memcontrol.c revision 0e574a932d2cab8eb3b02d21feb59f2c09154738
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_state(node, N_POSSIBLE) { 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 VM_BUG_ON(PageCgroupAcctLRU(pc)); 1044 /* 1045 * putback: charge: 1046 * SetPageLRU SetPageCgroupUsed 1047 * smp_mb smp_mb 1048 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1049 * 1050 * Ensure that one of the two sides adds the page to the memcg 1051 * LRU during a race. 1052 */ 1053 smp_mb(); 1054 /* 1055 * If the page is uncharged, it may be freed soon, but it 1056 * could also be swap cache (readahead, swapoff) that needs to 1057 * be reclaimable in the future. root_mem_cgroup will babysit 1058 * it for the time being. 1059 */ 1060 if (PageCgroupUsed(pc)) { 1061 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1062 smp_rmb(); 1063 memcg = pc->mem_cgroup; 1064 SetPageCgroupAcctLRU(pc); 1065 } else 1066 memcg = root_mem_cgroup; 1067 mz = page_cgroup_zoneinfo(memcg, page); 1068 /* compound_order() is stabilized through lru_lock */ 1069 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 1070 return &mz->lruvec; 1071} 1072 1073/** 1074 * mem_cgroup_lru_del_list - account for removing an lru page 1075 * @page: the page 1076 * @lru: target lru 1077 * 1078 * This function accounts for @page being removed from @lru. 1079 * 1080 * The callsite is then responsible for physically unlinking 1081 * @page->lru. 1082 */ 1083void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) 1084{ 1085 struct mem_cgroup_per_zone *mz; 1086 struct mem_cgroup *memcg; 1087 struct page_cgroup *pc; 1088 1089 if (mem_cgroup_disabled()) 1090 return; 1091 1092 pc = lookup_page_cgroup(page); 1093 /* 1094 * root_mem_cgroup babysits uncharged LRU pages, but 1095 * PageCgroupUsed is cleared when the page is about to get 1096 * freed. PageCgroupAcctLRU remembers whether the 1097 * LRU-accounting happened against pc->mem_cgroup or 1098 * root_mem_cgroup. 1099 */ 1100 if (TestClearPageCgroupAcctLRU(pc)) { 1101 VM_BUG_ON(!pc->mem_cgroup); 1102 memcg = pc->mem_cgroup; 1103 } else 1104 memcg = root_mem_cgroup; 1105 mz = page_cgroup_zoneinfo(memcg, page); 1106 /* huge page split is done under lru_lock. so, we have no races. */ 1107 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 1108} 1109 1110void mem_cgroup_lru_del(struct page *page) 1111{ 1112 mem_cgroup_lru_del_list(page, page_lru(page)); 1113} 1114 1115/** 1116 * mem_cgroup_lru_move_lists - account for moving a page between lrus 1117 * @zone: zone of the page 1118 * @page: the page 1119 * @from: current lru 1120 * @to: target lru 1121 * 1122 * This function accounts for @page being moved between the lrus @from 1123 * and @to, and returns the lruvec for the given @zone and the memcg 1124 * @page is charged to. 1125 * 1126 * The callsite is then responsible for physically relinking 1127 * @page->lru to the returned lruvec->lists[@to]. 1128 */ 1129struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, 1130 struct page *page, 1131 enum lru_list from, 1132 enum lru_list to) 1133{ 1134 /* XXX: Optimize this, especially for @from == @to */ 1135 mem_cgroup_lru_del_list(page, from); 1136 return mem_cgroup_lru_add_list(zone, page, to); 1137} 1138 1139/* 1140 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed 1141 * while it's linked to lru because the page may be reused after it's fully 1142 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again. 1143 * It's done under lock_page and expected that zone->lru_lock isnever held. 1144 */ 1145static void mem_cgroup_lru_del_before_commit(struct page *page) 1146{ 1147 enum lru_list lru; 1148 unsigned long flags; 1149 struct zone *zone = page_zone(page); 1150 struct page_cgroup *pc = lookup_page_cgroup(page); 1151 1152 /* 1153 * Doing this check without taking ->lru_lock seems wrong but this 1154 * is safe. Because if page_cgroup's USED bit is unset, the page 1155 * will not be added to any memcg's LRU. If page_cgroup's USED bit is 1156 * set, the commit after this will fail, anyway. 1157 * This all charge/uncharge is done under some mutual execustion. 1158 * So, we don't need to taking care of changes in USED bit. 1159 */ 1160 if (likely(!PageLRU(page))) 1161 return; 1162 1163 spin_lock_irqsave(&zone->lru_lock, flags); 1164 lru = page_lru(page); 1165 /* 1166 * The uncharged page could still be registered to the LRU of 1167 * the stale pc->mem_cgroup. 1168 * 1169 * As pc->mem_cgroup is about to get overwritten, the old LRU 1170 * accounting needs to be taken care of. Let root_mem_cgroup 1171 * babysit the page until the new memcg is responsible for it. 1172 * 1173 * The PCG_USED bit is guarded by lock_page() as the page is 1174 * swapcache/pagecache. 1175 */ 1176 if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) { 1177 del_page_from_lru_list(zone, page, lru); 1178 add_page_to_lru_list(zone, page, lru); 1179 } 1180 spin_unlock_irqrestore(&zone->lru_lock, flags); 1181} 1182 1183static void mem_cgroup_lru_add_after_commit(struct page *page) 1184{ 1185 enum lru_list lru; 1186 unsigned long flags; 1187 struct zone *zone = page_zone(page); 1188 struct page_cgroup *pc = lookup_page_cgroup(page); 1189 /* 1190 * putback: charge: 1191 * SetPageLRU SetPageCgroupUsed 1192 * smp_mb smp_mb 1193 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1194 * 1195 * Ensure that one of the two sides adds the page to the memcg 1196 * LRU during a race. 1197 */ 1198 smp_mb(); 1199 /* taking care of that the page is added to LRU while we commit it */ 1200 if (likely(!PageLRU(page))) 1201 return; 1202 spin_lock_irqsave(&zone->lru_lock, flags); 1203 lru = page_lru(page); 1204 /* 1205 * If the page is not on the LRU, someone will soon put it 1206 * there. If it is, and also already accounted for on the 1207 * memcg-side, it must be on the right lruvec as setting 1208 * pc->mem_cgroup and PageCgroupUsed is properly ordered. 1209 * Otherwise, root_mem_cgroup has been babysitting the page 1210 * during the charge. Move it to the new memcg now. 1211 */ 1212 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) { 1213 del_page_from_lru_list(zone, page, lru); 1214 add_page_to_lru_list(zone, page, lru); 1215 } 1216 spin_unlock_irqrestore(&zone->lru_lock, flags); 1217} 1218 1219/* 1220 * Checks whether given mem is same or in the root_mem_cgroup's 1221 * hierarchy subtree 1222 */ 1223static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, 1224 struct mem_cgroup *memcg) 1225{ 1226 if (root_memcg != memcg) { 1227 return (root_memcg->use_hierarchy && 1228 css_is_ancestor(&memcg->css, &root_memcg->css)); 1229 } 1230 1231 return true; 1232} 1233 1234int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) 1235{ 1236 int ret; 1237 struct mem_cgroup *curr = NULL; 1238 struct task_struct *p; 1239 1240 p = find_lock_task_mm(task); 1241 if (!p) 1242 return 0; 1243 curr = try_get_mem_cgroup_from_mm(p->mm); 1244 task_unlock(p); 1245 if (!curr) 1246 return 0; 1247 /* 1248 * We should check use_hierarchy of "memcg" not "curr". Because checking 1249 * use_hierarchy of "curr" here make this function true if hierarchy is 1250 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* 1251 * hierarchy(even if use_hierarchy is disabled in "memcg"). 1252 */ 1253 ret = mem_cgroup_same_or_subtree(memcg, curr); 1254 css_put(&curr->css); 1255 return ret; 1256} 1257 1258int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) 1259{ 1260 unsigned long inactive_ratio; 1261 int nid = zone_to_nid(zone); 1262 int zid = zone_idx(zone); 1263 unsigned long inactive; 1264 unsigned long active; 1265 unsigned long gb; 1266 1267 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1268 BIT(LRU_INACTIVE_ANON)); 1269 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1270 BIT(LRU_ACTIVE_ANON)); 1271 1272 gb = (inactive + active) >> (30 - PAGE_SHIFT); 1273 if (gb) 1274 inactive_ratio = int_sqrt(10 * gb); 1275 else 1276 inactive_ratio = 1; 1277 1278 return inactive * inactive_ratio < active; 1279} 1280 1281int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) 1282{ 1283 unsigned long active; 1284 unsigned long inactive; 1285 int zid = zone_idx(zone); 1286 int nid = zone_to_nid(zone); 1287 1288 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1289 BIT(LRU_INACTIVE_FILE)); 1290 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1291 BIT(LRU_ACTIVE_FILE)); 1292 1293 return (active > inactive); 1294} 1295 1296struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1297 struct zone *zone) 1298{ 1299 int nid = zone_to_nid(zone); 1300 int zid = zone_idx(zone); 1301 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1302 1303 return &mz->reclaim_stat; 1304} 1305 1306struct zone_reclaim_stat * 1307mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1308{ 1309 struct page_cgroup *pc; 1310 struct mem_cgroup_per_zone *mz; 1311 1312 if (mem_cgroup_disabled()) 1313 return NULL; 1314 1315 pc = lookup_page_cgroup(page); 1316 if (!PageCgroupUsed(pc)) 1317 return NULL; 1318 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1319 smp_rmb(); 1320 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1321 return &mz->reclaim_stat; 1322} 1323 1324#define mem_cgroup_from_res_counter(counter, member) \ 1325 container_of(counter, struct mem_cgroup, member) 1326 1327/** 1328 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1329 * @mem: the memory cgroup 1330 * 1331 * Returns the maximum amount of memory @mem can be charged with, in 1332 * pages. 1333 */ 1334static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) 1335{ 1336 unsigned long long margin; 1337 1338 margin = res_counter_margin(&memcg->res); 1339 if (do_swap_account) 1340 margin = min(margin, res_counter_margin(&memcg->memsw)); 1341 return margin >> PAGE_SHIFT; 1342} 1343 1344int mem_cgroup_swappiness(struct mem_cgroup *memcg) 1345{ 1346 struct cgroup *cgrp = memcg->css.cgroup; 1347 1348 /* root ? */ 1349 if (cgrp->parent == NULL) 1350 return vm_swappiness; 1351 1352 return memcg->swappiness; 1353} 1354 1355static void mem_cgroup_start_move(struct mem_cgroup *memcg) 1356{ 1357 int cpu; 1358 1359 get_online_cpus(); 1360 spin_lock(&memcg->pcp_counter_lock); 1361 for_each_online_cpu(cpu) 1362 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1363 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1364 spin_unlock(&memcg->pcp_counter_lock); 1365 put_online_cpus(); 1366 1367 synchronize_rcu(); 1368} 1369 1370static void mem_cgroup_end_move(struct mem_cgroup *memcg) 1371{ 1372 int cpu; 1373 1374 if (!memcg) 1375 return; 1376 get_online_cpus(); 1377 spin_lock(&memcg->pcp_counter_lock); 1378 for_each_online_cpu(cpu) 1379 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1380 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1381 spin_unlock(&memcg->pcp_counter_lock); 1382 put_online_cpus(); 1383} 1384/* 1385 * 2 routines for checking "mem" is under move_account() or not. 1386 * 1387 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1388 * for avoiding race in accounting. If true, 1389 * pc->mem_cgroup may be overwritten. 1390 * 1391 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1392 * under hierarchy of moving cgroups. This is for 1393 * waiting at hith-memory prressure caused by "move". 1394 */ 1395 1396static bool mem_cgroup_stealed(struct mem_cgroup *memcg) 1397{ 1398 VM_BUG_ON(!rcu_read_lock_held()); 1399 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1400} 1401 1402static bool mem_cgroup_under_move(struct mem_cgroup *memcg) 1403{ 1404 struct mem_cgroup *from; 1405 struct mem_cgroup *to; 1406 bool ret = false; 1407 /* 1408 * Unlike task_move routines, we access mc.to, mc.from not under 1409 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1410 */ 1411 spin_lock(&mc.lock); 1412 from = mc.from; 1413 to = mc.to; 1414 if (!from) 1415 goto unlock; 1416 1417 ret = mem_cgroup_same_or_subtree(memcg, from) 1418 || mem_cgroup_same_or_subtree(memcg, to); 1419unlock: 1420 spin_unlock(&mc.lock); 1421 return ret; 1422} 1423 1424static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) 1425{ 1426 if (mc.moving_task && current != mc.moving_task) { 1427 if (mem_cgroup_under_move(memcg)) { 1428 DEFINE_WAIT(wait); 1429 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1430 /* moving charge context might have finished. */ 1431 if (mc.moving_task) 1432 schedule(); 1433 finish_wait(&mc.waitq, &wait); 1434 return true; 1435 } 1436 } 1437 return false; 1438} 1439 1440/** 1441 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1442 * @memcg: The memory cgroup that went over limit 1443 * @p: Task that is going to be killed 1444 * 1445 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1446 * enabled 1447 */ 1448void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1449{ 1450 struct cgroup *task_cgrp; 1451 struct cgroup *mem_cgrp; 1452 /* 1453 * Need a buffer in BSS, can't rely on allocations. The code relies 1454 * on the assumption that OOM is serialized for memory controller. 1455 * If this assumption is broken, revisit this code. 1456 */ 1457 static char memcg_name[PATH_MAX]; 1458 int ret; 1459 1460 if (!memcg || !p) 1461 return; 1462 1463 1464 rcu_read_lock(); 1465 1466 mem_cgrp = memcg->css.cgroup; 1467 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1468 1469 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1470 if (ret < 0) { 1471 /* 1472 * Unfortunately, we are unable to convert to a useful name 1473 * But we'll still print out the usage information 1474 */ 1475 rcu_read_unlock(); 1476 goto done; 1477 } 1478 rcu_read_unlock(); 1479 1480 printk(KERN_INFO "Task in %s killed", memcg_name); 1481 1482 rcu_read_lock(); 1483 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1484 if (ret < 0) { 1485 rcu_read_unlock(); 1486 goto done; 1487 } 1488 rcu_read_unlock(); 1489 1490 /* 1491 * Continues from above, so we don't need an KERN_ level 1492 */ 1493 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1494done: 1495 1496 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1497 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1498 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1499 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1500 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1501 "failcnt %llu\n", 1502 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1503 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1504 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1505} 1506 1507/* 1508 * This function returns the number of memcg under hierarchy tree. Returns 1509 * 1(self count) if no children. 1510 */ 1511static int mem_cgroup_count_children(struct mem_cgroup *memcg) 1512{ 1513 int num = 0; 1514 struct mem_cgroup *iter; 1515 1516 for_each_mem_cgroup_tree(iter, memcg) 1517 num++; 1518 return num; 1519} 1520 1521/* 1522 * Return the memory (and swap, if configured) limit for a memcg. 1523 */ 1524u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1525{ 1526 u64 limit; 1527 u64 memsw; 1528 1529 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1530 limit += total_swap_pages << PAGE_SHIFT; 1531 1532 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1533 /* 1534 * If memsw is finite and limits the amount of swap space available 1535 * to this memcg, return that limit. 1536 */ 1537 return min(limit, memsw); 1538} 1539 1540static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, 1541 gfp_t gfp_mask, 1542 unsigned long flags) 1543{ 1544 unsigned long total = 0; 1545 bool noswap = false; 1546 int loop; 1547 1548 if (flags & MEM_CGROUP_RECLAIM_NOSWAP) 1549 noswap = true; 1550 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) 1551 noswap = true; 1552 1553 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { 1554 if (loop) 1555 drain_all_stock_async(memcg); 1556 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); 1557 /* 1558 * Allow limit shrinkers, which are triggered directly 1559 * by userspace, to catch signals and stop reclaim 1560 * after minimal progress, regardless of the margin. 1561 */ 1562 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) 1563 break; 1564 if (mem_cgroup_margin(memcg)) 1565 break; 1566 /* 1567 * If nothing was reclaimed after two attempts, there 1568 * may be no reclaimable pages in this hierarchy. 1569 */ 1570 if (loop && !total) 1571 break; 1572 } 1573 return total; 1574} 1575 1576/** 1577 * test_mem_cgroup_node_reclaimable 1578 * @mem: the target memcg 1579 * @nid: the node ID to be checked. 1580 * @noswap : specify true here if the user wants flle only information. 