memcontrol.c revision 90b3feaec8ffb167abd8903bf111605c2f035aa8
1/* memcontrol.c - Memory Controller 2 * 3 * Copyright IBM Corporation, 2007 4 * Author Balbir Singh <balbir@linux.vnet.ibm.com> 5 * 6 * Copyright 2007 OpenVZ SWsoft Inc 7 * Author: Pavel Emelianov <xemul@openvz.org> 8 * 9 * Memory thresholds 10 * Copyright (C) 2009 Nokia Corporation 11 * Author: Kirill A. Shutemov 12 * 13 * This program is free software; you can redistribute it and/or modify 14 * it under the terms of the GNU General Public License as published by 15 * the Free Software Foundation; either version 2 of the License, or 16 * (at your option) any later version. 17 * 18 * This program is distributed in the hope that it will be useful, 19 * but WITHOUT ANY WARRANTY; without even the implied warranty of 20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 21 * GNU General Public License for more details. 22 */ 23 24#include <linux/res_counter.h> 25#include <linux/memcontrol.h> 26#include <linux/cgroup.h> 27#include <linux/mm.h> 28#include <linux/hugetlb.h> 29#include <linux/pagemap.h> 30#include <linux/smp.h> 31#include <linux/page-flags.h> 32#include <linux/backing-dev.h> 33#include <linux/bit_spinlock.h> 34#include <linux/rcupdate.h> 35#include <linux/limits.h> 36#include <linux/export.h> 37#include <linux/mutex.h> 38#include <linux/rbtree.h> 39#include <linux/slab.h> 40#include <linux/swap.h> 41#include <linux/swapops.h> 42#include <linux/spinlock.h> 43#include <linux/eventfd.h> 44#include <linux/sort.h> 45#include <linux/fs.h> 46#include <linux/seq_file.h> 47#include <linux/vmalloc.h> 48#include <linux/mm_inline.h> 49#include <linux/page_cgroup.h> 50#include <linux/cpu.h> 51#include <linux/oom.h> 52#include "internal.h" 53#include <net/sock.h> 54#include <net/tcp_memcontrol.h> 55 56#include <asm/uaccess.h> 57 58#include <trace/events/vmscan.h> 59 60struct cgroup_subsys mem_cgroup_subsys __read_mostly; 61#define MEM_CGROUP_RECLAIM_RETRIES 5 62struct mem_cgroup *root_mem_cgroup __read_mostly; 63 64#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 65/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ 66int do_swap_account __read_mostly; 67 68/* for remember boot option*/ 69#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED 70static int really_do_swap_account __initdata = 1; 71#else 72static int really_do_swap_account __initdata = 0; 73#endif 74 75#else 76#define do_swap_account (0) 77#endif 78 79 80/* 81 * Statistics for memory cgroup. 82 */ 83enum mem_cgroup_stat_index { 84 /* 85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 86 */ 87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ 89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ 90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ 91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ 92 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ 93 MEM_CGROUP_STAT_NSTATS, 94}; 95 96enum mem_cgroup_events_index { 97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ 98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ 99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ 100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ 101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ 102 MEM_CGROUP_EVENTS_NSTATS, 103}; 104/* 105 * Per memcg event counter is incremented at every pagein/pageout. With THP, 106 * it will be incremated by the number of pages. This counter is used for 107 * for trigger some periodic events. This is straightforward and better 108 * than using jiffies etc. to handle periodic memcg event. 109 */ 110enum mem_cgroup_events_target { 111 MEM_CGROUP_TARGET_THRESH, 112 MEM_CGROUP_TARGET_SOFTLIMIT, 113 MEM_CGROUP_TARGET_NUMAINFO, 114 MEM_CGROUP_NTARGETS, 115}; 116#define THRESHOLDS_EVENTS_TARGET (128) 117#define SOFTLIMIT_EVENTS_TARGET (1024) 118#define NUMAINFO_EVENTS_TARGET (1024) 119 120struct mem_cgroup_stat_cpu { 121 long count[MEM_CGROUP_STAT_NSTATS]; 122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; 123 unsigned long targets[MEM_CGROUP_NTARGETS]; 124}; 125 126struct mem_cgroup_reclaim_iter { 127 /* css_id of the last scanned hierarchy member */ 128 int position; 129 /* scan generation, increased every round-trip */ 130 unsigned int generation; 131}; 132 133/* 134 * per-zone information in memory controller. 135 */ 136struct mem_cgroup_per_zone { 137 struct lruvec lruvec; 138 unsigned long count[NR_LRU_LISTS]; 139 140 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1]; 141 142 struct zone_reclaim_stat reclaim_stat; 143 struct rb_node tree_node; /* RB tree node */ 144 unsigned long long usage_in_excess;/* Set to the value by which */ 145 /* the soft limit is exceeded*/ 146 bool on_tree; 147 struct mem_cgroup *mem; /* Back pointer, we cannot */ 148 /* use container_of */ 149}; 150/* Macro for accessing counter */ 151#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 152 153struct mem_cgroup_per_node { 154 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 155}; 156 157struct mem_cgroup_lru_info { 158 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 159}; 160 161/* 162 * Cgroups above their limits are maintained in a RB-Tree, independent of 163 * their hierarchy representation 164 */ 165 166struct mem_cgroup_tree_per_zone { 167 struct rb_root rb_root; 168 spinlock_t lock; 169}; 170 171struct mem_cgroup_tree_per_node { 172 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; 173}; 174 175struct mem_cgroup_tree { 176 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; 177}; 178 179static struct mem_cgroup_tree soft_limit_tree __read_mostly; 180 181struct mem_cgroup_threshold { 182 struct eventfd_ctx *eventfd; 183 u64 threshold; 184}; 185 186/* For threshold */ 187struct mem_cgroup_threshold_ary { 188 /* An array index points to threshold just below usage. */ 189 int current_threshold; 190 /* Size of entries[] */ 191 unsigned int size; 192 /* Array of thresholds */ 193 struct mem_cgroup_threshold entries[0]; 194}; 195 196struct mem_cgroup_thresholds { 197 /* Primary thresholds array */ 198 struct mem_cgroup_threshold_ary *primary; 199 /* 200 * Spare threshold array. 201 * This is needed to make mem_cgroup_unregister_event() "never fail". 202 * It must be able to store at least primary->size - 1 entries. 203 */ 204 struct mem_cgroup_threshold_ary *spare; 205}; 206 207/* for OOM */ 208struct mem_cgroup_eventfd_list { 209 struct list_head list; 210 struct eventfd_ctx *eventfd; 211}; 212 213static void mem_cgroup_threshold(struct mem_cgroup *memcg); 214static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); 215 216/* 217 * The memory controller data structure. The memory controller controls both 218 * page cache and RSS per cgroup. We would eventually like to provide 219 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 220 * to help the administrator determine what knobs to tune. 221 * 222 * TODO: Add a water mark for the memory controller. Reclaim will begin when 223 * we hit the water mark. May be even add a low water mark, such that 224 * no reclaim occurs from a cgroup at it's low water mark, this is 225 * a feature that will be implemented much later in the future. 226 */ 227struct mem_cgroup { 228 struct cgroup_subsys_state css; 229 /* 230 * the counter to account for memory usage 231 */ 232 struct res_counter res; 233 /* 234 * the counter to account for mem+swap usage. 235 */ 236 struct res_counter memsw; 237 /* 238 * Per cgroup active and inactive list, similar to the 239 * per zone LRU lists. 240 */ 241 struct mem_cgroup_lru_info info; 242 int last_scanned_node; 243#if MAX_NUMNODES > 1 244 nodemask_t scan_nodes; 245 atomic_t numainfo_events; 246 atomic_t numainfo_updating; 247#endif 248 /* 249 * Should the accounting and control be hierarchical, per subtree? 250 */ 251 bool use_hierarchy; 252 253 bool oom_lock; 254 atomic_t under_oom; 255 256 atomic_t refcnt; 257 258 int swappiness; 259 /* OOM-Killer disable */ 260 int oom_kill_disable; 261 262 /* set when res.limit == memsw.limit */ 263 bool memsw_is_minimum; 264 265 /* protect arrays of thresholds */ 266 struct mutex thresholds_lock; 267 268 /* thresholds for memory usage. RCU-protected */ 269 struct mem_cgroup_thresholds thresholds; 270 271 /* thresholds for mem+swap usage. RCU-protected */ 272 struct mem_cgroup_thresholds memsw_thresholds; 273 274 /* For oom notifier event fd */ 275 struct list_head oom_notify; 276 277 /* 278 * Should we move charges of a task when a task is moved into this 279 * mem_cgroup ? And what type of charges should we move ? 280 */ 281 unsigned long move_charge_at_immigrate; 282 /* 283 * percpu counter. 284 */ 285 struct mem_cgroup_stat_cpu *stat; 286 /* 287 * used when a cpu is offlined or other synchronizations 288 * See mem_cgroup_read_stat(). 289 */ 290 struct mem_cgroup_stat_cpu nocpu_base; 291 spinlock_t pcp_counter_lock; 292 293#ifdef CONFIG_INET 294 struct tcp_memcontrol tcp_mem; 295#endif 296}; 297 298/* Stuffs for move charges at task migration. */ 299/* 300 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a 301 * left-shifted bitmap of these types. 302 */ 303enum move_type { 304 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ 305 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ 306 NR_MOVE_TYPE, 307}; 308 309/* "mc" and its members are protected by cgroup_mutex */ 310static struct move_charge_struct { 311 spinlock_t lock; /* for from, to */ 312 struct mem_cgroup *from; 313 struct mem_cgroup *to; 314 unsigned long precharge; 315 unsigned long moved_charge; 316 unsigned long moved_swap; 317 struct task_struct *moving_task; /* a task moving charges */ 318 wait_queue_head_t waitq; /* a waitq for other context */ 319} mc = { 320 .lock = __SPIN_LOCK_UNLOCKED(mc.lock), 321 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), 322}; 323 324static bool move_anon(void) 325{ 326 return test_bit(MOVE_CHARGE_TYPE_ANON, 327 &mc.to->move_charge_at_immigrate); 328} 329 330static bool move_file(void) 331{ 332 return test_bit(MOVE_CHARGE_TYPE_FILE, 333 &mc.to->move_charge_at_immigrate); 334} 335 336/* 337 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft 338 * limit reclaim to prevent infinite loops, if they ever occur. 339 */ 340#define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) 341#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) 342 343enum charge_type { 344 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 345 MEM_CGROUP_CHARGE_TYPE_MAPPED, 346 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 347 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 348 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 349 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ 350 NR_CHARGE_TYPE, 351}; 352 353/* for encoding cft->private value on file */ 354#define _MEM (0) 355#define _MEMSWAP (1) 356#define _OOM_TYPE (2) 357#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 358#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 359#define MEMFILE_ATTR(val) ((val) & 0xffff) 360/* Used for OOM nofiier */ 361#define OOM_CONTROL (0) 362 363/* 364 * Reclaim flags for mem_cgroup_hierarchical_reclaim 365 */ 366#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 367#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) 368#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 369#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) 370 371static void mem_cgroup_get(struct mem_cgroup *memcg); 372static void mem_cgroup_put(struct mem_cgroup *memcg); 373 374/* Writing them here to avoid exposing memcg's inner layout */ 375#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 376#ifdef CONFIG_INET 377#include <net/sock.h> 378#include <net/ip.h> 379 380static bool mem_cgroup_is_root(struct mem_cgroup *memcg); 381void sock_update_memcg(struct sock *sk) 382{ 383 if (static_branch(&memcg_socket_limit_enabled)) { 384 struct mem_cgroup *memcg; 385 386 BUG_ON(!sk->sk_prot->proto_cgroup); 387 388 /* Socket cloning can throw us here with sk_cgrp already 389 * filled. It won't however, necessarily happen from 390 * process context. So the test for root memcg given 391 * the current task's memcg won't help us in this case. 392 * 393 * Respecting the original socket's memcg is a better 394 * decision in this case. 395 */ 396 if (sk->sk_cgrp) { 397 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg)); 398 mem_cgroup_get(sk->sk_cgrp->memcg); 399 return; 400 } 401 402 rcu_read_lock(); 403 memcg = mem_cgroup_from_task(current); 404 if (!mem_cgroup_is_root(memcg)) { 405 mem_cgroup_get(memcg); 406 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg); 407 } 408 rcu_read_unlock(); 409 } 410} 411EXPORT_SYMBOL(sock_update_memcg); 412 413void sock_release_memcg(struct sock *sk) 414{ 415 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) { 416 struct mem_cgroup *memcg; 417 WARN_ON(!sk->sk_cgrp->memcg); 418 memcg = sk->sk_cgrp->memcg; 419 mem_cgroup_put(memcg); 420 } 421} 422 423struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg) 424{ 425 if (!memcg || mem_cgroup_is_root(memcg)) 426 return NULL; 427 428 return &memcg->tcp_mem.cg_proto; 429} 430EXPORT_SYMBOL(tcp_proto_cgroup); 431#endif /* CONFIG_INET */ 432#endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */ 433 434static void drain_all_stock_async(struct mem_cgroup *memcg); 435 436static struct mem_cgroup_per_zone * 437mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid) 438{ 439 return &memcg->info.nodeinfo[nid]->zoneinfo[zid]; 440} 441 442struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg) 443{ 444 return &memcg->css; 445} 446 447static struct mem_cgroup_per_zone * 448page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page) 449{ 450 int nid = page_to_nid(page); 451 int zid = page_zonenum(page); 452 453 return mem_cgroup_zoneinfo(memcg, nid, zid); 454} 455 456static struct mem_cgroup_tree_per_zone * 457soft_limit_tree_node_zone(int nid, int zid) 458{ 459 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 460} 461 462static struct mem_cgroup_tree_per_zone * 463soft_limit_tree_from_page(struct page *page) 464{ 465 int nid = page_to_nid(page); 466 int zid = page_zonenum(page); 467 468 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 469} 470 471static void 472__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg, 473 struct mem_cgroup_per_zone *mz, 474 struct mem_cgroup_tree_per_zone *mctz, 475 unsigned long long new_usage_in_excess) 476{ 477 struct rb_node **p = &mctz->rb_root.rb_node; 478 struct rb_node *parent = NULL; 479 struct mem_cgroup_per_zone *mz_node; 480 481 if (mz->on_tree) 482 return; 483 484 mz->usage_in_excess = new_usage_in_excess; 485 if (!mz->usage_in_excess) 486 return; 487 while (*p) { 488 parent = *p; 489 mz_node = rb_entry(parent, struct mem_cgroup_per_zone, 490 tree_node); 491 if (mz->usage_in_excess < mz_node->usage_in_excess) 492 p = &(*p)->rb_left; 493 /* 494 * We can't avoid mem cgroups that are over their soft 495 * limit by the same amount 496 */ 497 else if (mz->usage_in_excess >= mz_node->usage_in_excess) 498 p = &(*p)->rb_right; 499 } 500 rb_link_node(&mz->tree_node, parent, p); 501 rb_insert_color(&mz->tree_node, &mctz->rb_root); 502 mz->on_tree = true; 503} 504 505static void 506__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, 507 struct mem_cgroup_per_zone *mz, 508 struct mem_cgroup_tree_per_zone *mctz) 509{ 510 if (!mz->on_tree) 511 return; 512 rb_erase(&mz->tree_node, &mctz->rb_root); 513 mz->on_tree = false; 514} 515 516static void 517mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, 518 struct mem_cgroup_per_zone *mz, 519 struct mem_cgroup_tree_per_zone *mctz) 520{ 521 spin_lock(&mctz->lock); 522 __mem_cgroup_remove_exceeded(memcg, mz, mctz); 523 spin_unlock(&mctz->lock); 524} 525 526 527static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) 528{ 529 unsigned long long excess; 530 struct mem_cgroup_per_zone *mz; 531 struct mem_cgroup_tree_per_zone *mctz; 532 int nid = page_to_nid(page); 533 int zid = page_zonenum(page); 534 mctz = soft_limit_tree_from_page(page); 535 536 /* 537 * Necessary to update all ancestors when hierarchy is used. 538 * because their event counter is not touched. 539 */ 540 for (; memcg; memcg = parent_mem_cgroup(memcg)) { 541 mz = mem_cgroup_zoneinfo(memcg, nid, zid); 542 excess = res_counter_soft_limit_excess(&memcg->res); 543 /* 544 * We have to update the tree if mz is on RB-tree or 545 * mem is over its softlimit. 546 */ 547 if (excess || mz->on_tree) { 548 spin_lock(&mctz->lock); 549 /* if on-tree, remove it */ 550 if (mz->on_tree) 551 __mem_cgroup_remove_exceeded(memcg, mz, mctz); 552 /* 553 * Insert again. mz->usage_in_excess will be updated. 554 * If excess is 0, no tree ops. 555 */ 556 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess); 557 spin_unlock(&mctz->lock); 558 } 559 } 560} 561 562static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) 563{ 564 int node, zone; 565 struct mem_cgroup_per_zone *mz; 566 struct mem_cgroup_tree_per_zone *mctz; 567 568 for_each_node(node) { 569 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 570 mz = mem_cgroup_zoneinfo(memcg, node, zone); 571 mctz = soft_limit_tree_node_zone(node, zone); 572 mem_cgroup_remove_exceeded(memcg, mz, mctz); 573 } 574 } 575} 576 577static struct mem_cgroup_per_zone * 578__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 579{ 580 struct rb_node *rightmost = NULL; 581 struct mem_cgroup_per_zone *mz; 582 583retry: 584 mz = NULL; 585 rightmost = rb_last(&mctz->rb_root); 586 if (!rightmost) 587 goto done; /* Nothing to reclaim from */ 588 589 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); 590 /* 591 * Remove the node now but someone else can add it back, 592 * we will to add it back at the end of reclaim to its correct 593 * position in the tree. 594 */ 595 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 596 if (!res_counter_soft_limit_excess(&mz->mem->res) || 597 !css_tryget(&mz->mem->css)) 598 goto retry; 599done: 600 return mz; 601} 602 603static struct mem_cgroup_per_zone * 604mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 605{ 606 struct mem_cgroup_per_zone *mz; 607 608 spin_lock(&mctz->lock); 609 mz = __mem_cgroup_largest_soft_limit_node(mctz); 610 spin_unlock(&mctz->lock); 611 return mz; 612} 613 614/* 615 * Implementation Note: reading percpu statistics for memcg. 616 * 617 * Both of vmstat[] and percpu_counter has threshold and do periodic 618 * synchronization to implement "quick" read. There are trade-off between 619 * reading cost and precision of value. Then, we may have a chance to implement 620 * a periodic synchronizion of counter in memcg's counter. 621 * 622 * But this _read() function is used for user interface now. The user accounts 623 * memory usage by memory cgroup and he _always_ requires exact value because 624 * he accounts memory. Even if we provide quick-and-fuzzy read, we always 625 * have to visit all online cpus and make sum. So, for now, unnecessary 626 * synchronization is not implemented. (just implemented for cpu hotplug) 627 * 628 * If there are kernel internal actions which can make use of some not-exact 629 * value, and reading all cpu value can be performance bottleneck in some 630 * common workload, threashold and synchonization as vmstat[] should be 631 * implemented. 632 */ 633static long mem_cgroup_read_stat(struct mem_cgroup *memcg, 634 enum mem_cgroup_stat_index idx) 635{ 636 long val = 0; 637 int cpu; 638 639 get_online_cpus(); 640 for_each_online_cpu(cpu) 641 val += per_cpu(memcg->stat->count[idx], cpu); 642#ifdef CONFIG_HOTPLUG_CPU 643 spin_lock(&memcg->pcp_counter_lock); 644 val += memcg->nocpu_base.count[idx]; 645 spin_unlock(&memcg->pcp_counter_lock); 646#endif 647 put_online_cpus(); 648 return val; 649} 650 651static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg, 652 bool charge) 653{ 654 int val = (charge) ? 