memcontrol.c revision f53d7ce32e13dbd09573b176e6521a04c2c77803
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 658void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val) 659{ 660 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val); 661} 662 663void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val) 664{ 665 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val); 666} 667 668static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, 669 enum mem_cgroup_events_index idx) 670{ 671 unsigned long val = 0; 672 int cpu; 673 674 for_each_online_cpu(cpu) 675 val += per_cpu(memcg->stat->events[idx], cpu); 676#ifdef CONFIG_HOTPLUG_CPU 677 spin_lock(&memcg->pcp_counter_lock); 678 val += memcg->nocpu_base.events[idx]; 679 spin_unlock(&memcg->pcp_counter_lock); 680#endif 681 return val; 682} 683 684static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, 685 bool file, int nr_pages) 686{ 687 preempt_disable(); 688 689 if (file) 690 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], 691 nr_pages); 692 else 693 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], 694 nr_pages); 695 696 /* pagein of a big page is an event. So, ignore page size */ 697 if (nr_pages > 0) 698 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); 699 else { 700 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); 701 nr_pages = -nr_pages; /* for event */ 702 } 703 704 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); 705 706 preempt_enable(); 707} 708 709unsigned long 710mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, 711 unsigned int lru_mask) 712{ 713 struct mem_cgroup_per_zone *mz; 714 enum lru_list l; 715 unsigned long ret = 0; 716 717 mz = mem_cgroup_zoneinfo(memcg, nid, zid); 718 719 for_each_lru(l) { 720 if (BIT(l) & lru_mask) 721 ret += MEM_CGROUP_ZSTAT(mz, l); 722 } 723 return ret; 724} 725 726static unsigned long 727mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, 728 int nid, unsigned int lru_mask) 729{ 730 u64 total = 0; 731 int zid; 732 733 for (zid = 0; zid < MAX_NR_ZONES; zid++) 734 total += mem_cgroup_zone_nr_lru_pages(memcg, 735 nid, zid, lru_mask); 736 737 return total; 738} 739 740static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, 741 unsigned int lru_mask) 742{ 743 int nid; 744 u64 total = 0; 745 746 for_each_node_state(nid, N_HIGH_MEMORY) 747 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); 748 return total; 749} 750 751static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, 752 enum mem_cgroup_events_target target) 753{ 754 unsigned long val, next; 755 756 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]); 757 next = __this_cpu_read(memcg->stat->targets[target]); 758 /* from time_after() in jiffies.h */ 759 if ((long)next - (long)val < 0) { 760 switch (target) { 761 case MEM_CGROUP_TARGET_THRESH: 762 next = val + THRESHOLDS_EVENTS_TARGET; 763 break; 764 case MEM_CGROUP_TARGET_SOFTLIMIT: 765 next = val + SOFTLIMIT_EVENTS_TARGET; 766 break; 767 case MEM_CGROUP_TARGET_NUMAINFO: 768 next = val + NUMAINFO_EVENTS_TARGET; 769 break; 770 default: 771 break; 772 } 773 __this_cpu_write(memcg->stat->targets[target], next); 774 return true; 775 } 776 return false; 777} 778 779/* 780 * Check events in order. 781 * 782 */ 783static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) 784{ 785 preempt_disable(); 786 /* threshold event is triggered in finer grain than soft limit */ 787 if (unlikely(mem_cgroup_event_ratelimit(memcg, 788 MEM_CGROUP_TARGET_THRESH))) { 789 bool do_softlimit, do_numainfo; 790 791 do_softlimit = mem_cgroup_event_ratelimit(memcg, 792 MEM_CGROUP_TARGET_SOFTLIMIT); 793#if MAX_NUMNODES > 1 794 do_numainfo = mem_cgroup_event_ratelimit(memcg, 795 MEM_CGROUP_TARGET_NUMAINFO); 796#endif 797 preempt_enable(); 798 799 mem_cgroup_threshold(memcg); 800 if (unlikely(do_softlimit)) 801 mem_cgroup_update_tree(memcg, page); 802#if MAX_NUMNODES > 1 803 if (unlikely(do_numainfo)) 804 atomic_inc(&memcg->numainfo_events); 805#endif 806 } else 807 preempt_enable(); 808} 809 810struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 811{ 812 return container_of(cgroup_subsys_state(cont, 813 mem_cgroup_subsys_id), struct mem_cgroup, 814 css); 815} 816 817struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 818{ 819 /* 820 * mm_update_next_owner() may clear mm->owner to NULL 821 * if it races with swapoff, page migration, etc. 822 * So this can be called with p == NULL. 823 */ 824 if (unlikely(!p)) 825 return NULL; 826 827 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 828 struct mem_cgroup, css); 829} 830 831struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 832{ 833 struct mem_cgroup *memcg = NULL; 834 835 if (!mm) 836 return NULL; 837 /* 838 * Because we have no locks, mm->owner's may be being moved to other 839 * cgroup. We use css_tryget() here even if this looks 840 * pessimistic (rather than adding locks here). 841 */ 842 rcu_read_lock(); 843 do { 844 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 845 if (unlikely(!memcg)) 846 break; 847 } while (!css_tryget(&memcg->css)); 848 rcu_read_unlock(); 849 return memcg; 850} 851 852/** 853 * mem_cgroup_iter - iterate over memory cgroup hierarchy 854 * @root: hierarchy root 855 * @prev: previously returned memcg, NULL on first invocation 856 * @reclaim: cookie for shared reclaim walks, NULL for full walks 857 * 858 * Returns references to children of the hierarchy below @root, or 859 * @root itself, or %NULL after a full round-trip. 860 * 861 * Caller must pass the return value in @prev on subsequent 862 * invocations for reference counting, or use mem_cgroup_iter_break() 863 * to cancel a hierarchy walk before the round-trip is complete. 864 * 865 * Reclaimers can specify a zone and a priority level in @reclaim to 866 * divide up the memcgs in the hierarchy among all concurrent 867 * reclaimers operating on the same zone and priority. 868 */ 869struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, 870 struct mem_cgroup *prev, 871 struct mem_cgroup_reclaim_cookie *reclaim) 872{ 873 struct mem_cgroup *memcg = NULL; 874 int id = 0; 875 876 if (mem_cgroup_disabled()) 877 return NULL; 878 879 if (!root) 880 root = root_mem_cgroup; 881 882 if (prev && !reclaim) 883 id = css_id(&prev->css); 884 885 if (prev && prev != root) 886 css_put(&prev->css); 887 888 if (!root->use_hierarchy && root != root_mem_cgroup) { 889 if (prev) 890 return NULL; 891 return root; 892 } 893 894 while (!memcg) { 895 struct mem_cgroup_reclaim_iter *uninitialized_var(iter); 896 struct cgroup_subsys_state *css; 897 898 if (reclaim) { 899 int nid = zone_to_nid(reclaim->zone); 900 int zid = zone_idx(reclaim->zone); 901 struct mem_cgroup_per_zone *mz; 902 903 mz = mem_cgroup_zoneinfo(root, nid, zid); 904 iter = &mz->reclaim_iter[reclaim->priority]; 905 if (prev && reclaim->generation != iter->generation) 906 return NULL; 907 id = iter->position; 908 } 909 910 rcu_read_lock(); 911 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id); 912 if (css) { 913 if (css == &root->css || css_tryget(css)) 914 memcg = container_of(css, 915 struct mem_cgroup, css); 916 } else 917 id = 0; 918 rcu_read_unlock(); 919 920 if (reclaim) { 921 iter->position = id; 922 if (!css) 923 iter->generation++; 924 else if (!prev && memcg) 925 reclaim->generation = iter->generation; 926 } 927 928 if (prev && !css) 929 return NULL; 930 } 931 return memcg; 932} 933 934/** 935 * mem_cgroup_iter_break - abort a hierarchy walk prematurely 936 * @root: hierarchy root 937 * @prev: last visited hierarchy member as returned by mem_cgroup_iter() 938 */ 939void mem_cgroup_iter_break(struct mem_cgroup *root, 940 struct mem_cgroup *prev) 941{ 942 if (!root) 943 root = root_mem_cgroup; 944 if (prev && prev != root) 945 css_put(&prev->css); 946} 947 948/* 949 * Iteration constructs for visiting all cgroups (under a tree). If 950 * loops are exited prematurely (break), mem_cgroup_iter_break() must 951 * be used for reference counting. 952 */ 953#define for_each_mem_cgroup_tree(iter, root) \ 954 for (iter = mem_cgroup_iter(root, NULL, NULL); \ 955 iter != NULL; \ 956 iter = mem_cgroup_iter(root, iter, NULL)) 957 958#define for_each_mem_cgroup(iter) \ 959 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ 960 iter != NULL; \ 961 iter = mem_cgroup_iter(NULL, iter, NULL)) 962 963static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) 964{ 965 return (memcg == root_mem_cgroup); 966} 967 968void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) 969{ 970 struct mem_cgroup *memcg; 971 972 if (!mm) 973 return; 974 975 rcu_read_lock(); 976 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 977 if (unlikely(!memcg)) 978 goto out; 979 980 switch (idx) { 981 case PGMAJFAULT: 982 mem_cgroup_pgmajfault(memcg, 1); 983 break; 984 case PGFAULT: 985 mem_cgroup_pgfault(memcg, 1); 986 break; 987 default: 988 BUG(); 989 } 990out: 991 rcu_read_unlock(); 992} 993EXPORT_SYMBOL(mem_cgroup_count_vm_event); 994 995/** 996 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg 997 * @zone: zone of the wanted lruvec 998 * @mem: memcg of the wanted lruvec 999 * 1000 * Returns the lru list vector holding pages for the given @zone and 1001 * @mem. This can be the global zone lruvec, if the memory controller 1002 * is disabled. 1003 */ 1004struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, 1005 struct mem_cgroup *memcg) 1006{ 1007 struct mem_cgroup_per_zone *mz; 1008 1009 if (mem_cgroup_disabled()) 1010 return &zone->lruvec; 1011 1012 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); 1013 return &mz->lruvec; 1014} 1015 1016/* 1017 * Following LRU functions are allowed to be used without PCG_LOCK. 1018 * Operations are called by routine of global LRU independently from memcg. 1019 * What we have to take care of here is validness of pc->mem_cgroup. 1020 * 1021 * Changes to pc->mem_cgroup happens when 1022 * 1. charge 1023 * 2. moving account 1024 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 1025 * It is added to LRU before charge. 1026 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 1027 * When moving account, the page is not on LRU. It's isolated. 1028 */ 1029 1030/** 1031 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec 1032 * @zone: zone of the page 1033 * @page: the page 1034 * @lru: current lru 1035 * 1036 * This function accounts for @page being added to @lru, and returns 1037 * the lruvec for the given @zone and the memcg @page is charged to. 1038 * 1039 * The callsite is then responsible for physically linking the page to 1040 * the returned lruvec->lists[@lru]. 1041 */ 1042struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page, 1043 enum lru_list lru) 1044{ 1045 struct mem_cgroup_per_zone *mz; 1046 struct mem_cgroup *memcg; 1047 struct page_cgroup *pc; 1048 1049 if (mem_cgroup_disabled()) 1050 return &zone->lruvec; 1051 1052 pc = lookup_page_cgroup(page); 1053 VM_BUG_ON(PageCgroupAcctLRU(pc)); 1054 /* 1055 * putback: charge: 1056 * SetPageLRU SetPageCgroupUsed 1057 * smp_mb smp_mb 1058 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1059 * 1060 * Ensure that one of the two sides adds the page to the memcg 1061 * LRU during a race. 1062 */ 1063 smp_mb(); 1064 /* 1065 * If the page is uncharged, it may be freed soon, but it 1066 * could also be swap cache (readahead, swapoff) that needs to 1067 * be reclaimable in the future. root_mem_cgroup will babysit 1068 * it for the time being. 