1581 * 1582 * This function returns whether the specified memcg contains any 1583 * reclaimable pages on a node. Returns true if there are any reclaimable 1584 * pages in the node. 1585 */ 1586static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, 1587 int nid, bool noswap) 1588{ 1589 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) 1590 return true; 1591 if (noswap || !total_swap_pages) 1592 return false; 1593 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) 1594 return true; 1595 return false; 1596 1597} 1598#if MAX_NUMNODES > 1 1599 1600/* 1601 * Always updating the nodemask is not very good - even if we have an empty 1602 * list or the wrong list here, we can start from some node and traverse all 1603 * nodes based on the zonelist. So update the list loosely once per 10 secs. 1604 * 1605 */ 1606static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) 1607{ 1608 int nid; 1609 /* 1610 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET 1611 * pagein/pageout changes since the last update. 1612 */ 1613 if (!atomic_read(&memcg->numainfo_events)) 1614 return; 1615 if (atomic_inc_return(&memcg->numainfo_updating) > 1) 1616 return; 1617 1618 /* make a nodemask where this memcg uses memory from */ 1619 memcg->scan_nodes = node_states[N_HIGH_MEMORY]; 1620 1621 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { 1622 1623 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) 1624 node_clear(nid, memcg->scan_nodes); 1625 } 1626 1627 atomic_set(&memcg->numainfo_events, 0); 1628 atomic_set(&memcg->numainfo_updating, 0); 1629} 1630 1631/* 1632 * Selecting a node where we start reclaim from. Because what we need is just 1633 * reducing usage counter, start from anywhere is O,K. Considering 1634 * memory reclaim from current node, there are pros. and cons. 1635 * 1636 * Freeing memory from current node means freeing memory from a node which 1637 * we'll use or we've used. So, it may make LRU bad. And if several threads 1638 * hit limits, it will see a contention on a node. But freeing from remote 1639 * node means more costs for memory reclaim because of memory latency. 1640 * 1641 * Now, we use round-robin. Better algorithm is welcomed. 1642 */ 1643int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1644{ 1645 int node; 1646 1647 mem_cgroup_may_update_nodemask(memcg); 1648 node = memcg->last_scanned_node; 1649 1650 node = next_node(node, memcg->scan_nodes); 1651 if (node == MAX_NUMNODES) 1652 node = first_node(memcg->scan_nodes); 1653 /* 1654 * We call this when we hit limit, not when pages are added to LRU. 1655 * No LRU may hold pages because all pages are UNEVICTABLE or 1656 * memcg is too small and all pages are not on LRU. In that case, 1657 * we use curret node. 1658 */ 1659 if (unlikely(node == MAX_NUMNODES)) 1660 node = numa_node_id(); 1661 1662 memcg->last_scanned_node = node; 1663 return node; 1664} 1665 1666/* 1667 * Check all nodes whether it contains reclaimable pages or not. 1668 * For quick scan, we make use of scan_nodes. This will allow us to skip 1669 * unused nodes. But scan_nodes is lazily updated and may not cotain 1670 * enough new information. We need to do double check. 1671 */ 1672bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1673{ 1674 int nid; 1675 1676 /* 1677 * quick check...making use of scan_node. 1678 * We can skip unused nodes. 1679 */ 1680 if (!nodes_empty(memcg->scan_nodes)) { 1681 for (nid = first_node(memcg->scan_nodes); 1682 nid < MAX_NUMNODES; 1683 nid = next_node(nid, memcg->scan_nodes)) { 1684 1685 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1686 return true; 1687 } 1688 } 1689 /* 1690 * Check rest of nodes. 1691 */ 1692 for_each_node_state(nid, N_HIGH_MEMORY) { 1693 if (node_isset(nid, memcg->scan_nodes)) 1694 continue; 1695 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1696 return true; 1697 } 1698 return false; 1699} 1700 1701#else 1702int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1703{ 1704 return 0; 1705} 1706 1707bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1708{ 1709 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); 1710} 1711#endif 1712 1713static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, 1714 struct zone *zone, 1715 gfp_t gfp_mask, 1716 unsigned long *total_scanned) 1717{ 1718 struct mem_cgroup *victim = NULL; 1719 int total = 0; 1720 int loop = 0; 1721 unsigned long excess; 1722 unsigned long nr_scanned; 1723 struct mem_cgroup_reclaim_cookie reclaim = { 1724 .zone = zone, 1725 .priority = 0, 1726 }; 1727 1728 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; 1729 1730 while (1) { 1731 victim = mem_cgroup_iter(root_memcg, victim, &reclaim); 1732 if (!victim) { 1733 loop++; 1734 if (loop >= 2) { 1735 /* 1736 * If we have not been able to reclaim 1737 * anything, it might because there are 1738 * no reclaimable pages under this hierarchy 1739 */ 1740 if (!total) 1741 break; 1742 /* 1743 * We want to do more targeted reclaim. 1744 * excess >> 2 is not to excessive so as to 1745 * reclaim too much, nor too less that we keep 1746 * coming back to reclaim from this cgroup 1747 */ 1748 if (total >= (excess >> 2) || 1749 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) 1750 break; 1751 } 1752 continue; 1753 } 1754 if (!mem_cgroup_reclaimable(victim, false)) 1755 continue; 1756 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, 1757 zone, &nr_scanned); 1758 *total_scanned += nr_scanned; 1759 if (!res_counter_soft_limit_excess(&root_memcg->res)) 1760 break; 1761 } 1762 mem_cgroup_iter_break(root_memcg, victim); 1763 return total; 1764} 1765 1766/* 1767 * Check OOM-Killer is already running under our hierarchy. 1768 * If someone is running, return false. 1769 * Has to be called with memcg_oom_lock 1770 */ 1771static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) 1772{ 1773 struct mem_cgroup *iter, *failed = NULL; 1774 1775 for_each_mem_cgroup_tree(iter, memcg) { 1776 if (iter->oom_lock) { 1777 /* 1778 * this subtree of our hierarchy is already locked 1779 * so we cannot give a lock. 1780 */ 1781 failed = iter; 1782 mem_cgroup_iter_break(memcg, iter); 1783 break; 1784 } else 1785 iter->oom_lock = true; 1786 } 1787 1788 if (!failed) 1789 return true; 1790 1791 /* 1792 * OK, we failed to lock the whole subtree so we have to clean up 1793 * what we set up to the failing subtree 1794 */ 1795 for_each_mem_cgroup_tree(iter, memcg) { 1796 if (iter == failed) { 1797 mem_cgroup_iter_break(memcg, iter); 1798 break; 1799 } 1800 iter->oom_lock = false; 1801 } 1802 return false; 1803} 1804 1805/* 1806 * Has to be called with memcg_oom_lock 1807 */ 1808static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) 1809{ 1810 struct mem_cgroup *iter; 1811 1812 for_each_mem_cgroup_tree(iter, memcg) 1813 iter->oom_lock = false; 1814 return 0; 1815} 1816 1817static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) 1818{ 1819 struct mem_cgroup *iter; 1820 1821 for_each_mem_cgroup_tree(iter, memcg) 1822 atomic_inc(&iter->under_oom); 1823} 1824 1825static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) 1826{ 1827 struct mem_cgroup *iter; 1828 1829 /* 1830 * When a new child is created while the hierarchy is under oom, 1831 * mem_cgroup_oom_lock() may not be called. We have to use 1832 * atomic_add_unless() here. 1833 */ 1834 for_each_mem_cgroup_tree(iter, memcg) 1835 atomic_add_unless(&iter->under_oom, -1, 0); 1836} 1837 1838static DEFINE_SPINLOCK(memcg_oom_lock); 1839static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1840 1841struct oom_wait_info { 1842 struct mem_cgroup *mem; 1843 wait_queue_t wait; 1844}; 1845 1846static int memcg_oom_wake_function(wait_queue_t *wait, 1847 unsigned mode, int sync, void *arg) 1848{ 1849 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg, 1850 *oom_wait_memcg; 1851 struct oom_wait_info *oom_wait_info; 1852 1853 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1854 oom_wait_memcg = oom_wait_info->mem; 1855 1856 /* 1857 * Both of oom_wait_info->mem and wake_mem are stable under us. 1858 * Then we can use css_is_ancestor without taking care of RCU. 1859 */ 1860 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) 1861 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) 1862 return 0; 1863 return autoremove_wake_function(wait, mode, sync, arg); 1864} 1865 1866static void memcg_wakeup_oom(struct mem_cgroup *memcg) 1867{ 1868 /* for filtering, pass "memcg" as argument. */ 1869 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); 1870} 1871 1872static void memcg_oom_recover(struct mem_cgroup *memcg) 1873{ 1874 if (memcg && atomic_read(&memcg->under_oom)) 1875 memcg_wakeup_oom(memcg); 1876} 1877 1878/* 1879 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1880 */ 1881bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask) 1882{ 1883 struct oom_wait_info owait; 1884 bool locked, need_to_kill; 1885 1886 owait.mem = memcg; 1887 owait.wait.flags = 0; 1888 owait.wait.func = memcg_oom_wake_function; 1889 owait.wait.private = current; 1890 INIT_LIST_HEAD(&owait.wait.task_list); 1891 need_to_kill = true; 1892 mem_cgroup_mark_under_oom(memcg); 1893 1894 /* At first, try to OOM lock hierarchy under memcg.*/ 1895 spin_lock(&memcg_oom_lock); 1896 locked = mem_cgroup_oom_lock(memcg); 1897 /* 1898 * Even if signal_pending(), we can't quit charge() loop without 1899 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1900 * under OOM is always welcomed, use TASK_KILLABLE here. 1901 */ 1902 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1903 if (!locked || memcg->oom_kill_disable) 1904 need_to_kill = false; 1905 if (locked) 1906 mem_cgroup_oom_notify(memcg); 1907 spin_unlock(&memcg_oom_lock); 1908 1909 if (need_to_kill) { 1910 finish_wait(&memcg_oom_waitq, &owait.wait); 1911 mem_cgroup_out_of_memory(memcg, mask); 1912 } else { 1913 schedule(); 1914 finish_wait(&memcg_oom_waitq, &owait.wait); 1915 } 1916 spin_lock(&memcg_oom_lock); 1917 if (locked) 1918 mem_cgroup_oom_unlock(memcg); 1919 memcg_wakeup_oom(memcg); 1920 spin_unlock(&memcg_oom_lock); 1921 1922 mem_cgroup_unmark_under_oom(memcg); 1923 1924 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1925 return false; 1926 /* Give chance to dying process */ 1927 schedule_timeout_uninterruptible(1); 1928 return true; 1929} 1930 1931/* 1932 * Currently used to update mapped file statistics, but the routine can be 1933 * generalized to update other statistics as well. 1934 * 1935 * Notes: Race condition 1936 * 1937 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1938 * it tends to be costly. But considering some conditions, we doesn't need 1939 * to do so _always_. 1940 * 1941 * Considering "charge", lock_page_cgroup() is not required because all 1942 * file-stat operations happen after a page is attached to radix-tree. There 1943 * are no race with "charge". 1944 * 1945 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1946 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1947 * if there are race with "uncharge". Statistics itself is properly handled 1948 * by flags. 1949 * 1950 * Considering "move", this is an only case we see a race. To make the race 1951 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1952 * possibility of race condition. If there is, we take a lock. 1953 */ 1954 1955void mem_cgroup_update_page_stat(struct page *page, 1956 enum mem_cgroup_page_stat_item idx, int val) 1957{ 1958 struct mem_cgroup *memcg; 1959 struct page_cgroup *pc = lookup_page_cgroup(page); 1960 bool need_unlock = false; 1961 unsigned long uninitialized_var(flags); 1962 1963 if (unlikely(!pc)) 1964 return; 1965 1966 rcu_read_lock(); 1967 memcg = pc->mem_cgroup; 1968 if (unlikely(!memcg || !PageCgroupUsed(pc))) 1969 goto out; 1970 /* pc->mem_cgroup is unstable ? */ 1971 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) { 1972 /* take a lock against to access pc->mem_cgroup */ 1973 move_lock_page_cgroup(pc, &flags); 1974 need_unlock = true; 1975 memcg = pc->mem_cgroup; 1976 if (!memcg || !PageCgroupUsed(pc)) 1977 goto out; 1978 } 1979 1980 switch (idx) { 1981 case MEMCG_NR_FILE_MAPPED: 1982 if (val > 0) 1983 SetPageCgroupFileMapped(pc); 1984 else if (!page_mapped(page)) 1985 ClearPageCgroupFileMapped(pc); 1986 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1987 break; 1988 default: 1989 BUG(); 1990 } 1991 1992 this_cpu_add(memcg->stat->count[idx], val); 1993 1994out: 1995 if (unlikely(need_unlock)) 1996 move_unlock_page_cgroup(pc, &flags); 1997 rcu_read_unlock(); 1998 return; 1999} 2000EXPORT_SYMBOL(mem_cgroup_update_page_stat); 2001 2002/* 2003 * size of first charge trial. "32" comes from vmscan.c's magic value. 2004 * TODO: maybe necessary to use big numbers in big irons. 2005 */ 2006#define CHARGE_BATCH 32U 2007struct memcg_stock_pcp { 2008 struct mem_cgroup *cached; /* this never be root cgroup */ 2009 unsigned int nr_pages; 2010 struct work_struct work; 2011 unsigned long flags; 2012#define FLUSHING_CACHED_CHARGE (0) 2013}; 2014static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 2015static DEFINE_MUTEX(percpu_charge_mutex); 2016 2017/* 2018 * Try to consume stocked charge on this cpu. If success, one page is consumed 2019 * from local stock and true is returned. If the stock is 0 or charges from a 2020 * cgroup which is not current target, returns false. This stock will be 2021 * refilled. 2022 */ 2023static bool consume_stock(struct mem_cgroup *memcg) 2024{ 2025 struct memcg_stock_pcp *stock; 2026 bool ret = true; 2027 2028 stock = &get_cpu_var(memcg_stock); 2029 if (memcg == stock->cached && stock->nr_pages) 2030 stock->nr_pages--; 2031 else /* need to call res_counter_charge */ 2032 ret = false; 2033 put_cpu_var(memcg_stock); 2034 return ret; 2035} 2036 2037/* 2038 * Returns stocks cached in percpu to res_counter and reset cached information. 2039 */ 2040static void drain_stock(struct memcg_stock_pcp *stock) 2041{ 2042 struct mem_cgroup *old = stock->cached; 2043 2044 if (stock->nr_pages) { 2045 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 2046 2047 res_counter_uncharge(&old->res, bytes); 2048 if (do_swap_account) 2049 res_counter_uncharge(&old->memsw, bytes); 2050 stock->nr_pages = 0; 2051 } 2052 stock->cached = NULL; 2053} 2054 2055/* 2056 * This must be called under preempt disabled or must be called by 2057 * a thread which is pinned to local cpu. 2058 */ 2059static void drain_local_stock(struct work_struct *dummy) 2060{ 2061 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 2062 drain_stock(stock); 2063 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); 2064} 2065 2066/* 2067 * Cache charges(val) which is from res_counter, to local per_cpu area. 2068 * This will be consumed by consume_stock() function, later. 2069 */ 2070static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) 2071{ 2072 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 2073 2074 if (stock->cached != memcg) { /* reset if necessary */ 2075 drain_stock(stock); 2076 stock->cached = memcg; 2077 } 2078 stock->nr_pages += nr_pages; 2079 put_cpu_var(memcg_stock); 2080} 2081 2082/* 2083 * Drains all per-CPU charge caches for given root_memcg resp. subtree 2084 * of the hierarchy under it. sync flag says whether we should block 2085 * until the work is done. 2086 */ 2087static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) 2088{ 2089 int cpu, curcpu; 2090 2091 /* Notify other cpus that system-wide "drain" is running */ 2092 get_online_cpus(); 2093 curcpu = get_cpu(); 2094 for_each_online_cpu(cpu) { 2095 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2096 struct mem_cgroup *memcg; 2097 2098 memcg = stock->cached; 2099 if (!memcg || !stock->nr_pages) 2100 continue; 2101 if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) 2102 continue; 2103 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { 2104 if (cpu == curcpu) 2105 drain_local_stock(&stock->work); 2106 else 2107 schedule_work_on(cpu, &stock->work); 2108 } 2109 } 2110 put_cpu(); 2111 2112 if (!sync) 2113 goto out; 2114 2115 for_each_online_cpu(cpu) { 2116 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2117 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) 2118 flush_work(&stock->work); 2119 } 2120out: 2121 put_online_cpus(); 2122} 2123 2124/* 2125 * Tries to drain stocked charges in other cpus. This function is asynchronous 2126 * and just put a work per cpu for draining localy on each cpu. Caller can 2127 * expects some charges will be back to res_counter later but cannot wait for 2128 * it. 2129 */ 2130static void drain_all_stock_async(struct mem_cgroup *root_memcg) 2131{ 2132 /* 2133 * If someone calls draining, avoid adding more kworker runs. 2134 */ 2135 if (!mutex_trylock(&percpu_charge_mutex)) 2136 return; 2137 drain_all_stock(root_memcg, false); 2138 mutex_unlock(&percpu_charge_mutex); 2139} 2140 2141/* This is a synchronous drain interface. */ 2142static void drain_all_stock_sync(struct mem_cgroup *root_memcg) 2143{ 2144 /* called when force_empty is called */ 2145 mutex_lock(&percpu_charge_mutex); 2146 drain_all_stock(root_memcg, true); 2147 mutex_unlock(&percpu_charge_mutex); 2148} 2149 2150/* 2151 * This function drains percpu counter value from DEAD cpu and 2152 * move it to local cpu. Note that this function can be preempted. 