1 : -1; 655 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); 656} 657 658static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, 659 enum mem_cgroup_events_index idx) 660{ 661 unsigned long val = 0; 662 int cpu; 663 664 for_each_online_cpu(cpu) 665 val += per_cpu(memcg->stat->events[idx], cpu); 666#ifdef CONFIG_HOTPLUG_CPU 667 spin_lock(&memcg->pcp_counter_lock); 668 val += memcg->nocpu_base.events[idx]; 669 spin_unlock(&memcg->pcp_counter_lock); 670#endif 671 return val; 672} 673 674static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, 675 bool file, int nr_pages) 676{ 677 preempt_disable(); 678 679 if (file) 680 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], 681 nr_pages); 682 else 683 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], 684 nr_pages); 685 686 /* pagein of a big page is an event. So, ignore page size */ 687 if (nr_pages > 0) 688 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); 689 else { 690 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); 691 nr_pages = -nr_pages; /* for event */ 692 } 693 694 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); 695 696 preempt_enable(); 697} 698 699unsigned long 700mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, 701 unsigned int lru_mask) 702{ 703 struct mem_cgroup_per_zone *mz; 704 enum lru_list l; 705 unsigned long ret = 0; 706 707 mz = mem_cgroup_zoneinfo(memcg, nid, zid); 708 709 for_each_lru(l) { 710 if (BIT(l) & lru_mask) 711 ret += MEM_CGROUP_ZSTAT(mz, l); 712 } 713 return ret; 714} 715 716static unsigned long 717mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, 718 int nid, unsigned int lru_mask) 719{ 720 u64 total = 0; 721 int zid; 722 723 for (zid = 0; zid < MAX_NR_ZONES; zid++) 724 total += mem_cgroup_zone_nr_lru_pages(memcg, 725 nid, zid, lru_mask); 726 727 return total; 728} 729 730static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, 731 unsigned int lru_mask) 732{ 733 int nid; 734 u64 total = 0; 735 736 for_each_node_state(nid, N_HIGH_MEMORY) 737 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); 738 return total; 739} 740 741static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, 742 enum mem_cgroup_events_target target) 743{ 744 unsigned long val, next; 745 746 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]); 747 next = __this_cpu_read(memcg->stat->targets[target]); 748 /* from time_after() in jiffies.h */ 749 if ((long)next - (long)val < 0) { 750 switch (target) { 751 case MEM_CGROUP_TARGET_THRESH: 752 next = val + THRESHOLDS_EVENTS_TARGET; 753 break; 754 case MEM_CGROUP_TARGET_SOFTLIMIT: 755 next = val + SOFTLIMIT_EVENTS_TARGET; 756 break; 757 case MEM_CGROUP_TARGET_NUMAINFO: 758 next = val + NUMAINFO_EVENTS_TARGET; 759 break; 760 default: 761 break; 762 } 763 __this_cpu_write(memcg->stat->targets[target], next); 764 return true; 765 } 766 return false; 767} 768 769/* 770 * Check events in order. 771 * 772 */ 773static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) 774{ 775 preempt_disable(); 776 /* threshold event is triggered in finer grain than soft limit */ 777 if (unlikely(mem_cgroup_event_ratelimit(memcg, 778 MEM_CGROUP_TARGET_THRESH))) { 779 bool do_softlimit, do_numainfo; 780 781 do_softlimit = mem_cgroup_event_ratelimit(memcg, 782 MEM_CGROUP_TARGET_SOFTLIMIT); 783#if MAX_NUMNODES > 1 784 do_numainfo = mem_cgroup_event_ratelimit(memcg, 785 MEM_CGROUP_TARGET_NUMAINFO); 786#endif 787 preempt_enable(); 788 789 mem_cgroup_threshold(memcg); 790 if (unlikely(do_softlimit)) 791 mem_cgroup_update_tree(memcg, page); 792#if MAX_NUMNODES > 1 793 if (unlikely(do_numainfo)) 794 atomic_inc(&memcg->numainfo_events); 795#endif 796 } else 797 preempt_enable(); 798} 799 800struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 801{ 802 return container_of(cgroup_subsys_state(cont, 803 mem_cgroup_subsys_id), struct mem_cgroup, 804 css); 805} 806 807struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 808{ 809 /* 810 * mm_update_next_owner() may clear mm->owner to NULL 811 * if it races with swapoff, page migration, etc. 812 * So this can be called with p == NULL. 813 */ 814 if (unlikely(!p)) 815 return NULL; 816 817 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 818 struct mem_cgroup, css); 819} 820 821struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 822{ 823 struct mem_cgroup *memcg = NULL; 824 825 if (!mm) 826 return NULL; 827 /* 828 * Because we have no locks, mm->owner's may be being moved to other 829 * cgroup. We use css_tryget() here even if this looks 830 * pessimistic (rather than adding locks here). 831 */ 832 rcu_read_lock(); 833 do { 834 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 835 if (unlikely(!memcg)) 836 break; 837 } while (!css_tryget(&memcg->css)); 838 rcu_read_unlock(); 839 return memcg; 840} 841 842/** 843 * mem_cgroup_iter - iterate over memory cgroup hierarchy 844 * @root: hierarchy root 845 * @prev: previously returned memcg, NULL on first invocation 846 * @reclaim: cookie for shared reclaim walks, NULL for full walks 847 * 848 * Returns references to children of the hierarchy below @root, or 849 * @root itself, or %NULL after a full round-trip. 850 * 851 * Caller must pass the return value in @prev on subsequent 852 * invocations for reference counting, or use mem_cgroup_iter_break() 853 * to cancel a hierarchy walk before the round-trip is complete. 854 * 855 * Reclaimers can specify a zone and a priority level in @reclaim to 856 * divide up the memcgs in the hierarchy among all concurrent 857 * reclaimers operating on the same zone and priority. 858 */ 859struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, 860 struct mem_cgroup *prev, 861 struct mem_cgroup_reclaim_cookie *reclaim) 862{ 863 struct mem_cgroup *memcg = NULL; 864 int id = 0; 865 866 if (mem_cgroup_disabled()) 867 return NULL; 868 869 if (!root) 870 root = root_mem_cgroup; 871 872 if (prev && !reclaim) 873 id = css_id(&prev->css); 874 875 if (prev && prev != root) 876 css_put(&prev->css); 877 878 if (!root->use_hierarchy && root != root_mem_cgroup) { 879 if (prev) 880 return NULL; 881 return root; 882 } 883 884 while (!memcg) { 885 struct mem_cgroup_reclaim_iter *uninitialized_var(iter); 886 struct cgroup_subsys_state *css; 887 888 if (reclaim) { 889 int nid = zone_to_nid(reclaim->zone); 890 int zid = zone_idx(reclaim->zone); 891 struct mem_cgroup_per_zone *mz; 892 893 mz = mem_cgroup_zoneinfo(root, nid, zid); 894 iter = &mz->reclaim_iter[reclaim->priority]; 895 if (prev && reclaim->generation != iter->generation) 896 return NULL; 897 id = iter->position; 898 } 899 900 rcu_read_lock(); 901 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id); 902 if (css) { 903 if (css == &root->css || css_tryget(css)) 904 memcg = container_of(css, 905 struct mem_cgroup, css); 906 } else 907 id = 0; 908 rcu_read_unlock(); 909 910 if (reclaim) { 911 iter->position = id; 912 if (!css) 913 iter->generation++; 914 else if (!prev && memcg) 915 reclaim->generation = iter->generation; 916 } 917 918 if (prev && !css) 919 return NULL; 920 } 921 return memcg; 922} 923 924/** 925 * mem_cgroup_iter_break - abort a hierarchy walk prematurely 926 * @root: hierarchy root 927 * @prev: last visited hierarchy member as returned by mem_cgroup_iter() 928 */ 929void mem_cgroup_iter_break(struct mem_cgroup *root, 930 struct mem_cgroup *prev) 931{ 932 if (!root) 933 root = root_mem_cgroup; 934 if (prev && prev != root) 935 css_put(&prev->css); 936} 937 938/* 939 * Iteration constructs for visiting all cgroups (under a tree). If 940 * loops are exited prematurely (break), mem_cgroup_iter_break() must 941 * be used for reference counting. 942 */ 943#define for_each_mem_cgroup_tree(iter, root) \ 944 for (iter = mem_cgroup_iter(root, NULL, NULL); \ 945 iter != NULL; \ 946 iter = mem_cgroup_iter(root, iter, NULL)) 947 948#define for_each_mem_cgroup(iter) \ 949 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ 950 iter != NULL; \ 951 iter = mem_cgroup_iter(NULL, iter, NULL)) 952 953static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) 954{ 955 return (memcg == root_mem_cgroup); 956} 957 958void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) 959{ 960 struct mem_cgroup *memcg; 961 962 if (!mm) 963 return; 964 965 rcu_read_lock(); 966 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 967 if (unlikely(!memcg)) 968 goto out; 969 970 switch (idx) { 971 case PGFAULT: 972 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]); 973 break; 974 case PGMAJFAULT: 975 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]); 976 break; 977 default: 978 BUG(); 979 } 980out: 981 rcu_read_unlock(); 982} 983EXPORT_SYMBOL(mem_cgroup_count_vm_event); 984 985/** 986 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg 987 * @zone: zone of the wanted lruvec 988 * @mem: memcg of the wanted lruvec 989 * 990 * Returns the lru list vector holding pages for the given @zone and 991 * @mem. This can be the global zone lruvec, if the memory controller 992 * is disabled. 993 */ 994struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, 995 struct mem_cgroup *memcg) 996{ 997 struct mem_cgroup_per_zone *mz; 998 999 if (mem_cgroup_disabled()) 1000 return &zone->lruvec; 1001 1002 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); 1003 return &mz->lruvec; 1004} 1005 1006/* 1007 * Following LRU functions are allowed to be used without PCG_LOCK. 1008 * Operations are called by routine of global LRU independently from memcg. 1009 * What we have to take care of here is validness of pc->mem_cgroup. 1010 * 1011 * Changes to pc->mem_cgroup happens when 1012 * 1. charge 1013 * 2. moving account 1014 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 1015 * It is added to LRU before charge. 1016 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 1017 * When moving account, the page is not on LRU. It's isolated. 1018 */ 1019 1020/** 1021 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec 1022 * @zone: zone of the page 1023 * @page: the page 1024 * @lru: current lru 1025 * 1026 * This function accounts for @page being added to @lru, and returns 1027 * the lruvec for the given @zone and the memcg @page is charged to. 1028 * 1029 * The callsite is then responsible for physically linking the page to 1030 * the returned lruvec->lists[@lru]. 1031 */ 1032struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page, 1033 enum lru_list lru) 1034{ 1035 struct mem_cgroup_per_zone *mz; 1036 struct mem_cgroup *memcg; 1037 struct page_cgroup *pc; 1038 1039 if (mem_cgroup_disabled()) 1040 return &zone->lruvec; 1041 1042 pc = lookup_page_cgroup(page); 1043 memcg = pc->mem_cgroup; 1044 mz = page_cgroup_zoneinfo(memcg, page); 1045 /* compound_order() is stabilized through lru_lock */ 1046 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 1047 return &mz->lruvec; 1048} 1049 1050/** 1051 * mem_cgroup_lru_del_list - account for removing an lru page 1052 * @page: the page 1053 * @lru: target lru 1054 * 1055 * This function accounts for @page being removed from @lru. 1056 * 1057 * The callsite is then responsible for physically unlinking 1058 * @page->lru. 1059 */ 1060void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) 1061{ 1062 struct mem_cgroup_per_zone *mz; 1063 struct mem_cgroup *memcg; 1064 struct page_cgroup *pc; 1065 1066 if (mem_cgroup_disabled()) 1067 return; 1068 1069 pc = lookup_page_cgroup(page); 1070 memcg = pc->mem_cgroup; 1071 VM_BUG_ON(!memcg); 1072 mz = page_cgroup_zoneinfo(memcg, page); 1073 /* huge page split is done under lru_lock. so, we have no races. */ 1074 VM_BUG_ON(MEM_CGROUP_ZSTAT(mz, lru) < (1 << compound_order(page))); 1075 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 1076} 1077 1078void mem_cgroup_lru_del(struct page *page) 1079{ 1080 mem_cgroup_lru_del_list(page, page_lru(page)); 1081} 1082 1083/** 1084 * mem_cgroup_lru_move_lists - account for moving a page between lrus 1085 * @zone: zone of the page 1086 * @page: the page 1087 * @from: current lru 1088 * @to: target lru 1089 * 1090 * This function accounts for @page being moved between the lrus @from 1091 * and @to, and returns the lruvec for the given @zone and the memcg 1092 * @page is charged to. 1093 * 1094 * The callsite is then responsible for physically relinking 1095 * @page->lru to the returned lruvec->lists[@to]. 1096 */ 1097struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, 1098 struct page *page, 1099 enum lru_list from, 1100 enum lru_list to) 1101{ 1102 /* XXX: Optimize this, especially for @from == @to */ 1103 mem_cgroup_lru_del_list(page, from); 1104 return mem_cgroup_lru_add_list(zone, page, to); 1105} 1106 1107/* 1108 * Checks whether given mem is same or in the root_mem_cgroup's 1109 * hierarchy subtree 1110 */ 1111static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, 1112 struct mem_cgroup *memcg) 1113{ 1114 if (root_memcg != memcg) { 1115 return (root_memcg->use_hierarchy && 1116 css_is_ancestor(&memcg->css, &root_memcg->css)); 1117 } 1118 1119 return true; 1120} 1121 1122int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) 1123{ 1124 int ret; 1125 struct mem_cgroup *curr = NULL; 1126 struct task_struct *p; 1127 1128 p = find_lock_task_mm(task); 1129 if (p) { 1130 curr = try_get_mem_cgroup_from_mm(p->mm); 1131 task_unlock(p); 1132 } else { 1133 /* 1134 * All threads may have already detached their mm's, but the oom 1135 * killer still needs to detect if they have already been oom 1136 * killed to prevent needlessly killing additional tasks. 1137 */ 1138 task_lock(task); 1139 curr = mem_cgroup_from_task(task); 1140 if (curr) 1141 css_get(&curr->css); 1142 task_unlock(task); 1143 } 1144 if (!curr) 1145 return 0; 1146 /* 1147 * We should check use_hierarchy of "memcg" not "curr". Because checking 1148 * use_hierarchy of "curr" here make this function true if hierarchy is 1149 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* 1150 * hierarchy(even if use_hierarchy is disabled in "memcg"). 1151 */ 1152 ret = mem_cgroup_same_or_subtree(memcg, curr); 1153 css_put(&curr->css); 1154 return ret; 1155} 1156 1157int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) 1158{ 1159 unsigned long inactive_ratio; 1160 int nid = zone_to_nid(zone); 1161 int zid = zone_idx(zone); 1162 unsigned long inactive; 1163 unsigned long active; 1164 unsigned long gb; 1165 1166 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1167 BIT(LRU_INACTIVE_ANON)); 1168 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1169 BIT(LRU_ACTIVE_ANON)); 1170 1171 gb = (inactive + active) >> (30 - PAGE_SHIFT); 1172 if (gb) 1173 inactive_ratio = int_sqrt(10 * gb); 1174 else 1175 inactive_ratio = 1; 1176 1177 return inactive * inactive_ratio < active; 1178} 1179 1180int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) 1181{ 1182 unsigned long active; 1183 unsigned long inactive; 1184 int zid = zone_idx(zone); 1185 int nid = zone_to_nid(zone); 1186 1187 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1188 BIT(LRU_INACTIVE_FILE)); 1189 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1190 BIT(LRU_ACTIVE_FILE)); 1191 1192 return (active > inactive); 1193} 1194 1195struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1196 struct zone *zone) 1197{ 1198 int nid = zone_to_nid(zone); 1199 int zid = zone_idx(zone); 1200 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1201 1202 return &mz->reclaim_stat; 1203} 1204 1205struct zone_reclaim_stat * 1206mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1207{ 1208 struct page_cgroup *pc; 1209 struct mem_cgroup_per_zone *mz; 1210 1211 if (mem_cgroup_disabled()) 1212 return NULL; 1213 1214 pc = lookup_page_cgroup(page); 1215 if (!PageCgroupUsed(pc)) 1216 return NULL; 1217 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1218 smp_rmb(); 1219 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1220 return &mz->reclaim_stat; 1221} 1222 1223#define mem_cgroup_from_res_counter(counter, member) \ 1224 container_of(counter, struct mem_cgroup, member) 1225 1226/** 1227 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1228 * @mem: the memory cgroup 1229 * 1230 * Returns the maximum amount of memory @mem can be charged with, in 1231 * pages. 1232 */ 1233static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) 1234{ 1235 unsigned long long margin; 1236 1237 margin = res_counter_margin(&memcg->res); 1238 if (do_swap_account) 1239 margin = min(margin, res_counter_margin(&memcg->memsw)); 1240 return margin >> PAGE_SHIFT; 1241} 1242 1243int mem_cgroup_swappiness(struct mem_cgroup *memcg) 1244{ 1245 struct cgroup *cgrp = memcg->css.cgroup; 1246 1247 /* root ? */ 1248 if (cgrp->parent == NULL) 1249 return vm_swappiness; 1250 1251 return memcg->swappiness; 1252} 1253 1254static void mem_cgroup_start_move(struct mem_cgroup *memcg) 1255{ 1256 int cpu; 1257 1258 get_online_cpus(); 1259 spin_lock(&memcg->pcp_counter_lock); 1260 for_each_online_cpu(cpu) 1261 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1262 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1263 spin_unlock(&memcg->pcp_counter_lock); 1264 put_online_cpus(); 1265 1266 synchronize_rcu(); 1267} 1268 1269static void mem_cgroup_end_move(struct mem_cgroup *memcg) 1270{ 1271 int cpu; 1272 1273 if (!memcg) 1274 return; 1275 get_online_cpus(); 1276 spin_lock(&memcg->pcp_counter_lock); 1277 for_each_online_cpu(cpu) 1278 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1279 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1280 spin_unlock(&memcg->pcp_counter_lock); 1281 put_online_cpus(); 1282} 1283/* 1284 * 2 routines for checking "mem" is under move_account() or not. 1285 * 1286 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1287 * for avoiding race in accounting. If true, 1288 * pc->mem_cgroup may be overwritten. 1289 * 1290 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1291 * under hierarchy of moving cgroups. This is for 1292 * waiting at hith-memory prressure caused by "move". 1293 */ 1294 1295static bool mem_cgroup_stealed(struct mem_cgroup *memcg) 1296{ 1297 VM_BUG_ON(!rcu_read_lock_held()); 1298 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1299} 1300 1301static bool mem_cgroup_under_move(struct mem_cgroup *memcg) 1302{ 1303 struct mem_cgroup *from; 1304 struct mem_cgroup *to; 1305 bool ret = false; 1306 /* 1307 * Unlike task_move routines, we access mc.to, mc.from not under 1308 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1309 */ 1310 spin_lock(&mc.lock); 1311 from = mc.from; 1312 to = mc.to; 1313 if (!from) 1314 goto unlock; 1315 1316 ret = mem_cgroup_same_or_subtree(memcg, from) 1317 || mem_cgroup_same_or_subtree(memcg, to); 1318unlock: 1319 spin_unlock(&mc.lock); 1320 return ret; 1321} 1322 1323static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) 1324{ 1325 if (mc.moving_task && current != mc.moving_task) { 1326 if (mem_cgroup_under_move(memcg)) { 1327 DEFINE_WAIT(wait); 1328 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1329 /* moving charge context might have finished. */ 1330 if (mc.moving_task) 1331 schedule(); 1332 finish_wait(&mc.waitq, &wait); 1333 return true; 1334 } 1335 } 1336 return false; 1337} 1338 1339/** 1340 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1341 * @memcg: The memory cgroup that went over limit 1342 * @p: Task that is going to be killed 1343 * 1344 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1345 * enabled 1346 */ 1347void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1348{ 1349 struct cgroup *task_cgrp; 1350 struct cgroup *mem_cgrp; 1351 /* 1352 * Need a buffer in BSS, can't rely on allocations. The code relies 1353 * on the assumption that OOM is serialized for memory controller. 1354 * If this assumption is broken, revisit this code. 1355 */ 1356 static char memcg_name[PATH_MAX]; 1357 int ret; 1358 1359 if (!memcg || !p) 1360 return; 1361 1362 1363 rcu_read_lock(); 1364 1365 mem_cgrp = memcg->css.cgroup; 1366 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1367 1368 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1369 if (ret < 0) { 1370 /* 1371 * Unfortunately, we are unable to convert to a useful name 1372 * But we'll still print out the usage information 1373 */ 1374 rcu_read_unlock(); 1375 goto done; 1376 } 1377 rcu_read_unlock(); 1378 1379 printk(KERN_INFO "Task in %s killed", memcg_name); 1380 1381 rcu_read_lock(); 1382 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1383 if (ret < 0) { 1384 rcu_read_unlock(); 1385 goto done; 1386 } 1387 rcu_read_unlock(); 1388 1389 /* 1390 * Continues from above, so we don't need an KERN_ level 1391 */ 1392 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1393done: 1394 1395 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1396 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1397 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1398 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1399 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1400 "failcnt %llu\n", 1401 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1402 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1403 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1404} 1405 1406/* 1407 * This function returns the number of memcg under hierarchy tree. Returns 1408 * 1(self count) if no children. 