1069 */ 1070 if (PageCgroupUsed(pc)) { 1071 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1072 smp_rmb(); 1073 memcg = pc->mem_cgroup; 1074 SetPageCgroupAcctLRU(pc); 1075 } else 1076 memcg = root_mem_cgroup; 1077 mz = page_cgroup_zoneinfo(memcg, page); 1078 /* compound_order() is stabilized through lru_lock */ 1079 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 1080 return &mz->lruvec; 1081} 1082 1083/** 1084 * mem_cgroup_lru_del_list - account for removing an lru page 1085 * @page: the page 1086 * @lru: target lru 1087 * 1088 * This function accounts for @page being removed from @lru. 1089 * 1090 * The callsite is then responsible for physically unlinking 1091 * @page->lru. 1092 */ 1093void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) 1094{ 1095 struct mem_cgroup_per_zone *mz; 1096 struct mem_cgroup *memcg; 1097 struct page_cgroup *pc; 1098 1099 if (mem_cgroup_disabled()) 1100 return; 1101 1102 pc = lookup_page_cgroup(page); 1103 /* 1104 * root_mem_cgroup babysits uncharged LRU pages, but 1105 * PageCgroupUsed is cleared when the page is about to get 1106 * freed. PageCgroupAcctLRU remembers whether the 1107 * LRU-accounting happened against pc->mem_cgroup or 1108 * root_mem_cgroup. 1109 */ 1110 if (TestClearPageCgroupAcctLRU(pc)) { 1111 VM_BUG_ON(!pc->mem_cgroup); 1112 memcg = pc->mem_cgroup; 1113 } else 1114 memcg = root_mem_cgroup; 1115 mz = page_cgroup_zoneinfo(memcg, page); 1116 /* huge page split is done under lru_lock. so, we have no races. */ 1117 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 1118} 1119 1120void mem_cgroup_lru_del(struct page *page) 1121{ 1122 mem_cgroup_lru_del_list(page, page_lru(page)); 1123} 1124 1125/** 1126 * mem_cgroup_lru_move_lists - account for moving a page between lrus 1127 * @zone: zone of the page 1128 * @page: the page 1129 * @from: current lru 1130 * @to: target lru 1131 * 1132 * This function accounts for @page being moved between the lrus @from 1133 * and @to, and returns the lruvec for the given @zone and the memcg 1134 * @page is charged to. 1135 * 1136 * The callsite is then responsible for physically relinking 1137 * @page->lru to the returned lruvec->lists[@to]. 1138 */ 1139struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, 1140 struct page *page, 1141 enum lru_list from, 1142 enum lru_list to) 1143{ 1144 /* XXX: Optimize this, especially for @from == @to */ 1145 mem_cgroup_lru_del_list(page, from); 1146 return mem_cgroup_lru_add_list(zone, page, to); 1147} 1148 1149/* 1150 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed 1151 * while it's linked to lru because the page may be reused after it's fully 1152 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again. 1153 * It's done under lock_page and expected that zone->lru_lock isnever held. 1154 */ 1155static void mem_cgroup_lru_del_before_commit(struct page *page) 1156{ 1157 enum lru_list lru; 1158 unsigned long flags; 1159 struct zone *zone = page_zone(page); 1160 struct page_cgroup *pc = lookup_page_cgroup(page); 1161 1162 /* 1163 * Doing this check without taking ->lru_lock seems wrong but this 1164 * is safe. Because if page_cgroup's USED bit is unset, the page 1165 * will not be added to any memcg's LRU. If page_cgroup's USED bit is 1166 * set, the commit after this will fail, anyway. 1167 * This all charge/uncharge is done under some mutual execustion. 1168 * So, we don't need to taking care of changes in USED bit. 1169 */ 1170 if (likely(!PageLRU(page))) 1171 return; 1172 1173 spin_lock_irqsave(&zone->lru_lock, flags); 1174 lru = page_lru(page); 1175 /* 1176 * The uncharged page could still be registered to the LRU of 1177 * the stale pc->mem_cgroup. 1178 * 1179 * As pc->mem_cgroup is about to get overwritten, the old LRU 1180 * accounting needs to be taken care of. Let root_mem_cgroup 1181 * babysit the page until the new memcg is responsible for it. 1182 * 1183 * The PCG_USED bit is guarded by lock_page() as the page is 1184 * swapcache/pagecache. 1185 */ 1186 if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) { 1187 del_page_from_lru_list(zone, page, lru); 1188 add_page_to_lru_list(zone, page, lru); 1189 } 1190 spin_unlock_irqrestore(&zone->lru_lock, flags); 1191} 1192 1193static void mem_cgroup_lru_add_after_commit(struct page *page) 1194{ 1195 enum lru_list lru; 1196 unsigned long flags; 1197 struct zone *zone = page_zone(page); 1198 struct page_cgroup *pc = lookup_page_cgroup(page); 1199 /* 1200 * putback: charge: 1201 * SetPageLRU SetPageCgroupUsed 1202 * smp_mb smp_mb 1203 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1204 * 1205 * Ensure that one of the two sides adds the page to the memcg 1206 * LRU during a race. 1207 */ 1208 smp_mb(); 1209 /* taking care of that the page is added to LRU while we commit it */ 1210 if (likely(!PageLRU(page))) 1211 return; 1212 spin_lock_irqsave(&zone->lru_lock, flags); 1213 lru = page_lru(page); 1214 /* 1215 * If the page is not on the LRU, someone will soon put it 1216 * there. If it is, and also already accounted for on the 1217 * memcg-side, it must be on the right lruvec as setting 1218 * pc->mem_cgroup and PageCgroupUsed is properly ordered. 1219 * Otherwise, root_mem_cgroup has been babysitting the page 1220 * during the charge. Move it to the new memcg now. 1221 */ 1222 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) { 1223 del_page_from_lru_list(zone, page, lru); 1224 add_page_to_lru_list(zone, page, lru); 1225 } 1226 spin_unlock_irqrestore(&zone->lru_lock, flags); 1227} 1228 1229/* 1230 * Checks whether given mem is same or in the root_mem_cgroup's 1231 * hierarchy subtree 1232 */ 1233static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, 1234 struct mem_cgroup *memcg) 1235{ 1236 if (root_memcg != memcg) { 1237 return (root_memcg->use_hierarchy && 1238 css_is_ancestor(&memcg->css, &root_memcg->css)); 1239 } 1240 1241 return true; 1242} 1243 1244int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) 1245{ 1246 int ret; 1247 struct mem_cgroup *curr = NULL; 1248 struct task_struct *p; 1249 1250 p = find_lock_task_mm(task); 1251 if (!p) 1252 return 0; 1253 curr = try_get_mem_cgroup_from_mm(p->mm); 1254 task_unlock(p); 1255 if (!curr) 1256 return 0; 1257 /* 1258 * We should check use_hierarchy of "memcg" not "curr". Because checking 1259 * use_hierarchy of "curr" here make this function true if hierarchy is 1260 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* 1261 * hierarchy(even if use_hierarchy is disabled in "memcg"). 1262 */ 1263 ret = mem_cgroup_same_or_subtree(memcg, curr); 1264 css_put(&curr->css); 1265 return ret; 1266} 1267 1268int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) 1269{ 1270 unsigned long inactive_ratio; 1271 int nid = zone_to_nid(zone); 1272 int zid = zone_idx(zone); 1273 unsigned long inactive; 1274 unsigned long active; 1275 unsigned long gb; 1276 1277 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1278 BIT(LRU_INACTIVE_ANON)); 1279 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1280 BIT(LRU_ACTIVE_ANON)); 1281 1282 gb = (inactive + active) >> (30 - PAGE_SHIFT); 1283 if (gb) 1284 inactive_ratio = int_sqrt(10 * gb); 1285 else 1286 inactive_ratio = 1; 1287 1288 return inactive * inactive_ratio < active; 1289} 1290 1291int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) 1292{ 1293 unsigned long active; 1294 unsigned long inactive; 1295 int zid = zone_idx(zone); 1296 int nid = zone_to_nid(zone); 1297 1298 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1299 BIT(LRU_INACTIVE_FILE)); 1300 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1301 BIT(LRU_ACTIVE_FILE)); 1302 1303 return (active > inactive); 1304} 1305 1306struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1307 struct zone *zone) 1308{ 1309 int nid = zone_to_nid(zone); 1310 int zid = zone_idx(zone); 1311 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1312 1313 return &mz->reclaim_stat; 1314} 1315 1316struct zone_reclaim_stat * 1317mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1318{ 1319 struct page_cgroup *pc; 1320 struct mem_cgroup_per_zone *mz; 1321 1322 if (mem_cgroup_disabled()) 1323 return NULL; 1324 1325 pc = lookup_page_cgroup(page); 1326 if (!PageCgroupUsed(pc)) 1327 return NULL; 1328 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1329 smp_rmb(); 1330 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1331 return &mz->reclaim_stat; 1332} 1333 1334#define mem_cgroup_from_res_counter(counter, member) \ 1335 container_of(counter, struct mem_cgroup, member) 1336 1337/** 1338 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1339 * @mem: the memory cgroup 1340 * 1341 * Returns the maximum amount of memory @mem can be charged with, in 1342 * pages. 1343 */ 1344static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) 1345{ 1346 unsigned long long margin; 1347 1348 margin = res_counter_margin(&memcg->res); 1349 if (do_swap_account) 1350 margin = min(margin, res_counter_margin(&memcg->memsw)); 1351 return margin >> PAGE_SHIFT; 1352} 1353 1354int mem_cgroup_swappiness(struct mem_cgroup *memcg) 1355{ 1356 struct cgroup *cgrp = memcg->css.cgroup; 1357 1358 /* root ? */ 1359 if (cgrp->parent == NULL) 1360 return vm_swappiness; 1361 1362 return memcg->swappiness; 1363} 1364 1365static void mem_cgroup_start_move(struct mem_cgroup *memcg) 1366{ 1367 int cpu; 1368 1369 get_online_cpus(); 1370 spin_lock(&memcg->pcp_counter_lock); 1371 for_each_online_cpu(cpu) 1372 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1373 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1374 spin_unlock(&memcg->pcp_counter_lock); 1375 put_online_cpus(); 1376 1377 synchronize_rcu(); 1378} 1379 1380static void mem_cgroup_end_move(struct mem_cgroup *memcg) 1381{ 1382 int cpu; 1383 1384 if (!memcg) 1385 return; 1386 get_online_cpus(); 1387 spin_lock(&memcg->pcp_counter_lock); 1388 for_each_online_cpu(cpu) 1389 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1390 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1391 spin_unlock(&memcg->pcp_counter_lock); 1392 put_online_cpus(); 1393} 1394/* 1395 * 2 routines for checking "mem" is under move_account() or not. 1396 * 1397 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1398 * for avoiding race in accounting. If true, 1399 * pc->mem_cgroup may be overwritten. 1400 * 1401 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1402 * under hierarchy of moving cgroups. This is for 1403 * waiting at hith-memory prressure caused by "move". 