
4457 * asynchronous, so take an extra ref on @css.
4458 */
4459 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4460 &memory_cgrp_subsys);
4461 ret = -EINVAL;
4462 if (IS_ERR(cfile_css))
4463 goto out_put_cfile;
4464 if (cfile_css != css) {
4465 css_put(cfile_css);
4466 goto out_put_cfile;
4467 }
4468
4469 ret = event->register_event(memcg, event->eventfd, buf);
4470 if (ret)
4471 goto out_put_css;
4472
4473 vfs_poll(efile.file, &event->pt);
4474
4475 spin_lock(&memcg->event_list_lock);
4476 list_add(&event->list, &memcg->event_list);
4477 spin_unlock(&memcg->event_list_lock);
4478
4479 fdput(cfile);
4480 fdput(efile);
4481
4482 return nbytes;
4483
4484 out_put_css:
4485 css_put(css);
4486 out_put_cfile:
4487 fdput(cfile);
4488 out_put_eventfd:
4489 eventfd_ctx_put(event->eventfd);
4490 out_put_efile:
4491 fdput(efile);
4492 out_kfree:
4493 kfree(event);
4494
4495 return ret;
4496 }
4497
4498 static struct cftype mem_cgroup_legacy_files[] = {
4499 {
4500 .name = "usage_in_bytes",
4501 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4502 .read_u64 = mem_cgroup_read_u64,
4503 },
4504 {
4505 .name = "max_usage_in_bytes",
4506 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4507 .write = mem_cgroup_reset,
4508 .read_u64 = mem_cgroup_read_u64,
4509 },
4510 {
4511 .name = "limit_in_bytes",
4512 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4513 .write = mem_cgroup_write,
4514 .read_u64 = mem_cgroup_read_u64,
4515 },
4516 {
4517 .name = "soft_limit_in_bytes",
4518 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4519 .write = mem_cgroup_write,
4520 .read_u64 = mem_cgroup_read_u64,
4521 },
4522 {
4523 .name = "failcnt",
4524 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4525 .write = mem_cgroup_reset,
4526 .read_u64 = mem_cgroup_read_u64,
4527 },
4528 {
4529 .name = "stat",
4530 .seq_show = memcg_stat_show,
4531 },
4532 {
4533 .name = "force_empty",
4534 .write = mem_cgroup_force_empty_write,
4535 },
4536 {
4537 .name = "use_hierarchy",
4538 .write_u64 = mem_cgroup_hierarchy_write,
4539 .read_u64 = mem_cgroup_hierarchy_read,
4540 },
4541 {
4542 .name = "cgroup.event_control", /* XXX: for compat */
4543 .write = memcg_write_event_control,
4544 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4545 },
4546 {
4547 .name = "swappiness",
4548 .read_u64 = mem_cgroup_swappiness_read,
4549 .write_u64 = mem_cgroup_swappiness_write,
4550 },
4551 {
4552 .name = "move_charge_at_immigrate",
4553 .read_u64 = mem_cgroup_move_charge_read,
4554 .write_u64 = mem_cgroup_move_charge_write,
4555 },
4556 {
4557 .name = "oom_control",
4558 .seq_show = mem_cgroup_oom_control_read,
4559 .write_u64 = mem_cgroup_oom_control_write,
4560 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4561 },
4562 {
4563 .name = "pressure_level",
4564 },
4565 #ifdef CONFIG_NUMA
4566 {
4567 .name = "numa_stat",
4568 .seq_show = memcg_numa_stat_show,
4569 },
4570 #endif
4571 {
4572 .name = "kmem.limit_in_bytes",
4573 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4574 .write = mem_cgroup_write,
4575 .read_u64 = mem_cgroup_read_u64,
4576 },
4577 {
4578 .name = "kmem.usage_in_bytes",
4579 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4580 .read_u64 = mem_cgroup_read_u64,
4581 },
4582 {
4583 .name = "kmem.failcnt",
4584 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4585 .write = mem_cgroup_reset,
4586 .read_u64 = mem_cgroup_read_u64,
4587 },
4588 {
4589 .name = "kmem.max_usage_in_bytes",
4590 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4591 .write = mem_cgroup_reset,
4592 .read_u64 = mem_cgroup_read_u64,
4593 },
4594 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4595 {
4596 .name = "kmem.slabinfo",
4597 .seq_start = memcg_slab_start,
4598 .seq_next = memcg_slab_next,
4599 .seq_stop = memcg_slab_stop,
4600 .seq_show = memcg_slab_show,
4601 },
4602 #endif
4603 {
4604 .name = "kmem.tcp.limit_in_bytes",
4605 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4606 .write = mem_cgroup_write,
4607 .read_u64 = mem_cgroup_read_u64,
4608 },
4609 {
4610 .name = "kmem.tcp.usage_in_bytes",
4611 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4612 .read_u64 = mem_cgroup_read_u64,
4613 },
4614 {
4615 .name = "kmem.tcp.failcnt",
4616 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4617 .write = mem_cgroup_reset,
4618 .read_u64 = mem_cgroup_read_u64,
4619 },
4620 {
4621 .name = "kmem.tcp.max_usage_in_bytes",
4622 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4623 .write = mem_cgroup_reset,
4624 .read_u64 = mem_cgroup_read_u64,
4625 },
4626 { }, /* terminate */
4627 };
4628
4629 /*
4630 * Private memory cgroup IDR
4631 *
4632 * Swap-out records and page cache shadow entries need to store memcg
4633 * references in constrained space, so we maintain an ID space that is
4634 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4635 * memory-controlled cgroups to 64k.
4636 *
4637 * However, there usually are many references to the oflline CSS after
4638 * the cgroup has been destroyed, such as page cache or reclaimable
4639 * slab objects, that don't need to hang on to the ID. We want to keep
4640 * those dead CSS from occupying IDs, or we might quickly exhaust the
4641 * relatively small ID space and prevent the creation of new cgroups
4642 * even when there are much fewer than 64k cgroups - possibly none.
4643 *
4644 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4645 * be freed and recycled when it's no longer needed, which is usually
4646 * when the CSS is offlined.
4647 *
4648 * The only exception to that are records of swapped out tmpfs/shmem
4649 * pages that need to be attributed to live ancestors on swapin. But
4650 * those references are manageable from userspace.
4651 */
4652
4653 static DEFINE_IDR(mem_cgroup_idr);
4654
4655 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4656 {
4657 if (memcg->id.id > 0) {
4658 idr_remove(&mem_cgroup_idr, memcg->id.id);
4659 memcg->id.id = 0;
4660 }
4661 }
4662
4663 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4664 {
4665 refcount_add(n, &memcg->id.ref);
4666 }
4667
4668 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4669 {
4670 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4671 mem_cgroup_id_remove(memcg);
4672
4673 /* Memcg ID pins CSS */
4674 css_put(&memcg->css);
4675 }
4676 }
4677
4678 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4679 {
4680 mem_cgroup_id_get_many(memcg, 1);
4681 }
4682
4683 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4684 {
4685 mem_cgroup_id_put_many(memcg, 1);
4686 }
4687
4688 /**
4689 * mem_cgroup_from_id - look up a memcg from a memcg id
4690 * @id: the memcg id to look up
4691 *
4692 * Caller must hold rcu_read_lock().
4693 */
4694 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4695 {
4696 WARN_ON_ONCE(!rcu_read_lock_held());
4697 return idr_find(&mem_cgroup_idr, id);
4698 }
4699
4700 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4701 {
4702 struct mem_cgroup_per_node *pn;
4703 int tmp = node;
4704 /*
4705 * This routine is called against possible nodes.
4706 * But it's BUG to call kmalloc() against offline node.
4707 *
4708 * TODO: this routine can waste much memory for nodes which will
4709 * never be onlined. It's better to use memory hotplug callback
4710 * function.
4711 */
4712 if (!node_state(node, N_NORMAL_MEMORY))
4713 tmp = -1;
4714 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4715 if (!pn)
4716 return 1;
4717
4718 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4719 if (!pn->lruvec_stat_local) {
4720 kfree(pn);
4721 return 1;
4722 }
4723
4724 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4725 if (!pn->lruvec_stat_cpu) {
4726 free_percpu(pn->lruvec_stat_local);
4727 kfree(pn);
4728 return 1;
4729 }
4730
4731 lruvec_init(&pn->lruvec);
4732 pn->usage_in_excess = 0;
4733 pn->on_tree = false;
4734 pn->memcg = memcg;
4735
4736 memcg->nodeinfo[node] = pn;
4737 return 0;
4738 }
4739
4740 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4741 {
4742 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4743
4744 if (!pn)
4745 return;
4746
4747 free_percpu(pn->lruvec_stat_cpu);
4748 free_percpu(pn->lruvec_stat_local);
4749 kfree(pn);
4750 }
4751
4752 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4753 {
4754 int node;
4755
4756 /*
4757 * Flush percpu vmstats and vmevents to guarantee the value correctness
4758 * on parent's and all ancestor levels.