1409 */ 1410static int mem_cgroup_count_children(struct mem_cgroup *memcg) 1411{ 1412 int num = 0; 1413 struct mem_cgroup *iter; 1414 1415 for_each_mem_cgroup_tree(iter, memcg) 1416 num++; 1417 return num; 1418} 1419 1420/* 1421 * Return the memory (and swap, if configured) limit for a memcg. 1422 */ 1423u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1424{ 1425 u64 limit; 1426 u64 memsw; 1427 1428 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1429 limit += total_swap_pages << PAGE_SHIFT; 1430 1431 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1432 /* 1433 * If memsw is finite and limits the amount of swap space available 1434 * to this memcg, return that limit. 1435 */ 1436 return min(limit, memsw); 1437} 1438 1439static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, 1440 gfp_t gfp_mask, 1441 unsigned long flags) 1442{ 1443 unsigned long total = 0; 1444 bool noswap = false; 1445 int loop; 1446 1447 if (flags & MEM_CGROUP_RECLAIM_NOSWAP) 1448 noswap = true; 1449 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) 1450 noswap = true; 1451 1452 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { 1453 if (loop) 1454 drain_all_stock_async(memcg); 1455 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); 1456 /* 1457 * Allow limit shrinkers, which are triggered directly 1458 * by userspace, to catch signals and stop reclaim 1459 * after minimal progress, regardless of the margin. 1460 */ 1461 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) 1462 break; 1463 if (mem_cgroup_margin(memcg)) 1464 break; 1465 /* 1466 * If nothing was reclaimed after two attempts, there 1467 * may be no reclaimable pages in this hierarchy. 1468 */ 1469 if (loop && !total) 1470 break; 1471 } 1472 return total; 1473} 1474 1475/** 1476 * test_mem_cgroup_node_reclaimable 1477 * @mem: the target memcg 1478 * @nid: the node ID to be checked. 1479 * @noswap : specify true here if the user wants flle only information. 1480 * 1481 * This function returns whether the specified memcg contains any 1482 * reclaimable pages on a node. Returns true if there are any reclaimable 1483 * pages in the node. 1484 */ 1485static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, 1486 int nid, bool noswap) 1487{ 1488 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) 1489 return true; 1490 if (noswap || !total_swap_pages) 1491 return false; 1492 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) 1493 return true; 1494 return false; 1495 1496} 1497#if MAX_NUMNODES > 1 1498 1499/* 1500 * Always updating the nodemask is not very good - even if we have an empty 1501 * list or the wrong list here, we can start from some node and traverse all 1502 * nodes based on the zonelist. So update the list loosely once per 10 secs. 1503 * 1504 */ 1505static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) 1506{ 1507 int nid; 1508 /* 1509 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET 1510 * pagein/pageout changes since the last update. 1511 */ 1512 if (!atomic_read(&memcg->numainfo_events)) 1513 return; 1514 if (atomic_inc_return(&memcg->numainfo_updating) > 1) 1515 return; 1516 1517 /* make a nodemask where this memcg uses memory from */ 1518 memcg->scan_nodes = node_states[N_HIGH_MEMORY]; 1519 1520 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { 1521 1522 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) 1523 node_clear(nid, memcg->scan_nodes); 1524 } 1525 1526 atomic_set(&memcg->numainfo_events, 0); 1527 atomic_set(&memcg->numainfo_updating, 0); 1528} 1529 1530/* 1531 * Selecting a node where we start reclaim from. Because what we need is just 1532 * reducing usage counter, start from anywhere is O,K. Considering 1533 * memory reclaim from current node, there are pros. and cons. 1534 * 1535 * Freeing memory from current node means freeing memory from a node which 1536 * we'll use or we've used. So, it may make LRU bad. And if several threads 1537 * hit limits, it will see a contention on a node. But freeing from remote 1538 * node means more costs for memory reclaim because of memory latency. 1539 * 1540 * Now, we use round-robin. Better algorithm is welcomed. 1541 */ 1542int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1543{ 1544 int node; 1545 1546 mem_cgroup_may_update_nodemask(memcg); 1547 node = memcg->last_scanned_node; 1548 1549 node = next_node(node, memcg->scan_nodes); 1550 if (node == MAX_NUMNODES) 1551 node = first_node(memcg->scan_nodes); 1552 /* 1553 * We call this when we hit limit, not when pages are added to LRU. 1554 * No LRU may hold pages because all pages are UNEVICTABLE or 1555 * memcg is too small and all pages are not on LRU. In that case, 1556 * we use curret node. 1557 */ 1558 if (unlikely(node == MAX_NUMNODES)) 1559 node = numa_node_id(); 1560 1561 memcg->last_scanned_node = node; 1562 return node; 1563} 1564 1565/* 1566 * Check all nodes whether it contains reclaimable pages or not. 1567 * For quick scan, we make use of scan_nodes. This will allow us to skip 1568 * unused nodes. But scan_nodes is lazily updated and may not cotain 1569 * enough new information. We need to do double check. 1570 */ 1571bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1572{ 1573 int nid; 1574 1575 /* 1576 * quick check...making use of scan_node. 1577 * We can skip unused nodes. 1578 */ 1579 if (!nodes_empty(memcg->scan_nodes)) { 1580 for (nid = first_node(memcg->scan_nodes); 1581 nid < MAX_NUMNODES; 1582 nid = next_node(nid, memcg->scan_nodes)) { 1583 1584 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1585 return true; 1586 } 1587 } 1588 /* 1589 * Check rest of nodes. 1590 */ 1591 for_each_node_state(nid, N_HIGH_MEMORY) { 1592 if (node_isset(nid, memcg->scan_nodes)) 1593 continue; 1594 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1595 return true; 1596 } 1597 return false; 1598} 1599 1600#else 1601int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1602{ 1603 return 0; 1604} 1605 1606bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1607{ 1608 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); 1609} 1610#endif 1611 1612static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, 1613 struct zone *zone, 1614 gfp_t gfp_mask, 1615 unsigned long *total_scanned) 1616{ 1617 struct mem_cgroup *victim = NULL; 1618 int total = 0; 1619 int loop = 0; 1620 unsigned long excess; 1621 unsigned long nr_scanned; 1622 struct mem_cgroup_reclaim_cookie reclaim = { 1623 .zone = zone, 1624 .priority = 0, 1625 }; 1626 1627 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; 1628 1629 while (1) { 1630 victim = mem_cgroup_iter(root_memcg, victim, &reclaim); 1631 if (!victim) { 1632 loop++; 1633 if (loop >= 2) { 1634 /* 1635 * If we have not been able to reclaim 1636 * anything, it might because there are 1637 * no reclaimable pages under this hierarchy 1638 */ 1639 if (!total) 1640 break; 1641 /* 1642 * We want to do more targeted reclaim. 1643 * excess >> 2 is not to excessive so as to 1644 * reclaim too much, nor too less that we keep 1645 * coming back to reclaim from this cgroup 1646 */ 1647 if (total >= (excess >> 2) || 1648 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) 1649 break; 1650 } 1651 continue; 1652 } 1653 if (!mem_cgroup_reclaimable(victim, false)) 1654 continue; 1655 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, 1656 zone, &nr_scanned); 1657 *total_scanned += nr_scanned; 1658 if (!res_counter_soft_limit_excess(&root_memcg->res)) 1659 break; 1660 } 1661 mem_cgroup_iter_break(root_memcg, victim); 1662 return total; 1663} 1664 1665/* 1666 * Check OOM-Killer is already running under our hierarchy. 1667 * If someone is running, return false. 1668 * Has to be called with memcg_oom_lock 1669 */ 1670static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) 1671{ 1672 struct mem_cgroup *iter, *failed = NULL; 1673 1674 for_each_mem_cgroup_tree(iter, memcg) { 1675 if (iter->oom_lock) { 1676 /* 1677 * this subtree of our hierarchy is already locked 1678 * so we cannot give a lock. 1679 */ 1680 failed = iter; 1681 mem_cgroup_iter_break(memcg, iter); 1682 break; 1683 } else 1684 iter->oom_lock = true; 1685 } 1686 1687 if (!failed) 1688 return true; 1689 1690 /* 1691 * OK, we failed to lock the whole subtree so we have to clean up 1692 * what we set up to the failing subtree 1693 */ 1694 for_each_mem_cgroup_tree(iter, memcg) { 1695 if (iter == failed) { 1696 mem_cgroup_iter_break(memcg, iter); 1697 break; 1698 } 1699 iter->oom_lock = false; 1700 } 1701 return false; 1702} 1703 1704/* 1705 * Has to be called with memcg_oom_lock 1706 */ 1707static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) 1708{ 1709 struct mem_cgroup *iter; 1710 1711 for_each_mem_cgroup_tree(iter, memcg) 1712 iter->oom_lock = false; 1713 return 0; 1714} 1715 1716static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) 1717{ 1718 struct mem_cgroup *iter; 1719 1720 for_each_mem_cgroup_tree(iter, memcg) 1721 atomic_inc(&iter->under_oom); 1722} 1723 1724static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) 1725{ 1726 struct mem_cgroup *iter; 1727 1728 /* 1729 * When a new child is created while the hierarchy is under oom, 1730 * mem_cgroup_oom_lock() may not be called. We have to use 1731 * atomic_add_unless() here. 1732 */ 1733 for_each_mem_cgroup_tree(iter, memcg) 1734 atomic_add_unless(&iter->under_oom, -1, 0); 1735} 1736 1737static DEFINE_SPINLOCK(memcg_oom_lock); 1738static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1739 1740struct oom_wait_info { 1741 struct mem_cgroup *mem; 1742 wait_queue_t wait; 1743}; 1744 1745static int memcg_oom_wake_function(wait_queue_t *wait, 1746 unsigned mode, int sync, void *arg) 1747{ 1748 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg, 1749 *oom_wait_memcg; 1750 struct oom_wait_info *oom_wait_info; 1751 1752 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1753 oom_wait_memcg = oom_wait_info->mem; 1754 1755 /* 1756 * Both of oom_wait_info->mem and wake_mem are stable under us. 1757 * Then we can use css_is_ancestor without taking care of RCU. 1758 */ 1759 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) 1760 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) 1761 return 0; 1762 return autoremove_wake_function(wait, mode, sync, arg); 1763} 1764 1765static void memcg_wakeup_oom(struct mem_cgroup *memcg) 1766{ 1767 /* for filtering, pass "memcg" as argument. */ 1768 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); 1769} 1770 1771static void memcg_oom_recover(struct mem_cgroup *memcg) 1772{ 1773 if (memcg && atomic_read(&memcg->under_oom)) 1774 memcg_wakeup_oom(memcg); 1775} 1776 1777/* 1778 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1779 */ 1780bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask) 1781{ 1782 struct oom_wait_info owait; 1783 bool locked, need_to_kill; 1784 1785 owait.mem = memcg; 1786 owait.wait.flags = 0; 1787 owait.wait.func = memcg_oom_wake_function; 1788 owait.wait.private = current; 1789 INIT_LIST_HEAD(&owait.wait.task_list); 1790 need_to_kill = true; 1791 mem_cgroup_mark_under_oom(memcg); 1792 1793 /* At first, try to OOM lock hierarchy under memcg.*/ 1794 spin_lock(&memcg_oom_lock); 1795 locked = mem_cgroup_oom_lock(memcg); 1796 /* 1797 * Even if signal_pending(), we can't quit charge() loop without 1798 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1799 * under OOM is always welcomed, use TASK_KILLABLE here. 1800 */ 1801 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1802 if (!locked || memcg->oom_kill_disable) 1803 need_to_kill = false; 1804 if (locked) 1805 mem_cgroup_oom_notify(memcg); 1806 spin_unlock(&memcg_oom_lock); 1807 1808 if (need_to_kill) { 1809 finish_wait(&memcg_oom_waitq, &owait.wait); 1810 mem_cgroup_out_of_memory(memcg, mask); 1811 } else { 1812 schedule(); 1813 finish_wait(&memcg_oom_waitq, &owait.wait); 1814 } 1815 spin_lock(&memcg_oom_lock); 1816 if (locked) 1817 mem_cgroup_oom_unlock(memcg); 1818 memcg_wakeup_oom(memcg); 1819 spin_unlock(&memcg_oom_lock); 1820 1821 mem_cgroup_unmark_under_oom(memcg); 1822 1823 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1824 return false; 1825 /* Give chance to dying process */ 1826 schedule_timeout_uninterruptible(1); 1827 return true; 1828} 1829 1830/* 1831 * Currently used to update mapped file statistics, but the routine can be 1832 * generalized to update other statistics as well. 1833 * 1834 * Notes: Race condition 1835 * 1836 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1837 * it tends to be costly. But considering some conditions, we doesn't need 1838 * to do so _always_. 1839 * 1840 * Considering "charge", lock_page_cgroup() is not required because all 1841 * file-stat operations happen after a page is attached to radix-tree. There 1842 * are no race with "charge". 1843 * 1844 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1845 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1846 * if there are race with "uncharge". Statistics itself is properly handled 1847 * by flags. 1848 * 1849 * Considering "move", this is an only case we see a race. To make the race 1850 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1851 * possibility of race condition. If there is, we take a lock. 1852 */ 1853 1854void mem_cgroup_update_page_stat(struct page *page, 1855 enum mem_cgroup_page_stat_item idx, int val) 1856{ 1857 struct mem_cgroup *memcg; 1858 struct page_cgroup *pc = lookup_page_cgroup(page); 1859 bool need_unlock = false; 1860 unsigned long uninitialized_var(flags); 1861 1862 if (mem_cgroup_disabled()) 1863 return; 1864 1865 rcu_read_lock(); 1866 memcg = pc->mem_cgroup; 1867 if (unlikely(!memcg || !PageCgroupUsed(pc))) 1868 goto out; 1869 /* pc->mem_cgroup is unstable ? */ 1870 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) { 1871 /* take a lock against to access pc->mem_cgroup */ 1872 move_lock_page_cgroup(pc, &flags); 1873 need_unlock = true; 1874 memcg = pc->mem_cgroup; 1875 if (!memcg || !PageCgroupUsed(pc)) 1876 goto out; 1877 } 1878 1879 switch (idx) { 1880 case MEMCG_NR_FILE_MAPPED: 1881 if (val > 0) 1882 SetPageCgroupFileMapped(pc); 1883 else if (!page_mapped(page)) 1884 ClearPageCgroupFileMapped(pc); 1885 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1886 break; 1887 default: 1888 BUG(); 1889 } 1890 1891 this_cpu_add(memcg->stat->count[idx], val); 1892 1893out: 1894 if (unlikely(need_unlock)) 1895 move_unlock_page_cgroup(pc, &flags); 1896 rcu_read_unlock(); 1897 return; 1898} 1899EXPORT_SYMBOL(mem_cgroup_update_page_stat); 1900 1901/* 1902 * size of first charge trial. "32" comes from vmscan.c's magic value. 1903 * TODO: maybe necessary to use big numbers in big irons. 1904 */ 1905#define CHARGE_BATCH 32U 1906struct memcg_stock_pcp { 1907 struct mem_cgroup *cached; /* this never be root cgroup */ 1908 unsigned int nr_pages; 1909 struct work_struct work; 1910 unsigned long flags; 1911#define FLUSHING_CACHED_CHARGE (0) 1912}; 1913static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 1914static DEFINE_MUTEX(percpu_charge_mutex); 1915 1916/* 1917 * Try to consume stocked charge on this cpu. If success, one page is consumed 1918 * from local stock and true is returned. If the stock is 0 or charges from a 1919 * cgroup which is not current target, returns false. This stock will be 1920 * refilled. 1921 */ 1922static bool consume_stock(struct mem_cgroup *memcg) 1923{ 1924 struct memcg_stock_pcp *stock; 1925 bool ret = true; 1926 1927 stock = &get_cpu_var(memcg_stock); 1928 if (memcg == stock->cached && stock->nr_pages) 1929 stock->nr_pages--; 1930 else /* need to call res_counter_charge */ 1931 ret = false; 1932 put_cpu_var(memcg_stock); 1933 return ret; 1934} 1935 1936/* 1937 * Returns stocks cached in percpu to res_counter and reset cached information. 1938 */ 1939static void drain_stock(struct memcg_stock_pcp *stock) 1940{ 1941 struct mem_cgroup *old = stock->cached; 1942 1943 if (stock->nr_pages) { 1944 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 1945 1946 res_counter_uncharge(&old->res, bytes); 1947 if (do_swap_account) 1948 res_counter_uncharge(&old->memsw, bytes); 1949 stock->nr_pages = 0; 1950 } 1951 stock->cached = NULL; 1952} 1953 1954/* 1955 * This must be called under preempt disabled or must be called by 1956 * a thread which is pinned to local cpu. 1957 */ 1958static void drain_local_stock(struct work_struct *dummy) 1959{ 1960 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 1961 drain_stock(stock); 1962 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); 1963} 1964 1965/* 1966 * Cache charges(val) which is from res_counter, to local per_cpu area. 1967 * This will be consumed by consume_stock() function, later. 1968 */ 1969static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) 1970{ 1971 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 1972 1973 if (stock->cached != memcg) { /* reset if necessary */ 1974 drain_stock(stock); 1975 stock->cached = memcg; 1976 } 1977 stock->nr_pages += nr_pages; 1978 put_cpu_var(memcg_stock); 1979} 1980 1981/* 1982 * Drains all per-CPU charge caches for given root_memcg resp. subtree 1983 * of the hierarchy under it. sync flag says whether we should block 1984 * until the work is done. 1985 */ 1986static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) 1987{ 1988 int cpu, curcpu; 1989 1990 /* Notify other cpus that system-wide "drain" is running */ 1991 get_online_cpus(); 1992 curcpu = get_cpu(); 1993 for_each_online_cpu(cpu) { 1994 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 1995 struct mem_cgroup *memcg; 1996 1997 memcg = stock->cached; 1998 if (!memcg || !stock->nr_pages) 1999 continue; 2000 if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) 2001 continue; 2002 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { 2003 if (cpu == curcpu) 2004 drain_local_stock(&stock->work); 2005 else 2006 schedule_work_on(cpu, &stock->work); 2007 } 2008 } 2009 put_cpu(); 2010 2011 if (!sync) 2012 goto out; 2013 2014 for_each_online_cpu(cpu) { 2015 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2016 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) 2017 flush_work(&stock->work); 2018 } 2019out: 2020 put_online_cpus(); 2021} 2022 2023/* 2024 * Tries to drain stocked charges in other cpus. This function is asynchronous 2025 * and just put a work per cpu for draining localy on each cpu. Caller can 2026 * expects some charges will be back to res_counter later but cannot wait for 2027 * it. 2028 */ 2029static void drain_all_stock_async(struct mem_cgroup *root_memcg) 2030{ 2031 /* 2032 * If someone calls draining, avoid adding more kworker runs. 2033 */ 2034 if (!mutex_trylock(&percpu_charge_mutex)) 2035 return; 2036 drain_all_stock(root_memcg, false); 2037 mutex_unlock(&percpu_charge_mutex); 2038} 2039 2040/* This is a synchronous drain interface. */ 2041static void drain_all_stock_sync(struct mem_cgroup *root_memcg) 2042{ 2043 /* called when force_empty is called */ 2044 mutex_lock(&percpu_charge_mutex); 2045 drain_all_stock(root_memcg, true); 2046 mutex_unlock(&percpu_charge_mutex); 2047} 2048 2049/* 2050 * This function drains percpu counter value from DEAD cpu and 2051 * move it to local cpu. Note that this function can be preempted. 