1404 */ 1405 1406static bool mem_cgroup_stealed(struct mem_cgroup *memcg) 1407{ 1408 VM_BUG_ON(!rcu_read_lock_held()); 1409 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1410} 1411 1412static bool mem_cgroup_under_move(struct mem_cgroup *memcg) 1413{ 1414 struct mem_cgroup *from; 1415 struct mem_cgroup *to; 1416 bool ret = false; 1417 /* 1418 * Unlike task_move routines, we access mc.to, mc.from not under 1419 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1420 */ 1421 spin_lock(&mc.lock); 1422 from = mc.from; 1423 to = mc.to; 1424 if (!from) 1425 goto unlock; 1426 1427 ret = mem_cgroup_same_or_subtree(memcg, from) 1428 || mem_cgroup_same_or_subtree(memcg, to); 1429unlock: 1430 spin_unlock(&mc.lock); 1431 return ret; 1432} 1433 1434static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) 1435{ 1436 if (mc.moving_task && current != mc.moving_task) { 1437 if (mem_cgroup_under_move(memcg)) { 1438 DEFINE_WAIT(wait); 1439 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1440 /* moving charge context might have finished. */ 1441 if (mc.moving_task) 1442 schedule(); 1443 finish_wait(&mc.waitq, &wait); 1444 return true; 1445 } 1446 } 1447 return false; 1448} 1449 1450/** 1451 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1452 * @memcg: The memory cgroup that went over limit 1453 * @p: Task that is going to be killed 1454 * 1455 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1456 * enabled 1457 */ 1458void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1459{ 1460 struct cgroup *task_cgrp; 1461 struct cgroup *mem_cgrp; 1462 /* 1463 * Need a buffer in BSS, can't rely on allocations. The code relies 1464 * on the assumption that OOM is serialized for memory controller. 1465 * If this assumption is broken, revisit this code. 1466 */ 1467 static char memcg_name[PATH_MAX]; 1468 int ret; 1469 1470 if (!memcg || !p) 1471 return; 1472 1473 1474 rcu_read_lock(); 1475 1476 mem_cgrp = memcg->css.cgroup; 1477 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1478 1479 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1480 if (ret < 0) { 1481 /* 1482 * Unfortunately, we are unable to convert to a useful name 1483 * But we'll still print out the usage information 1484 */ 1485 rcu_read_unlock(); 1486 goto done; 1487 } 1488 rcu_read_unlock(); 1489 1490 printk(KERN_INFO "Task in %s killed", memcg_name); 1491 1492 rcu_read_lock(); 1493 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1494 if (ret < 0) { 1495 rcu_read_unlock(); 1496 goto done; 1497 } 1498 rcu_read_unlock(); 1499 1500 /* 1501 * Continues from above, so we don't need an KERN_ level 1502 */ 1503 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1504done: 1505 1506 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1507 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1508 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1509 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1510 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1511 "failcnt %llu\n", 1512 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1513 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1514 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1515} 1516 1517/* 1518 * This function returns the number of memcg under hierarchy tree. Returns 1519 * 1(self count) if no children. 1520 */ 1521static int mem_cgroup_count_children(struct mem_cgroup *memcg) 1522{ 1523 int num = 0; 1524 struct mem_cgroup *iter; 1525 1526 for_each_mem_cgroup_tree(iter, memcg) 1527 num++; 1528 return num; 1529} 1530 1531/* 1532 * Return the memory (and swap, if configured) limit for a memcg. 1533 */ 1534u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1535{ 1536 u64 limit; 1537 u64 memsw; 1538 1539 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1540 limit += total_swap_pages << PAGE_SHIFT; 1541 1542 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1543 /* 1544 * If memsw is finite and limits the amount of swap space available 1545 * to this memcg, return that limit. 1546 */ 1547 return min(limit, memsw); 1548} 1549 1550static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, 1551 gfp_t gfp_mask, 1552 unsigned long flags) 1553{ 1554 unsigned long total = 0; 1555 bool noswap = false; 1556 int loop; 1557 1558 if (flags & MEM_CGROUP_RECLAIM_NOSWAP) 1559 noswap = true; 1560 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) 1561 noswap = true; 1562 1563 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { 1564 if (loop) 1565 drain_all_stock_async(memcg); 1566 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); 1567 /* 1568 * Allow limit shrinkers, which are triggered directly 1569 * by userspace, to catch signals and stop reclaim 1570 * after minimal progress, regardless of the margin. 1571 */ 1572 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) 1573 break; 1574 if (mem_cgroup_margin(memcg)) 1575 break; 1576 /* 1577 * If nothing was reclaimed after two attempts, there 1578 * may be no reclaimable pages in this hierarchy. 1579 */ 1580 if (loop && !total) 1581 break; 1582 } 1583 return total; 1584} 1585 1586/** 1587 * test_mem_cgroup_node_reclaimable 1588 * @mem: the target memcg 1589 * @nid: the node ID to be checked. 1590 * @noswap : specify true here if the user wants flle only information. 1591 * 1592 * This function returns whether the specified memcg contains any 1593 * reclaimable pages on a node. Returns true if there are any reclaimable 1594 * pages in the node. 1595 */ 1596static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, 1597 int nid, bool noswap) 1598{ 1599 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) 1600 return true; 1601 if (noswap || !total_swap_pages) 1602 return false; 1603 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) 1604 return true; 1605 return false; 1606 1607} 1608#if MAX_NUMNODES > 1 1609 1610/* 1611 * Always updating the nodemask is not very good - even if we have an empty 1612 * list or the wrong list here, we can start from some node and traverse all 1613 * nodes based on the zonelist. So update the list loosely once per 10 secs. 1614 * 1615 */ 1616static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) 1617{ 1618 int nid; 1619 /* 1620 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET 1621 * pagein/pageout changes since the last update. 1622 */ 1623 if (!atomic_read(&memcg->numainfo_events)) 1624 return; 1625 if (atomic_inc_return(&memcg->numainfo_updating) > 1) 1626 return; 1627 1628 /* make a nodemask where this memcg uses memory from */ 1629 memcg->scan_nodes = node_states[N_HIGH_MEMORY]; 1630 1631 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { 1632 1633 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) 1634 node_clear(nid, memcg->scan_nodes); 1635 } 1636 1637 atomic_set(&memcg->numainfo_events, 0); 1638 atomic_set(&memcg->numainfo_updating, 0); 1639} 1640 1641/* 1642 * Selecting a node where we start reclaim from. Because what we need is just 1643 * reducing usage counter, start from anywhere is O,K. Considering 1644 * memory reclaim from current node, there are pros. and cons. 1645 * 1646 * Freeing memory from current node means freeing memory from a node which 1647 * we'll use or we've used. So, it may make LRU bad. And if several threads 1648 * hit limits, it will see a contention on a node. But freeing from remote 1649 * node means more costs for memory reclaim because of memory latency. 1650 * 1651 * Now, we use round-robin. Better algorithm is welcomed. 1652 */ 1653int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1654{ 1655 int node; 1656 1657 mem_cgroup_may_update_nodemask(memcg); 1658 node = memcg->last_scanned_node; 1659 1660 node = next_node(node, memcg->scan_nodes); 1661 if (node == MAX_NUMNODES) 1662 node = first_node(memcg->scan_nodes); 1663 /* 1664 * We call this when we hit limit, not when pages are added to LRU. 1665 * No LRU may hold pages because all pages are UNEVICTABLE or 1666 * memcg is too small and all pages are not on LRU. In that case, 1667 * we use curret node. 1668 */ 1669 if (unlikely(node == MAX_NUMNODES)) 1670 node = numa_node_id(); 1671 1672 memcg->last_scanned_node = node; 1673 return node; 1674} 1675 1676/* 1677 * Check all nodes whether it contains reclaimable pages or not. 1678 * For quick scan, we make use of scan_nodes. This will allow us to skip 1679 * unused nodes. But scan_nodes is lazily updated and may not cotain 1680 * enough new information. We need to do double check. 1681 */ 1682bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1683{ 1684 int nid; 1685 1686 /* 1687 * quick check...making use of scan_node. 1688 * We can skip unused nodes. 1689 */ 1690 if (!nodes_empty(memcg->scan_nodes)) { 1691 for (nid = first_node(memcg->scan_nodes); 1692 nid < MAX_NUMNODES; 1693 nid = next_node(nid, memcg->scan_nodes)) { 1694 1695 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1696 return true; 1697 } 1698 } 1699 /* 1700 * Check rest of nodes. 1701 */ 1702 for_each_node_state(nid, N_HIGH_MEMORY) { 1703 if (node_isset(nid, memcg->scan_nodes)) 1704 continue; 1705 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1706 return true; 1707 } 1708 return false; 1709} 1710 1711#else 1712int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1713{ 1714 return 0; 1715} 1716 1717bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1718{ 1719 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); 1720} 1721#endif 1722 1723static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, 1724 struct zone *zone, 1725 gfp_t gfp_mask, 1726 unsigned long *total_scanned) 1727{ 1728 struct mem_cgroup *victim = NULL; 1729 int total = 0; 1730 int loop = 0; 1731 unsigned long excess; 1732 unsigned long nr_scanned; 1733 struct mem_cgroup_reclaim_cookie reclaim = { 1734 .zone = zone, 1735 .priority = 0, 1736 }; 1737 1738 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; 1739 1740 while (1) { 1741 victim = mem_cgroup_iter(root_memcg, victim, &reclaim); 1742 if (!victim) { 1743 loop++; 1744 if (loop >= 2) { 1745 /* 1746 * If we have not been able to reclaim 1747 * anything, it might because there are 1748 * no reclaimable pages under this hierarchy 1749 */ 1750 if (!total) 1751 break; 1752 /* 1753 * We want to do more targeted reclaim. 1754 * excess >> 2 is not to excessive so as to 1755 * reclaim too much, nor too less that we keep 1756 * coming back to reclaim from this cgroup 1757 */ 1758 if (total >= (excess >> 2) || 1759 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) 1760 break; 1761 } 1762 continue; 1763 } 1764 if (!mem_cgroup_reclaimable(victim, false)) 1765 continue; 1766 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, 1767 zone, &nr_scanned); 1768 *total_scanned += nr_scanned; 1769 if (!res_counter_soft_limit_excess(&root_memcg->res)) 1770 break; 1771 } 1772 mem_cgroup_iter_break(root_memcg, victim); 1773 return total; 1774} 1775 1776/* 1777 * Check OOM-Killer is already running under our hierarchy. 1778 * If someone is running, return false. 