]>
Commit | Line | Data |
---|---|---|
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/pagewalk.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/psi.h> | |
61 | #include <linux/seq_buf.h> | |
62 | #include "internal.h" | |
63 | #include <net/sock.h> | |
64 | #include <net/ip.h> | |
65 | #include "slab.h" | |
66 | ||
67 | #include <linux/uaccess.h> | |
68 | ||
69 | #include <trace/events/vmscan.h> | |
70 | ||
71 | struct cgroup_subsys memory_cgrp_subsys __read_mostly; | |
72 | EXPORT_SYMBOL(memory_cgrp_subsys); | |
73 | ||
74 | struct mem_cgroup *root_mem_cgroup __read_mostly; | |
75 | ||
76 | #define MEM_CGROUP_RECLAIM_RETRIES 5 | |
77 | ||
78 | /* Socket memory accounting disabled? */ | |
79 | static bool cgroup_memory_nosocket; | |
80 | ||
81 | /* Kernel memory accounting disabled? */ | |
82 | static bool cgroup_memory_nokmem; | |
83 | ||
84 | /* Whether the swap controller is active */ | |
85 | #ifdef CONFIG_MEMCG_SWAP | |
86 | int do_swap_account __read_mostly; | |
87 | #else | |
88 | #define do_swap_account 0 | |
89 | #endif | |
90 | ||
91 | #ifdef CONFIG_CGROUP_WRITEBACK | |
92 | static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq); | |
93 | #endif | |
94 | ||
95 | /* Whether legacy memory+swap accounting is active */ | |
96 | static bool do_memsw_account(void) | |
97 | { | |
98 | return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account; | |
99 | } | |
100 | ||
101 | #define THRESHOLDS_EVENTS_TARGET 128 | |
102 | #define SOFTLIMIT_EVENTS_TARGET 1024 | |
103 | ||
104 | /* | |
105 | * Cgroups above their limits are maintained in a RB-Tree, independent of | |
106 | * their hierarchy representation | |
107 | */ | |
108 | ||
109 | struct mem_cgroup_tree_per_node { | |
110 | struct rb_root rb_root; | |
111 | struct rb_node *rb_rightmost; | |
112 | spinlock_t lock; | |
113 | }; | |
114 | ||
115 | struct mem_cgroup_tree { | |
116 | struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; | |
117 | }; | |
118 | ||
119 | static struct mem_cgroup_tree soft_limit_tree __read_mostly; | |
120 | ||
121 | /* for OOM */ | |
122 | struct mem_cgroup_eventfd_list { | |
123 | struct list_head list; | |
124 | struct eventfd_ctx *eventfd; | |
125 | }; | |
126 | ||
127 | /* | |
128 | * cgroup_event represents events which userspace want to receive. | |
129 | */ | |
130 | struct mem_cgroup_event { | |
131 | /* | |
132 | * memcg which the event belongs to. | |
133 | */ | |
134 | struct mem_cgroup *memcg; | |
135 | /* | |
136 | * eventfd to signal userspace about the event. | |
137 | */ | |
138 | struct eventfd_ctx *eventfd; | |
139 | /* | |
140 | * Each of these stored in a list by the cgroup. | |
141 | */ | |
142 | struct list_head list; | |
143 | /* | |
144 | * register_event() callback will be used to add new userspace | |
145 | * waiter for changes related to this event. Use eventfd_signal() | |
146 | * on eventfd to send notification to userspace. | |
147 | */ | |
148 | int (*register_event)(struct mem_cgroup *memcg, | |
149 | struct eventfd_ctx *eventfd, const char *args); | |
150 | /* | |
151 | * unregister_event() callback will be called when userspace closes | |
152 | * the eventfd or on cgroup removing. This callback must be set, | |
153 | * if you want provide notification functionality. | |
154 | */ | |
155 | void (*unregister_event)(struct mem_cgroup *memcg, | |
156 | struct eventfd_ctx *eventfd); | |
157 | /* | |
158 | * All fields below needed to unregister event when | |
159 | * userspace closes eventfd. | |
160 | */ | |
161 | poll_table pt; | |
162 | wait_queue_head_t *wqh; | |
163 | wait_queue_entry_t wait; | |
164 | struct work_struct remove; | |
165 | }; | |
166 | ||
167 | static void mem_cgroup_threshold(struct mem_cgroup *memcg); | |
168 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); | |
169 | ||
170 | /* Stuffs for move charges at task migration. */ | |
171 | /* | |
172 | * Types of charges to be moved. | |
173 | */ | |
174 | #define MOVE_ANON 0x1U | |
175 | #define MOVE_FILE 0x2U | |
176 | #define MOVE_MASK (MOVE_ANON | MOVE_FILE) | |
177 | ||
178 | /* "mc" and its members are protected by cgroup_mutex */ | |
179 | static struct move_charge_struct { | |
180 | spinlock_t lock; /* for from, to */ | |
181 | struct mm_struct *mm; | |
182 | struct mem_cgroup *from; | |
183 | struct mem_cgroup *to; | |
184 | unsigned long flags; | |
185 | unsigned long precharge; | |
186 | unsigned long moved_charge; | |
187 | unsigned long moved_swap; | |
188 | struct task_struct *moving_task; /* a task moving charges */ | |
189 | wait_queue_head_t waitq; /* a waitq for other context */ | |
190 | } mc = { | |
191 | .lock = __SPIN_LOCK_UNLOCKED(mc.lock), | |
192 | .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), | |
193 | }; | |
194 | ||
195 | /* | |
196 | * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft | |
197 | * limit reclaim to prevent infinite loops, if they ever occur. | |
198 | */ | |
199 | #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100 | |
200 | #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2 | |
201 | ||
202 | enum charge_type { | |
203 | MEM_CGROUP_CHARGE_TYPE_CACHE = 0, | |
204 | MEM_CGROUP_CHARGE_TYPE_ANON, | |
205 | MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ | |
206 | MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ | |
207 | NR_CHARGE_TYPE, | |
208 | }; | |
209 | ||
210 | /* for encoding cft->private value on file */ | |
211 | enum res_type { | |
212 | _MEM, | |
213 | _MEMSWAP, | |
214 | _OOM_TYPE, | |
215 | _KMEM, | |
216 | _TCP, | |
217 | }; | |
218 | ||
219 | #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val)) | |
220 | #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff) | |
221 | #define MEMFILE_ATTR(val) ((val) & 0xffff) | |
222 | /* Used for OOM nofiier */ | |
223 | #define OOM_CONTROL (0) | |
224 | ||
225 | /* | |
226 | * Iteration constructs for visiting all cgroups (under a tree). If | |
227 | * loops are exited prematurely (break), mem_cgroup_iter_break() must | |
228 | * be used for reference counting. | |
229 | */ | |
230 | #define for_each_mem_cgroup_tree(iter, root) \ | |
231 | for (iter = mem_cgroup_iter(root, NULL, NULL); \ | |
232 | iter != NULL; \ | |
233 | iter = mem_cgroup_iter(root, iter, NULL)) | |
234 | ||
235 | #define for_each_mem_cgroup(iter) \ | |
236 | for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ | |
237 | iter != NULL; \ | |
238 | iter = mem_cgroup_iter(NULL, iter, NULL)) | |
239 | ||
240 | static inline bool should_force_charge(void) | |
241 | { | |
242 | return tsk_is_oom_victim(current) || fatal_signal_pending(current) || | |
243 | (current->flags & PF_EXITING); | |
244 | } | |
245 | ||
246 | /* Some nice accessors for the vmpressure. */ | |
247 | struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg) | |
248 | { | |
249 | if (!memcg) | |
250 | memcg = root_mem_cgroup; | |
251 | return &memcg->vmpressure; | |
252 | } | |
253 | ||
254 | struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr) | |
255 | { | |
256 | return &container_of(vmpr, struct mem_cgroup, vmpressure)->css; | |
257 | } | |
258 | ||
259 | #ifdef CONFIG_MEMCG_KMEM | |
260 | /* | |
261 | * This will be the memcg's index in each cache's ->memcg_params.memcg_caches. | |
262 | * The main reason for not using cgroup id for this: | |
263 | * this works better in sparse environments, where we have a lot of memcgs, | |
264 | * but only a few kmem-limited. Or also, if we have, for instance, 200 | |
265 | * memcgs, and none but the 200th is kmem-limited, we'd have to have a | |
266 | * 200 entry array for that. | |
267 | * | |
268 | * The current size of the caches array is stored in memcg_nr_cache_ids. It | |
269 | * will double each time we have to increase it. | |
270 | */ | |
271 | static DEFINE_IDA(memcg_cache_ida); | |
272 | int memcg_nr_cache_ids; | |
273 | ||
274 | /* Protects memcg_nr_cache_ids */ | |
275 | static DECLARE_RWSEM(memcg_cache_ids_sem); | |
276 | ||
277 | void memcg_get_cache_ids(void) | |
278 | { | |
279 | down_read(&memcg_cache_ids_sem); | |
280 | } | |
281 | ||
282 | void memcg_put_cache_ids(void) | |
283 | { | |
284 | up_read(&memcg_cache_ids_sem); | |
285 | } | |
286 | ||
287 | /* | |
288 | * MIN_SIZE is different than 1, because we would like to avoid going through | |
289 | * the alloc/free process all the time. In a small machine, 4 kmem-limited | |
290 | * cgroups is a reasonable guess. In the future, it could be a parameter or | |
291 | * tunable, but that is strictly not necessary. | |
292 | * | |
293 | * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get | |
294 | * this constant directly from cgroup, but it is understandable that this is | |
295 | * better kept as an internal representation in cgroup.c. In any case, the | |
296 | * cgrp_id space is not getting any smaller, and we don't have to necessarily | |
297 | * increase ours as well if it increases. | |
298 | */ | |
299 | #define MEMCG_CACHES_MIN_SIZE 4 | |
300 | #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX | |
301 | ||
302 | /* | |
303 | * A lot of the calls to the cache allocation functions are expected to be | |
304 | * inlined by the compiler. Since the calls to memcg_kmem_get_cache are | |
305 | * conditional to this static branch, we'll have to allow modules that does | |
306 | * kmem_cache_alloc and the such to see this symbol as well | |
307 | */ | |
308 | DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key); | |
309 | EXPORT_SYMBOL(memcg_kmem_enabled_key); | |
310 | ||
311 | struct workqueue_struct *memcg_kmem_cache_wq; | |
312 | #endif | |
313 | ||
314 | static int memcg_shrinker_map_size; | |
315 | static DEFINE_MUTEX(memcg_shrinker_map_mutex); | |
316 | ||
317 | static void memcg_free_shrinker_map_rcu(struct rcu_head *head) | |
318 | { | |
319 | kvfree(container_of(head, struct memcg_shrinker_map, rcu)); | |
320 | } | |
321 | ||
322 | static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg, | |
323 | int size, int old_size) | |
324 | { | |
325 | struct memcg_shrinker_map *new, *old; | |
326 | int nid; | |
327 | ||
328 | lockdep_assert_held(&memcg_shrinker_map_mutex); | |
329 | ||
330 | for_each_node(nid) { | |
331 | old = rcu_dereference_protected( | |
332 | mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true); | |
333 | /* Not yet online memcg */ | |
334 | if (!old) | |
335 | return 0; | |
336 | ||
337 | new = kvmalloc(sizeof(*new) + size, GFP_KERNEL); | |
338 | if (!new) | |
339 | return -ENOMEM; | |
340 | ||
341 | /* Set all old bits, clear all new bits */ | |
342 | memset(new->map, (int)0xff, old_size); | |
343 | memset((void *)new->map + old_size, 0, size - old_size); | |
344 | ||
345 | rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new); | |
346 | call_rcu(&old->rcu, memcg_free_shrinker_map_rcu); | |
347 | } | |
348 | ||
349 | return 0; | |
350 | } | |
351 | ||
352 | static void memcg_free_shrinker_maps(struct mem_cgroup *memcg) | |
353 | { | |
354 | struct mem_cgroup_per_node *pn; | |
355 | struct memcg_shrinker_map *map; | |
356 | int nid; | |
357 | ||
358 | if (mem_cgroup_is_root(memcg)) | |
359 | return; | |
360 | ||
361 | for_each_node(nid) { | |
362 | pn = mem_cgroup_nodeinfo(memcg, nid); | |
363 | map = rcu_dereference_protected(pn->shrinker_map, true); | |
364 | if (map) | |
365 | kvfree(map); | |
366 | rcu_assign_pointer(pn->shrinker_map, NULL); | |
367 | } | |
368 | } | |
369 | ||
370 | static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg) | |
371 | { | |
372 | struct memcg_shrinker_map *map; | |
373 | int nid, size, ret = 0; | |
374 | ||
375 | if (mem_cgroup_is_root(memcg)) | |
376 | return 0; | |
377 | ||
378 | mutex_lock(&memcg_shrinker_map_mutex); | |
379 | size = memcg_shrinker_map_size; | |
380 | for_each_node(nid) { | |
381 | map = kvzalloc(sizeof(*map) + size, GFP_KERNEL); | |
382 | if (!map) { | |
383 | memcg_free_shrinker_maps(memcg); | |
384 | ret = -ENOMEM; | |
385 | break; | |
386 | } | |
387 | rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map); | |
388 | } | |
389 | mutex_unlock(&memcg_shrinker_map_mutex); | |
390 | ||
391 | return ret; | |
392 | } | |
393 | ||
394 | int memcg_expand_shrinker_maps(int new_id) | |
395 | { | |
396 | int size, old_size, ret = 0; | |
397 | struct mem_cgroup *memcg; | |
398 | ||
399 | size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long); | |
400 | old_size = memcg_shrinker_map_size; | |
401 | if (size <= old_size) | |
402 | return 0; | |
403 | ||
404 | mutex_lock(&memcg_shrinker_map_mutex); | |
405 | if (!root_mem_cgroup) | |
406 | goto unlock; | |
407 | ||
408 | for_each_mem_cgroup(memcg) { | |
409 | if (mem_cgroup_is_root(memcg)) | |
410 | continue; | |
411 | ret = memcg_expand_one_shrinker_map(memcg, size, old_size); | |
412 | if (ret) { | |
413 | mem_cgroup_iter_break(NULL, memcg); | |
414 | goto unlock; | |
415 | } | |
416 | } | |
417 | unlock: | |
418 | if (!ret) | |
419 | memcg_shrinker_map_size = size; | |
420 | mutex_unlock(&memcg_shrinker_map_mutex); | |
421 | return ret; | |
422 | } | |
423 | ||
424 | void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id) | |
425 | { | |
426 | if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) { | |
427 | struct memcg_shrinker_map *map; | |
428 | ||
429 | rcu_read_lock(); | |
430 | map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map); | |
431 | /* Pairs with smp mb in shrink_slab() */ | |
432 | smp_mb__before_atomic(); | |
433 | set_bit(shrinker_id, map->map); | |
434 | rcu_read_unlock(); | |
435 | } | |
436 | } | |
437 | ||
438 | /** | |
439 | * mem_cgroup_css_from_page - css of the memcg associated with a page | |
440 | * @page: page of interest | |
441 | * | |
442 | * If memcg is bound to the default hierarchy, css of the memcg associated | |
443 | * with @page is returned. The returned css remains associated with @page | |
444 | * until it is released. | |
445 | * | |
446 | * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup | |
447 | * is returned. | |
448 | */ | |
449 | struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page) | |
450 | { | |
451 | struct mem_cgroup *memcg; | |
452 | ||
453 | memcg = page->mem_cgroup; | |
454 | ||
455 | if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
456 | memcg = root_mem_cgroup; | |
457 | ||
458 | return &memcg->css; | |
459 | } | |
460 | ||
461 | /** | |
462 | * page_cgroup_ino - return inode number of the memcg a page is charged to | |
463 | * @page: the page | |
464 | * | |
465 | * Look up the closest online ancestor of the memory cgroup @page is charged to | |
466 | * and return its inode number or 0 if @page is not charged to any cgroup. It | |
467 | * is safe to call this function without holding a reference to @page. | |
468 | * | |
469 | * Note, this function is inherently racy, because there is nothing to prevent | |
470 | * the cgroup inode from getting torn down and potentially reallocated a moment | |
471 | * after page_cgroup_ino() returns, so it only should be used by callers that | |
472 | * do not care (such as procfs interfaces). | |
473 | */ | |
474 | ino_t page_cgroup_ino(struct page *page) | |
475 | { | |
476 | struct mem_cgroup *memcg; | |
477 | unsigned long ino = 0; | |
478 | ||
479 | rcu_read_lock(); | |
480 | if (PageSlab(page) && !PageTail(page)) | |
481 | memcg = memcg_from_slab_page(page); | |
482 | else | |
483 | memcg = READ_ONCE(page->mem_cgroup); | |
484 | while (memcg && !(memcg->css.flags & CSS_ONLINE)) | |
485 | memcg = parent_mem_cgroup(memcg); | |
486 | if (memcg) | |
487 | ino = cgroup_ino(memcg->css.cgroup); | |
488 | rcu_read_unlock(); | |
489 | return ino; | |
490 | } | |
491 | ||
492 | static struct mem_cgroup_per_node * | |
493 | mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page) | |
494 | { | |
495 | int nid = page_to_nid(page); | |
496 | ||
497 | return memcg->nodeinfo[nid]; | |
498 | } | |
499 | ||
500 | static struct mem_cgroup_tree_per_node * | |
501 | soft_limit_tree_node(int nid) | |
502 | { | |
503 | return soft_limit_tree.rb_tree_per_node[nid]; | |
504 | } | |
505 | ||
506 | static struct mem_cgroup_tree_per_node * | |
507 | soft_limit_tree_from_page(struct page *page) | |
508 | { | |
509 | int nid = page_to_nid(page); | |
510 | ||
511 | return soft_limit_tree.rb_tree_per_node[nid]; | |
512 | } | |
513 | ||
514 | static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz, | |
515 | struct mem_cgroup_tree_per_node *mctz, | |
516 | unsigned long new_usage_in_excess) | |
517 | { | |
518 | struct rb_node **p = &mctz->rb_root.rb_node; | |
519 | struct rb_node *parent = NULL; | |
520 | struct mem_cgroup_per_node *mz_node; | |
521 | bool rightmost = true; | |
522 | ||
523 | if (mz->on_tree) | |
524 | return; | |
525 | ||
526 | mz->usage_in_excess = new_usage_in_excess; | |
527 | if (!mz->usage_in_excess) | |
528 | return; | |
529 | while (*p) { | |
530 | parent = *p; | |
531 | mz_node = rb_entry(parent, struct mem_cgroup_per_node, | |
532 | tree_node); | |
533 | if (mz->usage_in_excess < mz_node->usage_in_excess) { | |
534 | p = &(*p)->rb_left; | |
535 | rightmost = false; | |
536 | } | |
537 | ||
538 | /* | |
539 | * We can't avoid mem cgroups that are over their soft | |
540 | * limit by the same amount | |
541 | */ | |
542 | else if (mz->usage_in_excess >= mz_node->usage_in_excess) | |
543 | p = &(*p)->rb_right; | |
544 | } | |
545 | ||
546 | if (rightmost) | |
547 | mctz->rb_rightmost = &mz->tree_node; | |
548 | ||
549 | rb_link_node(&mz->tree_node, parent, p); | |
550 | rb_insert_color(&mz->tree_node, &mctz->rb_root); | |
551 | mz->on_tree = true; | |
552 | } | |
553 | ||
554 | static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz, | |
555 | struct mem_cgroup_tree_per_node *mctz) | |
556 | { | |
557 | if (!mz->on_tree) | |
558 | return; | |
559 | ||
560 | if (&mz->tree_node == mctz->rb_rightmost) | |
561 | mctz->rb_rightmost = rb_prev(&mz->tree_node); | |
562 | ||
563 | rb_erase(&mz->tree_node, &mctz->rb_root); | |
564 | mz->on_tree = false; | |
565 | } | |
566 | ||
567 | static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz, | |
568 | struct mem_cgroup_tree_per_node *mctz) | |
569 | { | |
570 | unsigned long flags; | |
571 | ||
572 | spin_lock_irqsave(&mctz->lock, flags); | |
573 | __mem_cgroup_remove_exceeded(mz, mctz); | |
574 | spin_unlock_irqrestore(&mctz->lock, flags); | |
575 | } | |
576 | ||
577 | static unsigned long soft_limit_excess(struct mem_cgroup *memcg) | |
578 | { | |
579 | unsigned long nr_pages = page_counter_read(&memcg->memory); | |
580 | unsigned long soft_limit = READ_ONCE(memcg->soft_limit); | |
581 | unsigned long excess = 0; | |
582 | ||
583 | if (nr_pages > soft_limit) | |
584 | excess = nr_pages - soft_limit; | |
585 | ||
586 | return excess; | |
587 | } | |
588 | ||
589 | static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) | |
590 | { | |
591 | unsigned long excess; | |
592 | struct mem_cgroup_per_node *mz; | |
593 | struct mem_cgroup_tree_per_node *mctz; | |
594 | ||
595 | mctz = soft_limit_tree_from_page(page); | |
596 | if (!mctz) | |
597 | return; | |
598 | /* | |
599 | * Necessary to update all ancestors when hierarchy is used. | |
600 | * because their event counter is not touched. | |
601 | */ | |
602 | for (; memcg; memcg = parent_mem_cgroup(memcg)) { | |
603 | mz = mem_cgroup_page_nodeinfo(memcg, page); | |
604 | excess = soft_limit_excess(memcg); | |
605 | /* | |
606 | * We have to update the tree if mz is on RB-tree or | |
607 | * mem is over its softlimit. | |
608 | */ | |
609 | if (excess || mz->on_tree) { | |
610 | unsigned long flags; | |
611 | ||
612 | spin_lock_irqsave(&mctz->lock, flags); | |
613 | /* if on-tree, remove it */ | |
614 | if (mz->on_tree) | |
615 | __mem_cgroup_remove_exceeded(mz, mctz); | |
616 | /* | |
617 | * Insert again. mz->usage_in_excess will be updated. | |
618 | * If excess is 0, no tree ops. | |
619 | */ | |
620 | __mem_cgroup_insert_exceeded(mz, mctz, excess); | |
621 | spin_unlock_irqrestore(&mctz->lock, flags); | |
622 | } | |
623 | } | |
624 | } | |
625 | ||
626 | static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) | |
627 | { | |
628 | struct mem_cgroup_tree_per_node *mctz; | |
629 | struct mem_cgroup_per_node *mz; | |
630 | int nid; | |
631 | ||
632 | for_each_node(nid) { | |
633 | mz = mem_cgroup_nodeinfo(memcg, nid); | |
634 | mctz = soft_limit_tree_node(nid); | |
635 | if (mctz) | |
636 | mem_cgroup_remove_exceeded(mz, mctz); | |
637 | } | |
638 | } | |
639 | ||
640 | static struct mem_cgroup_per_node * | |
641 | __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz) | |
642 | { | |
643 | struct mem_cgroup_per_node *mz; | |
644 | ||
645 | retry: | |
646 | mz = NULL; | |
647 | if (!mctz->rb_rightmost) | |
648 | goto done; /* Nothing to reclaim from */ | |
649 | ||
650 | mz = rb_entry(mctz->rb_rightmost, | |
651 | struct mem_cgroup_per_node, tree_node); | |
652 | /* | |
653 | * Remove the node now but someone else can add it back, | |
654 | * we will to add it back at the end of reclaim to its correct | |
655 | * position in the tree. | |
656 | */ | |
657 | __mem_cgroup_remove_exceeded(mz, mctz); | |
658 | if (!soft_limit_excess(mz->memcg) || | |
659 | !css_tryget_online(&mz->memcg->css)) | |
660 | goto retry; | |
661 | done: | |
662 | return mz; | |
663 | } | |
664 | ||
665 | static struct mem_cgroup_per_node * | |
666 | mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz) | |
667 | { | |
668 | struct mem_cgroup_per_node *mz; | |
669 | ||
670 | spin_lock_irq(&mctz->lock); | |
671 | mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
672 | spin_unlock_irq(&mctz->lock); | |
673 | return mz; | |
674 | } | |
675 | ||
676 | /** | |
677 | * __mod_memcg_state - update cgroup memory statistics | |
678 | * @memcg: the memory cgroup | |
679 | * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item | |
680 | * @val: delta to add to the counter, can be negative | |
681 | */ | |
682 | void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val) | |
683 | { | |
684 | long x; | |
685 | ||
686 | if (mem_cgroup_disabled()) | |
687 | return; | |
688 | ||
689 | x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]); | |
690 | if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) { | |
691 | struct mem_cgroup *mi; | |
692 | ||
693 | /* | |
694 | * Batch local counters to keep them in sync with | |
695 | * the hierarchical ones. | |
696 | */ | |
697 | __this_cpu_add(memcg->vmstats_local->stat[idx], x); | |
698 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
699 | atomic_long_add(x, &mi->vmstats[idx]); | |
700 | x = 0; | |
701 | } | |
702 | __this_cpu_write(memcg->vmstats_percpu->stat[idx], x); | |
703 | } | |
704 | ||
705 | static struct mem_cgroup_per_node * | |
706 | parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid) | |
707 | { | |
708 | struct mem_cgroup *parent; | |
709 | ||
710 | parent = parent_mem_cgroup(pn->memcg); | |
711 | if (!parent) | |
712 | return NULL; | |
713 | return mem_cgroup_nodeinfo(parent, nid); | |
714 | } | |
715 | ||
716 | /** | |
717 | * __mod_lruvec_state - update lruvec memory statistics | |
718 | * @lruvec: the lruvec | |
719 | * @idx: the stat item | |
720 | * @val: delta to add to the counter, can be negative | |
721 | * | |
722 | * The lruvec is the intersection of the NUMA node and a cgroup. This | |
723 | * function updates the all three counters that are affected by a | |
724 | * change of state at this level: per-node, per-cgroup, per-lruvec. | |
725 | */ | |
726 | void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx, | |
727 | int val) | |
728 | { | |
729 | pg_data_t *pgdat = lruvec_pgdat(lruvec); | |
730 | struct mem_cgroup_per_node *pn; | |
731 | struct mem_cgroup *memcg; | |
732 | long x; | |
733 | ||
734 | /* Update node */ | |
735 | __mod_node_page_state(pgdat, idx, val); | |
736 | ||
737 | if (mem_cgroup_disabled()) | |
738 | return; | |
739 | ||
740 | pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec); | |
741 | memcg = pn->memcg; | |
742 | ||
743 | /* Update memcg */ | |
744 | __mod_memcg_state(memcg, idx, val); | |
745 | ||
746 | /* Update lruvec */ | |
747 | __this_cpu_add(pn->lruvec_stat_local->count[idx], val); | |
748 | ||
749 | x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]); | |
750 | if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) { | |
751 | struct mem_cgroup_per_node *pi; | |
752 | ||
753 | for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id)) | |
754 | atomic_long_add(x, &pi->lruvec_stat[idx]); | |
755 | x = 0; | |
756 | } | |
757 | __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x); | |
758 | } | |
759 | ||
760 | void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val) | |
761 | { | |
762 | struct page *page = virt_to_head_page(p); | |
763 | pg_data_t *pgdat = page_pgdat(page); | |
764 | struct mem_cgroup *memcg; | |
765 | struct lruvec *lruvec; | |
766 | ||
767 | rcu_read_lock(); | |
768 | memcg = memcg_from_slab_page(page); | |
769 | ||
770 | /* Untracked pages have no memcg, no lruvec. Update only the node */ | |
771 | if (!memcg || memcg == root_mem_cgroup) { | |
772 | __mod_node_page_state(pgdat, idx, val); | |
773 | } else { | |
774 | lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
775 | __mod_lruvec_state(lruvec, idx, val); | |
776 | } | |
777 | rcu_read_unlock(); | |
778 | } | |
779 | ||
780 | /** | |
781 | * __count_memcg_events - account VM events in a cgroup | |
782 | * @memcg: the memory cgroup | |
783 | * @idx: the event item | |
784 | * @count: the number of events that occured | |
785 | */ | |
786 | void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx, | |
787 | unsigned long count) | |
788 | { | |
789 | unsigned long x; | |
790 | ||
791 | if (mem_cgroup_disabled()) | |
792 | return; | |
793 | ||
794 | x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]); | |
795 | if (unlikely(x > MEMCG_CHARGE_BATCH)) { | |
796 | struct mem_cgroup *mi; | |
797 | ||
798 | /* | |
799 | * Batch local counters to keep them in sync with | |
800 | * the hierarchical ones. | |
801 | */ | |
802 | __this_cpu_add(memcg->vmstats_local->events[idx], x); | |
803 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
804 | atomic_long_add(x, &mi->vmevents[idx]); | |
805 | x = 0; | |
806 | } | |
807 | __this_cpu_write(memcg->vmstats_percpu->events[idx], x); | |
808 | } | |
809 | ||
810 | static unsigned long memcg_events(struct mem_cgroup *memcg, int event) | |
811 | { | |
812 | return atomic_long_read(&memcg->vmevents[event]); | |
813 | } | |
814 | ||
815 | static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event) | |
816 | { | |
817 | long x = 0; | |
818 | int cpu; | |
819 | ||
820 | for_each_possible_cpu(cpu) | |
821 | x += per_cpu(memcg->vmstats_local->events[event], cpu); | |
822 | return x; | |
823 | } | |
824 | ||
825 | static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, | |
826 | struct page *page, | |
827 | bool compound, int nr_pages) | |
828 | { | |
829 | /* | |
830 | * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is | |
831 | * counted as CACHE even if it's on ANON LRU. | |
832 | */ | |
833 | if (PageAnon(page)) | |
834 | __mod_memcg_state(memcg, MEMCG_RSS, nr_pages); | |
835 | else { | |
836 | __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages); | |
837 | if (PageSwapBacked(page)) | |
838 | __mod_memcg_state(memcg, NR_SHMEM, nr_pages); | |
839 | } | |
840 | ||
841 | if (compound) { | |
842 | VM_BUG_ON_PAGE(!PageTransHuge(page), page); | |
843 | __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages); | |
844 | } | |
845 | ||
846 | /* pagein of a big page is an event. So, ignore page size */ | |
847 | if (nr_pages > 0) | |
848 | __count_memcg_events(memcg, PGPGIN, 1); | |
849 | else { | |
850 | __count_memcg_events(memcg, PGPGOUT, 1); | |
851 | nr_pages = -nr_pages; /* for event */ | |
852 | } | |
853 | ||
854 | __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages); | |
855 | } | |
856 | ||
857 | static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, | |
858 | enum mem_cgroup_events_target target) | |
859 | { | |
860 | unsigned long val, next; | |
861 | ||
862 | val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events); | |
863 | next = __this_cpu_read(memcg->vmstats_percpu->targets[target]); | |
864 | /* from time_after() in jiffies.h */ | |
865 | if ((long)(next - val) < 0) { | |
866 | switch (target) { | |
867 | case MEM_CGROUP_TARGET_THRESH: | |
868 | next = val + THRESHOLDS_EVENTS_TARGET; | |
869 | break; | |
870 | case MEM_CGROUP_TARGET_SOFTLIMIT: | |
871 | next = val + SOFTLIMIT_EVENTS_TARGET; | |
872 | break; | |
873 | default: | |
874 | break; | |
875 | } | |
876 | __this_cpu_write(memcg->vmstats_percpu->targets[target], next); | |
877 | return true; | |
878 | } | |
879 | return false; | |
880 | } | |
881 | ||
882 | /* | |
883 | * Check events in order. | |
884 | * | |
885 | */ | |
886 | static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) | |
887 | { | |
888 | /* threshold event is triggered in finer grain than soft limit */ | |
889 | if (unlikely(mem_cgroup_event_ratelimit(memcg, | |
890 | MEM_CGROUP_TARGET_THRESH))) { | |
891 | bool do_softlimit; | |
892 | ||
893 | do_softlimit = mem_cgroup_event_ratelimit(memcg, | |
894 | MEM_CGROUP_TARGET_SOFTLIMIT); | |
895 | mem_cgroup_threshold(memcg); | |
896 | if (unlikely(do_softlimit)) | |
897 | mem_cgroup_update_tree(memcg, page); | |
898 | } | |
899 | } | |
900 | ||
901 | struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) | |
902 | { | |
903 | /* | |
904 | * mm_update_next_owner() may clear mm->owner to NULL | |
905 | * if it races with swapoff, page migration, etc. | |
906 | * So this can be called with p == NULL. | |
907 | */ | |
908 | if (unlikely(!p)) | |
909 | return NULL; | |
910 | ||
911 | return mem_cgroup_from_css(task_css(p, memory_cgrp_id)); | |
912 | } | |
913 | EXPORT_SYMBOL(mem_cgroup_from_task); | |
914 | ||
915 | /** | |
916 | * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg. | |
917 | * @mm: mm from which memcg should be extracted. It can be NULL. | |
918 | * | |
919 | * Obtain a reference on mm->memcg and returns it if successful. Otherwise | |
920 | * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is | |
921 | * returned. | |
922 | */ | |
923 | struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm) | |
924 | { | |
925 | struct mem_cgroup *memcg; | |
926 | ||
927 | if (mem_cgroup_disabled()) | |
928 | return NULL; | |
929 | ||
930 | rcu_read_lock(); | |
931 | do { | |
932 | /* | |
933 | * Page cache insertions can happen withou an | |
934 | * actual mm context, e.g. during disk probing | |
935 | * on boot, loopback IO, acct() writes etc. | |
936 | */ | |
937 | if (unlikely(!mm)) | |
938 | memcg = root_mem_cgroup; | |
939 | else { | |
940 | memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); | |
941 | if (unlikely(!memcg)) | |
942 | memcg = root_mem_cgroup; | |
943 | } | |
944 | } while (!css_tryget(&memcg->css)); | |
945 | rcu_read_unlock(); | |
946 | return memcg; | |
947 | } | |
948 | EXPORT_SYMBOL(get_mem_cgroup_from_mm); | |
949 | ||
950 | /** | |
951 | * get_mem_cgroup_from_page: Obtain a reference on given page's memcg. | |
952 | * @page: page from which memcg should be extracted. | |
953 | * | |
954 | * Obtain a reference on page->memcg and returns it if successful. Otherwise | |
955 | * root_mem_cgroup is returned. | |
956 | */ | |
957 | struct mem_cgroup *get_mem_cgroup_from_page(struct page *page) | |
958 | { | |
959 | struct mem_cgroup *memcg = page->mem_cgroup; | |
960 | ||
961 | if (mem_cgroup_disabled()) | |
962 | return NULL; | |
963 | ||