4759 */
4760 memcg_flush_percpu_vmstats(memcg, false);
4761 memcg_flush_percpu_vmevents(memcg);
4762 for_each_node(node)
4763 free_mem_cgroup_per_node_info(memcg, node);
4764 free_percpu(memcg->vmstats_percpu);
4765 free_percpu(memcg->vmstats_local);
4766 kfree(memcg);
4767 }
4768
4769 static void mem_cgroup_free(struct mem_cgroup *memcg)
4770 {
4771 memcg_wb_domain_exit(memcg);
4772 __mem_cgroup_free(memcg);
4773 }
4774
4775 static struct mem_cgroup *mem_cgroup_alloc(void)
4776 {
4777 struct mem_cgroup *memcg;
4778 unsigned int size;
4779 int node;
4780
4781 size = sizeof(struct mem_cgroup);
4782 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4783
4784 memcg = kzalloc(size, GFP_KERNEL);
4785 if (!memcg)
4786 return NULL;
4787
4788 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4789 1, MEM_CGROUP_ID_MAX,
4790 GFP_KERNEL);
4791 if (memcg->id.id < 0)
4792 goto fail;
4793
4794 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4795 if (!memcg->vmstats_local)
4796 goto fail;
4797
4798 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4799 if (!memcg->vmstats_percpu)
4800 goto fail;
4801
4802 for_each_node(node)
4803 if (alloc_mem_cgroup_per_node_info(memcg, node))
4804 goto fail;
4805
4806 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4807 goto fail;
4808
4809 INIT_WORK(&memcg->high_work, high_work_func);
4810 memcg->last_scanned_node = MAX_NUMNODES;
4811 INIT_LIST_HEAD(&memcg->oom_notify);
4812 mutex_init(&memcg->thresholds_lock);
4813 spin_lock_init(&memcg->move_lock);
4814 vmpressure_init(&memcg->vmpressure);
4815 INIT_LIST_HEAD(&memcg->event_list);
4816 spin_lock_init(&memcg->event_list_lock);
4817 memcg->socket_pressure = jiffies;
4818 #ifdef CONFIG_MEMCG_KMEM
4819 memcg->kmemcg_id = -1;
4820 #endif
4821 #ifdef CONFIG_CGROUP_WRITEBACK
4822 INIT_LIST_HEAD(&memcg->cgwb_list);
4823 #endif
4824 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4825 return memcg;
4826 fail:
4827 mem_cgroup_id_remove(memcg);
4828 __mem_cgroup_free(memcg);
4829 return NULL;
4830 }
4831
4832 static struct cgroup_subsys_state * __ref
4833 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4834 {
4835 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4836 struct mem_cgroup *memcg;
4837 long error = -ENOMEM;
4838
4839 memcg = mem_cgroup_alloc();
4840 if (!memcg)
4841 return ERR_PTR(error);
4842
4843 memcg->high = PAGE_COUNTER_MAX;
4844 memcg->soft_limit = PAGE_COUNTER_MAX;
4845 if (parent) {
4846 memcg->swappiness = mem_cgroup_swappiness(parent);
4847 memcg->oom_kill_disable = parent->oom_kill_disable;
4848 }
4849 if (parent && parent->use_hierarchy) {
4850 memcg->use_hierarchy = true;
4851 page_counter_init(&memcg->memory, &parent->memory);
4852 page_counter_init(&memcg->swap, &parent->swap);
4853 page_counter_init(&memcg->memsw, &parent->memsw);
4854 page_counter_init(&memcg->kmem, &parent->kmem);
4855 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4856 } else {
4857 page_counter_init(&memcg->memory, NULL);
4858 page_counter_init(&memcg->swap, NULL);
4859 page_counter_init(&memcg->memsw, NULL);
4860 page_counter_init(&memcg->kmem, NULL);
4861 page_counter_init(&memcg->tcpmem, NULL);
4862 /*
4863 * Deeper hierachy with use_hierarchy == false doesn't make
4864 * much sense so let cgroup subsystem know about this
4865 * unfortunate state in our controller.
4866 */
4867 if (parent != root_mem_cgroup)
4868 memory_cgrp_subsys.broken_hierarchy = true;
4869 }
4870
4871 /* The following stuff does not apply to the root */
4872 if (!parent) {
4873 #ifdef CONFIG_MEMCG_KMEM
4874 INIT_LIST_HEAD(&memcg->kmem_caches);
4875 #endif
4876 root_mem_cgroup = memcg;
4877 return &memcg->css;
4878 }
4879
4880 error = memcg_online_kmem(memcg);
4881 if (error)
4882 goto fail;
4883
4884 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4885 static_branch_inc(&memcg_sockets_enabled_key);
4886
4887 return &memcg->css;
4888 fail:
4889 mem_cgroup_id_remove(memcg);
4890 mem_cgroup_free(memcg);
4891 return ERR_PTR(-ENOMEM);
4892 }
4893
4894 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4895 {
4896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4897
4898 /*
4899 * A memcg must be visible for memcg_expand_shrinker_maps()
4900 * by the time the maps are allocated. So, we allocate maps
4901 * here, when for_each_mem_cgroup() can't skip it.
4902 */
4903 if (memcg_alloc_shrinker_maps(memcg)) {
4904 mem_cgroup_id_remove(memcg);
4905 return -ENOMEM;
4906 }
4907
4908 /* Online state pins memcg ID, memcg ID pins CSS */
4909 refcount_set(&memcg->id.ref, 1);
4910 css_get(css);
4911 return 0;
4912 }
4913
4914 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4915 {
4916 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4917 struct mem_cgroup_event *event, *tmp;
4918
4919 /*
4920 * Unregister events and notify userspace.
4921 * Notify userspace about cgroup removing only after rmdir of cgroup
4922 * directory to avoid race between userspace and kernelspace.
4923 */
4924 spin_lock(&memcg->event_list_lock);
4925 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4926 list_del_init(&event->list);
4927 schedule_work(&event->remove);
4928 }
4929 spin_unlock(&memcg->event_list_lock);
4930
4931 page_counter_set_min(&memcg->memory, 0);
4932 page_counter_set_low(&memcg->memory, 0);
4933
4934 memcg_offline_kmem(memcg);
4935 wb_memcg_offline(memcg);
4936
4937 drain_all_stock(memcg);
4938
4939 mem_cgroup_id_put(memcg);
4940 }
4941
4942 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4943 {
4944 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4945
4946 invalidate_reclaim_iterators(memcg);
4947 }
4948
4949 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4950 {
4951 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4952
4953 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4954 static_branch_dec(&memcg_sockets_enabled_key);
4955
4956 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4957 static_branch_dec(&memcg_sockets_enabled_key);
4958
4959 vmpressure_cleanup(&memcg->vmpressure);
4960 cancel_work_sync(&memcg->high_work);
4961 mem_cgroup_remove_from_trees(memcg);
4962 memcg_free_shrinker_maps(memcg);
4963 memcg_free_kmem(memcg);
4964 mem_cgroup_free(memcg);
4965 }
4966
4967 /**
4968 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4969 * @css: the target css
4970 *
4971 * Reset the states of the mem_cgroup associated with @css. This is
4972 * invoked when the userland requests disabling on the default hierarchy
4973 * but the memcg is pinned through dependency. The memcg should stop
4974 * applying policies and should revert to the vanilla state as it may be
4975 * made visible again.
4976 *
4977 * The current implementation only resets the essential configurations.
4978 * This needs to be expanded to cover all the visible parts.
4979 */
4980 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4981 {
4982 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4983
4984 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4985 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4986 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4987 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4988 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4989 page_counter_set_min(&memcg->memory, 0);
4990 page_counter_set_low(&memcg->memory, 0);
4991 memcg->high = PAGE_COUNTER_MAX;
4992 memcg->soft_limit = PAGE_COUNTER_MAX;
4993 memcg_wb_domain_size_changed(memcg);
4994 }
4995
4996 #ifdef CONFIG_MMU
4997 /* Handlers for move charge at task migration. */
4998 static int mem_cgroup_do_precharge(unsigned long count)
4999 {
5000 int ret;
5001
5002 /* Try a single bulk charge without reclaim first, kswapd may wake */
5003 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5004 if (!ret) {
5005 mc.precharge += count;
5006 return ret;
5007 }
5008
5009 /* Try charges one by one with reclaim, but do not retry */
5010 while (count--) {
5011 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5012 if (ret)
5013 return ret;
5014 mc.precharge++;
5015 cond_resched();
5016 }
5017 return 0;
5018 }
5019
5020 union mc_target {
5021 struct page *page;
5022 swp_entry_t ent;
5023 };
5024
5025 enum mc_target_type {
5026 MC_TARGET_NONE = 0,
5027 MC_TARGET_PAGE,
5028 MC_TARGET_SWAP,
5029 MC_TARGET_DEVICE,
5030 };
5031
5032 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5033 unsigned long addr, pte_t ptent)
5034 {
5035 struct page *page = vm_normal_page(vma, addr, ptent);
5036
5037 if (!page || !page_mapped(page))
5038 return NULL;
5039 if (PageAnon(page)) {
5040 if (!(mc.flags & MOVE_ANON))
5041 return NULL;
5042 } else {
5043 if (!(mc.flags & MOVE_FILE))
5044 return NULL;
5045 }
5046 if (!get_page_unless_zero(page))
5047 return NULL;
5048
5049 return page;
5050 }
5051
5052 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5053 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5054 pte_t ptent, swp_entry_t *entry)
5055 {
5056 struct page *page = NULL;
5057 swp_entry_t ent = pte_to_swp_entry(ptent);
5058
5059 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5060 return NULL;
5061
5062 /*
5063 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5064 * a device and because they are not accessible by CPU they are store
5065 * as special swap entry in the CPU page table.
5066 */
5067 if (is_device_private_entry(ent)) {
5068 page = device_private_entry_to_page(ent);
5069 /*
5070 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5071 * a refcount of 1 when free (unlike normal page)
5072 */
5073 if (!page_ref_add_unless(page, 1, 1))
5074 return NULL;
5075 return page;
5076 }
5077
5078 /*
5079 * Because lookup_swap_cache() updates some statistics counter,
5080 * we call find_get_page() with swapper_space directly.