2153 */ 2154static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) 2155{ 2156 int i; 2157 2158 spin_lock(&memcg->pcp_counter_lock); 2159 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 2160 long x = per_cpu(memcg->stat->count[i], cpu); 2161 2162 per_cpu(memcg->stat->count[i], cpu) = 0; 2163 memcg->nocpu_base.count[i] += x; 2164 } 2165 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { 2166 unsigned long x = per_cpu(memcg->stat->events[i], cpu); 2167 2168 per_cpu(memcg->stat->events[i], cpu) = 0; 2169 memcg->nocpu_base.events[i] += x; 2170 } 2171 /* need to clear ON_MOVE value, works as a kind of lock. */ 2172 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 2173 spin_unlock(&memcg->pcp_counter_lock); 2174} 2175 2176static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu) 2177{ 2178 int idx = MEM_CGROUP_ON_MOVE; 2179 2180 spin_lock(&memcg->pcp_counter_lock); 2181 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx]; 2182 spin_unlock(&memcg->pcp_counter_lock); 2183} 2184 2185static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 2186 unsigned long action, 2187 void *hcpu) 2188{ 2189 int cpu = (unsigned long)hcpu; 2190 struct memcg_stock_pcp *stock; 2191 struct mem_cgroup *iter; 2192 2193 if ((action == CPU_ONLINE)) { 2194 for_each_mem_cgroup(iter) 2195 synchronize_mem_cgroup_on_move(iter, cpu); 2196 return NOTIFY_OK; 2197 } 2198 2199 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 2200 return NOTIFY_OK; 2201 2202 for_each_mem_cgroup(iter) 2203 mem_cgroup_drain_pcp_counter(iter, cpu); 2204 2205 stock = &per_cpu(memcg_stock, cpu); 2206 drain_stock(stock); 2207 return NOTIFY_OK; 2208} 2209 2210 2211/* See __mem_cgroup_try_charge() for details */ 2212enum { 2213 CHARGE_OK, /* success */ 2214 CHARGE_RETRY, /* need to retry but retry is not bad */ 2215 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 2216 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 2217 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 2218}; 2219 2220static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, 2221 unsigned int nr_pages, bool oom_check) 2222{ 2223 unsigned long csize = nr_pages * PAGE_SIZE; 2224 struct mem_cgroup *mem_over_limit; 2225 struct res_counter *fail_res; 2226 unsigned long flags = 0; 2227 int ret; 2228 2229 ret = res_counter_charge(&memcg->res, csize, &fail_res); 2230 2231 if (likely(!ret)) { 2232 if (!do_swap_account) 2233 return CHARGE_OK; 2234 ret = res_counter_charge(&memcg->memsw, csize, &fail_res); 2235 if (likely(!ret)) 2236 return CHARGE_OK; 2237 2238 res_counter_uncharge(&memcg->res, csize); 2239 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 2240 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 2241 } else 2242 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 2243 /* 2244 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch 2245 * of regular pages (CHARGE_BATCH), or a single regular page (1). 2246 * 2247 * Never reclaim on behalf of optional batching, retry with a 2248 * single page instead. 2249 */ 2250 if (nr_pages == CHARGE_BATCH) 2251 return CHARGE_RETRY; 2252 2253 if (!(gfp_mask & __GFP_WAIT)) 2254 return CHARGE_WOULDBLOCK; 2255 2256 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); 2257 if (mem_cgroup_margin(mem_over_limit) >= nr_pages) 2258 return CHARGE_RETRY; 2259 /* 2260 * Even though the limit is exceeded at this point, reclaim 2261 * may have been able to free some pages. Retry the charge 2262 * before killing the task. 2263 * 2264 * Only for regular pages, though: huge pages are rather 2265 * unlikely to succeed so close to the limit, and we fall back 2266 * to regular pages anyway in case of failure. 2267 */ 2268 if (nr_pages == 1 && ret) 2269 return CHARGE_RETRY; 2270 2271 /* 2272 * At task move, charge accounts can be doubly counted. So, it's 2273 * better to wait until the end of task_move if something is going on. 2274 */ 2275 if (mem_cgroup_wait_acct_move(mem_over_limit)) 2276 return CHARGE_RETRY; 2277 2278 /* If we don't need to call oom-killer at el, return immediately */ 2279 if (!oom_check) 2280 return CHARGE_NOMEM; 2281 /* check OOM */ 2282 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 2283 return CHARGE_OOM_DIE; 2284 2285 return CHARGE_RETRY; 2286} 2287 2288/* 2289 * Unlike exported interface, "oom" parameter is added. if oom==true, 2290 * oom-killer can be invoked. 2291 */ 2292static int __mem_cgroup_try_charge(struct mm_struct *mm, 2293 gfp_t gfp_mask, 2294 unsigned int nr_pages, 2295 struct mem_cgroup **ptr, 2296 bool oom) 2297{ 2298 unsigned int batch = max(CHARGE_BATCH, nr_pages); 2299 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2300 struct mem_cgroup *memcg = NULL; 2301 int ret; 2302 2303 /* 2304 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 2305 * in system level. So, allow to go ahead dying process in addition to 2306 * MEMDIE process. 2307 */ 2308 if (unlikely(test_thread_flag(TIF_MEMDIE) 2309 || fatal_signal_pending(current))) 2310 goto bypass; 2311 2312 /* 2313 * We always charge the cgroup the mm_struct belongs to. 2314 * The mm_struct's mem_cgroup changes on task migration if the 2315 * thread group leader migrates. It's possible that mm is not 2316 * set, if so charge the init_mm (happens for pagecache usage). 2317 */ 2318 if (!*ptr && !mm) 2319 goto bypass; 2320again: 2321 if (*ptr) { /* css should be a valid one */ 2322 memcg = *ptr; 2323 VM_BUG_ON(css_is_removed(&memcg->css)); 2324 if (mem_cgroup_is_root(memcg)) 2325 goto done; 2326 if (nr_pages == 1 && consume_stock(memcg)) 2327 goto done; 2328 css_get(&memcg->css); 2329 } else { 2330 struct task_struct *p; 2331 2332 rcu_read_lock(); 2333 p = rcu_dereference(mm->owner); 2334 /* 2335 * Because we don't have task_lock(), "p" can exit. 2336 * In that case, "memcg" can point to root or p can be NULL with 2337 * race with swapoff. Then, we have small risk of mis-accouning. 2338 * But such kind of mis-account by race always happens because 2339 * we don't have cgroup_mutex(). It's overkill and we allo that 2340 * small race, here. 2341 * (*) swapoff at el will charge against mm-struct not against 2342 * task-struct. So, mm->owner can be NULL. 2343 */ 2344 memcg = mem_cgroup_from_task(p); 2345 if (!memcg || mem_cgroup_is_root(memcg)) { 2346 rcu_read_unlock(); 2347 goto done; 2348 } 2349 if (nr_pages == 1 && consume_stock(memcg)) { 2350 /* 2351 * It seems dagerous to access memcg without css_get(). 2352 * But considering how consume_stok works, it's not 2353 * necessary. If consume_stock success, some charges 2354 * from this memcg are cached on this cpu. So, we 2355 * don't need to call css_get()/css_tryget() before 2356 * calling consume_stock(). 2357 */ 2358 rcu_read_unlock(); 2359 goto done; 2360 } 2361 /* after here, we may be blocked. we need to get refcnt */ 2362 if (!css_tryget(&memcg->css)) { 2363 rcu_read_unlock(); 2364 goto again; 2365 } 2366 rcu_read_unlock(); 2367 } 2368 2369 do { 2370 bool oom_check; 2371 2372 /* If killed, bypass charge */ 2373 if (fatal_signal_pending(current)) { 2374 css_put(&memcg->css); 2375 goto bypass; 2376 } 2377 2378 oom_check = false; 2379 if (oom && !nr_oom_retries) { 2380 oom_check = true; 2381 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2382 } 2383 2384 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); 2385 switch (ret) { 2386 case CHARGE_OK: 2387 break; 2388 case CHARGE_RETRY: /* not in OOM situation but retry */ 2389 batch = nr_pages; 2390 css_put(&memcg->css); 2391 memcg = NULL; 2392 goto again; 2393 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2394 css_put(&memcg->css); 2395 goto nomem; 2396 case CHARGE_NOMEM: /* OOM routine works */ 2397 if (!oom) { 2398 css_put(&memcg->css); 2399 goto nomem; 2400 } 2401 /* If oom, we never return -ENOMEM */ 2402 nr_oom_retries--; 2403 break; 2404 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2405 css_put(&memcg->css); 2406 goto bypass; 2407 } 2408 } while (ret != CHARGE_OK); 2409 2410 if (batch > nr_pages) 2411 refill_stock(memcg, batch - nr_pages); 2412 css_put(&memcg->css); 2413done: 2414 *ptr = memcg; 2415 return 0; 2416nomem: 2417 *ptr = NULL; 2418 return -ENOMEM; 2419bypass: 2420 *ptr = NULL; 2421 return 0; 2422} 2423 2424/* 2425 * Somemtimes we have to undo a charge we got by try_charge(). 2426 * This function is for that and do uncharge, put css's refcnt. 2427 * gotten by try_charge(). 2428 */ 2429static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, 2430 unsigned int nr_pages) 2431{ 2432 if (!mem_cgroup_is_root(memcg)) { 2433 unsigned long bytes = nr_pages * PAGE_SIZE; 2434 2435 res_counter_uncharge(&memcg->res, bytes); 2436 if (do_swap_account) 2437 res_counter_uncharge(&memcg->memsw, bytes); 2438 } 2439} 2440 2441/* 2442 * A helper function to get mem_cgroup from ID. must be called under 2443 * rcu_read_lock(). The caller must check css_is_removed() or some if 2444 * it's concern. (dropping refcnt from swap can be called against removed 2445 * memcg.) 2446 */ 2447static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2448{ 2449 struct cgroup_subsys_state *css; 2450 2451 /* ID 0 is unused ID */ 2452 if (!id) 2453 return NULL; 2454 css = css_lookup(&mem_cgroup_subsys, id); 2455 if (!css) 2456 return NULL; 2457 return container_of(css, struct mem_cgroup, css); 2458} 2459 2460struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2461{ 2462 struct mem_cgroup *memcg = NULL; 2463 struct page_cgroup *pc; 2464 unsigned short id; 2465 swp_entry_t ent; 2466 2467 VM_BUG_ON(!PageLocked(page)); 2468 2469 pc = lookup_page_cgroup(page); 2470 lock_page_cgroup(pc); 2471 if (PageCgroupUsed(pc)) { 2472 memcg = pc->mem_cgroup; 2473 if (memcg && !css_tryget(&memcg->css)) 2474 memcg = NULL; 2475 } else if (PageSwapCache(page)) { 2476 ent.val = page_private(page); 2477 id = lookup_swap_cgroup(ent); 2478 rcu_read_lock(); 2479 memcg = mem_cgroup_lookup(id); 2480 if (memcg && !css_tryget(&memcg->css)) 2481 memcg = NULL; 2482 rcu_read_unlock(); 2483 } 2484 unlock_page_cgroup(pc); 2485 return memcg; 2486} 2487 2488static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, 2489 struct page *page, 2490 unsigned int nr_pages, 2491 struct page_cgroup *pc, 2492 enum charge_type ctype) 2493{ 2494 lock_page_cgroup(pc); 2495 if (unlikely(PageCgroupUsed(pc))) { 2496 unlock_page_cgroup(pc); 2497 __mem_cgroup_cancel_charge(memcg, nr_pages); 2498 return; 2499 } 2500 /* 2501 * we don't need page_cgroup_lock about tail pages, becase they are not 2502 * accessed by any other context at this point. 2503 */ 2504 pc->mem_cgroup = memcg; 2505 /* 2506 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2507 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2508 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2509 * before USED bit, we need memory barrier here. 2510 * See mem_cgroup_add_lru_list(), etc. 2511 */ 2512 smp_wmb(); 2513 switch (ctype) { 2514 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2515 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2516 SetPageCgroupCache(pc); 2517 SetPageCgroupUsed(pc); 2518 break; 2519 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2520 ClearPageCgroupCache(pc); 2521 SetPageCgroupUsed(pc); 2522 break; 2523 default: 2524 break; 2525 } 2526 2527 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages); 2528 unlock_page_cgroup(pc); 2529 /* 2530 * "charge_statistics" updated event counter. Then, check it. 2531 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2532 * if they exceeds softlimit. 2533 */ 2534 memcg_check_events(memcg, page); 2535} 2536 2537#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2538 2539#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2540 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) 2541/* 2542 * Because tail pages are not marked as "used", set it. We're under 2543 * zone->lru_lock, 'splitting on pmd' and compound_lock. 2544 * charge/uncharge will be never happen and move_account() is done under 2545 * compound_lock(), so we don't have to take care of races. 2546 */ 2547void mem_cgroup_split_huge_fixup(struct page *head) 2548{ 2549 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2550 struct page_cgroup *pc; 2551 int i; 2552 2553 if (mem_cgroup_disabled()) 2554 return; 2555 for (i = 1; i < HPAGE_PMD_NR; i++) { 2556 pc = head_pc + i; 2557 pc->mem_cgroup = head_pc->mem_cgroup; 2558 smp_wmb();/* see __commit_charge() */ 2559 /* 2560 * LRU flags cannot be copied because we need to add tail 2561 * page to LRU by generic call and our hooks will be called. 2562 */ 2563 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2564 } 2565 2566 if (PageCgroupAcctLRU(head_pc)) { 2567 enum lru_list lru; 2568 struct mem_cgroup_per_zone *mz; 2569 /* 2570 * We hold lru_lock, then, reduce counter directly. 2571 */ 2572 lru = page_lru(head); 2573 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); 2574 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1; 2575 } 2576} 2577#endif 2578 2579/** 2580 * mem_cgroup_move_account - move account of the page 2581 * @page: the page 2582 * @nr_pages: number of regular pages (>1 for huge pages) 2583 * @pc: page_cgroup of the page. 2584 * @from: mem_cgroup which the page is moved from. 2585 * @to: mem_cgroup which the page is moved to. @from != @to. 2586 * @uncharge: whether we should call uncharge and css_put against @from. 2587 * 2588 * The caller must confirm following. 2589 * - page is not on LRU (isolate_page() is useful.) 2590 * - compound_lock is held when nr_pages > 1 2591 * 2592 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2593 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is 2594 * true, this function does "uncharge" from old cgroup, but it doesn't if 2595 * @uncharge is false, so a caller should do "uncharge". 2596 */ 2597static int mem_cgroup_move_account(struct page *page, 2598 unsigned int nr_pages, 2599 struct page_cgroup *pc, 2600 struct mem_cgroup *from, 2601 struct mem_cgroup *to, 2602 bool uncharge) 2603{ 2604 unsigned long flags; 2605 int ret; 2606 2607 VM_BUG_ON(from == to); 2608 VM_BUG_ON(PageLRU(page)); 2609 /* 2610 * The page is isolated from LRU. So, collapse function 2611 * will not handle this page. But page splitting can happen. 2612 * Do this check under compound_page_lock(). The caller should 2613 * hold it. 2614 */ 2615 ret = -EBUSY; 2616 if (nr_pages > 1 && !PageTransHuge(page)) 2617 goto out; 2618 2619 lock_page_cgroup(pc); 2620 2621 ret = -EINVAL; 2622 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2623 goto unlock; 2624 2625 move_lock_page_cgroup(pc, &flags); 2626 2627 if (PageCgroupFileMapped(pc)) { 2628 /* Update mapped_file data for mem_cgroup */ 2629 preempt_disable(); 2630 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2631 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2632 preempt_enable(); 2633 } 2634 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2635 if (uncharge) 2636 /* This is not "cancel", but cancel_charge does all we need. */ 2637 __mem_cgroup_cancel_charge(from, nr_pages); 2638 2639 /* caller should have done css_get */ 2640 pc->mem_cgroup = to; 2641 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2642 /* 2643 * We charges against "to" which may not have any tasks. Then, "to" 2644 * can be under rmdir(). But in current implementation, caller of 2645 * this function is just force_empty() and move charge, so it's 2646 * guaranteed that "to" is never removed. So, we don't check rmdir 2647 * status here. 2648 */ 2649 move_unlock_page_cgroup(pc, &flags); 2650 ret = 0; 2651unlock: 2652 unlock_page_cgroup(pc); 2653 /* 2654 * check events 2655 */ 2656 memcg_check_events(to, page); 2657 memcg_check_events(from, page); 2658out: 2659 return ret; 2660} 2661 2662/* 2663 * move charges to its parent. 2664 */ 2665 2666static int mem_cgroup_move_parent(struct page *page, 2667 struct page_cgroup *pc, 2668 struct mem_cgroup *child, 2669 gfp_t gfp_mask) 2670{ 2671 struct cgroup *cg = child->css.cgroup; 2672 struct cgroup *pcg = cg->parent; 2673 struct mem_cgroup *parent; 2674 unsigned int nr_pages; 2675 unsigned long uninitialized_var(flags); 2676 int ret; 2677 2678 /* Is ROOT ? */ 2679 if (!pcg) 2680 return -EINVAL; 2681 2682 ret = -EBUSY; 2683 if (!get_page_unless_zero(page)) 2684 goto out; 2685 if (isolate_lru_page(page)) 2686 goto put; 2687 2688 nr_pages = hpage_nr_pages(page); 2689 2690 parent = mem_cgroup_from_cont(pcg); 2691 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); 2692 if (ret || !parent) 2693 goto put_back; 2694 2695 if (nr_pages > 1) 2696 flags = compound_lock_irqsave(page); 2697 2698 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); 2699 if (ret) 2700 __mem_cgroup_cancel_charge(parent, nr_pages); 2701 2702 if (nr_pages > 1) 2703 compound_unlock_irqrestore(page, flags); 2704put_back: 2705 putback_lru_page(page); 2706put: 2707 put_page(page); 2708out: 2709 return ret; 2710} 2711 2712/* 2713 * Charge the memory controller for page usage. 