964 | rcu_read_lock(); | |
965 | if (!memcg || !css_tryget_online(&memcg->css)) | |
966 | memcg = root_mem_cgroup; | |
967 | rcu_read_unlock(); | |
968 | return memcg; | |
969 | } | |
970 | EXPORT_SYMBOL(get_mem_cgroup_from_page); | |
971 | ||
972 | /** | |
973 | * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg. | |
974 | */ | |
975 | static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void) | |
976 | { | |
977 | if (unlikely(current->active_memcg)) { | |
978 | struct mem_cgroup *memcg = root_mem_cgroup; | |
979 | ||
980 | rcu_read_lock(); | |
981 | if (css_tryget_online(¤t->active_memcg->css)) | |
982 | memcg = current->active_memcg; | |
983 | rcu_read_unlock(); | |
984 | return memcg; | |
985 | } | |
986 | return get_mem_cgroup_from_mm(current->mm); | |
987 | } | |
988 | ||
989 | /** | |
990 | * mem_cgroup_iter - iterate over memory cgroup hierarchy | |
991 | * @root: hierarchy root | |
992 | * @prev: previously returned memcg, NULL on first invocation | |
993 | * @reclaim: cookie for shared reclaim walks, NULL for full walks | |
994 | * | |
995 | * Returns references to children of the hierarchy below @root, or | |
996 | * @root itself, or %NULL after a full round-trip. | |
997 | * | |
998 | * Caller must pass the return value in @prev on subsequent | |
999 | * invocations for reference counting, or use mem_cgroup_iter_break() | |
1000 | * to cancel a hierarchy walk before the round-trip is complete. | |
1001 | * | |
1002 | * Reclaimers can specify a node and a priority level in @reclaim to | |
1003 | * divide up the memcgs in the hierarchy among all concurrent | |
1004 | * reclaimers operating on the same node and priority. | |
1005 | */ | |
1006 | struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, | |
1007 | struct mem_cgroup *prev, | |
1008 | struct mem_cgroup_reclaim_cookie *reclaim) | |
1009 | { | |
1010 | struct mem_cgroup_reclaim_iter *uninitialized_var(iter); | |
1011 | struct cgroup_subsys_state *css = NULL; | |
1012 | struct mem_cgroup *memcg = NULL; | |
1013 | struct mem_cgroup *pos = NULL; | |
1014 | ||
1015 | if (mem_cgroup_disabled()) | |
1016 | return NULL; | |
1017 | ||
1018 | if (!root) | |
1019 | root = root_mem_cgroup; | |
1020 | ||
1021 | if (prev && !reclaim) | |
1022 | pos = prev; | |
1023 | ||
1024 | if (!root->use_hierarchy && root != root_mem_cgroup) { | |
1025 | if (prev) | |
1026 | goto out; | |
1027 | return root; | |
1028 | } | |
1029 | ||
1030 | rcu_read_lock(); | |
1031 | ||
1032 | if (reclaim) { | |
1033 | struct mem_cgroup_per_node *mz; | |
1034 | ||
1035 | mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id); | |
1036 | iter = &mz->iter; | |
1037 | ||
1038 | if (prev && reclaim->generation != iter->generation) | |
1039 | goto out_unlock; | |
1040 | ||
1041 | while (1) { | |
1042 | pos = READ_ONCE(iter->position); | |
1043 | if (!pos || css_tryget(&pos->css)) | |
1044 | break; | |
1045 | /* | |
1046 | * css reference reached zero, so iter->position will | |
1047 | * be cleared by ->css_released. However, we should not | |
1048 | * rely on this happening soon, because ->css_released | |
1049 | * is called from a work queue, and by busy-waiting we | |
1050 | * might block it. So we clear iter->position right | |
1051 | * away. | |
1052 | */ | |
1053 | (void)cmpxchg(&iter->position, pos, NULL); | |
1054 | } | |
1055 | } | |
1056 | ||
1057 | if (pos) | |
1058 | css = &pos->css; | |
1059 | ||
1060 | for (;;) { | |
1061 | css = css_next_descendant_pre(css, &root->css); | |
1062 | if (!css) { | |
1063 | /* | |
1064 | * Reclaimers share the hierarchy walk, and a | |
1065 | * new one might jump in right at the end of | |
1066 | * the hierarchy - make sure they see at least | |
1067 | * one group and restart from the beginning. | |
1068 | */ | |
1069 | if (!prev) | |
1070 | continue; | |
1071 | break; | |
1072 | } | |
1073 | ||
1074 | /* | |
1075 | * Verify the css and acquire a reference. The root | |
1076 | * is provided by the caller, so we know it's alive | |
1077 | * and kicking, and don't take an extra reference. | |
1078 | */ | |
1079 | memcg = mem_cgroup_from_css(css); | |
1080 | ||
1081 | if (css == &root->css) | |
1082 | break; | |
1083 | ||
1084 | if (css_tryget(css)) | |
1085 | break; | |
1086 | ||
1087 | memcg = NULL; | |
1088 | } | |
1089 | ||
1090 | if (reclaim) { | |
1091 | /* | |
1092 | * The position could have already been updated by a competing | |
1093 | * thread, so check that the value hasn't changed since we read | |
1094 | * it to avoid reclaiming from the same cgroup twice. | |
1095 | */ | |
1096 | (void)cmpxchg(&iter->position, pos, memcg); | |
1097 | ||
1098 | if (pos) | |
1099 | css_put(&pos->css); | |
1100 | ||
1101 | if (!memcg) | |
1102 | iter->generation++; | |
1103 | else if (!prev) | |
1104 | reclaim->generation = iter->generation; | |
1105 | } | |
1106 | ||
1107 | out_unlock: | |
1108 | rcu_read_unlock(); | |
1109 | out: | |
1110 | if (prev && prev != root) | |
1111 | css_put(&prev->css); | |
1112 | ||
1113 | return memcg; | |
1114 | } | |
1115 | ||
1116 | /** | |
1117 | * mem_cgroup_iter_break - abort a hierarchy walk prematurely | |
1118 | * @root: hierarchy root | |
1119 | * @prev: last visited hierarchy member as returned by mem_cgroup_iter() | |
1120 | */ | |
1121 | void mem_cgroup_iter_break(struct mem_cgroup *root, | |
1122 | struct mem_cgroup *prev) | |
1123 | { | |
1124 | if (!root) | |
1125 | root = root_mem_cgroup; | |
1126 | if (prev && prev != root) | |
1127 | css_put(&prev->css); | |
1128 | } | |
1129 | ||
1130 | static void __invalidate_reclaim_iterators(struct mem_cgroup *from, | |
1131 | struct mem_cgroup *dead_memcg) | |
1132 | { | |
1133 | struct mem_cgroup_reclaim_iter *iter; | |
1134 | struct mem_cgroup_per_node *mz; | |
1135 | int nid; | |
1136 | ||
1137 | for_each_node(nid) { | |
1138 | mz = mem_cgroup_nodeinfo(from, nid); | |
1139 | iter = &mz->iter; | |
1140 | cmpxchg(&iter->position, dead_memcg, NULL); | |
1141 | } | |
1142 | } | |
1143 | ||
1144 | static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg) | |
1145 | { | |
1146 | struct mem_cgroup *memcg = dead_memcg; | |
1147 | struct mem_cgroup *last; | |
1148 | ||
1149 | do { | |
1150 | __invalidate_reclaim_iterators(memcg, dead_memcg); | |
1151 | last = memcg; | |
1152 | } while ((memcg = parent_mem_cgroup(memcg))); | |
1153 | ||
1154 | /* | |
1155 | * When cgruop1 non-hierarchy mode is used, | |
1156 | * parent_mem_cgroup() does not walk all the way up to the | |
1157 | * cgroup root (root_mem_cgroup). So we have to handle | |
1158 | * dead_memcg from cgroup root separately. | |
1159 | */ | |
1160 | if (last != root_mem_cgroup) | |
1161 | __invalidate_reclaim_iterators(root_mem_cgroup, | |
1162 | dead_memcg); | |
1163 | } | |
1164 | ||
1165 | /** | |
1166 | * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy | |
1167 | * @memcg: hierarchy root | |
1168 | * @fn: function to call for each task | |
1169 | * @arg: argument passed to @fn | |
1170 | * | |
1171 | * This function iterates over tasks attached to @memcg or to any of its | |
1172 | * descendants and calls @fn for each task. If @fn returns a non-zero | |
1173 | * value, the function breaks the iteration loop and returns the value. | |
1174 | * Otherwise, it will iterate over all tasks and return 0. | |
1175 | * | |
1176 | * This function must not be called for the root memory cgroup. | |
1177 | */ | |
1178 | int mem_cgroup_scan_tasks(struct mem_cgroup *memcg, | |
1179 | int (*fn)(struct task_struct *, void *), void *arg) | |
1180 | { | |
1181 | struct mem_cgroup *iter; | |
1182 | int ret = 0; | |
1183 | ||
1184 | BUG_ON(memcg == root_mem_cgroup); | |
1185 | ||
1186 | for_each_mem_cgroup_tree(iter, memcg) { | |
1187 | struct css_task_iter it; | |
1188 | struct task_struct *task; | |
1189 | ||
1190 | css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it); | |
1191 | while (!ret && (task = css_task_iter_next(&it))) | |
1192 | ret = fn(task, arg); | |
1193 | css_task_iter_end(&it); | |
1194 | if (ret) { | |
1195 | mem_cgroup_iter_break(memcg, iter); | |
1196 | break; | |
1197 | } | |
1198 | } | |
1199 | return ret; | |
1200 | } | |
1201 | ||
1202 | /** | |
1203 | * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page | |
1204 | * @page: the page | |
1205 | * @pgdat: pgdat of the page | |
1206 | * | |
1207 | * This function is only safe when following the LRU page isolation | |
1208 | * and putback protocol: the LRU lock must be held, and the page must | |
1209 | * either be PageLRU() or the caller must have isolated/allocated it. | |
1210 | */ | |
1211 | struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat) | |
1212 | { | |
1213 | struct mem_cgroup_per_node *mz; | |
1214 | struct mem_cgroup *memcg; | |
1215 | struct lruvec *lruvec; | |
1216 | ||
1217 | if (mem_cgroup_disabled()) { | |
1218 | lruvec = &pgdat->__lruvec; | |
1219 | goto out; | |
1220 | } | |
1221 | ||
1222 | memcg = page->mem_cgroup; | |
1223 | /* | |
1224 | * Swapcache readahead pages are added to the LRU - and | |
1225 | * possibly migrated - before they are charged. | |
1226 | */ | |
1227 | if (!memcg) | |
1228 | memcg = root_mem_cgroup; | |
1229 | ||
1230 | mz = mem_cgroup_page_nodeinfo(memcg, page); | |
1231 | lruvec = &mz->lruvec; | |
1232 | out: | |
1233 | /* | |
1234 | * Since a node can be onlined after the mem_cgroup was created, | |
1235 | * we have to be prepared to initialize lruvec->zone here; | |
1236 | * and if offlined then reonlined, we need to reinitialize it. | |
1237 | */ | |
1238 | if (unlikely(lruvec->pgdat != pgdat)) | |
1239 | lruvec->pgdat = pgdat; | |
1240 | return lruvec; | |
1241 | } | |
1242 | ||
1243 | /** | |
1244 | * mem_cgroup_update_lru_size - account for adding or removing an lru page | |
1245 | * @lruvec: mem_cgroup per zone lru vector | |
1246 | * @lru: index of lru list the page is sitting on | |
1247 | * @zid: zone id of the accounted pages | |
1248 | * @nr_pages: positive when adding or negative when removing | |
1249 | * | |
1250 | * This function must be called under lru_lock, just before a page is added | |
1251 | * to or just after a page is removed from an lru list (that ordering being | |
1252 | * so as to allow it to check that lru_size 0 is consistent with list_empty). | |
1253 | */ | |
1254 | void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru, | |
1255 | int zid, int nr_pages) | |
1256 | { | |
1257 | struct mem_cgroup_per_node *mz; | |
1258 | unsigned long *lru_size; | |
1259 | long size; | |
1260 | ||
1261 | if (mem_cgroup_disabled()) | |
1262 | return; | |
1263 | ||
1264 | mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec); | |
1265 | lru_size = &mz->lru_zone_size[zid][lru]; | |
1266 | ||
1267 | if (nr_pages < 0) | |
1268 | *lru_size += nr_pages; | |
1269 | ||
1270 | size = *lru_size; | |
1271 | if (WARN_ONCE(size < 0, | |
1272 | "%s(%p, %d, %d): lru_size %ld\n", | |
1273 | __func__, lruvec, lru, nr_pages, size)) { | |
1274 | VM_BUG_ON(1); | |
1275 | *lru_size = 0; | |
1276 | } | |
1277 | ||
1278 | if (nr_pages > 0) | |
1279 | *lru_size += nr_pages; | |
1280 | } | |
1281 | ||
1282 | /** | |
1283 | * mem_cgroup_margin - calculate chargeable space of a memory cgroup | |
1284 | * @memcg: the memory cgroup | |
1285 | * | |
1286 | * Returns the maximum amount of memory @mem can be charged with, in | |
1287 | * pages. | |
1288 | */ | |
1289 | static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) | |
1290 | { | |
1291 | unsigned long margin = 0; | |
1292 | unsigned long count; | |
1293 | unsigned long limit; | |
1294 | ||
1295 | count = page_counter_read(&memcg->memory); | |
1296 | limit = READ_ONCE(memcg->memory.max); | |
1297 | if (count < limit) | |
1298 | margin = limit - count; | |
1299 | ||
1300 | if (do_memsw_account()) { | |
1301 | count = page_counter_read(&memcg->memsw); | |
1302 | limit = READ_ONCE(memcg->memsw.max); | |
1303 | if (count <= limit) | |
1304 | margin = min(margin, limit - count); | |
1305 | else | |
1306 | margin = 0; | |
1307 | } | |
1308 | ||
1309 | return margin; | |
1310 | } | |
1311 | ||
1312 | /* | |
1313 | * A routine for checking "mem" is under move_account() or not. | |
1314 | * | |
1315 | * Checking a cgroup is mc.from or mc.to or under hierarchy of | |
1316 | * moving cgroups. This is for waiting at high-memory pressure | |
1317 | * caused by "move". | |
1318 | */ | |
1319 | static bool mem_cgroup_under_move(struct mem_cgroup *memcg) | |
1320 | { | |
1321 | struct mem_cgroup *from; | |
1322 | struct mem_cgroup *to; | |
1323 | bool ret = false; | |
1324 | /* | |
1325 | * Unlike task_move routines, we access mc.to, mc.from not under | |
1326 | * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. | |
1327 | */ | |
1328 | spin_lock(&mc.lock); | |
1329 | from = mc.from; | |
1330 | to = mc.to; | |
1331 | if (!from) | |
1332 | goto unlock; | |
1333 | ||
1334 | ret = mem_cgroup_is_descendant(from, memcg) || | |
1335 | mem_cgroup_is_descendant(to, memcg); | |
1336 | unlock: | |
1337 | spin_unlock(&mc.lock); | |
1338 | return ret; | |
1339 | } | |
1340 | ||
1341 | static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) | |
1342 | { | |
1343 | if (mc.moving_task && current != mc.moving_task) { | |
1344 | if (mem_cgroup_under_move(memcg)) { | |
1345 | DEFINE_WAIT(wait); | |
1346 | prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); | |
1347 | /* moving charge context might have finished. */ | |
1348 | if (mc.moving_task) | |
1349 | schedule(); | |
1350 | finish_wait(&mc.waitq, &wait); | |
1351 | return true; | |
1352 | } | |
1353 | } | |
1354 | return false; | |
1355 | } | |
1356 | ||
1357 | static char *memory_stat_format(struct mem_cgroup *memcg) | |
1358 | { | |
1359 | struct seq_buf s; | |
1360 | int i; | |
1361 | ||
1362 | seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE); | |
1363 | if (!s.buffer) | |
1364 | return NULL; | |
1365 | ||
1366 | /* | |
1367 | * Provide statistics on the state of the memory subsystem as | |
1368 | * well as cumulative event counters that show past behavior. | |
1369 | * | |
1370 | * This list is ordered following a combination of these gradients: | |
1371 | * 1) generic big picture -> specifics and details | |
1372 | * 2) reflecting userspace activity -> reflecting kernel heuristics | |
1373 | * | |
1374 | * Current memory state: | |
1375 | */ | |
1376 | ||
1377 | seq_buf_printf(&s, "anon %llu\n", | |
1378 | (u64)memcg_page_state(memcg, MEMCG_RSS) * | |
1379 | PAGE_SIZE); | |
1380 | seq_buf_printf(&s, "file %llu\n", | |
1381 | (u64)memcg_page_state(memcg, MEMCG_CACHE) * | |
1382 | PAGE_SIZE); | |
1383 | seq_buf_printf(&s, "kernel_stack %llu\n", | |
1384 | (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) * | |
1385 | 1024); | |
1386 | seq_buf_printf(&s, "slab %llu\n", | |
1387 | (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) + | |
1388 | memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) * | |
1389 | PAGE_SIZE); | |
1390 | seq_buf_printf(&s, "sock %llu\n", | |
1391 | (u64)memcg_page_state(memcg, MEMCG_SOCK) * | |
1392 | PAGE_SIZE); | |
1393 | ||
1394 | seq_buf_printf(&s, "shmem %llu\n", | |
1395 | (u64)memcg_page_state(memcg, NR_SHMEM) * | |
1396 | PAGE_SIZE); | |
1397 | seq_buf_printf(&s, "file_mapped %llu\n", | |
1398 | (u64)memcg_page_state(memcg, NR_FILE_MAPPED) * | |
1399 | PAGE_SIZE); | |
1400 | seq_buf_printf(&s, "file_dirty %llu\n", | |
1401 | (u64)memcg_page_state(memcg, NR_FILE_DIRTY) * | |
1402 | PAGE_SIZE); | |
1403 | seq_buf_printf(&s, "file_writeback %llu\n", | |
1404 | (u64)memcg_page_state(memcg, NR_WRITEBACK) * | |
1405 | PAGE_SIZE); | |
1406 | ||
1407 | /* | |
1408 | * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter | |
1409 | * with the NR_ANON_THP vm counter, but right now it's a pain in the | |
1410 | * arse because it requires migrating the work out of rmap to a place | |
1411 | * where the page->mem_cgroup is set up and stable. | |
1412 | */ | |
1413 | seq_buf_printf(&s, "anon_thp %llu\n", | |
1414 | (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) * | |
1415 | PAGE_SIZE); | |
1416 | ||
1417 | for (i = 0; i < NR_LRU_LISTS; i++) | |
1418 | seq_buf_printf(&s, "%s %llu\n", lru_list_name(i), | |
1419 | (u64)memcg_page_state(memcg, NR_LRU_BASE + i) * | |
1420 | PAGE_SIZE); | |
1421 | ||
1422 | seq_buf_printf(&s, "slab_reclaimable %llu\n", | |
1423 | (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) * | |
1424 | PAGE_SIZE); | |
1425 | seq_buf_printf(&s, "slab_unreclaimable %llu\n", | |
1426 | (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) * | |
1427 | PAGE_SIZE); | |
1428 | ||
1429 | /* Accumulated memory events */ | |
1430 | ||
1431 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT), | |
1432 | memcg_events(memcg, PGFAULT)); | |
1433 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT), | |
1434 | memcg_events(memcg, PGMAJFAULT)); | |
1435 | ||
1436 | seq_buf_printf(&s, "workingset_refault %lu\n", | |
1437 | memcg_page_state(memcg, WORKINGSET_REFAULT)); | |
1438 | seq_buf_printf(&s, "workingset_activate %lu\n", | |
1439 | memcg_page_state(memcg, WORKINGSET_ACTIVATE)); | |
1440 | seq_buf_printf(&s, "workingset_nodereclaim %lu\n", | |
1441 | memcg_page_state(memcg, WORKINGSET_NODERECLAIM)); | |
1442 | ||
1443 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL), | |
1444 | memcg_events(memcg, PGREFILL)); | |
1445 | seq_buf_printf(&s, "pgscan %lu\n", | |
1446 | memcg_events(memcg, PGSCAN_KSWAPD) + | |
1447 | memcg_events(memcg, PGSCAN_DIRECT)); | |
1448 | seq_buf_printf(&s, "pgsteal %lu\n", | |
1449 | memcg_events(memcg, PGSTEAL_KSWAPD) + | |
1450 | memcg_events(memcg, PGSTEAL_DIRECT)); | |
1451 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE), | |
1452 | memcg_events(memcg, PGACTIVATE)); | |
1453 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE), | |
1454 | memcg_events(memcg, PGDEACTIVATE)); | |
1455 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE), | |
1456 | memcg_events(memcg, PGLAZYFREE)); | |
1457 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED), | |
1458 | memcg_events(memcg, PGLAZYFREED)); | |
1459 | ||
1460 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
1461 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC), | |
1462 | memcg_events(memcg, THP_FAULT_ALLOC)); | |
1463 | seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC), | |
1464 | memcg_events(memcg, THP_COLLAPSE_ALLOC)); | |
1465 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
1466 | ||
1467 | /* The above should easily fit into one page */ | |
1468 | WARN_ON_ONCE(seq_buf_has_overflowed(&s)); | |
1469 | ||
1470 | return s.buffer; | |
1471 | } | |
1472 | ||
1473 | #define K(x) ((x) << (PAGE_SHIFT-10)) | |
1474 | /** | |
1475 | * mem_cgroup_print_oom_context: Print OOM information relevant to | |
1476 | * memory controller. | |
1477 | * @memcg: The memory cgroup that went over limit | |
1478 | * @p: Task that is going to be killed | |
1479 | * | |
1480 | * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is | |
1481 | * enabled | |
1482 | */ | |
1483 | void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p) | |
1484 | { | |
1485 | rcu_read_lock(); | |
1486 | ||
1487 | if (memcg) { | |
1488 | pr_cont(",oom_memcg="); | |
1489 | pr_cont_cgroup_path(memcg->css.cgroup); | |
1490 | } else | |
1491 | pr_cont(",global_oom"); | |
1492 | if (p) { | |
1493 | pr_cont(",task_memcg="); | |
1494 | pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id)); | |
1495 | } | |
1496 | rcu_read_unlock(); | |
1497 | } | |
1498 | ||
1499 | /** | |
1500 | * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to | |
1501 | * memory controller. | |
1502 | * @memcg: The memory cgroup that went over limit | |
1503 | */ | |
1504 | void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg) | |
1505 | { | |
1506 | char *buf; | |
1507 | ||
1508 | pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n", | |
1509 | K((u64)page_counter_read(&memcg->memory)), | |
1510 | K((u64)memcg->memory.max), memcg->memory.failcnt); | |
1511 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
1512 | pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n", | |
1513 | K((u64)page_counter_read(&memcg->swap)), | |
1514 | K((u64)memcg->swap.max), memcg->swap.failcnt); | |
1515 | else { | |
1516 | pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n", | |
1517 | K((u64)page_counter_read(&memcg->memsw)), | |
1518 | K((u64)memcg->memsw.max), memcg->memsw.failcnt); | |
1519 | pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n", | |
1520 | K((u64)page_counter_read(&memcg->kmem)), | |
1521 | K((u64)memcg->kmem.max), memcg->kmem.failcnt); | |
1522 | } | |
1523 | ||
1524 | pr_info("Memory cgroup stats for "); | |
1525 | pr_cont_cgroup_path(memcg->css.cgroup); | |
1526 | pr_cont(":"); | |
1527 | buf = memory_stat_format(memcg); | |
1528 | if (!buf) | |
1529 | return; | |
1530 | pr_info("%s", buf); | |
1531 | kfree(buf); | |
1532 | } | |
1533 | ||
1534 | /* | |
1535 | * Return the memory (and swap, if configured) limit for a memcg. | |
1536 | */ | |
1537 | unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg) | |
1538 | { | |
1539 | unsigned long max; | |
1540 | ||
1541 | max = memcg->memory.max; | |
1542 | if (mem_cgroup_swappiness(memcg)) { | |
1543 | unsigned long memsw_max; | |
1544 | unsigned long swap_max; | |
1545 | ||
1546 | memsw_max = memcg->memsw.max; | |
1547 | swap_max = memcg->swap.max; | |
1548 | swap_max = min(swap_max, (unsigned long)total_swap_pages); | |
1549 | max = min(max + swap_max, memsw_max); | |
1550 | } | |
1551 | return max; | |
1552 | } | |
1553 | ||
1554 | unsigned long mem_cgroup_size(struct mem_cgroup *memcg) | |
1555 | { | |
1556 | return page_counter_read(&memcg->memory); | |
1557 | } | |
1558 | ||
1559 | static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask, | |
1560 | int order) | |
1561 | { | |
1562 | struct oom_control oc = { | |
1563 | .zonelist = NULL, | |
1564 | .nodemask = NULL, | |
1565 | .memcg = memcg, | |
1566 | .gfp_mask = gfp_mask, | |
1567 | .order = order, | |
1568 | }; | |
1569 | bool ret; | |
1570 | ||
1571 | if (mutex_lock_killable(&oom_lock)) | |
1572 | return true; | |
1573 | /* | |
1574 | * A few threads which were not waiting at mutex_lock_killable() can | |
1575 | * fail to bail out. Therefore, check again after holding oom_lock. | |
1576 | */ | |
1577 | ret = should_force_charge() || out_of_memory(&oc); | |
1578 | mutex_unlock(&oom_lock); | |
1579 | return ret; | |
1580 | } | |
1581 | ||
1582 | static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, | |
1583 | pg_data_t *pgdat, | |
1584 | gfp_t gfp_mask, | |
1585 | unsigned long *total_scanned) | |
1586 | { | |
1587 | struct mem_cgroup *victim = NULL; | |
1588 | int total = 0; | |
1589 | int loop = 0; | |
1590 | unsigned long excess; | |
1591 | unsigned long nr_scanned; | |
1592 | struct mem_cgroup_reclaim_cookie reclaim = { | |
1593 | .pgdat = pgdat, | |
1594 | }; | |
1595 | ||
1596 | excess = soft_limit_excess(root_memcg); | |
1597 | ||
1598 | while (1) { | |
1599 | victim = mem_cgroup_iter(root_memcg, victim, &reclaim); | |
1600 | if (!victim) { | |
1601 | loop++; | |
1602 | if (loop >= 2) { | |
1603 | /* | |
1604 | * If we have not been able to reclaim | |
1605 | * anything, it might because there are | |
1606 | * no reclaimable pages under this hierarchy | |
1607 | */ | |
1608 | if (!total) | |
1609 | break; | |
1610 | /* | |
1611 | * We want to do more targeted reclaim. | |
1612 | * excess >> 2 is not to excessive so as to | |
1613 | * reclaim too much, nor too less that we keep | |
1614 | * coming back to reclaim from this cgroup | |
1615 | */ | |
1616 | if (total >= (excess >> 2) || | |
1617 | (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) | |
1618 | break; | |
1619 | } | |
1620 | continue; | |
1621 | } | |
1622 | total += mem_cgroup_shrink_node(victim, gfp_mask, false, | |
1623 | pgdat, &nr_scanned); | |
1624 | *total_scanned += nr_scanned; | |
1625 | if (!soft_limit_excess(root_memcg)) | |
1626 | break; | |
1627 | } | |
1628 | mem_cgroup_iter_break(root_memcg, victim); | |
1629 | return total; | |
1630 | } | |
1631 | ||
1632 | #ifdef CONFIG_LOCKDEP | |
1633 | static struct lockdep_map memcg_oom_lock_dep_map = { | |
1634 | .name = "memcg_oom_lock", | |
1635 | }; | |
1636 | #endif | |
1637 | ||
1638 | static DEFINE_SPINLOCK(memcg_oom_lock); | |
1639 | ||
1640 | /* | |
1641 | * Check OOM-Killer is already running under our hierarchy. | |
1642 | * If someone is running, return false. | |
1643 | */ | |
1644 | static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg) | |
1645 | { | |
1646 | struct mem_cgroup *iter, *failed = NULL; | |
1647 | ||
1648 | spin_lock(&memcg_oom_lock); | |
1649 | ||
1650 | for_each_mem_cgroup_tree(iter, memcg) { | |
1651 | if (iter->oom_lock) { | |
1652 | /* | |
1653 | * this subtree of our hierarchy is already locked | |
1654 | * so we cannot give a lock. | |
1655 | */ | |
1656 | failed = iter; | |
1657 | mem_cgroup_iter_break(memcg, iter); | |
1658 | break; | |
1659 | } else | |
1660 | iter->oom_lock = true; | |
1661 | } | |
1662 | ||
1663 | if (failed) { | |
1664 | /* | |
1665 | * OK, we failed to lock the whole subtree so we have | |
1666 | * to clean up what we set up to the failing subtree | |
1667 | */ | |
1668 | for_each_mem_cgroup_tree(iter, memcg) { | |
1669 | if (iter == failed) { | |
1670 | mem_cgroup_iter_break(memcg, iter); | |
1671 | break; | |
1672 | } | |
1673 | iter->oom_lock = false; | |
1674 | } | |
1675 | } else | |
1676 | mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_); | |
1677 | ||
1678 | spin_unlock(&memcg_oom_lock); | |
1679 | ||
1680 | return !failed; | |
1681 | } | |
1682 | ||
1683 | static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg) | |
1684 | { | |
1685 | struct mem_cgroup *iter; | |
1686 | ||
1687 | spin_lock(&memcg_oom_lock); | |
1688 | mutex_release(&memcg_oom_lock_dep_map, _RET_IP_); | |
1689 | for_each_mem_cgroup_tree(iter, memcg) | |
1690 | iter->oom_lock = false; | |
1691 | spin_unlock(&memcg_oom_lock); | |
1692 | } | |
1693 | ||
1694 | static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) | |
1695 | { | |
1696 | struct mem_cgroup *iter; | |
1697 | ||
1698 | spin_lock(&memcg_oom_lock); | |
1699 | for_each_mem_cgroup_tree(iter, memcg) | |
1700 | iter->under_oom++; | |
1701 | spin_unlock(&memcg_oom_lock); | |
1702 | } | |
1703 | ||
1704 | static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) | |
1705 | { | |
1706 | struct mem_cgroup *iter; | |
1707 | ||
1708 | /* | |
1709 | * When a new child is created while the hierarchy is under oom, | |
1710 | * mem_cgroup_oom_lock() may not be called. Watch for underflow. | |
1711 | */ | |
1712 | spin_lock(&memcg_oom_lock); | |
1713 | for_each_mem_cgroup_tree(iter, memcg) | |
1714 | if (iter->under_oom > 0) | |
1715 | iter->under_oom--; | |
1716 | spin_unlock(&memcg_oom_lock); | |
1717 | } | |
1718 | ||
1719 | static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); | |
1720 | ||
1721 | struct oom_wait_info { | |
1722 | struct mem_cgroup *memcg; | |
1723 | wait_queue_entry_t wait; | |
1724 | }; | |
1725 | ||
1726 | static int memcg_oom_wake_function(wait_queue_entry_t *wait, | |
1727 | unsigned mode, int sync, void *arg) | |
1728 | { | |
1729 | struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg; | |
1730 | struct mem_cgroup *oom_wait_memcg; | |
1731 | struct oom_wait_info *oom_wait_info; | |
1732 | ||
1733 | oom_wait_info = container_of(wait, struct oom_wait_info, wait); | |
1734 | oom_wait_memcg = oom_wait_info->memcg; | |
1735 | ||
1736 | if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) && | |
1737 | !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg)) | |
1738 | return 0; | |
1739 | return autoremove_wake_function(wait, mode, sync, arg); | |
1740 | } | |
1741 | ||
1742 | static void memcg_oom_recover(struct mem_cgroup *memcg) | |
1743 | { | |
1744 | /* | |
1745 | * For the following lockless ->under_oom test, the only required | |
1746 | * guarantee is that it must see the state asserted by an OOM when | |
1747 | * this function is called as a result of userland actions | |
1748 | * triggered by the notification of the OOM. This is trivially | |
1749 | * achieved by invoking mem_cgroup_mark_under_oom() before | |
1750 | * triggering notification. | |
1751 | */ | |
1752 | if (memcg && memcg->under_oom) | |
1753 | __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); | |
1754 | } | |
1755 | ||
1756 | enum oom_status { | |
1757 | OOM_SUCCESS, | |
1758 | OOM_FAILED, | |
1759 | OOM_ASYNC, | |
1760 | OOM_SKIPPED | |
1761 | }; | |
1762 | ||
1763 | static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order) | |
1764 | { | |
1765 | enum oom_status ret; | |
1766 | bool locked; | |
1767 | ||
1768 | if (order > PAGE_ALLOC_COSTLY_ORDER) | |
1769 | return OOM_SKIPPED; | |
1770 | ||
1771 | memcg_memory_event(memcg, MEMCG_OOM); | |
1772 | ||
1773 | /* | |
1774 | * We are in the middle of the charge context here, so we | |
1775 | * don't want to block when potentially sitting on a callstack | |
1776 | * that holds all kinds of filesystem and mm locks. | |
1777 | * | |
1778 | * cgroup1 allows disabling the OOM killer and waiting for outside | |
1779 | * handling until the charge can succeed; remember the context and put | |
1780 | * the task to sleep at the end of the page fault when all locks are | |
1781 | * released. | |
1782 | * | |
1783 | * On the other hand, in-kernel OOM killer allows for an async victim | |
1784 | * memory reclaim (oom_reaper) and that means that we are not solely | |
1785 | * relying on the oom victim to make a forward progress and we can | |
1786 | * invoke the oom killer here. | |
1787 | * | |
1788 | * Please note that mem_cgroup_out_of_memory might fail to find a | |
1789 | * victim and then we have to bail out from the charge path. | |
1790 | */ | |
1791 | if (memcg->oom_kill_disable) { | |
1792 | if (!current->in_user_fault) | |
1793 | return OOM_SKIPPED; | |
1794 | css_get(&memcg->css); | |
1795 | current->memcg_in_oom = memcg; | |
1796 | current->memcg_oom_gfp_mask = mask; | |
1797 | current->memcg_oom_order = order; | |
1798 | ||
1799 | return OOM_ASYNC; | |
1800 | } | |
1801 | ||
1802 | mem_cgroup_mark_under_oom(memcg); | |
1803 | ||
1804 | locked = mem_cgroup_oom_trylock(memcg); | |
1805 | ||
1806 | if (locked) | |
1807 | mem_cgroup_oom_notify(memcg); | |
1808 | ||
1809 | mem_cgroup_unmark_under_oom(memcg); | |
1810 | if (mem_cgroup_out_of_memory(memcg, mask, order)) | |
1811 | ret = OOM_SUCCESS; | |
1812 | else | |
1813 | ret = OOM_FAILED; | |
1814 | ||
1815 | if (locked) | |
1816 | mem_cgroup_oom_unlock(memcg); | |
1817 | ||
1818 | return ret; | |
1819 | } | |
1820 | ||
1821 | /** | |
1822 | * mem_cgroup_oom_synchronize - complete memcg OOM handling | |
1823 | * @handle: actually kill/wait or just clean up the OOM state | |
1824 | * | |
1825 | * This has to be called at the end of a page fault if the memcg OOM | |
1826 | * handler was enabled. | |
1827 | * | |
1828 | * Memcg supports userspace OOM handling where failed allocations must | |
1829 | * sleep on a waitqueue until the userspace task resolves the | |
1830 | * situation. Sleeping directly in the charge context with all kinds | |
1831 | * of locks held is not a good idea, instead we remember an OOM state | |
1832 | * in the task and mem_cgroup_oom_synchronize() has to be called at | |
1833 | * the end of the page fault to complete the OOM handling. | |
1834 | * | |
1835 | * Returns %true if an ongoing memcg OOM situation was detected and | |
1836 | * completed, %false otherwise. | |
1837 | */ | |