2052 */ 2053static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) 2054{ 2055 int i; 2056 2057 spin_lock(&memcg->pcp_counter_lock); 2058 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 2059 long x = per_cpu(memcg->stat->count[i], cpu); 2060 2061 per_cpu(memcg->stat->count[i], cpu) = 0; 2062 memcg->nocpu_base.count[i] += x; 2063 } 2064 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { 2065 unsigned long x = per_cpu(memcg->stat->events[i], cpu); 2066 2067 per_cpu(memcg->stat->events[i], cpu) = 0; 2068 memcg->nocpu_base.events[i] += x; 2069 } 2070 /* need to clear ON_MOVE value, works as a kind of lock. */ 2071 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 2072 spin_unlock(&memcg->pcp_counter_lock); 2073} 2074 2075static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu) 2076{ 2077 int idx = MEM_CGROUP_ON_MOVE; 2078 2079 spin_lock(&memcg->pcp_counter_lock); 2080 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx]; 2081 spin_unlock(&memcg->pcp_counter_lock); 2082} 2083 2084static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 2085 unsigned long action, 2086 void *hcpu) 2087{ 2088 int cpu = (unsigned long)hcpu; 2089 struct memcg_stock_pcp *stock; 2090 struct mem_cgroup *iter; 2091 2092 if ((action == CPU_ONLINE)) { 2093 for_each_mem_cgroup(iter) 2094 synchronize_mem_cgroup_on_move(iter, cpu); 2095 return NOTIFY_OK; 2096 } 2097 2098 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 2099 return NOTIFY_OK; 2100 2101 for_each_mem_cgroup(iter) 2102 mem_cgroup_drain_pcp_counter(iter, cpu); 2103 2104 stock = &per_cpu(memcg_stock, cpu); 2105 drain_stock(stock); 2106 return NOTIFY_OK; 2107} 2108 2109 2110/* See __mem_cgroup_try_charge() for details */ 2111enum { 2112 CHARGE_OK, /* success */ 2113 CHARGE_RETRY, /* need to retry but retry is not bad */ 2114 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 2115 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 2116 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 2117}; 2118 2119static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, 2120 unsigned int nr_pages, bool oom_check) 2121{ 2122 unsigned long csize = nr_pages * PAGE_SIZE; 2123 struct mem_cgroup *mem_over_limit; 2124 struct res_counter *fail_res; 2125 unsigned long flags = 0; 2126 int ret; 2127 2128 ret = res_counter_charge(&memcg->res, csize, &fail_res); 2129 2130 if (likely(!ret)) { 2131 if (!do_swap_account) 2132 return CHARGE_OK; 2133 ret = res_counter_charge(&memcg->memsw, csize, &fail_res); 2134 if (likely(!ret)) 2135 return CHARGE_OK; 2136 2137 res_counter_uncharge(&memcg->res, csize); 2138 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 2139 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 2140 } else 2141 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 2142 /* 2143 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch 2144 * of regular pages (CHARGE_BATCH), or a single regular page (1). 2145 * 2146 * Never reclaim on behalf of optional batching, retry with a 2147 * single page instead. 2148 */ 2149 if (nr_pages == CHARGE_BATCH) 2150 return CHARGE_RETRY; 2151 2152 if (!(gfp_mask & __GFP_WAIT)) 2153 return CHARGE_WOULDBLOCK; 2154 2155 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); 2156 if (mem_cgroup_margin(mem_over_limit) >= nr_pages) 2157 return CHARGE_RETRY; 2158 /* 2159 * Even though the limit is exceeded at this point, reclaim 2160 * may have been able to free some pages. Retry the charge 2161 * before killing the task. 2162 * 2163 * Only for regular pages, though: huge pages are rather 2164 * unlikely to succeed so close to the limit, and we fall back 2165 * to regular pages anyway in case of failure. 2166 */ 2167 if (nr_pages == 1 && ret) 2168 return CHARGE_RETRY; 2169 2170 /* 2171 * At task move, charge accounts can be doubly counted. So, it's 2172 * better to wait until the end of task_move if something is going on. 2173 */ 2174 if (mem_cgroup_wait_acct_move(mem_over_limit)) 2175 return CHARGE_RETRY; 2176 2177 /* If we don't need to call oom-killer at el, return immediately */ 2178 if (!oom_check) 2179 return CHARGE_NOMEM; 2180 /* check OOM */ 2181 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 2182 return CHARGE_OOM_DIE; 2183 2184 return CHARGE_RETRY; 2185} 2186 2187/* 2188 * __mem_cgroup_try_charge() does 2189 * 1. detect memcg to be charged against from passed *mm and *ptr, 2190 * 2. update res_counter 2191 * 3. call memory reclaim if necessary. 2192 * 2193 * In some special case, if the task is fatal, fatal_signal_pending() or 2194 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup 2195 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon 2196 * as possible without any hazards. 2: all pages should have a valid 2197 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg 2198 * pointer, that is treated as a charge to root_mem_cgroup. 2199 * 2200 * So __mem_cgroup_try_charge() will return 2201 * 0 ... on success, filling *ptr with a valid memcg pointer. 2202 * -ENOMEM ... charge failure because of resource limits. 2203 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup. 2204 * 2205 * Unlike the exported interface, an "oom" parameter is added. if oom==true, 2206 * the oom-killer can be invoked. 2207 */ 2208static int __mem_cgroup_try_charge(struct mm_struct *mm, 2209 gfp_t gfp_mask, 2210 unsigned int nr_pages, 2211 struct mem_cgroup **ptr, 2212 bool oom) 2213{ 2214 unsigned int batch = max(CHARGE_BATCH, nr_pages); 2215 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2216 struct mem_cgroup *memcg = NULL; 2217 int ret; 2218 2219 /* 2220 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 2221 * in system level. So, allow to go ahead dying process in addition to 2222 * MEMDIE process. 2223 */ 2224 if (unlikely(test_thread_flag(TIF_MEMDIE) 2225 || fatal_signal_pending(current))) 2226 goto bypass; 2227 2228 /* 2229 * We always charge the cgroup the mm_struct belongs to. 2230 * The mm_struct's mem_cgroup changes on task migration if the 2231 * thread group leader migrates. It's possible that mm is not 2232 * set, if so charge the init_mm (happens for pagecache usage). 2233 */ 2234 if (!*ptr && !mm) 2235 *ptr = root_mem_cgroup; 2236again: 2237 if (*ptr) { /* css should be a valid one */ 2238 memcg = *ptr; 2239 VM_BUG_ON(css_is_removed(&memcg->css)); 2240 if (mem_cgroup_is_root(memcg)) 2241 goto done; 2242 if (nr_pages == 1 && consume_stock(memcg)) 2243 goto done; 2244 css_get(&memcg->css); 2245 } else { 2246 struct task_struct *p; 2247 2248 rcu_read_lock(); 2249 p = rcu_dereference(mm->owner); 2250 /* 2251 * Because we don't have task_lock(), "p" can exit. 2252 * In that case, "memcg" can point to root or p can be NULL with 2253 * race with swapoff. Then, we have small risk of mis-accouning. 2254 * But such kind of mis-account by race always happens because 2255 * we don't have cgroup_mutex(). It's overkill and we allo that 2256 * small race, here. 2257 * (*) swapoff at el will charge against mm-struct not against 2258 * task-struct. So, mm->owner can be NULL. 2259 */ 2260 memcg = mem_cgroup_from_task(p); 2261 if (!memcg) 2262 memcg = root_mem_cgroup; 2263 if (mem_cgroup_is_root(memcg)) { 2264 rcu_read_unlock(); 2265 goto done; 2266 } 2267 if (nr_pages == 1 && consume_stock(memcg)) { 2268 /* 2269 * It seems dagerous to access memcg without css_get(). 2270 * But considering how consume_stok works, it's not 2271 * necessary. If consume_stock success, some charges 2272 * from this memcg are cached on this cpu. So, we 2273 * don't need to call css_get()/css_tryget() before 2274 * calling consume_stock(). 2275 */ 2276 rcu_read_unlock(); 2277 goto done; 2278 } 2279 /* after here, we may be blocked. we need to get refcnt */ 2280 if (!css_tryget(&memcg->css)) { 2281 rcu_read_unlock(); 2282 goto again; 2283 } 2284 rcu_read_unlock(); 2285 } 2286 2287 do { 2288 bool oom_check; 2289 2290 /* If killed, bypass charge */ 2291 if (fatal_signal_pending(current)) { 2292 css_put(&memcg->css); 2293 goto bypass; 2294 } 2295 2296 oom_check = false; 2297 if (oom && !nr_oom_retries) { 2298 oom_check = true; 2299 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2300 } 2301 2302 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); 2303 switch (ret) { 2304 case CHARGE_OK: 2305 break; 2306 case CHARGE_RETRY: /* not in OOM situation but retry */ 2307 batch = nr_pages; 2308 css_put(&memcg->css); 2309 memcg = NULL; 2310 goto again; 2311 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2312 css_put(&memcg->css); 2313 goto nomem; 2314 case CHARGE_NOMEM: /* OOM routine works */ 2315 if (!oom) { 2316 css_put(&memcg->css); 2317 goto nomem; 2318 } 2319 /* If oom, we never return -ENOMEM */ 2320 nr_oom_retries--; 2321 break; 2322 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2323 css_put(&memcg->css); 2324 goto bypass; 2325 } 2326 } while (ret != CHARGE_OK); 2327 2328 if (batch > nr_pages) 2329 refill_stock(memcg, batch - nr_pages); 2330 css_put(&memcg->css); 2331done: 2332 *ptr = memcg; 2333 return 0; 2334nomem: 2335 *ptr = NULL; 2336 return -ENOMEM; 2337bypass: 2338 *ptr = root_mem_cgroup; 2339 return -EINTR; 2340} 2341 2342/* 2343 * Somemtimes we have to undo a charge we got by try_charge(). 2344 * This function is for that and do uncharge, put css's refcnt. 2345 * gotten by try_charge(). 2346 */ 2347static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, 2348 unsigned int nr_pages) 2349{ 2350 if (!mem_cgroup_is_root(memcg)) { 2351 unsigned long bytes = nr_pages * PAGE_SIZE; 2352 2353 res_counter_uncharge(&memcg->res, bytes); 2354 if (do_swap_account) 2355 res_counter_uncharge(&memcg->memsw, bytes); 2356 } 2357} 2358 2359/* 2360 * A helper function to get mem_cgroup from ID. must be called under 2361 * rcu_read_lock(). The caller must check css_is_removed() or some if 2362 * it's concern. (dropping refcnt from swap can be called against removed 2363 * memcg.) 2364 */ 2365static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2366{ 2367 struct cgroup_subsys_state *css; 2368 2369 /* ID 0 is unused ID */ 2370 if (!id) 2371 return NULL; 2372 css = css_lookup(&mem_cgroup_subsys, id); 2373 if (!css) 2374 return NULL; 2375 return container_of(css, struct mem_cgroup, css); 2376} 2377 2378struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2379{ 2380 struct mem_cgroup *memcg = NULL; 2381 struct page_cgroup *pc; 2382 unsigned short id; 2383 swp_entry_t ent; 2384 2385 VM_BUG_ON(!PageLocked(page)); 2386 2387 pc = lookup_page_cgroup(page); 2388 lock_page_cgroup(pc); 2389 if (PageCgroupUsed(pc)) { 2390 memcg = pc->mem_cgroup; 2391 if (memcg && !css_tryget(&memcg->css)) 2392 memcg = NULL; 2393 } else if (PageSwapCache(page)) { 2394 ent.val = page_private(page); 2395 id = lookup_swap_cgroup_id(ent); 2396 rcu_read_lock(); 2397 memcg = mem_cgroup_lookup(id); 2398 if (memcg && !css_tryget(&memcg->css)) 2399 memcg = NULL; 2400 rcu_read_unlock(); 2401 } 2402 unlock_page_cgroup(pc); 2403 return memcg; 2404} 2405 2406static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, 2407 struct page *page, 2408 unsigned int nr_pages, 2409 struct page_cgroup *pc, 2410 enum charge_type ctype) 2411{ 2412 lock_page_cgroup(pc); 2413 if (unlikely(PageCgroupUsed(pc))) { 2414 unlock_page_cgroup(pc); 2415 __mem_cgroup_cancel_charge(memcg, nr_pages); 2416 return; 2417 } 2418 /* 2419 * we don't need page_cgroup_lock about tail pages, becase they are not 2420 * accessed by any other context at this point. 2421 */ 2422 pc->mem_cgroup = memcg; 2423 /* 2424 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2425 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2426 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2427 * before USED bit, we need memory barrier here. 2428 * See mem_cgroup_add_lru_list(), etc. 2429 */ 2430 smp_wmb(); 2431 switch (ctype) { 2432 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2433 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2434 SetPageCgroupCache(pc); 2435 SetPageCgroupUsed(pc); 2436 break; 2437 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2438 ClearPageCgroupCache(pc); 2439 SetPageCgroupUsed(pc); 2440 break; 2441 default: 2442 break; 2443 } 2444 2445 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages); 2446 unlock_page_cgroup(pc); 2447 WARN_ON_ONCE(PageLRU(page)); 2448 /* 2449 * "charge_statistics" updated event counter. Then, check it. 2450 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2451 * if they exceeds softlimit. 2452 */ 2453 memcg_check_events(memcg, page); 2454} 2455 2456#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2457 2458#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2459 (1 << PCG_MIGRATION)) 2460/* 2461 * Because tail pages are not marked as "used", set it. We're under 2462 * zone->lru_lock, 'splitting on pmd' and compound_lock. 2463 * charge/uncharge will be never happen and move_account() is done under 2464 * compound_lock(), so we don't have to take care of races. 2465 */ 2466void mem_cgroup_split_huge_fixup(struct page *head) 2467{ 2468 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2469 struct page_cgroup *pc; 2470 int i; 2471 2472 if (mem_cgroup_disabled()) 2473 return; 2474 for (i = 1; i < HPAGE_PMD_NR; i++) { 2475 pc = head_pc + i; 2476 pc->mem_cgroup = head_pc->mem_cgroup; 2477 smp_wmb();/* see __commit_charge() */ 2478 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2479 } 2480} 2481#endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 2482 2483/** 2484 * mem_cgroup_move_account - move account of the page 2485 * @page: the page 2486 * @nr_pages: number of regular pages (>1 for huge pages) 2487 * @pc: page_cgroup of the page. 2488 * @from: mem_cgroup which the page is moved from. 2489 * @to: mem_cgroup which the page is moved to. @from != @to. 2490 * @uncharge: whether we should call uncharge and css_put against @from. 2491 * 2492 * The caller must confirm following. 2493 * - page is not on LRU (isolate_page() is useful.) 2494 * - compound_lock is held when nr_pages > 1 2495 * 2496 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2497 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is 2498 * true, this function does "uncharge" from old cgroup, but it doesn't if 2499 * @uncharge is false, so a caller should do "uncharge". 2500 */ 2501static int mem_cgroup_move_account(struct page *page, 2502 unsigned int nr_pages, 2503 struct page_cgroup *pc, 2504 struct mem_cgroup *from, 2505 struct mem_cgroup *to, 2506 bool uncharge) 2507{ 2508 unsigned long flags; 2509 int ret; 2510 2511 VM_BUG_ON(from == to); 2512 VM_BUG_ON(PageLRU(page)); 2513 /* 2514 * The page is isolated from LRU. So, collapse function 2515 * will not handle this page. But page splitting can happen. 2516 * Do this check under compound_page_lock(). The caller should 2517 * hold it. 2518 */ 2519 ret = -EBUSY; 2520 if (nr_pages > 1 && !PageTransHuge(page)) 2521 goto out; 2522 2523 lock_page_cgroup(pc); 2524 2525 ret = -EINVAL; 2526 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2527 goto unlock; 2528 2529 move_lock_page_cgroup(pc, &flags); 2530 2531 if (PageCgroupFileMapped(pc)) { 2532 /* Update mapped_file data for mem_cgroup */ 2533 preempt_disable(); 2534 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2535 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2536 preempt_enable(); 2537 } 2538 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2539 if (uncharge) 2540 /* This is not "cancel", but cancel_charge does all we need. */ 2541 __mem_cgroup_cancel_charge(from, nr_pages); 2542 2543 /* caller should have done css_get */ 2544 pc->mem_cgroup = to; 2545 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2546 /* 2547 * We charges against "to" which may not have any tasks. Then, "to" 2548 * can be under rmdir(). But in current implementation, caller of 2549 * this function is just force_empty() and move charge, so it's 2550 * guaranteed that "to" is never removed. So, we don't check rmdir 2551 * status here. 2552 */ 2553 move_unlock_page_cgroup(pc, &flags); 2554 ret = 0; 2555unlock: 2556 unlock_page_cgroup(pc); 2557 /* 2558 * check events 2559 */ 2560 memcg_check_events(to, page); 2561 memcg_check_events(from, page); 2562out: 2563 return ret; 2564} 2565 2566/* 2567 * move charges to its parent. 2568 */ 2569 2570static int mem_cgroup_move_parent(struct page *page, 2571 struct page_cgroup *pc, 2572 struct mem_cgroup *child, 2573 gfp_t gfp_mask) 2574{ 2575 struct cgroup *cg = child->css.cgroup; 2576 struct cgroup *pcg = cg->parent; 2577 struct mem_cgroup *parent; 2578 unsigned int nr_pages; 2579 unsigned long uninitialized_var(flags); 2580 int ret; 2581 2582 /* Is ROOT ? */ 2583 if (!pcg) 2584 return -EINVAL; 2585 2586 ret = -EBUSY; 2587 if (!get_page_unless_zero(page)) 2588 goto out; 2589 if (isolate_lru_page(page)) 2590 goto put; 2591 2592 nr_pages = hpage_nr_pages(page); 2593 2594 parent = mem_cgroup_from_cont(pcg); 2595 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); 2596 if (ret) 2597 goto put_back; 2598 2599 if (nr_pages > 1) 2600 flags = compound_lock_irqsave(page); 2601 2602 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); 2603 if (ret) 2604 __mem_cgroup_cancel_charge(parent, nr_pages); 2605 2606 if (nr_pages > 1) 2607 compound_unlock_irqrestore(page, flags); 2608put_back: 2609 putback_lru_page(page); 2610put: 2611 put_page(page); 2612out: 2613 return ret; 2614} 2615 2616/* 2617 * Charge the memory controller for page usage. 2618 * Return 2619 * 0 if the charge was successful 2620 * < 0 if the cgroup is over its limit 2621 */ 2622static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2623 gfp_t gfp_mask, enum charge_type ctype) 2624{ 2625 struct mem_cgroup *memcg = NULL; 2626 unsigned int nr_pages = 1; 2627 struct page_cgroup *pc; 2628 bool oom = true; 2629 int ret; 2630 2631 if (PageTransHuge(page)) { 2632 nr_pages <<= compound_order(page); 2633 VM_BUG_ON(!PageTransHuge(page)); 2634 /* 2635 * Never OOM-kill a process for a huge page. The 2636 * fault handler will fall back to regular pages. 2637 */ 2638 oom = false; 2639 } 2640 2641 pc = lookup_page_cgroup(page); 2642 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); 2643 if (ret == -ENOMEM) 2644 return ret; 2645 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype); 2646 return 0; 2647} 2648 2649int mem_cgroup_newpage_charge(struct page *page, 2650 struct mm_struct *mm, gfp_t gfp_mask) 2651{ 2652 if (mem_cgroup_disabled()) 2653 return 0; 2654 VM_BUG_ON(page_mapped(page)); 2655 VM_BUG_ON(page->mapping && !PageAnon(page)); 2656 VM_BUG_ON(!mm); 2657 return mem_cgroup_charge_common(page, mm, gfp_mask, 2658 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2659} 2660 2661static void 2662__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2663 enum charge_type ctype); 2664 2665static void 2666__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg, 2667 enum charge_type ctype) 2668{ 2669 struct page_cgroup *pc = lookup_page_cgroup(page); 2670 struct zone *zone = page_zone(page); 2671 unsigned long flags; 2672 bool removed = false; 2673 2674 /* 2675 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page 2676 * is already on LRU. It means the page may on some other page_cgroup's 2677 * LRU. Take care of it. 2678 */ 2679 spin_lock_irqsave(&zone->lru_lock, flags); 2680 if (PageLRU(page)) { 2681 del_page_from_lru_list(zone, page, page_lru(page)); 2682 ClearPageLRU(page); 2683 removed = true; 2684 } 2685 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 2686 if (removed) { 2687 add_page_to_lru_list(zone, page, page_lru(page)); 2688 SetPageLRU(page); 2689 } 2690 spin_unlock_irqrestore(&zone->lru_lock, flags); 2691 return; 2692} 2693 2694int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2695 gfp_t gfp_mask) 2696{ 2697 struct mem_cgroup *memcg = NULL; 2698 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; 2699 int ret; 2700 2701 if (mem_cgroup_disabled()) 2702 return 0; 2703 if (PageCompound(page)) 2704 return 0; 2705 2706 if (unlikely(!mm)) 2707 mm = &init_mm; 2708 if (!page_is_file_cache(page)) 2709 type = MEM_CGROUP_CHARGE_TYPE_SHMEM; 2710 2711 if (!PageSwapCache(page)) 2712 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type); 2713 else { /* page is swapcache/shmem */ 2714 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); 2715 if (!ret) 2716 __mem_cgroup_commit_charge_swapin(page, memcg, type); 2717 } 2718 return ret; 2719} 2720 2721/* 2722 * While swap-in, try_charge -> commit or cancel, the page is locked. 