2714 * Return 2715 * 0 if the charge was successful 2716 * < 0 if the cgroup is over its limit 2717 */ 2718static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2719 gfp_t gfp_mask, enum charge_type ctype) 2720{ 2721 struct mem_cgroup *memcg = NULL; 2722 unsigned int nr_pages = 1; 2723 struct page_cgroup *pc; 2724 bool oom = true; 2725 int ret; 2726 2727 if (PageTransHuge(page)) { 2728 nr_pages <<= compound_order(page); 2729 VM_BUG_ON(!PageTransHuge(page)); 2730 /* 2731 * Never OOM-kill a process for a huge page. The 2732 * fault handler will fall back to regular pages. 2733 */ 2734 oom = false; 2735 } 2736 2737 pc = lookup_page_cgroup(page); 2738 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */ 2739 2740 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); 2741 if (ret || !memcg) 2742 return ret; 2743 2744 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype); 2745 return 0; 2746} 2747 2748int mem_cgroup_newpage_charge(struct page *page, 2749 struct mm_struct *mm, gfp_t gfp_mask) 2750{ 2751 if (mem_cgroup_disabled()) 2752 return 0; 2753 /* 2754 * If already mapped, we don't have to account. 2755 * If page cache, page->mapping has address_space. 2756 * But page->mapping may have out-of-use anon_vma pointer, 2757 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 2758 * is NULL. 2759 */ 2760 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 2761 return 0; 2762 if (unlikely(!mm)) 2763 mm = &init_mm; 2764 return mem_cgroup_charge_common(page, mm, gfp_mask, 2765 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2766} 2767 2768static void 2769__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2770 enum charge_type ctype); 2771 2772static void 2773__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg, 2774 enum charge_type ctype) 2775{ 2776 struct page_cgroup *pc = lookup_page_cgroup(page); 2777 /* 2778 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page 2779 * is already on LRU. It means the page may on some other page_cgroup's 2780 * LRU. Take care of it. 2781 */ 2782 mem_cgroup_lru_del_before_commit(page); 2783 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 2784 mem_cgroup_lru_add_after_commit(page); 2785 return; 2786} 2787 2788int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2789 gfp_t gfp_mask) 2790{ 2791 struct mem_cgroup *memcg = NULL; 2792 int ret; 2793 2794 if (mem_cgroup_disabled()) 2795 return 0; 2796 if (PageCompound(page)) 2797 return 0; 2798 2799 if (unlikely(!mm)) 2800 mm = &init_mm; 2801 2802 if (page_is_file_cache(page)) { 2803 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true); 2804 if (ret || !memcg) 2805 return ret; 2806 2807 /* 2808 * FUSE reuses pages without going through the final 2809 * put that would remove them from the LRU list, make 2810 * sure that they get relinked properly. 2811 */ 2812 __mem_cgroup_commit_charge_lrucare(page, memcg, 2813 MEM_CGROUP_CHARGE_TYPE_CACHE); 2814 return ret; 2815 } 2816 /* shmem */ 2817 if (PageSwapCache(page)) { 2818 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); 2819 if (!ret) 2820 __mem_cgroup_commit_charge_swapin(page, memcg, 2821 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2822 } else 2823 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 2824 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2825 2826 return ret; 2827} 2828 2829/* 2830 * While swap-in, try_charge -> commit or cancel, the page is locked. 2831 * And when try_charge() successfully returns, one refcnt to memcg without 2832 * struct page_cgroup is acquired. This refcnt will be consumed by 2833 * "commit()" or removed by "cancel()" 2834 */ 2835int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2836 struct page *page, 2837 gfp_t mask, struct mem_cgroup **memcgp) 2838{ 2839 struct mem_cgroup *memcg; 2840 int ret; 2841 2842 *memcgp = NULL; 2843 2844 if (mem_cgroup_disabled()) 2845 return 0; 2846 2847 if (!do_swap_account) 2848 goto charge_cur_mm; 2849 /* 2850 * A racing thread's fault, or swapoff, may have already updated 2851 * the pte, and even removed page from swap cache: in those cases 2852 * do_swap_page()'s pte_same() test will fail; but there's also a 2853 * KSM case which does need to charge the page. 2854 */ 2855 if (!PageSwapCache(page)) 2856 goto charge_cur_mm; 2857 memcg = try_get_mem_cgroup_from_page(page); 2858 if (!memcg) 2859 goto charge_cur_mm; 2860 *memcgp = memcg; 2861 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); 2862 css_put(&memcg->css); 2863 return ret; 2864charge_cur_mm: 2865 if (unlikely(!mm)) 2866 mm = &init_mm; 2867 return __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); 2868} 2869 2870static void 2871__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, 2872 enum charge_type ctype) 2873{ 2874 if (mem_cgroup_disabled()) 2875 return; 2876 if (!memcg) 2877 return; 2878 cgroup_exclude_rmdir(&memcg->css); 2879 2880 __mem_cgroup_commit_charge_lrucare(page, memcg, ctype); 2881 /* 2882 * Now swap is on-memory. This means this page may be 2883 * counted both as mem and swap....double count. 2884 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2885 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2886 * may call delete_from_swap_cache() before reach here. 2887 */ 2888 if (do_swap_account && PageSwapCache(page)) { 2889 swp_entry_t ent = {.val = page_private(page)}; 2890 struct mem_cgroup *swap_memcg; 2891 unsigned short id; 2892 2893 id = swap_cgroup_record(ent, 0); 2894 rcu_read_lock(); 2895 swap_memcg = mem_cgroup_lookup(id); 2896 if (swap_memcg) { 2897 /* 2898 * This recorded memcg can be obsolete one. So, avoid 2899 * calling css_tryget 2900 */ 2901 if (!mem_cgroup_is_root(swap_memcg)) 2902 res_counter_uncharge(&swap_memcg->memsw, 2903 PAGE_SIZE); 2904 mem_cgroup_swap_statistics(swap_memcg, false); 2905 mem_cgroup_put(swap_memcg); 2906 } 2907 rcu_read_unlock(); 2908 } 2909 /* 2910 * At swapin, we may charge account against cgroup which has no tasks. 2911 * So, rmdir()->pre_destroy() can be called while we do this charge. 2912 * In that case, we need to call pre_destroy() again. check it here. 2913 */ 2914 cgroup_release_and_wakeup_rmdir(&memcg->css); 2915} 2916 2917void mem_cgroup_commit_charge_swapin(struct page *page, 2918 struct mem_cgroup *memcg) 2919{ 2920 __mem_cgroup_commit_charge_swapin(page, memcg, 2921 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2922} 2923 2924void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) 2925{ 2926 if (mem_cgroup_disabled()) 2927 return; 2928 if (!memcg) 2929 return; 2930 __mem_cgroup_cancel_charge(memcg, 1); 2931} 2932 2933static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, 2934 unsigned int nr_pages, 2935 const enum charge_type ctype) 2936{ 2937 struct memcg_batch_info *batch = NULL; 2938 bool uncharge_memsw = true; 2939 2940 /* If swapout, usage of swap doesn't decrease */ 2941 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2942 uncharge_memsw = false; 2943 2944 batch = ¤t->memcg_batch; 2945 /* 2946 * In usual, we do css_get() when we remember memcg pointer. 2947 * But in this case, we keep res->usage until end of a series of 2948 * uncharges. Then, it's ok to ignore memcg's refcnt. 2949 */ 2950 if (!batch->memcg) 2951 batch->memcg = memcg; 2952 /* 2953 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2954 * In those cases, all pages freed continuously can be expected to be in 2955 * the same cgroup and we have chance to coalesce uncharges. 2956 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2957 * because we want to do uncharge as soon as possible. 2958 */ 2959 2960 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2961 goto direct_uncharge; 2962 2963 if (nr_pages > 1) 2964 goto direct_uncharge; 2965 2966 /* 2967 * In typical case, batch->memcg == mem. This means we can 2968 * merge a series of uncharges to an uncharge of res_counter. 2969 * If not, we uncharge res_counter ony by one. 2970 */ 2971 if (batch->memcg != memcg) 2972 goto direct_uncharge; 2973 /* remember freed charge and uncharge it later */ 2974 batch->nr_pages++; 2975 if (uncharge_memsw) 2976 batch->memsw_nr_pages++; 2977 return; 2978direct_uncharge: 2979 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); 2980 if (uncharge_memsw) 2981 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); 2982 if (unlikely(batch->memcg != memcg)) 2983 memcg_oom_recover(memcg); 2984 return; 2985} 2986 2987/* 2988 * uncharge if !page_mapped(page) 2989 */ 2990static struct mem_cgroup * 2991__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 2992{ 2993 struct mem_cgroup *memcg = NULL; 2994 unsigned int nr_pages = 1; 2995 struct page_cgroup *pc; 2996 2997 if (mem_cgroup_disabled()) 2998 return NULL; 2999 3000 if (PageSwapCache(page)) 3001 return NULL; 3002 3003 if (PageTransHuge(page)) { 3004 nr_pages <<= compound_order(page); 3005 VM_BUG_ON(!PageTransHuge(page)); 3006 } 3007 /* 3008 * Check if our page_cgroup is valid 3009 */ 3010 pc = lookup_page_cgroup(page); 3011 if (unlikely(!pc || !PageCgroupUsed(pc))) 3012 return NULL; 3013 3014 lock_page_cgroup(pc); 3015 3016 memcg = pc->mem_cgroup; 3017 3018 if (!PageCgroupUsed(pc)) 3019 goto unlock_out; 3020 3021 switch (ctype) { 3022 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 3023 case MEM_CGROUP_CHARGE_TYPE_DROP: 3024 /* See mem_cgroup_prepare_migration() */ 3025 if (page_mapped(page) || PageCgroupMigration(pc)) 3026 goto unlock_out; 3027 break; 3028 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 3029 if (!PageAnon(page)) { /* Shared memory */ 3030 if (page->mapping && !page_is_file_cache(page)) 3031 goto unlock_out; 3032 } else if (page_mapped(page)) /* Anon */ 3033 goto unlock_out; 3034 break; 3035 default: 3036 break; 3037 } 3038 3039 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages); 3040 3041 ClearPageCgroupUsed(pc); 3042 /* 3043 * pc->mem_cgroup is not cleared here. It will be accessed when it's 3044 * freed from LRU. This is safe because uncharged page is expected not 3045 * to be reused (freed soon). Exception is SwapCache, it's handled by 3046 * special functions. 3047 */ 3048 3049 unlock_page_cgroup(pc); 3050 /* 3051 * even after unlock, we have memcg->res.usage here and this memcg 3052 * will never be freed. 3053 */ 3054 memcg_check_events(memcg, page); 3055 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 3056 mem_cgroup_swap_statistics(memcg, true); 3057 mem_cgroup_get(memcg); 3058 } 3059 if (!mem_cgroup_is_root(memcg)) 3060 mem_cgroup_do_uncharge(memcg, nr_pages, ctype); 3061 3062 return memcg; 3063 3064unlock_out: 3065 unlock_page_cgroup(pc); 3066 return NULL; 3067} 3068 3069void mem_cgroup_uncharge_page(struct page *page) 3070{ 3071 /* early check. */ 3072 if (page_mapped(page)) 3073 return; 3074 if (page->mapping && !PageAnon(page)) 3075 return; 3076 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 3077} 3078 3079void mem_cgroup_uncharge_cache_page(struct page *page) 3080{ 3081 VM_BUG_ON(page_mapped(page)); 3082 VM_BUG_ON(page->mapping); 3083 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 3084} 3085 3086/* 3087 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 3088 * In that cases, pages are freed continuously and we can expect pages 3089 * are in the same memcg. All these calls itself limits the number of 3090 * pages freed at once, then uncharge_start/end() is called properly. 3091 * This may be called prural(2) times in a context, 3092 */ 3093 3094void mem_cgroup_uncharge_start(void) 3095{ 3096 current->memcg_batch.do_batch++; 3097 /* We can do nest. */ 3098 if (current->memcg_batch.do_batch == 1) { 3099 current->memcg_batch.memcg = NULL; 3100 current->memcg_batch.nr_pages = 0; 3101 current->memcg_batch.memsw_nr_pages = 0; 3102 } 3103} 3104 3105void mem_cgroup_uncharge_end(void) 3106{ 3107 struct memcg_batch_info *batch = ¤t->memcg_batch; 3108 3109 if (!batch->do_batch) 3110 return; 3111 3112 batch->do_batch--; 3113 if (batch->do_batch) /* If stacked, do nothing. */ 3114 return; 3115 3116 if (!batch->memcg) 3117 return; 3118 /* 3119 * This "batch->memcg" is valid without any css_get/put etc... 3120 * bacause we hide charges behind us. 3121 */ 3122 if (batch->nr_pages) 3123 res_counter_uncharge(&batch->memcg->res, 3124 batch->nr_pages * PAGE_SIZE); 3125 if (batch->memsw_nr_pages) 3126 res_counter_uncharge(&batch->memcg->memsw, 3127 batch->memsw_nr_pages * PAGE_SIZE); 3128 memcg_oom_recover(batch->memcg); 3129 /* forget this pointer (for sanity check) */ 3130 batch->memcg = NULL; 3131} 3132 3133#ifdef CONFIG_SWAP 3134/* 3135 * called after __delete_from_swap_cache() and drop "page" account. 3136 * memcg information is recorded to swap_cgroup of "ent" 3137 */ 3138void 3139mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 3140{ 3141 struct mem_cgroup *memcg; 3142 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 3143 3144 if (!swapout) /* this was a swap cache but the swap is unused ! */ 3145 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 3146 3147 memcg = __mem_cgroup_uncharge_common(page, ctype); 3148 3149 /* 3150 * record memcg information, if swapout && memcg != NULL, 3151 * mem_cgroup_get() was called in uncharge(). 3152 */ 3153 if (do_swap_account && swapout && memcg) 3154 swap_cgroup_record(ent, css_id(&memcg->css)); 3155} 3156#endif 3157 3158#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3159/* 3160 * called from swap_entry_free(). remove record in swap_cgroup and 3161 * uncharge "memsw" account. 3162 */ 3163void mem_cgroup_uncharge_swap(swp_entry_t ent) 3164{ 3165 struct mem_cgroup *memcg; 3166 unsigned short id; 3167 3168 if (!do_swap_account) 3169 return; 3170 3171 id = swap_cgroup_record(ent, 0); 3172 rcu_read_lock(); 3173 memcg = mem_cgroup_lookup(id); 3174 if (memcg) { 3175 /* 3176 * We uncharge this because swap is freed. 3177 * This memcg can be obsolete one. We avoid calling css_tryget 3178 */ 3179 if (!mem_cgroup_is_root(memcg)) 3180 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 3181 mem_cgroup_swap_statistics(memcg, false); 3182 mem_cgroup_put(memcg); 3183 } 3184 rcu_read_unlock(); 3185} 3186 3187/** 3188 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 3189 * @entry: swap entry to be moved 3190 * @from: mem_cgroup which the entry is moved from 3191 * @to: mem_cgroup which the entry is moved to 3192 * @need_fixup: whether we should fixup res_counters and refcounts. 3193 * 3194 * It succeeds only when the swap_cgroup's record for this entry is the same 3195 * as the mem_cgroup's id of @from. 3196 * 3197 * Returns 0 on success, -EINVAL on failure. 3198 * 3199 * The caller must have charged to @to, IOW, called res_counter_charge() about 3200 * both res and memsw, and called css_get(). 3201 */ 3202static int mem_cgroup_move_swap_account(swp_entry_t entry, 3203 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3204{ 3205 unsigned short old_id, new_id; 3206 3207 old_id = css_id(&from->css); 3208 new_id = css_id(&to->css); 3209 3210 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 3211 mem_cgroup_swap_statistics(from, false); 3212 mem_cgroup_swap_statistics(to, true); 3213 /* 3214 * This function is only called from task migration context now. 3215 * It postpones res_counter and refcount handling till the end 3216 * of task migration(mem_cgroup_clear_mc()) for performance 3217 * improvement. But we cannot postpone mem_cgroup_get(to) 3218 * because if the process that has been moved to @to does 3219 * swap-in, the refcount of @to might be decreased to 0. 3220 */ 3221 mem_cgroup_get(to); 3222 if (need_fixup) { 3223 if (!mem_cgroup_is_root(from)) 3224 res_counter_uncharge(&from->memsw, PAGE_SIZE); 3225 mem_cgroup_put(from); 3226 /* 3227 * we charged both to->res and to->memsw, so we should 3228 * uncharge to->res. 3229 */ 3230 if (!mem_cgroup_is_root(to)) 3231 res_counter_uncharge(&to->res, PAGE_SIZE); 3232 } 3233 return 0; 3234 } 3235 return -EINVAL; 3236} 3237#else 3238static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 3239 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3240{ 3241 return -EINVAL; 3242} 3243#endif 3244 3245/* 3246 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 3247 * page belongs to. 3248 */ 3249int mem_cgroup_prepare_migration(struct page *page, 3250 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) 3251{ 3252 struct mem_cgroup *memcg = NULL; 3253 struct page_cgroup *pc; 3254 enum charge_type ctype; 3255 int ret = 0; 3256 3257 *memcgp = NULL; 3258 3259 VM_BUG_ON(PageTransHuge(page)); 3260 if (mem_cgroup_disabled()) 3261 return 0; 3262 3263 pc = lookup_page_cgroup(page); 3264 lock_page_cgroup(pc); 3265 if (PageCgroupUsed(pc)) { 3266 memcg = pc->mem_cgroup; 3267 css_get(&memcg->css); 3268 /* 3269 * At migrating an anonymous page, its mapcount goes down 3270 * to 0 and uncharge() will be called. But, even if it's fully 3271 * unmapped, migration may fail and this page has to be 3272 * charged again. We set MIGRATION flag here and delay uncharge 3273 * until end_migration() is called 3274 * 3275 * Corner Case Thinking 3276 * A) 3277 * When the old page was mapped as Anon and it's unmap-and-freed 3278 * while migration was ongoing. 