1838 | bool mem_cgroup_oom_synchronize(bool handle) | |
1839 | { | |
1840 | struct mem_cgroup *memcg = current->memcg_in_oom; | |
1841 | struct oom_wait_info owait; | |
1842 | bool locked; | |
1843 | ||
1844 | /* OOM is global, do not handle */ | |
1845 | if (!memcg) | |
1846 | return false; | |
1847 | ||
1848 | if (!handle) | |
1849 | goto cleanup; | |
1850 | ||
1851 | owait.memcg = memcg; | |
1852 | owait.wait.flags = 0; | |
1853 | owait.wait.func = memcg_oom_wake_function; | |
1854 | owait.wait.private = current; | |
1855 | INIT_LIST_HEAD(&owait.wait.entry); | |
1856 | ||
1857 | prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); | |
1858 | mem_cgroup_mark_under_oom(memcg); | |
1859 | ||
1860 | locked = mem_cgroup_oom_trylock(memcg); | |
1861 | ||
1862 | if (locked) | |
1863 | mem_cgroup_oom_notify(memcg); | |
1864 | ||
1865 | if (locked && !memcg->oom_kill_disable) { | |
1866 | mem_cgroup_unmark_under_oom(memcg); | |
1867 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1868 | mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask, | |
1869 | current->memcg_oom_order); | |
1870 | } else { | |
1871 | schedule(); | |
1872 | mem_cgroup_unmark_under_oom(memcg); | |
1873 | finish_wait(&memcg_oom_waitq, &owait.wait); | |
1874 | } | |
1875 | ||
1876 | if (locked) { | |
1877 | mem_cgroup_oom_unlock(memcg); | |
1878 | /* | |
1879 | * There is no guarantee that an OOM-lock contender | |
1880 | * sees the wakeups triggered by the OOM kill | |
1881 | * uncharges. Wake any sleepers explicitely. | |
1882 | */ | |
1883 | memcg_oom_recover(memcg); | |
1884 | } | |
1885 | cleanup: | |
1886 | current->memcg_in_oom = NULL; | |
1887 | css_put(&memcg->css); | |
1888 | return true; | |
1889 | } | |
1890 | ||
1891 | /** | |
1892 | * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM | |
1893 | * @victim: task to be killed by the OOM killer | |
1894 | * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM | |
1895 | * | |
1896 | * Returns a pointer to a memory cgroup, which has to be cleaned up | |
1897 | * by killing all belonging OOM-killable tasks. | |
1898 | * | |
1899 | * Caller has to call mem_cgroup_put() on the returned non-NULL memcg. | |
1900 | */ | |
1901 | struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim, | |
1902 | struct mem_cgroup *oom_domain) | |
1903 | { | |
1904 | struct mem_cgroup *oom_group = NULL; | |
1905 | struct mem_cgroup *memcg; | |
1906 | ||
1907 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
1908 | return NULL; | |
1909 | ||
1910 | if (!oom_domain) | |
1911 | oom_domain = root_mem_cgroup; | |
1912 | ||
1913 | rcu_read_lock(); | |
1914 | ||
1915 | memcg = mem_cgroup_from_task(victim); | |
1916 | if (memcg == root_mem_cgroup) | |
1917 | goto out; | |
1918 | ||
1919 | /* | |
1920 | * Traverse the memory cgroup hierarchy from the victim task's | |
1921 | * cgroup up to the OOMing cgroup (or root) to find the | |
1922 | * highest-level memory cgroup with oom.group set. | |
1923 | */ | |
1924 | for (; memcg; memcg = parent_mem_cgroup(memcg)) { | |
1925 | if (memcg->oom_group) | |
1926 | oom_group = memcg; | |
1927 | ||
1928 | if (memcg == oom_domain) | |
1929 | break; | |
1930 | } | |
1931 | ||
1932 | if (oom_group) | |
1933 | css_get(&oom_group->css); | |
1934 | out: | |
1935 | rcu_read_unlock(); | |
1936 | ||
1937 | return oom_group; | |
1938 | } | |
1939 | ||
1940 | void mem_cgroup_print_oom_group(struct mem_cgroup *memcg) | |
1941 | { | |
1942 | pr_info("Tasks in "); | |
1943 | pr_cont_cgroup_path(memcg->css.cgroup); | |
1944 | pr_cont(" are going to be killed due to memory.oom.group set\n"); | |
1945 | } | |
1946 | ||
1947 | /** | |
1948 | * lock_page_memcg - lock a page->mem_cgroup binding | |
1949 | * @page: the page | |
1950 | * | |
1951 | * This function protects unlocked LRU pages from being moved to | |
1952 | * another cgroup. | |
1953 | * | |
1954 | * It ensures lifetime of the returned memcg. Caller is responsible | |
1955 | * for the lifetime of the page; __unlock_page_memcg() is available | |
1956 | * when @page might get freed inside the locked section. | |
1957 | */ | |
1958 | struct mem_cgroup *lock_page_memcg(struct page *page) | |
1959 | { | |
1960 | struct mem_cgroup *memcg; | |
1961 | unsigned long flags; | |
1962 | ||
1963 | /* | |
1964 | * The RCU lock is held throughout the transaction. The fast | |
1965 | * path can get away without acquiring the memcg->move_lock | |
1966 | * because page moving starts with an RCU grace period. | |
1967 | * | |
1968 | * The RCU lock also protects the memcg from being freed when | |
1969 | * the page state that is going to change is the only thing | |
1970 | * preventing the page itself from being freed. E.g. writeback | |
1971 | * doesn't hold a page reference and relies on PG_writeback to | |
1972 | * keep off truncation, migration and so forth. | |
1973 | */ | |
1974 | rcu_read_lock(); | |
1975 | ||
1976 | if (mem_cgroup_disabled()) | |
1977 | return NULL; | |
1978 | again: | |
1979 | memcg = page->mem_cgroup; | |
1980 | if (unlikely(!memcg)) | |
1981 | return NULL; | |
1982 | ||
1983 | if (atomic_read(&memcg->moving_account) <= 0) | |
1984 | return memcg; | |
1985 | ||
1986 | spin_lock_irqsave(&memcg->move_lock, flags); | |
1987 | if (memcg != page->mem_cgroup) { | |
1988 | spin_unlock_irqrestore(&memcg->move_lock, flags); | |
1989 | goto again; | |
1990 | } | |
1991 | ||
1992 | /* | |
1993 | * When charge migration first begins, we can have locked and | |
1994 | * unlocked page stat updates happening concurrently. Track | |
1995 | * the task who has the lock for unlock_page_memcg(). | |
1996 | */ | |
1997 | memcg->move_lock_task = current; | |
1998 | memcg->move_lock_flags = flags; | |
1999 | ||
2000 | return memcg; | |
2001 | } | |
2002 | EXPORT_SYMBOL(lock_page_memcg); | |
2003 | ||
2004 | /** | |
2005 | * __unlock_page_memcg - unlock and unpin a memcg | |
2006 | * @memcg: the memcg | |
2007 | * | |
2008 | * Unlock and unpin a memcg returned by lock_page_memcg(). | |
2009 | */ | |
2010 | void __unlock_page_memcg(struct mem_cgroup *memcg) | |
2011 | { | |
2012 | if (memcg && memcg->move_lock_task == current) { | |
2013 | unsigned long flags = memcg->move_lock_flags; | |
2014 | ||
2015 | memcg->move_lock_task = NULL; | |
2016 | memcg->move_lock_flags = 0; | |
2017 | ||
2018 | spin_unlock_irqrestore(&memcg->move_lock, flags); | |
2019 | } | |
2020 | ||
2021 | rcu_read_unlock(); | |
2022 | } | |
2023 | ||
2024 | /** | |
2025 | * unlock_page_memcg - unlock a page->mem_cgroup binding | |
2026 | * @page: the page | |
2027 | */ | |
2028 | void unlock_page_memcg(struct page *page) | |
2029 | { | |
2030 | __unlock_page_memcg(page->mem_cgroup); | |
2031 | } | |
2032 | EXPORT_SYMBOL(unlock_page_memcg); | |
2033 | ||
2034 | struct memcg_stock_pcp { | |
2035 | struct mem_cgroup *cached; /* this never be root cgroup */ | |
2036 | unsigned int nr_pages; | |
2037 | struct work_struct work; | |
2038 | unsigned long flags; | |
2039 | #define FLUSHING_CACHED_CHARGE 0 | |
2040 | }; | |
2041 | static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); | |
2042 | static DEFINE_MUTEX(percpu_charge_mutex); | |
2043 | ||
2044 | /** | |
2045 | * consume_stock: Try to consume stocked charge on this cpu. | |
2046 | * @memcg: memcg to consume from. | |
2047 | * @nr_pages: how many pages to charge. | |
2048 | * | |
2049 | * The charges will only happen if @memcg matches the current cpu's memcg | |
2050 | * stock, and at least @nr_pages are available in that stock. Failure to | |
2051 | * service an allocation will refill the stock. | |
2052 | * | |
2053 | * returns true if successful, false otherwise. | |
2054 | */ | |
2055 | static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2056 | { | |
2057 | struct memcg_stock_pcp *stock; | |
2058 | unsigned long flags; | |
2059 | bool ret = false; | |
2060 | ||
2061 | if (nr_pages > MEMCG_CHARGE_BATCH) | |
2062 | return ret; | |
2063 | ||
2064 | local_irq_save(flags); | |
2065 | ||
2066 | stock = this_cpu_ptr(&memcg_stock); | |
2067 | if (memcg == stock->cached && stock->nr_pages >= nr_pages) { | |
2068 | stock->nr_pages -= nr_pages; | |
2069 | ret = true; | |
2070 | } | |
2071 | ||
2072 | local_irq_restore(flags); | |
2073 | ||
2074 | return ret; | |
2075 | } | |
2076 | ||
2077 | /* | |
2078 | * Returns stocks cached in percpu and reset cached information. | |
2079 | */ | |
2080 | static void drain_stock(struct memcg_stock_pcp *stock) | |
2081 | { | |
2082 | struct mem_cgroup *old = stock->cached; | |
2083 | ||
2084 | if (stock->nr_pages) { | |
2085 | page_counter_uncharge(&old->memory, stock->nr_pages); | |
2086 | if (do_memsw_account()) | |
2087 | page_counter_uncharge(&old->memsw, stock->nr_pages); | |
2088 | css_put_many(&old->css, stock->nr_pages); | |
2089 | stock->nr_pages = 0; | |
2090 | } | |
2091 | stock->cached = NULL; | |
2092 | } | |
2093 | ||
2094 | static void drain_local_stock(struct work_struct *dummy) | |
2095 | { | |
2096 | struct memcg_stock_pcp *stock; | |
2097 | unsigned long flags; | |
2098 | ||
2099 | /* | |
2100 | * The only protection from memory hotplug vs. drain_stock races is | |
2101 | * that we always operate on local CPU stock here with IRQ disabled | |
2102 | */ | |
2103 | local_irq_save(flags); | |
2104 | ||
2105 | stock = this_cpu_ptr(&memcg_stock); | |
2106 | drain_stock(stock); | |
2107 | clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); | |
2108 | ||
2109 | local_irq_restore(flags); | |
2110 | } | |
2111 | ||
2112 | /* | |
2113 | * Cache charges(val) to local per_cpu area. | |
2114 | * This will be consumed by consume_stock() function, later. | |
2115 | */ | |
2116 | static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2117 | { | |
2118 | struct memcg_stock_pcp *stock; | |
2119 | unsigned long flags; | |
2120 | ||
2121 | local_irq_save(flags); | |
2122 | ||
2123 | stock = this_cpu_ptr(&memcg_stock); | |
2124 | if (stock->cached != memcg) { /* reset if necessary */ | |
2125 | drain_stock(stock); | |
2126 | stock->cached = memcg; | |
2127 | } | |
2128 | stock->nr_pages += nr_pages; | |
2129 | ||
2130 | if (stock->nr_pages > MEMCG_CHARGE_BATCH) | |
2131 | drain_stock(stock); | |
2132 | ||
2133 | local_irq_restore(flags); | |
2134 | } | |
2135 | ||
2136 | /* | |
2137 | * Drains all per-CPU charge caches for given root_memcg resp. subtree | |
2138 | * of the hierarchy under it. | |
2139 | */ | |
2140 | static void drain_all_stock(struct mem_cgroup *root_memcg) | |
2141 | { | |
2142 | int cpu, curcpu; | |
2143 | ||
2144 | /* If someone's already draining, avoid adding running more workers. */ | |
2145 | if (!mutex_trylock(&percpu_charge_mutex)) | |
2146 | return; | |
2147 | /* | |
2148 | * Notify other cpus that system-wide "drain" is running | |
2149 | * We do not care about races with the cpu hotplug because cpu down | |
2150 | * as well as workers from this path always operate on the local | |
2151 | * per-cpu data. CPU up doesn't touch memcg_stock at all. | |
2152 | */ | |
2153 | curcpu = get_cpu(); | |
2154 | for_each_online_cpu(cpu) { | |
2155 | struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); | |
2156 | struct mem_cgroup *memcg; | |
2157 | bool flush = false; | |
2158 | ||
2159 | rcu_read_lock(); | |
2160 | memcg = stock->cached; | |
2161 | if (memcg && stock->nr_pages && | |
2162 | mem_cgroup_is_descendant(memcg, root_memcg)) | |
2163 | flush = true; | |
2164 | rcu_read_unlock(); | |
2165 | ||
2166 | if (flush && | |
2167 | !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { | |
2168 | if (cpu == curcpu) | |
2169 | drain_local_stock(&stock->work); | |
2170 | else | |
2171 | schedule_work_on(cpu, &stock->work); | |
2172 | } | |
2173 | } | |
2174 | put_cpu(); | |
2175 | mutex_unlock(&percpu_charge_mutex); | |
2176 | } | |
2177 | ||
2178 | static int memcg_hotplug_cpu_dead(unsigned int cpu) | |
2179 | { | |
2180 | struct memcg_stock_pcp *stock; | |
2181 | struct mem_cgroup *memcg, *mi; | |
2182 | ||
2183 | stock = &per_cpu(memcg_stock, cpu); | |
2184 | drain_stock(stock); | |
2185 | ||
2186 | for_each_mem_cgroup(memcg) { | |
2187 | int i; | |
2188 | ||
2189 | for (i = 0; i < MEMCG_NR_STAT; i++) { | |
2190 | int nid; | |
2191 | long x; | |
2192 | ||
2193 | x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0); | |
2194 | if (x) | |
2195 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
2196 | atomic_long_add(x, &memcg->vmstats[i]); | |
2197 | ||
2198 | if (i >= NR_VM_NODE_STAT_ITEMS) | |
2199 | continue; | |
2200 | ||
2201 | for_each_node(nid) { | |
2202 | struct mem_cgroup_per_node *pn; | |
2203 | ||
2204 | pn = mem_cgroup_nodeinfo(memcg, nid); | |
2205 | x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0); | |
2206 | if (x) | |
2207 | do { | |
2208 | atomic_long_add(x, &pn->lruvec_stat[i]); | |
2209 | } while ((pn = parent_nodeinfo(pn, nid))); | |
2210 | } | |
2211 | } | |
2212 | ||
2213 | for (i = 0; i < NR_VM_EVENT_ITEMS; i++) { | |
2214 | long x; | |
2215 | ||
2216 | x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0); | |
2217 | if (x) | |
2218 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
2219 | atomic_long_add(x, &memcg->vmevents[i]); | |
2220 | } | |
2221 | } | |
2222 | ||
2223 | return 0; | |
2224 | } | |
2225 | ||
2226 | static void reclaim_high(struct mem_cgroup *memcg, | |
2227 | unsigned int nr_pages, | |
2228 | gfp_t gfp_mask) | |
2229 | { | |
2230 | do { | |
2231 | if (page_counter_read(&memcg->memory) <= memcg->high) | |
2232 | continue; | |
2233 | memcg_memory_event(memcg, MEMCG_HIGH); | |
2234 | try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true); | |
2235 | } while ((memcg = parent_mem_cgroup(memcg))); | |
2236 | } | |
2237 | ||
2238 | static void high_work_func(struct work_struct *work) | |
2239 | { | |
2240 | struct mem_cgroup *memcg; | |
2241 | ||
2242 | memcg = container_of(work, struct mem_cgroup, high_work); | |
2243 | reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL); | |
2244 | } | |
2245 | ||
2246 | /* | |
2247 | * Clamp the maximum sleep time per allocation batch to 2 seconds. This is | |
2248 | * enough to still cause a significant slowdown in most cases, while still | |
2249 | * allowing diagnostics and tracing to proceed without becoming stuck. | |
2250 | */ | |
2251 | #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ) | |
2252 | ||
2253 | /* | |
2254 | * When calculating the delay, we use these either side of the exponentiation to | |
2255 | * maintain precision and scale to a reasonable number of jiffies (see the table | |
2256 | * below. | |
2257 | * | |
2258 | * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the | |
2259 | * overage ratio to a delay. | |
2260 | * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the | |
2261 | * proposed penalty in order to reduce to a reasonable number of jiffies, and | |
2262 | * to produce a reasonable delay curve. | |
2263 | * | |
2264 | * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a | |
2265 | * reasonable delay curve compared to precision-adjusted overage, not | |
2266 | * penalising heavily at first, but still making sure that growth beyond the | |
2267 | * limit penalises misbehaviour cgroups by slowing them down exponentially. For | |
2268 | * example, with a high of 100 megabytes: | |
2269 | * | |
2270 | * +-------+------------------------+ | |
2271 | * | usage | time to allocate in ms | | |
2272 | * +-------+------------------------+ | |
2273 | * | 100M | 0 | | |
2274 | * | 101M | 6 | | |
2275 | * | 102M | 25 | | |
2276 | * | 103M | 57 | | |
2277 | * | 104M | 102 | | |
2278 | * | 105M | 159 | | |
2279 | * | 106M | 230 | | |
2280 | * | 107M | 313 | | |
2281 | * | 108M | 409 | | |
2282 | * | 109M | 518 | | |
2283 | * | 110M | 639 | | |
2284 | * | 111M | 774 | | |
2285 | * | 112M | 921 | | |
2286 | * | 113M | 1081 | | |
2287 | * | 114M | 1254 | | |
2288 | * | 115M | 1439 | | |
2289 | * | 116M | 1638 | | |
2290 | * | 117M | 1849 | | |
2291 | * | 118M | 2000 | | |
2292 | * | 119M | 2000 | | |
2293 | * | 120M | 2000 | | |
2294 | * +-------+------------------------+ | |
2295 | */ | |
2296 | #define MEMCG_DELAY_PRECISION_SHIFT 20 | |
2297 | #define MEMCG_DELAY_SCALING_SHIFT 14 | |
2298 | ||
2299 | /* | |
2300 | * Get the number of jiffies that we should penalise a mischievous cgroup which | |
2301 | * is exceeding its memory.high by checking both it and its ancestors. | |
2302 | */ | |
2303 | static unsigned long calculate_high_delay(struct mem_cgroup *memcg, | |
2304 | unsigned int nr_pages) | |
2305 | { | |
2306 | unsigned long penalty_jiffies; | |
2307 | u64 max_overage = 0; | |
2308 | ||
2309 | do { | |
2310 | unsigned long usage, high; | |
2311 | u64 overage; | |
2312 | ||
2313 | usage = page_counter_read(&memcg->memory); | |
2314 | high = READ_ONCE(memcg->high); | |
2315 | ||
2316 | /* | |
2317 | * Prevent division by 0 in overage calculation by acting as if | |
2318 | * it was a threshold of 1 page | |
2319 | */ | |
2320 | high = max(high, 1UL); | |
2321 | ||
2322 | overage = usage - high; | |
2323 | overage <<= MEMCG_DELAY_PRECISION_SHIFT; | |
2324 | overage = div64_u64(overage, high); | |
2325 | ||
2326 | if (overage > max_overage) | |
2327 | max_overage = overage; | |
2328 | } while ((memcg = parent_mem_cgroup(memcg)) && | |
2329 | !mem_cgroup_is_root(memcg)); | |
2330 | ||
2331 | if (!max_overage) | |
2332 | return 0; | |
2333 | ||
2334 | /* | |
2335 | * We use overage compared to memory.high to calculate the number of | |
2336 | * jiffies to sleep (penalty_jiffies). Ideally this value should be | |
2337 | * fairly lenient on small overages, and increasingly harsh when the | |
2338 | * memcg in question makes it clear that it has no intention of stopping | |
2339 | * its crazy behaviour, so we exponentially increase the delay based on | |
2340 | * overage amount. | |
2341 | */ | |
2342 | penalty_jiffies = max_overage * max_overage * HZ; | |
2343 | penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT; | |
2344 | penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT; | |
2345 | ||
2346 | /* | |
2347 | * Factor in the task's own contribution to the overage, such that four | |
2348 | * N-sized allocations are throttled approximately the same as one | |
2349 | * 4N-sized allocation. | |
2350 | * | |
2351 | * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or | |
2352 | * larger the current charge patch is than that. | |
2353 | */ | |
2354 | penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH; | |
2355 | ||
2356 | /* | |
2357 | * Clamp the max delay per usermode return so as to still keep the | |
2358 | * application moving forwards and also permit diagnostics, albeit | |
2359 | * extremely slowly. | |
2360 | */ | |
2361 | return min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES); | |
2362 | } | |
2363 | ||
2364 | /* | |
2365 | * Scheduled by try_charge() to be executed from the userland return path | |
2366 | * and reclaims memory over the high limit. | |
2367 | */ | |
2368 | void mem_cgroup_handle_over_high(void) | |
2369 | { | |
2370 | unsigned long penalty_jiffies; | |
2371 | unsigned long pflags; | |
2372 | unsigned int nr_pages = current->memcg_nr_pages_over_high; | |
2373 | struct mem_cgroup *memcg; | |
2374 | ||
2375 | if (likely(!nr_pages)) | |
2376 | return; | |
2377 | ||
2378 | memcg = get_mem_cgroup_from_mm(current->mm); | |
2379 | reclaim_high(memcg, nr_pages, GFP_KERNEL); | |
2380 | current->memcg_nr_pages_over_high = 0; | |
2381 | ||
2382 | /* | |
2383 | * memory.high is breached and reclaim is unable to keep up. Throttle | |
2384 | * allocators proactively to slow down excessive growth. | |
2385 | */ | |
2386 | penalty_jiffies = calculate_high_delay(memcg, nr_pages); | |
2387 | ||
2388 | /* | |
2389 | * Don't sleep if the amount of jiffies this memcg owes us is so low | |
2390 | * that it's not even worth doing, in an attempt to be nice to those who | |
2391 | * go only a small amount over their memory.high value and maybe haven't | |
2392 | * been aggressively reclaimed enough yet. | |
2393 | */ | |
2394 | if (penalty_jiffies <= HZ / 100) | |
2395 | goto out; | |
2396 | ||
2397 | /* | |
2398 | * If we exit early, we're guaranteed to die (since | |
2399 | * schedule_timeout_killable sets TASK_KILLABLE). This means we don't | |
2400 | * need to account for any ill-begotten jiffies to pay them off later. | |
2401 | */ | |
2402 | psi_memstall_enter(&pflags); | |
2403 | schedule_timeout_killable(penalty_jiffies); | |
2404 | psi_memstall_leave(&pflags); | |
2405 | ||
2406 | out: | |
2407 | css_put(&memcg->css); | |
2408 | } | |
2409 | ||
2410 | static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, | |
2411 | unsigned int nr_pages) | |
2412 | { | |
2413 | unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages); | |
2414 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2415 | struct mem_cgroup *mem_over_limit; | |
2416 | struct page_counter *counter; | |
2417 | unsigned long nr_reclaimed; | |
2418 | bool may_swap = true; | |
2419 | bool drained = false; | |
2420 | enum oom_status oom_status; | |
2421 | ||
2422 | if (mem_cgroup_is_root(memcg)) | |
2423 | return 0; | |
2424 | retry: | |
2425 | if (consume_stock(memcg, nr_pages)) | |
2426 | return 0; | |
2427 | ||
2428 | if (!do_memsw_account() || | |
2429 | page_counter_try_charge(&memcg->memsw, batch, &counter)) { | |
2430 | if (page_counter_try_charge(&memcg->memory, batch, &counter)) | |
2431 | goto done_restock; | |
2432 | if (do_memsw_account()) | |
2433 | page_counter_uncharge(&memcg->memsw, batch); | |
2434 | mem_over_limit = mem_cgroup_from_counter(counter, memory); | |
2435 | } else { | |
2436 | mem_over_limit = mem_cgroup_from_counter(counter, memsw); | |
2437 | may_swap = false; | |
2438 | } | |
2439 | ||
2440 | if (batch > nr_pages) { | |
2441 | batch = nr_pages; | |
2442 | goto retry; | |
2443 | } | |
2444 | ||
2445 | /* | |
2446 | * Memcg doesn't have a dedicated reserve for atomic | |
2447 | * allocations. But like the global atomic pool, we need to | |
2448 | * put the burden of reclaim on regular allocation requests | |
2449 | * and let these go through as privileged allocations. | |
2450 | */ | |
2451 | if (gfp_mask & __GFP_ATOMIC) | |
2452 | goto force; | |
2453 | ||
2454 | /* | |
2455 | * Unlike in global OOM situations, memcg is not in a physical | |
2456 | * memory shortage. Allow dying and OOM-killed tasks to | |
2457 | * bypass the last charges so that they can exit quickly and | |
2458 | * free their memory. | |
2459 | */ | |
2460 | if (unlikely(should_force_charge())) | |
2461 | goto force; | |
2462 | ||
2463 | /* | |
2464 | * Prevent unbounded recursion when reclaim operations need to | |
2465 | * allocate memory. This might exceed the limits temporarily, | |
2466 | * but we prefer facilitating memory reclaim and getting back | |
2467 | * under the limit over triggering OOM kills in these cases. | |
2468 | */ | |
2469 | if (unlikely(current->flags & PF_MEMALLOC)) | |
2470 | goto force; | |
2471 | ||
2472 | if (unlikely(task_in_memcg_oom(current))) | |
2473 | goto nomem; | |
2474 | ||
2475 | if (!gfpflags_allow_blocking(gfp_mask)) | |
2476 | goto nomem; | |
2477 | ||
2478 | memcg_memory_event(mem_over_limit, MEMCG_MAX); | |
2479 | ||
2480 | nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages, | |
2481 | gfp_mask, may_swap); | |
2482 | ||
2483 | if (mem_cgroup_margin(mem_over_limit) >= nr_pages) | |
2484 | goto retry; | |
2485 | ||
2486 | if (!drained) { | |
2487 | drain_all_stock(mem_over_limit); | |
2488 | drained = true; | |
2489 | goto retry; | |
2490 | } | |
2491 | ||
2492 | if (gfp_mask & __GFP_NORETRY) | |
2493 | goto nomem; | |
2494 | /* | |
2495 | * Even though the limit is exceeded at this point, reclaim | |
2496 | * may have been able to free some pages. Retry the charge | |
2497 | * before killing the task. | |
2498 | * | |
2499 | * Only for regular pages, though: huge pages are rather | |
2500 | * unlikely to succeed so close to the limit, and we fall back | |
2501 | * to regular pages anyway in case of failure. | |
2502 | */ | |
2503 | if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER)) | |
2504 | goto retry; | |
2505 | /* | |
2506 | * At task move, charge accounts can be doubly counted. So, it's | |
2507 | * better to wait until the end of task_move if something is going on. | |
2508 | */ | |
2509 | if (mem_cgroup_wait_acct_move(mem_over_limit)) | |
2510 | goto retry; | |
2511 | ||
2512 | if (nr_retries--) | |
2513 | goto retry; | |
2514 | ||
2515 | if (gfp_mask & __GFP_RETRY_MAYFAIL) | |
2516 | goto nomem; | |
2517 | ||
2518 | if (gfp_mask & __GFP_NOFAIL) | |
2519 | goto force; | |
2520 | ||
2521 | if (fatal_signal_pending(current)) | |
2522 | goto force; | |
2523 | ||
2524 | /* | |
2525 | * keep retrying as long as the memcg oom killer is able to make | |
2526 | * a forward progress or bypass the charge if the oom killer | |
2527 | * couldn't make any progress. | |
2528 | */ | |
2529 | oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask, | |
2530 | get_order(nr_pages * PAGE_SIZE)); | |
2531 | switch (oom_status) { | |
2532 | case OOM_SUCCESS: | |
2533 | nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
2534 | goto retry; | |
2535 | case OOM_FAILED: | |
2536 | goto force; | |
2537 | default: | |
2538 | goto nomem; | |
2539 | } | |
2540 | nomem: | |
2541 | if (!(gfp_mask & __GFP_NOFAIL)) | |
2542 | return -ENOMEM; | |
2543 | force: | |
2544 | /* | |
2545 | * The allocation either can't fail or will lead to more memory | |
2546 | * being freed very soon. Allow memory usage go over the limit | |
2547 | * temporarily by force charging it. | |
2548 | */ | |
2549 | page_counter_charge(&memcg->memory, nr_pages); | |
2550 | if (do_memsw_account()) | |
2551 | page_counter_charge(&memcg->memsw, nr_pages); | |
2552 | css_get_many(&memcg->css, nr_pages); | |
2553 | ||
2554 | return 0; | |
2555 | ||
2556 | done_restock: | |
2557 | css_get_many(&memcg->css, batch); | |
2558 | if (batch > nr_pages) | |
2559 | refill_stock(memcg, batch - nr_pages); | |
2560 | ||
2561 | /* | |
2562 | * If the hierarchy is above the normal consumption range, schedule | |
2563 | * reclaim on returning to userland. We can perform reclaim here | |
2564 | * if __GFP_RECLAIM but let's always punt for simplicity and so that | |
2565 | * GFP_KERNEL can consistently be used during reclaim. @memcg is | |
2566 | * not recorded as it most likely matches current's and won't | |
2567 | * change in the meantime. As high limit is checked again before | |
2568 | * reclaim, the cost of mismatch is negligible. | |
2569 | */ | |
2570 | do { | |
2571 | if (page_counter_read(&memcg->memory) > memcg->high) { | |
2572 | /* Don't bother a random interrupted task */ | |
2573 | if (in_interrupt()) { | |
2574 | schedule_work(&memcg->high_work); | |
2575 | break; | |
2576 | } | |
2577 | current->memcg_nr_pages_over_high += batch; | |
2578 | set_notify_resume(current); | |
2579 | break; | |
2580 | } | |
2581 | } while ((memcg = parent_mem_cgroup(memcg))); | |
2582 | ||
2583 | return 0; | |
2584 | } | |
2585 | ||
2586 | static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages) | |
2587 | { | |
2588 | if (mem_cgroup_is_root(memcg)) | |
2589 | return; | |
2590 | ||
2591 | page_counter_uncharge(&memcg->memory, nr_pages); | |
2592 | if (do_memsw_account()) | |
2593 | page_counter_uncharge(&memcg->memsw, nr_pages); | |
2594 | ||
2595 | css_put_many(&memcg->css, nr_pages); | |
2596 | } | |
2597 | ||
2598 | static void lock_page_lru(struct page *page, int *isolated) | |
2599 | { | |
2600 | pg_data_t *pgdat = page_pgdat(page); | |
2601 | ||
2602 | spin_lock_irq(&pgdat->lru_lock); | |
2603 | if (PageLRU(page)) { | |
2604 | struct lruvec *lruvec; | |
2605 | ||
2606 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
2607 | ClearPageLRU(page); | |
2608 | del_page_from_lru_list(page, lruvec, page_lru(page)); | |
2609 | *isolated = 1; | |
2610 | } else | |
2611 | *isolated = 0; | |
2612 | } | |
2613 | ||
2614 | static void unlock_page_lru(struct page *page, int isolated) | |
2615 | { | |
2616 | pg_data_t *pgdat = page_pgdat(page); | |
2617 | ||
2618 | if (isolated) { | |
2619 | struct lruvec *lruvec; | |
2620 | ||
2621 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
2622 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
2623 | SetPageLRU(page); | |
2624 | add_page_to_lru_list(page, lruvec, page_lru(page)); | |
2625 | } | |
2626 | spin_unlock_irq(&pgdat->lru_lock); | |
2627 | } | |
2628 | ||
2629 | static void commit_charge(struct page *page, struct mem_cgroup *memcg, | |
2630 | bool lrucare) | |
2631 | { | |
2632 | int isolated; | |
2633 | ||
2634 | VM_BUG_ON_PAGE(page->mem_cgroup, page); | |
2635 | ||
2636 | /* | |
2637 | * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page | |
2638 | * may already be on some other mem_cgroup's LRU. Take care of it. | |
2639 | */ | |
2640 | if (lrucare) | |
2641 | lock_page_lru(page, &isolated); | |
2642 | ||
2643 | /* | |
2644 | * Nobody should be changing or seriously looking at | |
2645 | * page->mem_cgroup at this point: | |
2646 | * | |
2647 | * - the page is uncharged | |
2648 | * | |
2649 | * - the page is off-LRU | |
2650 | * | |
2651 | * - an anonymous fault has exclusive page access, except for | |
2652 | * a locked page table | |
2653 | * | |
2654 | * - a page cache insertion, a swapin fault, or a migration | |
2655 | * have the page locked | |
2656 | */ | |
2657 | page->mem_cgroup = memcg; | |
2658 | ||
2659 | if (lrucare) | |
2660 | unlock_page_lru(page, isolated); | |
2661 | } | |
2662 | ||
2663 | #ifdef CONFIG_MEMCG_KMEM | |
2664 | static int memcg_alloc_cache_id(void) | |
2665 | { | |
2666 | int id, size; | |
2667 | int err; | |
2668 | ||
2669 | id = ida_simple_get(&memcg_cache_ida, | |
2670 | 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL); | |
2671 | if (id < 0) | |
2672 | return id; | |
2673 | ||
2674 | if (id < memcg_nr_cache_ids) | |
2675 | return id; | |
2676 | ||
2677 | /* | |
2678 | * There's no space for the new id in memcg_caches arrays, | |
2679 | * so we have to grow them. | |
2680 | */ | |
2681 | down_write(&memcg_cache_ids_sem); | |
2682 | ||
2683 | size = 2 * (id + 1); | |
2684 | if (size < MEMCG_CACHES_MIN_SIZE) | |
2685 | size = MEMCG_CACHES_MIN_SIZE; | |
2686 | else if (size > MEMCG_CACHES_MAX_SIZE) | |
2687 | size = MEMCG_CACHES_MAX_SIZE; | |
2688 | ||
2689 | err = memcg_update_all_caches(size); | |
2690 | if (!err) | |
2691 | err = memcg_update_all_list_lrus(size); | |
2692 | if (!err) | |
2693 | memcg_nr_cache_ids = size; | |
2694 | ||
2695 | up_write(&memcg_cache_ids_sem); | |
2696 | ||
2697 | if (err) { | |
2698 | ida_simple_remove(&memcg_cache_ida, id); | |
2699 | return err; | |
2700 | } | |
2701 | return id; | |
2702 | } | |
2703 | ||
2704 | static void memcg_free_cache_id(int id) | |
2705 | { | |
2706 | ida_simple_remove(&memcg_cache_ida, id); | |
2707 | } | |
2708 | ||
2709 | struct memcg_kmem_cache_create_work { | |
2710 | struct mem_cgroup *memcg; | |
2711 | struct kmem_cache *cachep; | |
2712 | struct work_struct work; | |
2713 | }; | |
2714 | ||
2715 | static void memcg_kmem_cache_create_func(struct work_struct *w) | |
2716 | { | |
2717 | struct memcg_kmem_cache_create_work *cw = | |
2718 | container_of(w, struct memcg_kmem_cache_create_work, work); | |
2719 | struct mem_cgroup *memcg = cw->memcg; | |
2720 | struct kmem_cache *cachep = cw->cachep; | |
2721 | ||
2722 | memcg_create_kmem_cache(memcg, cachep); | |
2723 | ||
2724 | css_put(&memcg->css); | |
2725 | kfree(cw); | |
2726 | } | |
2727 | ||
2728 | /* | |
2729 | * Enqueue the creation of a per-memcg kmem_cache. | |
2730 | */ | |