2723 * And when try_charge() successfully returns, one refcnt to memcg without 2724 * struct page_cgroup is acquired. This refcnt will be consumed by 2725 * "commit()" or removed by "cancel()" 2726 */ 2727int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2728 struct page *page, 2729 gfp_t mask, struct mem_cgroup **memcgp) 2730{ 2731 struct mem_cgroup *memcg; 2732 int ret; 2733 2734 *memcgp = NULL; 2735 2736 if (mem_cgroup_disabled()) 2737 return 0; 2738 2739 if (!do_swap_account) 2740 goto charge_cur_mm; 2741 /* 2742 * A racing thread's fault, or swapoff, may have already updated 2743 * the pte, and even removed page from swap cache: in those cases 2744 * do_swap_page()'s pte_same() test will fail; but there's also a 2745 * KSM case which does need to charge the page. 2746 */ 2747 if (!PageSwapCache(page)) 2748 goto charge_cur_mm; 2749 memcg = try_get_mem_cgroup_from_page(page); 2750 if (!memcg) 2751 goto charge_cur_mm; 2752 *memcgp = memcg; 2753 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); 2754 css_put(&memcg->css); 2755 if (ret == -EINTR) 2756 ret = 0; 2757 return ret; 2758charge_cur_mm: 2759 if (unlikely(!mm)) 2760 mm = &init_mm; 2761 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); 2762 if (ret == -EINTR) 2763 ret = 0; 2764 return ret; 2765} 2766 2767static void 2768__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, 2769 enum charge_type ctype) 2770{ 2771 if (mem_cgroup_disabled()) 2772 return; 2773 if (!memcg) 2774 return; 2775 cgroup_exclude_rmdir(&memcg->css); 2776 2777 __mem_cgroup_commit_charge_lrucare(page, memcg, ctype); 2778 /* 2779 * Now swap is on-memory. This means this page may be 2780 * counted both as mem and swap....double count. 2781 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2782 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2783 * may call delete_from_swap_cache() before reach here. 2784 */ 2785 if (do_swap_account && PageSwapCache(page)) { 2786 swp_entry_t ent = {.val = page_private(page)}; 2787 struct mem_cgroup *swap_memcg; 2788 unsigned short id; 2789 2790 id = swap_cgroup_record(ent, 0); 2791 rcu_read_lock(); 2792 swap_memcg = mem_cgroup_lookup(id); 2793 if (swap_memcg) { 2794 /* 2795 * This recorded memcg can be obsolete one. So, avoid 2796 * calling css_tryget 2797 */ 2798 if (!mem_cgroup_is_root(swap_memcg)) 2799 res_counter_uncharge(&swap_memcg->memsw, 2800 PAGE_SIZE); 2801 mem_cgroup_swap_statistics(swap_memcg, false); 2802 mem_cgroup_put(swap_memcg); 2803 } 2804 rcu_read_unlock(); 2805 } 2806 /* 2807 * At swapin, we may charge account against cgroup which has no tasks. 2808 * So, rmdir()->pre_destroy() can be called while we do this charge. 2809 * In that case, we need to call pre_destroy() again. check it here. 2810 */ 2811 cgroup_release_and_wakeup_rmdir(&memcg->css); 2812} 2813 2814void mem_cgroup_commit_charge_swapin(struct page *page, 2815 struct mem_cgroup *memcg) 2816{ 2817 __mem_cgroup_commit_charge_swapin(page, memcg, 2818 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2819} 2820 2821void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) 2822{ 2823 if (mem_cgroup_disabled()) 2824 return; 2825 if (!memcg) 2826 return; 2827 __mem_cgroup_cancel_charge(memcg, 1); 2828} 2829 2830static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, 2831 unsigned int nr_pages, 2832 const enum charge_type ctype) 2833{ 2834 struct memcg_batch_info *batch = NULL; 2835 bool uncharge_memsw = true; 2836 2837 /* If swapout, usage of swap doesn't decrease */ 2838 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2839 uncharge_memsw = false; 2840 2841 batch = ¤t->memcg_batch; 2842 /* 2843 * In usual, we do css_get() when we remember memcg pointer. 2844 * But in this case, we keep res->usage until end of a series of 2845 * uncharges. Then, it's ok to ignore memcg's refcnt. 2846 */ 2847 if (!batch->memcg) 2848 batch->memcg = memcg; 2849 /* 2850 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2851 * In those cases, all pages freed continuously can be expected to be in 2852 * the same cgroup and we have chance to coalesce uncharges. 2853 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2854 * because we want to do uncharge as soon as possible. 2855 */ 2856 2857 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2858 goto direct_uncharge; 2859 2860 if (nr_pages > 1) 2861 goto direct_uncharge; 2862 2863 /* 2864 * In typical case, batch->memcg == mem. This means we can 2865 * merge a series of uncharges to an uncharge of res_counter. 2866 * If not, we uncharge res_counter ony by one. 2867 */ 2868 if (batch->memcg != memcg) 2869 goto direct_uncharge; 2870 /* remember freed charge and uncharge it later */ 2871 batch->nr_pages++; 2872 if (uncharge_memsw) 2873 batch->memsw_nr_pages++; 2874 return; 2875direct_uncharge: 2876 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); 2877 if (uncharge_memsw) 2878 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); 2879 if (unlikely(batch->memcg != memcg)) 2880 memcg_oom_recover(memcg); 2881 return; 2882} 2883 2884/* 2885 * uncharge if !page_mapped(page) 2886 */ 2887static struct mem_cgroup * 2888__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 2889{ 2890 struct mem_cgroup *memcg = NULL; 2891 unsigned int nr_pages = 1; 2892 struct page_cgroup *pc; 2893 2894 if (mem_cgroup_disabled()) 2895 return NULL; 2896 2897 if (PageSwapCache(page)) 2898 return NULL; 2899 2900 if (PageTransHuge(page)) { 2901 nr_pages <<= compound_order(page); 2902 VM_BUG_ON(!PageTransHuge(page)); 2903 } 2904 /* 2905 * Check if our page_cgroup is valid 2906 */ 2907 pc = lookup_page_cgroup(page); 2908 if (unlikely(!PageCgroupUsed(pc))) 2909 return NULL; 2910 2911 lock_page_cgroup(pc); 2912 2913 memcg = pc->mem_cgroup; 2914 2915 if (!PageCgroupUsed(pc)) 2916 goto unlock_out; 2917 2918 switch (ctype) { 2919 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2920 case MEM_CGROUP_CHARGE_TYPE_DROP: 2921 /* See mem_cgroup_prepare_migration() */ 2922 if (page_mapped(page) || PageCgroupMigration(pc)) 2923 goto unlock_out; 2924 break; 2925 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 2926 if (!PageAnon(page)) { /* Shared memory */ 2927 if (page->mapping && !page_is_file_cache(page)) 2928 goto unlock_out; 2929 } else if (page_mapped(page)) /* Anon */ 2930 goto unlock_out; 2931 break; 2932 default: 2933 break; 2934 } 2935 2936 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages); 2937 2938 ClearPageCgroupUsed(pc); 2939 /* 2940 * pc->mem_cgroup is not cleared here. It will be accessed when it's 2941 * freed from LRU. This is safe because uncharged page is expected not 2942 * to be reused (freed soon). Exception is SwapCache, it's handled by 2943 * special functions. 2944 */ 2945 2946 unlock_page_cgroup(pc); 2947 /* 2948 * even after unlock, we have memcg->res.usage here and this memcg 2949 * will never be freed. 2950 */ 2951 memcg_check_events(memcg, page); 2952 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 2953 mem_cgroup_swap_statistics(memcg, true); 2954 mem_cgroup_get(memcg); 2955 } 2956 if (!mem_cgroup_is_root(memcg)) 2957 mem_cgroup_do_uncharge(memcg, nr_pages, ctype); 2958 2959 return memcg; 2960 2961unlock_out: 2962 unlock_page_cgroup(pc); 2963 return NULL; 2964} 2965 2966void mem_cgroup_uncharge_page(struct page *page) 2967{ 2968 /* early check. */ 2969 if (page_mapped(page)) 2970 return; 2971 VM_BUG_ON(page->mapping && !PageAnon(page)); 2972 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 2973} 2974 2975void mem_cgroup_uncharge_cache_page(struct page *page) 2976{ 2977 VM_BUG_ON(page_mapped(page)); 2978 VM_BUG_ON(page->mapping); 2979 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 2980} 2981 2982/* 2983 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 2984 * In that cases, pages are freed continuously and we can expect pages 2985 * are in the same memcg. All these calls itself limits the number of 2986 * pages freed at once, then uncharge_start/end() is called properly. 2987 * This may be called prural(2) times in a context, 2988 */ 2989 2990void mem_cgroup_uncharge_start(void) 2991{ 2992 current->memcg_batch.do_batch++; 2993 /* We can do nest. */ 2994 if (current->memcg_batch.do_batch == 1) { 2995 current->memcg_batch.memcg = NULL; 2996 current->memcg_batch.nr_pages = 0; 2997 current->memcg_batch.memsw_nr_pages = 0; 2998 } 2999} 3000 3001void mem_cgroup_uncharge_end(void) 3002{ 3003 struct memcg_batch_info *batch = ¤t->memcg_batch; 3004 3005 if (!batch->do_batch) 3006 return; 3007 3008 batch->do_batch--; 3009 if (batch->do_batch) /* If stacked, do nothing. */ 3010 return; 3011 3012 if (!batch->memcg) 3013 return; 3014 /* 3015 * This "batch->memcg" is valid without any css_get/put etc... 3016 * bacause we hide charges behind us. 3017 */ 3018 if (batch->nr_pages) 3019 res_counter_uncharge(&batch->memcg->res, 3020 batch->nr_pages * PAGE_SIZE); 3021 if (batch->memsw_nr_pages) 3022 res_counter_uncharge(&batch->memcg->memsw, 3023 batch->memsw_nr_pages * PAGE_SIZE); 3024 memcg_oom_recover(batch->memcg); 3025 /* forget this pointer (for sanity check) */ 3026 batch->memcg = NULL; 3027} 3028 3029/* 3030 * A function for resetting pc->mem_cgroup for newly allocated pages. 3031 * This function should be called if the newpage will be added to LRU 3032 * before start accounting. 3033 */ 3034void mem_cgroup_reset_owner(struct page *newpage) 3035{ 3036 struct page_cgroup *pc; 3037 3038 if (mem_cgroup_disabled()) 3039 return; 3040 3041 pc = lookup_page_cgroup(newpage); 3042 VM_BUG_ON(PageCgroupUsed(pc)); 3043 pc->mem_cgroup = root_mem_cgroup; 3044} 3045 3046#ifdef CONFIG_SWAP 3047/* 3048 * called after __delete_from_swap_cache() and drop "page" account. 3049 * memcg information is recorded to swap_cgroup of "ent" 3050 */ 3051void 3052mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 3053{ 3054 struct mem_cgroup *memcg; 3055 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 3056 3057 if (!swapout) /* this was a swap cache but the swap is unused ! */ 3058 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 3059 3060 memcg = __mem_cgroup_uncharge_common(page, ctype); 3061 3062 /* 3063 * record memcg information, if swapout && memcg != NULL, 3064 * mem_cgroup_get() was called in uncharge(). 3065 */ 3066 if (do_swap_account && swapout && memcg) 3067 swap_cgroup_record(ent, css_id(&memcg->css)); 3068} 3069#endif 3070 3071#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3072/* 3073 * called from swap_entry_free(). remove record in swap_cgroup and 3074 * uncharge "memsw" account. 3075 */ 3076void mem_cgroup_uncharge_swap(swp_entry_t ent) 3077{ 3078 struct mem_cgroup *memcg; 3079 unsigned short id; 3080 3081 if (!do_swap_account) 3082 return; 3083 3084 id = swap_cgroup_record(ent, 0); 3085 rcu_read_lock(); 3086 memcg = mem_cgroup_lookup(id); 3087 if (memcg) { 3088 /* 3089 * We uncharge this because swap is freed. 3090 * This memcg can be obsolete one. We avoid calling css_tryget 3091 */ 3092 if (!mem_cgroup_is_root(memcg)) 3093 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 3094 mem_cgroup_swap_statistics(memcg, false); 3095 mem_cgroup_put(memcg); 3096 } 3097 rcu_read_unlock(); 3098} 3099 3100/** 3101 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 3102 * @entry: swap entry to be moved 3103 * @from: mem_cgroup which the entry is moved from 3104 * @to: mem_cgroup which the entry is moved to 3105 * @need_fixup: whether we should fixup res_counters and refcounts. 3106 * 3107 * It succeeds only when the swap_cgroup's record for this entry is the same 3108 * as the mem_cgroup's id of @from. 3109 * 3110 * Returns 0 on success, -EINVAL on failure. 3111 * 3112 * The caller must have charged to @to, IOW, called res_counter_charge() about 3113 * both res and memsw, and called css_get(). 3114 */ 3115static int mem_cgroup_move_swap_account(swp_entry_t entry, 3116 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3117{ 3118 unsigned short old_id, new_id; 3119 3120 old_id = css_id(&from->css); 3121 new_id = css_id(&to->css); 3122 3123 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 3124 mem_cgroup_swap_statistics(from, false); 3125 mem_cgroup_swap_statistics(to, true); 3126 /* 3127 * This function is only called from task migration context now. 3128 * It postpones res_counter and refcount handling till the end 3129 * of task migration(mem_cgroup_clear_mc()) for performance 3130 * improvement. But we cannot postpone mem_cgroup_get(to) 3131 * because if the process that has been moved to @to does 3132 * swap-in, the refcount of @to might be decreased to 0. 3133 */ 3134 mem_cgroup_get(to); 3135 if (need_fixup) { 3136 if (!mem_cgroup_is_root(from)) 3137 res_counter_uncharge(&from->memsw, PAGE_SIZE); 3138 mem_cgroup_put(from); 3139 /* 3140 * we charged both to->res and to->memsw, so we should 3141 * uncharge to->res. 3142 */ 3143 if (!mem_cgroup_is_root(to)) 3144 res_counter_uncharge(&to->res, PAGE_SIZE); 3145 } 3146 return 0; 3147 } 3148 return -EINVAL; 3149} 3150#else 3151static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 3152 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3153{ 3154 return -EINVAL; 3155} 3156#endif 3157 3158/* 3159 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 3160 * page belongs to. 3161 */ 3162int mem_cgroup_prepare_migration(struct page *page, 3163 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) 3164{ 3165 struct mem_cgroup *memcg = NULL; 3166 struct page_cgroup *pc; 3167 enum charge_type ctype; 3168 int ret = 0; 3169 3170 *memcgp = NULL; 3171 3172 VM_BUG_ON(PageTransHuge(page)); 3173 if (mem_cgroup_disabled()) 3174 return 0; 3175 3176 pc = lookup_page_cgroup(page); 3177 lock_page_cgroup(pc); 3178 if (PageCgroupUsed(pc)) { 3179 memcg = pc->mem_cgroup; 3180 css_get(&memcg->css); 3181 /* 3182 * At migrating an anonymous page, its mapcount goes down 3183 * to 0 and uncharge() will be called. But, even if it's fully 3184 * unmapped, migration may fail and this page has to be 3185 * charged again. We set MIGRATION flag here and delay uncharge 3186 * until end_migration() is called 3187 * 3188 * Corner Case Thinking 3189 * A) 3190 * When the old page was mapped as Anon and it's unmap-and-freed 3191 * while migration was ongoing. 3192 * If unmap finds the old page, uncharge() of it will be delayed 3193 * until end_migration(). If unmap finds a new page, it's 3194 * uncharged when it make mapcount to be 1->0. If unmap code 3195 * finds swap_migration_entry, the new page will not be mapped 3196 * and end_migration() will find it(mapcount==0). 3197 * 3198 * B) 3199 * When the old page was mapped but migraion fails, the kernel 3200 * remaps it. A charge for it is kept by MIGRATION flag even 3201 * if mapcount goes down to 0. We can do remap successfully 3202 * without charging it again. 3203 * 3204 * C) 3205 * The "old" page is under lock_page() until the end of 3206 * migration, so, the old page itself will not be swapped-out. 3207 * If the new page is swapped out before end_migraton, our 3208 * hook to usual swap-out path will catch the event. 3209 */ 3210 if (PageAnon(page)) 3211 SetPageCgroupMigration(pc); 3212 } 3213 unlock_page_cgroup(pc); 3214 /* 3215 * If the page is not charged at this point, 3216 * we return here. 3217 */ 3218 if (!memcg) 3219 return 0; 3220 3221 *memcgp = memcg; 3222 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); 3223 css_put(&memcg->css);/* drop extra refcnt */ 3224 if (ret) { 3225 if (PageAnon(page)) { 3226 lock_page_cgroup(pc); 3227 ClearPageCgroupMigration(pc); 3228 unlock_page_cgroup(pc); 3229 /* 3230 * The old page may be fully unmapped while we kept it. 3231 */ 3232 mem_cgroup_uncharge_page(page); 3233 } 3234 /* we'll need to revisit this error code (we have -EINTR) */ 3235 return -ENOMEM; 3236 } 3237 /* 3238 * We charge new page before it's used/mapped. So, even if unlock_page() 3239 * is called before end_migration, we can catch all events on this new 3240 * page. In the case new page is migrated but not remapped, new page's 3241 * mapcount will be finally 0 and we call uncharge in end_migration(). 3242 */ 3243 pc = lookup_page_cgroup(newpage); 3244 if (PageAnon(page)) 3245 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 3246 else if (page_is_file_cache(page)) 3247 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 3248 else 3249 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3250 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 3251 return ret; 3252} 3253 3254/* remove redundant charge if migration failed*/ 3255void mem_cgroup_end_migration(struct mem_cgroup *memcg, 3256 struct page *oldpage, struct page *newpage, bool migration_ok) 3257{ 3258 struct page *used, *unused; 3259 struct page_cgroup *pc; 3260 3261 if (!memcg) 3262 return; 3263 /* blocks rmdir() */ 3264 cgroup_exclude_rmdir(&memcg->css); 3265 if (!migration_ok) { 3266 used = oldpage; 3267 unused = newpage; 3268 } else { 3269 used = newpage; 3270 unused = oldpage; 3271 } 3272 /* 3273 * We disallowed uncharge of pages under migration because mapcount 3274 * of the page goes down to zero, temporarly. 3275 * Clear the flag and check the page should be charged. 3276 */ 3277 pc = lookup_page_cgroup(oldpage); 3278 lock_page_cgroup(pc); 3279 ClearPageCgroupMigration(pc); 3280 unlock_page_cgroup(pc); 3281 3282 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 3283 3284 /* 3285 * If a page is a file cache, radix-tree replacement is very atomic 3286 * and we can skip this check. When it was an Anon page, its mapcount 3287 * goes down to 0. But because we added MIGRATION flage, it's not 3288 * uncharged yet. There are several case but page->mapcount check 3289 * and USED bit check in mem_cgroup_uncharge_page() will do enough 3290 * check. (see prepare_charge() also) 3291 */ 3292 if (PageAnon(used)) 3293 mem_cgroup_uncharge_page(used); 3294 /* 3295 * At migration, we may charge account against cgroup which has no 3296 * tasks. 3297 * So, rmdir()->pre_destroy() can be called while we do this charge. 3298 * In that case, we need to call pre_destroy() again. check it here. 3299 */ 3300 cgroup_release_and_wakeup_rmdir(&memcg->css); 3301} 3302 3303/* 3304 * At replace page cache, newpage is not under any memcg but it's on 3305 * LRU. So, this function doesn't touch res_counter but handles LRU 3306 * in correct way. Both pages are locked so we cannot race with uncharge. 3307 */ 3308void mem_cgroup_replace_page_cache(struct page *oldpage, 3309 struct page *newpage) 3310{ 3311 struct mem_cgroup *memcg; 3312 struct page_cgroup *pc; 3313 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; 3314 3315 if (mem_cgroup_disabled()) 3316 return; 3317 3318 pc = lookup_page_cgroup(oldpage); 3319 /* fix accounting on old pages */ 3320 lock_page_cgroup(pc); 3321 memcg = pc->mem_cgroup; 3322 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1); 3323 ClearPageCgroupUsed(pc); 3324 unlock_page_cgroup(pc); 3325 3326 if (PageSwapBacked(oldpage)) 3327 type = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3328 3329 /* 3330 * Even if newpage->mapping was NULL before starting replacement, 3331 * the newpage may be on LRU(or pagevec for LRU) already. We lock 3332 * LRU while we overwrite pc->mem_cgroup. 3333 */ 3334 __mem_cgroup_commit_charge_lrucare(newpage, memcg, type); 3335} 3336 3337#ifdef CONFIG_DEBUG_VM 3338static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3339{ 3340 struct page_cgroup *pc; 3341 3342 pc = lookup_page_cgroup(page); 3343 /* 3344 * Can be NULL while feeding pages into the page allocator for 3345 * the first time, i.e. during boot or memory hotplug; 3346 * or when mem_cgroup_disabled(). 