3279 * If unmap finds the old page, uncharge() of it will be delayed 3280 * until end_migration(). If unmap finds a new page, it's 3281 * uncharged when it make mapcount to be 1->0. If unmap code 3282 * finds swap_migration_entry, the new page will not be mapped 3283 * and end_migration() will find it(mapcount==0). 3284 * 3285 * B) 3286 * When the old page was mapped but migraion fails, the kernel 3287 * remaps it. A charge for it is kept by MIGRATION flag even 3288 * if mapcount goes down to 0. We can do remap successfully 3289 * without charging it again. 3290 * 3291 * C) 3292 * The "old" page is under lock_page() until the end of 3293 * migration, so, the old page itself will not be swapped-out. 3294 * If the new page is swapped out before end_migraton, our 3295 * hook to usual swap-out path will catch the event. 3296 */ 3297 if (PageAnon(page)) 3298 SetPageCgroupMigration(pc); 3299 } 3300 unlock_page_cgroup(pc); 3301 /* 3302 * If the page is not charged at this point, 3303 * we return here. 3304 */ 3305 if (!memcg) 3306 return 0; 3307 3308 *memcgp = memcg; 3309 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); 3310 css_put(&memcg->css);/* drop extra refcnt */ 3311 if (ret || *memcgp == NULL) { 3312 if (PageAnon(page)) { 3313 lock_page_cgroup(pc); 3314 ClearPageCgroupMigration(pc); 3315 unlock_page_cgroup(pc); 3316 /* 3317 * The old page may be fully unmapped while we kept it. 3318 */ 3319 mem_cgroup_uncharge_page(page); 3320 } 3321 return -ENOMEM; 3322 } 3323 /* 3324 * We charge new page before it's used/mapped. So, even if unlock_page() 3325 * is called before end_migration, we can catch all events on this new 3326 * page. In the case new page is migrated but not remapped, new page's 3327 * mapcount will be finally 0 and we call uncharge in end_migration(). 3328 */ 3329 pc = lookup_page_cgroup(newpage); 3330 if (PageAnon(page)) 3331 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 3332 else if (page_is_file_cache(page)) 3333 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 3334 else 3335 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3336 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 3337 return ret; 3338} 3339 3340/* remove redundant charge if migration failed*/ 3341void mem_cgroup_end_migration(struct mem_cgroup *memcg, 3342 struct page *oldpage, struct page *newpage, bool migration_ok) 3343{ 3344 struct page *used, *unused; 3345 struct page_cgroup *pc; 3346 3347 if (!memcg) 3348 return; 3349 /* blocks rmdir() */ 3350 cgroup_exclude_rmdir(&memcg->css); 3351 if (!migration_ok) { 3352 used = oldpage; 3353 unused = newpage; 3354 } else { 3355 used = newpage; 3356 unused = oldpage; 3357 } 3358 /* 3359 * We disallowed uncharge of pages under migration because mapcount 3360 * of the page goes down to zero, temporarly. 3361 * Clear the flag and check the page should be charged. 3362 */ 3363 pc = lookup_page_cgroup(oldpage); 3364 lock_page_cgroup(pc); 3365 ClearPageCgroupMigration(pc); 3366 unlock_page_cgroup(pc); 3367 3368 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 3369 3370 /* 3371 * If a page is a file cache, radix-tree replacement is very atomic 3372 * and we can skip this check. When it was an Anon page, its mapcount 3373 * goes down to 0. But because we added MIGRATION flage, it's not 3374 * uncharged yet. There are several case but page->mapcount check 3375 * and USED bit check in mem_cgroup_uncharge_page() will do enough 3376 * check. (see prepare_charge() also) 3377 */ 3378 if (PageAnon(used)) 3379 mem_cgroup_uncharge_page(used); 3380 /* 3381 * At migration, we may charge account against cgroup which has no 3382 * tasks. 3383 * So, rmdir()->pre_destroy() can be called while we do this charge. 3384 * In that case, we need to call pre_destroy() again. check it here. 3385 */ 3386 cgroup_release_and_wakeup_rmdir(&memcg->css); 3387} 3388 3389/* 3390 * At replace page cache, newpage is not under any memcg but it's on 3391 * LRU. So, this function doesn't touch res_counter but handles LRU 3392 * in correct way. Both pages are locked so we cannot race with uncharge. 3393 */ 3394void mem_cgroup_replace_page_cache(struct page *oldpage, 3395 struct page *newpage) 3396{ 3397 struct mem_cgroup *memcg; 3398 struct page_cgroup *pc; 3399 struct zone *zone; 3400 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; 3401 unsigned long flags; 3402 3403 if (mem_cgroup_disabled()) 3404 return; 3405 3406 pc = lookup_page_cgroup(oldpage); 3407 /* fix accounting on old pages */ 3408 lock_page_cgroup(pc); 3409 memcg = pc->mem_cgroup; 3410 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1); 3411 ClearPageCgroupUsed(pc); 3412 unlock_page_cgroup(pc); 3413 3414 if (PageSwapBacked(oldpage)) 3415 type = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3416 3417 zone = page_zone(newpage); 3418 pc = lookup_page_cgroup(newpage); 3419 /* 3420 * Even if newpage->mapping was NULL before starting replacement, 3421 * the newpage may be on LRU(or pagevec for LRU) already. We lock 3422 * LRU while we overwrite pc->mem_cgroup. 3423 */ 3424 spin_lock_irqsave(&zone->lru_lock, flags); 3425 if (PageLRU(newpage)) 3426 del_page_from_lru_list(zone, newpage, page_lru(newpage)); 3427 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type); 3428 if (PageLRU(newpage)) 3429 add_page_to_lru_list(zone, newpage, page_lru(newpage)); 3430 spin_unlock_irqrestore(&zone->lru_lock, flags); 3431} 3432 3433#ifdef CONFIG_DEBUG_VM 3434static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3435{ 3436 struct page_cgroup *pc; 3437 3438 pc = lookup_page_cgroup(page); 3439 if (likely(pc) && PageCgroupUsed(pc)) 3440 return pc; 3441 return NULL; 3442} 3443 3444bool mem_cgroup_bad_page_check(struct page *page) 3445{ 3446 if (mem_cgroup_disabled()) 3447 return false; 3448 3449 return lookup_page_cgroup_used(page) != NULL; 3450} 3451 3452void mem_cgroup_print_bad_page(struct page *page) 3453{ 3454 struct page_cgroup *pc; 3455 3456 pc = lookup_page_cgroup_used(page); 3457 if (pc) { 3458 int ret = -1; 3459 char *path; 3460 3461 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", 3462 pc, pc->flags, pc->mem_cgroup); 3463 3464 path = kmalloc(PATH_MAX, GFP_KERNEL); 3465 if (path) { 3466 rcu_read_lock(); 3467 ret = cgroup_path(pc->mem_cgroup->css.cgroup, 3468 path, PATH_MAX); 3469 rcu_read_unlock(); 3470 } 3471 3472 printk(KERN_CONT "(%s)\n", 3473 (ret < 0) ? "cannot get the path" : path); 3474 kfree(path); 3475 } 3476} 3477#endif 3478 3479static DEFINE_MUTEX(set_limit_mutex); 3480 3481static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3482 unsigned long long val) 3483{ 3484 int retry_count; 3485 u64 memswlimit, memlimit; 3486 int ret = 0; 3487 int children = mem_cgroup_count_children(memcg); 3488 u64 curusage, oldusage; 3489 int enlarge; 3490 3491 /* 3492 * For keeping hierarchical_reclaim simple, how long we should retry 3493 * is depends on callers. We set our retry-count to be function 3494 * of # of children which we should visit in this loop. 3495 */ 3496 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3497 3498 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3499 3500 enlarge = 0; 3501 while (retry_count) { 3502 if (signal_pending(current)) { 3503 ret = -EINTR; 3504 break; 3505 } 3506 /* 3507 * Rather than hide all in some function, I do this in 3508 * open coded manner. You see what this really does. 3509 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3510 */ 3511 mutex_lock(&set_limit_mutex); 3512 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3513 if (memswlimit < val) { 3514 ret = -EINVAL; 3515 mutex_unlock(&set_limit_mutex); 3516 break; 3517 } 3518 3519 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3520 if (memlimit < val) 3521 enlarge = 1; 3522 3523 ret = res_counter_set_limit(&memcg->res, val); 3524 if (!ret) { 3525 if (memswlimit == val) 3526 memcg->memsw_is_minimum = true; 3527 else 3528 memcg->memsw_is_minimum = false; 3529 } 3530 mutex_unlock(&set_limit_mutex); 3531 3532 if (!ret) 3533 break; 3534 3535 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3536 MEM_CGROUP_RECLAIM_SHRINK); 3537 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3538 /* Usage is reduced ? */ 3539 if (curusage >= oldusage) 3540 retry_count--; 3541 else 3542 oldusage = curusage; 3543 } 3544 if (!ret && enlarge) 3545 memcg_oom_recover(memcg); 3546 3547 return ret; 3548} 3549 3550static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3551 unsigned long long val) 3552{ 3553 int retry_count; 3554 u64 memlimit, memswlimit, oldusage, curusage; 3555 int children = mem_cgroup_count_children(memcg); 3556 int ret = -EBUSY; 3557 int enlarge = 0; 3558 3559 /* see mem_cgroup_resize_res_limit */ 3560 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3561 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3562 while (retry_count) { 3563 if (signal_pending(current)) { 3564 ret = -EINTR; 3565 break; 3566 } 3567 /* 3568 * Rather than hide all in some function, I do this in 3569 * open coded manner. You see what this really does. 3570 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3571 */ 3572 mutex_lock(&set_limit_mutex); 3573 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3574 if (memlimit > val) { 3575 ret = -EINVAL; 3576 mutex_unlock(&set_limit_mutex); 3577 break; 3578 } 3579 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3580 if (memswlimit < val) 3581 enlarge = 1; 3582 ret = res_counter_set_limit(&memcg->memsw, val); 3583 if (!ret) { 3584 if (memlimit == val) 3585 memcg->memsw_is_minimum = true; 3586 else 3587 memcg->memsw_is_minimum = false; 3588 } 3589 mutex_unlock(&set_limit_mutex); 3590 3591 if (!ret) 3592 break; 3593 3594 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3595 MEM_CGROUP_RECLAIM_NOSWAP | 3596 MEM_CGROUP_RECLAIM_SHRINK); 3597 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3598 /* Usage is reduced ? */ 3599 if (curusage >= oldusage) 3600 retry_count--; 3601 else 3602 oldusage = curusage; 3603 } 3604 if (!ret && enlarge) 3605 memcg_oom_recover(memcg); 3606 return ret; 3607} 3608 3609unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3610 gfp_t gfp_mask, 3611 unsigned long *total_scanned) 3612{ 3613 unsigned long nr_reclaimed = 0; 3614 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3615 unsigned long reclaimed; 3616 int loop = 0; 3617 struct mem_cgroup_tree_per_zone *mctz; 3618 unsigned long long excess; 3619 unsigned long nr_scanned; 3620 3621 if (order > 0) 3622 return 0; 3623 3624 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3625 /* 3626 * This loop can run a while, specially if mem_cgroup's continuously 3627 * keep exceeding their soft limit and putting the system under 3628 * pressure 3629 */ 3630 do { 3631 if (next_mz) 3632 mz = next_mz; 3633 else 3634 mz = mem_cgroup_largest_soft_limit_node(mctz); 3635 if (!mz) 3636 break; 3637 3638 nr_scanned = 0; 3639 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone, 3640 gfp_mask, &nr_scanned); 3641 nr_reclaimed += reclaimed; 3642 *total_scanned += nr_scanned; 3643 spin_lock(&mctz->lock); 3644 3645 /* 3646 * If we failed to reclaim anything from this memory cgroup 3647 * it is time to move on to the next cgroup 3648 */ 3649 next_mz = NULL; 3650 if (!reclaimed) { 3651 do { 3652 /* 3653 * Loop until we find yet another one. 3654 * 3655 * By the time we get the soft_limit lock 3656 * again, someone might have aded the 3657 * group back on the RB tree. Iterate to 3658 * make sure we get a different mem. 3659 * mem_cgroup_largest_soft_limit_node returns 3660 * NULL if no other cgroup is present on 3661 * the tree 3662 */ 3663 next_mz = 3664 __mem_cgroup_largest_soft_limit_node(mctz); 3665 if (next_mz == mz) 3666 css_put(&next_mz->mem->css); 3667 else /* next_mz == NULL or other memcg */ 3668 break; 3669 } while (1); 3670 } 3671 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3672 excess = res_counter_soft_limit_excess(&mz->mem->res); 3673 /* 3674 * One school of thought says that we should not add 3675 * back the node to the tree if reclaim returns 0. 3676 * But our reclaim could return 0, simply because due 3677 * to priority we are exposing a smaller subset of 3678 * memory to reclaim from. Consider this as a longer 3679 * term TODO. 3680 */ 3681 /* If excess == 0, no tree ops */ 3682 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3683 spin_unlock(&mctz->lock); 3684 css_put(&mz->mem->css); 3685 loop++; 3686 /* 3687 * Could not reclaim anything and there are no more 3688 * mem cgroups to try or we seem to be looping without 3689 * reclaiming anything. 3690 */ 3691 if (!nr_reclaimed && 3692 (next_mz == NULL || 3693 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3694 break; 3695 } while (!nr_reclaimed); 3696 if (next_mz) 3697 css_put(&next_mz->mem->css); 3698 return nr_reclaimed; 3699} 3700 3701/* 3702 * This routine traverse page_cgroup in given list and drop them all. 3703 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3704 */ 3705static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, 3706 int node, int zid, enum lru_list lru) 3707{ 3708 struct mem_cgroup_per_zone *mz; 3709 unsigned long flags, loop; 3710 struct list_head *list; 3711 struct page *busy; 3712 struct zone *zone; 3713 int ret = 0; 3714 3715 zone = &NODE_DATA(node)->node_zones[zid]; 3716 mz = mem_cgroup_zoneinfo(memcg, node, zid); 3717 list = &mz->lruvec.lists[lru]; 3718 3719 loop = MEM_CGROUP_ZSTAT(mz, lru); 3720 /* give some margin against EBUSY etc...*/ 3721 loop += 256; 3722 busy = NULL; 3723 while (loop--) { 3724 struct page_cgroup *pc; 3725 struct page *page; 3726 3727 ret = 0; 3728 spin_lock_irqsave(&zone->lru_lock, flags); 3729 if (list_empty(list)) { 3730 spin_unlock_irqrestore(&zone->lru_lock, flags); 3731 break; 3732 } 3733 page = list_entry(list->prev, struct page, lru); 3734 if (busy == page) { 3735 list_move(&page->lru, list); 3736 busy = NULL; 3737 spin_unlock_irqrestore(&zone->lru_lock, flags); 3738 continue; 3739 } 3740 spin_unlock_irqrestore(&zone->lru_lock, flags); 3741 3742 pc = lookup_page_cgroup(page); 3743 3744 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); 3745 if (ret == -ENOMEM) 3746 break; 3747 3748 if (ret == -EBUSY || ret == -EINVAL) { 3749 /* found lock contention or "pc" is obsolete. */ 3750 busy = page; 3751 cond_resched(); 3752 } else 3753 busy = NULL; 3754 } 3755 3756 if (!ret && !list_empty(list)) 3757 return -EBUSY; 3758 return ret; 3759} 3760 3761/* 3762 * make mem_cgroup's charge to be 0 if there is no task. 3763 * This enables deleting this mem_cgroup. 3764 */ 3765static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) 3766{ 3767 int ret; 3768 int node, zid, shrink; 3769 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3770 struct cgroup *cgrp = memcg->css.cgroup; 3771 3772 css_get(&memcg->css); 3773 3774 shrink = 0; 3775 /* should free all ? */ 3776 if (free_all) 3777 goto try_to_free; 3778move_account: 3779 do { 3780 ret = -EBUSY; 3781 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3782 goto out; 3783 ret = -EINTR; 3784 if (signal_pending(current)) 3785 goto out; 3786 /* This is for making all *used* pages to be on LRU. */ 3787 lru_add_drain_all(); 3788 drain_all_stock_sync(memcg); 3789 ret = 0; 3790 mem_cgroup_start_move(memcg); 3791 for_each_node_state(node, N_HIGH_MEMORY) { 3792 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3793 enum lru_list l; 3794 for_each_lru(l) { 3795 ret = mem_cgroup_force_empty_list(memcg, 3796 node, zid, l); 3797 if (ret) 3798 break; 3799 } 3800 } 3801 if (ret) 3802 break; 3803 } 3804 mem_cgroup_end_move(memcg); 3805 memcg_oom_recover(memcg); 3806 /* it seems parent cgroup doesn't have enough mem */ 3807 if (ret == -ENOMEM) 3808 goto try_to_free; 3809 cond_resched(); 3810 /* "ret" should also be checked to ensure all lists are empty. */ 3811 } while (memcg->res.usage > 0 || ret); 3812out: 3813 css_put(&memcg->css); 3814 return ret; 3815 3816try_to_free: 3817 /* returns EBUSY if there is a task or if we come here twice. */ 3818 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3819 ret = -EBUSY; 3820 goto out; 3821 } 3822 /* we call try-to-free pages for make this cgroup empty */ 3823 lru_add_drain_all(); 3824 /* try to free all pages in this cgroup */ 3825 shrink = 1; 3826 while (nr_retries && memcg->res.usage > 0) { 3827 int progress; 3828 3829 if (signal_pending(current)) { 3830 ret = -EINTR; 3831 goto out; 3832 } 3833 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, 3834 false); 3835 if (!progress) { 3836 nr_retries--; 3837 /* maybe some writeback is necessary */ 3838 congestion_wait(BLK_RW_ASYNC, HZ/10); 3839 } 3840 3841 } 3842 lru_add_drain(); 3843 /* try move_account...there may be some *locked* pages. */ 3844 goto move_account; 3845} 3846 3847int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3848{ 3849 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3850} 3851 3852 3853static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3854{ 3855 return mem_cgroup_from_cont(cont)->use_hierarchy; 3856} 3857 3858static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3859 u64 val) 3860{ 3861 int retval = 0; 3862 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3863 struct cgroup *parent = cont->parent; 3864 struct mem_cgroup *parent_memcg = NULL; 3865 3866 if (parent) 3867 parent_memcg = mem_cgroup_from_cont(parent); 3868 3869 cgroup_lock(); 3870 /* 3871 * If parent's use_hierarchy is set, we can't make any modifications 3872 * in the child subtrees. If it is unset, then the change can 3873 * occur, provided the current cgroup has no children. 