2731 | static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg, | |
2732 | struct kmem_cache *cachep) | |
2733 | { | |
2734 | struct memcg_kmem_cache_create_work *cw; | |
2735 | ||
2736 | if (!css_tryget_online(&memcg->css)) | |
2737 | return; | |
2738 | ||
2739 | cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN); | |
2740 | if (!cw) | |
2741 | return; | |
2742 | ||
2743 | cw->memcg = memcg; | |
2744 | cw->cachep = cachep; | |
2745 | INIT_WORK(&cw->work, memcg_kmem_cache_create_func); | |
2746 | ||
2747 | queue_work(memcg_kmem_cache_wq, &cw->work); | |
2748 | } | |
2749 | ||
2750 | static inline bool memcg_kmem_bypass(void) | |
2751 | { | |
2752 | if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD)) | |
2753 | return true; | |
2754 | return false; | |
2755 | } | |
2756 | ||
2757 | /** | |
2758 | * memcg_kmem_get_cache: select the correct per-memcg cache for allocation | |
2759 | * @cachep: the original global kmem cache | |
2760 | * | |
2761 | * Return the kmem_cache we're supposed to use for a slab allocation. | |
2762 | * We try to use the current memcg's version of the cache. | |
2763 | * | |
2764 | * If the cache does not exist yet, if we are the first user of it, we | |
2765 | * create it asynchronously in a workqueue and let the current allocation | |
2766 | * go through with the original cache. | |
2767 | * | |
2768 | * This function takes a reference to the cache it returns to assure it | |
2769 | * won't get destroyed while we are working with it. Once the caller is | |
2770 | * done with it, memcg_kmem_put_cache() must be called to release the | |
2771 | * reference. | |
2772 | */ | |
2773 | struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep) | |
2774 | { | |
2775 | struct mem_cgroup *memcg; | |
2776 | struct kmem_cache *memcg_cachep; | |
2777 | struct memcg_cache_array *arr; | |
2778 | int kmemcg_id; | |
2779 | ||
2780 | VM_BUG_ON(!is_root_cache(cachep)); | |
2781 | ||
2782 | if (memcg_kmem_bypass()) | |
2783 | return cachep; | |
2784 | ||
2785 | rcu_read_lock(); | |
2786 | ||
2787 | if (unlikely(current->active_memcg)) | |
2788 | memcg = current->active_memcg; | |
2789 | else | |
2790 | memcg = mem_cgroup_from_task(current); | |
2791 | ||
2792 | if (!memcg || memcg == root_mem_cgroup) | |
2793 | goto out_unlock; | |
2794 | ||
2795 | kmemcg_id = READ_ONCE(memcg->kmemcg_id); | |
2796 | if (kmemcg_id < 0) | |
2797 | goto out_unlock; | |
2798 | ||
2799 | arr = rcu_dereference(cachep->memcg_params.memcg_caches); | |
2800 | ||
2801 | /* | |
2802 | * Make sure we will access the up-to-date value. The code updating | |
2803 | * memcg_caches issues a write barrier to match the data dependency | |
2804 | * barrier inside READ_ONCE() (see memcg_create_kmem_cache()). | |
2805 | */ | |
2806 | memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]); | |
2807 | ||
2808 | /* | |
2809 | * If we are in a safe context (can wait, and not in interrupt | |
2810 | * context), we could be be predictable and return right away. | |
2811 | * This would guarantee that the allocation being performed | |
2812 | * already belongs in the new cache. | |
2813 | * | |
2814 | * However, there are some clashes that can arrive from locking. | |
2815 | * For instance, because we acquire the slab_mutex while doing | |
2816 | * memcg_create_kmem_cache, this means no further allocation | |
2817 | * could happen with the slab_mutex held. So it's better to | |
2818 | * defer everything. | |
2819 | * | |
2820 | * If the memcg is dying or memcg_cache is about to be released, | |
2821 | * don't bother creating new kmem_caches. Because memcg_cachep | |
2822 | * is ZEROed as the fist step of kmem offlining, we don't need | |
2823 | * percpu_ref_tryget_live() here. css_tryget_online() check in | |
2824 | * memcg_schedule_kmem_cache_create() will prevent us from | |
2825 | * creation of a new kmem_cache. | |
2826 | */ | |
2827 | if (unlikely(!memcg_cachep)) | |
2828 | memcg_schedule_kmem_cache_create(memcg, cachep); | |
2829 | else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt)) | |
2830 | cachep = memcg_cachep; | |
2831 | out_unlock: | |
2832 | rcu_read_unlock(); | |
2833 | return cachep; | |
2834 | } | |
2835 | ||
2836 | /** | |
2837 | * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache | |
2838 | * @cachep: the cache returned by memcg_kmem_get_cache | |
2839 | */ | |
2840 | void memcg_kmem_put_cache(struct kmem_cache *cachep) | |
2841 | { | |
2842 | if (!is_root_cache(cachep)) | |
2843 | percpu_ref_put(&cachep->memcg_params.refcnt); | |
2844 | } | |
2845 | ||
2846 | /** | |
2847 | * __memcg_kmem_charge_memcg: charge a kmem page | |
2848 | * @page: page to charge | |
2849 | * @gfp: reclaim mode | |
2850 | * @order: allocation order | |
2851 | * @memcg: memory cgroup to charge | |
2852 | * | |
2853 | * Returns 0 on success, an error code on failure. | |
2854 | */ | |
2855 | int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order, | |
2856 | struct mem_cgroup *memcg) | |
2857 | { | |
2858 | unsigned int nr_pages = 1 << order; | |
2859 | struct page_counter *counter; | |
2860 | int ret; | |
2861 | ||
2862 | ret = try_charge(memcg, gfp, nr_pages); | |
2863 | if (ret) | |
2864 | return ret; | |
2865 | ||
2866 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && | |
2867 | !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) { | |
2868 | ||
2869 | /* | |
2870 | * Enforce __GFP_NOFAIL allocation because callers are not | |
2871 | * prepared to see failures and likely do not have any failure | |
2872 | * handling code. | |
2873 | */ | |
2874 | if (gfp & __GFP_NOFAIL) { | |
2875 | page_counter_charge(&memcg->kmem, nr_pages); | |
2876 | return 0; | |
2877 | } | |
2878 | cancel_charge(memcg, nr_pages); | |
2879 | return -ENOMEM; | |
2880 | } | |
2881 | return 0; | |
2882 | } | |
2883 | ||
2884 | /** | |
2885 | * __memcg_kmem_charge: charge a kmem page to the current memory cgroup | |
2886 | * @page: page to charge | |
2887 | * @gfp: reclaim mode | |
2888 | * @order: allocation order | |
2889 | * | |
2890 | * Returns 0 on success, an error code on failure. | |
2891 | */ | |
2892 | int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order) | |
2893 | { | |
2894 | struct mem_cgroup *memcg; | |
2895 | int ret = 0; | |
2896 | ||
2897 | if (memcg_kmem_bypass()) | |
2898 | return 0; | |
2899 | ||
2900 | memcg = get_mem_cgroup_from_current(); | |
2901 | if (!mem_cgroup_is_root(memcg)) { | |
2902 | ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg); | |
2903 | if (!ret) { | |
2904 | page->mem_cgroup = memcg; | |
2905 | __SetPageKmemcg(page); | |
2906 | } | |
2907 | } | |
2908 | css_put(&memcg->css); | |
2909 | return ret; | |
2910 | } | |
2911 | ||
2912 | /** | |
2913 | * __memcg_kmem_uncharge_memcg: uncharge a kmem page | |
2914 | * @memcg: memcg to uncharge | |
2915 | * @nr_pages: number of pages to uncharge | |
2916 | */ | |
2917 | void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg, | |
2918 | unsigned int nr_pages) | |
2919 | { | |
2920 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
2921 | page_counter_uncharge(&memcg->kmem, nr_pages); | |
2922 | ||
2923 | page_counter_uncharge(&memcg->memory, nr_pages); | |
2924 | if (do_memsw_account()) | |
2925 | page_counter_uncharge(&memcg->memsw, nr_pages); | |
2926 | } | |
2927 | /** | |
2928 | * __memcg_kmem_uncharge: uncharge a kmem page | |
2929 | * @page: page to uncharge | |
2930 | * @order: allocation order | |
2931 | */ | |
2932 | void __memcg_kmem_uncharge(struct page *page, int order) | |
2933 | { | |
2934 | struct mem_cgroup *memcg = page->mem_cgroup; | |
2935 | unsigned int nr_pages = 1 << order; | |
2936 | ||
2937 | if (!memcg) | |
2938 | return; | |
2939 | ||
2940 | VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page); | |
2941 | __memcg_kmem_uncharge_memcg(memcg, nr_pages); | |
2942 | page->mem_cgroup = NULL; | |
2943 | ||
2944 | /* slab pages do not have PageKmemcg flag set */ | |
2945 | if (PageKmemcg(page)) | |
2946 | __ClearPageKmemcg(page); | |
2947 | ||
2948 | css_put_many(&memcg->css, nr_pages); | |
2949 | } | |
2950 | #endif /* CONFIG_MEMCG_KMEM */ | |
2951 | ||
2952 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
2953 | ||
2954 | /* | |
2955 | * Because tail pages are not marked as "used", set it. We're under | |
2956 | * pgdat->lru_lock and migration entries setup in all page mappings. | |
2957 | */ | |
2958 | void mem_cgroup_split_huge_fixup(struct page *head) | |
2959 | { | |
2960 | int i; | |
2961 | ||
2962 | if (mem_cgroup_disabled()) | |
2963 | return; | |
2964 | ||
2965 | for (i = 1; i < HPAGE_PMD_NR; i++) | |
2966 | head[i].mem_cgroup = head->mem_cgroup; | |
2967 | ||
2968 | __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR); | |
2969 | } | |
2970 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
2971 | ||
2972 | #ifdef CONFIG_MEMCG_SWAP | |
2973 | /** | |
2974 | * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. | |
2975 | * @entry: swap entry to be moved | |
2976 | * @from: mem_cgroup which the entry is moved from | |
2977 | * @to: mem_cgroup which the entry is moved to | |
2978 | * | |
2979 | * It succeeds only when the swap_cgroup's record for this entry is the same | |
2980 | * as the mem_cgroup's id of @from. | |
2981 | * | |
2982 | * Returns 0 on success, -EINVAL on failure. | |
2983 | * | |
2984 | * The caller must have charged to @to, IOW, called page_counter_charge() about | |
2985 | * both res and memsw, and called css_get(). | |
2986 | */ | |
2987 | static int mem_cgroup_move_swap_account(swp_entry_t entry, | |
2988 | struct mem_cgroup *from, struct mem_cgroup *to) | |
2989 | { | |
2990 | unsigned short old_id, new_id; | |
2991 | ||
2992 | old_id = mem_cgroup_id(from); | |
2993 | new_id = mem_cgroup_id(to); | |
2994 | ||
2995 | if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { | |
2996 | mod_memcg_state(from, MEMCG_SWAP, -1); | |
2997 | mod_memcg_state(to, MEMCG_SWAP, 1); | |
2998 | return 0; | |
2999 | } | |
3000 | return -EINVAL; | |
3001 | } | |
3002 | #else | |
3003 | static inline int mem_cgroup_move_swap_account(swp_entry_t entry, | |
3004 | struct mem_cgroup *from, struct mem_cgroup *to) | |
3005 | { | |
3006 | return -EINVAL; | |
3007 | } | |
3008 | #endif | |
3009 | ||
3010 | static DEFINE_MUTEX(memcg_max_mutex); | |
3011 | ||
3012 | static int mem_cgroup_resize_max(struct mem_cgroup *memcg, | |
3013 | unsigned long max, bool memsw) | |
3014 | { | |
3015 | bool enlarge = false; | |
3016 | bool drained = false; | |
3017 | int ret; | |
3018 | bool limits_invariant; | |
3019 | struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory; | |
3020 | ||
3021 | do { | |
3022 | if (signal_pending(current)) { | |
3023 | ret = -EINTR; | |
3024 | break; | |
3025 | } | |
3026 | ||
3027 | mutex_lock(&memcg_max_mutex); | |
3028 | /* | |
3029 | * Make sure that the new limit (memsw or memory limit) doesn't | |
3030 | * break our basic invariant rule memory.max <= memsw.max. | |
3031 | */ | |
3032 | limits_invariant = memsw ? max >= memcg->memory.max : | |
3033 | max <= memcg->memsw.max; | |
3034 | if (!limits_invariant) { | |
3035 | mutex_unlock(&memcg_max_mutex); | |
3036 | ret = -EINVAL; | |
3037 | break; | |
3038 | } | |
3039 | if (max > counter->max) | |
3040 | enlarge = true; | |
3041 | ret = page_counter_set_max(counter, max); | |
3042 | mutex_unlock(&memcg_max_mutex); | |
3043 | ||
3044 | if (!ret) | |
3045 | break; | |
3046 | ||
3047 | if (!drained) { | |
3048 | drain_all_stock(memcg); | |
3049 | drained = true; | |
3050 | continue; | |
3051 | } | |
3052 | ||
3053 | if (!try_to_free_mem_cgroup_pages(memcg, 1, | |
3054 | GFP_KERNEL, !memsw)) { | |
3055 | ret = -EBUSY; | |
3056 | break; | |
3057 | } | |
3058 | } while (true); | |
3059 | ||
3060 | if (!ret && enlarge) | |
3061 | memcg_oom_recover(memcg); | |
3062 | ||
3063 | return ret; | |
3064 | } | |
3065 | ||
3066 | unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order, | |
3067 | gfp_t gfp_mask, | |
3068 | unsigned long *total_scanned) | |
3069 | { | |
3070 | unsigned long nr_reclaimed = 0; | |
3071 | struct mem_cgroup_per_node *mz, *next_mz = NULL; | |
3072 | unsigned long reclaimed; | |
3073 | int loop = 0; | |
3074 | struct mem_cgroup_tree_per_node *mctz; | |
3075 | unsigned long excess; | |
3076 | unsigned long nr_scanned; | |
3077 | ||
3078 | if (order > 0) | |
3079 | return 0; | |
3080 | ||
3081 | mctz = soft_limit_tree_node(pgdat->node_id); | |
3082 | ||
3083 | /* | |
3084 | * Do not even bother to check the largest node if the root | |
3085 | * is empty. Do it lockless to prevent lock bouncing. Races | |
3086 | * are acceptable as soft limit is best effort anyway. | |
3087 | */ | |
3088 | if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root)) | |
3089 | return 0; | |
3090 | ||
3091 | /* | |
3092 | * This loop can run a while, specially if mem_cgroup's continuously | |
3093 | * keep exceeding their soft limit and putting the system under | |
3094 | * pressure | |
3095 | */ | |
3096 | do { | |
3097 | if (next_mz) | |
3098 | mz = next_mz; | |
3099 | else | |
3100 | mz = mem_cgroup_largest_soft_limit_node(mctz); | |
3101 | if (!mz) | |
3102 | break; | |
3103 | ||
3104 | nr_scanned = 0; | |
3105 | reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat, | |
3106 | gfp_mask, &nr_scanned); | |
3107 | nr_reclaimed += reclaimed; | |
3108 | *total_scanned += nr_scanned; | |
3109 | spin_lock_irq(&mctz->lock); | |
3110 | __mem_cgroup_remove_exceeded(mz, mctz); | |
3111 | ||
3112 | /* | |
3113 | * If we failed to reclaim anything from this memory cgroup | |
3114 | * it is time to move on to the next cgroup | |
3115 | */ | |
3116 | next_mz = NULL; | |
3117 | if (!reclaimed) | |
3118 | next_mz = __mem_cgroup_largest_soft_limit_node(mctz); | |
3119 | ||
3120 | excess = soft_limit_excess(mz->memcg); | |
3121 | /* | |
3122 | * One school of thought says that we should not add | |
3123 | * back the node to the tree if reclaim returns 0. | |
3124 | * But our reclaim could return 0, simply because due | |
3125 | * to priority we are exposing a smaller subset of | |
3126 | * memory to reclaim from. Consider this as a longer | |
3127 | * term TODO. | |
3128 | */ | |
3129 | /* If excess == 0, no tree ops */ | |
3130 | __mem_cgroup_insert_exceeded(mz, mctz, excess); | |
3131 | spin_unlock_irq(&mctz->lock); | |
3132 | css_put(&mz->memcg->css); | |
3133 | loop++; | |
3134 | /* | |
3135 | * Could not reclaim anything and there are no more | |
3136 | * mem cgroups to try or we seem to be looping without | |
3137 | * reclaiming anything. | |
3138 | */ | |
3139 | if (!nr_reclaimed && | |
3140 | (next_mz == NULL || | |
3141 | loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) | |
3142 | break; | |
3143 | } while (!nr_reclaimed); | |
3144 | if (next_mz) | |
3145 | css_put(&next_mz->memcg->css); | |
3146 | return nr_reclaimed; | |
3147 | } | |
3148 | ||
3149 | /* | |
3150 | * Test whether @memcg has children, dead or alive. Note that this | |
3151 | * function doesn't care whether @memcg has use_hierarchy enabled and | |
3152 | * returns %true if there are child csses according to the cgroup | |
3153 | * hierarchy. Testing use_hierarchy is the caller's responsiblity. | |
3154 | */ | |
3155 | static inline bool memcg_has_children(struct mem_cgroup *memcg) | |
3156 | { | |
3157 | bool ret; | |
3158 | ||
3159 | rcu_read_lock(); | |
3160 | ret = css_next_child(NULL, &memcg->css); | |
3161 | rcu_read_unlock(); | |
3162 | return ret; | |
3163 | } | |
3164 | ||
3165 | /* | |
3166 | * Reclaims as many pages from the given memcg as possible. | |
3167 | * | |
3168 | * Caller is responsible for holding css reference for memcg. | |
3169 | */ | |
3170 | static int mem_cgroup_force_empty(struct mem_cgroup *memcg) | |
3171 | { | |
3172 | int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
3173 | ||
3174 | /* we call try-to-free pages for make this cgroup empty */ | |
3175 | lru_add_drain_all(); | |
3176 | ||
3177 | drain_all_stock(memcg); | |
3178 | ||
3179 | /* try to free all pages in this cgroup */ | |
3180 | while (nr_retries && page_counter_read(&memcg->memory)) { | |
3181 | int progress; | |
3182 | ||
3183 | if (signal_pending(current)) | |
3184 | return -EINTR; | |
3185 | ||
3186 | progress = try_to_free_mem_cgroup_pages(memcg, 1, | |
3187 | GFP_KERNEL, true); | |
3188 | if (!progress) { | |
3189 | nr_retries--; | |
3190 | /* maybe some writeback is necessary */ | |
3191 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
3192 | } | |
3193 | ||
3194 | } | |
3195 | ||
3196 | return 0; | |
3197 | } | |
3198 | ||
3199 | static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of, | |
3200 | char *buf, size_t nbytes, | |
3201 | loff_t off) | |
3202 | { | |
3203 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
3204 | ||
3205 | if (mem_cgroup_is_root(memcg)) | |
3206 | return -EINVAL; | |
3207 | return mem_cgroup_force_empty(memcg) ?: nbytes; | |
3208 | } | |
3209 | ||
3210 | static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css, | |
3211 | struct cftype *cft) | |
3212 | { | |
3213 | return mem_cgroup_from_css(css)->use_hierarchy; | |
3214 | } | |
3215 | ||
3216 | static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css, | |
3217 | struct cftype *cft, u64 val) | |
3218 | { | |
3219 | int retval = 0; | |
3220 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
3221 | struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent); | |
3222 | ||
3223 | if (memcg->use_hierarchy == val) | |
3224 | return 0; | |
3225 | ||
3226 | /* | |
3227 | * If parent's use_hierarchy is set, we can't make any modifications | |
3228 | * in the child subtrees. If it is unset, then the change can | |
3229 | * occur, provided the current cgroup has no children. | |
3230 | * | |
3231 | * For the root cgroup, parent_mem is NULL, we allow value to be | |
3232 | * set if there are no children. | |
3233 | */ | |
3234 | if ((!parent_memcg || !parent_memcg->use_hierarchy) && | |
3235 | (val == 1 || val == 0)) { | |
3236 | if (!memcg_has_children(memcg)) | |
3237 | memcg->use_hierarchy = val; | |
3238 | else | |
3239 | retval = -EBUSY; | |
3240 | } else | |
3241 | retval = -EINVAL; | |
3242 | ||
3243 | return retval; | |
3244 | } | |
3245 | ||
3246 | static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) | |
3247 | { | |
3248 | unsigned long val; | |
3249 | ||
3250 | if (mem_cgroup_is_root(memcg)) { | |
3251 | val = memcg_page_state(memcg, MEMCG_CACHE) + | |
3252 | memcg_page_state(memcg, MEMCG_RSS); | |
3253 | if (swap) | |
3254 | val += memcg_page_state(memcg, MEMCG_SWAP); | |
3255 | } else { | |
3256 | if (!swap) | |
3257 | val = page_counter_read(&memcg->memory); | |
3258 | else | |
3259 | val = page_counter_read(&memcg->memsw); | |
3260 | } | |
3261 | return val; | |
3262 | } | |
3263 | ||
3264 | enum { | |
3265 | RES_USAGE, | |
3266 | RES_LIMIT, | |
3267 | RES_MAX_USAGE, | |
3268 | RES_FAILCNT, | |
3269 | RES_SOFT_LIMIT, | |
3270 | }; | |
3271 | ||
3272 | static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css, | |
3273 | struct cftype *cft) | |
3274 | { | |
3275 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
3276 | struct page_counter *counter; | |
3277 | ||
3278 | switch (MEMFILE_TYPE(cft->private)) { | |
3279 | case _MEM: | |
3280 | counter = &memcg->memory; | |
3281 | break; | |
3282 | case _MEMSWAP: | |
3283 | counter = &memcg->memsw; | |
3284 | break; | |
3285 | case _KMEM: | |
3286 | counter = &memcg->kmem; | |
3287 | break; | |
3288 | case _TCP: | |
3289 | counter = &memcg->tcpmem; | |
3290 | break; | |
3291 | default: | |
3292 | BUG(); | |
3293 | } | |
3294 | ||
3295 | switch (MEMFILE_ATTR(cft->private)) { | |
3296 | case RES_USAGE: | |
3297 | if (counter == &memcg->memory) | |
3298 | return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE; | |
3299 | if (counter == &memcg->memsw) | |
3300 | return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE; | |
3301 | return (u64)page_counter_read(counter) * PAGE_SIZE; | |
3302 | case RES_LIMIT: | |
3303 | return (u64)counter->max * PAGE_SIZE; | |
3304 | case RES_MAX_USAGE: | |
3305 | return (u64)counter->watermark * PAGE_SIZE; | |
3306 | case RES_FAILCNT: | |
3307 | return counter->failcnt; | |
3308 | case RES_SOFT_LIMIT: | |
3309 | return (u64)memcg->soft_limit * PAGE_SIZE; | |
3310 | default: | |
3311 | BUG(); | |
3312 | } | |
3313 | } | |
3314 | ||
3315 | static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg) | |
3316 | { | |
3317 | unsigned long stat[MEMCG_NR_STAT] = {0}; | |
3318 | struct mem_cgroup *mi; | |
3319 | int node, cpu, i; | |
3320 | ||
3321 | for_each_online_cpu(cpu) | |
3322 | for (i = 0; i < MEMCG_NR_STAT; i++) | |
3323 | stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu); | |
3324 | ||
3325 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
3326 | for (i = 0; i < MEMCG_NR_STAT; i++) | |
3327 | atomic_long_add(stat[i], &mi->vmstats[i]); | |
3328 | ||
3329 | for_each_node(node) { | |
3330 | struct mem_cgroup_per_node *pn = memcg->nodeinfo[node]; | |
3331 | struct mem_cgroup_per_node *pi; | |
3332 | ||
3333 | for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) | |
3334 | stat[i] = 0; | |
3335 | ||
3336 | for_each_online_cpu(cpu) | |
3337 | for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) | |
3338 | stat[i] += per_cpu( | |
3339 | pn->lruvec_stat_cpu->count[i], cpu); | |
3340 | ||
3341 | for (pi = pn; pi; pi = parent_nodeinfo(pi, node)) | |
3342 | for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) | |
3343 | atomic_long_add(stat[i], &pi->lruvec_stat[i]); | |
3344 | } | |
3345 | } | |
3346 | ||
3347 | static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg) | |
3348 | { | |
3349 | unsigned long events[NR_VM_EVENT_ITEMS]; | |
3350 | struct mem_cgroup *mi; | |
3351 | int cpu, i; | |
3352 | ||
3353 | for (i = 0; i < NR_VM_EVENT_ITEMS; i++) | |
3354 | events[i] = 0; | |
3355 | ||
3356 | for_each_online_cpu(cpu) | |
3357 | for (i = 0; i < NR_VM_EVENT_ITEMS; i++) | |
3358 | events[i] += per_cpu(memcg->vmstats_percpu->events[i], | |
3359 | cpu); | |
3360 | ||
3361 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) | |
3362 | for (i = 0; i < NR_VM_EVENT_ITEMS; i++) | |
3363 | atomic_long_add(events[i], &mi->vmevents[i]); | |
3364 | } | |
3365 | ||
3366 | #ifdef CONFIG_MEMCG_KMEM | |
3367 | static int memcg_online_kmem(struct mem_cgroup *memcg) | |
3368 | { | |
3369 | int memcg_id; | |
3370 | ||
3371 | if (cgroup_memory_nokmem) | |
3372 | return 0; | |
3373 | ||
3374 | BUG_ON(memcg->kmemcg_id >= 0); | |
3375 | BUG_ON(memcg->kmem_state); | |
3376 | ||
3377 | memcg_id = memcg_alloc_cache_id(); | |
3378 | if (memcg_id < 0) | |
3379 | return memcg_id; | |
3380 | ||
3381 | static_branch_inc(&memcg_kmem_enabled_key); | |
3382 | /* | |
3383 | * A memory cgroup is considered kmem-online as soon as it gets | |
3384 | * kmemcg_id. Setting the id after enabling static branching will | |
3385 | * guarantee no one starts accounting before all call sites are | |
3386 | * patched. | |
3387 | */ | |
3388 | memcg->kmemcg_id = memcg_id; | |
3389 | memcg->kmem_state = KMEM_ONLINE; | |
3390 | INIT_LIST_HEAD(&memcg->kmem_caches); | |
3391 | ||
3392 | return 0; | |
3393 | } | |
3394 | ||
3395 | static void memcg_offline_kmem(struct mem_cgroup *memcg) | |
3396 | { | |
3397 | struct cgroup_subsys_state *css; | |
3398 | struct mem_cgroup *parent, *child; | |
3399 | int kmemcg_id; | |
3400 | ||
3401 | if (memcg->kmem_state != KMEM_ONLINE) | |
3402 | return; | |
3403 | /* | |
3404 | * Clear the online state before clearing memcg_caches array | |
3405 | * entries. The slab_mutex in memcg_deactivate_kmem_caches() | |
3406 | * guarantees that no cache will be created for this cgroup | |
3407 | * after we are done (see memcg_create_kmem_cache()). | |
3408 | */ | |
3409 | memcg->kmem_state = KMEM_ALLOCATED; | |
3410 | ||
3411 | parent = parent_mem_cgroup(memcg); | |
3412 | if (!parent) | |
3413 | parent = root_mem_cgroup; | |
3414 | ||
3415 | /* | |
3416 | * Deactivate and reparent kmem_caches. | |
3417 | */ | |
3418 | memcg_deactivate_kmem_caches(memcg, parent); | |
3419 | ||
3420 | kmemcg_id = memcg->kmemcg_id; | |
3421 | BUG_ON(kmemcg_id < 0); | |
3422 | ||
3423 | /* | |
3424 | * Change kmemcg_id of this cgroup and all its descendants to the | |
3425 | * parent's id, and then move all entries from this cgroup's list_lrus | |
3426 | * to ones of the parent. After we have finished, all list_lrus | |
3427 | * corresponding to this cgroup are guaranteed to remain empty. The | |
3428 | * ordering is imposed by list_lru_node->lock taken by | |
3429 | * memcg_drain_all_list_lrus(). | |
3430 | */ | |
3431 | rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */ | |
3432 | css_for_each_descendant_pre(css, &memcg->css) { | |
3433 | child = mem_cgroup_from_css(css); | |
3434 | BUG_ON(child->kmemcg_id != kmemcg_id); | |
3435 | child->kmemcg_id = parent->kmemcg_id; | |
3436 | if (!memcg->use_hierarchy) | |
3437 | break; | |
3438 | } | |
3439 | rcu_read_unlock(); | |
3440 | ||
3441 | memcg_drain_all_list_lrus(kmemcg_id, parent); | |
3442 | ||
3443 | memcg_free_cache_id(kmemcg_id); | |
3444 | } | |
3445 | ||
3446 | static void memcg_free_kmem(struct mem_cgroup *memcg) | |
3447 | { | |
3448 | /* css_alloc() failed, offlining didn't happen */ | |
3449 | if (unlikely(memcg->kmem_state == KMEM_ONLINE)) | |
3450 | memcg_offline_kmem(memcg); | |
3451 | ||
3452 | if (memcg->kmem_state == KMEM_ALLOCATED) { | |
3453 | WARN_ON(!list_empty(&memcg->kmem_caches)); | |
3454 | static_branch_dec(&memcg_kmem_enabled_key); | |
3455 | } | |
3456 | } | |
3457 | #else | |
3458 | static int memcg_online_kmem(struct mem_cgroup *memcg) | |
3459 | { | |
3460 | return 0; | |
3461 | } | |
3462 | static void memcg_offline_kmem(struct mem_cgroup *memcg) | |
3463 | { | |
3464 | } | |
3465 | static void memcg_free_kmem(struct mem_cgroup *memcg) | |
3466 | { | |
3467 | } | |
3468 | #endif /* CONFIG_MEMCG_KMEM */ | |
3469 | ||
3470 | static int memcg_update_kmem_max(struct mem_cgroup *memcg, | |
3471 | unsigned long max) | |
3472 | { | |
3473 | int ret; | |
3474 | ||
3475 | mutex_lock(&memcg_max_mutex); | |
3476 | ret = page_counter_set_max(&memcg->kmem, max); | |
3477 | mutex_unlock(&memcg_max_mutex); | |
3478 | return ret; | |
3479 | } | |
3480 | ||
3481 | static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max) | |
3482 | { | |
3483 | int ret; | |
3484 | ||
3485 | mutex_lock(&memcg_max_mutex); | |
3486 | ||
3487 | ret = page_counter_set_max(&memcg->tcpmem, max); | |
3488 | if (ret) | |
3489 | goto out; | |
3490 | ||
3491 | if (!memcg->tcpmem_active) { | |
3492 | /* | |
3493 | * The active flag needs to be written after the static_key | |
3494 | * update. This is what guarantees that the socket activation | |
3495 | * function is the last one to run. See mem_cgroup_sk_alloc() | |
3496 | * for details, and note that we don't mark any socket as | |
3497 | * belonging to this memcg until that flag is up. | |
3498 | * | |
3499 | * We need to do this, because static_keys will span multiple | |
3500 | * sites, but we can't control their order. If we mark a socket | |
3501 | * as accounted, but the accounting functions are not patched in | |
3502 | * yet, we'll lose accounting. | |
3503 | * | |
3504 | * We never race with the readers in mem_cgroup_sk_alloc(), | |
3505 | * because when this value change, the code to process it is not | |
3506 | * patched in yet. | |
3507 | */ | |
3508 | static_branch_inc(&memcg_sockets_enabled_key); | |
3509 | memcg->tcpmem_active = true; | |
3510 | } | |
3511 | out: | |
3512 | mutex_unlock(&memcg_max_mutex); | |
3513 | return ret; | |
3514 | } | |
3515 | ||
3516 | /* | |
3517 | * The user of this function is... | |
3518 | * RES_LIMIT. | |
3519 | */ | |
3520 | static ssize_t mem_cgroup_write(struct kernfs_open_file *of, | |
3521 | char *buf, size_t nbytes, loff_t off) | |
3522 | { | |
3523 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
3524 | unsigned long nr_pages; | |
3525 | int ret; | |
3526 | ||
3527 | buf = strstrip(buf); | |
3528 | ret = page_counter_memparse(buf, "-1", &nr_pages); | |
3529 | if (ret) | |
3530 | return ret; | |
3531 | ||
3532 | switch (MEMFILE_ATTR(of_cft(of)->private)) { | |
3533 | case RES_LIMIT: | |
3534 | if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ | |
3535 | ret = -EINVAL; | |
3536 | break; | |
3537 | } | |
3538 | switch (MEMFILE_TYPE(of_cft(of)->private)) { | |
3539 | case _MEM: | |
3540 | ret = mem_cgroup_resize_max(memcg, nr_pages, false); | |
3541 | break; | |
3542 | case _MEMSWAP: | |
3543 | ret = mem_cgroup_resize_max(memcg, nr_pages, true); | |
3544 | break; | |
3545 | case _KMEM: | |
3546 | pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. " | |
3547 | "Please report your usecase to linux-mm@kvack.org if you " | |
3548 | "depend on this functionality.\n"); | |
3549 | ret = memcg_update_kmem_max(memcg, nr_pages); | |
3550 | break; | |
3551 | case _TCP: | |
3552 | ret = memcg_update_tcp_max(memcg, nr_pages); | |
3553 | break; | |
3554 | } | |
3555 | break; | |
3556 | case RES_SOFT_LIMIT: | |
3557 | memcg->soft_limit = nr_pages; | |
3558 | ret = 0; | |
3559 | break; | |
3560 | } | |
3561 | return ret ?: nbytes; | |
3562 | } | |
3563 | ||
3564 | static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf, | |
3565 | size_t nbytes, loff_t off) | |
3566 | { | |
3567 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
3568 | struct page_counter *counter; | |
3569 | ||
3570 | switch (MEMFILE_TYPE(of_cft(of)->private)) { | |
3571 | case _MEM: | |
3572 | counter = &memcg->memory; | |
3573 | break; | |
3574 | case _MEMSWAP: | |
3575 | counter = &memcg->memsw; | |
3576 | break; | |
3577 | case _KMEM: | |
3578 | counter = &memcg->kmem; | |
3579 | break; | |
3580 | case _TCP: | |
3581 | counter = &memcg->tcpmem; | |
3582 | break; | |
3583 | default: | |
3584 | BUG(); | |
3585 | } | |
3586 | ||
3587 | switch (MEMFILE_ATTR(of_cft(of)->private)) { | |
3588 | case RES_MAX_USAGE: | |
3589 | page_counter_reset_watermark(counter); | |
3590 | break; | |
3591 | case RES_FAILCNT: | |
3592 | counter->failcnt = 0; | |
3593 | break; | |
3594 | default: | |
3595 | BUG(); | |
3596 | } | |
3597 | ||
3598 | return nbytes; | |
3599 | } | |
3600 | ||
3601 | static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css, | |
3602 | struct cftype *cft) | |
3603 | { | |
3604 | return mem_cgroup_from_css(css)->move_charge_at_immigrate; | |
3605 | } | |
3606 | ||
3607 | #ifdef CONFIG_MMU | |
3608 | static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css, | |
3609 | struct cftype *cft, u64 val) | |
3610 | { | |
3611 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
3612 | ||
3613 | if (val & ~MOVE_MASK) | |
3614 | return -EINVAL; | |
3615 | ||
3616 | /* | |
3617 | * No kind of locking is needed in here, because ->can_attach() will | |
3618 | * check this value once in the beginning of the process, and then carry | |
3619 | * on with stale data. This means that changes to this value will only | |
3620 | * affect task migrations starting after the change. | |
3621 | */ | |
3622 | memcg->move_charge_at_immigrate = val; | |
3623 | return 0; | |
3624 | } | |
3625 | #else | |
3626 | static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css, | |
3627 | struct cftype *cft, u64 val) | |
3628 | { | |
3629 | return -ENOSYS; | |
3630 | } | |
3631 | #endif | |
3632 | ||
3633 | #ifdef CONFIG_NUMA | |
3634 | ||
3635 | #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE)) | |
3636 | #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON)) | |
3637 | #define LRU_ALL ((1 << NR_LRU_LISTS) - 1) | |