1779 * Has to be called with memcg_oom_lock 1780 */ 1781static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) 1782{ 1783 struct mem_cgroup *iter, *failed = NULL; 1784 1785 for_each_mem_cgroup_tree(iter, memcg) { 1786 if (iter->oom_lock) { 1787 /* 1788 * this subtree of our hierarchy is already locked 1789 * so we cannot give a lock. 1790 */ 1791 failed = iter; 1792 mem_cgroup_iter_break(memcg, iter); 1793 break; 1794 } else 1795 iter->oom_lock = true; 1796 } 1797 1798 if (!failed) 1799 return true; 1800 1801 /* 1802 * OK, we failed to lock the whole subtree so we have to clean up 1803 * what we set up to the failing subtree 1804 */ 1805 for_each_mem_cgroup_tree(iter, memcg) { 1806 if (iter == failed) { 1807 mem_cgroup_iter_break(memcg, iter); 1808 break; 1809 } 1810 iter->oom_lock = false; 1811 } 1812 return false; 1813} 1814 1815/* 1816 * Has to be called with memcg_oom_lock 1817 */ 1818static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) 1819{ 1820 struct mem_cgroup *iter; 1821 1822 for_each_mem_cgroup_tree(iter, memcg) 1823 iter->oom_lock = false; 1824 return 0; 1825} 1826 1827static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) 1828{ 1829 struct mem_cgroup *iter; 1830 1831 for_each_mem_cgroup_tree(iter, memcg) 1832 atomic_inc(&iter->under_oom); 1833} 1834 1835static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) 1836{ 1837 struct mem_cgroup *iter; 1838 1839 /* 1840 * When a new child is created while the hierarchy is under oom, 1841 * mem_cgroup_oom_lock() may not be called. We have to use 1842 * atomic_add_unless() here. 1843 */ 1844 for_each_mem_cgroup_tree(iter, memcg) 1845 atomic_add_unless(&iter->under_oom, -1, 0); 1846} 1847 1848static DEFINE_SPINLOCK(memcg_oom_lock); 1849static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1850 1851struct oom_wait_info { 1852 struct mem_cgroup *mem; 1853 wait_queue_t wait; 1854}; 1855 1856static int memcg_oom_wake_function(wait_queue_t *wait, 1857 unsigned mode, int sync, void *arg) 1858{ 1859 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg, 1860 *oom_wait_memcg; 1861 struct oom_wait_info *oom_wait_info; 1862 1863 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1864 oom_wait_memcg = oom_wait_info->mem; 1865 1866 /* 1867 * Both of oom_wait_info->mem and wake_mem are stable under us. 1868 * Then we can use css_is_ancestor without taking care of RCU. 1869 */ 1870 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) 1871 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) 1872 return 0; 1873 return autoremove_wake_function(wait, mode, sync, arg); 1874} 1875 1876static void memcg_wakeup_oom(struct mem_cgroup *memcg) 1877{ 1878 /* for filtering, pass "memcg" as argument. */ 1879 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); 1880} 1881 1882static void memcg_oom_recover(struct mem_cgroup *memcg) 1883{ 1884 if (memcg && atomic_read(&memcg->under_oom)) 1885 memcg_wakeup_oom(memcg); 1886} 1887 1888/* 1889 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1890 */ 1891bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask) 1892{ 1893 struct oom_wait_info owait; 1894 bool locked, need_to_kill; 1895 1896 owait.mem = memcg; 1897 owait.wait.flags = 0; 1898 owait.wait.func = memcg_oom_wake_function; 1899 owait.wait.private = current; 1900 INIT_LIST_HEAD(&owait.wait.task_list); 1901 need_to_kill = true; 1902 mem_cgroup_mark_under_oom(memcg); 1903 1904 /* At first, try to OOM lock hierarchy under memcg.*/ 1905 spin_lock(&memcg_oom_lock); 1906 locked = mem_cgroup_oom_lock(memcg); 1907 /* 1908 * Even if signal_pending(), we can't quit charge() loop without 1909 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1910 * under OOM is always welcomed, use TASK_KILLABLE here. 1911 */ 1912 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1913 if (!locked || memcg->oom_kill_disable) 1914 need_to_kill = false; 1915 if (locked) 1916 mem_cgroup_oom_notify(memcg); 1917 spin_unlock(&memcg_oom_lock); 1918 1919 if (need_to_kill) { 1920 finish_wait(&memcg_oom_waitq, &owait.wait); 1921 mem_cgroup_out_of_memory(memcg, mask); 1922 } else { 1923 schedule(); 1924 finish_wait(&memcg_oom_waitq, &owait.wait); 1925 } 1926 spin_lock(&memcg_oom_lock); 1927 if (locked) 1928 mem_cgroup_oom_unlock(memcg); 1929 memcg_wakeup_oom(memcg); 1930 spin_unlock(&memcg_oom_lock); 1931 1932 mem_cgroup_unmark_under_oom(memcg); 1933 1934 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1935 return false; 1936 /* Give chance to dying process */ 1937 schedule_timeout_uninterruptible(1); 1938 return true; 1939} 1940 1941/* 1942 * Currently used to update mapped file statistics, but the routine can be 1943 * generalized to update other statistics as well. 1944 * 1945 * Notes: Race condition 1946 * 1947 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1948 * it tends to be costly. But considering some conditions, we doesn't need 1949 * to do so _always_. 1950 * 1951 * Considering "charge", lock_page_cgroup() is not required because all 1952 * file-stat operations happen after a page is attached to radix-tree. There 1953 * are no race with "charge". 1954 * 1955 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1956 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1957 * if there are race with "uncharge". Statistics itself is properly handled 1958 * by flags. 1959 * 1960 * Considering "move", this is an only case we see a race. To make the race 1961 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1962 * possibility of race condition. If there is, we take a lock. 1963 */ 1964 1965void mem_cgroup_update_page_stat(struct page *page, 1966 enum mem_cgroup_page_stat_item idx, int val) 1967{ 1968 struct mem_cgroup *memcg; 1969 struct page_cgroup *pc = lookup_page_cgroup(page); 1970 bool need_unlock = false; 1971 unsigned long uninitialized_var(flags); 1972 1973 if (unlikely(!pc)) 1974 return; 1975 1976 rcu_read_lock(); 1977 memcg = pc->mem_cgroup; 1978 if (unlikely(!memcg || !PageCgroupUsed(pc))) 1979 goto out; 1980 /* pc->mem_cgroup is unstable ? */ 1981 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) { 1982 /* take a lock against to access pc->mem_cgroup */ 1983 move_lock_page_cgroup(pc, &flags); 1984 need_unlock = true; 1985 memcg = pc->mem_cgroup; 1986 if (!memcg || !PageCgroupUsed(pc)) 1987 goto out; 1988 } 1989 1990 switch (idx) { 1991 case MEMCG_NR_FILE_MAPPED: 1992 if (val > 0) 1993 SetPageCgroupFileMapped(pc); 1994 else if (!page_mapped(page)) 1995 ClearPageCgroupFileMapped(pc); 1996 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1997 break; 1998 default: 1999 BUG(); 2000 } 2001 2002 this_cpu_add(memcg->stat->count[idx], val); 2003 2004out: 2005 if (unlikely(need_unlock)) 2006 move_unlock_page_cgroup(pc, &flags); 2007 rcu_read_unlock(); 2008 return; 2009} 2010EXPORT_SYMBOL(mem_cgroup_update_page_stat); 2011 2012/* 2013 * size of first charge trial. "32" comes from vmscan.c's magic value. 2014 * TODO: maybe necessary to use big numbers in big irons. 2015 */ 2016#define CHARGE_BATCH 32U 2017struct memcg_stock_pcp { 2018 struct mem_cgroup *cached; /* this never be root cgroup */ 2019 unsigned int nr_pages; 2020 struct work_struct work; 2021 unsigned long flags; 2022#define FLUSHING_CACHED_CHARGE (0) 2023}; 2024static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 2025static DEFINE_MUTEX(percpu_charge_mutex); 2026 2027/* 2028 * Try to consume stocked charge on this cpu. If success, one page is consumed 2029 * from local stock and true is returned. If the stock is 0 or charges from a 2030 * cgroup which is not current target, returns false. This stock will be 2031 * refilled. 2032 */ 2033static bool consume_stock(struct mem_cgroup *memcg) 2034{ 2035 struct memcg_stock_pcp *stock; 2036 bool ret = true; 2037 2038 stock = &get_cpu_var(memcg_stock); 2039 if (memcg == stock->cached && stock->nr_pages) 2040 stock->nr_pages--; 2041 else /* need to call res_counter_charge */ 2042 ret = false; 2043 put_cpu_var(memcg_stock); 2044 return ret; 2045} 2046 2047/* 2048 * Returns stocks cached in percpu to res_counter and reset cached information. 2049 */ 2050static void drain_stock(struct memcg_stock_pcp *stock) 2051{ 2052 struct mem_cgroup *old = stock->cached; 2053 2054 if (stock->nr_pages) { 2055 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 2056 2057 res_counter_uncharge(&old->res, bytes); 2058 if (do_swap_account) 2059 res_counter_uncharge(&old->memsw, bytes); 2060 stock->nr_pages = 0; 2061 } 2062 stock->cached = NULL; 2063} 2064 2065/* 2066 * This must be called under preempt disabled or must be called by 2067 * a thread which is pinned to local cpu. 2068 */ 2069static void drain_local_stock(struct work_struct *dummy) 2070{ 2071 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 2072 drain_stock(stock); 2073 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); 2074} 2075 2076/* 2077 * Cache charges(val) which is from res_counter, to local per_cpu area. 2078 * This will be consumed by consume_stock() function, later. 2079 */ 2080static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) 2081{ 2082 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 2083 2084 if (stock->cached != memcg) { /* reset if necessary */ 2085 drain_stock(stock); 2086 stock->cached = memcg; 2087 } 2088 stock->nr_pages += nr_pages; 2089 put_cpu_var(memcg_stock); 2090} 2091 2092/* 2093 * Drains all per-CPU charge caches for given root_memcg resp. subtree 2094 * of the hierarchy under it. sync flag says whether we should block 2095 * until the work is done. 2096 */ 2097static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) 2098{ 2099 int cpu, curcpu; 2100 2101 /* Notify other cpus that system-wide "drain" is running */ 2102 get_online_cpus(); 2103 curcpu = get_cpu(); 2104 for_each_online_cpu(cpu) { 2105 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2106 struct mem_cgroup *memcg; 2107 2108 memcg = stock->cached; 2109 if (!memcg || !stock->nr_pages) 2110 continue; 2111 if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) 2112 continue; 2113 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { 2114 if (cpu == curcpu) 2115 drain_local_stock(&stock->work); 2116 else 2117 schedule_work_on(cpu, &stock->work); 2118 } 2119 } 2120 put_cpu(); 2121 2122 if (!sync) 2123 goto out; 2124 2125 for_each_online_cpu(cpu) { 2126 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2127 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) 2128 flush_work(&stock->work); 2129 } 2130out: 2131 put_online_cpus(); 2132} 2133 2134/* 2135 * Tries to drain stocked charges in other cpus. This function is asynchronous 2136 * and just put a work per cpu for draining localy on each cpu. Caller can 2137 * expects some charges will be back to res_counter later but cannot wait for 2138 * it. 2139 */ 2140static void drain_all_stock_async(struct mem_cgroup *root_memcg) 2141{ 2142 /* 2143 * If someone calls draining, avoid adding more kworker runs. 2144 */ 2145 if (!mutex_trylock(&percpu_charge_mutex)) 2146 return; 2147 drain_all_stock(root_memcg, false); 2148 mutex_unlock(&percpu_charge_mutex); 2149} 2150 2151/* This is a synchronous drain interface. */ 2152static void drain_all_stock_sync(struct mem_cgroup *root_memcg) 2153{ 2154 /* called when force_empty is called */ 2155 mutex_lock(&percpu_charge_mutex); 2156 drain_all_stock(root_memcg, true); 2157 mutex_unlock(&percpu_charge_mutex); 2158} 2159 2160/* 2161 * This function drains percpu counter value from DEAD cpu and 2162 * move it to local cpu. Note that this function can be preempted. 