3874 * 3875 * For the root cgroup, parent_mem is NULL, we allow value to be 3876 * set if there are no children. 3877 */ 3878 if ((!parent_memcg || !parent_memcg->use_hierarchy) && 3879 (val == 1 || val == 0)) { 3880 if (list_empty(&cont->children)) 3881 memcg->use_hierarchy = val; 3882 else 3883 retval = -EBUSY; 3884 } else 3885 retval = -EINVAL; 3886 cgroup_unlock(); 3887 3888 return retval; 3889} 3890 3891 3892static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, 3893 enum mem_cgroup_stat_index idx) 3894{ 3895 struct mem_cgroup *iter; 3896 long val = 0; 3897 3898 /* Per-cpu values can be negative, use a signed accumulator */ 3899 for_each_mem_cgroup_tree(iter, memcg) 3900 val += mem_cgroup_read_stat(iter, idx); 3901 3902 if (val < 0) /* race ? */ 3903 val = 0; 3904 return val; 3905} 3906 3907static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) 3908{ 3909 u64 val; 3910 3911 if (!mem_cgroup_is_root(memcg)) { 3912 if (!swap) 3913 return res_counter_read_u64(&memcg->res, RES_USAGE); 3914 else 3915 return res_counter_read_u64(&memcg->memsw, RES_USAGE); 3916 } 3917 3918 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); 3919 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); 3920 3921 if (swap) 3922 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 3923 3924 return val << PAGE_SHIFT; 3925} 3926 3927static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3928{ 3929 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3930 u64 val; 3931 int type, name; 3932 3933 type = MEMFILE_TYPE(cft->private); 3934 name = MEMFILE_ATTR(cft->private); 3935 switch (type) { 3936 case _MEM: 3937 if (name == RES_USAGE) 3938 val = mem_cgroup_usage(memcg, false); 3939 else 3940 val = res_counter_read_u64(&memcg->res, name); 3941 break; 3942 case _MEMSWAP: 3943 if (name == RES_USAGE) 3944 val = mem_cgroup_usage(memcg, true); 3945 else 3946 val = res_counter_read_u64(&memcg->memsw, name); 3947 break; 3948 default: 3949 BUG(); 3950 break; 3951 } 3952 return val; 3953} 3954/* 3955 * The user of this function is... 3956 * RES_LIMIT. 3957 */ 3958static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3959 const char *buffer) 3960{ 3961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3962 int type, name; 3963 unsigned long long val; 3964 int ret; 3965 3966 type = MEMFILE_TYPE(cft->private); 3967 name = MEMFILE_ATTR(cft->private); 3968 switch (name) { 3969 case RES_LIMIT: 3970 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3971 ret = -EINVAL; 3972 break; 3973 } 3974 /* This function does all necessary parse...reuse it */ 3975 ret = res_counter_memparse_write_strategy(buffer, &val); 3976 if (ret) 3977 break; 3978 if (type == _MEM) 3979 ret = mem_cgroup_resize_limit(memcg, val); 3980 else 3981 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3982 break; 3983 case RES_SOFT_LIMIT: 3984 ret = res_counter_memparse_write_strategy(buffer, &val); 3985 if (ret) 3986 break; 3987 /* 3988 * For memsw, soft limits are hard to implement in terms 3989 * of semantics, for now, we support soft limits for 3990 * control without swap 3991 */ 3992 if (type == _MEM) 3993 ret = res_counter_set_soft_limit(&memcg->res, val); 3994 else 3995 ret = -EINVAL; 3996 break; 3997 default: 3998 ret = -EINVAL; /* should be BUG() ? */ 3999 break; 4000 } 4001 return ret; 4002} 4003 4004static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 4005 unsigned long long *mem_limit, unsigned long long *memsw_limit) 4006{ 4007 struct cgroup *cgroup; 4008 unsigned long long min_limit, min_memsw_limit, tmp; 4009 4010 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 4011 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4012 cgroup = memcg->css.cgroup; 4013 if (!memcg->use_hierarchy) 4014 goto out; 4015 4016 while (cgroup->parent) { 4017 cgroup = cgroup->parent; 4018 memcg = mem_cgroup_from_cont(cgroup); 4019 if (!memcg->use_hierarchy) 4020 break; 4021 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 4022 min_limit = min(min_limit, tmp); 4023 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4024 min_memsw_limit = min(min_memsw_limit, tmp); 4025 } 4026out: 4027 *mem_limit = min_limit; 4028 *memsw_limit = min_memsw_limit; 4029 return; 4030} 4031 4032static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 4033{ 4034 struct mem_cgroup *memcg; 4035 int type, name; 4036 4037 memcg = mem_cgroup_from_cont(cont); 4038 type = MEMFILE_TYPE(event); 4039 name = MEMFILE_ATTR(event); 4040 switch (name) { 4041 case RES_MAX_USAGE: 4042 if (type == _MEM) 4043 res_counter_reset_max(&memcg->res); 4044 else 4045 res_counter_reset_max(&memcg->memsw); 4046 break; 4047 case RES_FAILCNT: 4048 if (type == _MEM) 4049 res_counter_reset_failcnt(&memcg->res); 4050 else 4051 res_counter_reset_failcnt(&memcg->memsw); 4052 break; 4053 } 4054 4055 return 0; 4056} 4057 4058static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 4059 struct cftype *cft) 4060{ 4061 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 4062} 4063 4064#ifdef CONFIG_MMU 4065static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4066 struct cftype *cft, u64 val) 4067{ 4068 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4069 4070 if (val >= (1 << NR_MOVE_TYPE)) 4071 return -EINVAL; 4072 /* 4073 * We check this value several times in both in can_attach() and 4074 * attach(), so we need cgroup lock to prevent this value from being 4075 * inconsistent. 4076 */ 4077 cgroup_lock(); 4078 memcg->move_charge_at_immigrate = val; 4079 cgroup_unlock(); 4080 4081 return 0; 4082} 4083#else 4084static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4085 struct cftype *cft, u64 val) 4086{ 4087 return -ENOSYS; 4088} 4089#endif 4090 4091 4092/* For read statistics */ 4093enum { 4094 MCS_CACHE, 4095 MCS_RSS, 4096 MCS_FILE_MAPPED, 4097 MCS_PGPGIN, 4098 MCS_PGPGOUT, 4099 MCS_SWAP, 4100 MCS_PGFAULT, 4101 MCS_PGMAJFAULT, 4102 MCS_INACTIVE_ANON, 4103 MCS_ACTIVE_ANON, 4104 MCS_INACTIVE_FILE, 4105 MCS_ACTIVE_FILE, 4106 MCS_UNEVICTABLE, 4107 NR_MCS_STAT, 4108}; 4109 4110struct mcs_total_stat { 4111 s64 stat[NR_MCS_STAT]; 4112}; 4113 4114struct { 4115 char *local_name; 4116 char *total_name; 4117} memcg_stat_strings[NR_MCS_STAT] = { 4118 {"cache", "total_cache"}, 4119 {"rss", "total_rss"}, 4120 {"mapped_file", "total_mapped_file"}, 4121 {"pgpgin", "total_pgpgin"}, 4122 {"pgpgout", "total_pgpgout"}, 4123 {"swap", "total_swap"}, 4124 {"pgfault", "total_pgfault"}, 4125 {"pgmajfault", "total_pgmajfault"}, 4126 {"inactive_anon", "total_inactive_anon"}, 4127 {"active_anon", "total_active_anon"}, 4128 {"inactive_file", "total_inactive_file"}, 4129 {"active_file", "total_active_file"}, 4130 {"unevictable", "total_unevictable"} 4131}; 4132 4133 4134static void 4135mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4136{ 4137 s64 val; 4138 4139 /* per cpu stat */ 4140 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); 4141 s->stat[MCS_CACHE] += val * PAGE_SIZE; 4142 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); 4143 s->stat[MCS_RSS] += val * PAGE_SIZE; 4144 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); 4145 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 4146 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); 4147 s->stat[MCS_PGPGIN] += val; 4148 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); 4149 s->stat[MCS_PGPGOUT] += val; 4150 if (do_swap_account) { 4151 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 4152 s->stat[MCS_SWAP] += val * PAGE_SIZE; 4153 } 4154 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); 4155 s->stat[MCS_PGFAULT] += val; 4156 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); 4157 s->stat[MCS_PGMAJFAULT] += val; 4158 4159 /* per zone stat */ 4160 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); 4161 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 4162 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); 4163 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 4164 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); 4165 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 4166 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); 4167 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 4168 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); 4169 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 4170} 4171 4172static void 4173mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4174{ 4175 struct mem_cgroup *iter; 4176 4177 for_each_mem_cgroup_tree(iter, memcg) 4178 mem_cgroup_get_local_stat(iter, s); 4179} 4180 4181#ifdef CONFIG_NUMA 4182static int mem_control_numa_stat_show(struct seq_file *m, void *arg) 4183{ 4184 int nid; 4185 unsigned long total_nr, file_nr, anon_nr, unevictable_nr; 4186 unsigned long node_nr; 4187 struct cgroup *cont = m->private; 4188 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4189 4190 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL); 4191 seq_printf(m, "total=%lu", total_nr); 4192 for_each_node_state(nid, N_HIGH_MEMORY) { 4193 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL); 4194 seq_printf(m, " N%d=%lu", nid, node_nr); 4195 } 4196 seq_putc(m, '\n'); 4197 4198 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE); 4199 seq_printf(m, "file=%lu", file_nr); 4200 for_each_node_state(nid, N_HIGH_MEMORY) { 4201 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4202 LRU_ALL_FILE); 4203 seq_printf(m, " N%d=%lu", nid, node_nr); 4204 } 4205 seq_putc(m, '\n'); 4206 4207 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON); 4208 seq_printf(m, "anon=%lu", anon_nr); 4209 for_each_node_state(nid, N_HIGH_MEMORY) { 4210 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4211 LRU_ALL_ANON); 4212 seq_printf(m, " N%d=%lu", nid, node_nr); 4213 } 4214 seq_putc(m, '\n'); 4215 4216 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE)); 4217 seq_printf(m, "unevictable=%lu", unevictable_nr); 4218 for_each_node_state(nid, N_HIGH_MEMORY) { 4219 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4220 BIT(LRU_UNEVICTABLE)); 4221 seq_printf(m, " N%d=%lu", nid, node_nr); 4222 } 4223 seq_putc(m, '\n'); 4224 return 0; 4225} 4226#endif /* CONFIG_NUMA */ 4227 4228static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 4229 struct cgroup_map_cb *cb) 4230{ 4231 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4232 struct mcs_total_stat mystat; 4233 int i; 4234 4235 memset(&mystat, 0, sizeof(mystat)); 4236 mem_cgroup_get_local_stat(mem_cont, &mystat); 4237 4238 4239 for (i = 0; i < NR_MCS_STAT; i++) { 4240 if (i == MCS_SWAP && !do_swap_account) 4241 continue; 4242 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 4243 } 4244 4245 /* Hierarchical information */ 4246 { 4247 unsigned long long limit, memsw_limit; 4248 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 4249 cb->fill(cb, "hierarchical_memory_limit", limit); 4250 if (do_swap_account) 4251 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 4252 } 4253 4254 memset(&mystat, 0, sizeof(mystat)); 4255 mem_cgroup_get_total_stat(mem_cont, &mystat); 4256 for (i = 0; i < NR_MCS_STAT; i++) { 4257 if (i == MCS_SWAP && !do_swap_account) 4258 continue; 4259 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 4260 } 4261 4262#ifdef CONFIG_DEBUG_VM 4263 { 4264 int nid, zid; 4265 struct mem_cgroup_per_zone *mz; 4266 unsigned long recent_rotated[2] = {0, 0}; 4267 unsigned long recent_scanned[2] = {0, 0}; 4268 4269 for_each_online_node(nid) 4270 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 4271 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 4272 4273 recent_rotated[0] += 4274 mz->reclaim_stat.recent_rotated[0]; 4275 recent_rotated[1] += 4276 mz->reclaim_stat.recent_rotated[1]; 4277 recent_scanned[0] += 4278 mz->reclaim_stat.recent_scanned[0]; 4279 recent_scanned[1] += 4280 mz->reclaim_stat.recent_scanned[1]; 4281 } 4282 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 4283 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 4284 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 4285 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 4286 } 4287#endif 4288 4289 return 0; 4290} 4291 4292static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 4293{ 4294 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4295 4296 return mem_cgroup_swappiness(memcg); 4297} 4298 4299static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 4300 u64 val) 4301{ 4302 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4303 struct mem_cgroup *parent; 4304 4305 if (val > 100) 4306 return -EINVAL; 4307 4308 if (cgrp->parent == NULL) 4309 return -EINVAL; 4310 4311 parent = mem_cgroup_from_cont(cgrp->parent); 4312 4313 cgroup_lock(); 4314 4315 /* If under hierarchy, only empty-root can set this value */ 4316 if ((parent->use_hierarchy) || 4317 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4318 cgroup_unlock(); 4319 return -EINVAL; 4320 } 4321 4322 memcg->swappiness = val; 4323 4324 cgroup_unlock(); 4325 4326 return 0; 4327} 4328 4329static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 4330{ 4331 struct mem_cgroup_threshold_ary *t; 4332 u64 usage; 4333 int i; 4334 4335 rcu_read_lock(); 4336 if (!swap) 4337 t = rcu_dereference(memcg->thresholds.primary); 4338 else 4339 t = rcu_dereference(memcg->memsw_thresholds.primary); 4340 4341 if (!t) 4342 goto unlock; 4343 4344 usage = mem_cgroup_usage(memcg, swap); 4345 4346 /* 4347 * current_threshold points to threshold just below usage. 4348 * If it's not true, a threshold was crossed after last 4349 * call of __mem_cgroup_threshold(). 4350 */ 4351 i = t->current_threshold; 4352 4353 /* 4354 * Iterate backward over array of thresholds starting from 4355 * current_threshold and check if a threshold is crossed. 4356 * If none of thresholds below usage is crossed, we read 4357 * only one element of the array here. 4358 */ 4359 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 4360 eventfd_signal(t->entries[i].eventfd, 1); 4361 4362 /* i = current_threshold + 1 */ 4363 i++; 4364 4365 /* 4366 * Iterate forward over array of thresholds starting from 4367 * current_threshold+1 and check if a threshold is crossed. 