3638 | ||
3639 | static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, | |
3640 | int nid, unsigned int lru_mask) | |
3641 | { | |
3642 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid)); | |
3643 | unsigned long nr = 0; | |
3644 | enum lru_list lru; | |
3645 | ||
3646 | VM_BUG_ON((unsigned)nid >= nr_node_ids); | |
3647 | ||
3648 | for_each_lru(lru) { | |
3649 | if (!(BIT(lru) & lru_mask)) | |
3650 | continue; | |
3651 | nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru); | |
3652 | } | |
3653 | return nr; | |
3654 | } | |
3655 | ||
3656 | static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, | |
3657 | unsigned int lru_mask) | |
3658 | { | |
3659 | unsigned long nr = 0; | |
3660 | enum lru_list lru; | |
3661 | ||
3662 | for_each_lru(lru) { | |
3663 | if (!(BIT(lru) & lru_mask)) | |
3664 | continue; | |
3665 | nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru); | |
3666 | } | |
3667 | return nr; | |
3668 | } | |
3669 | ||
3670 | static int memcg_numa_stat_show(struct seq_file *m, void *v) | |
3671 | { | |
3672 | struct numa_stat { | |
3673 | const char *name; | |
3674 | unsigned int lru_mask; | |
3675 | }; | |
3676 | ||
3677 | static const struct numa_stat stats[] = { | |
3678 | { "total", LRU_ALL }, | |
3679 | { "file", LRU_ALL_FILE }, | |
3680 | { "anon", LRU_ALL_ANON }, | |
3681 | { "unevictable", BIT(LRU_UNEVICTABLE) }, | |
3682 | }; | |
3683 | const struct numa_stat *stat; | |
3684 | int nid; | |
3685 | unsigned long nr; | |
3686 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
3687 | ||
3688 | for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) { | |
3689 | nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask); | |
3690 | seq_printf(m, "%s=%lu", stat->name, nr); | |
3691 | for_each_node_state(nid, N_MEMORY) { | |
3692 | nr = mem_cgroup_node_nr_lru_pages(memcg, nid, | |
3693 | stat->lru_mask); | |
3694 | seq_printf(m, " N%d=%lu", nid, nr); | |
3695 | } | |
3696 | seq_putc(m, '\n'); | |
3697 | } | |
3698 | ||
3699 | for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) { | |
3700 | struct mem_cgroup *iter; | |
3701 | ||
3702 | nr = 0; | |
3703 | for_each_mem_cgroup_tree(iter, memcg) | |
3704 | nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask); | |
3705 | seq_printf(m, "hierarchical_%s=%lu", stat->name, nr); | |
3706 | for_each_node_state(nid, N_MEMORY) { | |
3707 | nr = 0; | |
3708 | for_each_mem_cgroup_tree(iter, memcg) | |
3709 | nr += mem_cgroup_node_nr_lru_pages( | |
3710 | iter, nid, stat->lru_mask); | |
3711 | seq_printf(m, " N%d=%lu", nid, nr); | |
3712 | } | |
3713 | seq_putc(m, '\n'); | |
3714 | } | |
3715 | ||
3716 | return 0; | |
3717 | } | |
3718 | #endif /* CONFIG_NUMA */ | |
3719 | ||
3720 | static const unsigned int memcg1_stats[] = { | |
3721 | MEMCG_CACHE, | |
3722 | MEMCG_RSS, | |
3723 | MEMCG_RSS_HUGE, | |
3724 | NR_SHMEM, | |
3725 | NR_FILE_MAPPED, | |
3726 | NR_FILE_DIRTY, | |
3727 | NR_WRITEBACK, | |
3728 | MEMCG_SWAP, | |
3729 | }; | |
3730 | ||
3731 | static const char *const memcg1_stat_names[] = { | |
3732 | "cache", | |
3733 | "rss", | |
3734 | "rss_huge", | |
3735 | "shmem", | |
3736 | "mapped_file", | |
3737 | "dirty", | |
3738 | "writeback", | |
3739 | "swap", | |
3740 | }; | |
3741 | ||
3742 | /* Universal VM events cgroup1 shows, original sort order */ | |
3743 | static const unsigned int memcg1_events[] = { | |
3744 | PGPGIN, | |
3745 | PGPGOUT, | |
3746 | PGFAULT, | |
3747 | PGMAJFAULT, | |
3748 | }; | |
3749 | ||
3750 | static int memcg_stat_show(struct seq_file *m, void *v) | |
3751 | { | |
3752 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
3753 | unsigned long memory, memsw; | |
3754 | struct mem_cgroup *mi; | |
3755 | unsigned int i; | |
3756 | ||
3757 | BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats)); | |
3758 | ||
3759 | for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) { | |
3760 | if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account()) | |
3761 | continue; | |
3762 | seq_printf(m, "%s %lu\n", memcg1_stat_names[i], | |
3763 | memcg_page_state_local(memcg, memcg1_stats[i]) * | |
3764 | PAGE_SIZE); | |
3765 | } | |
3766 | ||
3767 | for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) | |
3768 | seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]), | |
3769 | memcg_events_local(memcg, memcg1_events[i])); | |
3770 | ||
3771 | for (i = 0; i < NR_LRU_LISTS; i++) | |
3772 | seq_printf(m, "%s %lu\n", lru_list_name(i), | |
3773 | memcg_page_state_local(memcg, NR_LRU_BASE + i) * | |
3774 | PAGE_SIZE); | |
3775 | ||
3776 | /* Hierarchical information */ | |
3777 | memory = memsw = PAGE_COUNTER_MAX; | |
3778 | for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) { | |
3779 | memory = min(memory, mi->memory.max); | |
3780 | memsw = min(memsw, mi->memsw.max); | |
3781 | } | |
3782 | seq_printf(m, "hierarchical_memory_limit %llu\n", | |
3783 | (u64)memory * PAGE_SIZE); | |
3784 | if (do_memsw_account()) | |
3785 | seq_printf(m, "hierarchical_memsw_limit %llu\n", | |
3786 | (u64)memsw * PAGE_SIZE); | |
3787 | ||
3788 | for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) { | |
3789 | if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account()) | |
3790 | continue; | |
3791 | seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i], | |
3792 | (u64)memcg_page_state(memcg, memcg1_stats[i]) * | |
3793 | PAGE_SIZE); | |
3794 | } | |
3795 | ||
3796 | for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) | |
3797 | seq_printf(m, "total_%s %llu\n", | |
3798 | vm_event_name(memcg1_events[i]), | |
3799 | (u64)memcg_events(memcg, memcg1_events[i])); | |
3800 | ||
3801 | for (i = 0; i < NR_LRU_LISTS; i++) | |
3802 | seq_printf(m, "total_%s %llu\n", lru_list_name(i), | |
3803 | (u64)memcg_page_state(memcg, NR_LRU_BASE + i) * | |
3804 | PAGE_SIZE); | |
3805 | ||
3806 | #ifdef CONFIG_DEBUG_VM | |
3807 | { | |
3808 | pg_data_t *pgdat; | |
3809 | struct mem_cgroup_per_node *mz; | |
3810 | struct zone_reclaim_stat *rstat; | |
3811 | unsigned long recent_rotated[2] = {0, 0}; | |
3812 | unsigned long recent_scanned[2] = {0, 0}; | |
3813 | ||
3814 | for_each_online_pgdat(pgdat) { | |
3815 | mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id); | |
3816 | rstat = &mz->lruvec.reclaim_stat; | |
3817 | ||
3818 | recent_rotated[0] += rstat->recent_rotated[0]; | |
3819 | recent_rotated[1] += rstat->recent_rotated[1]; | |
3820 | recent_scanned[0] += rstat->recent_scanned[0]; | |
3821 | recent_scanned[1] += rstat->recent_scanned[1]; | |
3822 | } | |
3823 | seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]); | |
3824 | seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]); | |
3825 | seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]); | |
3826 | seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]); | |
3827 | } | |
3828 | #endif | |
3829 | ||
3830 | return 0; | |
3831 | } | |
3832 | ||
3833 | static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css, | |
3834 | struct cftype *cft) | |
3835 | { | |
3836 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
3837 | ||
3838 | return mem_cgroup_swappiness(memcg); | |
3839 | } | |
3840 | ||
3841 | static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css, | |
3842 | struct cftype *cft, u64 val) | |
3843 | { | |
3844 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
3845 | ||
3846 | if (val > 100) | |
3847 | return -EINVAL; | |
3848 | ||
3849 | if (css->parent) | |
3850 | memcg->swappiness = val; | |
3851 | else | |
3852 | vm_swappiness = val; | |
3853 | ||
3854 | return 0; | |
3855 | } | |
3856 | ||
3857 | static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) | |
3858 | { | |
3859 | struct mem_cgroup_threshold_ary *t; | |
3860 | unsigned long usage; | |
3861 | int i; | |
3862 | ||
3863 | rcu_read_lock(); | |
3864 | if (!swap) | |
3865 | t = rcu_dereference(memcg->thresholds.primary); | |
3866 | else | |
3867 | t = rcu_dereference(memcg->memsw_thresholds.primary); | |
3868 | ||
3869 | if (!t) | |
3870 | goto unlock; | |
3871 | ||
3872 | usage = mem_cgroup_usage(memcg, swap); | |
3873 | ||
3874 | /* | |
3875 | * current_threshold points to threshold just below or equal to usage. | |
3876 | * If it's not true, a threshold was crossed after last | |
3877 | * call of __mem_cgroup_threshold(). | |
3878 | */ | |
3879 | i = t->current_threshold; | |
3880 | ||
3881 | /* | |
3882 | * Iterate backward over array of thresholds starting from | |
3883 | * current_threshold and check if a threshold is crossed. | |
3884 | * If none of thresholds below usage is crossed, we read | |
3885 | * only one element of the array here. | |
3886 | */ | |
3887 | for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) | |
3888 | eventfd_signal(t->entries[i].eventfd, 1); | |
3889 | ||
3890 | /* i = current_threshold + 1 */ | |
3891 | i++; | |
3892 | ||
3893 | /* | |
3894 | * Iterate forward over array of thresholds starting from | |
3895 | * current_threshold+1 and check if a threshold is crossed. | |
3896 | * If none of thresholds above usage is crossed, we read | |
3897 | * only one element of the array here. | |
3898 | */ | |
3899 | for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) | |
3900 | eventfd_signal(t->entries[i].eventfd, 1); | |
3901 | ||
3902 | /* Update current_threshold */ | |
3903 | t->current_threshold = i - 1; | |
3904 | unlock: | |
3905 | rcu_read_unlock(); | |
3906 | } | |
3907 | ||
3908 | static void mem_cgroup_threshold(struct mem_cgroup *memcg) | |
3909 | { | |
3910 | while (memcg) { | |
3911 | __mem_cgroup_threshold(memcg, false); | |
3912 | if (do_memsw_account()) | |
3913 | __mem_cgroup_threshold(memcg, true); | |
3914 | ||
3915 | memcg = parent_mem_cgroup(memcg); | |
3916 | } | |
3917 | } | |
3918 | ||
3919 | static int compare_thresholds(const void *a, const void *b) | |
3920 | { | |
3921 | const struct mem_cgroup_threshold *_a = a; | |
3922 | const struct mem_cgroup_threshold *_b = b; | |
3923 | ||
3924 | if (_a->threshold > _b->threshold) | |
3925 | return 1; | |
3926 | ||
3927 | if (_a->threshold < _b->threshold) | |
3928 | return -1; | |
3929 | ||
3930 | return 0; | |
3931 | } | |
3932 | ||
3933 | static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) | |
3934 | { | |
3935 | struct mem_cgroup_eventfd_list *ev; | |
3936 | ||
3937 | spin_lock(&memcg_oom_lock); | |
3938 | ||
3939 | list_for_each_entry(ev, &memcg->oom_notify, list) | |
3940 | eventfd_signal(ev->eventfd, 1); | |
3941 | ||
3942 | spin_unlock(&memcg_oom_lock); | |
3943 | return 0; | |
3944 | } | |
3945 | ||
3946 | static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) | |
3947 | { | |
3948 | struct mem_cgroup *iter; | |
3949 | ||
3950 | for_each_mem_cgroup_tree(iter, memcg) | |
3951 | mem_cgroup_oom_notify_cb(iter); | |
3952 | } | |
3953 | ||
3954 | static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
3955 | struct eventfd_ctx *eventfd, const char *args, enum res_type type) | |
3956 | { | |
3957 | struct mem_cgroup_thresholds *thresholds; | |
3958 | struct mem_cgroup_threshold_ary *new; | |
3959 | unsigned long threshold; | |
3960 | unsigned long usage; | |
3961 | int i, size, ret; | |
3962 | ||
3963 | ret = page_counter_memparse(args, "-1", &threshold); | |
3964 | if (ret) | |
3965 | return ret; | |
3966 | ||
3967 | mutex_lock(&memcg->thresholds_lock); | |
3968 | ||
3969 | if (type == _MEM) { | |
3970 | thresholds = &memcg->thresholds; | |
3971 | usage = mem_cgroup_usage(memcg, false); | |
3972 | } else if (type == _MEMSWAP) { | |
3973 | thresholds = &memcg->memsw_thresholds; | |
3974 | usage = mem_cgroup_usage(memcg, true); | |
3975 | } else | |
3976 | BUG(); | |
3977 | ||
3978 | /* Check if a threshold crossed before adding a new one */ | |
3979 | if (thresholds->primary) | |
3980 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
3981 | ||
3982 | size = thresholds->primary ? thresholds->primary->size + 1 : 1; | |
3983 | ||
3984 | /* Allocate memory for new array of thresholds */ | |
3985 | new = kmalloc(struct_size(new, entries, size), GFP_KERNEL); | |
3986 | if (!new) { | |
3987 | ret = -ENOMEM; | |
3988 | goto unlock; | |
3989 | } | |
3990 | new->size = size; | |
3991 | ||
3992 | /* Copy thresholds (if any) to new array */ | |
3993 | if (thresholds->primary) { | |
3994 | memcpy(new->entries, thresholds->primary->entries, (size - 1) * | |
3995 | sizeof(struct mem_cgroup_threshold)); | |
3996 | } | |
3997 | ||
3998 | /* Add new threshold */ | |
3999 | new->entries[size - 1].eventfd = eventfd; | |
4000 | new->entries[size - 1].threshold = threshold; | |
4001 | ||
4002 | /* Sort thresholds. Registering of new threshold isn't time-critical */ | |
4003 | sort(new->entries, size, sizeof(struct mem_cgroup_threshold), | |
4004 | compare_thresholds, NULL); | |
4005 | ||
4006 | /* Find current threshold */ | |
4007 | new->current_threshold = -1; | |
4008 | for (i = 0; i < size; i++) { | |
4009 | if (new->entries[i].threshold <= usage) { | |
4010 | /* | |
4011 | * new->current_threshold will not be used until | |
4012 | * rcu_assign_pointer(), so it's safe to increment | |
4013 | * it here. | |
4014 | */ | |
4015 | ++new->current_threshold; | |
4016 | } else | |
4017 | break; | |
4018 | } | |
4019 | ||
4020 | /* Free old spare buffer and save old primary buffer as spare */ | |
4021 | kfree(thresholds->spare); | |
4022 | thresholds->spare = thresholds->primary; | |
4023 | ||
4024 | rcu_assign_pointer(thresholds->primary, new); | |
4025 | ||
4026 | /* To be sure that nobody uses thresholds */ | |
4027 | synchronize_rcu(); | |
4028 | ||
4029 | unlock: | |
4030 | mutex_unlock(&memcg->thresholds_lock); | |
4031 | ||
4032 | return ret; | |
4033 | } | |
4034 | ||
4035 | static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
4036 | struct eventfd_ctx *eventfd, const char *args) | |
4037 | { | |
4038 | return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM); | |
4039 | } | |
4040 | ||
4041 | static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg, | |
4042 | struct eventfd_ctx *eventfd, const char *args) | |
4043 | { | |
4044 | return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP); | |
4045 | } | |
4046 | ||
4047 | static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
4048 | struct eventfd_ctx *eventfd, enum res_type type) | |
4049 | { | |
4050 | struct mem_cgroup_thresholds *thresholds; | |
4051 | struct mem_cgroup_threshold_ary *new; | |
4052 | unsigned long usage; | |
4053 | int i, j, size, entries; | |
4054 | ||
4055 | mutex_lock(&memcg->thresholds_lock); | |
4056 | ||
4057 | if (type == _MEM) { | |
4058 | thresholds = &memcg->thresholds; | |
4059 | usage = mem_cgroup_usage(memcg, false); | |
4060 | } else if (type == _MEMSWAP) { | |
4061 | thresholds = &memcg->memsw_thresholds; | |
4062 | usage = mem_cgroup_usage(memcg, true); | |
4063 | } else | |
4064 | BUG(); | |
4065 | ||
4066 | if (!thresholds->primary) | |
4067 | goto unlock; | |
4068 | ||
4069 | /* Check if a threshold crossed before removing */ | |
4070 | __mem_cgroup_threshold(memcg, type == _MEMSWAP); | |
4071 | ||
4072 | /* Calculate new number of threshold */ | |
4073 | size = entries = 0; | |
4074 | for (i = 0; i < thresholds->primary->size; i++) { | |
4075 | if (thresholds->primary->entries[i].eventfd != eventfd) | |
4076 | size++; | |
4077 | else | |
4078 | entries++; | |
4079 | } | |
4080 | ||
4081 | new = thresholds->spare; | |
4082 | ||
4083 | /* If no items related to eventfd have been cleared, nothing to do */ | |
4084 | if (!entries) | |
4085 | goto unlock; | |
4086 | ||
4087 | /* Set thresholds array to NULL if we don't have thresholds */ | |
4088 | if (!size) { | |
4089 | kfree(new); | |
4090 | new = NULL; | |
4091 | goto swap_buffers; | |
4092 | } | |
4093 | ||
4094 | new->size = size; | |
4095 | ||
4096 | /* Copy thresholds and find current threshold */ | |
4097 | new->current_threshold = -1; | |
4098 | for (i = 0, j = 0; i < thresholds->primary->size; i++) { | |
4099 | if (thresholds->primary->entries[i].eventfd == eventfd) | |
4100 | continue; | |
4101 | ||
4102 | new->entries[j] = thresholds->primary->entries[i]; | |
4103 | if (new->entries[j].threshold <= usage) { | |
4104 | /* | |
4105 | * new->current_threshold will not be used | |
4106 | * until rcu_assign_pointer(), so it's safe to increment | |
4107 | * it here. | |
4108 | */ | |
4109 | ++new->current_threshold; | |
4110 | } | |
4111 | j++; | |
4112 | } | |
4113 | ||
4114 | swap_buffers: | |
4115 | /* Swap primary and spare array */ | |
4116 | thresholds->spare = thresholds->primary; | |
4117 | ||
4118 | rcu_assign_pointer(thresholds->primary, new); | |
4119 | ||
4120 | /* To be sure that nobody uses thresholds */ | |
4121 | synchronize_rcu(); | |
4122 | ||
4123 | /* If all events are unregistered, free the spare array */ | |
4124 | if (!new) { | |
4125 | kfree(thresholds->spare); | |
4126 | thresholds->spare = NULL; | |
4127 | } | |
4128 | unlock: | |
4129 | mutex_unlock(&memcg->thresholds_lock); | |
4130 | } | |
4131 | ||
4132 | static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
4133 | struct eventfd_ctx *eventfd) | |
4134 | { | |
4135 | return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM); | |
4136 | } | |
4137 | ||
4138 | static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg, | |
4139 | struct eventfd_ctx *eventfd) | |
4140 | { | |
4141 | return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP); | |
4142 | } | |
4143 | ||
4144 | static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg, | |
4145 | struct eventfd_ctx *eventfd, const char *args) | |
4146 | { | |
4147 | struct mem_cgroup_eventfd_list *event; | |
4148 | ||
4149 | event = kmalloc(sizeof(*event), GFP_KERNEL); | |
4150 | if (!event) | |
4151 | return -ENOMEM; | |
4152 | ||
4153 | spin_lock(&memcg_oom_lock); | |
4154 | ||
4155 | event->eventfd = eventfd; | |
4156 | list_add(&event->list, &memcg->oom_notify); | |
4157 | ||
4158 | /* already in OOM ? */ | |
4159 | if (memcg->under_oom) | |
4160 | eventfd_signal(eventfd, 1); | |
4161 | spin_unlock(&memcg_oom_lock); | |
4162 | ||
4163 | return 0; | |
4164 | } | |
4165 | ||
4166 | static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg, | |
4167 | struct eventfd_ctx *eventfd) | |
4168 | { | |
4169 | struct mem_cgroup_eventfd_list *ev, *tmp; | |
4170 | ||
4171 | spin_lock(&memcg_oom_lock); | |
4172 | ||
4173 | list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { | |
4174 | if (ev->eventfd == eventfd) { | |
4175 | list_del(&ev->list); | |
4176 | kfree(ev); | |
4177 | } | |
4178 | } | |
4179 | ||
4180 | spin_unlock(&memcg_oom_lock); | |
4181 | } | |
4182 | ||
4183 | static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v) | |
4184 | { | |
4185 | struct mem_cgroup *memcg = mem_cgroup_from_seq(sf); | |
4186 | ||
4187 | seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable); | |
4188 | seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom); | |
4189 | seq_printf(sf, "oom_kill %lu\n", | |
4190 | atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL])); | |
4191 | return 0; | |
4192 | } | |
4193 | ||
4194 | static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css, | |
4195 | struct cftype *cft, u64 val) | |
4196 | { | |
4197 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
4198 | ||
4199 | /* cannot set to root cgroup and only 0 and 1 are allowed */ | |
4200 | if (!css->parent || !((val == 0) || (val == 1))) | |
4201 | return -EINVAL; | |
4202 | ||
4203 | memcg->oom_kill_disable = val; | |
4204 | if (!val) | |
4205 | memcg_oom_recover(memcg); | |
4206 | ||
4207 | return 0; | |
4208 | } | |
4209 | ||
4210 | #ifdef CONFIG_CGROUP_WRITEBACK | |
4211 | ||
4212 | #include <trace/events/writeback.h> | |
4213 | ||
4214 | static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp) | |
4215 | { | |
4216 | return wb_domain_init(&memcg->cgwb_domain, gfp); | |
4217 | } | |
4218 | ||
4219 | static void memcg_wb_domain_exit(struct mem_cgroup *memcg) | |
4220 | { | |
4221 | wb_domain_exit(&memcg->cgwb_domain); | |
4222 | } | |
4223 | ||
4224 | static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg) | |
4225 | { | |
4226 | wb_domain_size_changed(&memcg->cgwb_domain); | |
4227 | } | |
4228 | ||
4229 | struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb) | |
4230 | { | |
4231 | struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css); | |
4232 | ||
4233 | if (!memcg->css.parent) | |
4234 | return NULL; | |
4235 | ||
4236 | return &memcg->cgwb_domain; | |
4237 | } | |
4238 | ||
4239 | /* | |
4240 | * idx can be of type enum memcg_stat_item or node_stat_item. | |
4241 | * Keep in sync with memcg_exact_page(). | |
4242 | */ | |
4243 | static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx) | |
4244 | { | |
4245 | long x = atomic_long_read(&memcg->vmstats[idx]); | |
4246 | int cpu; | |
4247 | ||
4248 | for_each_online_cpu(cpu) | |
4249 | x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx]; | |
4250 | if (x < 0) | |
4251 | x = 0; | |
4252 | return x; | |
4253 | } | |
4254 | ||
4255 | /** | |
4256 | * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg | |
4257 | * @wb: bdi_writeback in question | |
4258 | * @pfilepages: out parameter for number of file pages | |
4259 | * @pheadroom: out parameter for number of allocatable pages according to memcg | |
4260 | * @pdirty: out parameter for number of dirty pages | |
4261 | * @pwriteback: out parameter for number of pages under writeback | |
4262 | * | |
4263 | * Determine the numbers of file, headroom, dirty, and writeback pages in | |
4264 | * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom | |
4265 | * is a bit more involved. | |
4266 | * | |
4267 | * A memcg's headroom is "min(max, high) - used". In the hierarchy, the | |
4268 | * headroom is calculated as the lowest headroom of itself and the | |
4269 | * ancestors. Note that this doesn't consider the actual amount of | |
4270 | * available memory in the system. The caller should further cap | |
4271 | * *@pheadroom accordingly. | |
4272 | */ | |
4273 | void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages, | |
4274 | unsigned long *pheadroom, unsigned long *pdirty, | |
4275 | unsigned long *pwriteback) | |
4276 | { | |
4277 | struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css); | |
4278 | struct mem_cgroup *parent; | |
4279 | ||
4280 | *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY); | |
4281 | ||
4282 | /* this should eventually include NR_UNSTABLE_NFS */ | |
4283 | *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK); | |
4284 | *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) + | |
4285 | memcg_exact_page_state(memcg, NR_ACTIVE_FILE); | |
4286 | *pheadroom = PAGE_COUNTER_MAX; | |
4287 | ||
4288 | while ((parent = parent_mem_cgroup(memcg))) { | |
4289 | unsigned long ceiling = min(memcg->memory.max, memcg->high); | |
4290 | unsigned long used = page_counter_read(&memcg->memory); | |
4291 | ||
4292 | *pheadroom = min(*pheadroom, ceiling - min(ceiling, used)); | |
4293 | memcg = parent; | |
4294 | } | |
4295 | } | |
4296 | ||
4297 | /* | |
4298 | * Foreign dirty flushing | |
4299 | * | |
4300 | * There's an inherent mismatch between memcg and writeback. The former | |
4301 | * trackes ownership per-page while the latter per-inode. This was a | |
4302 | * deliberate design decision because honoring per-page ownership in the | |
4303 | * writeback path is complicated, may lead to higher CPU and IO overheads | |
4304 | * and deemed unnecessary given that write-sharing an inode across | |
4305 | * different cgroups isn't a common use-case. | |
4306 | * | |
4307 | * Combined with inode majority-writer ownership switching, this works well | |
4308 | * enough in most cases but there are some pathological cases. For | |
4309 | * example, let's say there are two cgroups A and B which keep writing to | |
4310 | * different but confined parts of the same inode. B owns the inode and | |
4311 | * A's memory is limited far below B's. A's dirty ratio can rise enough to | |
4312 | * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid | |
4313 | * triggering background writeback. A will be slowed down without a way to | |
4314 | * make writeback of the dirty pages happen. | |
4315 | * | |
4316 | * Conditions like the above can lead to a cgroup getting repatedly and | |
4317 | * severely throttled after making some progress after each | |
4318 | * dirty_expire_interval while the underyling IO device is almost | |
4319 | * completely idle. | |
4320 | * | |
4321 | * Solving this problem completely requires matching the ownership tracking | |
4322 | * granularities between memcg and writeback in either direction. However, | |
4323 | * the more egregious behaviors can be avoided by simply remembering the | |
4324 | * most recent foreign dirtying events and initiating remote flushes on | |
4325 | * them when local writeback isn't enough to keep the memory clean enough. | |
4326 | * | |
4327 | * The following two functions implement such mechanism. When a foreign | |
4328 | * page - a page whose memcg and writeback ownerships don't match - is | |
4329 | * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning | |
4330 | * bdi_writeback on the page owning memcg. When balance_dirty_pages() | |
4331 | * decides that the memcg needs to sleep due to high dirty ratio, it calls | |
4332 | * mem_cgroup_flush_foreign() which queues writeback on the recorded | |
4333 | * foreign bdi_writebacks which haven't expired. Both the numbers of | |
4334 | * recorded bdi_writebacks and concurrent in-flight foreign writebacks are | |
4335 | * limited to MEMCG_CGWB_FRN_CNT. | |
4336 | * | |
4337 | * The mechanism only remembers IDs and doesn't hold any object references. | |
4338 | * As being wrong occasionally doesn't matter, updates and accesses to the | |
4339 | * records are lockless and racy. | |
4340 | */ | |
4341 | void mem_cgroup_track_foreign_dirty_slowpath(struct page *page, | |
4342 | struct bdi_writeback *wb) | |
4343 | { | |
4344 | struct mem_cgroup *memcg = page->mem_cgroup; | |
4345 | struct memcg_cgwb_frn *frn; | |
4346 | u64 now = get_jiffies_64(); | |
4347 | u64 oldest_at = now; | |
4348 | int oldest = -1; | |
4349 | int i; | |
4350 | ||
4351 | trace_track_foreign_dirty(page, wb); | |
4352 | ||
4353 | /* | |
4354 | * Pick the slot to use. If there is already a slot for @wb, keep | |
4355 | * using it. If not replace the oldest one which isn't being | |
4356 | * written out. | |
4357 | */ | |
4358 | for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) { | |
4359 | frn = &memcg->cgwb_frn[i]; | |
4360 | if (frn->bdi_id == wb->bdi->id && | |
4361 | frn->memcg_id == wb->memcg_css->id) | |
4362 | break; | |
4363 | if (time_before64(frn->at, oldest_at) && | |
4364 | atomic_read(&frn->done.cnt) == 1) { | |
4365 | oldest = i; | |
4366 | oldest_at = frn->at; | |
4367 | } | |
4368 | } | |
4369 | ||
4370 | if (i < MEMCG_CGWB_FRN_CNT) { | |
4371 | /* | |
4372 | * Re-using an existing one. Update timestamp lazily to | |
4373 | * avoid making the cacheline hot. We want them to be | |
4374 | * reasonably up-to-date and significantly shorter than | |
4375 | * dirty_expire_interval as that's what expires the record. | |
4376 | * Use the shorter of 1s and dirty_expire_interval / 8. | |
4377 | */ | |
4378 | unsigned long update_intv = | |
4379 | min_t(unsigned long, HZ, | |
4380 | msecs_to_jiffies(dirty_expire_interval * 10) / 8); | |
4381 | ||
4382 | if (time_before64(frn->at, now - update_intv)) | |
4383 | frn->at = now; | |
4384 | } else if (oldest >= 0) { | |
4385 | /* replace the oldest free one */ | |
4386 | frn = &memcg->cgwb_frn[oldest]; | |
4387 | frn->bdi_id = wb->bdi->id; | |
4388 | frn->memcg_id = wb->memcg_css->id; | |
4389 | frn->at = now; | |
4390 | } | |
4391 | } | |
4392 | ||
4393 | /* issue foreign writeback flushes for recorded foreign dirtying events */ | |
4394 | void mem_cgroup_flush_foreign(struct bdi_writeback *wb) | |
4395 | { | |
4396 | struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css); | |
4397 | unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10); | |
4398 | u64 now = jiffies_64; | |
4399 | int i; | |
4400 | ||
4401 | for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) { | |
4402 | struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i]; | |
4403 | ||
4404 | /* | |
4405 | * If the record is older than dirty_expire_interval, | |
4406 | * writeback on it has already started. No need to kick it | |
4407 | * off again. Also, don't start a new one if there's | |
4408 | * already one in flight. | |
4409 | */ | |
4410 | if (time_after64(frn->at, now - intv) && | |
4411 | atomic_read(&frn->done.cnt) == 1) { | |
4412 | frn->at = 0; | |
4413 | trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id); | |
4414 | cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0, | |
4415 | WB_REASON_FOREIGN_FLUSH, | |
4416 | &frn->done); | |
4417 | } | |
4418 | } | |
4419 | } | |
4420 | ||
4421 | #else /* CONFIG_CGROUP_WRITEBACK */ | |
4422 | ||
4423 | static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp) | |
4424 | { | |
4425 | return 0; | |
4426 | } | |
4427 | ||
4428 | static void memcg_wb_domain_exit(struct mem_cgroup *memcg) | |
4429 | { | |
4430 | } | |
4431 | ||
4432 | static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg) | |
4433 | { | |
4434 | } | |
4435 | ||
4436 | #endif /* CONFIG_CGROUP_WRITEBACK */ | |
4437 | ||
4438 | /* | |
4439 | * DO NOT USE IN NEW FILES. | |
4440 | * | |
4441 | * "cgroup.event_control" implementation. | |
4442 | * | |
4443 | * This is way over-engineered. It tries to support fully configurable | |
4444 | * events for each user. Such level of flexibility is completely | |
4445 | * unnecessary especially in the light of the planned unified hierarchy. | |
4446 | * | |
4447 | * Please deprecate this and replace with something simpler if at all | |
4448 | * possible. | |
4449 | */ | |
4450 | ||
4451 | /* | |
4452 | * Unregister event and free resources. | |
4453 | * | |
4454 | * Gets called from workqueue. | |
4455 | */ | |
4456 | static void memcg_event_remove(struct work_struct *work) | |
4457 | { | |
4458 | struct mem_cgroup_event *event = | |
4459 | container_of(work, struct mem_cgroup_event, remove); | |
4460 | struct mem_cgroup *memcg = event->memcg; | |
4461 | ||
4462 | remove_wait_queue(event->wqh, &event->wait); | |
4463 | ||
4464 | event->unregister_event(memcg, event->eventfd); | |
4465 | ||
4466 | /* Notify userspace the event is going away. */ | |
4467 | eventfd_signal(event->eventfd, 1); | |
4468 | ||
4469 | eventfd_ctx_put(event->eventfd); | |
4470 | kfree(event); | |
4471 | css_put(&memcg->css); | |
4472 | } | |
4473 | ||
4474 | /* | |
4475 | * Gets called on EPOLLHUP on eventfd when user closes it. | |
4476 | * | |
4477 | * Called with wqh->lock held and interrupts disabled. | |
4478 | */ | |
4479 | static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode, | |
4480 | int sync, void *key) | |
4481 | { | |
4482 | struct mem_cgroup_event *event = | |
4483 | container_of(wait, struct mem_cgroup_event, wait); | |
4484 | struct mem_cgroup *memcg = event->memcg; | |
4485 | __poll_t flags = key_to_poll(key); | |
4486 | ||
4487 | if (flags & EPOLLHUP) { | |
4488 | /* | |
4489 | * If the event has been detached at cgroup removal, we | |
4490 | * can simply return knowing the other side will cleanup | |
4491 | * for us. | |
4492 | * | |
4493 | * We can't race against event freeing since the other | |
4494 | * side will require wqh->lock via remove_wait_queue(), | |
4495 | * which we hold. | |
4496 | */ | |
4497 | spin_lock(&memcg->event_list_lock); | |