5081 */
5082 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5083 if (do_memsw_account())
5084 entry->val = ent.val;
5085
5086 return page;
5087 }
5088 #else
5089 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5090 pte_t ptent, swp_entry_t *entry)
5091 {
5092 return NULL;
5093 }
5094 #endif
5095
5096 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5097 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5098 {
5099 struct page *page = NULL;
5100 struct address_space *mapping;
5101 pgoff_t pgoff;
5102
5103 if (!vma->vm_file) /* anonymous vma */
5104 return NULL;
5105 if (!(mc.flags & MOVE_FILE))
5106 return NULL;
5107
5108 mapping = vma->vm_file->f_mapping;
5109 pgoff = linear_page_index(vma, addr);
5110
5111 /* page is moved even if it's not RSS of this task(page-faulted). */
5112 #ifdef CONFIG_SWAP
5113 /* shmem/tmpfs may report page out on swap: account for that too. */
5114 if (shmem_mapping(mapping)) {
5115 page = find_get_entry(mapping, pgoff);
5116 if (xa_is_value(page)) {
5117 swp_entry_t swp = radix_to_swp_entry(page);
5118 if (do_memsw_account())
5119 *entry = swp;
5120 page = find_get_page(swap_address_space(swp),
5121 swp_offset(swp));
5122 }
5123 } else
5124 page = find_get_page(mapping, pgoff);
5125 #else
5126 page = find_get_page(mapping, pgoff);
5127 #endif
5128 return page;
5129 }
5130
5131 /**
5132 * mem_cgroup_move_account - move account of the page
5133 * @page: the page
5134 * @compound: charge the page as compound or small page
5135 * @from: mem_cgroup which the page is moved from.
5136 * @to: mem_cgroup which the page is moved to. @from != @to.
5137 *
5138 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5139 *
5140 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5141 * from old cgroup.
5142 */
5143 static int mem_cgroup_move_account(struct page *page,
5144 bool compound,
5145 struct mem_cgroup *from,
5146 struct mem_cgroup *to)
5147 {
5148 unsigned long flags;
5149 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5150 int ret;
5151 bool anon;
5152
5153 VM_BUG_ON(from == to);
5154 VM_BUG_ON_PAGE(PageLRU(page), page);
5155 VM_BUG_ON(compound && !PageTransHuge(page));
5156
5157 /*
5158 * Prevent mem_cgroup_migrate() from looking at
5159 * page->mem_cgroup of its source page while we change it.
5160 */
5161 ret = -EBUSY;
5162 if (!trylock_page(page))
5163 goto out;
5164
5165 ret = -EINVAL;
5166 if (page->mem_cgroup != from)
5167 goto out_unlock;
5168
5169 anon = PageAnon(page);
5170
5171 spin_lock_irqsave(&from->move_lock, flags);
5172
5173 if (!anon && page_mapped(page)) {
5174 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
5175 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
5176 }
5177
5178 /*
5179 * move_lock grabbed above and caller set from->moving_account, so
5180 * mod_memcg_page_state will serialize updates to PageDirty.
5181 * So mapping should be stable for dirty pages.
5182 */
5183 if (!anon && PageDirty(page)) {
5184 struct address_space *mapping = page_mapping(page);
5185
5186 if (mapping_cap_account_dirty(mapping)) {
5187 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
5188 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
5189 }
5190 }
5191
5192 if (PageWriteback(page)) {
5193 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
5194 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
5195 }
5196
5197 /*
5198 * It is safe to change page->mem_cgroup here because the page
5199 * is referenced, charged, and isolated - we can't race with
5200 * uncharging, charging, migration, or LRU putback.
5201 */
5202
5203 /* caller should have done css_get */
5204 page->mem_cgroup = to;
5205 spin_unlock_irqrestore(&from->move_lock, flags);
5206
5207 ret = 0;
5208
5209 local_irq_disable();
5210 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5211 memcg_check_events(to, page);
5212 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5213 memcg_check_events(from, page);
5214 local_irq_enable();
5215 out_unlock:
5216 unlock_page(page);
5217 out:
5218 return ret;
5219 }
5220
5221 /**
5222 * get_mctgt_type - get target type of moving charge
5223 * @vma: the vma the pte to be checked belongs
5224 * @addr: the address corresponding to the pte to be checked
5225 * @ptent: the pte to be checked
5226 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5227 *
5228 * Returns
5229 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5230 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5231 * move charge. if @target is not NULL, the page is stored in target->page
5232 * with extra refcnt got(Callers should handle it).
5233 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5234 * target for charge migration. if @target is not NULL, the entry is stored
5235 * in target->ent.
5236 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5237 * (so ZONE_DEVICE page and thus not on the lru).
5238 * For now we such page is charge like a regular page would be as for all
5239 * intent and purposes it is just special memory taking the place of a
5240 * regular page.
5241 *
5242 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5243 *
5244 * Called with pte lock held.
5245 */
5246
5247 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5248 unsigned long addr, pte_t ptent, union mc_target *target)
5249 {
5250 struct page *page = NULL;
5251 enum mc_target_type ret = MC_TARGET_NONE;
5252 swp_entry_t ent = { .val = 0 };
5253
5254 if (pte_present(ptent))
5255 page = mc_handle_present_pte(vma, addr, ptent);
5256 else if (is_swap_pte(ptent))
5257 page = mc_handle_swap_pte(vma, ptent, &ent);
5258 else if (pte_none(ptent))
5259 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5260
5261 if (!page && !ent.val)
5262 return ret;
5263 if (page) {
5264 /*
5265 * Do only loose check w/o serialization.
5266 * mem_cgroup_move_account() checks the page is valid or
5267 * not under LRU exclusion.
5268 */
5269 if (page->mem_cgroup == mc.from) {
5270 ret = MC_TARGET_PAGE;
5271 if (is_device_private_page(page))
5272 ret = MC_TARGET_DEVICE;
5273 if (target)
5274 target->page = page;
5275 }
5276 if (!ret || !target)
5277 put_page(page);
5278 }
5279 /*
5280 * There is a swap entry and a page doesn't exist or isn't charged.
5281 * But we cannot move a tail-page in a THP.
5282 */
5283 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5284 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5285 ret = MC_TARGET_SWAP;
5286 if (target)
5287 target->ent = ent;
5288 }
5289 return ret;
5290 }
5291
5292 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5293 /*
5294 * We don't consider PMD mapped swapping or file mapped pages because THP does
5295 * not support them for now.
5296 * Caller should make sure that pmd_trans_huge(pmd) is true.
5297 */
5298 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5299 unsigned long addr, pmd_t pmd, union mc_target *target)
5300 {
5301 struct page *page = NULL;
5302 enum mc_target_type ret = MC_TARGET_NONE;
5303
5304 if (unlikely(is_swap_pmd(pmd))) {
5305 VM_BUG_ON(thp_migration_supported() &&
5306 !is_pmd_migration_entry(pmd));
5307 return ret;
5308 }
5309 page = pmd_page(pmd);
5310 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5311 if (!(mc.flags & MOVE_ANON))
5312 return ret;
5313 if (page->mem_cgroup == mc.from) {
5314 ret = MC_TARGET_PAGE;
5315 if (target) {
5316 get_page(page);
5317 target->page = page;
5318 }
5319 }
5320 return ret;
5321 }
5322 #else
5323 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5324 unsigned long addr, pmd_t pmd, union mc_target *target)
5325 {
5326 return MC_TARGET_NONE;
5327 }
5328 #endif
5329
5330 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5331 unsigned long addr, unsigned long end,
5332 struct mm_walk *walk)
5333 {
5334 struct vm_area_struct *vma = walk->vma;
5335 pte_t *pte;
5336 spinlock_t *ptl;
5337
5338 ptl = pmd_trans_huge_lock(pmd, vma);
5339 if (ptl) {
5340 /*
5341 * Note their can not be MC_TARGET_DEVICE for now as we do not
5342 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5343 * this might change.
5344 */
5345 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5346 mc.precharge += HPAGE_PMD_NR;
5347 spin_unlock(ptl);
5348 return 0;
5349 }
5350
5351 if (pmd_trans_unstable(pmd))
5352 return 0;
5353 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5354 for (; addr != end; pte++, addr += PAGE_SIZE)
5355 if (get_mctgt_type(vma, addr, *pte, NULL))
5356 mc.precharge++; /* increment precharge temporarily */
5357 pte_unmap_unlock(pte - 1, ptl);
5358 cond_resched();
5359
5360 return 0;
5361 }
5362
5363 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5364 {
5365 unsigned long precharge;
5366
5367 struct mm_walk mem_cgroup_count_precharge_walk = {
5368 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5369 .mm = mm,
5370 };
5371 down_read(&mm->mmap_sem);
5372 walk_page_range(0, mm->highest_vm_end,
5373 &mem_cgroup_count_precharge_walk);
5374 up_read(&mm->mmap_sem);
5375
5376 precharge = mc.precharge;
5377 mc.precharge = 0;
5378
5379 return precharge;
5380 }
5381
5382 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5383 {
5384 unsigned long precharge = mem_cgroup_count_precharge(mm);
5385
5386 VM_BUG_ON(mc.moving_task);
5387 mc.moving_task = current;
5388 return mem_cgroup_do_precharge(precharge);
5389 }
5390
5391 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5392 static void __mem_cgroup_clear_mc(void)
5393 {
5394 struct mem_cgroup *from = mc.from;
5395 struct mem_cgroup *to = mc.to;
5396
5397 /* we must uncharge all the leftover precharges from mc.to */
5398 if (mc.precharge) {
5399 cancel_charge(mc.to, mc.precharge);
5400 mc.precharge = 0;
5401 }
5402 /*
5403 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5404 * we must uncharge here.