4368 * If none of thresholds above usage is crossed, we read 4369 * only one element of the array here. 4370 */ 4371 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 4372 eventfd_signal(t->entries[i].eventfd, 1); 4373 4374 /* Update current_threshold */ 4375 t->current_threshold = i - 1; 4376unlock: 4377 rcu_read_unlock(); 4378} 4379 4380static void mem_cgroup_threshold(struct mem_cgroup *memcg) 4381{ 4382 while (memcg) { 4383 __mem_cgroup_threshold(memcg, false); 4384 if (do_swap_account) 4385 __mem_cgroup_threshold(memcg, true); 4386 4387 memcg = parent_mem_cgroup(memcg); 4388 } 4389} 4390 4391static int compare_thresholds(const void *a, const void *b) 4392{ 4393 const struct mem_cgroup_threshold *_a = a; 4394 const struct mem_cgroup_threshold *_b = b; 4395 4396 return _a->threshold - _b->threshold; 4397} 4398 4399static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) 4400{ 4401 struct mem_cgroup_eventfd_list *ev; 4402 4403 list_for_each_entry(ev, &memcg->oom_notify, list) 4404 eventfd_signal(ev->eventfd, 1); 4405 return 0; 4406} 4407 4408static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) 4409{ 4410 struct mem_cgroup *iter; 4411 4412 for_each_mem_cgroup_tree(iter, memcg) 4413 mem_cgroup_oom_notify_cb(iter); 4414} 4415 4416static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 4417 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4418{ 4419 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4420 struct mem_cgroup_thresholds *thresholds; 4421 struct mem_cgroup_threshold_ary *new; 4422 int type = MEMFILE_TYPE(cft->private); 4423 u64 threshold, usage; 4424 int i, size, ret; 4425 4426 ret = res_counter_memparse_write_strategy(args, &threshold); 4427 if (ret) 4428 return ret; 4429 4430 mutex_lock(&memcg->thresholds_lock); 4431 4432 if (type == _MEM) 4433 thresholds = &memcg->thresholds; 4434 else if (type == _MEMSWAP) 4435 thresholds = &memcg->memsw_thresholds; 4436 else 4437 BUG(); 4438 4439 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4440 4441 /* Check if a threshold crossed before adding a new one */ 4442 if (thresholds->primary) 4443 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4444 4445 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 4446 4447 /* Allocate memory for new array of thresholds */ 4448 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 4449 GFP_KERNEL); 4450 if (!new) { 4451 ret = -ENOMEM; 4452 goto unlock; 4453 } 4454 new->size = size; 4455 4456 /* Copy thresholds (if any) to new array */ 4457 if (thresholds->primary) { 4458 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 4459 sizeof(struct mem_cgroup_threshold)); 4460 } 4461 4462 /* Add new threshold */ 4463 new->entries[size - 1].eventfd = eventfd; 4464 new->entries[size - 1].threshold = threshold; 4465 4466 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4467 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4468 compare_thresholds, NULL); 4469 4470 /* Find current threshold */ 4471 new->current_threshold = -1; 4472 for (i = 0; i < size; i++) { 4473 if (new->entries[i].threshold < usage) { 4474 /* 4475 * new->current_threshold will not be used until 4476 * rcu_assign_pointer(), so it's safe to increment 4477 * it here. 4478 */ 4479 ++new->current_threshold; 4480 } 4481 } 4482 4483 /* Free old spare buffer and save old primary buffer as spare */ 4484 kfree(thresholds->spare); 4485 thresholds->spare = thresholds->primary; 4486 4487 rcu_assign_pointer(thresholds->primary, new); 4488 4489 /* To be sure that nobody uses thresholds */ 4490 synchronize_rcu(); 4491 4492unlock: 4493 mutex_unlock(&memcg->thresholds_lock); 4494 4495 return ret; 4496} 4497 4498static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4499 struct cftype *cft, struct eventfd_ctx *eventfd) 4500{ 4501 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4502 struct mem_cgroup_thresholds *thresholds; 4503 struct mem_cgroup_threshold_ary *new; 4504 int type = MEMFILE_TYPE(cft->private); 4505 u64 usage; 4506 int i, j, size; 4507 4508 mutex_lock(&memcg->thresholds_lock); 4509 if (type == _MEM) 4510 thresholds = &memcg->thresholds; 4511 else if (type == _MEMSWAP) 4512 thresholds = &memcg->memsw_thresholds; 4513 else 4514 BUG(); 4515 4516 /* 4517 * Something went wrong if we trying to unregister a threshold 4518 * if we don't have thresholds 4519 */ 4520 BUG_ON(!thresholds); 4521 4522 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4523 4524 /* Check if a threshold crossed before removing */ 4525 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4526 4527 /* Calculate new number of threshold */ 4528 size = 0; 4529 for (i = 0; i < thresholds->primary->size; i++) { 4530 if (thresholds->primary->entries[i].eventfd != eventfd) 4531 size++; 4532 } 4533 4534 new = thresholds->spare; 4535 4536 /* Set thresholds array to NULL if we don't have thresholds */ 4537 if (!size) { 4538 kfree(new); 4539 new = NULL; 4540 goto swap_buffers; 4541 } 4542 4543 new->size = size; 4544 4545 /* Copy thresholds and find current threshold */ 4546 new->current_threshold = -1; 4547 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4548 if (thresholds->primary->entries[i].eventfd == eventfd) 4549 continue; 4550 4551 new->entries[j] = thresholds->primary->entries[i]; 4552 if (new->entries[j].threshold < usage) { 4553 /* 4554 * new->current_threshold will not be used 4555 * until rcu_assign_pointer(), so it's safe to increment 4556 * it here. 4557 */ 4558 ++new->current_threshold; 4559 } 4560 j++; 4561 } 4562 4563swap_buffers: 4564 /* Swap primary and spare array */ 4565 thresholds->spare = thresholds->primary; 4566 rcu_assign_pointer(thresholds->primary, new); 4567 4568 /* To be sure that nobody uses thresholds */ 4569 synchronize_rcu(); 4570 4571 mutex_unlock(&memcg->thresholds_lock); 4572} 4573 4574static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4575 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4576{ 4577 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4578 struct mem_cgroup_eventfd_list *event; 4579 int type = MEMFILE_TYPE(cft->private); 4580 4581 BUG_ON(type != _OOM_TYPE); 4582 event = kmalloc(sizeof(*event), GFP_KERNEL); 4583 if (!event) 4584 return -ENOMEM; 4585 4586 spin_lock(&memcg_oom_lock); 4587 4588 event->eventfd = eventfd; 4589 list_add(&event->list, &memcg->oom_notify); 4590 4591 /* already in OOM ? */ 4592 if (atomic_read(&memcg->under_oom)) 4593 eventfd_signal(eventfd, 1); 4594 spin_unlock(&memcg_oom_lock); 4595 4596 return 0; 4597} 4598 4599static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4600 struct cftype *cft, struct eventfd_ctx *eventfd) 4601{ 4602 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4603 struct mem_cgroup_eventfd_list *ev, *tmp; 4604 int type = MEMFILE_TYPE(cft->private); 4605 4606 BUG_ON(type != _OOM_TYPE); 4607 4608 spin_lock(&memcg_oom_lock); 4609 4610 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { 4611 if (ev->eventfd == eventfd) { 4612 list_del(&ev->list); 4613 kfree(ev); 4614 } 4615 } 4616 4617 spin_unlock(&memcg_oom_lock); 4618} 4619 4620static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4621 struct cftype *cft, struct cgroup_map_cb *cb) 4622{ 4623 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4624 4625 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); 4626 4627 if (atomic_read(&memcg->under_oom)) 4628 cb->fill(cb, "under_oom", 1); 4629 else 4630 cb->fill(cb, "under_oom", 0); 4631 return 0; 4632} 4633 4634static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4635 struct cftype *cft, u64 val) 4636{ 4637 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4638 struct mem_cgroup *parent; 4639 4640 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4641 if (!cgrp->parent || !((val == 0) || (val == 1))) 4642 return -EINVAL; 4643 4644 parent = mem_cgroup_from_cont(cgrp->parent); 4645 4646 cgroup_lock(); 4647 /* oom-kill-disable is a flag for subhierarchy. */ 4648 if ((parent->use_hierarchy) || 4649 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4650 cgroup_unlock(); 4651 return -EINVAL; 4652 } 4653 memcg->oom_kill_disable = val; 4654 if (!val) 4655 memcg_oom_recover(memcg); 4656 cgroup_unlock(); 4657 return 0; 4658} 4659 4660#ifdef CONFIG_NUMA 4661static const struct file_operations mem_control_numa_stat_file_operations = { 4662 .read = seq_read, 4663 .llseek = seq_lseek, 4664 .release = single_release, 4665}; 4666 4667static int mem_control_numa_stat_open(struct inode *unused, struct file *file) 4668{ 4669 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; 4670 4671 file->f_op = &mem_control_numa_stat_file_operations; 4672 return single_open(file, mem_control_numa_stat_show, cont); 4673} 4674#endif /* CONFIG_NUMA */ 4675 4676#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 4677static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4678{ 4679 /* 4680 * Part of this would be better living in a separate allocation 4681 * function, leaving us with just the cgroup tree population work. 4682 * We, however, depend on state such as network's proto_list that 4683 * is only initialized after cgroup creation. I found the less 4684 * cumbersome way to deal with it to defer it all to populate time 4685 */ 4686 return mem_cgroup_sockets_init(cont, ss); 4687}; 4688 4689static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4690 struct cgroup *cont) 4691{ 4692 mem_cgroup_sockets_destroy(cont, ss); 4693} 4694#else 4695static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4696{ 4697 return 0; 4698} 4699 4700static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4701 struct cgroup *cont) 4702{ 4703} 4704#endif 4705 4706static struct cftype mem_cgroup_files[] = { 4707 { 4708 .name = "usage_in_bytes", 4709 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4710 .read_u64 = mem_cgroup_read, 4711 .register_event = mem_cgroup_usage_register_event, 4712 .unregister_event = mem_cgroup_usage_unregister_event, 4713 }, 4714 { 4715 .name = "max_usage_in_bytes", 4716 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4717 .trigger = mem_cgroup_reset, 4718 .read_u64 = mem_cgroup_read, 4719 }, 4720 { 4721 .name = "limit_in_bytes", 4722 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4723 .write_string = mem_cgroup_write, 4724 .read_u64 = mem_cgroup_read, 4725 }, 4726 { 4727 .name = "soft_limit_in_bytes", 4728 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4729 .write_string = mem_cgroup_write, 4730 .read_u64 = mem_cgroup_read, 4731 }, 4732 { 4733 .name = "failcnt", 4734 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4735 .trigger = mem_cgroup_reset, 4736 .read_u64 = mem_cgroup_read, 4737 }, 4738 { 4739 .name = "stat", 4740 .read_map = mem_control_stat_show, 4741 }, 4742 { 4743 .name = "force_empty", 4744 .trigger = mem_cgroup_force_empty_write, 4745 }, 4746 { 4747 .name = "use_hierarchy", 4748 .write_u64 = mem_cgroup_hierarchy_write, 4749 .read_u64 = mem_cgroup_hierarchy_read, 4750 }, 4751 { 4752 .name = "swappiness", 4753 .read_u64 = mem_cgroup_swappiness_read, 4754 .write_u64 = mem_cgroup_swappiness_write, 4755 }, 4756 { 4757 .name = "move_charge_at_immigrate", 4758 .read_u64 = mem_cgroup_move_charge_read, 4759 .write_u64 = mem_cgroup_move_charge_write, 4760 }, 4761 { 4762 .name = "oom_control", 4763 .read_map = mem_cgroup_oom_control_read, 4764 .write_u64 = mem_cgroup_oom_control_write, 4765 .register_event = mem_cgroup_oom_register_event, 4766 .unregister_event = mem_cgroup_oom_unregister_event, 4767 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4768 }, 4769#ifdef CONFIG_NUMA 4770 { 4771 .name = "numa_stat", 4772 .open = mem_control_numa_stat_open, 4773 .mode = S_IRUGO, 4774 }, 4775#endif 4776}; 4777 4778#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4779static struct cftype memsw_cgroup_files[] = { 4780 { 4781 .name = "memsw.usage_in_bytes", 4782 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4783 .read_u64 = mem_cgroup_read, 4784 .register_event = mem_cgroup_usage_register_event, 4785 .unregister_event = mem_cgroup_usage_unregister_event, 4786 }, 4787 { 4788 .name = "memsw.max_usage_in_bytes", 4789 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4790 .trigger = mem_cgroup_reset, 4791 .read_u64 = mem_cgroup_read, 4792 }, 4793 { 4794 .name = "memsw.limit_in_bytes", 4795 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4796 .write_string = mem_cgroup_write, 4797 .read_u64 = mem_cgroup_read, 4798 }, 4799 { 4800 .name = "memsw.failcnt", 4801 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4802 .trigger = mem_cgroup_reset, 4803 .read_u64 = mem_cgroup_read, 4804 }, 4805}; 4806 4807static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4808{ 4809 if (!do_swap_account) 4810 return 0; 4811 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4812 ARRAY_SIZE(memsw_cgroup_files)); 4813}; 4814#else 4815static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4816{ 4817 return 0; 4818} 4819#endif 4820 4821static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4822{ 4823 struct mem_cgroup_per_node *pn; 4824 struct mem_cgroup_per_zone *mz; 4825 enum lru_list l; 4826 int zone, tmp = node; 4827 /* 4828 * This routine is called against possible nodes. 4829 * But it's BUG to call kmalloc() against offline node. 4830 * 4831 * TODO: this routine can waste much memory for nodes which will 4832 * never be onlined. It's better to use memory hotplug callback 4833 * function. 4834 */ 4835 if (!node_state(node, N_NORMAL_MEMORY)) 4836 tmp = -1; 4837 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4838 if (!pn) 4839 return 1; 4840 4841 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4842 mz = &pn->zoneinfo[zone]; 4843 for_each_lru(l) 4844 INIT_LIST_HEAD(&mz->lruvec.lists[l]); 4845 mz->usage_in_excess = 0; 4846 mz->on_tree = false; 4847 mz->mem = memcg; 4848 } 4849 memcg->info.nodeinfo[node] = pn; 4850 return 0; 4851} 4852 4853static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4854{ 4855 kfree(memcg->info.nodeinfo[node]); 4856} 4857 4858static struct mem_cgroup *mem_cgroup_alloc(void) 4859{ 4860 struct mem_cgroup *mem; 4861 int size = sizeof(struct mem_cgroup); 4862 4863 /* Can be very big if MAX_NUMNODES is very big */ 4864 if (size < PAGE_SIZE) 4865 mem = kzalloc(size, GFP_KERNEL); 4866 else 4867 mem = vzalloc(size); 4868 4869 if (!mem) 4870 return NULL; 4871 4872 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4873 if (!mem->stat) 4874 goto out_free; 4875 spin_lock_init(&mem->pcp_counter_lock); 4876 return mem; 4877 4878out_free: 4879 if (size < PAGE_SIZE) 4880 kfree(mem); 4881 else 4882 vfree(mem); 4883 return NULL; 4884} 4885 4886/* 4887 * At destroying mem_cgroup, references from swap_cgroup can remain. 4888 * (scanning all at force_empty is too costly...) 4889 * 4890 * Instead of clearing all references at force_empty, we remember 4891 * the number of reference from swap_cgroup and free mem_cgroup when 4892 * it goes down to 0. 4893 * 4894 * Removal of cgroup itself succeeds regardless of refs from swap. 4895 */ 4896 4897static void __mem_cgroup_free(struct mem_cgroup *memcg) 4898{ 4899 int node; 4900 4901 mem_cgroup_remove_from_trees(memcg); 4902 free_css_id(&mem_cgroup_subsys, &memcg->css); 4903 4904 for_each_node_state(node, N_POSSIBLE) 4905 free_mem_cgroup_per_zone_info(memcg, node); 4906 4907 free_percpu(memcg->stat); 4908 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4909 kfree(memcg); 4910 else 4911 vfree(memcg); 4912} 4913 4914static void mem_cgroup_get(struct mem_cgroup *memcg) 4915{ 4916 atomic_inc(&memcg->refcnt); 4917} 4918 4919static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) 4920{ 4921 if (atomic_sub_and_test(count, &memcg->refcnt)) { 4922 struct mem_cgroup *parent = parent_mem_cgroup(memcg); 4923 __mem_cgroup_free(memcg); 4924 if (parent) 4925 mem_cgroup_put(parent); 4926 } 4927} 4928 4929static void mem_cgroup_put(struct mem_cgroup *memcg) 4930{ 4931 __mem_cgroup_put(memcg, 1); 4932} 4933 4934/* 4935 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4936 */ 4937struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) 4938{ 4939 if (!memcg->res.parent) 4940 return NULL; 4941 return mem_cgroup_from_res_counter(memcg->res.parent, res); 4942} 4943EXPORT_SYMBOL(parent_mem_cgroup); 4944 4945#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4946static void __init enable_swap_cgroup(void) 4947{ 4948 if (!mem_cgroup_disabled() && really_do_swap_account) 4949 do_swap_account = 1; 4950} 4951#else 4952static void __init enable_swap_cgroup(void) 4953{ 4954} 4955#endif 4956 4957static int mem_cgroup_soft_limit_tree_init(void) 4958{ 4959 struct mem_cgroup_tree_per_node *rtpn; 4960 struct mem_cgroup_tree_per_zone *rtpz; 4961 int tmp, node, zone; 4962 4963 for_each_node_state(node, N_POSSIBLE) { 4964 tmp = node; 4965 if (!node_state(node, N_NORMAL_MEMORY)) 4966 tmp = -1; 4967 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4968 if (!rtpn) 4969 return 1; 4970 4971 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4972 4973 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4974 rtpz = &rtpn->rb_tree_per_zone[zone]; 4975 rtpz->rb_root = RB_ROOT; 4976 spin_lock_init(&rtpz->lock); 4977 } 4978 } 4979 return 0; 4980} 4981 4982static struct cgroup_subsys_state * __ref 4983mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4984{ 4985 struct mem_cgroup *memcg, *parent; 4986 long error = -ENOMEM; 4987 int node; 4988 4989 memcg = mem_cgroup_alloc(); 4990 if (!memcg) 4991 return ERR_PTR(error); 4992 4993 for_each_node_state(node, N_POSSIBLE) 4994 if (alloc_mem_cgroup_per_zone_info(memcg, node)) 4995 goto free_out; 4996 4997 /* root ? */ 4998 if (cont->parent == NULL) { 4999 int cpu; 5000 enable_swap_cgroup(); 5001 parent = NULL; 5002 if (mem_cgroup_soft_limit_tree_init()) 5003 goto free_out; 5004 root_mem_cgroup = memcg; 5005 for_each_possible_cpu(cpu) { 5006 struct memcg_stock_pcp *stock = 5007 &per_cpu(memcg_stock, cpu); 5008 INIT_WORK(&stock->work, drain_local_stock); 5009 } 5010 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 5011 } else { 5012 parent = mem_cgroup_from_cont(cont->parent); 5013 memcg->use_hierarchy = parent->use_hierarchy; 5014 memcg->oom_kill_disable = parent->oom_kill_disable; 5015 } 5016 5017 if (parent && parent->use_hierarchy) { 5018 res_counter_init(&memcg->res, &parent->res); 5019 res_counter_init(&memcg->memsw, &parent->memsw); 5020 /* 5021 * We increment refcnt of the parent to ensure that we can 5022 * safely access it on res_counter_charge/uncharge. 