2163 */ 2164static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) 2165{ 2166 int i; 2167 2168 spin_lock(&memcg->pcp_counter_lock); 2169 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 2170 long x = per_cpu(memcg->stat->count[i], cpu); 2171 2172 per_cpu(memcg->stat->count[i], cpu) = 0; 2173 memcg->nocpu_base.count[i] += x; 2174 } 2175 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { 2176 unsigned long x = per_cpu(memcg->stat->events[i], cpu); 2177 2178 per_cpu(memcg->stat->events[i], cpu) = 0; 2179 memcg->nocpu_base.events[i] += x; 2180 } 2181 /* need to clear ON_MOVE value, works as a kind of lock. */ 2182 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 2183 spin_unlock(&memcg->pcp_counter_lock); 2184} 2185 2186static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu) 2187{ 2188 int idx = MEM_CGROUP_ON_MOVE; 2189 2190 spin_lock(&memcg->pcp_counter_lock); 2191 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx]; 2192 spin_unlock(&memcg->pcp_counter_lock); 2193} 2194 2195static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 2196 unsigned long action, 2197 void *hcpu) 2198{ 2199 int cpu = (unsigned long)hcpu; 2200 struct memcg_stock_pcp *stock; 2201 struct mem_cgroup *iter; 2202 2203 if ((action == CPU_ONLINE)) { 2204 for_each_mem_cgroup(iter) 2205 synchronize_mem_cgroup_on_move(iter, cpu); 2206 return NOTIFY_OK; 2207 } 2208 2209 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 2210 return NOTIFY_OK; 2211 2212 for_each_mem_cgroup(iter) 2213 mem_cgroup_drain_pcp_counter(iter, cpu); 2214 2215 stock = &per_cpu(memcg_stock, cpu); 2216 drain_stock(stock); 2217 return NOTIFY_OK; 2218} 2219 2220 2221/* See __mem_cgroup_try_charge() for details */ 2222enum { 2223 CHARGE_OK, /* success */ 2224 CHARGE_RETRY, /* need to retry but retry is not bad */ 2225 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 2226 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 2227 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 2228}; 2229 2230static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, 2231 unsigned int nr_pages, bool oom_check) 2232{ 2233 unsigned long csize = nr_pages * PAGE_SIZE; 2234 struct mem_cgroup *mem_over_limit; 2235 struct res_counter *fail_res; 2236 unsigned long flags = 0; 2237 int ret; 2238 2239 ret = res_counter_charge(&memcg->res, csize, &fail_res); 2240 2241 if (likely(!ret)) { 2242 if (!do_swap_account) 2243 return CHARGE_OK; 2244 ret = res_counter_charge(&memcg->memsw, csize, &fail_res); 2245 if (likely(!ret)) 2246 return CHARGE_OK; 2247 2248 res_counter_uncharge(&memcg->res, csize); 2249 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 2250 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 2251 } else 2252 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 2253 /* 2254 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch 2255 * of regular pages (CHARGE_BATCH), or a single regular page (1). 2256 * 2257 * Never reclaim on behalf of optional batching, retry with a 2258 * single page instead. 2259 */ 2260 if (nr_pages == CHARGE_BATCH) 2261 return CHARGE_RETRY; 2262 2263 if (!(gfp_mask & __GFP_WAIT)) 2264 return CHARGE_WOULDBLOCK; 2265 2266 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); 2267 if (mem_cgroup_margin(mem_over_limit) >= nr_pages) 2268 return CHARGE_RETRY; 2269 /* 2270 * Even though the limit is exceeded at this point, reclaim 2271 * may have been able to free some pages. Retry the charge 2272 * before killing the task. 2273 * 2274 * Only for regular pages, though: huge pages are rather 2275 * unlikely to succeed so close to the limit, and we fall back 2276 * to regular pages anyway in case of failure. 2277 */ 2278 if (nr_pages == 1 && ret) 2279 return CHARGE_RETRY; 2280 2281 /* 2282 * At task move, charge accounts can be doubly counted. So, it's 2283 * better to wait until the end of task_move if something is going on. 2284 */ 2285 if (mem_cgroup_wait_acct_move(mem_over_limit)) 2286 return CHARGE_RETRY; 2287 2288 /* If we don't need to call oom-killer at el, return immediately */ 2289 if (!oom_check) 2290 return CHARGE_NOMEM; 2291 /* check OOM */ 2292 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 2293 return CHARGE_OOM_DIE; 2294 2295 return CHARGE_RETRY; 2296} 2297 2298/* 2299 * Unlike exported interface, "oom" parameter is added. if oom==true, 2300 * oom-killer can be invoked. 2301 */ 2302static int __mem_cgroup_try_charge(struct mm_struct *mm, 2303 gfp_t gfp_mask, 2304 unsigned int nr_pages, 2305 struct mem_cgroup **ptr, 2306 bool oom) 2307{ 2308 unsigned int batch = max(CHARGE_BATCH, nr_pages); 2309 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2310 struct mem_cgroup *memcg = NULL; 2311 int ret; 2312 2313 /* 2314 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 2315 * in system level. So, allow to go ahead dying process in addition to 2316 * MEMDIE process. 2317 */ 2318 if (unlikely(test_thread_flag(TIF_MEMDIE) 2319 || fatal_signal_pending(current))) 2320 goto bypass; 2321 2322 /* 2323 * We always charge the cgroup the mm_struct belongs to. 2324 * The mm_struct's mem_cgroup changes on task migration if the 2325 * thread group leader migrates. It's possible that mm is not 2326 * set, if so charge the init_mm (happens for pagecache usage). 2327 */ 2328 if (!*ptr && !mm) 2329 goto bypass; 2330again: 2331 if (*ptr) { /* css should be a valid one */ 2332 memcg = *ptr; 2333 VM_BUG_ON(css_is_removed(&memcg->css)); 2334 if (mem_cgroup_is_root(memcg)) 2335 goto done; 2336 if (nr_pages == 1 && consume_stock(memcg)) 2337 goto done; 2338 css_get(&memcg->css); 2339 } else { 2340 struct task_struct *p; 2341 2342 rcu_read_lock(); 2343 p = rcu_dereference(mm->owner); 2344 /* 2345 * Because we don't have task_lock(), "p" can exit. 2346 * In that case, "memcg" can point to root or p can be NULL with 2347 * race with swapoff. Then, we have small risk of mis-accouning. 2348 * But such kind of mis-account by race always happens because 2349 * we don't have cgroup_mutex(). It's overkill and we allo that 2350 * small race, here. 2351 * (*) swapoff at el will charge against mm-struct not against 2352 * task-struct. So, mm->owner can be NULL. 2353 */ 2354 memcg = mem_cgroup_from_task(p); 2355 if (!memcg || mem_cgroup_is_root(memcg)) { 2356 rcu_read_unlock(); 2357 goto done; 2358 } 2359 if (nr_pages == 1 && consume_stock(memcg)) { 2360 /* 2361 * It seems dagerous to access memcg without css_get(). 2362 * But considering how consume_stok works, it's not 2363 * necessary. If consume_stock success, some charges 2364 * from this memcg are cached on this cpu. So, we 2365 * don't need to call css_get()/css_tryget() before 2366 * calling consume_stock(). 2367 */ 2368 rcu_read_unlock(); 2369 goto done; 2370 } 2371 /* after here, we may be blocked. we need to get refcnt */ 2372 if (!css_tryget(&memcg->css)) { 2373 rcu_read_unlock(); 2374 goto again; 2375 } 2376 rcu_read_unlock(); 2377 } 2378 2379 do { 2380 bool oom_check; 2381 2382 /* If killed, bypass charge */ 2383 if (fatal_signal_pending(current)) { 2384 css_put(&memcg->css); 2385 goto bypass; 2386 } 2387 2388 oom_check = false; 2389 if (oom && !nr_oom_retries) { 2390 oom_check = true; 2391 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2392 } 2393 2394 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); 2395 switch (ret) { 2396 case CHARGE_OK: 2397 break; 2398 case CHARGE_RETRY: /* not in OOM situation but retry */ 2399 batch = nr_pages; 2400 css_put(&memcg->css); 2401 memcg = NULL; 2402 goto again; 2403 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2404 css_put(&memcg->css); 2405 goto nomem; 2406 case CHARGE_NOMEM: /* OOM routine works */ 2407 if (!oom) { 2408 css_put(&memcg->css); 2409 goto nomem; 2410 } 2411 /* If oom, we never return -ENOMEM */ 2412 nr_oom_retries--; 2413 break; 2414 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2415 css_put(&memcg->css); 2416 goto bypass; 2417 } 2418 } while (ret != CHARGE_OK); 2419 2420 if (batch > nr_pages) 2421 refill_stock(memcg, batch - nr_pages); 2422 css_put(&memcg->css); 2423done: 2424 *ptr = memcg; 2425 return 0; 2426nomem: 2427 *ptr = NULL; 2428 return -ENOMEM; 2429bypass: 2430 *ptr = NULL; 2431 return 0; 2432} 2433 2434/* 2435 * Somemtimes we have to undo a charge we got by try_charge(). 2436 * This function is for that and do uncharge, put css's refcnt. 2437 * gotten by try_charge(). 2438 */ 2439static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, 2440 unsigned int nr_pages) 2441{ 2442 if (!mem_cgroup_is_root(memcg)) { 2443 unsigned long bytes = nr_pages * PAGE_SIZE; 2444 2445 res_counter_uncharge(&memcg->res, bytes); 2446 if (do_swap_account) 2447 res_counter_uncharge(&memcg->memsw, bytes); 2448 } 2449} 2450 2451/* 2452 * A helper function to get mem_cgroup from ID. must be called under 2453 * rcu_read_lock(). The caller must check css_is_removed() or some if 2454 * it's concern. (dropping refcnt from swap can be called against removed 2455 * memcg.) 2456 */ 2457static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2458{ 2459 struct cgroup_subsys_state *css; 2460 2461 /* ID 0 is unused ID */ 2462 if (!id) 2463 return NULL; 2464 css = css_lookup(&mem_cgroup_subsys, id); 2465 if (!css) 2466 return NULL; 2467 return container_of(css, struct mem_cgroup, css); 2468} 2469 2470struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2471{ 2472 struct mem_cgroup *memcg = NULL; 2473 struct page_cgroup *pc; 2474 unsigned short id; 2475 swp_entry_t ent; 2476 2477 VM_BUG_ON(!PageLocked(page)); 2478 2479 pc = lookup_page_cgroup(page); 2480 lock_page_cgroup(pc); 2481 if (PageCgroupUsed(pc)) { 2482 memcg = pc->mem_cgroup; 2483 if (memcg && !css_tryget(&memcg->css)) 2484 memcg = NULL; 2485 } else if (PageSwapCache(page)) { 2486 ent.val = page_private(page); 2487 id = lookup_swap_cgroup(ent); 2488 rcu_read_lock(); 2489 memcg = mem_cgroup_lookup(id); 2490 if (memcg && !css_tryget(&memcg->css)) 2491 memcg = NULL; 2492 rcu_read_unlock(); 2493 } 2494 unlock_page_cgroup(pc); 2495 return memcg; 2496} 2497 2498static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, 2499 struct page *page, 2500 unsigned int nr_pages, 2501 struct page_cgroup *pc, 2502 enum charge_type ctype) 2503{ 2504 lock_page_cgroup(pc); 2505 if (unlikely(PageCgroupUsed(pc))) { 2506 unlock_page_cgroup(pc); 2507 __mem_cgroup_cancel_charge(memcg, nr_pages); 2508 return; 2509 } 2510 /* 2511 * we don't need page_cgroup_lock about tail pages, becase they are not 2512 * accessed by any other context at this point. 2513 */ 2514 pc->mem_cgroup = memcg; 2515 /* 2516 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2517 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2518 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2519 * before USED bit, we need memory barrier here. 2520 * See mem_cgroup_add_lru_list(), etc. 2521 */ 2522 smp_wmb(); 2523 switch (ctype) { 2524 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2525 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2526 SetPageCgroupCache(pc); 2527 SetPageCgroupUsed(pc); 2528 break; 2529 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2530 ClearPageCgroupCache(pc); 2531 SetPageCgroupUsed(pc); 2532 break; 2533 default: 2534 break; 2535 } 2536 2537 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages); 2538 unlock_page_cgroup(pc); 2539 /* 2540 * "charge_statistics" updated event counter. Then, check it. 2541 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2542 * if they exceeds softlimit. 