4498 | if (!list_empty(&event->list)) { | |
4499 | list_del_init(&event->list); | |
4500 | /* | |
4501 | * We are in atomic context, but cgroup_event_remove() | |
4502 | * may sleep, so we have to call it in workqueue. | |
4503 | */ | |
4504 | schedule_work(&event->remove); | |
4505 | } | |
4506 | spin_unlock(&memcg->event_list_lock); | |
4507 | } | |
4508 | ||
4509 | return 0; | |
4510 | } | |
4511 | ||
4512 | static void memcg_event_ptable_queue_proc(struct file *file, | |
4513 | wait_queue_head_t *wqh, poll_table *pt) | |
4514 | { | |
4515 | struct mem_cgroup_event *event = | |
4516 | container_of(pt, struct mem_cgroup_event, pt); | |
4517 | ||
4518 | event->wqh = wqh; | |
4519 | add_wait_queue(wqh, &event->wait); | |
4520 | } | |
4521 | ||
4522 | /* | |
4523 | * DO NOT USE IN NEW FILES. | |
4524 | * | |
4525 | * Parse input and register new cgroup event handler. | |
4526 | * | |
4527 | * Input must be in format '<event_fd> <control_fd> <args>'. | |
4528 | * Interpretation of args is defined by control file implementation. | |
4529 | */ | |
4530 | static ssize_t memcg_write_event_control(struct kernfs_open_file *of, | |
4531 | char *buf, size_t nbytes, loff_t off) | |
4532 | { | |
4533 | struct cgroup_subsys_state *css = of_css(of); | |
4534 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
4535 | struct mem_cgroup_event *event; | |
4536 | struct cgroup_subsys_state *cfile_css; | |
4537 | unsigned int efd, cfd; | |
4538 | struct fd efile; | |
4539 | struct fd cfile; | |
4540 | const char *name; | |
4541 | char *endp; | |
4542 | int ret; | |
4543 | ||
4544 | buf = strstrip(buf); | |
4545 | ||
4546 | efd = simple_strtoul(buf, &endp, 10); | |
4547 | if (*endp != ' ') | |
4548 | return -EINVAL; | |
4549 | buf = endp + 1; | |
4550 | ||
4551 | cfd = simple_strtoul(buf, &endp, 10); | |
4552 | if ((*endp != ' ') && (*endp != '\0')) | |
4553 | return -EINVAL; | |
4554 | buf = endp + 1; | |
4555 | ||
4556 | event = kzalloc(sizeof(*event), GFP_KERNEL); | |
4557 | if (!event) | |
4558 | return -ENOMEM; | |
4559 | ||
4560 | event->memcg = memcg; | |
4561 | INIT_LIST_HEAD(&event->list); | |
4562 | init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc); | |
4563 | init_waitqueue_func_entry(&event->wait, memcg_event_wake); | |
4564 | INIT_WORK(&event->remove, memcg_event_remove); | |
4565 | ||
4566 | efile = fdget(efd); | |
4567 | if (!efile.file) { | |
4568 | ret = -EBADF; | |
4569 | goto out_kfree; | |
4570 | } | |
4571 | ||
4572 | event->eventfd = eventfd_ctx_fileget(efile.file); | |
4573 | if (IS_ERR(event->eventfd)) { | |
4574 | ret = PTR_ERR(event->eventfd); | |
4575 | goto out_put_efile; | |
4576 | } | |
4577 | ||
4578 | cfile = fdget(cfd); | |
4579 | if (!cfile.file) { | |
4580 | ret = -EBADF; | |
4581 | goto out_put_eventfd; | |
4582 | } | |
4583 | ||
4584 | /* the process need read permission on control file */ | |
4585 | /* AV: shouldn't we check that it's been opened for read instead? */ | |
4586 | ret = inode_permission(file_inode(cfile.file), MAY_READ); | |
4587 | if (ret < 0) | |
4588 | goto out_put_cfile; | |
4589 | ||
4590 | /* | |
4591 | * Determine the event callbacks and set them in @event. This used | |
4592 | * to be done via struct cftype but cgroup core no longer knows | |
4593 | * about these events. The following is crude but the whole thing | |
4594 | * is for compatibility anyway. | |
4595 | * | |
4596 | * DO NOT ADD NEW FILES. | |
4597 | */ | |
4598 | name = cfile.file->f_path.dentry->d_name.name; | |
4599 | ||
4600 | if (!strcmp(name, "memory.usage_in_bytes")) { | |
4601 | event->register_event = mem_cgroup_usage_register_event; | |
4602 | event->unregister_event = mem_cgroup_usage_unregister_event; | |
4603 | } else if (!strcmp(name, "memory.oom_control")) { | |
4604 | event->register_event = mem_cgroup_oom_register_event; | |
4605 | event->unregister_event = mem_cgroup_oom_unregister_event; | |
4606 | } else if (!strcmp(name, "memory.pressure_level")) { | |
4607 | event->register_event = vmpressure_register_event; | |
4608 | event->unregister_event = vmpressure_unregister_event; | |
4609 | } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) { | |
4610 | event->register_event = memsw_cgroup_usage_register_event; | |
4611 | event->unregister_event = memsw_cgroup_usage_unregister_event; | |
4612 | } else { | |
4613 | ret = -EINVAL; | |
4614 | goto out_put_cfile; | |
4615 | } | |
4616 | ||
4617 | /* | |
4618 | * Verify @cfile should belong to @css. Also, remaining events are | |
4619 | * automatically removed on cgroup destruction but the removal is | |
4620 | * asynchronous, so take an extra ref on @css. | |
4621 | */ | |
4622 | cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent, | |
4623 | &memory_cgrp_subsys); | |
4624 | ret = -EINVAL; | |
4625 | if (IS_ERR(cfile_css)) | |
4626 | goto out_put_cfile; | |
4627 | if (cfile_css != css) { | |
4628 | css_put(cfile_css); | |
4629 | goto out_put_cfile; | |
4630 | } | |
4631 | ||
4632 | ret = event->register_event(memcg, event->eventfd, buf); | |
4633 | if (ret) | |
4634 | goto out_put_css; | |
4635 | ||
4636 | vfs_poll(efile.file, &event->pt); | |
4637 | ||
4638 | spin_lock(&memcg->event_list_lock); | |
4639 | list_add(&event->list, &memcg->event_list); | |
4640 | spin_unlock(&memcg->event_list_lock); | |
4641 | ||
4642 | fdput(cfile); | |
4643 | fdput(efile); | |
4644 | ||
4645 | return nbytes; | |
4646 | ||
4647 | out_put_css: | |
4648 | css_put(css); | |
4649 | out_put_cfile: | |
4650 | fdput(cfile); | |
4651 | out_put_eventfd: | |
4652 | eventfd_ctx_put(event->eventfd); | |
4653 | out_put_efile: | |
4654 | fdput(efile); | |
4655 | out_kfree: | |
4656 | kfree(event); | |
4657 | ||
4658 | return ret; | |
4659 | } | |
4660 | ||
4661 | static struct cftype mem_cgroup_legacy_files[] = { | |
4662 | { | |
4663 | .name = "usage_in_bytes", | |
4664 | .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), | |
4665 | .read_u64 = mem_cgroup_read_u64, | |
4666 | }, | |
4667 | { | |
4668 | .name = "max_usage_in_bytes", | |
4669 | .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), | |
4670 | .write = mem_cgroup_reset, | |
4671 | .read_u64 = mem_cgroup_read_u64, | |
4672 | }, | |
4673 | { | |
4674 | .name = "limit_in_bytes", | |
4675 | .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), | |
4676 | .write = mem_cgroup_write, | |
4677 | .read_u64 = mem_cgroup_read_u64, | |
4678 | }, | |
4679 | { | |
4680 | .name = "soft_limit_in_bytes", | |
4681 | .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), | |
4682 | .write = mem_cgroup_write, | |
4683 | .read_u64 = mem_cgroup_read_u64, | |
4684 | }, | |
4685 | { | |
4686 | .name = "failcnt", | |
4687 | .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), | |
4688 | .write = mem_cgroup_reset, | |
4689 | .read_u64 = mem_cgroup_read_u64, | |
4690 | }, | |
4691 | { | |
4692 | .name = "stat", | |
4693 | .seq_show = memcg_stat_show, | |
4694 | }, | |
4695 | { | |
4696 | .name = "force_empty", | |
4697 | .write = mem_cgroup_force_empty_write, | |
4698 | }, | |
4699 | { | |
4700 | .name = "use_hierarchy", | |
4701 | .write_u64 = mem_cgroup_hierarchy_write, | |
4702 | .read_u64 = mem_cgroup_hierarchy_read, | |
4703 | }, | |
4704 | { | |
4705 | .name = "cgroup.event_control", /* XXX: for compat */ | |
4706 | .write = memcg_write_event_control, | |
4707 | .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE, | |
4708 | }, | |
4709 | { | |
4710 | .name = "swappiness", | |
4711 | .read_u64 = mem_cgroup_swappiness_read, | |
4712 | .write_u64 = mem_cgroup_swappiness_write, | |
4713 | }, | |
4714 | { | |
4715 | .name = "move_charge_at_immigrate", | |
4716 | .read_u64 = mem_cgroup_move_charge_read, | |
4717 | .write_u64 = mem_cgroup_move_charge_write, | |
4718 | }, | |
4719 | { | |
4720 | .name = "oom_control", | |
4721 | .seq_show = mem_cgroup_oom_control_read, | |
4722 | .write_u64 = mem_cgroup_oom_control_write, | |
4723 | .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), | |
4724 | }, | |
4725 | { | |
4726 | .name = "pressure_level", | |
4727 | }, | |
4728 | #ifdef CONFIG_NUMA | |
4729 | { | |
4730 | .name = "numa_stat", | |
4731 | .seq_show = memcg_numa_stat_show, | |
4732 | }, | |
4733 | #endif | |
4734 | { | |
4735 | .name = "kmem.limit_in_bytes", | |
4736 | .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT), | |
4737 | .write = mem_cgroup_write, | |
4738 | .read_u64 = mem_cgroup_read_u64, | |
4739 | }, | |
4740 | { | |
4741 | .name = "kmem.usage_in_bytes", | |
4742 | .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE), | |
4743 | .read_u64 = mem_cgroup_read_u64, | |
4744 | }, | |
4745 | { | |
4746 | .name = "kmem.failcnt", | |
4747 | .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT), | |
4748 | .write = mem_cgroup_reset, | |
4749 | .read_u64 = mem_cgroup_read_u64, | |
4750 | }, | |
4751 | { | |
4752 | .name = "kmem.max_usage_in_bytes", | |
4753 | .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE), | |
4754 | .write = mem_cgroup_reset, | |
4755 | .read_u64 = mem_cgroup_read_u64, | |
4756 | }, | |
4757 | #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG) | |
4758 | { | |
4759 | .name = "kmem.slabinfo", | |
4760 | .seq_start = memcg_slab_start, | |
4761 | .seq_next = memcg_slab_next, | |
4762 | .seq_stop = memcg_slab_stop, | |
4763 | .seq_show = memcg_slab_show, | |
4764 | }, | |
4765 | #endif | |
4766 | { | |
4767 | .name = "kmem.tcp.limit_in_bytes", | |
4768 | .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT), | |
4769 | .write = mem_cgroup_write, | |
4770 | .read_u64 = mem_cgroup_read_u64, | |
4771 | }, | |
4772 | { | |
4773 | .name = "kmem.tcp.usage_in_bytes", | |
4774 | .private = MEMFILE_PRIVATE(_TCP, RES_USAGE), | |
4775 | .read_u64 = mem_cgroup_read_u64, | |
4776 | }, | |
4777 | { | |
4778 | .name = "kmem.tcp.failcnt", | |
4779 | .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT), | |
4780 | .write = mem_cgroup_reset, | |
4781 | .read_u64 = mem_cgroup_read_u64, | |
4782 | }, | |
4783 | { | |
4784 | .name = "kmem.tcp.max_usage_in_bytes", | |
4785 | .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE), | |
4786 | .write = mem_cgroup_reset, | |
4787 | .read_u64 = mem_cgroup_read_u64, | |
4788 | }, | |
4789 | { }, /* terminate */ | |
4790 | }; | |
4791 | ||
4792 | /* | |
4793 | * Private memory cgroup IDR | |
4794 | * | |
4795 | * Swap-out records and page cache shadow entries need to store memcg | |
4796 | * references in constrained space, so we maintain an ID space that is | |
4797 | * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of | |
4798 | * memory-controlled cgroups to 64k. | |
4799 | * | |
4800 | * However, there usually are many references to the oflline CSS after | |
4801 | * the cgroup has been destroyed, such as page cache or reclaimable | |
4802 | * slab objects, that don't need to hang on to the ID. We want to keep | |
4803 | * those dead CSS from occupying IDs, or we might quickly exhaust the | |
4804 | * relatively small ID space and prevent the creation of new cgroups | |
4805 | * even when there are much fewer than 64k cgroups - possibly none. | |
4806 | * | |
4807 | * Maintain a private 16-bit ID space for memcg, and allow the ID to | |
4808 | * be freed and recycled when it's no longer needed, which is usually | |
4809 | * when the CSS is offlined. | |
4810 | * | |
4811 | * The only exception to that are records of swapped out tmpfs/shmem | |
4812 | * pages that need to be attributed to live ancestors on swapin. But | |
4813 | * those references are manageable from userspace. | |
4814 | */ | |
4815 | ||
4816 | static DEFINE_IDR(mem_cgroup_idr); | |
4817 | ||
4818 | static void mem_cgroup_id_remove(struct mem_cgroup *memcg) | |
4819 | { | |
4820 | if (memcg->id.id > 0) { | |
4821 | idr_remove(&mem_cgroup_idr, memcg->id.id); | |
4822 | memcg->id.id = 0; | |
4823 | } | |
4824 | } | |
4825 | ||
4826 | static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n) | |
4827 | { | |
4828 | refcount_add(n, &memcg->id.ref); | |
4829 | } | |
4830 | ||
4831 | static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n) | |
4832 | { | |
4833 | if (refcount_sub_and_test(n, &memcg->id.ref)) { | |
4834 | mem_cgroup_id_remove(memcg); | |
4835 | ||
4836 | /* Memcg ID pins CSS */ | |
4837 | css_put(&memcg->css); | |
4838 | } | |
4839 | } | |
4840 | ||
4841 | static inline void mem_cgroup_id_put(struct mem_cgroup *memcg) | |
4842 | { | |
4843 | mem_cgroup_id_put_many(memcg, 1); | |
4844 | } | |
4845 | ||
4846 | /** | |
4847 | * mem_cgroup_from_id - look up a memcg from a memcg id | |
4848 | * @id: the memcg id to look up | |
4849 | * | |
4850 | * Caller must hold rcu_read_lock(). | |
4851 | */ | |
4852 | struct mem_cgroup *mem_cgroup_from_id(unsigned short id) | |
4853 | { | |
4854 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
4855 | return idr_find(&mem_cgroup_idr, id); | |
4856 | } | |
4857 | ||
4858 | static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node) | |
4859 | { | |
4860 | struct mem_cgroup_per_node *pn; | |
4861 | int tmp = node; | |
4862 | /* | |
4863 | * This routine is called against possible nodes. | |
4864 | * But it's BUG to call kmalloc() against offline node. | |
4865 | * | |
4866 | * TODO: this routine can waste much memory for nodes which will | |
4867 | * never be onlined. It's better to use memory hotplug callback | |
4868 | * function. | |
4869 | */ | |
4870 | if (!node_state(node, N_NORMAL_MEMORY)) | |
4871 | tmp = -1; | |
4872 | pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); | |
4873 | if (!pn) | |
4874 | return 1; | |
4875 | ||
4876 | pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat); | |
4877 | if (!pn->lruvec_stat_local) { | |
4878 | kfree(pn); | |
4879 | return 1; | |
4880 | } | |
4881 | ||
4882 | pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat); | |
4883 | if (!pn->lruvec_stat_cpu) { | |
4884 | free_percpu(pn->lruvec_stat_local); | |
4885 | kfree(pn); | |
4886 | return 1; | |
4887 | } | |
4888 | ||
4889 | lruvec_init(&pn->lruvec); | |
4890 | pn->usage_in_excess = 0; | |
4891 | pn->on_tree = false; | |
4892 | pn->memcg = memcg; | |
4893 | ||
4894 | memcg->nodeinfo[node] = pn; | |
4895 | return 0; | |
4896 | } | |
4897 | ||
4898 | static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node) | |
4899 | { | |
4900 | struct mem_cgroup_per_node *pn = memcg->nodeinfo[node]; | |
4901 | ||
4902 | if (!pn) | |
4903 | return; | |
4904 | ||
4905 | free_percpu(pn->lruvec_stat_cpu); | |
4906 | free_percpu(pn->lruvec_stat_local); | |
4907 | kfree(pn); | |
4908 | } | |
4909 | ||
4910 | static void __mem_cgroup_free(struct mem_cgroup *memcg) | |
4911 | { | |
4912 | int node; | |
4913 | ||
4914 | for_each_node(node) | |
4915 | free_mem_cgroup_per_node_info(memcg, node); | |
4916 | free_percpu(memcg->vmstats_percpu); | |
4917 | free_percpu(memcg->vmstats_local); | |
4918 | kfree(memcg); | |
4919 | } | |
4920 | ||
4921 | static void mem_cgroup_free(struct mem_cgroup *memcg) | |
4922 | { | |
4923 | memcg_wb_domain_exit(memcg); | |
4924 | /* | |
4925 | * Flush percpu vmstats and vmevents to guarantee the value correctness | |
4926 | * on parent's and all ancestor levels. | |
4927 | */ | |
4928 | memcg_flush_percpu_vmstats(memcg); | |
4929 | memcg_flush_percpu_vmevents(memcg); | |
4930 | __mem_cgroup_free(memcg); | |
4931 | } | |
4932 | ||
4933 | static struct mem_cgroup *mem_cgroup_alloc(void) | |
4934 | { | |
4935 | struct mem_cgroup *memcg; | |
4936 | unsigned int size; | |
4937 | int node; | |
4938 | int __maybe_unused i; | |
4939 | ||
4940 | size = sizeof(struct mem_cgroup); | |
4941 | size += nr_node_ids * sizeof(struct mem_cgroup_per_node *); | |
4942 | ||
4943 | memcg = kzalloc(size, GFP_KERNEL); | |
4944 | if (!memcg) | |
4945 | return NULL; | |
4946 | ||
4947 | memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL, | |
4948 | 1, MEM_CGROUP_ID_MAX, | |
4949 | GFP_KERNEL); | |
4950 | if (memcg->id.id < 0) | |
4951 | goto fail; | |
4952 | ||
4953 | memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu); | |
4954 | if (!memcg->vmstats_local) | |
4955 | goto fail; | |
4956 | ||
4957 | memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu); | |
4958 | if (!memcg->vmstats_percpu) | |
4959 | goto fail; | |
4960 | ||
4961 | for_each_node(node) | |
4962 | if (alloc_mem_cgroup_per_node_info(memcg, node)) | |
4963 | goto fail; | |
4964 | ||
4965 | if (memcg_wb_domain_init(memcg, GFP_KERNEL)) | |
4966 | goto fail; | |
4967 | ||
4968 | INIT_WORK(&memcg->high_work, high_work_func); | |
4969 | INIT_LIST_HEAD(&memcg->oom_notify); | |
4970 | mutex_init(&memcg->thresholds_lock); | |
4971 | spin_lock_init(&memcg->move_lock); | |
4972 | vmpressure_init(&memcg->vmpressure); | |
4973 | INIT_LIST_HEAD(&memcg->event_list); | |
4974 | spin_lock_init(&memcg->event_list_lock); | |
4975 | memcg->socket_pressure = jiffies; | |
4976 | #ifdef CONFIG_MEMCG_KMEM | |
4977 | memcg->kmemcg_id = -1; | |
4978 | #endif | |
4979 | #ifdef CONFIG_CGROUP_WRITEBACK | |
4980 | INIT_LIST_HEAD(&memcg->cgwb_list); | |
4981 | for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) | |
4982 | memcg->cgwb_frn[i].done = | |
4983 | __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq); | |
4984 | #endif | |
4985 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
4986 | spin_lock_init(&memcg->deferred_split_queue.split_queue_lock); | |
4987 | INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue); | |
4988 | memcg->deferred_split_queue.split_queue_len = 0; | |
4989 | #endif | |
4990 | idr_replace(&mem_cgroup_idr, memcg, memcg->id.id); | |
4991 | return memcg; | |
4992 | fail: | |
4993 | mem_cgroup_id_remove(memcg); | |
4994 | __mem_cgroup_free(memcg); | |
4995 | return NULL; | |
4996 | } | |
4997 | ||
4998 | static struct cgroup_subsys_state * __ref | |
4999 | mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) | |
5000 | { | |
5001 | struct mem_cgroup *parent = mem_cgroup_from_css(parent_css); | |
5002 | struct mem_cgroup *memcg; | |
5003 | long error = -ENOMEM; | |
5004 | ||
5005 | memcg = mem_cgroup_alloc(); | |
5006 | if (!memcg) | |
5007 | return ERR_PTR(error); | |
5008 | ||
5009 | memcg->high = PAGE_COUNTER_MAX; | |
5010 | memcg->soft_limit = PAGE_COUNTER_MAX; | |
5011 | if (parent) { | |
5012 | memcg->swappiness = mem_cgroup_swappiness(parent); | |
5013 | memcg->oom_kill_disable = parent->oom_kill_disable; | |
5014 | } | |
5015 | if (parent && parent->use_hierarchy) { | |
5016 | memcg->use_hierarchy = true; | |
5017 | page_counter_init(&memcg->memory, &parent->memory); | |
5018 | page_counter_init(&memcg->swap, &parent->swap); | |
5019 | page_counter_init(&memcg->memsw, &parent->memsw); | |
5020 | page_counter_init(&memcg->kmem, &parent->kmem); | |
5021 | page_counter_init(&memcg->tcpmem, &parent->tcpmem); | |
5022 | } else { | |
5023 | page_counter_init(&memcg->memory, NULL); | |
5024 | page_counter_init(&memcg->swap, NULL); | |
5025 | page_counter_init(&memcg->memsw, NULL); | |
5026 | page_counter_init(&memcg->kmem, NULL); | |
5027 | page_counter_init(&memcg->tcpmem, NULL); | |
5028 | /* | |
5029 | * Deeper hierachy with use_hierarchy == false doesn't make | |
5030 | * much sense so let cgroup subsystem know about this | |
5031 | * unfortunate state in our controller. | |
5032 | */ | |
5033 | if (parent != root_mem_cgroup) | |
5034 | memory_cgrp_subsys.broken_hierarchy = true; | |
5035 | } | |
5036 | ||
5037 | /* The following stuff does not apply to the root */ | |
5038 | if (!parent) { | |
5039 | #ifdef CONFIG_MEMCG_KMEM | |
5040 | INIT_LIST_HEAD(&memcg->kmem_caches); | |
5041 | #endif | |
5042 | root_mem_cgroup = memcg; | |
5043 | return &memcg->css; | |
5044 | } | |
5045 | ||
5046 | error = memcg_online_kmem(memcg); | |
5047 | if (error) | |
5048 | goto fail; | |
5049 | ||
5050 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket) | |
5051 | static_branch_inc(&memcg_sockets_enabled_key); | |
5052 | ||
5053 | return &memcg->css; | |
5054 | fail: | |
5055 | mem_cgroup_id_remove(memcg); | |
5056 | mem_cgroup_free(memcg); | |
5057 | return ERR_PTR(-ENOMEM); | |
5058 | } | |
5059 | ||
5060 | static int mem_cgroup_css_online(struct cgroup_subsys_state *css) | |
5061 | { | |
5062 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5063 | ||
5064 | /* | |
5065 | * A memcg must be visible for memcg_expand_shrinker_maps() | |
5066 | * by the time the maps are allocated. So, we allocate maps | |
5067 | * here, when for_each_mem_cgroup() can't skip it. | |
5068 | */ | |
5069 | if (memcg_alloc_shrinker_maps(memcg)) { | |
5070 | mem_cgroup_id_remove(memcg); | |
5071 | return -ENOMEM; | |
5072 | } | |
5073 | ||
5074 | /* Online state pins memcg ID, memcg ID pins CSS */ | |
5075 | refcount_set(&memcg->id.ref, 1); | |
5076 | css_get(css); | |
5077 | return 0; | |
5078 | } | |
5079 | ||
5080 | static void mem_cgroup_css_offline(struct cgroup_subsys_state *css) | |
5081 | { | |
5082 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5083 | struct mem_cgroup_event *event, *tmp; | |
5084 | ||
5085 | /* | |
5086 | * Unregister events and notify userspace. | |
5087 | * Notify userspace about cgroup removing only after rmdir of cgroup | |
5088 | * directory to avoid race between userspace and kernelspace. | |
5089 | */ | |
5090 | spin_lock(&memcg->event_list_lock); | |
5091 | list_for_each_entry_safe(event, tmp, &memcg->event_list, list) { | |
5092 | list_del_init(&event->list); | |
5093 | schedule_work(&event->remove); | |
5094 | } | |
5095 | spin_unlock(&memcg->event_list_lock); | |
5096 | ||
5097 | page_counter_set_min(&memcg->memory, 0); | |
5098 | page_counter_set_low(&memcg->memory, 0); | |
5099 | ||
5100 | memcg_offline_kmem(memcg); | |
5101 | wb_memcg_offline(memcg); | |
5102 | ||
5103 | drain_all_stock(memcg); | |
5104 | ||
5105 | mem_cgroup_id_put(memcg); | |
5106 | } | |
5107 | ||
5108 | static void mem_cgroup_css_released(struct cgroup_subsys_state *css) | |
5109 | { | |
5110 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5111 | ||
5112 | invalidate_reclaim_iterators(memcg); | |
5113 | } | |
5114 | ||
5115 | static void mem_cgroup_css_free(struct cgroup_subsys_state *css) | |
5116 | { | |
5117 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5118 | int __maybe_unused i; | |
5119 | ||
5120 | #ifdef CONFIG_CGROUP_WRITEBACK | |
5121 | for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) | |
5122 | wb_wait_for_completion(&memcg->cgwb_frn[i].done); | |
5123 | #endif | |
5124 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket) | |
5125 | static_branch_dec(&memcg_sockets_enabled_key); | |
5126 | ||
5127 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active) | |
5128 | static_branch_dec(&memcg_sockets_enabled_key); | |
5129 | ||
5130 | vmpressure_cleanup(&memcg->vmpressure); | |
5131 | cancel_work_sync(&memcg->high_work); | |
5132 | mem_cgroup_remove_from_trees(memcg); | |
5133 | memcg_free_shrinker_maps(memcg); | |
5134 | memcg_free_kmem(memcg); | |
5135 | mem_cgroup_free(memcg); | |
5136 | } | |
5137 | ||
5138 | /** | |
5139 | * mem_cgroup_css_reset - reset the states of a mem_cgroup | |
5140 | * @css: the target css | |
5141 | * | |
5142 | * Reset the states of the mem_cgroup associated with @css. This is | |
5143 | * invoked when the userland requests disabling on the default hierarchy | |
5144 | * but the memcg is pinned through dependency. The memcg should stop | |
5145 | * applying policies and should revert to the vanilla state as it may be | |
5146 | * made visible again. | |
5147 | * | |
5148 | * The current implementation only resets the essential configurations. | |
5149 | * This needs to be expanded to cover all the visible parts. | |
5150 | */ | |
5151 | static void mem_cgroup_css_reset(struct cgroup_subsys_state *css) | |
5152 | { | |
5153 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5154 | ||
5155 | page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX); | |
5156 | page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX); | |
5157 | page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX); | |
5158 | page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX); | |
5159 | page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX); | |
5160 | page_counter_set_min(&memcg->memory, 0); | |
5161 | page_counter_set_low(&memcg->memory, 0); | |
5162 | memcg->high = PAGE_COUNTER_MAX; | |
5163 | memcg->soft_limit = PAGE_COUNTER_MAX; | |
5164 | memcg_wb_domain_size_changed(memcg); | |
5165 | } | |
5166 | ||
5167 | #ifdef CONFIG_MMU | |
5168 | /* Handlers for move charge at task migration. */ | |
5169 | static int mem_cgroup_do_precharge(unsigned long count) | |
5170 | { | |
5171 | int ret; | |
5172 | ||
5173 | /* Try a single bulk charge without reclaim first, kswapd may wake */ | |
5174 | ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count); | |
5175 | if (!ret) { | |
5176 | mc.precharge += count; | |
5177 | return ret; | |
5178 | } | |
5179 | ||
5180 | /* Try charges one by one with reclaim, but do not retry */ | |
5181 | while (count--) { | |
5182 | ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1); | |
5183 | if (ret) | |
5184 | return ret; | |
5185 | mc.precharge++; | |
5186 | cond_resched(); | |
5187 | } | |
5188 | return 0; | |
5189 | } | |
5190 | ||
5191 | union mc_target { | |
5192 | struct page *page; | |
5193 | swp_entry_t ent; | |
5194 | }; | |
5195 | ||
5196 | enum mc_target_type { | |
5197 | MC_TARGET_NONE = 0, | |
5198 | MC_TARGET_PAGE, | |
5199 | MC_TARGET_SWAP, | |
5200 | MC_TARGET_DEVICE, | |
5201 | }; | |
5202 | ||
5203 | static struct page *mc_handle_present_pte(struct vm_area_struct *vma, | |
5204 | unsigned long addr, pte_t ptent) | |
5205 | { | |
5206 | struct page *page = vm_normal_page(vma, addr, ptent); | |
5207 | ||
5208 | if (!page || !page_mapped(page)) | |
5209 | return NULL; | |
5210 | if (PageAnon(page)) { | |
5211 | if (!(mc.flags & MOVE_ANON)) | |
5212 | return NULL; | |
5213 | } else { | |
5214 | if (!(mc.flags & MOVE_FILE)) | |
5215 | return NULL; | |
5216 | } | |
5217 | if (!get_page_unless_zero(page)) | |
5218 | return NULL; | |
5219 | ||
5220 | return page; | |
5221 | } | |
5222 | ||
5223 | #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE) | |
5224 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
5225 | pte_t ptent, swp_entry_t *entry) | |
5226 | { | |
5227 | struct page *page = NULL; | |
5228 | swp_entry_t ent = pte_to_swp_entry(ptent); | |
5229 | ||
5230 | if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent)) | |
5231 | return NULL; | |
5232 | ||
5233 | /* | |
5234 | * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to | |
5235 | * a device and because they are not accessible by CPU they are store | |
5236 | * as special swap entry in the CPU page table. | |
5237 | */ | |
5238 | if (is_device_private_entry(ent)) { | |
5239 | page = device_private_entry_to_page(ent); | |
5240 | /* | |
5241 | * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have | |
5242 | * a refcount of 1 when free (unlike normal page) | |
5243 | */ | |
5244 | if (!page_ref_add_unless(page, 1, 1)) | |
5245 | return NULL; | |
5246 | return page; | |
5247 | } | |
5248 | ||
5249 | /* | |
5250 | * Because lookup_swap_cache() updates some statistics counter, | |
5251 | * we call find_get_page() with swapper_space directly. | |
5252 | */ | |
5253 | page = find_get_page(swap_address_space(ent), swp_offset(ent)); | |
5254 | if (do_memsw_account()) | |
5255 | entry->val = ent.val; | |
5256 | ||
5257 | return page; | |
5258 | } | |
5259 | #else | |
5260 | static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, | |
5261 | pte_t ptent, swp_entry_t *entry) | |
5262 | { | |
5263 | return NULL; | |
5264 | } | |
5265 | #endif | |
5266 | ||
5267 | static struct page *mc_handle_file_pte(struct vm_area_struct *vma, | |
5268 | unsigned long addr, pte_t ptent, swp_entry_t *entry) | |
5269 | { | |
5270 | struct page *page = NULL; | |
5271 | struct address_space *mapping; | |
5272 | pgoff_t pgoff; | |
5273 | ||
5274 | if (!vma->vm_file) /* anonymous vma */ | |
5275 | return NULL; | |
5276 | if (!(mc.flags & MOVE_FILE)) | |
5277 | return NULL; | |
5278 | ||
5279 | mapping = vma->vm_file->f_mapping; | |
5280 | pgoff = linear_page_index(vma, addr); | |
5281 | ||
5282 | /* page is moved even if it's not RSS of this task(page-faulted). */ | |
5283 | #ifdef CONFIG_SWAP | |
5284 | /* shmem/tmpfs may report page out on swap: account for that too. */ | |
5285 | if (shmem_mapping(mapping)) { | |
5286 | page = find_get_entry(mapping, pgoff); | |
5287 | if (xa_is_value(page)) { | |
5288 | swp_entry_t swp = radix_to_swp_entry(page); | |
5289 | if (do_memsw_account()) | |
5290 | *entry = swp; | |
5291 | page = find_get_page(swap_address_space(swp), | |
5292 | swp_offset(swp)); | |
5293 | } | |
5294 | } else | |
5295 | page = find_get_page(mapping, pgoff); | |
5296 | #else | |
5297 | page = find_get_page(mapping, pgoff); | |
5298 | #endif | |
5299 | return page; | |
5300 | } | |
5301 | ||
5302 | /** | |
5303 | * mem_cgroup_move_account - move account of the page | |
5304 | * @page: the page | |
5305 | * @compound: charge the page as compound or small page | |
5306 | * @from: mem_cgroup which the page is moved from. | |
5307 | * @to: mem_cgroup which the page is moved to. @from != @to. | |
5308 | * | |
5309 | * The caller must make sure the page is not on LRU (isolate_page() is useful.) | |
5310 | * | |
5311 | * This function doesn't do "charge" to new cgroup and doesn't do "uncharge" | |
5312 | * from old cgroup. | |
5313 | */ | |
5314 | static int mem_cgroup_move_account(struct page *page, | |
5315 | bool compound, | |
5316 | struct mem_cgroup *from, | |
5317 | struct mem_cgroup *to) | |
5318 | { | |
5319 | struct lruvec *from_vec, *to_vec; | |
5320 | struct pglist_data *pgdat; | |
5321 | unsigned long flags; | |
5322 | unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1; | |
5323 | int ret; | |
5324 | bool anon; | |
5325 | ||
5326 | VM_BUG_ON(from == to); | |
5327 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
5328 | VM_BUG_ON(compound && !PageTransHuge(page)); | |
5329 | ||
5330 | /* | |
5331 | * Prevent mem_cgroup_migrate() from looking at | |
5332 | * page->mem_cgroup of its source page while we change it. | |
5333 | */ | |
5334 | ret = -EBUSY; | |
5335 | if (!trylock_page(page)) | |
5336 | goto out; | |
5337 | ||
5338 | ret = -EINVAL; | |
5339 | if (page->mem_cgroup != from) | |
5340 | goto out_unlock; | |
5341 | ||
5342 | anon = PageAnon(page); | |
5343 | ||
5344 | pgdat = page_pgdat(page); | |
5345 | from_vec = mem_cgroup_lruvec(from, pgdat); | |
5346 | to_vec = mem_cgroup_lruvec(to, pgdat); | |
5347 | ||
5348 | spin_lock_irqsave(&from->move_lock, flags); | |
5349 | ||
5350 | if (!anon && page_mapped(page)) { | |
5351 | __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages); | |
5352 | __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages); | |
5353 | } | |
5354 | ||
5355 | /* | |
5356 | * move_lock grabbed above and caller set from->moving_account, so | |
5357 | * mod_memcg_page_state will serialize updates to PageDirty. | |
5358 | * So mapping should be stable for dirty pages. | |
5359 | */ | |
5360 | if (!anon && PageDirty(page)) { | |
5361 | struct address_space *mapping = page_mapping(page); | |
5362 | ||
5363 | if (mapping_cap_account_dirty(mapping)) { | |
5364 | __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages); | |
5365 | __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages); | |
5366 | } | |
5367 | } | |
5368 | ||
5369 | if (PageWriteback(page)) { | |
5370 | __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages); | |