5405 */
5406 if (mc.moved_charge) {
5407 cancel_charge(mc.from, mc.moved_charge);
5408 mc.moved_charge = 0;
5409 }
5410 /* we must fixup refcnts and charges */
5411 if (mc.moved_swap) {
5412 /* uncharge swap account from the old cgroup */
5413 if (!mem_cgroup_is_root(mc.from))
5414 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5415
5416 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5417
5418 /*
5419 * we charged both to->memory and to->memsw, so we
5420 * should uncharge to->memory.
5421 */
5422 if (!mem_cgroup_is_root(mc.to))
5423 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5424
5425 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5426 css_put_many(&mc.to->css, mc.moved_swap);
5427
5428 mc.moved_swap = 0;
5429 }
5430 memcg_oom_recover(from);
5431 memcg_oom_recover(to);
5432 wake_up_all(&mc.waitq);
5433 }
5434
5435 static void mem_cgroup_clear_mc(void)
5436 {
5437 struct mm_struct *mm = mc.mm;
5438
5439 /*
5440 * we must clear moving_task before waking up waiters at the end of
5441 * task migration.
5442 */
5443 mc.moving_task = NULL;
5444 __mem_cgroup_clear_mc();
5445 spin_lock(&mc.lock);
5446 mc.from = NULL;
5447 mc.to = NULL;
5448 mc.mm = NULL;
5449 spin_unlock(&mc.lock);
5450
5451 mmput(mm);
5452 }
5453
5454 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5455 {
5456 struct cgroup_subsys_state *css;
5457 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5458 struct mem_cgroup *from;
5459 struct task_struct *leader, *p;
5460 struct mm_struct *mm;
5461 unsigned long move_flags;
5462 int ret = 0;
5463
5464 /* charge immigration isn't supported on the default hierarchy */
5465 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5466 return 0;
5467
5468 /*
5469 * Multi-process migrations only happen on the default hierarchy
5470 * where charge immigration is not used. Perform charge
5471 * immigration if @tset contains a leader and whine if there are
5472 * multiple.
5473 */
5474 p = NULL;
5475 cgroup_taskset_for_each_leader(leader, css, tset) {
5476 WARN_ON_ONCE(p);
5477 p = leader;
5478 memcg = mem_cgroup_from_css(css);
5479 }
5480 if (!p)
5481 return 0;
5482
5483 /*
5484 * We are now commited to this value whatever it is. Changes in this
5485 * tunable will only affect upcoming migrations, not the current one.
5486 * So we need to save it, and keep it going.
5487 */
5488 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5489 if (!move_flags)
5490 return 0;
5491
5492 from = mem_cgroup_from_task(p);
5493
5494 VM_BUG_ON(from == memcg);
5495
5496 mm = get_task_mm(p);
5497 if (!mm)
5498 return 0;
5499 /* We move charges only when we move a owner of the mm */
5500 if (mm->owner == p) {
5501 VM_BUG_ON(mc.from);
5502 VM_BUG_ON(mc.to);
5503 VM_BUG_ON(mc.precharge);
5504 VM_BUG_ON(mc.moved_charge);
5505 VM_BUG_ON(mc.moved_swap);
5506
5507 spin_lock(&mc.lock);
5508 mc.mm = mm;
5509 mc.from = from;
5510 mc.to = memcg;
5511 mc.flags = move_flags;
5512 spin_unlock(&mc.lock);
5513 /* We set mc.moving_task later */
5514
5515 ret = mem_cgroup_precharge_mc(mm);
5516 if (ret)
5517 mem_cgroup_clear_mc();
5518 } else {
5519 mmput(mm);
5520 }
5521 return ret;
5522 }
5523
5524 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5525 {
5526 if (mc.to)
5527 mem_cgroup_clear_mc();
5528 }
5529
5530 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5531 unsigned long addr, unsigned long end,
5532 struct mm_walk *walk)
5533 {
5534 int ret = 0;
5535 struct vm_area_struct *vma = walk->vma;
5536 pte_t *pte;
5537 spinlock_t *ptl;
5538 enum mc_target_type target_type;
5539 union mc_target target;
5540 struct page *page;
5541
5542 ptl = pmd_trans_huge_lock(pmd, vma);
5543 if (ptl) {
5544 if (mc.precharge < HPAGE_PMD_NR) {
5545 spin_unlock(ptl);
5546 return 0;
5547 }
5548 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5549 if (target_type == MC_TARGET_PAGE) {
5550 page = target.page;
5551 if (!isolate_lru_page(page)) {
5552 if (!mem_cgroup_move_account(page, true,
5553 mc.from, mc.to)) {
5554 mc.precharge -= HPAGE_PMD_NR;
5555 mc.moved_charge += HPAGE_PMD_NR;
5556 }
5557 putback_lru_page(page);
5558 }
5559 put_page(page);
5560 } else if (target_type == MC_TARGET_DEVICE) {
5561 page = target.page;
5562 if (!mem_cgroup_move_account(page, true,
5563 mc.from, mc.to)) {
5564 mc.precharge -= HPAGE_PMD_NR;
5565 mc.moved_charge += HPAGE_PMD_NR;
5566 }
5567 put_page(page);
5568 }
5569 spin_unlock(ptl);
5570 return 0;
5571 }
5572
5573 if (pmd_trans_unstable(pmd))
5574 return 0;
5575 retry:
5576 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5577 for (; addr != end; addr += PAGE_SIZE) {
5578 pte_t ptent = *(pte++);
5579 bool device = false;
5580 swp_entry_t ent;
5581
5582 if (!mc.precharge)
5583 break;
5584
5585 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5586 case MC_TARGET_DEVICE:
5587 device = true;
5588 /* fall through */
5589 case MC_TARGET_PAGE:
5590 page = target.page;
5591 /*
5592 * We can have a part of the split pmd here. Moving it
5593 * can be done but it would be too convoluted so simply
5594 * ignore such a partial THP and keep it in original
5595 * memcg. There should be somebody mapping the head.
5596 */
5597 if (PageTransCompound(page))
5598 goto put;
5599 if (!device && isolate_lru_page(page))
5600 goto put;
5601 if (!mem_cgroup_move_account(page, false,
5602 mc.from, mc.to)) {
5603 mc.precharge--;
5604 /* we uncharge from mc.from later. */
5605 mc.moved_charge++;
5606 }
5607 if (!device)
5608 putback_lru_page(page);
5609 put: /* get_mctgt_type() gets the page */
5610 put_page(page);
5611 break;
5612 case MC_TARGET_SWAP:
5613 ent = target.ent;
5614 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5615 mc.precharge--;
5616 /* we fixup refcnts and charges later. */
5617 mc.moved_swap++;
5618 }
5619 break;
5620 default:
5621 break;
5622 }
5623 }
5624 pte_unmap_unlock(pte - 1, ptl);
5625 cond_resched();
5626
5627 if (addr != end) {
5628 /*
5629 * We have consumed all precharges we got in can_attach().
5630 * We try charge one by one, but don't do any additional
5631 * charges to mc.to if we have failed in charge once in attach()
5632 * phase.
5633 */
5634 ret = mem_cgroup_do_precharge(1);
5635 if (!ret)
5636 goto retry;
5637 }
5638
5639 return ret;
5640 }
5641
5642 static void mem_cgroup_move_charge(void)
5643 {
5644 struct mm_walk mem_cgroup_move_charge_walk = {
5645 .pmd_entry = mem_cgroup_move_charge_pte_range,
5646 .mm = mc.mm,
5647 };
5648
5649 lru_add_drain_all();
5650 /*
5651 * Signal lock_page_memcg() to take the memcg's move_lock
5652 * while we're moving its pages to another memcg. Then wait
5653 * for already started RCU-only updates to finish.
5654 */
5655 atomic_inc(&mc.from->moving_account);
5656 synchronize_rcu();
5657 retry:
5658 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5659 /*
5660 * Someone who are holding the mmap_sem might be waiting in
5661 * waitq. So we cancel all extra charges, wake up all waiters,
5662 * and retry. Because we cancel precharges, we might not be able
5663 * to move enough charges, but moving charge is a best-effort
5664 * feature anyway, so it wouldn't be a big problem.
5665 */
5666 __mem_cgroup_clear_mc();
5667 cond_resched();
5668 goto retry;
5669 }
5670 /*
5671 * When we have consumed all precharges and failed in doing
5672 * additional charge, the page walk just aborts.
5673 */
5674 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5675
5676 up_read(&mc.mm->mmap_sem);
5677 atomic_dec(&mc.from->moving_account);
5678 }
5679
5680 static void mem_cgroup_move_task(void)
5681 {
5682 if (mc.to) {
5683 mem_cgroup_move_charge();
5684 mem_cgroup_clear_mc();
5685 }
5686 }
5687 #else /* !CONFIG_MMU */
5688 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5689 {
5690 return 0;
5691 }
5692 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5693 {
5694 }
5695 static void mem_cgroup_move_task(void)
5696 {
5697 }
5698 #endif
5699
5700 /*
5701 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5702 * to verify whether we're attached to the default hierarchy on each mount
5703 * attempt.