2543 */ 2544 memcg_check_events(memcg, page); 2545} 2546 2547#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2548 2549#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2550 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) 2551/* 2552 * Because tail pages are not marked as "used", set it. We're under 2553 * zone->lru_lock, 'splitting on pmd' and compound_lock. 2554 * charge/uncharge will be never happen and move_account() is done under 2555 * compound_lock(), so we don't have to take care of races. 2556 */ 2557void mem_cgroup_split_huge_fixup(struct page *head) 2558{ 2559 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2560 struct page_cgroup *pc; 2561 int i; 2562 2563 if (mem_cgroup_disabled()) 2564 return; 2565 for (i = 1; i < HPAGE_PMD_NR; i++) { 2566 pc = head_pc + i; 2567 pc->mem_cgroup = head_pc->mem_cgroup; 2568 smp_wmb();/* see __commit_charge() */ 2569 /* 2570 * LRU flags cannot be copied because we need to add tail 2571 * page to LRU by generic call and our hooks will be called. 2572 */ 2573 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2574 } 2575 2576 if (PageCgroupAcctLRU(head_pc)) { 2577 enum lru_list lru; 2578 struct mem_cgroup_per_zone *mz; 2579 /* 2580 * We hold lru_lock, then, reduce counter directly. 2581 */ 2582 lru = page_lru(head); 2583 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); 2584 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1; 2585 } 2586} 2587#endif 2588 2589/** 2590 * mem_cgroup_move_account - move account of the page 2591 * @page: the page 2592 * @nr_pages: number of regular pages (>1 for huge pages) 2593 * @pc: page_cgroup of the page. 2594 * @from: mem_cgroup which the page is moved from. 2595 * @to: mem_cgroup which the page is moved to. @from != @to. 2596 * @uncharge: whether we should call uncharge and css_put against @from. 2597 * 2598 * The caller must confirm following. 2599 * - page is not on LRU (isolate_page() is useful.) 2600 * - compound_lock is held when nr_pages > 1 2601 * 2602 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2603 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is 2604 * true, this function does "uncharge" from old cgroup, but it doesn't if 2605 * @uncharge is false, so a caller should do "uncharge". 2606 */ 2607static int mem_cgroup_move_account(struct page *page, 2608 unsigned int nr_pages, 2609 struct page_cgroup *pc, 2610 struct mem_cgroup *from, 2611 struct mem_cgroup *to, 2612 bool uncharge) 2613{ 2614 unsigned long flags; 2615 int ret; 2616 2617 VM_BUG_ON(from == to); 2618 VM_BUG_ON(PageLRU(page)); 2619 /* 2620 * The page is isolated from LRU. So, collapse function 2621 * will not handle this page. But page splitting can happen. 2622 * Do this check under compound_page_lock(). The caller should 2623 * hold it. 2624 */ 2625 ret = -EBUSY; 2626 if (nr_pages > 1 && !PageTransHuge(page)) 2627 goto out; 2628 2629 lock_page_cgroup(pc); 2630 2631 ret = -EINVAL; 2632 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2633 goto unlock; 2634 2635 move_lock_page_cgroup(pc, &flags); 2636 2637 if (PageCgroupFileMapped(pc)) { 2638 /* Update mapped_file data for mem_cgroup */ 2639 preempt_disable(); 2640 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2641 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2642 preempt_enable(); 2643 } 2644 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2645 if (uncharge) 2646 /* This is not "cancel", but cancel_charge does all we need. */ 2647 __mem_cgroup_cancel_charge(from, nr_pages); 2648 2649 /* caller should have done css_get */ 2650 pc->mem_cgroup = to; 2651 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2652 /* 2653 * We charges against "to" which may not have any tasks. Then, "to" 2654 * can be under rmdir(). But in current implementation, caller of 2655 * this function is just force_empty() and move charge, so it's 2656 * guaranteed that "to" is never removed. So, we don't check rmdir 2657 * status here. 2658 */ 2659 move_unlock_page_cgroup(pc, &flags); 2660 ret = 0; 2661unlock: 2662 unlock_page_cgroup(pc); 2663 /* 2664 * check events 2665 */ 2666 memcg_check_events(to, page); 2667 memcg_check_events(from, page); 2668out: 2669 return ret; 2670} 2671 2672/* 2673 * move charges to its parent. 2674 */ 2675 2676static int mem_cgroup_move_parent(struct page *page, 2677 struct page_cgroup *pc, 2678 struct mem_cgroup *child, 2679 gfp_t gfp_mask) 2680{ 2681 struct cgroup *cg = child->css.cgroup; 2682 struct cgroup *pcg = cg->parent; 2683 struct mem_cgroup *parent; 2684 unsigned int nr_pages; 2685 unsigned long uninitialized_var(flags); 2686 int ret; 2687 2688 /* Is ROOT ? */ 2689 if (!pcg) 2690 return -EINVAL; 2691 2692 ret = -EBUSY; 2693 if (!get_page_unless_zero(page)) 2694 goto out; 2695 if (isolate_lru_page(page)) 2696 goto put; 2697 2698 nr_pages = hpage_nr_pages(page); 2699 2700 parent = mem_cgroup_from_cont(pcg); 2701 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); 2702 if (ret || !parent) 2703 goto put_back; 2704 2705 if (nr_pages > 1) 2706 flags = compound_lock_irqsave(page); 2707 2708 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); 2709 if (ret) 2710 __mem_cgroup_cancel_charge(parent, nr_pages); 2711 2712 if (nr_pages > 1) 2713 compound_unlock_irqrestore(page, flags); 2714put_back: 2715 putback_lru_page(page); 2716put: 2717 put_page(page); 2718out: 2719 return ret; 2720} 2721 2722/* 2723 * Charge the memory controller for page usage. 2724 * Return 2725 * 0 if the charge was successful 2726 * < 0 if the cgroup is over its limit 2727 */ 2728static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2729 gfp_t gfp_mask, enum charge_type ctype) 2730{ 2731 struct mem_cgroup *memcg = NULL; 2732 unsigned int nr_pages = 1; 2733 struct page_cgroup *pc; 2734 bool oom = true; 2735 int ret; 2736 2737 if (PageTransHuge(page)) { 2738 nr_pages <<= compound_order(page); 2739 VM_BUG_ON(!PageTransHuge(page)); 2740 /* 2741 * Never OOM-kill a process for a huge page. The 2742 * fault handler will fall back to regular pages. 2743 */ 2744 oom = false; 2745 } 2746 2747 pc = lookup_page_cgroup(page); 2748 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */ 2749 2750 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); 2751 if (ret || !memcg) 2752 return ret; 2753 2754 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype); 2755 return 0; 2756} 2757 2758int mem_cgroup_newpage_charge(struct page *page, 2759 struct mm_struct *mm, gfp_t gfp_mask) 2760{ 2761 if (mem_cgroup_disabled()) 2762 return 0; 2763 /* 2764 * If already mapped, we don't have to account. 2765 * If page cache, page->mapping has address_space. 2766 * But page->mapping may have out-of-use anon_vma pointer, 2767 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 2768 * is NULL. 2769 */ 2770 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 2771 return 0; 2772 if (unlikely(!mm)) 2773 mm = &init_mm; 2774 return mem_cgroup_charge_common(page, mm, gfp_mask, 2775 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2776} 2777 2778static void 2779__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2780 enum charge_type ctype); 2781 2782static void 2783__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg, 2784 enum charge_type ctype) 2785{ 2786 struct page_cgroup *pc = lookup_page_cgroup(page); 2787 /* 2788 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page 2789 * is already on LRU. It means the page may on some other page_cgroup's 2790 * LRU. Take care of it. 2791 */ 2792 mem_cgroup_lru_del_before_commit(page); 2793 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 2794 mem_cgroup_lru_add_after_commit(page); 2795 return; 2796} 2797 2798int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2799 gfp_t gfp_mask) 2800{ 2801 struct mem_cgroup *memcg = NULL; 2802 int ret; 2803 2804 if (mem_cgroup_disabled()) 2805 return 0; 2806 if (PageCompound(page)) 2807 return 0; 2808 2809 if (unlikely(!mm)) 2810 mm = &init_mm; 2811 2812 if (page_is_file_cache(page)) { 2813 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true); 2814 if (ret || !memcg) 2815 return ret; 2816 2817 /* 2818 * FUSE reuses pages without going through the final 2819 * put that would remove them from the LRU list, make 2820 * sure that they get relinked properly. 2821 */ 2822 __mem_cgroup_commit_charge_lrucare(page, memcg, 2823 MEM_CGROUP_CHARGE_TYPE_CACHE); 2824 return ret; 2825 } 2826 /* shmem */ 2827 if (PageSwapCache(page)) { 2828 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); 2829 if (!ret) 2830 __mem_cgroup_commit_charge_swapin(page, memcg, 2831 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2832 } else 2833 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 2834 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2835 2836 return ret; 2837} 2838 2839/* 2840 * While swap-in, try_charge -> commit or cancel, the page is locked. 2841 * And when try_charge() successfully returns, one refcnt to memcg without 2842 * struct page_cgroup is acquired. This refcnt will be consumed by 2843 * "commit()" or removed by "cancel()" 2844 */ 2845int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2846 struct page *page, 2847 gfp_t mask, struct mem_cgroup **ptr) 2848{ 2849 struct mem_cgroup *memcg; 2850 int ret; 2851 2852 *ptr = NULL; 2853 2854 if (mem_cgroup_disabled()) 2855 return 0; 2856 2857 if (!do_swap_account) 2858 goto charge_cur_mm; 2859 /* 2860 * A racing thread's fault, or swapoff, may have already updated 2861 * the pte, and even removed page from swap cache: in those cases 2862 * do_swap_page()'s pte_same() test will fail; but there's also a 2863 * KSM case which does need to charge the page. 2864 */ 2865 if (!PageSwapCache(page)) 2866 goto charge_cur_mm; 2867 memcg = try_get_mem_cgroup_from_page(page); 2868 if (!memcg) 2869 goto charge_cur_mm; 2870 *ptr = memcg; 2871 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true); 2872 css_put(&memcg->css); 2873 return ret; 2874charge_cur_mm: 2875 if (unlikely(!mm)) 2876 mm = &init_mm; 2877 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true); 2878} 2879 2880static void 2881__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2882 enum charge_type ctype) 2883{ 2884 if (mem_cgroup_disabled()) 2885 return; 2886 if (!ptr) 2887 return; 2888 cgroup_exclude_rmdir(&ptr->css); 2889 2890 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype); 2891 /* 2892 * Now swap is on-memory. This means this page may be 2893 * counted both as mem and swap....double count. 