5371 | __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages); | |
5372 | } | |
5373 | ||
5374 | /* | |
5375 | * It is safe to change page->mem_cgroup here because the page | |
5376 | * is referenced, charged, and isolated - we can't race with | |
5377 | * uncharging, charging, migration, or LRU putback. | |
5378 | */ | |
5379 | ||
5380 | /* caller should have done css_get */ | |
5381 | page->mem_cgroup = to; | |
5382 | ||
5383 | spin_unlock_irqrestore(&from->move_lock, flags); | |
5384 | ||
5385 | ret = 0; | |
5386 | ||
5387 | local_irq_disable(); | |
5388 | mem_cgroup_charge_statistics(to, page, compound, nr_pages); | |
5389 | memcg_check_events(to, page); | |
5390 | mem_cgroup_charge_statistics(from, page, compound, -nr_pages); | |
5391 | memcg_check_events(from, page); | |
5392 | local_irq_enable(); | |
5393 | out_unlock: | |
5394 | unlock_page(page); | |
5395 | out: | |
5396 | return ret; | |
5397 | } | |
5398 | ||
5399 | /** | |
5400 | * get_mctgt_type - get target type of moving charge | |
5401 | * @vma: the vma the pte to be checked belongs | |
5402 | * @addr: the address corresponding to the pte to be checked | |
5403 | * @ptent: the pte to be checked | |
5404 | * @target: the pointer the target page or swap ent will be stored(can be NULL) | |
5405 | * | |
5406 | * Returns | |
5407 | * 0(MC_TARGET_NONE): if the pte is not a target for move charge. | |
5408 | * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for | |
5409 | * move charge. if @target is not NULL, the page is stored in target->page | |
5410 | * with extra refcnt got(Callers should handle it). | |
5411 | * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a | |
5412 | * target for charge migration. if @target is not NULL, the entry is stored | |
5413 | * in target->ent. | |
5414 | * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE | |
5415 | * (so ZONE_DEVICE page and thus not on the lru). | |
5416 | * For now we such page is charge like a regular page would be as for all | |
5417 | * intent and purposes it is just special memory taking the place of a | |
5418 | * regular page. | |
5419 | * | |
5420 | * See Documentations/vm/hmm.txt and include/linux/hmm.h | |
5421 | * | |
5422 | * Called with pte lock held. | |
5423 | */ | |
5424 | ||
5425 | static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma, | |
5426 | unsigned long addr, pte_t ptent, union mc_target *target) | |
5427 | { | |
5428 | struct page *page = NULL; | |
5429 | enum mc_target_type ret = MC_TARGET_NONE; | |
5430 | swp_entry_t ent = { .val = 0 }; | |
5431 | ||
5432 | if (pte_present(ptent)) | |
5433 | page = mc_handle_present_pte(vma, addr, ptent); | |
5434 | else if (is_swap_pte(ptent)) | |
5435 | page = mc_handle_swap_pte(vma, ptent, &ent); | |
5436 | else if (pte_none(ptent)) | |
5437 | page = mc_handle_file_pte(vma, addr, ptent, &ent); | |
5438 | ||
5439 | if (!page && !ent.val) | |
5440 | return ret; | |
5441 | if (page) { | |
5442 | /* | |
5443 | * Do only loose check w/o serialization. | |
5444 | * mem_cgroup_move_account() checks the page is valid or | |
5445 | * not under LRU exclusion. | |
5446 | */ | |
5447 | if (page->mem_cgroup == mc.from) { | |
5448 | ret = MC_TARGET_PAGE; | |
5449 | if (is_device_private_page(page)) | |
5450 | ret = MC_TARGET_DEVICE; | |
5451 | if (target) | |
5452 | target->page = page; | |
5453 | } | |
5454 | if (!ret || !target) | |
5455 | put_page(page); | |
5456 | } | |
5457 | /* | |
5458 | * There is a swap entry and a page doesn't exist or isn't charged. | |
5459 | * But we cannot move a tail-page in a THP. | |
5460 | */ | |
5461 | if (ent.val && !ret && (!page || !PageTransCompound(page)) && | |
5462 | mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) { | |
5463 | ret = MC_TARGET_SWAP; | |
5464 | if (target) | |
5465 | target->ent = ent; | |
5466 | } | |
5467 | return ret; | |
5468 | } | |
5469 | ||
5470 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
5471 | /* | |
5472 | * We don't consider PMD mapped swapping or file mapped pages because THP does | |
5473 | * not support them for now. | |
5474 | * Caller should make sure that pmd_trans_huge(pmd) is true. | |
5475 | */ | |
5476 | static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
5477 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
5478 | { | |
5479 | struct page *page = NULL; | |
5480 | enum mc_target_type ret = MC_TARGET_NONE; | |
5481 | ||
5482 | if (unlikely(is_swap_pmd(pmd))) { | |
5483 | VM_BUG_ON(thp_migration_supported() && | |
5484 | !is_pmd_migration_entry(pmd)); | |
5485 | return ret; | |
5486 | } | |
5487 | page = pmd_page(pmd); | |
5488 | VM_BUG_ON_PAGE(!page || !PageHead(page), page); | |
5489 | if (!(mc.flags & MOVE_ANON)) | |
5490 | return ret; | |
5491 | if (page->mem_cgroup == mc.from) { | |
5492 | ret = MC_TARGET_PAGE; | |
5493 | if (target) { | |
5494 | get_page(page); | |
5495 | target->page = page; | |
5496 | } | |
5497 | } | |
5498 | return ret; | |
5499 | } | |
5500 | #else | |
5501 | static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma, | |
5502 | unsigned long addr, pmd_t pmd, union mc_target *target) | |
5503 | { | |
5504 | return MC_TARGET_NONE; | |
5505 | } | |
5506 | #endif | |
5507 | ||
5508 | static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, | |
5509 | unsigned long addr, unsigned long end, | |
5510 | struct mm_walk *walk) | |
5511 | { | |
5512 | struct vm_area_struct *vma = walk->vma; | |
5513 | pte_t *pte; | |
5514 | spinlock_t *ptl; | |
5515 | ||
5516 | ptl = pmd_trans_huge_lock(pmd, vma); | |
5517 | if (ptl) { | |
5518 | /* | |
5519 | * Note their can not be MC_TARGET_DEVICE for now as we do not | |
5520 | * support transparent huge page with MEMORY_DEVICE_PRIVATE but | |
5521 | * this might change. | |
5522 | */ | |
5523 | if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE) | |
5524 | mc.precharge += HPAGE_PMD_NR; | |
5525 | spin_unlock(ptl); | |
5526 | return 0; | |
5527 | } | |
5528 | ||
5529 | if (pmd_trans_unstable(pmd)) | |
5530 | return 0; | |
5531 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
5532 | for (; addr != end; pte++, addr += PAGE_SIZE) | |
5533 | if (get_mctgt_type(vma, addr, *pte, NULL)) | |
5534 | mc.precharge++; /* increment precharge temporarily */ | |
5535 | pte_unmap_unlock(pte - 1, ptl); | |
5536 | cond_resched(); | |
5537 | ||
5538 | return 0; | |
5539 | } | |
5540 | ||
5541 | static const struct mm_walk_ops precharge_walk_ops = { | |
5542 | .pmd_entry = mem_cgroup_count_precharge_pte_range, | |
5543 | }; | |
5544 | ||
5545 | static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) | |
5546 | { | |
5547 | unsigned long precharge; | |
5548 | ||
5549 | down_read(&mm->mmap_sem); | |
5550 | walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL); | |
5551 | up_read(&mm->mmap_sem); | |
5552 | ||
5553 | precharge = mc.precharge; | |
5554 | mc.precharge = 0; | |
5555 | ||
5556 | return precharge; | |
5557 | } | |
5558 | ||
5559 | static int mem_cgroup_precharge_mc(struct mm_struct *mm) | |
5560 | { | |
5561 | unsigned long precharge = mem_cgroup_count_precharge(mm); | |
5562 | ||
5563 | VM_BUG_ON(mc.moving_task); | |
5564 | mc.moving_task = current; | |
5565 | return mem_cgroup_do_precharge(precharge); | |
5566 | } | |
5567 | ||
5568 | /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ | |
5569 | static void __mem_cgroup_clear_mc(void) | |
5570 | { | |
5571 | struct mem_cgroup *from = mc.from; | |
5572 | struct mem_cgroup *to = mc.to; | |
5573 | ||
5574 | /* we must uncharge all the leftover precharges from mc.to */ | |
5575 | if (mc.precharge) { | |
5576 | cancel_charge(mc.to, mc.precharge); | |
5577 | mc.precharge = 0; | |
5578 | } | |
5579 | /* | |
5580 | * we didn't uncharge from mc.from at mem_cgroup_move_account(), so | |
5581 | * we must uncharge here. | |
5582 | */ | |
5583 | if (mc.moved_charge) { | |
5584 | cancel_charge(mc.from, mc.moved_charge); | |
5585 | mc.moved_charge = 0; | |
5586 | } | |
5587 | /* we must fixup refcnts and charges */ | |
5588 | if (mc.moved_swap) { | |
5589 | /* uncharge swap account from the old cgroup */ | |
5590 | if (!mem_cgroup_is_root(mc.from)) | |
5591 | page_counter_uncharge(&mc.from->memsw, mc.moved_swap); | |
5592 | ||
5593 | mem_cgroup_id_put_many(mc.from, mc.moved_swap); | |
5594 | ||
5595 | /* | |
5596 | * we charged both to->memory and to->memsw, so we | |
5597 | * should uncharge to->memory. | |
5598 | */ | |
5599 | if (!mem_cgroup_is_root(mc.to)) | |
5600 | page_counter_uncharge(&mc.to->memory, mc.moved_swap); | |
5601 | ||
5602 | mem_cgroup_id_get_many(mc.to, mc.moved_swap); | |
5603 | css_put_many(&mc.to->css, mc.moved_swap); | |
5604 | ||
5605 | mc.moved_swap = 0; | |
5606 | } | |
5607 | memcg_oom_recover(from); | |
5608 | memcg_oom_recover(to); | |
5609 | wake_up_all(&mc.waitq); | |
5610 | } | |
5611 | ||
5612 | static void mem_cgroup_clear_mc(void) | |
5613 | { | |
5614 | struct mm_struct *mm = mc.mm; | |
5615 | ||
5616 | /* | |
5617 | * we must clear moving_task before waking up waiters at the end of | |
5618 | * task migration. | |
5619 | */ | |
5620 | mc.moving_task = NULL; | |
5621 | __mem_cgroup_clear_mc(); | |
5622 | spin_lock(&mc.lock); | |
5623 | mc.from = NULL; | |
5624 | mc.to = NULL; | |
5625 | mc.mm = NULL; | |
5626 | spin_unlock(&mc.lock); | |
5627 | ||
5628 | mmput(mm); | |
5629 | } | |
5630 | ||
5631 | static int mem_cgroup_can_attach(struct cgroup_taskset *tset) | |
5632 | { | |
5633 | struct cgroup_subsys_state *css; | |
5634 | struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */ | |
5635 | struct mem_cgroup *from; | |
5636 | struct task_struct *leader, *p; | |
5637 | struct mm_struct *mm; | |
5638 | unsigned long move_flags; | |
5639 | int ret = 0; | |
5640 | ||
5641 | /* charge immigration isn't supported on the default hierarchy */ | |
5642 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
5643 | return 0; | |
5644 | ||
5645 | /* | |
5646 | * Multi-process migrations only happen on the default hierarchy | |
5647 | * where charge immigration is not used. Perform charge | |
5648 | * immigration if @tset contains a leader and whine if there are | |
5649 | * multiple. | |
5650 | */ | |
5651 | p = NULL; | |
5652 | cgroup_taskset_for_each_leader(leader, css, tset) { | |
5653 | WARN_ON_ONCE(p); | |
5654 | p = leader; | |
5655 | memcg = mem_cgroup_from_css(css); | |
5656 | } | |
5657 | if (!p) | |
5658 | return 0; | |
5659 | ||
5660 | /* | |
5661 | * We are now commited to this value whatever it is. Changes in this | |
5662 | * tunable will only affect upcoming migrations, not the current one. | |
5663 | * So we need to save it, and keep it going. | |
5664 | */ | |
5665 | move_flags = READ_ONCE(memcg->move_charge_at_immigrate); | |
5666 | if (!move_flags) | |
5667 | return 0; | |
5668 | ||
5669 | from = mem_cgroup_from_task(p); | |
5670 | ||
5671 | VM_BUG_ON(from == memcg); | |
5672 | ||
5673 | mm = get_task_mm(p); | |
5674 | if (!mm) | |
5675 | return 0; | |
5676 | /* We move charges only when we move a owner of the mm */ | |
5677 | if (mm->owner == p) { | |
5678 | VM_BUG_ON(mc.from); | |
5679 | VM_BUG_ON(mc.to); | |
5680 | VM_BUG_ON(mc.precharge); | |
5681 | VM_BUG_ON(mc.moved_charge); | |
5682 | VM_BUG_ON(mc.moved_swap); | |
5683 | ||
5684 | spin_lock(&mc.lock); | |
5685 | mc.mm = mm; | |
5686 | mc.from = from; | |
5687 | mc.to = memcg; | |
5688 | mc.flags = move_flags; | |
5689 | spin_unlock(&mc.lock); | |
5690 | /* We set mc.moving_task later */ | |
5691 | ||
5692 | ret = mem_cgroup_precharge_mc(mm); | |
5693 | if (ret) | |
5694 | mem_cgroup_clear_mc(); | |
5695 | } else { | |
5696 | mmput(mm); | |
5697 | } | |
5698 | return ret; | |
5699 | } | |
5700 | ||
5701 | static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset) | |
5702 | { | |
5703 | if (mc.to) | |
5704 | mem_cgroup_clear_mc(); | |
5705 | } | |
5706 | ||
5707 | static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, | |
5708 | unsigned long addr, unsigned long end, | |
5709 | struct mm_walk *walk) | |
5710 | { | |
5711 | int ret = 0; | |
5712 | struct vm_area_struct *vma = walk->vma; | |
5713 | pte_t *pte; | |
5714 | spinlock_t *ptl; | |
5715 | enum mc_target_type target_type; | |
5716 | union mc_target target; | |
5717 | struct page *page; | |
5718 | ||
5719 | ptl = pmd_trans_huge_lock(pmd, vma); | |
5720 | if (ptl) { | |
5721 | if (mc.precharge < HPAGE_PMD_NR) { | |
5722 | spin_unlock(ptl); | |
5723 | return 0; | |
5724 | } | |
5725 | target_type = get_mctgt_type_thp(vma, addr, *pmd, &target); | |
5726 | if (target_type == MC_TARGET_PAGE) { | |
5727 | page = target.page; | |
5728 | if (!isolate_lru_page(page)) { | |
5729 | if (!mem_cgroup_move_account(page, true, | |
5730 | mc.from, mc.to)) { | |
5731 | mc.precharge -= HPAGE_PMD_NR; | |
5732 | mc.moved_charge += HPAGE_PMD_NR; | |
5733 | } | |
5734 | putback_lru_page(page); | |
5735 | } | |
5736 | put_page(page); | |
5737 | } else if (target_type == MC_TARGET_DEVICE) { | |
5738 | page = target.page; | |
5739 | if (!mem_cgroup_move_account(page, true, | |
5740 | mc.from, mc.to)) { | |
5741 | mc.precharge -= HPAGE_PMD_NR; | |
5742 | mc.moved_charge += HPAGE_PMD_NR; | |
5743 | } | |
5744 | put_page(page); | |
5745 | } | |
5746 | spin_unlock(ptl); | |
5747 | return 0; | |
5748 | } | |
5749 | ||
5750 | if (pmd_trans_unstable(pmd)) | |
5751 | return 0; | |
5752 | retry: | |
5753 | pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); | |
5754 | for (; addr != end; addr += PAGE_SIZE) { | |
5755 | pte_t ptent = *(pte++); | |
5756 | bool device = false; | |
5757 | swp_entry_t ent; | |
5758 | ||
5759 | if (!mc.precharge) | |
5760 | break; | |
5761 | ||
5762 | switch (get_mctgt_type(vma, addr, ptent, &target)) { | |
5763 | case MC_TARGET_DEVICE: | |
5764 | device = true; | |
5765 | /* fall through */ | |
5766 | case MC_TARGET_PAGE: | |
5767 | page = target.page; | |
5768 | /* | |
5769 | * We can have a part of the split pmd here. Moving it | |
5770 | * can be done but it would be too convoluted so simply | |
5771 | * ignore such a partial THP and keep it in original | |
5772 | * memcg. There should be somebody mapping the head. | |
5773 | */ | |
5774 | if (PageTransCompound(page)) | |
5775 | goto put; | |
5776 | if (!device && isolate_lru_page(page)) | |
5777 | goto put; | |
5778 | if (!mem_cgroup_move_account(page, false, | |
5779 | mc.from, mc.to)) { | |
5780 | mc.precharge--; | |
5781 | /* we uncharge from mc.from later. */ | |
5782 | mc.moved_charge++; | |
5783 | } | |
5784 | if (!device) | |
5785 | putback_lru_page(page); | |
5786 | put: /* get_mctgt_type() gets the page */ | |
5787 | put_page(page); | |
5788 | break; | |
5789 | case MC_TARGET_SWAP: | |
5790 | ent = target.ent; | |
5791 | if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) { | |
5792 | mc.precharge--; | |
5793 | /* we fixup refcnts and charges later. */ | |
5794 | mc.moved_swap++; | |
5795 | } | |
5796 | break; | |
5797 | default: | |
5798 | break; | |
5799 | } | |
5800 | } | |
5801 | pte_unmap_unlock(pte - 1, ptl); | |
5802 | cond_resched(); | |
5803 | ||
5804 | if (addr != end) { | |
5805 | /* | |
5806 | * We have consumed all precharges we got in can_attach(). | |
5807 | * We try charge one by one, but don't do any additional | |
5808 | * charges to mc.to if we have failed in charge once in attach() | |
5809 | * phase. | |
5810 | */ | |
5811 | ret = mem_cgroup_do_precharge(1); | |
5812 | if (!ret) | |
5813 | goto retry; | |
5814 | } | |
5815 | ||
5816 | return ret; | |
5817 | } | |
5818 | ||
5819 | static const struct mm_walk_ops charge_walk_ops = { | |
5820 | .pmd_entry = mem_cgroup_move_charge_pte_range, | |
5821 | }; | |
5822 | ||
5823 | static void mem_cgroup_move_charge(void) | |
5824 | { | |
5825 | lru_add_drain_all(); | |
5826 | /* | |
5827 | * Signal lock_page_memcg() to take the memcg's move_lock | |
5828 | * while we're moving its pages to another memcg. Then wait | |
5829 | * for already started RCU-only updates to finish. | |
5830 | */ | |
5831 | atomic_inc(&mc.from->moving_account); | |
5832 | synchronize_rcu(); | |
5833 | retry: | |
5834 | if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) { | |
5835 | /* | |
5836 | * Someone who are holding the mmap_sem might be waiting in | |
5837 | * waitq. So we cancel all extra charges, wake up all waiters, | |
5838 | * and retry. Because we cancel precharges, we might not be able | |
5839 | * to move enough charges, but moving charge is a best-effort | |
5840 | * feature anyway, so it wouldn't be a big problem. | |
5841 | */ | |
5842 | __mem_cgroup_clear_mc(); | |
5843 | cond_resched(); | |
5844 | goto retry; | |
5845 | } | |
5846 | /* | |
5847 | * When we have consumed all precharges and failed in doing | |
5848 | * additional charge, the page walk just aborts. | |
5849 | */ | |
5850 | walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops, | |
5851 | NULL); | |
5852 | ||
5853 | up_read(&mc.mm->mmap_sem); | |
5854 | atomic_dec(&mc.from->moving_account); | |
5855 | } | |
5856 | ||
5857 | static void mem_cgroup_move_task(void) | |
5858 | { | |
5859 | if (mc.to) { | |
5860 | mem_cgroup_move_charge(); | |
5861 | mem_cgroup_clear_mc(); | |
5862 | } | |
5863 | } | |
5864 | #else /* !CONFIG_MMU */ | |
5865 | static int mem_cgroup_can_attach(struct cgroup_taskset *tset) | |
5866 | { | |
5867 | return 0; | |
5868 | } | |
5869 | static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset) | |
5870 | { | |
5871 | } | |
5872 | static void mem_cgroup_move_task(void) | |
5873 | { | |
5874 | } | |
5875 | #endif | |
5876 | ||
5877 | /* | |
5878 | * Cgroup retains root cgroups across [un]mount cycles making it necessary | |
5879 | * to verify whether we're attached to the default hierarchy on each mount | |
5880 | * attempt. | |
5881 | */ | |
5882 | static void mem_cgroup_bind(struct cgroup_subsys_state *root_css) | |
5883 | { | |
5884 | /* | |
5885 | * use_hierarchy is forced on the default hierarchy. cgroup core | |
5886 | * guarantees that @root doesn't have any children, so turning it | |
5887 | * on for the root memcg is enough. | |
5888 | */ | |
5889 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
5890 | root_mem_cgroup->use_hierarchy = true; | |
5891 | else | |
5892 | root_mem_cgroup->use_hierarchy = false; | |
5893 | } | |
5894 | ||
5895 | static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value) | |
5896 | { | |
5897 | if (value == PAGE_COUNTER_MAX) | |
5898 | seq_puts(m, "max\n"); | |
5899 | else | |
5900 | seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE); | |
5901 | ||
5902 | return 0; | |
5903 | } | |
5904 | ||
5905 | static u64 memory_current_read(struct cgroup_subsys_state *css, | |
5906 | struct cftype *cft) | |
5907 | { | |
5908 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
5909 | ||
5910 | return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE; | |
5911 | } | |
5912 | ||
5913 | static int memory_min_show(struct seq_file *m, void *v) | |
5914 | { | |
5915 | return seq_puts_memcg_tunable(m, | |
5916 | READ_ONCE(mem_cgroup_from_seq(m)->memory.min)); | |
5917 | } | |
5918 | ||
5919 | static ssize_t memory_min_write(struct kernfs_open_file *of, | |
5920 | char *buf, size_t nbytes, loff_t off) | |
5921 | { | |
5922 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
5923 | unsigned long min; | |
5924 | int err; | |
5925 | ||
5926 | buf = strstrip(buf); | |
5927 | err = page_counter_memparse(buf, "max", &min); | |
5928 | if (err) | |
5929 | return err; | |
5930 | ||
5931 | page_counter_set_min(&memcg->memory, min); | |
5932 | ||
5933 | return nbytes; | |
5934 | } | |
5935 | ||
5936 | static int memory_low_show(struct seq_file *m, void *v) | |
5937 | { | |
5938 | return seq_puts_memcg_tunable(m, | |
5939 | READ_ONCE(mem_cgroup_from_seq(m)->memory.low)); | |
5940 | } | |
5941 | ||
5942 | static ssize_t memory_low_write(struct kernfs_open_file *of, | |
5943 | char *buf, size_t nbytes, loff_t off) | |
5944 | { | |
5945 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
5946 | unsigned long low; | |
5947 | int err; | |
5948 | ||
5949 | buf = strstrip(buf); | |
5950 | err = page_counter_memparse(buf, "max", &low); | |
5951 | if (err) | |
5952 | return err; | |
5953 | ||
5954 | page_counter_set_low(&memcg->memory, low); | |
5955 | ||
5956 | return nbytes; | |
5957 | } | |
5958 | ||
5959 | static int memory_high_show(struct seq_file *m, void *v) | |
5960 | { | |
5961 | return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high)); | |
5962 | } | |
5963 | ||
5964 | static ssize_t memory_high_write(struct kernfs_open_file *of, | |
5965 | char *buf, size_t nbytes, loff_t off) | |
5966 | { | |
5967 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
5968 | unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; | |
5969 | bool drained = false; | |
5970 | unsigned long high; | |
5971 | int err; | |
5972 | ||
5973 | buf = strstrip(buf); | |
5974 | err = page_counter_memparse(buf, "max", &high); | |
5975 | if (err) | |
5976 | return err; | |
5977 | ||
5978 | memcg->high = high; | |
5979 | ||
5980 | for (;;) { | |
5981 | unsigned long nr_pages = page_counter_read(&memcg->memory); | |
5982 | unsigned long reclaimed; | |
5983 | ||
5984 | if (nr_pages <= high) | |
5985 | break; | |
5986 | ||
5987 | if (signal_pending(current)) | |
5988 | break; | |
5989 | ||
5990 | if (!drained) { | |
5991 | drain_all_stock(memcg); | |
5992 | drained = true; | |
5993 | continue; | |
5994 | } | |
5995 | ||
5996 | reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high, | |
5997 | GFP_KERNEL, true); | |
5998 | ||
5999 | if (!reclaimed && !nr_retries--) | |
6000 | break; | |
6001 | } | |
6002 | ||
6003 | return nbytes; | |
6004 | } | |
6005 | ||
6006 | static int memory_max_show(struct seq_file *m, void *v) | |
6007 | { | |
6008 | return seq_puts_memcg_tunable(m, | |
6009 | READ_ONCE(mem_cgroup_from_seq(m)->memory.max)); | |
6010 | } | |
6011 | ||
6012 | static ssize_t memory_max_write(struct kernfs_open_file *of, | |
6013 | char *buf, size_t nbytes, loff_t off) | |
6014 | { | |
6015 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
6016 | unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES; | |
6017 | bool drained = false; | |
6018 | unsigned long max; | |
6019 | int err; | |
6020 | ||
6021 | buf = strstrip(buf); | |
6022 | err = page_counter_memparse(buf, "max", &max); | |
6023 | if (err) | |
6024 | return err; | |
6025 | ||
6026 | xchg(&memcg->memory.max, max); | |
6027 | ||
6028 | for (;;) { | |
6029 | unsigned long nr_pages = page_counter_read(&memcg->memory); | |
6030 | ||
6031 | if (nr_pages <= max) | |
6032 | break; | |
6033 | ||
6034 | if (signal_pending(current)) | |
6035 | break; | |
6036 | ||
6037 | if (!drained) { | |
6038 | drain_all_stock(memcg); | |
6039 | drained = true; | |
6040 | continue; | |
6041 | } | |
6042 | ||
6043 | if (nr_reclaims) { | |
6044 | if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max, | |
6045 | GFP_KERNEL, true)) | |
6046 | nr_reclaims--; | |
6047 | continue; | |
6048 | } | |
6049 | ||
6050 | memcg_memory_event(memcg, MEMCG_OOM); | |
6051 | if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0)) | |
6052 | break; | |
6053 | } | |
6054 | ||
6055 | memcg_wb_domain_size_changed(memcg); | |
6056 | return nbytes; | |
6057 | } | |
6058 | ||
6059 | static void __memory_events_show(struct seq_file *m, atomic_long_t *events) | |
6060 | { | |
6061 | seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW])); | |
6062 | seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH])); | |
6063 | seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX])); | |
6064 | seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM])); | |
6065 | seq_printf(m, "oom_kill %lu\n", | |
6066 | atomic_long_read(&events[MEMCG_OOM_KILL])); | |
6067 | } | |
6068 | ||
6069 | static int memory_events_show(struct seq_file *m, void *v) | |
6070 | { | |
6071 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
6072 | ||
6073 | __memory_events_show(m, memcg->memory_events); | |
6074 | return 0; | |
6075 | } | |
6076 | ||
6077 | static int memory_events_local_show(struct seq_file *m, void *v) | |
6078 | { | |
6079 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
6080 | ||
6081 | __memory_events_show(m, memcg->memory_events_local); | |
6082 | return 0; | |
6083 | } | |
6084 | ||
6085 | static int memory_stat_show(struct seq_file *m, void *v) | |
6086 | { | |
6087 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
6088 | char *buf; | |
6089 | ||
6090 | buf = memory_stat_format(memcg); | |
6091 | if (!buf) | |
6092 | return -ENOMEM; | |
6093 | seq_puts(m, buf); | |
6094 | kfree(buf); | |
6095 | return 0; | |
6096 | } | |
6097 | ||
6098 | static int memory_oom_group_show(struct seq_file *m, void *v) | |
6099 | { | |
6100 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
6101 | ||
6102 | seq_printf(m, "%d\n", memcg->oom_group); | |
6103 | ||
6104 | return 0; | |
6105 | } | |
6106 | ||
6107 | static ssize_t memory_oom_group_write(struct kernfs_open_file *of, | |
6108 | char *buf, size_t nbytes, loff_t off) | |
6109 | { | |
6110 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
6111 | int ret, oom_group; | |
6112 | ||
6113 | buf = strstrip(buf); | |
6114 | if (!buf) | |
6115 | return -EINVAL; | |
6116 | ||
6117 | ret = kstrtoint(buf, 0, &oom_group); | |
6118 | if (ret) | |
6119 | return ret; | |
6120 | ||
6121 | if (oom_group != 0 && oom_group != 1) | |
6122 | return -EINVAL; | |
6123 | ||
6124 | memcg->oom_group = oom_group; | |
6125 | ||
6126 | return nbytes; | |
6127 | } | |
6128 | ||
6129 | static struct cftype memory_files[] = { | |
6130 | { | |
6131 | .name = "current", | |
6132 | .flags = CFTYPE_NOT_ON_ROOT, | |
6133 | .read_u64 = memory_current_read, | |
6134 | }, | |
6135 | { | |
6136 | .name = "min", | |
6137 | .flags = CFTYPE_NOT_ON_ROOT, | |
6138 | .seq_show = memory_min_show, | |
6139 | .write = memory_min_write, | |
6140 | }, | |
6141 | { | |
6142 | .name = "low", | |
6143 | .flags = CFTYPE_NOT_ON_ROOT, | |
6144 | .seq_show = memory_low_show, | |
6145 | .write = memory_low_write, | |
6146 | }, | |
6147 | { | |
6148 | .name = "high", | |
6149 | .flags = CFTYPE_NOT_ON_ROOT, | |
6150 | .seq_show = memory_high_show, | |
6151 | .write = memory_high_write, | |
6152 | }, | |
6153 | { | |
6154 | .name = "max", | |
6155 | .flags = CFTYPE_NOT_ON_ROOT, | |
6156 | .seq_show = memory_max_show, | |
6157 | .write = memory_max_write, | |
6158 | }, | |
6159 | { | |
6160 | .name = "events", | |
6161 | .flags = CFTYPE_NOT_ON_ROOT, | |
6162 | .file_offset = offsetof(struct mem_cgroup, events_file), | |
6163 | .seq_show = memory_events_show, | |
6164 | }, | |
6165 | { | |
6166 | .name = "events.local", | |
6167 | .flags = CFTYPE_NOT_ON_ROOT, | |
6168 | .file_offset = offsetof(struct mem_cgroup, events_local_file), | |
6169 | .seq_show = memory_events_local_show, | |
6170 | }, | |
6171 | { | |
6172 | .name = "stat", | |
6173 | .flags = CFTYPE_NOT_ON_ROOT, | |
6174 | .seq_show = memory_stat_show, | |
6175 | }, | |
6176 | { | |
6177 | .name = "oom.group", | |
6178 | .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE, | |
6179 | .seq_show = memory_oom_group_show, | |
6180 | .write = memory_oom_group_write, | |
6181 | }, | |
6182 | { } /* terminate */ | |
6183 | }; | |
6184 | ||
6185 | struct cgroup_subsys memory_cgrp_subsys = { | |
6186 | .css_alloc = mem_cgroup_css_alloc, | |
6187 | .css_online = mem_cgroup_css_online, | |
6188 | .css_offline = mem_cgroup_css_offline, | |
6189 | .css_released = mem_cgroup_css_released, | |
6190 | .css_free = mem_cgroup_css_free, | |
6191 | .css_reset = mem_cgroup_css_reset, | |
6192 | .can_attach = mem_cgroup_can_attach, | |
6193 | .cancel_attach = mem_cgroup_cancel_attach, | |
6194 | .post_attach = mem_cgroup_move_task, | |
6195 | .bind = mem_cgroup_bind, | |
6196 | .dfl_cftypes = memory_files, | |
6197 | .legacy_cftypes = mem_cgroup_legacy_files, | |
6198 | .early_init = 0, | |
6199 | }; | |
6200 | ||
6201 | /** | |
6202 | * mem_cgroup_protected - check if memory consumption is in the normal range | |
6203 | * @root: the top ancestor of the sub-tree being checked | |
6204 | * @memcg: the memory cgroup to check | |
6205 | * | |
6206 | * WARNING: This function is not stateless! It can only be used as part | |
6207 | * of a top-down tree iteration, not for isolated queries. | |
6208 | * | |
6209 | * Returns one of the following: | |
6210 | * MEMCG_PROT_NONE: cgroup memory is not protected | |
6211 | * MEMCG_PROT_LOW: cgroup memory is protected as long there is | |
6212 | * an unprotected supply of reclaimable memory from other cgroups. | |
6213 | * MEMCG_PROT_MIN: cgroup memory is protected | |
6214 | * | |
6215 | * @root is exclusive; it is never protected when looked at directly | |
6216 | * | |
6217 | * To provide a proper hierarchical behavior, effective memory.min/low values | |
6218 | * are used. Below is the description of how effective memory.low is calculated. | |
6219 | * Effective memory.min values is calculated in the same way. | |
6220 | * | |
6221 | * Effective memory.low is always equal or less than the original memory.low. | |
6222 | * If there is no memory.low overcommittment (which is always true for | |
6223 | * top-level memory cgroups), these two values are equal. | |
6224 | * Otherwise, it's a part of parent's effective memory.low, | |
6225 | * calculated as a cgroup's memory.low usage divided by sum of sibling's | |
6226 | * memory.low usages, where memory.low usage is the size of actually | |
6227 | * protected memory. | |
6228 | * | |
6229 | * low_usage | |
6230 | * elow = min( memory.low, parent->elow * ------------------ ), | |
6231 | * siblings_low_usage | |
6232 | * | |
6233 | * | memory.current, if memory.current < memory.low | |
6234 | * low_usage = | | |
6235 | * | 0, otherwise. | |
6236 | * | |
6237 | * | |
6238 | * Such definition of the effective memory.low provides the expected | |
6239 | * hierarchical behavior: parent's memory.low value is limiting | |
6240 | * children, unprotected memory is reclaimed first and cgroups, | |
6241 | * which are not using their guarantee do not affect actual memory | |
6242 | * distribution. | |
6243 | * | |
6244 | * For example, if there are memcgs A, A/B, A/C, A/D and A/E: | |
6245 | * | |
6246 | * A A/memory.low = 2G, A/memory.current = 6G | |
6247 | * //\\ | |
6248 | * BC DE B/memory.low = 3G B/memory.current = 2G | |
6249 | * C/memory.low = 1G C/memory.current = 2G | |
6250 | * D/memory.low = 0 D/memory.current = 2G | |
6251 | * E/memory.low = 10G E/memory.current = 0 | |
6252 | * | |
6253 | * and the memory pressure is applied, the following memory distribution | |
6254 | * is expected (approximately): | |
6255 | * | |
6256 | * A/memory.current = 2G | |
6257 | * | |
6258 | * B/memory.current = 1.3G | |
6259 | * C/memory.current = 0.6G | |
6260 | * D/memory.current = 0 | |
6261 | * E/memory.current = 0 | |
6262 | * | |
6263 | * These calculations require constant tracking of the actual low usages | |
6264 | * (see propagate_protected_usage()), as well as recursive calculation of | |
6265 | * effective memory.low values. But as we do call mem_cgroup_protected() | |
6266 | * path for each memory cgroup top-down from the reclaim, | |
6267 | * it's possible to optimize this part, and save calculated elow | |
6268 | * for next usage. This part is intentionally racy, but it's ok, | |
6269 | * as memory.low is a best-effort mechanism. | |
6270 | */ | |
6271 | enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root, | |
6272 | struct mem_cgroup *memcg) | |
6273 | { | |
6274 | struct mem_cgroup *parent; | |
6275 | unsigned long emin, parent_emin; | |