5704 */
5705 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5706 {
5707 /*
5708 * use_hierarchy is forced on the default hierarchy. cgroup core
5709 * guarantees that @root doesn't have any children, so turning it
5710 * on for the root memcg is enough.
5711 */
5712 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5713 root_mem_cgroup->use_hierarchy = true;
5714 else
5715 root_mem_cgroup->use_hierarchy = false;
5716 }
5717
5718 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5719 {
5720 if (value == PAGE_COUNTER_MAX)
5721 seq_puts(m, "max\n");
5722 else
5723 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5724
5725 return 0;
5726 }
5727
5728 static u64 memory_current_read(struct cgroup_subsys_state *css,
5729 struct cftype *cft)
5730 {
5731 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5732
5733 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5734 }
5735
5736 static int memory_min_show(struct seq_file *m, void *v)
5737 {
5738 return seq_puts_memcg_tunable(m,
5739 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5740 }
5741
5742 static ssize_t memory_min_write(struct kernfs_open_file *of,
5743 char *buf, size_t nbytes, loff_t off)
5744 {
5745 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5746 unsigned long min;
5747 int err;
5748
5749 buf = strstrip(buf);
5750 err = page_counter_memparse(buf, "max", &min);
5751 if (err)
5752 return err;
5753
5754 page_counter_set_min(&memcg->memory, min);
5755
5756 return nbytes;
5757 }
5758
5759 static int memory_low_show(struct seq_file *m, void *v)
5760 {
5761 return seq_puts_memcg_tunable(m,
5762 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5763 }
5764
5765 static ssize_t memory_low_write(struct kernfs_open_file *of,
5766 char *buf, size_t nbytes, loff_t off)
5767 {
5768 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5769 unsigned long low;
5770 int err;
5771
5772 buf = strstrip(buf);
5773 err = page_counter_memparse(buf, "max", &low);
5774 if (err)
5775 return err;
5776
5777 page_counter_set_low(&memcg->memory, low);
5778
5779 return nbytes;
5780 }
5781
5782 static int memory_high_show(struct seq_file *m, void *v)
5783 {
5784 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5785 }
5786
5787 static ssize_t memory_high_write(struct kernfs_open_file *of,
5788 char *buf, size_t nbytes, loff_t off)
5789 {
5790 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5791 unsigned long nr_pages;
5792 unsigned long high;
5793 int err;
5794
5795 buf = strstrip(buf);
5796 err = page_counter_memparse(buf, "max", &high);
5797 if (err)
5798 return err;
5799
5800 memcg->high = high;
5801
5802 nr_pages = page_counter_read(&memcg->memory);
5803 if (nr_pages > high)
5804 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5805 GFP_KERNEL, true);
5806
5807 memcg_wb_domain_size_changed(memcg);
5808 return nbytes;
5809 }
5810
5811 static int memory_max_show(struct seq_file *m, void *v)
5812 {
5813 return seq_puts_memcg_tunable(m,
5814 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
5815 }
5816
5817 static ssize_t memory_max_write(struct kernfs_open_file *of,
5818 char *buf, size_t nbytes, loff_t off)
5819 {
5820 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5821 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5822 bool drained = false;
5823 unsigned long max;
5824 int err;
5825
5826 buf = strstrip(buf);
5827 err = page_counter_memparse(buf, "max", &max);
5828 if (err)
5829 return err;
5830
5831 xchg(&memcg->memory.max, max);
5832
5833 for (;;) {
5834 unsigned long nr_pages = page_counter_read(&memcg->memory);
5835
5836 if (nr_pages <= max)
5837 break;
5838
5839 if (signal_pending(current)) {
5840 err = -EINTR;
5841 break;
5842 }
5843
5844 if (!drained) {
5845 drain_all_stock(memcg);
5846 drained = true;
5847 continue;
5848 }
5849
5850 if (nr_reclaims) {
5851 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5852 GFP_KERNEL, true))
5853 nr_reclaims--;
5854 continue;
5855 }
5856
5857 memcg_memory_event(memcg, MEMCG_OOM);
5858 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5859 break;
5860 }
5861
5862 memcg_wb_domain_size_changed(memcg);
5863 return nbytes;
5864 }
5865
5866 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
5867 {
5868 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
5869 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
5870 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
5871 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
5872 seq_printf(m, "oom_kill %lu\n",
5873 atomic_long_read(&events[MEMCG_OOM_KILL]));
5874 }
5875
5876 static int memory_events_show(struct seq_file *m, void *v)
5877 {
5878 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5879
5880 __memory_events_show(m, memcg->memory_events);
5881 return 0;
5882 }
5883
5884 static int memory_events_local_show(struct seq_file *m, void *v)
5885 {
5886 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5887
5888 __memory_events_show(m, memcg->memory_events_local);
5889 return 0;
5890 }
5891
5892 static int memory_stat_show(struct seq_file *m, void *v)
5893 {
5894 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5895 char *buf;
5896
5897 buf = memory_stat_format(memcg);
5898 if (!buf)
5899 return -ENOMEM;
5900 seq_puts(m, buf);
5901 kfree(buf);
5902 return 0;
5903 }
5904
5905 static int memory_oom_group_show(struct seq_file *m, void *v)
5906 {
5907 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
5908
5909 seq_printf(m, "%d\n", memcg->oom_group);
5910
5911 return 0;
5912 }
5913
5914 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
5915 char *buf, size_t nbytes, loff_t off)
5916 {
5917 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5918 int ret, oom_group;
5919
5920 buf = strstrip(buf);
5921 if (!buf)
5922 return -EINVAL;
5923
5924 ret = kstrtoint(buf, 0, &oom_group);
5925 if (ret)
5926 return ret;
5927
5928 if (oom_group != 0 && oom_group != 1)
5929 return -EINVAL;
5930
5931 memcg->oom_group = oom_group;
5932
5933 return nbytes;
5934 }
5935
5936 static struct cftype memory_files[] = {
5937 {
5938 .name = "current",
5939 .flags = CFTYPE_NOT_ON_ROOT,
5940 .read_u64 = memory_current_read,
5941 },
5942 {
5943 .name = "min",
5944 .flags = CFTYPE_NOT_ON_ROOT,
5945 .seq_show = memory_min_show,
5946 .write = memory_min_write,
5947 },
5948 {
5949 .name = "low",
5950 .flags = CFTYPE_NOT_ON_ROOT,
5951 .seq_show = memory_low_show,
5952 .write = memory_low_write,
5953 },
5954 {
5955 .name = "high",
5956 .flags = CFTYPE_NOT_ON_ROOT,
5957 .seq_show = memory_high_show,
5958 .write = memory_high_write,
5959 },
5960 {
5961 .name = "max",
5962 .flags = CFTYPE_NOT_ON_ROOT,
5963 .seq_show = memory_max_show,
5964 .write = memory_max_write,
5965 },
5966 {
5967 .name = "events",
5968 .flags = CFTYPE_NOT_ON_ROOT,
5969 .file_offset = offsetof(struct mem_cgroup, events_file),
5970 .seq_show = memory_events_show,
5971 },
5972 {
5973 .name = "events.local",
5974 .flags = CFTYPE_NOT_ON_ROOT,
5975 .file_offset = offsetof(struct mem_cgroup, events_local_file),
5976 .seq_show = memory_events_local_show,
5977 },
5978 {
5979 .name = "stat",
5980 .flags = CFTYPE_NOT_ON_ROOT,
5981 .seq_show = memory_stat_show,
5982 },
5983 {
5984 .name = "oom.group",
5985 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
5986 .seq_show = memory_oom_group_show,
5987 .write = memory_oom_group_write,
5988 },
5989 { } /* terminate */
5990 };
5991
5992 struct cgroup_subsys memory_cgrp_subsys = {
5993 .css_alloc = mem_cgroup_css_alloc,
5994 .css_online = mem_cgroup_css_online,
5995 .css_offline = mem_cgroup_css_offline,
5996 .css_released = mem_cgroup_css_released,
5997 .css_free = mem_cgroup_css_free,
5998 .css_reset = mem_cgroup_css_reset,
5999 .can_attach = mem_cgroup_can_attach,
6000 .cancel_attach = mem_cgroup_cancel_attach,
6001 .post_attach = mem_cgroup_move_task,
6002 .bind = mem_cgroup_bind,
6003 .dfl_cftypes = memory_files,
6004 .legacy_cftypes = mem_cgroup_legacy_files,
6005 .early_init = 0,
6006 };
6007
6008 /**
6009 * mem_cgroup_protected - check if memory consumption is in the normal range
6010 * @root: the top ancestor of the sub-tree being checked
6011 * @memcg: the memory cgroup to check
6012 *
6013 * WARNING: This function is not stateless! It can only be used as part
6014 * of a top-down tree iteration, not for isolated queries.
6015 *
6016 * Returns one of the following:
6017 * MEMCG_PROT_NONE: cgroup memory is not protected
6018 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6019 * an unprotected supply of reclaimable memory from other cgroups.
6020 * MEMCG_PROT_MIN: cgroup memory is protected
6021 *
6022 * @root is exclusive; it is never protected when looked at directly
6023 *
6024 * To provide a proper hierarchical behavior, effective memory.min/low values
6025 * are used. Below is the description of how effective memory.low is calculated.
6026 * Effective memory.min values is calculated in the same way.
6027 *
6028 * Effective memory.low is always equal or less than the original memory.low.