2894 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2895 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2896 * may call delete_from_swap_cache() before reach here. 2897 */ 2898 if (do_swap_account && PageSwapCache(page)) { 2899 swp_entry_t ent = {.val = page_private(page)}; 2900 unsigned short id; 2901 struct mem_cgroup *memcg; 2902 2903 id = swap_cgroup_record(ent, 0); 2904 rcu_read_lock(); 2905 memcg = mem_cgroup_lookup(id); 2906 if (memcg) { 2907 /* 2908 * This recorded memcg can be obsolete one. So, avoid 2909 * calling css_tryget 2910 */ 2911 if (!mem_cgroup_is_root(memcg)) 2912 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 2913 mem_cgroup_swap_statistics(memcg, false); 2914 mem_cgroup_put(memcg); 2915 } 2916 rcu_read_unlock(); 2917 } 2918 /* 2919 * At swapin, we may charge account against cgroup which has no tasks. 2920 * So, rmdir()->pre_destroy() can be called while we do this charge. 2921 * In that case, we need to call pre_destroy() again. check it here. 2922 */ 2923 cgroup_release_and_wakeup_rmdir(&ptr->css); 2924} 2925 2926void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) 2927{ 2928 __mem_cgroup_commit_charge_swapin(page, ptr, 2929 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2930} 2931 2932void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) 2933{ 2934 if (mem_cgroup_disabled()) 2935 return; 2936 if (!memcg) 2937 return; 2938 __mem_cgroup_cancel_charge(memcg, 1); 2939} 2940 2941static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, 2942 unsigned int nr_pages, 2943 const enum charge_type ctype) 2944{ 2945 struct memcg_batch_info *batch = NULL; 2946 bool uncharge_memsw = true; 2947 2948 /* If swapout, usage of swap doesn't decrease */ 2949 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2950 uncharge_memsw = false; 2951 2952 batch = ¤t->memcg_batch; 2953 /* 2954 * In usual, we do css_get() when we remember memcg pointer. 2955 * But in this case, we keep res->usage until end of a series of 2956 * uncharges. Then, it's ok to ignore memcg's refcnt. 2957 */ 2958 if (!batch->memcg) 2959 batch->memcg = memcg; 2960 /* 2961 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2962 * In those cases, all pages freed continuously can be expected to be in 2963 * the same cgroup and we have chance to coalesce uncharges. 2964 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2965 * because we want to do uncharge as soon as possible. 2966 */ 2967 2968 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2969 goto direct_uncharge; 2970 2971 if (nr_pages > 1) 2972 goto direct_uncharge; 2973 2974 /* 2975 * In typical case, batch->memcg == mem. This means we can 2976 * merge a series of uncharges to an uncharge of res_counter. 2977 * If not, we uncharge res_counter ony by one. 2978 */ 2979 if (batch->memcg != memcg) 2980 goto direct_uncharge; 2981 /* remember freed charge and uncharge it later */ 2982 batch->nr_pages++; 2983 if (uncharge_memsw) 2984 batch->memsw_nr_pages++; 2985 return; 2986direct_uncharge: 2987 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); 2988 if (uncharge_memsw) 2989 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); 2990 if (unlikely(batch->memcg != memcg)) 2991 memcg_oom_recover(memcg); 2992 return; 2993} 2994 2995/* 2996 * uncharge if !page_mapped(page) 2997 */ 2998static struct mem_cgroup * 2999__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 3000{ 3001 struct mem_cgroup *memcg = NULL; 3002 unsigned int nr_pages = 1; 3003 struct page_cgroup *pc; 3004 3005 if (mem_cgroup_disabled()) 3006 return NULL; 3007 3008 if (PageSwapCache(page)) 3009 return NULL; 3010 3011 if (PageTransHuge(page)) { 3012 nr_pages <<= compound_order(page); 3013 VM_BUG_ON(!PageTransHuge(page)); 3014 } 3015 /* 3016 * Check if our page_cgroup is valid 3017 */ 3018 pc = lookup_page_cgroup(page); 3019 if (unlikely(!pc || !PageCgroupUsed(pc))) 3020 return NULL; 3021 3022 lock_page_cgroup(pc); 3023 3024 memcg = pc->mem_cgroup; 3025 3026 if (!PageCgroupUsed(pc)) 3027 goto unlock_out; 3028 3029 switch (ctype) { 3030 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 3031 case MEM_CGROUP_CHARGE_TYPE_DROP: 3032 /* See mem_cgroup_prepare_migration() */ 3033 if (page_mapped(page) || PageCgroupMigration(pc)) 3034 goto unlock_out; 3035 break; 3036 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 3037 if (!PageAnon(page)) { /* Shared memory */ 3038 if (page->mapping && !page_is_file_cache(page)) 3039 goto unlock_out; 3040 } else if (page_mapped(page)) /* Anon */ 3041 goto unlock_out; 3042 break; 3043 default: 3044 break; 3045 } 3046 3047 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages); 3048 3049 ClearPageCgroupUsed(pc); 3050 /* 3051 * pc->mem_cgroup is not cleared here. It will be accessed when it's 3052 * freed from LRU. This is safe because uncharged page is expected not 3053 * to be reused (freed soon). Exception is SwapCache, it's handled by 3054 * special functions. 3055 */ 3056 3057 unlock_page_cgroup(pc); 3058 /* 3059 * even after unlock, we have memcg->res.usage here and this memcg 3060 * will never be freed. 3061 */ 3062 memcg_check_events(memcg, page); 3063 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 3064 mem_cgroup_swap_statistics(memcg, true); 3065 mem_cgroup_get(memcg); 3066 } 3067 if (!mem_cgroup_is_root(memcg)) 3068 mem_cgroup_do_uncharge(memcg, nr_pages, ctype); 3069 3070 return memcg; 3071 3072unlock_out: 3073 unlock_page_cgroup(pc); 3074 return NULL; 3075} 3076 3077void mem_cgroup_uncharge_page(struct page *page) 3078{ 3079 /* early check. */ 3080 if (page_mapped(page)) 3081 return; 3082 if (page->mapping && !PageAnon(page)) 3083 return; 3084 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 3085} 3086 3087void mem_cgroup_uncharge_cache_page(struct page *page) 3088{ 3089 VM_BUG_ON(page_mapped(page)); 3090 VM_BUG_ON(page->mapping); 3091 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 3092} 3093 3094/* 3095 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 3096 * In that cases, pages are freed continuously and we can expect pages 3097 * are in the same memcg. All these calls itself limits the number of 3098 * pages freed at once, then uncharge_start/end() is called properly. 3099 * This may be called prural(2) times in a context, 3100 */ 3101 3102void mem_cgroup_uncharge_start(void) 3103{ 3104 current->memcg_batch.do_batch++; 3105 /* We can do nest. */ 3106 if (current->memcg_batch.do_batch == 1) { 3107 current->memcg_batch.memcg = NULL; 3108 current->memcg_batch.nr_pages = 0; 3109 current->memcg_batch.memsw_nr_pages = 0; 3110 } 3111} 3112 3113void mem_cgroup_uncharge_end(void) 3114{ 3115 struct memcg_batch_info *batch = ¤t->memcg_batch; 3116 3117 if (!batch->do_batch) 3118 return; 3119 3120 batch->do_batch--; 3121 if (batch->do_batch) /* If stacked, do nothing. */ 3122 return; 3123 3124 if (!batch->memcg) 3125 return; 3126 /* 3127 * This "batch->memcg" is valid without any css_get/put etc... 3128 * bacause we hide charges behind us. 3129 */ 3130 if (batch->nr_pages) 3131 res_counter_uncharge(&batch->memcg->res, 3132 batch->nr_pages * PAGE_SIZE); 3133 if (batch->memsw_nr_pages) 3134 res_counter_uncharge(&batch->memcg->memsw, 3135 batch->memsw_nr_pages * PAGE_SIZE); 3136 memcg_oom_recover(batch->memcg); 3137 /* forget this pointer (for sanity check) */ 3138 batch->memcg = NULL; 3139} 3140 3141#ifdef CONFIG_SWAP 3142/* 3143 * called after __delete_from_swap_cache() and drop "page" account. 3144 * memcg information is recorded to swap_cgroup of "ent" 3145 */ 3146void 3147mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 3148{ 3149 struct mem_cgroup *memcg; 3150 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 3151 3152 if (!swapout) /* this was a swap cache but the swap is unused ! */ 3153 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 3154 3155 memcg = __mem_cgroup_uncharge_common(page, ctype); 3156 3157 /* 3158 * record memcg information, if swapout && memcg != NULL, 3159 * mem_cgroup_get() was called in uncharge(). 3160 */ 3161 if (do_swap_account && swapout && memcg) 3162 swap_cgroup_record(ent, css_id(&memcg->css)); 3163} 3164#endif 3165 3166#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3167/* 3168 * called from swap_entry_free(). remove record in swap_cgroup and 3169 * uncharge "memsw" account. 3170 */ 3171void mem_cgroup_uncharge_swap(swp_entry_t ent) 3172{ 3173 struct mem_cgroup *memcg; 3174 unsigned short id; 3175 3176 if (!do_swap_account) 3177 return; 3178 3179 id = swap_cgroup_record(ent, 0); 3180 rcu_read_lock(); 3181 memcg = mem_cgroup_lookup(id); 3182 if (memcg) { 3183 /* 3184 * We uncharge this because swap is freed. 3185 * This memcg can be obsolete one. We avoid calling css_tryget 3186 */ 3187 if (!mem_cgroup_is_root(memcg)) 3188 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 3189 mem_cgroup_swap_statistics(memcg, false); 3190 mem_cgroup_put(memcg); 3191 } 3192 rcu_read_unlock(); 3193} 3194 3195/** 3196 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 3197 * @entry: swap entry to be moved 3198 * @from: mem_cgroup which the entry is moved from 3199 * @to: mem_cgroup which the entry is moved to 3200 * @need_fixup: whether we should fixup res_counters and refcounts. 3201 * 3202 * It succeeds only when the swap_cgroup's record for this entry is the same 3203 * as the mem_cgroup's id of @from. 3204 * 3205 * Returns 0 on success, -EINVAL on failure. 3206 * 3207 * The caller must have charged to @to, IOW, called res_counter_charge() about 3208 * both res and memsw, and called css_get(). 3209 */ 3210static int mem_cgroup_move_swap_account(swp_entry_t entry, 3211 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3212{ 3213 unsigned short old_id, new_id; 3214 3215 old_id = css_id(&from->css); 3216 new_id = css_id(&to->css); 3217 3218 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 3219 mem_cgroup_swap_statistics(from, false); 3220 mem_cgroup_swap_statistics(to, true); 3221 /* 3222 * This function is only called from task migration context now. 3223 * It postpones res_counter and refcount handling till the end 3224 * of task migration(mem_cgroup_clear_mc()) for performance 3225 * improvement. But we cannot postpone mem_cgroup_get(to) 3226 * because if the process that has been moved to @to does 3227 * swap-in, the refcount of @to might be decreased to 0. 3228 */ 3229 mem_cgroup_get(to); 3230 if (need_fixup) { 3231 if (!mem_cgroup_is_root(from)) 3232 res_counter_uncharge(&from->memsw, PAGE_SIZE); 3233 mem_cgroup_put(from); 3234 /* 3235 * we charged both to->res and to->memsw, so we should 3236 * uncharge to->res. 3237 */ 3238 if (!mem_cgroup_is_root(to)) 3239 res_counter_uncharge(&to->res, PAGE_SIZE); 3240 } 3241 return 0; 3242 } 3243 return -EINVAL; 3244} 3245#else 3246static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 3247 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3248{ 3249 return -EINVAL; 3250} 3251#endif 3252 3253/* 3254 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 3255 * page belongs to. 3256 */ 3257int mem_cgroup_prepare_migration(struct page *page, 3258 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask) 3259{ 3260 struct mem_cgroup *memcg = NULL; 3261 struct page_cgroup *pc; 3262 enum charge_type ctype; 3263 int ret = 0; 3264 3265 *ptr = NULL; 3266 3267 VM_BUG_ON(PageTransHuge(page)); 3268 if (mem_cgroup_disabled()) 3269 return 0; 3270 3271 pc = lookup_page_cgroup(page); 3272 lock_page_cgroup(pc); 3273 if (PageCgroupUsed(pc)) { 3274 memcg = pc->mem_cgroup; 3275 css_get(&memcg->css); 3276 /* 3277 * At migrating an anonymous page, its mapcount goes down 3278 * to 0 and uncharge() will be called. But, even if it's fully 3279 * unmapped, migration may fail and this page has to be 3280 * charged again. We set MIGRATION flag here and delay uncharge 3281 * until end_migration() is called 3282 * 3283 * Corner Case Thinking 3284 * A) 3285 * When the old page was mapped as Anon and it's unmap-and-freed 3286 * while migration was ongoing. 