5023 * This refcnt will be decremented when freeing this 5024 * mem_cgroup(see mem_cgroup_put). 5025 */ 5026 mem_cgroup_get(parent); 5027 } else { 5028 res_counter_init(&memcg->res, NULL); 5029 res_counter_init(&memcg->memsw, NULL); 5030 } 5031 memcg->last_scanned_node = MAX_NUMNODES; 5032 INIT_LIST_HEAD(&memcg->oom_notify); 5033 5034 if (parent) 5035 memcg->swappiness = mem_cgroup_swappiness(parent); 5036 atomic_set(&memcg->refcnt, 1); 5037 memcg->move_charge_at_immigrate = 0; 5038 mutex_init(&memcg->thresholds_lock); 5039 return &memcg->css; 5040free_out: 5041 __mem_cgroup_free(memcg); 5042 return ERR_PTR(error); 5043} 5044 5045static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 5046 struct cgroup *cont) 5047{ 5048 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5049 5050 return mem_cgroup_force_empty(memcg, false); 5051} 5052 5053static void mem_cgroup_destroy(struct cgroup_subsys *ss, 5054 struct cgroup *cont) 5055{ 5056 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5057 5058 kmem_cgroup_destroy(ss, cont); 5059 5060 mem_cgroup_put(memcg); 5061} 5062 5063static int mem_cgroup_populate(struct cgroup_subsys *ss, 5064 struct cgroup *cont) 5065{ 5066 int ret; 5067 5068 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 5069 ARRAY_SIZE(mem_cgroup_files)); 5070 5071 if (!ret) 5072 ret = register_memsw_files(cont, ss); 5073 5074 if (!ret) 5075 ret = register_kmem_files(cont, ss); 5076 5077 return ret; 5078} 5079 5080#ifdef CONFIG_MMU 5081/* Handlers for move charge at task migration. */ 5082#define PRECHARGE_COUNT_AT_ONCE 256 5083static int mem_cgroup_do_precharge(unsigned long count) 5084{ 5085 int ret = 0; 5086 int batch_count = PRECHARGE_COUNT_AT_ONCE; 5087 struct mem_cgroup *memcg = mc.to; 5088 5089 if (mem_cgroup_is_root(memcg)) { 5090 mc.precharge += count; 5091 /* we don't need css_get for root */ 5092 return ret; 5093 } 5094 /* try to charge at once */ 5095 if (count > 1) { 5096 struct res_counter *dummy; 5097 /* 5098 * "memcg" cannot be under rmdir() because we've already checked 5099 * by cgroup_lock_live_cgroup() that it is not removed and we 5100 * are still under the same cgroup_mutex. So we can postpone 5101 * css_get(). 5102 */ 5103 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) 5104 goto one_by_one; 5105 if (do_swap_account && res_counter_charge(&memcg->memsw, 5106 PAGE_SIZE * count, &dummy)) { 5107 res_counter_uncharge(&memcg->res, PAGE_SIZE * count); 5108 goto one_by_one; 5109 } 5110 mc.precharge += count; 5111 return ret; 5112 } 5113one_by_one: 5114 /* fall back to one by one charge */ 5115 while (count--) { 5116 if (signal_pending(current)) { 5117 ret = -EINTR; 5118 break; 5119 } 5120 if (!batch_count--) { 5121 batch_count = PRECHARGE_COUNT_AT_ONCE; 5122 cond_resched(); 5123 } 5124 ret = __mem_cgroup_try_charge(NULL, 5125 GFP_KERNEL, 1, &memcg, false); 5126 if (ret || !memcg) 5127 /* mem_cgroup_clear_mc() will do uncharge later */ 5128 return -ENOMEM; 5129 mc.precharge++; 5130 } 5131 return ret; 5132} 5133 5134/** 5135 * is_target_pte_for_mc - check a pte whether it is valid for move charge 5136 * @vma: the vma the pte to be checked belongs 5137 * @addr: the address corresponding to the pte to be checked 5138 * @ptent: the pte to be checked 5139 * @target: the pointer the target page or swap ent will be stored(can be NULL) 5140 * 5141 * Returns 5142 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 5143 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 5144 * move charge. if @target is not NULL, the page is stored in target->page 5145 * with extra refcnt got(Callers should handle it). 5146 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 5147 * target for charge migration. if @target is not NULL, the entry is stored 5148 * in target->ent. 5149 * 5150 * Called with pte lock held. 5151 */ 5152union mc_target { 5153 struct page *page; 5154 swp_entry_t ent; 5155}; 5156 5157enum mc_target_type { 5158 MC_TARGET_NONE, /* not used */ 5159 MC_TARGET_PAGE, 5160 MC_TARGET_SWAP, 5161}; 5162 5163static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 5164 unsigned long addr, pte_t ptent) 5165{ 5166 struct page *page = vm_normal_page(vma, addr, ptent); 5167 5168 if (!page || !page_mapped(page)) 5169 return NULL; 5170 if (PageAnon(page)) { 5171 /* we don't move shared anon */ 5172 if (!move_anon() || page_mapcount(page) > 2) 5173 return NULL; 5174 } else if (!move_file()) 5175 /* we ignore mapcount for file pages */ 5176 return NULL; 5177 if (!get_page_unless_zero(page)) 5178 return NULL; 5179 5180 return page; 5181} 5182 5183static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 5184 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5185{ 5186 int usage_count; 5187 struct page *page = NULL; 5188 swp_entry_t ent = pte_to_swp_entry(ptent); 5189 5190 if (!move_anon() || non_swap_entry(ent)) 5191 return NULL; 5192 usage_count = mem_cgroup_count_swap_user(ent, &page); 5193 if (usage_count > 1) { /* we don't move shared anon */ 5194 if (page) 5195 put_page(page); 5196 return NULL; 5197 } 5198 if (do_swap_account) 5199 entry->val = ent.val; 5200 5201 return page; 5202} 5203 5204static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 5205 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5206{ 5207 struct page *page = NULL; 5208 struct inode *inode; 5209 struct address_space *mapping; 5210 pgoff_t pgoff; 5211 5212 if (!vma->vm_file) /* anonymous vma */ 5213 return NULL; 5214 if (!move_file()) 5215 return NULL; 5216 5217 inode = vma->vm_file->f_path.dentry->d_inode; 5218 mapping = vma->vm_file->f_mapping; 5219 if (pte_none(ptent)) 5220 pgoff = linear_page_index(vma, addr); 5221 else /* pte_file(ptent) is true */ 5222 pgoff = pte_to_pgoff(ptent); 5223 5224 /* page is moved even if it's not RSS of this task(page-faulted). */ 5225 page = find_get_page(mapping, pgoff); 5226 5227#ifdef CONFIG_SWAP 5228 /* shmem/tmpfs may report page out on swap: account for that too. */ 5229 if (radix_tree_exceptional_entry(page)) { 5230 swp_entry_t swap = radix_to_swp_entry(page); 5231 if (do_swap_account) 5232 *entry = swap; 5233 page = find_get_page(&swapper_space, swap.val); 5234 } 5235#endif 5236 return page; 5237} 5238 5239static int is_target_pte_for_mc(struct vm_area_struct *vma, 5240 unsigned long addr, pte_t ptent, union mc_target *target) 5241{ 5242 struct page *page = NULL; 5243 struct page_cgroup *pc; 5244 int ret = 0; 5245 swp_entry_t ent = { .val = 0 }; 5246 5247 if (pte_present(ptent)) 5248 page = mc_handle_present_pte(vma, addr, ptent); 5249 else if (is_swap_pte(ptent)) 5250 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 5251 else if (pte_none(ptent) || pte_file(ptent)) 5252 page = mc_handle_file_pte(vma, addr, ptent, &ent); 5253 5254 if (!page && !ent.val) 5255 return 0; 5256 if (page) { 5257 pc = lookup_page_cgroup(page); 5258 /* 5259 * Do only loose check w/o page_cgroup lock. 5260 * mem_cgroup_move_account() checks the pc is valid or not under 5261 * the lock. 5262 */ 5263 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 5264 ret = MC_TARGET_PAGE; 5265 if (target) 5266 target->page = page; 5267 } 5268 if (!ret || !target) 5269 put_page(page); 5270 } 5271 /* There is a swap entry and a page doesn't exist or isn't charged */ 5272 if (ent.val && !ret && 5273 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { 5274 ret = MC_TARGET_SWAP; 5275 if (target) 5276 target->ent = ent; 5277 } 5278 return ret; 5279} 5280 5281static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 5282 unsigned long addr, unsigned long end, 5283 struct mm_walk *walk) 5284{ 5285 struct vm_area_struct *vma = walk->private; 5286 pte_t *pte; 5287 spinlock_t *ptl; 5288 5289 split_huge_page_pmd(walk->mm, pmd); 5290 5291 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5292 for (; addr != end; pte++, addr += PAGE_SIZE) 5293 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 5294 mc.precharge++; /* increment precharge temporarily */ 5295 pte_unmap_unlock(pte - 1, ptl); 5296 cond_resched(); 5297 5298 return 0; 5299} 5300 5301static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 5302{ 5303 unsigned long precharge; 5304 struct vm_area_struct *vma; 5305 5306 down_read(&mm->mmap_sem); 5307 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5308 struct mm_walk mem_cgroup_count_precharge_walk = { 5309 .pmd_entry = mem_cgroup_count_precharge_pte_range, 5310 .mm = mm, 5311 .private = vma, 5312 }; 5313 if (is_vm_hugetlb_page(vma)) 5314 continue; 5315 walk_page_range(vma->vm_start, vma->vm_end, 5316 &mem_cgroup_count_precharge_walk); 5317 } 5318 up_read(&mm->mmap_sem); 5319 5320 precharge = mc.precharge; 5321 mc.precharge = 0; 5322 5323 return precharge; 5324} 5325 5326static int mem_cgroup_precharge_mc(struct mm_struct *mm) 5327{ 5328 unsigned long precharge = mem_cgroup_count_precharge(mm); 5329 5330 VM_BUG_ON(mc.moving_task); 5331 mc.moving_task = current; 5332 return mem_cgroup_do_precharge(precharge); 5333} 5334 5335/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 5336static void __mem_cgroup_clear_mc(void) 5337{ 5338 struct mem_cgroup *from = mc.from; 5339 struct mem_cgroup *to = mc.to; 5340 5341 /* we must uncharge all the leftover precharges from mc.to */ 5342 if (mc.precharge) { 5343 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 5344 mc.precharge = 0; 5345 } 5346 /* 5347 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 5348 * we must uncharge here. 5349 */ 5350 if (mc.moved_charge) { 5351 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 5352 mc.moved_charge = 0; 5353 } 5354 /* we must fixup refcnts and charges */ 5355 if (mc.moved_swap) { 5356 /* uncharge swap account from the old cgroup */ 5357 if (!mem_cgroup_is_root(mc.from)) 5358 res_counter_uncharge(&mc.from->memsw, 5359 PAGE_SIZE * mc.moved_swap); 5360 __mem_cgroup_put(mc.from, mc.moved_swap); 5361 5362 if (!mem_cgroup_is_root(mc.to)) { 5363 /* 5364 * we charged both to->res and to->memsw, so we should 5365 * uncharge to->res. 5366 */ 5367 res_counter_uncharge(&mc.to->res, 5368 PAGE_SIZE * mc.moved_swap); 5369 } 5370 /* we've already done mem_cgroup_get(mc.to) */ 5371 mc.moved_swap = 0; 5372 } 5373 memcg_oom_recover(from); 5374 memcg_oom_recover(to); 5375 wake_up_all(&mc.waitq); 5376} 5377 5378static void mem_cgroup_clear_mc(void) 5379{ 5380 struct mem_cgroup *from = mc.from; 5381 5382 /* 5383 * we must clear moving_task before waking up waiters at the end of 5384 * task migration. 5385 */ 5386 mc.moving_task = NULL; 5387 __mem_cgroup_clear_mc(); 5388 spin_lock(&mc.lock); 5389 mc.from = NULL; 5390 mc.to = NULL; 5391 spin_unlock(&mc.lock); 5392 mem_cgroup_end_move(from); 5393} 5394 5395static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5396 struct cgroup *cgroup, 5397 struct cgroup_taskset *tset) 5398{ 5399 struct task_struct *p = cgroup_taskset_first(tset); 5400 int ret = 0; 5401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); 5402 5403 if (memcg->move_charge_at_immigrate) { 5404 struct mm_struct *mm; 5405 struct mem_cgroup *from = mem_cgroup_from_task(p); 5406 5407 VM_BUG_ON(from == memcg); 5408 5409 mm = get_task_mm(p); 5410 if (!mm) 5411 return 0; 5412 /* We move charges only when we move a owner of the mm */ 5413 if (mm->owner == p) { 5414 VM_BUG_ON(mc.from); 5415 VM_BUG_ON(mc.to); 5416 VM_BUG_ON(mc.precharge); 5417 VM_BUG_ON(mc.moved_charge); 5418 VM_BUG_ON(mc.moved_swap); 5419 mem_cgroup_start_move(from); 5420 spin_lock(&mc.lock); 5421 mc.from = from; 5422 mc.to = memcg; 5423 spin_unlock(&mc.lock); 5424 /* We set mc.moving_task later */ 5425 5426 ret = mem_cgroup_precharge_mc(mm); 5427 if (ret) 5428 mem_cgroup_clear_mc(); 5429 } 5430 mmput(mm); 5431 } 5432 return ret; 5433} 5434 5435static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5436 struct cgroup *cgroup, 5437 struct cgroup_taskset *tset) 5438{ 5439 mem_cgroup_clear_mc(); 5440} 5441 5442static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 5443 unsigned long addr, unsigned long end, 5444 struct mm_walk *walk) 5445{ 5446 int ret = 0; 5447 struct vm_area_struct *vma = walk->private; 5448 pte_t *pte; 5449 spinlock_t *ptl; 5450 5451 split_huge_page_pmd(walk->mm, pmd); 5452retry: 5453 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5454 for (; addr != end; addr += PAGE_SIZE) { 5455 pte_t ptent = *(pte++); 5456 union mc_target target; 5457 int type; 5458 struct page *page; 5459 struct page_cgroup *pc; 5460 swp_entry_t ent; 5461 5462 if (!mc.precharge) 5463 break; 5464 5465 type = is_target_pte_for_mc(vma, addr, ptent, &target); 5466 switch (type) { 5467 case MC_TARGET_PAGE: 5468 page = target.page; 5469 if (isolate_lru_page(page)) 5470 goto put; 5471 pc = lookup_page_cgroup(page); 5472 if (!mem_cgroup_move_account(page, 1, pc, 5473 mc.from, mc.to, false)) { 5474 mc.precharge--; 5475 /* we uncharge from mc.from later. */ 5476 mc.moved_charge++; 5477 } 5478 putback_lru_page(page); 5479put: /* is_target_pte_for_mc() gets the page */ 5480 put_page(page); 5481 break; 5482 case MC_TARGET_SWAP: 5483 ent = target.ent; 5484 if (!mem_cgroup_move_swap_account(ent, 5485 mc.from, mc.to, false)) { 5486 mc.precharge--; 5487 /* we fixup refcnts and charges later. */ 5488 mc.moved_swap++; 5489 } 5490 break; 5491 default: 5492 break; 5493 } 5494 } 5495 pte_unmap_unlock(pte - 1, ptl); 5496 cond_resched(); 5497 5498 if (addr != end) { 5499 /* 5500 * We have consumed all precharges we got in can_attach(). 5501 * We try charge one by one, but don't do any additional 5502 * charges to mc.to if we have failed in charge once in attach() 5503 * phase. 5504 */ 5505 ret = mem_cgroup_do_precharge(1); 5506 if (!ret) 5507 goto retry; 5508 } 5509 5510 return ret; 5511} 5512 5513static void mem_cgroup_move_charge(struct mm_struct *mm) 5514{ 5515 struct vm_area_struct *vma; 5516 5517 lru_add_drain_all(); 5518retry: 5519 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 5520 /* 5521 * Someone who are holding the mmap_sem might be waiting in 5522 * waitq. So we cancel all extra charges, wake up all waiters, 5523 * and retry. Because we cancel precharges, we might not be able 5524 * to move enough charges, but moving charge is a best-effort 5525 * feature anyway, so it wouldn't be a big problem. 5526 */ 5527 __mem_cgroup_clear_mc(); 5528 cond_resched(); 5529 goto retry; 5530 } 5531 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5532 int ret; 5533 struct mm_walk mem_cgroup_move_charge_walk = { 5534 .pmd_entry = mem_cgroup_move_charge_pte_range, 5535 .mm = mm, 5536 .private = vma, 5537 }; 5538 if (is_vm_hugetlb_page(vma)) 5539 continue; 5540 ret = walk_page_range(vma->vm_start, vma->vm_end, 5541 &mem_cgroup_move_charge_walk); 5542 if (ret) 5543 /* 5544 * means we have consumed all precharges and failed in 5545 * doing additional charge. Just abandon here. 5546 */ 5547 break; 5548 } 5549 up_read(&mm->mmap_sem); 5550} 5551 5552static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5553 struct cgroup *cont, 5554 struct cgroup_taskset *tset) 5555{ 5556 struct task_struct *p = cgroup_taskset_first(tset); 5557 struct mm_struct *mm = get_task_mm(p); 5558 5559 if (mm) { 5560 if (mc.to) 5561 mem_cgroup_move_charge(mm); 5562 put_swap_token(mm); 5563 mmput(mm); 5564 } 5565 if (mc.to) 5566 mem_cgroup_clear_mc(); 5567} 5568#else /* !CONFIG_MMU */ 5569static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5570 struct cgroup *cgroup, 5571 struct cgroup_taskset *tset) 5572{ 5573 return 0; 5574} 5575static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5576 struct cgroup *cgroup, 5577 struct cgroup_taskset *tset) 5578{ 5579} 5580static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5581 struct cgroup *cont, 5582 struct cgroup_taskset *tset) 5583{ 5584} 5585#endif 5586 5587struct cgroup_subsys mem_cgroup_subsys = { 5588 .name = "memory", 5589 .subsys_id = mem_cgroup_subsys_id, 5590 .create = mem_cgroup_create, 5591 .pre_destroy = mem_cgroup_pre_destroy, 5592 .destroy = mem_cgroup_destroy, 5593 .populate = mem_cgroup_populate, 5594 .can_attach = mem_cgroup_can_attach, 5595 .cancel_attach = mem_cgroup_cancel_attach, 5596 .attach = mem_cgroup_move_task, 5597 .early_init = 0, 5598 .use_id = 1, 5599}; 5600 5601#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5602static int __init enable_swap_account(char *s) 5603{ 5604 /* consider enabled if no parameter or 1 is given */ 5605 if (!strcmp(s, "1")) 5606 really_do_swap_account = 1; 5607 else if (!strcmp(s, "0")) 5608 really_do_swap_account = 0; 5609 return 1; 5610} 5611__setup("swapaccount=", enable_swap_account); 5612 5613#endif 5614