6029 * If there is no memory.low overcommittment (which is always true for
6030 * top-level memory cgroups), these two values are equal.
6031 * Otherwise, it's a part of parent's effective memory.low,
6032 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6033 * memory.low usages, where memory.low usage is the size of actually
6034 * protected memory.
6035 *
6036 * low_usage
6037 * elow = min( memory.low, parent->elow * ------------------ ),
6038 * siblings_low_usage
6039 *
6040 * | memory.current, if memory.current < memory.low
6041 * low_usage = |
6042 * | 0, otherwise.
6043 *
6044 *
6045 * Such definition of the effective memory.low provides the expected
6046 * hierarchical behavior: parent's memory.low value is limiting
6047 * children, unprotected memory is reclaimed first and cgroups,
6048 * which are not using their guarantee do not affect actual memory
6049 * distribution.
6050 *
6051 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6052 *
6053 * A A/memory.low = 2G, A/memory.current = 6G
6054 * //\\
6055 * BC DE B/memory.low = 3G B/memory.current = 2G
6056 * C/memory.low = 1G C/memory.current = 2G
6057 * D/memory.low = 0 D/memory.current = 2G
6058 * E/memory.low = 10G E/memory.current = 0
6059 *
6060 * and the memory pressure is applied, the following memory distribution
6061 * is expected (approximately):
6062 *
6063 * A/memory.current = 2G
6064 *
6065 * B/memory.current = 1.3G
6066 * C/memory.current = 0.6G
6067 * D/memory.current = 0
6068 * E/memory.current = 0
6069 *
6070 * These calculations require constant tracking of the actual low usages
6071 * (see propagate_protected_usage()), as well as recursive calculation of
6072 * effective memory.low values. But as we do call mem_cgroup_protected()
6073 * path for each memory cgroup top-down from the reclaim,
6074 * it's possible to optimize this part, and save calculated elow
6075 * for next usage. This part is intentionally racy, but it's ok,
6076 * as memory.low is a best-effort mechanism.
6077 */
6078 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6079 struct mem_cgroup *memcg)
6080 {
6081 struct mem_cgroup *parent;
6082 unsigned long emin, parent_emin;
6083 unsigned long elow, parent_elow;
6084 unsigned long usage;
6085
6086 if (mem_cgroup_disabled())
6087 return MEMCG_PROT_NONE;
6088
6089 if (!root)
6090 root = root_mem_cgroup;
6091 if (memcg == root)
6092 return MEMCG_PROT_NONE;
6093
6094 usage = page_counter_read(&memcg->memory);
6095 if (!usage)
6096 return MEMCG_PROT_NONE;
6097
6098 emin = memcg->memory.min;
6099 elow = memcg->memory.low;
6100
6101 parent = parent_mem_cgroup(memcg);
6102 /* No parent means a non-hierarchical mode on v1 memcg */
6103 if (!parent)
6104 return MEMCG_PROT_NONE;
6105
6106 if (parent == root)
6107 goto exit;
6108
6109 parent_emin = READ_ONCE(parent->memory.emin);
6110 emin = min(emin, parent_emin);
6111 if (emin && parent_emin) {
6112 unsigned long min_usage, siblings_min_usage;
6113
6114 min_usage = min(usage, memcg->memory.min);
6115 siblings_min_usage = atomic_long_read(
6116 &parent->memory.children_min_usage);
6117
6118 if (min_usage && siblings_min_usage)
6119 emin = min(emin, parent_emin * min_usage /
6120 siblings_min_usage);
6121 }
6122
6123 parent_elow = READ_ONCE(parent->memory.elow);
6124 elow = min(elow, parent_elow);
6125 if (elow && parent_elow) {
6126 unsigned long low_usage, siblings_low_usage;
6127
6128 low_usage = min(usage, memcg->memory.low);
6129 siblings_low_usage = atomic_long_read(
6130 &parent->memory.children_low_usage);
6131
6132 if (low_usage && siblings_low_usage)
6133 elow = min(elow, parent_elow * low_usage /
6134 siblings_low_usage);
6135 }
6136
6137 exit:
6138 memcg->memory.emin = emin;
6139 memcg->memory.elow = elow;
6140
6141 if (usage <= emin)
6142 return MEMCG_PROT_MIN;
6143 else if (usage <= elow)
6144 return MEMCG_PROT_LOW;
6145 else
6146 return MEMCG_PROT_NONE;
6147 }
6148
6149 /**
6150 * mem_cgroup_try_charge - try charging a page
6151 * @page: page to charge
6152 * @mm: mm context of the victim
6153 * @gfp_mask: reclaim mode
6154 * @memcgp: charged memcg return
6155 * @compound: charge the page as compound or small page
6156 *
6157 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6158 * pages according to @gfp_mask if necessary.
6159 *
6160 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6161 * Otherwise, an error code is returned.
6162 *
6163 * After page->mapping has been set up, the caller must finalize the
6164 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6165 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6166 */
6167 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6168 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6169 bool compound)
6170 {
6171 struct mem_cgroup *memcg = NULL;
6172 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6173 int ret = 0;
6174
6175 if (mem_cgroup_disabled())
6176 goto out;
6177
6178 if (PageSwapCache(page)) {
6179 /*
6180 * Every swap fault against a single page tries to charge the
6181 * page, bail as early as possible. shmem_unuse() encounters
6182 * already charged pages, too. The USED bit is protected by
6183 * the page lock, which serializes swap cache removal, which
6184 * in turn serializes uncharging.
6185 */
6186 VM_BUG_ON_PAGE(!PageLocked(page), page);
6187 if (compound_head(page)->mem_cgroup)
6188 goto out;
6189
6190 if (do_swap_account) {
6191 swp_entry_t ent = { .val = page_private(page), };
6192 unsigned short id = lookup_swap_cgroup_id(ent);
6193
6194 rcu_read_lock();
6195 memcg = mem_cgroup_from_id(id);
6196 if (memcg && !css_tryget_online(&memcg->css))
6197 memcg = NULL;
6198 rcu_read_unlock();
6199 }
6200 }
6201
6202 if (!memcg)
6203 memcg = get_mem_cgroup_from_mm(mm);
6204
6205 ret = try_charge(memcg, gfp_mask, nr_pages);
6206
6207 css_put(&memcg->css);
6208 out:
6209 *memcgp = memcg;
6210 return ret;
6211 }
6212
6213 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6214 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6215 bool compound)
6216 {
6217 struct mem_cgroup *memcg;
6218 int ret;
6219
6220 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6221 memcg = *memcgp;
6222 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6223 return ret;
6224 }
6225
6226 /**
6227 * mem_cgroup_commit_charge - commit a page charge
6228 * @page: page to charge
6229 * @memcg: memcg to charge the page to
6230 * @lrucare: page might be on LRU already
6231 * @compound: charge the page as compound or small page
6232 *
6233 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6234 * after page->mapping has been set up. This must happen atomically
6235 * as part of the page instantiation, i.e. under the page table lock
6236 * for anonymous pages, under the page lock for page and swap cache.
6237 *
6238 * In addition, the page must not be on the LRU during the commit, to
6239 * prevent racing with task migration. If it might be, use @lrucare.
6240 *
6241 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6242 */
6243 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6244 bool lrucare, bool compound)
6245 {
6246 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6247
6248 VM_BUG_ON_PAGE(!page->mapping, page);
6249 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6250
6251 if (mem_cgroup_disabled())
6252 return;
6253 /*
6254 * Swap faults will attempt to charge the same page multiple
6255 * times. But reuse_swap_page() might have removed the page
6256 * from swapcache already, so we can't check PageSwapCache().
6257 */
6258 if (!memcg)
6259 return;
6260
6261 commit_charge(page, memcg, lrucare);
6262
6263 local_irq_disable();
6264 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6265 memcg_check_events(memcg, page);
6266 local_irq_enable();
6267
6268 if (do_memsw_account() && PageSwapCache(page)) {
6269 swp_entry_t entry = { .val = page_private(page) };
6270 /*
6271 * The swap entry might not get freed for a long time,
6272 * let's not wait for it. The page already received a
6273 * memory+swap charge, drop the swap entry duplicate.
6274 */
6275 mem_cgroup_uncharge_swap(entry, nr_pages);
6276 }
6277 }
6278
6279 /**
6280 * mem_cgroup_cancel_charge - cancel a page charge
6281 * @page: page to charge
6282 * @memcg: memcg to charge the page to
6283 * @compound: charge the page as compound or small page
6284 *
6285 * Cancel a charge transaction started by mem_cgroup_try_charge().
6286 */
6287 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6288 bool compound)
6289 {
6290 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6291
6292 if (mem_cgroup_disabled())
6293 return;
6294 /*
6295 * Swap faults will attempt to charge the same page multiple
6296 * times. But reuse_swap_page() might have removed the page
6297 * from swapcache already, so we can't check PageSwapCache().