6276 | unsigned long elow, parent_elow; | |
6277 | unsigned long usage; | |
6278 | ||
6279 | if (mem_cgroup_disabled()) | |
6280 | return MEMCG_PROT_NONE; | |
6281 | ||
6282 | if (!root) | |
6283 | root = root_mem_cgroup; | |
6284 | if (memcg == root) | |
6285 | return MEMCG_PROT_NONE; | |
6286 | ||
6287 | usage = page_counter_read(&memcg->memory); | |
6288 | if (!usage) | |
6289 | return MEMCG_PROT_NONE; | |
6290 | ||
6291 | emin = memcg->memory.min; | |
6292 | elow = memcg->memory.low; | |
6293 | ||
6294 | parent = parent_mem_cgroup(memcg); | |
6295 | /* No parent means a non-hierarchical mode on v1 memcg */ | |
6296 | if (!parent) | |
6297 | return MEMCG_PROT_NONE; | |
6298 | ||
6299 | if (parent == root) | |
6300 | goto exit; | |
6301 | ||
6302 | parent_emin = READ_ONCE(parent->memory.emin); | |
6303 | emin = min(emin, parent_emin); | |
6304 | if (emin && parent_emin) { | |
6305 | unsigned long min_usage, siblings_min_usage; | |
6306 | ||
6307 | min_usage = min(usage, memcg->memory.min); | |
6308 | siblings_min_usage = atomic_long_read( | |
6309 | &parent->memory.children_min_usage); | |
6310 | ||
6311 | if (min_usage && siblings_min_usage) | |
6312 | emin = min(emin, parent_emin * min_usage / | |
6313 | siblings_min_usage); | |
6314 | } | |
6315 | ||
6316 | parent_elow = READ_ONCE(parent->memory.elow); | |
6317 | elow = min(elow, parent_elow); | |
6318 | if (elow && parent_elow) { | |
6319 | unsigned long low_usage, siblings_low_usage; | |
6320 | ||
6321 | low_usage = min(usage, memcg->memory.low); | |
6322 | siblings_low_usage = atomic_long_read( | |
6323 | &parent->memory.children_low_usage); | |
6324 | ||
6325 | if (low_usage && siblings_low_usage) | |
6326 | elow = min(elow, parent_elow * low_usage / | |
6327 | siblings_low_usage); | |
6328 | } | |
6329 | ||
6330 | exit: | |
6331 | memcg->memory.emin = emin; | |
6332 | memcg->memory.elow = elow; | |
6333 | ||
6334 | if (usage <= emin) | |
6335 | return MEMCG_PROT_MIN; | |
6336 | else if (usage <= elow) | |
6337 | return MEMCG_PROT_LOW; | |
6338 | else | |
6339 | return MEMCG_PROT_NONE; | |
6340 | } | |
6341 | ||
6342 | /** | |
6343 | * mem_cgroup_try_charge - try charging a page | |
6344 | * @page: page to charge | |
6345 | * @mm: mm context of the victim | |
6346 | * @gfp_mask: reclaim mode | |
6347 | * @memcgp: charged memcg return | |
6348 | * @compound: charge the page as compound or small page | |
6349 | * | |
6350 | * Try to charge @page to the memcg that @mm belongs to, reclaiming | |
6351 | * pages according to @gfp_mask if necessary. | |
6352 | * | |
6353 | * Returns 0 on success, with *@memcgp pointing to the charged memcg. | |
6354 | * Otherwise, an error code is returned. | |
6355 | * | |
6356 | * After page->mapping has been set up, the caller must finalize the | |
6357 | * charge with mem_cgroup_commit_charge(). Or abort the transaction | |
6358 | * with mem_cgroup_cancel_charge() in case page instantiation fails. | |
6359 | */ | |
6360 | int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm, | |
6361 | gfp_t gfp_mask, struct mem_cgroup **memcgp, | |
6362 | bool compound) | |
6363 | { | |
6364 | struct mem_cgroup *memcg = NULL; | |
6365 | unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1; | |
6366 | int ret = 0; | |
6367 | ||
6368 | if (mem_cgroup_disabled()) | |
6369 | goto out; | |
6370 | ||
6371 | if (PageSwapCache(page)) { | |
6372 | /* | |
6373 | * Every swap fault against a single page tries to charge the | |
6374 | * page, bail as early as possible. shmem_unuse() encounters | |
6375 | * already charged pages, too. The USED bit is protected by | |
6376 | * the page lock, which serializes swap cache removal, which | |
6377 | * in turn serializes uncharging. | |
6378 | */ | |
6379 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
6380 | if (compound_head(page)->mem_cgroup) | |
6381 | goto out; | |
6382 | ||
6383 | if (do_swap_account) { | |
6384 | swp_entry_t ent = { .val = page_private(page), }; | |
6385 | unsigned short id = lookup_swap_cgroup_id(ent); | |
6386 | ||
6387 | rcu_read_lock(); | |
6388 | memcg = mem_cgroup_from_id(id); | |
6389 | if (memcg && !css_tryget_online(&memcg->css)) | |
6390 | memcg = NULL; | |
6391 | rcu_read_unlock(); | |
6392 | } | |
6393 | } | |
6394 | ||
6395 | if (!memcg) | |
6396 | memcg = get_mem_cgroup_from_mm(mm); | |
6397 | ||
6398 | ret = try_charge(memcg, gfp_mask, nr_pages); | |
6399 | ||
6400 | css_put(&memcg->css); | |
6401 | out: | |
6402 | *memcgp = memcg; | |
6403 | return ret; | |
6404 | } | |
6405 | ||
6406 | int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm, | |
6407 | gfp_t gfp_mask, struct mem_cgroup **memcgp, | |
6408 | bool compound) | |
6409 | { | |
6410 | struct mem_cgroup *memcg; | |
6411 | int ret; | |
6412 | ||
6413 | ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound); | |
6414 | memcg = *memcgp; | |
6415 | mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask); | |
6416 | return ret; | |
6417 | } | |
6418 | ||
6419 | /** | |
6420 | * mem_cgroup_commit_charge - commit a page charge | |
6421 | * @page: page to charge | |
6422 | * @memcg: memcg to charge the page to | |
6423 | * @lrucare: page might be on LRU already | |
6424 | * @compound: charge the page as compound or small page | |
6425 | * | |
6426 | * Finalize a charge transaction started by mem_cgroup_try_charge(), | |
6427 | * after page->mapping has been set up. This must happen atomically | |
6428 | * as part of the page instantiation, i.e. under the page table lock | |
6429 | * for anonymous pages, under the page lock for page and swap cache. | |
6430 | * | |
6431 | * In addition, the page must not be on the LRU during the commit, to | |
6432 | * prevent racing with task migration. If it might be, use @lrucare. | |
6433 | * | |
6434 | * Use mem_cgroup_cancel_charge() to cancel the transaction instead. | |
6435 | */ | |
6436 | void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg, | |
6437 | bool lrucare, bool compound) | |
6438 | { | |
6439 | unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1; | |
6440 | ||
6441 | VM_BUG_ON_PAGE(!page->mapping, page); | |
6442 | VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page); | |
6443 | ||
6444 | if (mem_cgroup_disabled()) | |
6445 | return; | |
6446 | /* | |
6447 | * Swap faults will attempt to charge the same page multiple | |
6448 | * times. But reuse_swap_page() might have removed the page | |
6449 | * from swapcache already, so we can't check PageSwapCache(). | |
6450 | */ | |
6451 | if (!memcg) | |
6452 | return; | |
6453 | ||
6454 | commit_charge(page, memcg, lrucare); | |
6455 | ||
6456 | local_irq_disable(); | |
6457 | mem_cgroup_charge_statistics(memcg, page, compound, nr_pages); | |
6458 | memcg_check_events(memcg, page); | |
6459 | local_irq_enable(); | |
6460 | ||
6461 | if (do_memsw_account() && PageSwapCache(page)) { | |
6462 | swp_entry_t entry = { .val = page_private(page) }; | |
6463 | /* | |
6464 | * The swap entry might not get freed for a long time, | |
6465 | * let's not wait for it. The page already received a | |
6466 | * memory+swap charge, drop the swap entry duplicate. | |
6467 | */ | |
6468 | mem_cgroup_uncharge_swap(entry, nr_pages); | |
6469 | } | |
6470 | } | |
6471 | ||
6472 | /** | |
6473 | * mem_cgroup_cancel_charge - cancel a page charge | |
6474 | * @page: page to charge | |
6475 | * @memcg: memcg to charge the page to | |
6476 | * @compound: charge the page as compound or small page | |
6477 | * | |
6478 | * Cancel a charge transaction started by mem_cgroup_try_charge(). | |
6479 | */ | |
6480 | void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg, | |
6481 | bool compound) | |
6482 | { | |
6483 | unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1; | |
6484 | ||
6485 | if (mem_cgroup_disabled()) | |
6486 | return; | |
6487 | /* | |
6488 | * Swap faults will attempt to charge the same page multiple | |
6489 | * times. But reuse_swap_page() might have removed the page | |
6490 | * from swapcache already, so we can't check PageSwapCache(). | |
6491 | */ | |
6492 | if (!memcg) | |
6493 | return; | |
6494 | ||
6495 | cancel_charge(memcg, nr_pages); | |
6496 | } | |
6497 | ||
6498 | struct uncharge_gather { | |
6499 | struct mem_cgroup *memcg; | |
6500 | unsigned long pgpgout; | |
6501 | unsigned long nr_anon; | |
6502 | unsigned long nr_file; | |
6503 | unsigned long nr_kmem; | |
6504 | unsigned long nr_huge; | |
6505 | unsigned long nr_shmem; | |
6506 | struct page *dummy_page; | |
6507 | }; | |
6508 | ||
6509 | static inline void uncharge_gather_clear(struct uncharge_gather *ug) | |
6510 | { | |
6511 | memset(ug, 0, sizeof(*ug)); | |
6512 | } | |
6513 | ||
6514 | static void uncharge_batch(const struct uncharge_gather *ug) | |
6515 | { | |
6516 | unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem; | |
6517 | unsigned long flags; | |
6518 | ||
6519 | if (!mem_cgroup_is_root(ug->memcg)) { | |
6520 | page_counter_uncharge(&ug->memcg->memory, nr_pages); | |
6521 | if (do_memsw_account()) | |
6522 | page_counter_uncharge(&ug->memcg->memsw, nr_pages); | |
6523 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem) | |
6524 | page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem); | |
6525 | memcg_oom_recover(ug->memcg); | |
6526 | } | |
6527 | ||
6528 | local_irq_save(flags); | |
6529 | __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon); | |
6530 | __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file); | |
6531 | __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge); | |
6532 | __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem); | |
6533 | __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout); | |
6534 | __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages); | |
6535 | memcg_check_events(ug->memcg, ug->dummy_page); | |
6536 | local_irq_restore(flags); | |
6537 | ||
6538 | if (!mem_cgroup_is_root(ug->memcg)) | |
6539 | css_put_many(&ug->memcg->css, nr_pages); | |
6540 | } | |
6541 | ||
6542 | static void uncharge_page(struct page *page, struct uncharge_gather *ug) | |
6543 | { | |
6544 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
6545 | VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) && | |
6546 | !PageHWPoison(page) , page); | |
6547 | ||
6548 | if (!page->mem_cgroup) | |
6549 | return; | |
6550 | ||
6551 | /* | |
6552 | * Nobody should be changing or seriously looking at | |
6553 | * page->mem_cgroup at this point, we have fully | |
6554 | * exclusive access to the page. | |
6555 | */ | |
6556 | ||
6557 | if (ug->memcg != page->mem_cgroup) { | |
6558 | if (ug->memcg) { | |
6559 | uncharge_batch(ug); | |
6560 | uncharge_gather_clear(ug); | |
6561 | } | |
6562 | ug->memcg = page->mem_cgroup; | |
6563 | } | |
6564 | ||
6565 | if (!PageKmemcg(page)) { | |
6566 | unsigned int nr_pages = 1; | |
6567 | ||
6568 | if (PageTransHuge(page)) { | |
6569 | nr_pages = compound_nr(page); | |
6570 | ug->nr_huge += nr_pages; | |
6571 | } | |
6572 | if (PageAnon(page)) | |
6573 | ug->nr_anon += nr_pages; | |
6574 | else { | |
6575 | ug->nr_file += nr_pages; | |
6576 | if (PageSwapBacked(page)) | |
6577 | ug->nr_shmem += nr_pages; | |
6578 | } | |
6579 | ug->pgpgout++; | |
6580 | } else { | |
6581 | ug->nr_kmem += compound_nr(page); | |
6582 | __ClearPageKmemcg(page); | |
6583 | } | |
6584 | ||
6585 | ug->dummy_page = page; | |
6586 | page->mem_cgroup = NULL; | |
6587 | } | |
6588 | ||
6589 | static void uncharge_list(struct list_head *page_list) | |
6590 | { | |
6591 | struct uncharge_gather ug; | |
6592 | struct list_head *next; | |
6593 | ||
6594 | uncharge_gather_clear(&ug); | |
6595 | ||
6596 | /* | |
6597 | * Note that the list can be a single page->lru; hence the | |
6598 | * do-while loop instead of a simple list_for_each_entry(). | |
6599 | */ | |
6600 | next = page_list->next; | |
6601 | do { | |
6602 | struct page *page; | |
6603 | ||
6604 | page = list_entry(next, struct page, lru); | |
6605 | next = page->lru.next; | |
6606 | ||
6607 | uncharge_page(page, &ug); | |
6608 | } while (next != page_list); | |
6609 | ||
6610 | if (ug.memcg) | |
6611 | uncharge_batch(&ug); | |
6612 | } | |
6613 | ||
6614 | /** | |
6615 | * mem_cgroup_uncharge - uncharge a page | |
6616 | * @page: page to uncharge | |
6617 | * | |
6618 | * Uncharge a page previously charged with mem_cgroup_try_charge() and | |
6619 | * mem_cgroup_commit_charge(). | |
6620 | */ | |
6621 | void mem_cgroup_uncharge(struct page *page) | |
6622 | { | |
6623 | struct uncharge_gather ug; | |
6624 | ||
6625 | if (mem_cgroup_disabled()) | |
6626 | return; | |
6627 | ||
6628 | /* Don't touch page->lru of any random page, pre-check: */ | |
6629 | if (!page->mem_cgroup) | |
6630 | return; | |
6631 | ||
6632 | uncharge_gather_clear(&ug); | |
6633 | uncharge_page(page, &ug); | |
6634 | uncharge_batch(&ug); | |
6635 | } | |
6636 | ||
6637 | /** | |
6638 | * mem_cgroup_uncharge_list - uncharge a list of page | |
6639 | * @page_list: list of pages to uncharge | |
6640 | * | |
6641 | * Uncharge a list of pages previously charged with | |
6642 | * mem_cgroup_try_charge() and mem_cgroup_commit_charge(). | |
6643 | */ | |
6644 | void mem_cgroup_uncharge_list(struct list_head *page_list) | |
6645 | { | |
6646 | if (mem_cgroup_disabled()) | |
6647 | return; | |
6648 | ||
6649 | if (!list_empty(page_list)) | |
6650 | uncharge_list(page_list); | |
6651 | } | |
6652 | ||
6653 | /** | |
6654 | * mem_cgroup_migrate - charge a page's replacement | |
6655 | * @oldpage: currently circulating page | |
6656 | * @newpage: replacement page | |
6657 | * | |
6658 | * Charge @newpage as a replacement page for @oldpage. @oldpage will | |
6659 | * be uncharged upon free. | |
6660 | * | |
6661 | * Both pages must be locked, @newpage->mapping must be set up. | |
6662 | */ | |
6663 | void mem_cgroup_migrate(struct page *oldpage, struct page *newpage) | |
6664 | { | |
6665 | struct mem_cgroup *memcg; | |
6666 | unsigned int nr_pages; | |
6667 | unsigned long flags; | |
6668 | ||
6669 | VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage); | |
6670 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); | |
6671 | VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage); | |
6672 | VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage), | |
6673 | newpage); | |
6674 | ||
6675 | if (mem_cgroup_disabled()) | |
6676 | return; | |
6677 | ||
6678 | /* Page cache replacement: new page already charged? */ | |
6679 | if (newpage->mem_cgroup) | |
6680 | return; | |
6681 | ||
6682 | /* Swapcache readahead pages can get replaced before being charged */ | |
6683 | memcg = oldpage->mem_cgroup; | |
6684 | if (!memcg) | |
6685 | return; | |
6686 | ||
6687 | /* Force-charge the new page. The old one will be freed soon */ | |
6688 | nr_pages = hpage_nr_pages(newpage); | |
6689 | ||
6690 | page_counter_charge(&memcg->memory, nr_pages); | |
6691 | if (do_memsw_account()) | |
6692 | page_counter_charge(&memcg->memsw, nr_pages); | |
6693 | css_get_many(&memcg->css, nr_pages); | |
6694 | ||
6695 | commit_charge(newpage, memcg, false); | |
6696 | ||
6697 | local_irq_save(flags); | |
6698 | mem_cgroup_charge_statistics(memcg, newpage, PageTransHuge(newpage), | |
6699 | nr_pages); | |
6700 | memcg_check_events(memcg, newpage); | |
6701 | local_irq_restore(flags); | |
6702 | } | |
6703 | ||
6704 | DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key); | |
6705 | EXPORT_SYMBOL(memcg_sockets_enabled_key); | |
6706 | ||
6707 | void mem_cgroup_sk_alloc(struct sock *sk) | |
6708 | { | |
6709 | struct mem_cgroup *memcg; | |
6710 | ||
6711 | if (!mem_cgroup_sockets_enabled) | |
6712 | return; | |
6713 | ||
6714 | /* Do not associate the sock with unrelated interrupted task's memcg. */ | |
6715 | if (in_interrupt()) | |
6716 | return; | |
6717 | ||
6718 | rcu_read_lock(); | |
6719 | memcg = mem_cgroup_from_task(current); | |
6720 | if (memcg == root_mem_cgroup) | |
6721 | goto out; | |
6722 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active) | |
6723 | goto out; | |
6724 | if (css_tryget_online(&memcg->css)) | |
6725 | sk->sk_memcg = memcg; | |
6726 | out: | |
6727 | rcu_read_unlock(); | |
6728 | } | |
6729 | ||
6730 | void mem_cgroup_sk_free(struct sock *sk) | |
6731 | { | |
6732 | if (sk->sk_memcg) | |
6733 | css_put(&sk->sk_memcg->css); | |
6734 | } | |
6735 | ||
6736 | /** | |
6737 | * mem_cgroup_charge_skmem - charge socket memory | |
6738 | * @memcg: memcg to charge | |
6739 | * @nr_pages: number of pages to charge | |
6740 | * | |
6741 | * Charges @nr_pages to @memcg. Returns %true if the charge fit within | |
6742 | * @memcg's configured limit, %false if the charge had to be forced. | |
6743 | */ | |
6744 | bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages) | |
6745 | { | |
6746 | gfp_t gfp_mask = GFP_KERNEL; | |
6747 | ||
6748 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) { | |
6749 | struct page_counter *fail; | |
6750 | ||
6751 | if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) { | |
6752 | memcg->tcpmem_pressure = 0; | |
6753 | return true; | |
6754 | } | |
6755 | page_counter_charge(&memcg->tcpmem, nr_pages); | |
6756 | memcg->tcpmem_pressure = 1; | |
6757 | return false; | |
6758 | } | |
6759 | ||
6760 | /* Don't block in the packet receive path */ | |
6761 | if (in_softirq()) | |
6762 | gfp_mask = GFP_NOWAIT; | |
6763 | ||
6764 | mod_memcg_state(memcg, MEMCG_SOCK, nr_pages); | |
6765 | ||
6766 | if (try_charge(memcg, gfp_mask, nr_pages) == 0) | |
6767 | return true; | |
6768 | ||
6769 | try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages); | |
6770 | return false; | |
6771 | } | |
6772 | ||
6773 | /** | |
6774 | * mem_cgroup_uncharge_skmem - uncharge socket memory | |
6775 | * @memcg: memcg to uncharge | |
6776 | * @nr_pages: number of pages to uncharge | |
6777 | */ | |
6778 | void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages) | |
6779 | { | |
6780 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) { | |
6781 | page_counter_uncharge(&memcg->tcpmem, nr_pages); | |
6782 | return; | |
6783 | } | |
6784 | ||
6785 | mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages); | |
6786 | ||
6787 | refill_stock(memcg, nr_pages); | |
6788 | } | |
6789 | ||
6790 | static int __init cgroup_memory(char *s) | |
6791 | { | |
6792 | char *token; | |
6793 | ||
6794 | while ((token = strsep(&s, ",")) != NULL) { | |
6795 | if (!*token) | |
6796 | continue; | |
6797 | if (!strcmp(token, "nosocket")) | |
6798 | cgroup_memory_nosocket = true; | |
6799 | if (!strcmp(token, "nokmem")) | |
6800 | cgroup_memory_nokmem = true; | |
6801 | } | |
6802 | return 0; | |
6803 | } | |
6804 | __setup("cgroup.memory=", cgroup_memory); | |
6805 | ||
6806 | /* | |
6807 | * subsys_initcall() for memory controller. | |
6808 | * | |
6809 | * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this | |
6810 | * context because of lock dependencies (cgroup_lock -> cpu hotplug) but | |
6811 | * basically everything that doesn't depend on a specific mem_cgroup structure | |
6812 | * should be initialized from here. | |
6813 | */ | |
6814 | static int __init mem_cgroup_init(void) | |
6815 | { | |
6816 | int cpu, node; | |
6817 | ||
6818 | #ifdef CONFIG_MEMCG_KMEM | |
6819 | /* | |
6820 | * Kmem cache creation is mostly done with the slab_mutex held, | |
6821 | * so use a workqueue with limited concurrency to avoid stalling | |
6822 | * all worker threads in case lots of cgroups are created and | |
6823 | * destroyed simultaneously. | |
6824 | */ | |
6825 | memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1); | |
6826 | BUG_ON(!memcg_kmem_cache_wq); | |
6827 | #endif | |
6828 | ||
6829 | cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL, | |
6830 | memcg_hotplug_cpu_dead); | |
6831 | ||
6832 | for_each_possible_cpu(cpu) | |
6833 | INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work, | |
6834 | drain_local_stock); | |
6835 | ||
6836 | for_each_node(node) { | |
6837 | struct mem_cgroup_tree_per_node *rtpn; | |
6838 | ||
6839 | rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, | |
6840 | node_online(node) ? node : NUMA_NO_NODE); | |
6841 | ||
6842 | rtpn->rb_root = RB_ROOT; | |
6843 | rtpn->rb_rightmost = NULL; | |
6844 | spin_lock_init(&rtpn->lock); | |
6845 | soft_limit_tree.rb_tree_per_node[node] = rtpn; | |
6846 | } | |
6847 | ||
6848 | return 0; | |
6849 | } | |
6850 | subsys_initcall(mem_cgroup_init); | |
6851 | ||
6852 | #ifdef CONFIG_MEMCG_SWAP | |
6853 | static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg) | |
6854 | { | |
6855 | while (!refcount_inc_not_zero(&memcg->id.ref)) { | |
6856 | /* | |
6857 | * The root cgroup cannot be destroyed, so it's refcount must | |
6858 | * always be >= 1. | |
6859 | */ | |
6860 | if (WARN_ON_ONCE(memcg == root_mem_cgroup)) { | |
6861 | VM_BUG_ON(1); | |
6862 | break; | |
6863 | } | |
6864 | memcg = parent_mem_cgroup(memcg); | |
6865 | if (!memcg) | |
6866 | memcg = root_mem_cgroup; | |
6867 | } | |
6868 | return memcg; | |
6869 | } | |
6870 | ||
6871 | /** | |
6872 | * mem_cgroup_swapout - transfer a memsw charge to swap | |
6873 | * @page: page whose memsw charge to transfer | |
6874 | * @entry: swap entry to move the charge to | |
6875 | * | |
6876 | * Transfer the memsw charge of @page to @entry. | |
6877 | */ | |
6878 | void mem_cgroup_swapout(struct page *page, swp_entry_t entry) | |
6879 | { | |
6880 | struct mem_cgroup *memcg, *swap_memcg; | |
6881 | unsigned int nr_entries; | |
6882 | unsigned short oldid; | |
6883 | ||
6884 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
6885 | VM_BUG_ON_PAGE(page_count(page), page); | |
6886 | ||
6887 | if (!do_memsw_account()) | |
6888 | return; | |
6889 | ||
6890 | memcg = page->mem_cgroup; | |
6891 | ||
6892 | /* Readahead page, never charged */ | |
6893 | if (!memcg) | |
6894 | return; | |
6895 | ||
6896 | /* | |
6897 | * In case the memcg owning these pages has been offlined and doesn't | |
6898 | * have an ID allocated to it anymore, charge the closest online | |
6899 | * ancestor for the swap instead and transfer the memory+swap charge. | |
6900 | */ | |
6901 | swap_memcg = mem_cgroup_id_get_online(memcg); | |
6902 | nr_entries = hpage_nr_pages(page); | |
6903 | /* Get references for the tail pages, too */ | |
6904 | if (nr_entries > 1) | |
6905 | mem_cgroup_id_get_many(swap_memcg, nr_entries - 1); | |
6906 | oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg), | |
6907 | nr_entries); | |
6908 | VM_BUG_ON_PAGE(oldid, page); | |
6909 | mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries); | |
6910 | ||
6911 | page->mem_cgroup = NULL; | |
6912 | ||
6913 | if (!mem_cgroup_is_root(memcg)) | |
6914 | page_counter_uncharge(&memcg->memory, nr_entries); | |
6915 | ||
6916 | if (memcg != swap_memcg) { | |
6917 | if (!mem_cgroup_is_root(swap_memcg)) | |
6918 | page_counter_charge(&swap_memcg->memsw, nr_entries); | |
6919 | page_counter_uncharge(&memcg->memsw, nr_entries); | |
6920 | } | |
6921 | ||
6922 | /* | |
6923 | * Interrupts should be disabled here because the caller holds the | |
6924 | * i_pages lock which is taken with interrupts-off. It is | |
6925 | * important here to have the interrupts disabled because it is the | |
6926 | * only synchronisation we have for updating the per-CPU variables. | |
6927 | */ | |
6928 | VM_BUG_ON(!irqs_disabled()); | |
6929 | mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page), | |
6930 | -nr_entries); | |
6931 | memcg_check_events(memcg, page); | |
6932 | ||
6933 | if (!mem_cgroup_is_root(memcg)) | |
6934 | css_put_many(&memcg->css, nr_entries); | |
6935 | } | |
6936 | ||
6937 | /** | |
6938 | * mem_cgroup_try_charge_swap - try charging swap space for a page | |
6939 | * @page: page being added to swap | |
6940 | * @entry: swap entry to charge | |
6941 | * | |
6942 | * Try to charge @page's memcg for the swap space at @entry. | |
6943 | * | |
6944 | * Returns 0 on success, -ENOMEM on failure. | |
6945 | */ | |
6946 | int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry) | |
6947 | { | |
6948 | unsigned int nr_pages = hpage_nr_pages(page); | |
6949 | struct page_counter *counter; | |
6950 | struct mem_cgroup *memcg; | |
6951 | unsigned short oldid; | |
6952 | ||
6953 | if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account) | |
6954 | return 0; | |
6955 | ||
6956 | memcg = page->mem_cgroup; | |
6957 | ||
6958 | /* Readahead page, never charged */ | |
6959 | if (!memcg) | |
6960 | return 0; | |
6961 | ||
6962 | if (!entry.val) { | |
6963 | memcg_memory_event(memcg, MEMCG_SWAP_FAIL); | |
6964 | return 0; | |
6965 | } | |
6966 | ||
6967 | memcg = mem_cgroup_id_get_online(memcg); | |
6968 | ||
6969 | if (!mem_cgroup_is_root(memcg) && | |
6970 | !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) { | |
6971 | memcg_memory_event(memcg, MEMCG_SWAP_MAX); | |
6972 | memcg_memory_event(memcg, MEMCG_SWAP_FAIL); | |
6973 | mem_cgroup_id_put(memcg); | |
6974 | return -ENOMEM; | |
6975 | } | |
6976 | ||
6977 | /* Get references for the tail pages, too */ | |
6978 | if (nr_pages > 1) | |
6979 | mem_cgroup_id_get_many(memcg, nr_pages - 1); | |
6980 | oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages); | |
6981 | VM_BUG_ON_PAGE(oldid, page); | |
6982 | mod_memcg_state(memcg, MEMCG_SWAP, nr_pages); | |
6983 | ||
6984 | return 0; | |
6985 | } | |
6986 | ||
6987 | /** | |
6988 | * mem_cgroup_uncharge_swap - uncharge swap space | |
6989 | * @entry: swap entry to uncharge | |
6990 | * @nr_pages: the amount of swap space to uncharge | |
6991 | */ | |
6992 | void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages) | |
6993 | { | |
6994 | struct mem_cgroup *memcg; | |
6995 | unsigned short id; | |
6996 | ||
6997 | if (!do_swap_account) | |
6998 | return; | |
6999 | ||
7000 | id = swap_cgroup_record(entry, 0, nr_pages); | |
7001 | rcu_read_lock(); | |
7002 | memcg = mem_cgroup_from_id(id); | |
7003 | if (memcg) { | |
7004 | if (!mem_cgroup_is_root(memcg)) { | |
7005 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
7006 | page_counter_uncharge(&memcg->swap, nr_pages); | |
7007 | else | |
7008 | page_counter_uncharge(&memcg->memsw, nr_pages); | |
7009 | } | |
7010 | mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages); | |
7011 | mem_cgroup_id_put_many(memcg, nr_pages); | |
7012 | } | |
7013 | rcu_read_unlock(); | |
7014 | } | |
7015 | ||
7016 | long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg) | |
7017 | { | |
7018 | long nr_swap_pages = get_nr_swap_pages(); | |
7019 | ||
7020 | if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
7021 | return nr_swap_pages; | |
7022 | for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) | |
7023 | nr_swap_pages = min_t(long, nr_swap_pages, | |
7024 | READ_ONCE(memcg->swap.max) - | |
7025 | page_counter_read(&memcg->swap)); | |
7026 | return nr_swap_pages; | |
7027 | } | |
7028 | ||
7029 | bool mem_cgroup_swap_full(struct page *page) | |
7030 | { | |
7031 | struct mem_cgroup *memcg; | |
7032 | ||
7033 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
7034 | ||
7035 | if (vm_swap_full()) | |
7036 | return true; | |
7037 | if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys)) | |
7038 | return false; | |
7039 | ||
7040 | memcg = page->mem_cgroup; | |
7041 | if (!memcg) | |
7042 | return false; | |
7043 | ||
7044 | for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) | |
7045 | if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max) | |
7046 | return true; | |
7047 | ||
7048 | return false; | |
7049 | } | |
7050 | ||
7051 | /* for remember boot option*/ | |
7052 | #ifdef CONFIG_MEMCG_SWAP_ENABLED | |
7053 | static int really_do_swap_account __initdata = 1; | |
7054 | #else | |
7055 | static int really_do_swap_account __initdata; | |
7056 | #endif | |
7057 | ||
7058 | static int __init enable_swap_account(char *s) | |
7059 | { | |
7060 | if (!strcmp(s, "1")) | |
7061 | really_do_swap_account = 1; | |
7062 | else if (!strcmp(s, "0")) | |
7063 | really_do_swap_account = 0; | |
7064 | return 1; | |
7065 | } | |
7066 | __setup("swapaccount=", enable_swap_account); | |
7067 | ||
7068 | static u64 swap_current_read(struct cgroup_subsys_state *css, | |
7069 | struct cftype *cft) | |
7070 | { | |
7071 | struct mem_cgroup *memcg = mem_cgroup_from_css(css); | |
7072 | ||
7073 | return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE; | |
7074 | } | |
7075 | ||
7076 | static int swap_max_show(struct seq_file *m, void *v) | |
7077 | { | |
7078 | return seq_puts_memcg_tunable(m, | |
7079 | READ_ONCE(mem_cgroup_from_seq(m)->swap.max)); | |
7080 | } | |
7081 | ||
7082 | static ssize_t swap_max_write(struct kernfs_open_file *of, | |
7083 | char *buf, size_t nbytes, loff_t off) | |
7084 | { | |
7085 | struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of)); | |
7086 | unsigned long max; | |
7087 | int err; | |
7088 | ||
7089 | buf = strstrip(buf); | |
7090 | err = page_counter_memparse(buf, "max", &max); | |
7091 | if (err) | |
7092 | return err; | |
7093 | ||
7094 | xchg(&memcg->swap.max, max); | |
7095 | ||
7096 | return nbytes; | |
7097 | } | |
7098 | ||
7099 | static int swap_events_show(struct seq_file *m, void *v) | |
7100 | { | |
7101 | struct mem_cgroup *memcg = mem_cgroup_from_seq(m); | |
7102 | ||
7103 | seq_printf(m, "max %lu\n", | |
7104 | atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX])); | |
7105 | seq_printf(m, "fail %lu\n", | |
7106 | atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL])); | |
7107 | ||
7108 | return 0; | |
7109 | } | |
7110 | ||
7111 | static struct cftype swap_files[] = { | |
7112 | { | |
7113 | .name = "swap.current", | |
7114 | .flags = CFTYPE_NOT_ON_ROOT, | |
7115 | .read_u64 = swap_current_read, | |
7116 | }, | |
7117 | { | |
7118 | .name = "swap.max", | |
7119 | .flags = CFTYPE_NOT_ON_ROOT, | |
7120 | .seq_show = swap_max_show, | |
7121 | .write = swap_max_write, | |
7122 | }, | |
7123 | { | |
7124 | .name = "swap.events", | |
7125 | .flags = CFTYPE_NOT_ON_ROOT, | |
7126 | .file_offset = offsetof(struct mem_cgroup, swap_events_file), | |
7127 | .seq_show = swap_events_show, | |
7128 | }, | |
7129 | { } /* terminate */ | |
7130 | }; | |
7131 | ||
7132 | static struct cftype memsw_cgroup_files[] = { | |
7133 | { | |
7134 | .name = "memsw.usage_in_bytes", | |
7135 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), | |
7136 | .read_u64 = mem_cgroup_read_u64, | |
7137 | }, | |
7138 | { | |
7139 | .name = "memsw.max_usage_in_bytes", | |
7140 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), | |
7141 | .write = mem_cgroup_reset, | |
7142 | .read_u64 = mem_cgroup_read_u64, | |
7143 | }, | |
7144 | { | |
7145 | .name = "memsw.limit_in_bytes", | |
7146 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), | |
7147 | .write = mem_cgroup_write, | |
7148 | .read_u64 = mem_cgroup_read_u64, | |
7149 | }, | |
7150 | { | |
7151 | .name = "memsw.failcnt", | |
7152 | .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), | |
7153 | .write = mem_cgroup_reset, | |
7154 | .read_u64 = mem_cgroup_read_u64, | |
7155 | }, | |
7156 | { }, /* terminate */ | |
7157 | }; | |
7158 | ||
7159 | static int __init mem_cgroup_swap_init(void) | |
7160 | { | |
7161 | if (!mem_cgroup_disabled() && really_do_swap_account) { | |
7162 | do_swap_account = 1; | |
7163 | WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, | |
7164 | swap_files)); | |
7165 | WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, | |
7166 | memsw_cgroup_files)); | |
7167 | } | |
7168 | return 0; | |
7169 | } | |
7170 | subsys_initcall(mem_cgroup_swap_init); | |
7171 | ||
7172 | #endif /* CONFIG_MEMCG_SWAP */ |