3347 */ 3348 if (likely(pc) && PageCgroupUsed(pc)) 3349 return pc; 3350 return NULL; 3351} 3352 3353bool mem_cgroup_bad_page_check(struct page *page) 3354{ 3355 if (mem_cgroup_disabled()) 3356 return false; 3357 3358 return lookup_page_cgroup_used(page) != NULL; 3359} 3360 3361void mem_cgroup_print_bad_page(struct page *page) 3362{ 3363 struct page_cgroup *pc; 3364 3365 pc = lookup_page_cgroup_used(page); 3366 if (pc) { 3367 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n", 3368 pc, pc->flags, pc->mem_cgroup); 3369 } 3370} 3371#endif 3372 3373static DEFINE_MUTEX(set_limit_mutex); 3374 3375static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3376 unsigned long long val) 3377{ 3378 int retry_count; 3379 u64 memswlimit, memlimit; 3380 int ret = 0; 3381 int children = mem_cgroup_count_children(memcg); 3382 u64 curusage, oldusage; 3383 int enlarge; 3384 3385 /* 3386 * For keeping hierarchical_reclaim simple, how long we should retry 3387 * is depends on callers. We set our retry-count to be function 3388 * of # of children which we should visit in this loop. 3389 */ 3390 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3391 3392 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3393 3394 enlarge = 0; 3395 while (retry_count) { 3396 if (signal_pending(current)) { 3397 ret = -EINTR; 3398 break; 3399 } 3400 /* 3401 * Rather than hide all in some function, I do this in 3402 * open coded manner. You see what this really does. 3403 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3404 */ 3405 mutex_lock(&set_limit_mutex); 3406 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3407 if (memswlimit < val) { 3408 ret = -EINVAL; 3409 mutex_unlock(&set_limit_mutex); 3410 break; 3411 } 3412 3413 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3414 if (memlimit < val) 3415 enlarge = 1; 3416 3417 ret = res_counter_set_limit(&memcg->res, val); 3418 if (!ret) { 3419 if (memswlimit == val) 3420 memcg->memsw_is_minimum = true; 3421 else 3422 memcg->memsw_is_minimum = false; 3423 } 3424 mutex_unlock(&set_limit_mutex); 3425 3426 if (!ret) 3427 break; 3428 3429 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3430 MEM_CGROUP_RECLAIM_SHRINK); 3431 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3432 /* Usage is reduced ? */ 3433 if (curusage >= oldusage) 3434 retry_count--; 3435 else 3436 oldusage = curusage; 3437 } 3438 if (!ret && enlarge) 3439 memcg_oom_recover(memcg); 3440 3441 return ret; 3442} 3443 3444static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3445 unsigned long long val) 3446{ 3447 int retry_count; 3448 u64 memlimit, memswlimit, oldusage, curusage; 3449 int children = mem_cgroup_count_children(memcg); 3450 int ret = -EBUSY; 3451 int enlarge = 0; 3452 3453 /* see mem_cgroup_resize_res_limit */ 3454 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3455 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3456 while (retry_count) { 3457 if (signal_pending(current)) { 3458 ret = -EINTR; 3459 break; 3460 } 3461 /* 3462 * Rather than hide all in some function, I do this in 3463 * open coded manner. You see what this really does. 3464 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3465 */ 3466 mutex_lock(&set_limit_mutex); 3467 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3468 if (memlimit > val) { 3469 ret = -EINVAL; 3470 mutex_unlock(&set_limit_mutex); 3471 break; 3472 } 3473 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3474 if (memswlimit < val) 3475 enlarge = 1; 3476 ret = res_counter_set_limit(&memcg->memsw, val); 3477 if (!ret) { 3478 if (memlimit == val) 3479 memcg->memsw_is_minimum = true; 3480 else 3481 memcg->memsw_is_minimum = false; 3482 } 3483 mutex_unlock(&set_limit_mutex); 3484 3485 if (!ret) 3486 break; 3487 3488 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3489 MEM_CGROUP_RECLAIM_NOSWAP | 3490 MEM_CGROUP_RECLAIM_SHRINK); 3491 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3492 /* Usage is reduced ? */ 3493 if (curusage >= oldusage) 3494 retry_count--; 3495 else 3496 oldusage = curusage; 3497 } 3498 if (!ret && enlarge) 3499 memcg_oom_recover(memcg); 3500 return ret; 3501} 3502 3503unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3504 gfp_t gfp_mask, 3505 unsigned long *total_scanned) 3506{ 3507 unsigned long nr_reclaimed = 0; 3508 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3509 unsigned long reclaimed; 3510 int loop = 0; 3511 struct mem_cgroup_tree_per_zone *mctz; 3512 unsigned long long excess; 3513 unsigned long nr_scanned; 3514 3515 if (order > 0) 3516 return 0; 3517 3518 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3519 /* 3520 * This loop can run a while, specially if mem_cgroup's continuously 3521 * keep exceeding their soft limit and putting the system under 3522 * pressure 3523 */ 3524 do { 3525 if (next_mz) 3526 mz = next_mz; 3527 else 3528 mz = mem_cgroup_largest_soft_limit_node(mctz); 3529 if (!mz) 3530 break; 3531 3532 nr_scanned = 0; 3533 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone, 3534 gfp_mask, &nr_scanned); 3535 nr_reclaimed += reclaimed; 3536 *total_scanned += nr_scanned; 3537 spin_lock(&mctz->lock); 3538 3539 /* 3540 * If we failed to reclaim anything from this memory cgroup 3541 * it is time to move on to the next cgroup 3542 */ 3543 next_mz = NULL; 3544 if (!reclaimed) { 3545 do { 3546 /* 3547 * Loop until we find yet another one. 3548 * 3549 * By the time we get the soft_limit lock 3550 * again, someone might have aded the 3551 * group back on the RB tree. Iterate to 3552 * make sure we get a different mem. 3553 * mem_cgroup_largest_soft_limit_node returns 3554 * NULL if no other cgroup is present on 3555 * the tree 3556 */ 3557 next_mz = 3558 __mem_cgroup_largest_soft_limit_node(mctz); 3559 if (next_mz == mz) 3560 css_put(&next_mz->mem->css); 3561 else /* next_mz == NULL or other memcg */ 3562 break; 3563 } while (1); 3564 } 3565 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3566 excess = res_counter_soft_limit_excess(&mz->mem->res); 3567 /* 3568 * One school of thought says that we should not add 3569 * back the node to the tree if reclaim returns 0. 3570 * But our reclaim could return 0, simply because due 3571 * to priority we are exposing a smaller subset of 3572 * memory to reclaim from. Consider this as a longer 3573 * term TODO. 3574 */ 3575 /* If excess == 0, no tree ops */ 3576 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3577 spin_unlock(&mctz->lock); 3578 css_put(&mz->mem->css); 3579 loop++; 3580 /* 3581 * Could not reclaim anything and there are no more 3582 * mem cgroups to try or we seem to be looping without 3583 * reclaiming anything. 3584 */ 3585 if (!nr_reclaimed && 3586 (next_mz == NULL || 3587 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3588 break; 3589 } while (!nr_reclaimed); 3590 if (next_mz) 3591 css_put(&next_mz->mem->css); 3592 return nr_reclaimed; 3593} 3594 3595/* 3596 * This routine traverse page_cgroup in given list and drop them all. 3597 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3598 */ 3599static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, 3600 int node, int zid, enum lru_list lru) 3601{ 3602 struct mem_cgroup_per_zone *mz; 3603 unsigned long flags, loop; 3604 struct list_head *list; 3605 struct page *busy; 3606 struct zone *zone; 3607 int ret = 0; 3608 3609 zone = &NODE_DATA(node)->node_zones[zid]; 3610 mz = mem_cgroup_zoneinfo(memcg, node, zid); 3611 list = &mz->lruvec.lists[lru]; 3612 3613 loop = MEM_CGROUP_ZSTAT(mz, lru); 3614 /* give some margin against EBUSY etc...*/ 3615 loop += 256; 3616 busy = NULL; 3617 while (loop--) { 3618 struct page_cgroup *pc; 3619 struct page *page; 3620 3621 ret = 0; 3622 spin_lock_irqsave(&zone->lru_lock, flags); 3623 if (list_empty(list)) { 3624 spin_unlock_irqrestore(&zone->lru_lock, flags); 3625 break; 3626 } 3627 page = list_entry(list->prev, struct page, lru); 3628 if (busy == page) { 3629 list_move(&page->lru, list); 3630 busy = NULL; 3631 spin_unlock_irqrestore(&zone->lru_lock, flags); 3632 continue; 3633 } 3634 spin_unlock_irqrestore(&zone->lru_lock, flags); 3635 3636 pc = lookup_page_cgroup(page); 3637 3638 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); 3639 if (ret == -ENOMEM || ret == -EINTR) 3640 break; 3641 3642 if (ret == -EBUSY || ret == -EINVAL) { 3643 /* found lock contention or "pc" is obsolete. */ 3644 busy = page; 3645 cond_resched(); 3646 } else 3647 busy = NULL; 3648 } 3649 3650 if (!ret && !list_empty(list)) 3651 return -EBUSY; 3652 return ret; 3653} 3654 3655/* 3656 * make mem_cgroup's charge to be 0 if there is no task. 3657 * This enables deleting this mem_cgroup. 3658 */ 3659static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) 3660{ 3661 int ret; 3662 int node, zid, shrink; 3663 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3664 struct cgroup *cgrp = memcg->css.cgroup; 3665 3666 css_get(&memcg->css); 3667 3668 shrink = 0; 3669 /* should free all ? */ 3670 if (free_all) 3671 goto try_to_free; 3672move_account: 3673 do { 3674 ret = -EBUSY; 3675 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3676 goto out; 3677 ret = -EINTR; 3678 if (signal_pending(current)) 3679 goto out; 3680 /* This is for making all *used* pages to be on LRU. */ 3681 lru_add_drain_all(); 3682 drain_all_stock_sync(memcg); 3683 ret = 0; 3684 mem_cgroup_start_move(memcg); 3685 for_each_node_state(node, N_HIGH_MEMORY) { 3686 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3687 enum lru_list l; 3688 for_each_lru(l) { 3689 ret = mem_cgroup_force_empty_list(memcg, 3690 node, zid, l); 3691 if (ret) 3692 break; 3693 } 3694 } 3695 if (ret) 3696 break; 3697 } 3698 mem_cgroup_end_move(memcg); 3699 memcg_oom_recover(memcg); 3700 /* it seems parent cgroup doesn't have enough mem */ 3701 if (ret == -ENOMEM) 3702 goto try_to_free; 3703 cond_resched(); 3704 /* "ret" should also be checked to ensure all lists are empty. */ 3705 } while (memcg->res.usage > 0 || ret); 3706out: 3707 css_put(&memcg->css); 3708 return ret; 3709 3710try_to_free: 3711 /* returns EBUSY if there is a task or if we come here twice. */ 3712 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3713 ret = -EBUSY; 3714 goto out; 3715 } 3716 /* we call try-to-free pages for make this cgroup empty */ 3717 lru_add_drain_all(); 3718 /* try to free all pages in this cgroup */ 3719 shrink = 1; 3720 while (nr_retries && memcg->res.usage > 0) { 3721 int progress; 3722 3723 if (signal_pending(current)) { 3724 ret = -EINTR; 3725 goto out; 3726 } 3727 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, 3728 false); 3729 if (!progress) { 3730 nr_retries--; 3731 /* maybe some writeback is necessary */ 3732 congestion_wait(BLK_RW_ASYNC, HZ/10); 3733 } 3734 3735 } 3736 lru_add_drain(); 3737 /* try move_account...there may be some *locked* pages. */ 3738 goto move_account; 3739} 3740 3741int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3742{ 3743 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3744} 3745 3746 3747static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3748{ 3749 return mem_cgroup_from_cont(cont)->use_hierarchy; 3750} 3751 3752static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3753 u64 val) 3754{ 3755 int retval = 0; 3756 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3757 struct cgroup *parent = cont->parent; 3758 struct mem_cgroup *parent_memcg = NULL; 3759 3760 if (parent) 3761 parent_memcg = mem_cgroup_from_cont(parent); 3762 3763 cgroup_lock(); 3764 /* 3765 * If parent's use_hierarchy is set, we can't make any modifications 3766 * in the child subtrees. If it is unset, then the change can 3767 * occur, provided the current cgroup has no children. 3768 * 3769 * For the root cgroup, parent_mem is NULL, we allow value to be 3770 * set if there are no children. 3771 */ 3772 if ((!parent_memcg || !parent_memcg->use_hierarchy) && 3773 (val == 1 || val == 0)) { 3774 if (list_empty(&cont->children)) 3775 memcg->use_hierarchy = val; 3776 else 3777 retval = -EBUSY; 3778 } else 3779 retval = -EINVAL; 3780 cgroup_unlock(); 3781 3782 return retval; 3783} 3784 3785 3786static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, 3787 enum mem_cgroup_stat_index idx) 3788{ 3789 struct mem_cgroup *iter; 3790 long val = 0; 3791 3792 /* Per-cpu values can be negative, use a signed accumulator */ 3793 for_each_mem_cgroup_tree(iter, memcg) 3794 val += mem_cgroup_read_stat(iter, idx); 3795 3796 if (val < 0) /* race ? */ 3797 val = 0; 3798 return val; 3799} 3800 3801static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) 3802{ 3803 u64 val; 3804 3805 if (!mem_cgroup_is_root(memcg)) { 3806 if (!swap) 3807 return res_counter_read_u64(&memcg->res, RES_USAGE); 3808 else 3809 return res_counter_read_u64(&memcg->memsw, RES_USAGE); 3810 } 3811 3812 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); 3813 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); 3814 3815 if (swap) 3816 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 3817 3818 return val << PAGE_SHIFT; 3819} 3820 3821static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3822{ 3823 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3824 u64 val; 3825 int type, name; 3826 3827 type = MEMFILE_TYPE(cft->private); 3828 name = MEMFILE_ATTR(cft->private); 3829 switch (type) { 3830 case _MEM: 3831 if (name == RES_USAGE) 3832 val = mem_cgroup_usage(memcg, false); 3833 else 3834 val = res_counter_read_u64(&memcg->res, name); 3835 break; 3836 case _MEMSWAP: 3837 if (name == RES_USAGE) 3838 val = mem_cgroup_usage(memcg, true); 3839 else 3840 val = res_counter_read_u64(&memcg->memsw, name); 3841 break; 3842 default: 3843 BUG(); 3844 break; 3845 } 3846 return val; 3847} 3848/* 3849 * The user of this function is... 3850 * RES_LIMIT. 3851 */ 3852static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3853 const char *buffer) 3854{ 3855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3856 int type, name; 3857 unsigned long long val; 3858 int ret; 3859 3860 type = MEMFILE_TYPE(cft->private); 3861 name = MEMFILE_ATTR(cft->private); 3862 switch (name) { 3863 case RES_LIMIT: 3864 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3865 ret = -EINVAL; 3866 break; 3867 } 3868 /* This function does all necessary parse...reuse it */ 3869 ret = res_counter_memparse_write_strategy(buffer, &val); 3870 if (ret) 3871 break; 3872 if (type == _MEM) 3873 ret = mem_cgroup_resize_limit(memcg, val); 3874 else 3875 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3876 break; 3877 case RES_SOFT_LIMIT: 3878 ret = res_counter_memparse_write_strategy(buffer, &val); 3879 if (ret) 3880 break; 3881 /* 3882 * For memsw, soft limits are hard to implement in terms 3883 * of semantics, for now, we support soft limits for 3884 * control without swap 3885 */ 3886 if (type == _MEM) 3887 ret = res_counter_set_soft_limit(&memcg->res, val); 3888 else 3889 ret = -EINVAL; 3890 break; 3891 default: 3892 ret = -EINVAL; /* should be BUG() ? */ 3893 break; 3894 } 3895 return ret; 3896} 3897 3898static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 3899 unsigned long long *mem_limit, unsigned long long *memsw_limit) 3900{ 3901 struct cgroup *cgroup; 3902 unsigned long long min_limit, min_memsw_limit, tmp; 3903 3904 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3905 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3906 cgroup = memcg->css.cgroup; 3907 if (!memcg->use_hierarchy) 3908 goto out; 3909 3910 while (cgroup->parent) { 3911 cgroup = cgroup->parent; 3912 memcg = mem_cgroup_from_cont(cgroup); 3913 if (!memcg->use_hierarchy) 3914 break; 3915 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 3916 min_limit = min(min_limit, tmp); 3917 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3918 min_memsw_limit = min(min_memsw_limit, tmp); 3919 } 3920out: 3921 *mem_limit = min_limit; 3922 *memsw_limit = min_memsw_limit; 3923 return; 3924} 3925 3926static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 3927{ 3928 struct mem_cgroup *memcg; 3929 int type, name; 3930 3931 memcg = mem_cgroup_from_cont(cont); 3932 type = MEMFILE_TYPE(event); 3933 name = MEMFILE_ATTR(event); 3934 switch (name) { 3935 case RES_MAX_USAGE: 3936 if (type == _MEM) 3937 res_counter_reset_max(&memcg->res); 3938 else 3939 res_counter_reset_max(&memcg->memsw); 3940 break; 3941 case RES_FAILCNT: 3942 if (type == _MEM) 3943 res_counter_reset_failcnt(&memcg->res); 3944 else 3945 res_counter_reset_failcnt(&memcg->memsw); 3946 break; 3947 } 3948 3949 return 0; 3950} 3951 3952static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 3953 struct cftype *cft) 3954{ 3955 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 3956} 3957 3958#ifdef CONFIG_MMU 3959static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3960 struct cftype *cft, u64 val) 3961{ 3962 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 3963 3964 if (val >= (1 << NR_MOVE_TYPE)) 3965 return -EINVAL; 3966 /* 3967 * We check this value several times in both in can_attach() and 3968 * attach(), so we need cgroup lock to prevent this value from being 3969 * inconsistent. 3970 */ 3971 cgroup_lock(); 3972 memcg->move_charge_at_immigrate = val; 3973 cgroup_unlock(); 3974 3975 return 0; 3976} 3977#else 3978static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3979 struct cftype *cft, u64 val) 3980{ 3981 return -ENOSYS; 3982} 3983#endif 3984 3985 3986/* For read statistics */ 3987enum { 3988 MCS_CACHE, 3989 MCS_RSS, 3990 MCS_FILE_MAPPED, 3991 MCS_PGPGIN, 3992 MCS_PGPGOUT, 3993 MCS_SWAP, 3994 MCS_PGFAULT, 3995 MCS_PGMAJFAULT, 3996 MCS_INACTIVE_ANON, 3997 MCS_ACTIVE_ANON, 3998 MCS_INACTIVE_FILE, 3999 MCS_ACTIVE_FILE, 4000 MCS_UNEVICTABLE, 4001 NR_MCS_STAT, 4002}; 4003 4004struct mcs_total_stat { 4005 s64 stat[NR_MCS_STAT]; 4006}; 4007 4008struct { 4009 char *local_name; 4010 char *total_name; 4011} memcg_stat_strings[NR_MCS_STAT] = { 4012 {"cache", "total_cache"}, 4013 {"rss", "total_rss"}, 4014 {"mapped_file", "total_mapped_file"}, 4015 {"pgpgin", "total_pgpgin"}, 4016 {"pgpgout", "total_pgpgout"}, 4017 {"swap", "total_swap"}, 4018 {"pgfault", "total_pgfault"}, 4019 {"pgmajfault", "total_pgmajfault"}, 4020 {"inactive_anon", "total_inactive_anon"}, 4021 {"active_anon", "total_active_anon"}, 4022 {"inactive_file", "total_inactive_file"}, 4023 {"active_file", "total_active_file"}, 4024 {"unevictable", "total_unevictable"} 4025}; 4026 4027 4028static void 4029mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4030{ 4031 s64 val; 4032 4033 /* per cpu stat */ 4034 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); 4035 s->stat[MCS_CACHE] += val * PAGE_SIZE; 4036 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); 4037 s->stat[MCS_RSS] += val * PAGE_SIZE; 4038 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); 4039 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 4040 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); 4041 s->stat[MCS_PGPGIN] += val; 4042 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); 4043 s->stat[MCS_PGPGOUT] += val; 4044 if (do_swap_account) { 4045 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 4046 s->stat[MCS_SWAP] += val * PAGE_SIZE; 4047 } 4048 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); 4049 s->stat[MCS_PGFAULT] += val; 4050 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); 4051 s->stat[MCS_PGMAJFAULT] += val; 4052 4053 /* per zone stat */ 4054 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); 4055 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 4056 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); 4057 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 4058 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); 4059 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 4060 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); 4061 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 4062 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); 4063 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 4064} 4065 4066static void 4067mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4068{ 4069 struct mem_cgroup *iter; 4070 4071 for_each_mem_cgroup_tree(iter, memcg) 4072 mem_cgroup_get_local_stat(iter, s); 4073} 4074 4075#ifdef CONFIG_NUMA 4076static int mem_control_numa_stat_show(struct seq_file *m, void *arg) 4077{ 4078 int nid; 4079 unsigned long total_nr, file_nr, anon_nr, unevictable_nr; 4080 unsigned long node_nr; 4081 struct cgroup *cont = m->private; 4082 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4083 4084 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL); 4085 seq_printf(m, "total=%lu", total_nr); 4086 