6298 */
6299 if (!memcg)
6300 return;
6301
6302 cancel_charge(memcg, nr_pages);
6303 }
6304
6305 struct uncharge_gather {
6306 struct mem_cgroup *memcg;
6307 unsigned long pgpgout;
6308 unsigned long nr_anon;
6309 unsigned long nr_file;
6310 unsigned long nr_kmem;
6311 unsigned long nr_huge;
6312 unsigned long nr_shmem;
6313 struct page *dummy_page;
6314 };
6315
6316 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6317 {
6318 memset(ug, 0, sizeof(*ug));
6319 }
6320
6321 static void uncharge_batch(const struct uncharge_gather *ug)
6322 {
6323 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6324 unsigned long flags;
6325
6326 if (!mem_cgroup_is_root(ug->memcg)) {
6327 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6328 if (do_memsw_account())
6329 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6330 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6331 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6332 memcg_oom_recover(ug->memcg);
6333 }
6334
6335 local_irq_save(flags);
6336 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6337 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6338 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6339 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6340 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6341 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6342 memcg_check_events(ug->memcg, ug->dummy_page);
6343 local_irq_restore(flags);
6344
6345 if (!mem_cgroup_is_root(ug->memcg))
6346 css_put_many(&ug->memcg->css, nr_pages);
6347 }
6348
6349 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6350 {
6351 VM_BUG_ON_PAGE(PageLRU(page), page);
6352 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6353 !PageHWPoison(page) , page);
6354
6355 if (!page->mem_cgroup)
6356 return;
6357
6358 /*
6359 * Nobody should be changing or seriously looking at
6360 * page->mem_cgroup at this point, we have fully
6361 * exclusive access to the page.
6362 */
6363
6364 if (ug->memcg != page->mem_cgroup) {
6365 if (ug->memcg) {
6366 uncharge_batch(ug);
6367 uncharge_gather_clear(ug);
6368 }
6369 ug->memcg = page->mem_cgroup;
6370 }
6371
6372 if (!PageKmemcg(page)) {
6373 unsigned int nr_pages = 1;
6374
6375 if (PageTransHuge(page)) {
6376 nr_pages <<= compound_order(page);
6377 ug->nr_huge += nr_pages;
6378 }
6379 if (PageAnon(page))
6380 ug->nr_anon += nr_pages;
6381 else {
6382 ug->nr_file += nr_pages;
6383 if (PageSwapBacked(page))
6384 ug->nr_shmem += nr_pages;
6385 }
6386 ug->pgpgout++;
6387 } else {
6388 ug->nr_kmem += 1 << compound_order(page);
6389 __ClearPageKmemcg(page);
6390 }
6391
6392 ug->dummy_page = page;
6393 page->mem_cgroup = NULL;
6394 }
6395
6396 static void uncharge_list(struct list_head *page_list)
6397 {
6398 struct uncharge_gather ug;
6399 struct list_head *next;
6400
6401 uncharge_gather_clear(&ug);
6402
6403 /*
6404 * Note that the list can be a single page->lru; hence the
6405 * do-while loop instead of a simple list_for_each_entry().
6406 */
6407 next = page_list->next;
6408 do {
6409 struct page *page;
6410
6411 page = list_entry(next, struct page, lru);
6412 next = page->lru.next;
6413
6414 uncharge_page(page, &ug);
6415 } while (next != page_list);
6416
6417 if (ug.memcg)
6418 uncharge_batch(&ug);
6419 }
6420
6421 /**
6422 * mem_cgroup_uncharge - uncharge a page
6423 * @page: page to uncharge
6424 *
6425 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6426 * mem_cgroup_commit_charge().
6427 */
6428 void mem_cgroup_uncharge(struct page *page)
6429 {
6430 struct uncharge_gather ug;
6431
6432 if (mem_cgroup_disabled())
6433 return;
6434
6435 /* Don't touch page->lru of any random page, pre-check: */
6436 if (!page->mem_cgroup)
6437 return;
6438
6439 uncharge_gather_clear(&ug);
6440 uncharge_page(page, &ug);
6441 uncharge_batch(&ug);
6442 }
6443
6444 /**
6445 * mem_cgroup_uncharge_list - uncharge a list of page
6446 * @page_list: list of pages to uncharge
6447 *
6448 * Uncharge a list of pages previously charged with
6449 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6450 */
6451 void mem_cgroup_uncharge_list(struct list_head *page_list)
6452 {
6453 if (mem_cgroup_disabled())
6454 return;
6455
6456 if (!list_empty(page_list))
6457 uncharge_list(page_list);
6458 }
6459
6460 /**
6461 * mem_cgroup_migrate - charge a page's replacement
6462 * @oldpage: currently circulating page
6463 * @newpage: replacement page
6464 *
6465 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6466 * be uncharged upon free.
6467 *
6468 * Both pages must be locked, @newpage->mapping must be set up.
6469 */
6470 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6471 {
6472 struct mem_cgroup *memcg;
6473 unsigned int nr_pages;
6474 bool compound;
6475 unsigned long flags;
6476
6477 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6478 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6479 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6480 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6481 newpage);
6482
6483 if (mem_cgroup_disabled())
6484 return;
6485
6486 /* Page cache replacement: new page already charged? */
6487 if (newpage->mem_cgroup)
6488 return;
6489
6490 /* Swapcache readahead pages can get replaced before being charged */
6491 memcg = oldpage->mem_cgroup;
6492 if (!memcg)
6493 return;
6494
6495 /* Force-charge the new page. The old one will be freed soon */
6496 compound = PageTransHuge(newpage);
6497 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6498
6499 page_counter_charge(&memcg->memory, nr_pages);
6500 if (do_memsw_account())
6501 page_counter_charge(&memcg->memsw, nr_pages);
6502 css_get_many(&memcg->css, nr_pages);
6503
6504 commit_charge(newpage, memcg, false);
6505
6506 local_irq_save(flags);
6507 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6508 memcg_check_events(memcg, newpage);
6509 local_irq_restore(flags);
6510 }
6511
6512 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6513 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6514
6515 void mem_cgroup_sk_alloc(struct sock *sk)
6516 {
6517 struct mem_cgroup *memcg;
6518
6519 if (!mem_cgroup_sockets_enabled)
6520 return;
6521
6522 /*
6523 * Socket cloning can throw us here with sk_memcg already
6524 * filled. It won't however, necessarily happen from
6525 * process context. So the test for root memcg given
6526 * the current task's memcg won't help us in this case.
6527 *
6528 * Respecting the original socket's memcg is a better
6529 * decision in this case.
6530 */
6531 if (sk->sk_memcg) {
6532 css_get(&sk->sk_memcg->css);
6533 return;
6534 }
6535
6536 rcu_read_lock();
6537 memcg = mem_cgroup_from_task(current);
6538 if (memcg == root_mem_cgroup)
6539 goto out;
6540 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6541 goto out;
6542 if (css_tryget_online(&memcg->css))
6543 sk->sk_memcg = memcg;
6544 out:
6545 rcu_read_unlock();
6546 }
6547
6548 void mem_cgroup_sk_free(struct sock *sk)
6549 {
6550 if (sk->sk_memcg)
6551 css_put(&sk->sk_memcg->css);
6552 }
6553
6554 /**
6555 * mem_cgroup_charge_skmem - charge socket memory
6556 * @memcg: memcg to charge
6557 * @nr_pages: number of pages to charge
6558 *
6559 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6560 * @memcg's configured limit, %false if the charge had to be forced.
6561 */
6562 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6563 {
6564 gfp_t gfp_mask = GFP_KERNEL;
6565
6566 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6567 struct page_counter *fail;
6568
6569 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6570 memcg->tcpmem_pressure = 0;
6571 return true;
6572 }
6573 page_counter_charge(&memcg->tcpmem, nr_pages);
6574 memcg->tcpmem_pressure = 1;
6575 return false;
6576 }
6577
6578 /* Don't block in the packet receive path */
6579 if (in_softirq())
6580 gfp_mask = GFP_NOWAIT;
6581
6582 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6583
6584 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6585 return true;
6586
6587 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6588 return false;
6589 }
6590
6591 /**
6592 * mem_cgroup_uncharge_skmem - uncharge socket memory
6593 * @memcg: memcg to uncharge
6594 * @nr_pages: number of pages to uncharge
6595 */
6596 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6597 {
6598 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6599 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6600 return;
6601 }
6602
6603 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6604
6605 refill_stock(memcg, nr_pages);
6606 }
6607
6608 static int __init cgroup_memory(char *s)
6609 {
6610 char *token;
6611
6612 while ((token = strsep(&s, ",")) != NULL) {
6613 if (!*token)
6614 continue;
6615 if (!strcmp(token, "nosocket"))
6616 cgroup_memory_nosocket = true;
6617 if (!strcmp(token, "nokmem"))
6618 cgroup_memory_nokmem = true;
6619 }
6620 return 0;
6621 }
6622 __setup("cgroup.memory=", cgroup_memory);
6623
6624 /*
6625 * subsys_initcall() for memory controller.
6626 *
6627 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6628 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6629 * basically everything that doesn't depend on a specific mem_cgroup structure
6630 * should be initialized from here.
6631 */
6632 static int __init mem_cgroup_init(void)
6633 {
6634 int cpu, node;
6635
6636 #ifdef CONFIG_MEMCG_KMEM
6637 /*
6638 * Kmem cache creation is mostly done with the slab_mutex held,
6639 * so use a workqueue with limited concurrency to avoid stalling
6640 * all worker threads in case lots of cgroups are created and
6641 * destroyed simultaneously.
6642 */
6643 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6644 BUG_ON(!memcg_kmem_cache_wq);
6645 #endif
6646
6647 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6648 memcg_hotplug_cpu_dead);
6649
6650 for_each_possible_cpu(cpu)
6651 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6652 drain_local_stock);
6653
6654 for_each_node(node) {
6655 struct mem_cgroup_tree_per_node *rtpn;
6656
6657 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6658 node_online(node) ? node : NUMA_NO_NODE);
6659
6660 rtpn->rb_root = RB_ROOT;
6661 rtpn->rb_rightmost = NULL;
6662 spin_lock_init(&rtpn->lock);
6663 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6664 }
6665
6666 return 0;
6667 }
6668 subsys_initcall(mem_cgroup_init);
6669
6670 #ifdef CONFIG_MEMCG_SWAP
6671 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6672 {
6673 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6674 /*
6675 * The root cgroup cannot be destroyed, so it's refcount must
6676 * always be >= 1.
6677 */
6678 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6679 VM_BUG_ON(1);
6680 break;
6681 }
6682 memcg = parent_mem_cgroup(memcg);
6683 if (!memcg)
6684 memcg = root_mem_cgroup;
6685 }
6686 return memcg;
6687 }
6688
6689 /**
6690 * mem_cgroup_swapout - transfer a memsw charge to swap
6691 * @page: page whose memsw charge to transfer
6692 * @entry: swap entry to move the charge to
6693 *
6694 * Transfer the memsw charge of @page to @entry.
6695 */
6696 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6697 {
6698 struct mem_cgroup *memcg, *swap_memcg;
6699 unsigned int nr_entries;
6700 unsigned short oldid;
6701
6702 VM_BUG_ON_PAGE(PageLRU(page), page);
6703 VM_BUG_ON_PAGE(page_count(page), page);
6704
6705 if (!do_memsw_account())
6706 return;
6707
6708 memcg = page->mem_cgroup;
6709
6710 /* Readahead page, never charged */
6711 if (!memcg)
6712 return;
6713
6714 /*
6715 * In case the memcg owning these pages has been offlined and doesn't
6716 * have an ID allocated to it anymore, charge the closest online
6717 * ancestor for the swap instead and transfer the memory+swap charge.
6718 */
6719 swap_memcg = mem_cgroup_id_get_online(memcg);
6720 nr_entries = hpage_nr_pages(page);
6721 /* Get references for the tail pages, too */
6722 if (nr_entries > 1)
6723 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6724 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6725 nr_entries);
6726 VM_BUG_ON_PAGE(oldid, page);
6727 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6728
6729 page->mem_cgroup = NULL;
6730
6731 if (!mem_cgroup_is_root(memcg))
6732 page_counter_uncharge(&memcg->memory, nr_entries);
6733
6734 if (memcg != swap_memcg) {
6735 if (!mem_cgroup_is_root(swap_memcg))
6736 page_counter_charge(&swap_memcg->memsw, nr_entries);
6737 page_counter_uncharge(&memcg->memsw, nr_entries);
6738 }
6739
6740 /*
6741 * Interrupts should be disabled here because the caller holds the
6742 * i_pages lock which is taken with interrupts-off. It is
6743 * important here to have the interrupts disabled because it is the
6744 * only synchronisation we have for updating the per-CPU variables.
6745 */
6746 VM_BUG_ON(!irqs_disabled());
6747 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6748 -nr_entries);
6749 memcg_check_events(memcg, page);
6750
6751 if (!mem_cgroup_is_root(memcg))
6752 css_put_many(&memcg->css, nr_entries);
6753 }
6754
6755 /**
6756 * mem_cgroup_try_charge_swap - try charging swap space for a page
6757 * @page: page being added to swap
6758 * @entry: swap entry to charge
6759 *
6760 * Try to charge @page's memcg for the swap space at @entry.
6761 *
6762 * Returns 0 on success, -ENOMEM on failure.
6763 */
6764 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6765 {
6766 unsigned int nr_pages = hpage_nr_pages(page);
6767 struct page_counter *counter;
6768 struct mem_cgroup *memcg;
6769 unsigned short oldid;
6770
6771 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6772 return 0;
6773
6774 memcg = page->mem_cgroup;
6775
6776 /* Readahead page, never charged */
6777 if (!memcg)
6778 return 0;
6779
6780 if (!entry.val) {
6781 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6782 return 0;
6783 }
6784
6785 memcg = mem_cgroup_id_get_online(memcg);
6786
6787 if (!mem_cgroup_is_root(memcg) &&
6788 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6789 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6790 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6791 mem_cgroup_id_put(memcg);
6792 return -ENOMEM;
6793 }
6794
6795 /* Get references for the tail pages, too */
6796 if (nr_pages > 1)
6797 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6798 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6799 VM_BUG_ON_PAGE(oldid, page);
6800 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6801
6802 return 0;
6803 }
6804
6805 /**
6806 * mem_cgroup_uncharge_swap - uncharge swap space
6807 * @entry: swap entry to uncharge
6808 * @nr_pages: the amount of swap space to uncharge
6809 */
6810 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6811 {
6812 struct mem_cgroup *memcg;
6813 unsigned short id;
6814
6815 if (!do_swap_account)
6816 return;
6817
6818 id = swap_cgroup_record(entry, 0, nr_pages);
6819 rcu_read_lock();
6820 memcg = mem_cgroup_from_id(id);
6821 if (memcg) {
6822 if (!mem_cgroup_is_root(memcg)) {
6823 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6824 page_counter_uncharge(&memcg->swap, nr_pages);
6825 else
6826 page_counter_uncharge(&memcg->memsw, nr_pages);
6827 }
6828 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6829 mem_cgroup_id_put_many(memcg, nr_pages);
6830 }
6831 rcu_read_unlock();
6832 }
6833
6834 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6835 {
6836 long nr_swap_pages = get_nr_swap_pages();
6837
6838 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6839 return nr_swap_pages;
6840 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6841 nr_swap_pages = min_t(long, nr_swap_pages,
6842 READ_ONCE(memcg->swap.max) -
6843 page_counter_read(&memcg->swap));
6844 return nr_swap_pages;
6845 }
6846
6847 bool mem_cgroup_swap_full(struct page *page)
6848 {
6849 struct mem_cgroup *memcg;
6850
6851 VM_BUG_ON_PAGE(!PageLocked(page), page);
6852
6853 if (vm_swap_full())
6854 return true;
6855 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6856 return false;
6857
6858 memcg = page->mem_cgroup;
6859 if (!memcg)
6860 return false;
6861
6862 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6863 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6864 return true;
6865
6866 return false;
6867 }
6868
6869 /* for remember boot option*/
6870 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6871 static int really_do_swap_account __initdata = 1;
6872 #else
6873 static int really_do_swap_account __initdata;
6874 #endif
6875
6876 static int __init enable_swap_account(char *s)
6877 {
6878 if (!strcmp(s, "1"))
6879 really_do_swap_account = 1;
6880 else if (!strcmp(s, "0"))
6881 really_do_swap_account = 0;
6882 return 1;
6883 }
6884 __setup("swapaccount=", enable_swap_account);
6885
6886 static u64 swap_current_read(struct cgroup_subsys_state *css,
6887 struct cftype *cft)
6888 {
6889 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6890
6891 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6892 }
6893
6894 static int swap_max_show(struct seq_file *m, void *v)
6895 {
6896 return seq_puts_memcg_tunable(m,
6897 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
6898 }
6899
6900 static ssize_t swap_max_write(struct kernfs_open_file *of,
6901 char *buf, size_t nbytes, loff_t off)
6902 {
6903 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6904 unsigned long max;
6905 int err;
6906
6907 buf = strstrip(buf);
6908 err = page_counter_memparse(buf, "max", &max);
6909 if (err)
6910 return err;
6911
6912 xchg(&memcg->swap.max, max);
6913
6914 return nbytes;
6915 }
6916
6917 static int swap_events_show(struct seq_file *m, void *v)
6918 {
6919 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6920
6921 seq_printf(m, "max %lu\n",
6922 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6923 seq_printf(m, "fail %lu\n",
6924 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6925
6926 return 0;
6927 }
6928
6929 static struct cftype swap_files[] = {
6930 {
6931 .name = "swap.current",
6932 .flags = CFTYPE_NOT_ON_ROOT,
6933 .read_u64 = swap_current_read,
6934 },
6935 {
6936 .name = "swap.max",
6937 .flags = CFTYPE_NOT_ON_ROOT,
6938 .seq_show = swap_max_show,
6939 .write = swap_max_write,
6940 },
6941 {
6942 .name = "swap.events",
6943 .flags = CFTYPE_NOT_ON_ROOT,
6944 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6945 .seq_show = swap_events_show,
6946 },
6947 { } /* terminate */
6948 };
6949
6950 static struct cftype memsw_cgroup_files[] = {
6951 {
6952 .name = "memsw.usage_in_bytes",
6953 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6954 .read_u64 = mem_cgroup_read_u64,
6955 },
6956 {
6957 .name = "memsw.max_usage_in_bytes",
6958 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6959 .write = mem_cgroup_reset,
6960 .read_u64 = mem_cgroup_read_u64,
6961 },
6962 {
6963 .name = "memsw.limit_in_bytes",
6964 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6965 .write = mem_cgroup_write,
6966 .read_u64 = mem_cgroup_read_u64,
6967 },
6968 {
6969 .name = "memsw.failcnt",
6970 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6971 .write = mem_cgroup_reset,
6972 .read_u64 = mem_cgroup_read_u64,
6973 },
6974 { }, /* terminate */
6975 };
6976
6977 static int __init mem_cgroup_swap_init(void)
6978 {
6979 if (!mem_cgroup_disabled() && really_do_swap_account) {
6980 do_swap_account = 1;
6981 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6982 swap_files));
6983 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6984 memsw_cgroup_files));
6985 }
6986 return 0;
6987 }
6988 subsys_initcall(mem_cgroup_swap_init);
6989
6990 #endif /* CONFIG_MEMCG_SWAP */