for_each_node_state(nid, N_HIGH_MEMORY) { 4087 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL); 4088 seq_printf(m, " N%d=%lu", nid, node_nr); 4089 } 4090 seq_putc(m, '\n'); 4091 4092 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE); 4093 seq_printf(m, "file=%lu", file_nr); 4094 for_each_node_state(nid, N_HIGH_MEMORY) { 4095 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4096 LRU_ALL_FILE); 4097 seq_printf(m, " N%d=%lu", nid, node_nr); 4098 } 4099 seq_putc(m, '\n'); 4100 4101 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON); 4102 seq_printf(m, "anon=%lu", anon_nr); 4103 for_each_node_state(nid, N_HIGH_MEMORY) { 4104 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4105 LRU_ALL_ANON); 4106 seq_printf(m, " N%d=%lu", nid, node_nr); 4107 } 4108 seq_putc(m, '\n'); 4109 4110 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE)); 4111 seq_printf(m, "unevictable=%lu", unevictable_nr); 4112 for_each_node_state(nid, N_HIGH_MEMORY) { 4113 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4114 BIT(LRU_UNEVICTABLE)); 4115 seq_printf(m, " N%d=%lu", nid, node_nr); 4116 } 4117 seq_putc(m, '\n'); 4118 return 0; 4119} 4120#endif /* CONFIG_NUMA */ 4121 4122static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 4123 struct cgroup_map_cb *cb) 4124{ 4125 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4126 struct mcs_total_stat mystat; 4127 int i; 4128 4129 memset(&mystat, 0, sizeof(mystat)); 4130 mem_cgroup_get_local_stat(mem_cont, &mystat); 4131 4132 4133 for (i = 0; i < NR_MCS_STAT; i++) { 4134 if (i == MCS_SWAP && !do_swap_account) 4135 continue; 4136 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 4137 } 4138 4139 /* Hierarchical information */ 4140 { 4141 unsigned long long limit, memsw_limit; 4142 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 4143 cb->fill(cb, "hierarchical_memory_limit", limit); 4144 if (do_swap_account) 4145 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 4146 } 4147 4148 memset(&mystat, 0, sizeof(mystat)); 4149 mem_cgroup_get_total_stat(mem_cont, &mystat); 4150 for (i = 0; i < NR_MCS_STAT; i++) { 4151 if (i == MCS_SWAP && !do_swap_account) 4152 continue; 4153 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 4154 } 4155 4156#ifdef CONFIG_DEBUG_VM 4157 { 4158 int nid, zid; 4159 struct mem_cgroup_per_zone *mz; 4160 unsigned long recent_rotated[2] = {0, 0}; 4161 unsigned long recent_scanned[2] = {0, 0}; 4162 4163 for_each_online_node(nid) 4164 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 4165 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 4166 4167 recent_rotated[0] += 4168 mz->reclaim_stat.recent_rotated[0]; 4169 recent_rotated[1] += 4170 mz->reclaim_stat.recent_rotated[1]; 4171 recent_scanned[0] += 4172 mz->reclaim_stat.recent_scanned[0]; 4173 recent_scanned[1] += 4174 mz->reclaim_stat.recent_scanned[1]; 4175 } 4176 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 4177 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 4178 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 4179 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 4180 } 4181#endif 4182 4183 return 0; 4184} 4185 4186static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 4187{ 4188 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4189 4190 return mem_cgroup_swappiness(memcg); 4191} 4192 4193static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 4194 u64 val) 4195{ 4196 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4197 struct mem_cgroup *parent; 4198 4199 if (val > 100) 4200 return -EINVAL; 4201 4202 if (cgrp->parent == NULL) 4203 return -EINVAL; 4204 4205 parent = mem_cgroup_from_cont(cgrp->parent); 4206 4207 cgroup_lock(); 4208 4209 /* If under hierarchy, only empty-root can set this value */ 4210 if ((parent->use_hierarchy) || 4211 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4212 cgroup_unlock(); 4213 return -EINVAL; 4214 } 4215 4216 memcg->swappiness = val; 4217 4218 cgroup_unlock(); 4219 4220 return 0; 4221} 4222 4223static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 4224{ 4225 struct mem_cgroup_threshold_ary *t; 4226 u64 usage; 4227 int i; 4228 4229 rcu_read_lock(); 4230 if (!swap) 4231 t = rcu_dereference(memcg->thresholds.primary); 4232 else 4233 t = rcu_dereference(memcg->memsw_thresholds.primary); 4234 4235 if (!t) 4236 goto unlock; 4237 4238 usage = mem_cgroup_usage(memcg, swap); 4239 4240 /* 4241 * current_threshold points to threshold just below usage. 4242 * If it's not true, a threshold was crossed after last 4243 * call of __mem_cgroup_threshold(). 4244 */ 4245 i = t->current_threshold; 4246 4247 /* 4248 * Iterate backward over array of thresholds starting from 4249 * current_threshold and check if a threshold is crossed. 4250 * If none of thresholds below usage is crossed, we read 4251 * only one element of the array here. 4252 */ 4253 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 4254 eventfd_signal(t->entries[i].eventfd, 1); 4255 4256 /* i = current_threshold + 1 */ 4257 i++; 4258 4259 /* 4260 * Iterate forward over array of thresholds starting from 4261 * current_threshold+1 and check if a threshold is crossed. 4262 * If none of thresholds above usage is crossed, we read 4263 * only one element of the array here. 4264 */ 4265 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 4266 eventfd_signal(t->entries[i].eventfd, 1); 4267 4268 /* Update current_threshold */ 4269 t->current_threshold = i - 1; 4270unlock: 4271 rcu_read_unlock(); 4272} 4273 4274static void mem_cgroup_threshold(struct mem_cgroup *memcg) 4275{ 4276 while (memcg) { 4277 __mem_cgroup_threshold(memcg, false); 4278 if (do_swap_account) 4279 __mem_cgroup_threshold(memcg, true); 4280 4281 memcg = parent_mem_cgroup(memcg); 4282 } 4283} 4284 4285static int compare_thresholds(const void *a, const void *b) 4286{ 4287 const struct mem_cgroup_threshold *_a = a; 4288 const struct mem_cgroup_threshold *_b = b; 4289 4290 return _a->threshold - _b->threshold; 4291} 4292 4293static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) 4294{ 4295 struct mem_cgroup_eventfd_list *ev; 4296 4297 list_for_each_entry(ev, &memcg->oom_notify, list) 4298 eventfd_signal(ev->eventfd, 1); 4299 return 0; 4300} 4301 4302static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) 4303{ 4304 struct mem_cgroup *iter; 4305 4306 for_each_mem_cgroup_tree(iter, memcg) 4307 mem_cgroup_oom_notify_cb(iter); 4308} 4309 4310static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 4311 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4312{ 4313 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4314 struct mem_cgroup_thresholds *thresholds; 4315 struct mem_cgroup_threshold_ary *new; 4316 int type = MEMFILE_TYPE(cft->private); 4317 u64 threshold, usage; 4318 int i, size, ret; 4319 4320 ret = res_counter_memparse_write_strategy(args, &threshold); 4321 if (ret) 4322 return ret; 4323 4324 mutex_lock(&memcg->thresholds_lock); 4325 4326 if (type == _MEM) 4327 thresholds = &memcg->thresholds; 4328 else if (type == _MEMSWAP) 4329 thresholds = &memcg->memsw_thresholds; 4330 else 4331 BUG(); 4332 4333 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4334 4335 /* Check if a threshold crossed before adding a new one */ 4336 if (thresholds->primary) 4337 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4338 4339 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 4340 4341 /* Allocate memory for new array of thresholds */ 4342 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 4343 GFP_KERNEL); 4344 if (!new) { 4345 ret = -ENOMEM; 4346 goto unlock; 4347 } 4348 new->size = size; 4349 4350 /* Copy thresholds (if any) to new array */ 4351 if (thresholds->primary) { 4352 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 4353 sizeof(struct mem_cgroup_threshold)); 4354 } 4355 4356 /* Add new threshold */ 4357 new->entries[size - 1].eventfd = eventfd; 4358 new->entries[size - 1].threshold = threshold; 4359 4360 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4361 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4362 compare_thresholds, NULL); 4363 4364 /* Find current threshold */ 4365 new->current_threshold = -1; 4366 for (i = 0; i < size; i++) { 4367 if (new->entries[i].threshold < usage) { 4368 /* 4369 * new->current_threshold will not be used until 4370 * rcu_assign_pointer(), so it's safe to increment 4371 * it here. 4372 */ 4373 ++new->current_threshold; 4374 } 4375 } 4376 4377 /* Free old spare buffer and save old primary buffer as spare */ 4378 kfree(thresholds->spare); 4379 thresholds->spare = thresholds->primary; 4380 4381 rcu_assign_pointer(thresholds->primary, new); 4382 4383 /* To be sure that nobody uses thresholds */ 4384 synchronize_rcu(); 4385 4386unlock: 4387 mutex_unlock(&memcg->thresholds_lock); 4388 4389 return ret; 4390} 4391 4392static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4393 struct cftype *cft, struct eventfd_ctx *eventfd) 4394{ 4395 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4396 struct mem_cgroup_thresholds *thresholds; 4397 struct mem_cgroup_threshold_ary *new; 4398 int type = MEMFILE_TYPE(cft->private); 4399 u64 usage; 4400 int i, j, size; 4401 4402 mutex_lock(&memcg->thresholds_lock); 4403 if (type == _MEM) 4404 thresholds = &memcg->thresholds; 4405 else if (type == _MEMSWAP) 4406 thresholds = &memcg->memsw_thresholds; 4407 else 4408 BUG(); 4409 4410 /* 4411 * Something went wrong if we trying to unregister a threshold 4412 * if we don't have thresholds 4413 */ 4414 BUG_ON(!thresholds); 4415 4416 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4417 4418 /* Check if a threshold crossed before removing */ 4419 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4420 4421 /* Calculate new number of threshold */ 4422 size = 0; 4423 for (i = 0; i < thresholds->primary->size; i++) { 4424 if (thresholds->primary->entries[i].eventfd != eventfd) 4425 size++; 4426 } 4427 4428 new = thresholds->spare; 4429 4430 /* Set thresholds array to NULL if we don't have thresholds */ 4431 if (!size) { 4432 kfree(new); 4433 new = NULL; 4434 goto swap_buffers; 4435 } 4436 4437 new->size = size; 4438 4439 /* Copy thresholds and find current threshold */ 4440 new->current_threshold = -1; 4441 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4442 if (thresholds->primary->entries[i].eventfd == eventfd) 4443 continue; 4444 4445 new->entries[j] = thresholds->primary->entries[i]; 4446 if (new->entries[j].threshold < usage) { 4447 /* 4448 * new->current_threshold will not be used 4449 * until rcu_assign_pointer(), so it's safe to increment 4450 * it here. 4451 */ 4452 ++new->current_threshold; 4453 } 4454 j++; 4455 } 4456 4457swap_buffers: 4458 /* Swap primary and spare array */ 4459 thresholds->spare = thresholds->primary; 4460 rcu_assign_pointer(thresholds->primary, new); 4461 4462 /* To be sure that nobody uses thresholds */ 4463 synchronize_rcu(); 4464 4465 mutex_unlock(&memcg->thresholds_lock); 4466} 4467 4468static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4469 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4470{ 4471 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4472 struct mem_cgroup_eventfd_list *event; 4473 int type = MEMFILE_TYPE(cft->private); 4474 4475 BUG_ON(type != _OOM_TYPE); 4476 event = kmalloc(sizeof(*event), GFP_KERNEL); 4477 if (!event) 4478 return -ENOMEM; 4479 4480 spin_lock(&memcg_oom_lock); 4481 4482 event->eventfd = eventfd; 4483 list_add(&event->list, &memcg->oom_notify); 4484 4485 /* already in OOM ? */ 4486 if (atomic_read(&memcg->under_oom)) 4487 eventfd_signal(eventfd, 1); 4488 spin_unlock(&memcg_oom_lock); 4489 4490 return 0; 4491} 4492 4493static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4494 struct cftype *cft, struct eventfd_ctx *eventfd) 4495{ 4496 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4497 struct mem_cgroup_eventfd_list *ev, *tmp; 4498 int type = MEMFILE_TYPE(cft->private); 4499 4500 BUG_ON(type != _OOM_TYPE); 4501 4502 spin_lock(&memcg_oom_lock); 4503 4504 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { 4505 if (ev->eventfd == eventfd) { 4506 list_del(&ev->list); 4507 kfree(ev); 4508 } 4509 } 4510 4511 spin_unlock(&memcg_oom_lock); 4512} 4513 4514static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4515 struct cftype *cft, struct cgroup_map_cb *cb) 4516{ 4517 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4518 4519 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); 4520 4521 if (atomic_read(&memcg->under_oom)) 4522 cb->fill(cb, "under_oom", 1); 4523 else 4524 cb->fill(cb, "under_oom", 0); 4525 return 0; 4526} 4527 4528static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4529 struct cftype *cft, u64 val) 4530{ 4531 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4532 struct mem_cgroup *parent; 4533 4534 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4535 if (!cgrp->parent || !((val == 0) || (val == 1))) 4536 return -EINVAL; 4537 4538 parent = mem_cgroup_from_cont(cgrp->parent); 4539 4540 cgroup_lock(); 4541 /* oom-kill-disable is a flag for subhierarchy. */ 4542 if ((parent->use_hierarchy) || 4543 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4544 cgroup_unlock(); 4545 return -EINVAL; 4546 } 4547 memcg->oom_kill_disable = val; 4548 if (!val) 4549 memcg_oom_recover(memcg); 4550 cgroup_unlock(); 4551 return 0; 4552} 4553 4554#ifdef CONFIG_NUMA 4555static const struct file_operations mem_control_numa_stat_file_operations = { 4556 .read = seq_read, 4557 .llseek = seq_lseek, 4558 .release = single_release, 4559}; 4560 4561static int mem_control_numa_stat_open(struct inode *unused, struct file *file) 4562{ 4563 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; 4564 4565 file->f_op = &mem_control_numa_stat_file_operations; 4566 return single_open(file, mem_control_numa_stat_show, cont); 4567} 4568#endif /* CONFIG_NUMA */ 4569 4570#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 4571static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4572{ 4573 /* 4574 * Part of this would be better living in a separate allocation 4575 * function, leaving us with just the cgroup tree population work. 4576 * We, however, depend on state such as network's proto_list that 4577 * is only initialized after cgroup creation. I found the less 4578 * cumbersome way to deal with it to defer it all to populate time 4579 */ 4580 return mem_cgroup_sockets_init(cont, ss); 4581}; 4582 4583static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4584 struct cgroup *cont) 4585{ 4586 mem_cgroup_sockets_destroy(cont, ss); 4587} 4588#else 4589static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4590{ 4591 return 0; 4592} 4593 4594static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4595 struct cgroup *cont) 4596{ 4597} 4598#endif 4599 4600static struct cftype mem_cgroup_files[] = { 4601 { 4602 .name = "usage_in_bytes", 4603 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4604 .read_u64 = mem_cgroup_read, 4605 .register_event = mem_cgroup_usage_register_event, 4606 .unregister_event = mem_cgroup_usage_unregister_event, 4607 }, 4608 { 4609 .name = "max_usage_in_bytes", 4610 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4611 .trigger = mem_cgroup_reset, 4612 .read_u64 = mem_cgroup_read, 4613 }, 4614 { 4615 .name = "limit_in_bytes", 4616 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4617 .write_string = mem_cgroup_write, 4618 .read_u64 = mem_cgroup_read, 4619 }, 4620 { 4621 .name = "soft_limit_in_bytes", 4622 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4623 .write_string = mem_cgroup_write, 4624 .read_u64 = mem_cgroup_read, 4625 }, 4626 { 4627 .name = "failcnt", 4628 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4629 .trigger = mem_cgroup_reset, 4630 .read_u64 = mem_cgroup_read, 4631 }, 4632 { 4633 .name = "stat", 4634 .read_map = mem_control_stat_show, 4635 }, 4636 { 4637 .name = "force_empty", 4638 .trigger = mem_cgroup_force_empty_write, 4639 }, 4640 { 4641 .name = "use_hierarchy", 4642 .write_u64 = mem_cgroup_hierarchy_write, 4643 .read_u64 = mem_cgroup_hierarchy_read, 4644 }, 4645 { 4646 .name = "swappiness", 4647 .read_u64 = mem_cgroup_swappiness_read, 4648 .write_u64 = mem_cgroup_swappiness_write, 4649 }, 4650 { 4651 .name = "move_charge_at_immigrate", 4652 .read_u64 = mem_cgroup_move_charge_read, 4653 .write_u64 = mem_cgroup_move_charge_write, 4654 }, 4655 { 4656 .name = "oom_control", 4657 .read_map = mem_cgroup_oom_control_read, 4658 .write_u64 = mem_cgroup_oom_control_write, 4659 .register_event = mem_cgroup_oom_register_event, 4660 .unregister_event = mem_cgroup_oom_unregister_event, 4661 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4662 }, 4663#ifdef CONFIG_NUMA 4664 { 4665 .name = "numa_stat", 4666 .open = mem_control_numa_stat_open, 4667 .mode = S_IRUGO, 4668 }, 4669#endif 4670}; 4671 4672#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4673static struct cftype memsw_cgroup_files[] = { 4674 { 4675 .name = "memsw.usage_in_bytes", 4676 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4677 .read_u64 = mem_cgroup_read, 4678 .register_event = mem_cgroup_usage_register_event, 4679 .unregister_event = mem_cgroup_usage_unregister_event, 4680 }, 4681 { 4682 .name = "memsw.max_usage_in_bytes", 4683 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4684 .trigger = mem_cgroup_reset, 4685 .read_u64 = mem_cgroup_read, 4686 }, 4687 { 4688 .name = "memsw.limit_in_bytes", 4689 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4690 .write_string = mem_cgroup_write, 4691 .read_u64 = mem_cgroup_read, 4692 }, 4693 { 4694 .name = "memsw.failcnt", 4695 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4696 .trigger = mem_cgroup_reset, 4697 .read_u64 = mem_cgroup_read, 4698 }, 4699}; 4700 4701static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4702{ 4703 if (!do_swap_account) 4704 return 0; 4705 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4706 ARRAY_SIZE(memsw_cgroup_files)); 4707}; 4708#else 4709static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4710{ 4711 return 0; 4712} 4713#endif 4714 4715static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4716{ 4717 struct mem_cgroup_per_node *pn; 4718 struct mem_cgroup_per_zone *mz; 4719 enum lru_list l; 4720 int zone, tmp = node; 4721 /* 4722 * This routine is called against possible nodes. 4723 * But it's BUG to call kmalloc() against offline node. 4724 * 4725 * TODO: this routine can waste much memory for nodes which will 4726 * never be onlined. It's better to use memory hotplug callback 4727 * function. 4728 */ 4729 if (!node_state(node, N_NORMAL_MEMORY)) 4730 tmp = -1; 4731 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4732 if (!pn) 4733 return 1; 4734 4735 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4736 mz = &pn->zoneinfo[zone]; 4737 for_each_lru(l) 4738 INIT_LIST_HEAD(&mz->lruvec.lists[l]); 4739 mz->usage_in_excess = 0; 4740 mz->on_tree = false; 4741 mz->mem = memcg; 4742 } 4743 memcg->info.nodeinfo[node] = pn; 4744 return 0; 4745} 4746 4747static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4748{ 4749 kfree(memcg->info.nodeinfo[node]); 4750} 4751 4752static struct mem_cgroup *mem_cgroup_alloc(void) 4753{ 4754 struct mem_cgroup *mem; 4755 int size = sizeof(struct mem_cgroup); 4756 4757 /* Can be very big if MAX_NUMNODES is very big */ 4758 if (size < PAGE_SIZE) 4759 mem = kzalloc(size, GFP_KERNEL); 4760 else 4761 mem = vzalloc(size); 4762 4763 if (!mem) 4764 return NULL; 4765 4766 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4767 if (!mem->stat) 4768 goto out_free; 4769 spin_lock_init(&mem->pcp_counter_lock); 4770 return mem; 4771 4772out_free: 4773 if (size < PAGE_SIZE) 4774 kfree(mem); 4775 else 4776 vfree(mem); 4777 return NULL; 4778} 4779 4780/* 4781 * At destroying mem_cgroup, references from swap_cgroup can remain. 4782 * (scanning all at force_empty is too costly...) 4783 * 4784 * Instead of clearing all references at force_empty, we remember 4785 * the number of reference from swap_cgroup and free mem_cgroup when 4786 * it goes down to 0. 4787 * 4788 * Removal of cgroup itself succeeds regardless of refs from swap. 4789 */ 4790 4791static void __mem_cgroup_free(struct mem_cgroup *memcg) 4792{ 4793 int node; 4794 4795 mem_cgroup_remove_from_trees(memcg); 4796 free_css_id(&mem_cgroup_subsys, &memcg->css); 4797 4798 for_each_node(node) 4799 free_mem_cgroup_per_zone_info(memcg, node); 4800 4801 free_percpu(memcg->stat); 4802 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4803 kfree(memcg); 4804 else 4805 vfree(memcg); 4806} 4807 4808static void mem_cgroup_get(struct mem_cgroup *memcg) 4809{ 4810 atomic_inc(&memcg->refcnt); 4811} 4812 4813static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) 4814{ 4815 if (atomic_sub_and_test(count, &memcg->refcnt)) { 4816 struct mem_cgroup *parent = parent_mem_cgroup(memcg); 4817 __mem_cgroup_free(memcg); 4818 if (parent) 4819 mem_cgroup_put(parent); 4820 } 4821} 4822 4823static void mem_cgroup_put(struct mem_cgroup *memcg) 4824{ 4825 __mem_cgroup_put(memcg, 1); 4826} 4827 4828/* 4829 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4830 */ 4831struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) 4832{ 4833 if (!memcg->res.parent) 4834 return NULL; 4835 return mem_cgroup_from_res_counter(memcg->res.parent, res); 4836} 4837EXPORT_SYMBOL(parent_mem_cgroup); 4838 4839#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4840static void __init enable_swap_cgroup(void) 4841{ 4842 if (!mem_cgroup_disabled() && really_do_swap_account) 4843 do_swap_account = 1; 4844} 4845#else 4846static void __init enable_swap_cgroup(void) 4847{ 4848} 4849#endif 4850 4851static int mem_cgroup_soft_limit_tree_init(void) 4852{ 4853 struct mem_cgroup_tree_per_node *rtpn; 4854 struct mem_cgroup_tree_per_zone *rtpz; 4855 int tmp, node, zone; 4856 4857 for_each_node(node) { 4858 tmp = node; 4859 if (!node_state(node, N_NORMAL_MEMORY)) 4860 tmp = -1; 4861 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4862 if (!rtpn) 4863 goto err_cleanup; 4864 4865 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4866 4867 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4868 rtpz = &rtpn->rb_tree_per_zone[zone]; 4869 rtpz->rb_root = RB_ROOT; 4870 spin_lock_init(&rtpz->lock); 4871 } 4872 } 4873 return 0; 4874 4875err_cleanup: 4876 for_each_node(node) { 4877 if (!soft_limit_tree.rb_tree_per_node[node]) 4878 break; 4879 kfree(soft_limit_tree.rb_tree_per_node[node]); 4880 soft_limit_tree.rb_tree_per_node[node] = NULL; 4881 } 4882 return 1; 4883 4884} 4885 4886static struct cgroup_subsys_state * __ref 4887mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4888{ 4889 struct mem_cgroup *memcg, *parent; 4890 long error = -ENOMEM; 4891 int node; 4892 4893 memcg = mem_cgroup_alloc(); 4894 if (!memcg) 4895 return ERR_PTR(error); 4896 4897 for_each_node(node) 4898 if (alloc_mem_cgroup_per_zone_info(memcg, node)) 4899 goto free_out; 4900 4901 /* root ? */ 4902 if (cont->parent == NULL) { 4903 int cpu; 4904 enable_swap_cgroup(); 4905 parent = NULL; 4906 if (mem_cgroup_soft_limit_tree_init()) 4907 goto free_out; 4908 root_mem_cgroup = memcg; 4909 for_each_possible_cpu(cpu) { 4910 struct memcg_stock_pcp *stock = 4911 &per_cpu(memcg_stock, cpu); 4912 INIT_WORK(&stock->work, drain_local_stock); 4913 } 4914 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 4915 } else { 4916 parent = mem_cgroup_from_cont(cont->parent); 4917 memcg->use_hierarchy = parent->use_hierarchy; 4918 memcg->oom_kill_disable = parent->oom_kill_disable; 4919 } 4920 4921 if (parent && parent->use_hierarchy) { 4922 res_counter_init(&memcg->res, &parent->res); 4923 res_counter_init(&memcg->memsw, &parent->memsw); 4924 /* 4925 * We increment refcnt of the parent to ensure that we can 4926 * safely access it on res_counter_charge/uncharge. 4927 * This refcnt will be decremented when freeing this 4928 * mem_cgroup(see mem_cgroup_put). 4929 */ 4930 mem_cgroup_get(parent); 4931 } else { 4932 res_counter_init(&memcg->res, NULL); 4933 res_counter_init(&memcg->memsw, NULL); 4934 } 4935 memcg->last_scanned_node = MAX_NUMNODES; 4936 INIT_LIST_HEAD(&memcg->oom_notify); 4937 4938 if (parent) 4939 memcg->swappiness = mem_cgroup_swappiness(parent); 4940 atomic_set(&memcg->refcnt, 1); 4941 memcg->move_charge_at_immigrate = 0; 4942 mutex_init(&memcg->thresholds_lock); 4943 return &memcg->css; 4944free_out: 4945 __mem_cgroup_free(memcg); 4946 return ERR_PTR(error); 4947} 4948 4949static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 4950 struct cgroup *cont) 4951{ 4952 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 4953 4954 return mem_cgroup_force_empty(memcg, false); 4955} 4956 4957static void mem_cgroup_destroy(struct cgroup_subsys *ss, 4958 struct cgroup *cont) 4959{ 4960 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 4961 4962 kmem_cgroup_destroy(ss, cont); 4963 4964 mem_cgroup_put(memcg); 4965} 4966 4967static int mem_cgroup_populate(struct cgroup_subsys *ss, 4968 struct cgroup *cont) 4969{ 4970 int ret; 4971 4972 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 4973 ARRAY_SIZE(mem_cgroup_files)); 4974 4975 if (!ret) 4976 ret = register_memsw_files(cont, ss); 4977 4978 if (!ret) 4979 ret = register_kmem_files(cont, ss); 4980 4981 return ret; 4982} 4983 4984#ifdef CONFIG_MMU 4985/* Handlers for move charge at task migration. */ 4986#define PRECHARGE_COUNT_AT_ONCE 256 4987static int mem_cgroup_do_precharge(unsigned long count) 4988{ 4989 int ret = 0; 4990 int batch_count = PRECHARGE_COUNT_AT_ONCE; 4991 struct mem_cgroup *memcg = mc.to; 4992 4993 if (mem_cgroup_is_root(memcg)) { 4994 mc.precharge += count; 4995 /* we don't need css_get for root */ 4996 return ret; 4997 } 4998 /* try to charge at once */ 4999 if (count > 1) { 5000 struct res_counter *dummy; 5001 /* 5002 * "memcg" cannot be under rmdir() because we've already checked 5003 * by cgroup_lock_live_cgroup() that it is not removed and we 5004 * are still under the same cgroup_mutex. So we can postpone 5005 * css_get(). 5006 */ 5007 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) 5008 goto one_by_one; 5009 if (do_swap_account && res_counter_charge(&memcg->memsw, 5010 PAGE_SIZE * count, &dummy)) { 5011 res_counter_uncharge(&memcg->res, PAGE_SIZE * count); 5012 goto one_by_one; 5013 } 5014 mc.precharge += count; 5015 return ret; 5016 } 5017one_by_one: 5018 /* fall back to one by one charge */ 5019 while (count--) { 5020 if (signal_pending(current)) { 5021 ret = -EINTR; 5022 break; 5023 } 5024 if (!batch_count--) { 5025 batch_count = PRECHARGE_COUNT_AT_ONCE; 5026 cond_resched(); 5027 } 5028 ret = __mem_cgroup_try_charge(NULL, 5029 GFP_KERNEL, 1, &memcg, false); 5030 if (ret) 5031 /* mem_cgroup_clear_mc() will do uncharge later */ 5032 return ret; 5033 mc.precharge++; 5034 } 5035 return ret; 5036} 5037 5038/** 5039 * is_target_pte_for_mc - check a pte whether it is valid for move charge 5040 * @vma: the vma the pte to be checked belongs 5041 * @addr: the address corresponding to the pte to be checked 5042 * @ptent: the pte to be checked 5043 * @target: the pointer the target page or swap ent will be stored(can be NULL) 5044 * 5045 * Returns 5046 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 5047 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 5048 * move charge. if @target is not NULL, the page is stored in target->page 5049 * with extra refcnt got(Callers should handle it). 5050 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 5051 * target for charge migration. if @target is not NULL, the entry is stored 5052 * in target->ent. 5053 * 5054 * Called with pte lock held. 5055 */ 5056union mc_target { 5057 struct page *page; 5058 swp_entry_t ent; 5059}; 5060 5061enum mc_target_type { 5062 MC_TARGET_NONE, /* not used */ 5063 MC_TARGET_PAGE, 5064 MC_TARGET_SWAP, 5065}; 5066 5067static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 5068 unsigned long addr, pte_t ptent) 5069{ 5070 struct page *page = vm_normal_page(vma, addr, ptent); 5071 5072 if (!page || !page_mapped(page)) 5073 return NULL; 5074 if (PageAnon(page)) { 5075 /* we don't move shared anon */ 5076 if (!move_anon() || page_mapcount(page) > 2) 5077 return NULL; 5078 } else if (!move_file()) 5079 /* we ignore mapcount for file pages */ 5080 return NULL; 5081 if (!get_page_unless_zero(page)) 5082 return NULL; 5083 5084 return page; 5085} 5086 5087static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 5088 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5089{ 5090 int usage_count; 5091 struct page *page = NULL; 5092 swp_entry_t ent = pte_to_swp_entry(ptent); 5093 5094 if (!move_anon() || non_swap_entry(ent)) 5095 return NULL; 5096 usage_count = mem_cgroup_count_swap_user(ent, &page); 5097 if (usage_count > 1) { /* we don't move shared anon */ 5098 if (page) 5099 put_page(page); 5100 return NULL; 5101 } 5102 if (do_swap_account) 5103 entry->val = ent.val; 5104 5105 return page; 5106} 5107 5108static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 5109 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5110{ 5111 struct page *page = NULL; 5112 struct inode *inode; 5113 struct address_space *mapping; 5114 pgoff_t pgoff; 5115 5116 if (!vma->vm_file) /* anonymous vma */ 5117 return NULL; 5118 if (!move_file()) 5119 return NULL; 5120 5121 inode = vma->vm_file->f_path.dentry->d_inode; 5122 mapping = vma->vm_file->f_mapping; 5123 if (pte_none(ptent)) 5124 pgoff = linear_page_index(vma, addr); 5125 else /* pte_file(ptent) is true */ 5126 pgoff = pte_to_pgoff(ptent); 5127 5128 /* page is moved even if it's not RSS of this task(page-faulted). */ 5129 page = find_get_page(mapping, pgoff); 5130 5131#ifdef CONFIG_SWAP 5132 /* shmem/tmpfs may report page out on swap: account for that too. */ 5133 if (radix_tree_exceptional_entry(page)) { 5134 swp_entry_t swap = radix_to_swp_entry(page); 5135 if (do_swap_account) 5136 *entry = swap; 5137 page = find_get_page(&swapper_space, swap.val); 5138 } 5139#endif 5140 return page; 5141} 5142 5143static int is_target_pte_for_mc(struct vm_area_struct *vma, 5144 unsigned long addr, pte_t ptent, union mc_target *target) 5145{ 5146 struct page *page = NULL; 5147 struct page_cgroup *pc; 5148 int ret = 0; 5149 swp_entry_t ent = { .val = 0 }; 5150 5151 if (pte_present(ptent)) 5152 page = mc_handle_present_pte(vma, addr, ptent); 5153 else if (is_swap_pte(ptent)) 5154 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 5155 else if (pte_none(ptent) || pte_file(ptent)) 5156 page = mc_handle_file_pte(vma, addr, ptent, &ent); 5157 5158 if (!page && !ent.val) 5159 return 0; 5160 if (page) { 5161 pc = lookup_page_cgroup(page); 5162 /* 5163 * Do only loose check w/o page_cgroup lock. 5164 * mem_cgroup_move_account() checks the pc is valid or not under 5165 * the lock. 5166 */ 5167 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 5168 ret = MC_TARGET_PAGE; 5169 if (target) 5170 target->page = page; 5171 } 5172 if (!ret || !target) 5173 put_page(page); 5174 } 5175 /* There is a swap entry and a page doesn't exist or isn't charged */ 5176 if (ent.val && !ret && 5177 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) { 5178 ret = MC_TARGET_SWAP; 5179 if (target) 5180 target->ent = ent; 5181 } 5182 return ret; 5183} 5184 5185static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 5186 unsigned long addr, unsigned long end, 5187 struct mm_walk *walk) 5188{ 5189 struct vm_area_struct *vma = walk->private; 5190 pte_t *pte; 5191 spinlock_t *ptl; 5192 5193 split_huge_page_pmd(walk->mm, pmd); 5194 5195 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5196 for (; addr != end; pte++, addr += PAGE_SIZE) 5197 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 5198 mc.precharge++; /* increment precharge temporarily */ 5199 pte_unmap_unlock(pte - 1, ptl); 5200 cond_resched(); 5201 5202 return 0; 5203} 5204 5205static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 5206{ 5207 unsigned long precharge; 5208 struct vm_area_struct *vma; 5209 5210 down_read(&mm->mmap_sem); 5211 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5212 struct mm_walk mem_cgroup_count_precharge_walk = { 5213 .pmd_entry = mem_cgroup_count_precharge_pte_range, 5214 .mm = mm, 5215 .private = vma, 5216 }; 5217 if (is_vm_hugetlb_page(vma)) 5218 continue; 5219 walk_page_range(vma->vm_start, vma->vm_end, 5220 &mem_cgroup_count_precharge_walk); 5221 } 5222 up_read(&mm->mmap_sem); 5223 5224 precharge = mc.precharge; 5225 mc.precharge = 0; 5226 5227 return precharge; 5228} 5229 5230static int mem_cgroup_precharge_mc(struct mm_struct *mm) 5231{ 5232 unsigned long precharge = mem_cgroup_count_precharge(mm); 5233 5234 VM_BUG_ON(mc.moving_task); 5235 mc.moving_task = current; 5236 return mem_cgroup_do_precharge(precharge); 5237} 5238 5239/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 5240static void __mem_cgroup_clear_mc(void) 5241{ 5242 struct mem_cgroup *from = mc.from; 5243 struct mem_cgroup *to = mc.to; 5244 5245 /* we must uncharge all the leftover precharges from mc.to */ 5246 if (mc.precharge) { 5247 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 5248 mc.precharge = 0; 5249 } 5250 /* 5251 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 5252 * we must uncharge here. 5253 */ 5254 if (mc.moved_charge) { 5255 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 5256 mc.moved_charge = 0; 5257 } 5258 /* we must fixup refcnts and charges */ 5259 if (mc.moved_swap) { 5260 /* uncharge swap account from the old cgroup */ 5261 if (!mem_cgroup_is_root(mc.from)) 5262 res_counter_uncharge(&mc.from->memsw, 5263 PAGE_SIZE * mc.moved_swap); 5264 __mem_cgroup_put(mc.from, mc.moved_swap); 5265 5266 if (!mem_cgroup_is_root(mc.to)) { 5267 /* 5268 * we charged both to->res and to->memsw, so we should 5269 * uncharge to->res. 5270 */ 5271 res_counter_uncharge(&mc.to->res, 5272 PAGE_SIZE * mc.moved_swap); 5273 } 5274 /* we've already done mem_cgroup_get(mc.to) */ 5275 mc.moved_swap = 0; 5276 } 5277 memcg_oom_recover(from); 5278 memcg_oom_recover(to); 5279 wake_up_all(&mc.waitq); 5280} 5281 5282static void mem_cgroup_clear_mc(void) 5283{ 5284 struct mem_cgroup *from = mc.from; 5285 5286 /* 5287 * we must clear moving_task before waking up waiters at the end of 5288 * task migration. 5289 */ 5290 mc.moving_task = NULL; 5291 __mem_cgroup_clear_mc(); 5292 spin_lock(&mc.lock); 5293 mc.from = NULL; 5294 mc.to = NULL; 5295 spin_unlock(&mc.lock); 5296 mem_cgroup_end_move(from); 5297} 5298 5299static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5300 struct cgroup *cgroup, 5301 struct cgroup_taskset *tset) 5302{ 5303 struct task_struct *p = cgroup_taskset_first(tset); 5304 int ret = 0; 5305 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); 5306 5307 if (memcg->move_charge_at_immigrate) { 5308 struct mm_struct *mm; 5309 struct mem_cgroup *from = mem_cgroup_from_task(p); 5310 5311 VM_BUG_ON(from == memcg); 5312 5313 mm = get_task_mm(p); 5314 if (!mm) 5315 return 0; 5316 /* We move charges only when we move a owner of the mm */ 5317 if (mm->owner == p) { 5318 VM_BUG_ON(mc.from); 5319 VM_BUG_ON(mc.to); 5320 VM_BUG_ON(mc.precharge); 5321 VM_BUG_ON(mc.moved_charge); 5322 VM_BUG_ON(mc.moved_swap); 5323 mem_cgroup_start_move(from); 5324 spin_lock(&mc.lock); 5325 mc.from = from; 5326 mc.to = memcg; 5327 spin_unlock(&mc.lock); 5328 /* We set mc.moving_task later */ 5329 5330 ret = mem_cgroup_precharge_mc(mm); 5331 if (ret) 5332 mem_cgroup_clear_mc(); 5333 } 5334 mmput(mm); 5335 } 5336 return ret; 5337} 5338 5339static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5340 struct cgroup *cgroup, 5341 struct cgroup_taskset *tset) 5342{ 5343 mem_cgroup_clear_mc(); 5344} 5345 5346static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 5347 unsigned long addr, unsigned long end, 5348 struct mm_walk *walk) 5349{ 5350 int ret = 0; 5351 struct vm_area_struct *vma = walk->private; 5352 pte_t *pte; 5353 spinlock_t *ptl; 5354 5355 split_huge_page_pmd(walk->mm, pmd); 5356retry: 5357 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5358 for (; addr != end; addr += PAGE_SIZE) { 5359 pte_t ptent = *(pte++); 5360 union mc_target target; 5361 int type; 5362 struct page *page; 5363 struct page_cgroup *pc; 5364 swp_entry_t ent; 5365 5366 if (!mc.precharge) 5367 break; 5368 5369 type = is_target_pte_for_mc(vma, addr, ptent, &target); 5370 switch (type) { 5371 case MC_TARGET_PAGE: 5372 page = target.page; 5373 if (isolate_lru_page(page)) 5374 goto put; 5375 pc = lookup_page_cgroup(page); 5376 if (!mem_cgroup_move_account(page, 1, pc, 5377 mc.from, mc.to, false)) { 5378 mc.precharge--; 5379 /* we uncharge from mc.from later. */ 5380 mc.moved_charge++; 5381 } 5382 putback_lru_page(page); 5383put: /* is_target_pte_for_mc() gets the page */ 5384 put_page(page); 5385 break; 5386 case MC_TARGET_SWAP: 5387 ent = target.ent; 5388 if (!mem_cgroup_move_swap_account(ent, 5389 mc.from, mc.to, false)) { 5390 mc.precharge--; 5391 /* we fixup refcnts and charges later. */ 5392 mc.moved_swap++; 5393 } 5394 break; 5395 default: 5396 break; 5397 } 5398 } 5399 pte_unmap_unlock(pte - 1, ptl); 5400 cond_resched(); 5401 5402 if (addr != end) { 5403 /* 5404 * We have consumed all precharges we got in can_attach(). 5405 * We try charge one by one, but don't do any additional 5406 * charges to mc.to if we have failed in charge once in attach() 5407 * phase. 5408 */ 5409 ret = mem_cgroup_do_precharge(1); 5410 if (!ret) 5411 goto retry; 5412 } 5413 5414 return ret; 5415} 5416 5417static void mem_cgroup_move_charge(struct mm_struct *mm) 5418{ 5419 struct vm_area_struct *vma; 5420 5421 lru_add_drain_all(); 5422retry: 5423 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 5424 /* 5425 * Someone who are holding the mmap_sem might be waiting in 5426 * waitq. So we cancel all extra charges, wake up all waiters, 5427 * and retry. Because we cancel precharges, we might not be able 5428 * to move enough charges, but moving charge is a best-effort 5429 * feature anyway, so it wouldn't be a big problem. 5430 */ 5431 __mem_cgroup_clear_mc(); 5432 cond_resched(); 5433 goto retry; 5434 } 5435 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5436 int ret; 5437 struct mm_walk mem_cgroup_move_charge_walk = { 5438 .pmd_entry = mem_cgroup_move_charge_pte_range, 5439 .mm = mm, 5440 .private = vma, 5441 }; 5442 if (is_vm_hugetlb_page(vma)) 5443 continue; 5444 ret = walk_page_range(vma->vm_start, vma->vm_end, 5445 &mem_cgroup_move_charge_walk); 5446 if (ret) 5447 /* 5448 * means we have consumed all precharges and failed in 5449 * doing additional charge. Just abandon here. 5450 */ 5451 break; 5452 } 5453 up_read(&mm->mmap_sem); 5454} 5455 5456static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5457 struct cgroup *cont, 5458 struct cgroup_taskset *tset) 5459{ 5460 struct task_struct *p = cgroup_taskset_first(tset); 5461 struct mm_struct *mm = get_task_mm(p); 5462 5463 if (mm) { 5464 if (mc.to) 5465 mem_cgroup_move_charge(mm); 5466 put_swap_token(mm); 5467 mmput(mm); 5468 } 5469 if (mc.to) 5470 mem_cgroup_clear_mc(); 5471} 5472#else /* !CONFIG_MMU */ 5473static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5474 struct cgroup *cgroup, 5475 struct cgroup_taskset *tset) 5476{ 5477 return 0; 5478} 5479static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5480 struct cgroup *cgroup, 5481 struct cgroup_taskset *tset) 5482{ 5483} 5484static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5485 struct cgroup *cont, 5486 struct cgroup_taskset *tset) 5487{ 5488} 5489#endif 5490 5491struct cgroup_subsys mem_cgroup_subsys = { 5492 .name = "memory", 5493 .subsys_id = mem_cgroup_subsys_id, 5494 .create = mem_cgroup_create, 5495 .pre_destroy = mem_cgroup_pre_destroy, 5496 .destroy = mem_cgroup_destroy, 5497 .populate = mem_cgroup_populate, 5498 .can_attach = mem_cgroup_can_attach, 5499 .cancel_attach = mem_cgroup_cancel_attach, 5500 .attach = mem_cgroup_move_task, 5501 .early_init = 0, 5502 .use_id = 1, 5503}; 5504 5505#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5506static int __init enable_swap_account(char *s) 5507{ 5508 /* consider enabled if no parameter or 1 is given */ 5509 if (!strcmp(s, "1")) 5510 really_do_swap_account = 1; 5511 else if (!strcmp(s, "0")) 5512 really_do_swap_account = 0; 5513 return 1; 5514} 5515__setup("swapaccount=", enable_swap_account); 5516 5517#endif 5518