3287 * If unmap finds the old page, uncharge() of it will be delayed 3288 * until end_migration(). If unmap finds a new page, it's 3289 * uncharged when it make mapcount to be 1->0. If unmap code 3290 * finds swap_migration_entry, the new page will not be mapped 3291 * and end_migration() will find it(mapcount==0). 3292 * 3293 * B) 3294 * When the old page was mapped but migraion fails, the kernel 3295 * remaps it. A charge for it is kept by MIGRATION flag even 3296 * if mapcount goes down to 0. We can do remap successfully 3297 * without charging it again. 3298 * 3299 * C) 3300 * The "old" page is under lock_page() until the end of 3301 * migration, so, the old page itself will not be swapped-out. 3302 * If the new page is swapped out before end_migraton, our 3303 * hook to usual swap-out path will catch the event. 3304 */ 3305 if (PageAnon(page)) 3306 SetPageCgroupMigration(pc); 3307 } 3308 unlock_page_cgroup(pc); 3309 /* 3310 * If the page is not charged at this point, 3311 * we return here. 3312 */ 3313 if (!memcg) 3314 return 0; 3315 3316 *ptr = memcg; 3317 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false); 3318 css_put(&memcg->css);/* drop extra refcnt */ 3319 if (ret || *ptr == NULL) { 3320 if (PageAnon(page)) { 3321 lock_page_cgroup(pc); 3322 ClearPageCgroupMigration(pc); 3323 unlock_page_cgroup(pc); 3324 /* 3325 * The old page may be fully unmapped while we kept it. 3326 */ 3327 mem_cgroup_uncharge_page(page); 3328 } 3329 return -ENOMEM; 3330 } 3331 /* 3332 * We charge new page before it's used/mapped. So, even if unlock_page() 3333 * is called before end_migration, we can catch all events on this new 3334 * page. In the case new page is migrated but not remapped, new page's 3335 * mapcount will be finally 0 and we call uncharge in end_migration(). 3336 */ 3337 pc = lookup_page_cgroup(newpage); 3338 if (PageAnon(page)) 3339 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 3340 else if (page_is_file_cache(page)) 3341 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 3342 else 3343 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3344 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 3345 return ret; 3346} 3347 3348/* remove redundant charge if migration failed*/ 3349void mem_cgroup_end_migration(struct mem_cgroup *memcg, 3350 struct page *oldpage, struct page *newpage, bool migration_ok) 3351{ 3352 struct page *used, *unused; 3353 struct page_cgroup *pc; 3354 3355 if (!memcg) 3356 return; 3357 /* blocks rmdir() */ 3358 cgroup_exclude_rmdir(&memcg->css); 3359 if (!migration_ok) { 3360 used = oldpage; 3361 unused = newpage; 3362 } else { 3363 used = newpage; 3364 unused = oldpage; 3365 } 3366 /* 3367 * We disallowed uncharge of pages under migration because mapcount 3368 * of the page goes down to zero, temporarly. 3369 * Clear the flag and check the page should be charged. 3370 */ 3371 pc = lookup_page_cgroup(oldpage); 3372 lock_page_cgroup(pc); 3373 ClearPageCgroupMigration(pc); 3374 unlock_page_cgroup(pc); 3375 3376 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 3377 3378 /* 3379 * If a page is a file cache, radix-tree replacement is very atomic 3380 * and we can skip this check. When it was an Anon page, its mapcount 3381 * goes down to 0. But because we added MIGRATION flage, it's not 3382 * uncharged yet. There are several case but page->mapcount check 3383 * and USED bit check in mem_cgroup_uncharge_page() will do enough 3384 * check. (see prepare_charge() also) 3385 */ 3386 if (PageAnon(used)) 3387 mem_cgroup_uncharge_page(used); 3388 /* 3389 * At migration, we may charge account against cgroup which has no 3390 * tasks. 3391 * So, rmdir()->pre_destroy() can be called while we do this charge. 3392 * In that case, we need to call pre_destroy() again. check it here. 3393 */ 3394 cgroup_release_and_wakeup_rmdir(&memcg->css); 3395} 3396 3397/* 3398 * At replace page cache, newpage is not under any memcg but it's on 3399 * LRU. So, this function doesn't touch res_counter but handles LRU 3400 * in correct way. Both pages are locked so we cannot race with uncharge. 3401 */ 3402void mem_cgroup_replace_page_cache(struct page *oldpage, 3403 struct page *newpage) 3404{ 3405 struct mem_cgroup *memcg; 3406 struct page_cgroup *pc; 3407 struct zone *zone; 3408 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; 3409 unsigned long flags; 3410 3411 if (mem_cgroup_disabled()) 3412 return; 3413 3414 pc = lookup_page_cgroup(oldpage); 3415 /* fix accounting on old pages */ 3416 lock_page_cgroup(pc); 3417 memcg = pc->mem_cgroup; 3418 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1); 3419 ClearPageCgroupUsed(pc); 3420 unlock_page_cgroup(pc); 3421 3422 if (PageSwapBacked(oldpage)) 3423 type = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3424 3425 zone = page_zone(newpage); 3426 pc = lookup_page_cgroup(newpage); 3427 /* 3428 * Even if newpage->mapping was NULL before starting replacement, 3429 * the newpage may be on LRU(or pagevec for LRU) already. We lock 3430 * LRU while we overwrite pc->mem_cgroup. 3431 */ 3432 spin_lock_irqsave(&zone->lru_lock, flags); 3433 if (PageLRU(newpage)) 3434 del_page_from_lru_list(zone, newpage, page_lru(newpage)); 3435 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type); 3436 if (PageLRU(newpage)) 3437 add_page_to_lru_list(zone, newpage, page_lru(newpage)); 3438 spin_unlock_irqrestore(&zone->lru_lock, flags); 3439} 3440 3441#ifdef CONFIG_DEBUG_VM 3442static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3443{ 3444 struct page_cgroup *pc; 3445 3446 pc = lookup_page_cgroup(page); 3447 if (likely(pc) && PageCgroupUsed(pc)) 3448 return pc; 3449 return NULL; 3450} 3451 3452bool mem_cgroup_bad_page_check(struct page *page) 3453{ 3454 if (mem_cgroup_disabled()) 3455 return false; 3456 3457 return lookup_page_cgroup_used(page) != NULL; 3458} 3459 3460void mem_cgroup_print_bad_page(struct page *page) 3461{ 3462 struct page_cgroup *pc; 3463 3464 pc = lookup_page_cgroup_used(page); 3465 if (pc) { 3466 int ret = -1; 3467 char *path; 3468 3469 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", 3470 pc, pc->flags, pc->mem_cgroup); 3471 3472 path = kmalloc(PATH_MAX, GFP_KERNEL); 3473 if (path) { 3474 rcu_read_lock(); 3475 ret = cgroup_path(pc->mem_cgroup->css.cgroup, 3476 path, PATH_MAX); 3477 rcu_read_unlock(); 3478 } 3479 3480 printk(KERN_CONT "(%s)\n", 3481 (ret < 0) ? "cannot get the path" : path); 3482 kfree(path); 3483 } 3484} 3485#endif 3486 3487static DEFINE_MUTEX(set_limit_mutex); 3488 3489static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3490 unsigned long long val) 3491{ 3492 int retry_count; 3493 u64 memswlimit, memlimit; 3494 int ret = 0; 3495 int children = mem_cgroup_count_children(memcg); 3496 u64 curusage, oldusage; 3497 int enlarge; 3498 3499 /* 3500 * For keeping hierarchical_reclaim simple, how long we should retry 3501 * is depends on callers. We set our retry-count to be function 3502 * of # of children which we should visit in this loop. 3503 */ 3504 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3505 3506 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3507 3508 enlarge = 0; 3509 while (retry_count) { 3510 if (signal_pending(current)) { 3511 ret = -EINTR; 3512 break; 3513 } 3514 /* 3515 * Rather than hide all in some function, I do this in 3516 * open coded manner. You see what this really does. 3517 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3518 */ 3519 mutex_lock(&set_limit_mutex); 3520 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3521 if (memswlimit < val) { 3522 ret = -EINVAL; 3523 mutex_unlock(&set_limit_mutex); 3524 break; 3525 } 3526 3527 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3528 if (memlimit < val) 3529 enlarge = 1; 3530 3531 ret = res_counter_set_limit(&memcg->res, val); 3532 if (!ret) { 3533 if (memswlimit == val) 3534 memcg->memsw_is_minimum = true; 3535 else 3536 memcg->memsw_is_minimum = false; 3537 } 3538 mutex_unlock(&set_limit_mutex); 3539 3540 if (!ret) 3541 break; 3542 3543 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3544 MEM_CGROUP_RECLAIM_SHRINK); 3545 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3546 /* Usage is reduced ? */ 3547 if (curusage >= oldusage) 3548 retry_count--; 3549 else 3550 oldusage = curusage; 3551 } 3552 if (!ret && enlarge) 3553 memcg_oom_recover(memcg); 3554 3555 return ret; 3556} 3557 3558static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3559 unsigned long long val) 3560{ 3561 int retry_count; 3562 u64 memlimit, memswlimit, oldusage, curusage; 3563 int children = mem_cgroup_count_children(memcg); 3564 int ret = -EBUSY; 3565 int enlarge = 0; 3566 3567 /* see mem_cgroup_resize_res_limit */ 3568 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3569 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3570 while (retry_count) { 3571 if (signal_pending(current)) { 3572 ret = -EINTR; 3573 break; 3574 } 3575 /* 3576 * Rather than hide all in some function, I do this in 3577 * open coded manner. You see what this really does. 3578 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3579 */ 3580 mutex_lock(&set_limit_mutex); 3581 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3582 if (memlimit > val) { 3583 ret = -EINVAL; 3584 mutex_unlock(&set_limit_mutex); 3585 break; 3586 } 3587 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3588 if (memswlimit < val) 3589 enlarge = 1; 3590 ret = res_counter_set_limit(&memcg->memsw, val); 3591 if (!ret) { 3592 if (memlimit == val) 3593 memcg->memsw_is_minimum = true; 3594 else 3595 memcg->memsw_is_minimum = false; 3596 } 3597 mutex_unlock(&set_limit_mutex); 3598 3599 if (!ret) 3600 break; 3601 3602 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3603 MEM_CGROUP_RECLAIM_NOSWAP | 3604 MEM_CGROUP_RECLAIM_SHRINK); 3605 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3606 /* Usage is reduced ? */ 3607 if (curusage >= oldusage) 3608 retry_count--; 3609 else 3610 oldusage = curusage; 3611 } 3612 if (!ret && enlarge) 3613 memcg_oom_recover(memcg); 3614 return ret; 3615} 3616 3617unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3618 gfp_t gfp_mask, 3619 unsigned long *total_scanned) 3620{ 3621 unsigned long nr_reclaimed = 0; 3622 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3623 unsigned long reclaimed; 3624 int loop = 0; 3625 struct mem_cgroup_tree_per_zone *mctz; 3626 unsigned long long excess; 3627 unsigned long nr_scanned; 3628 3629 if (order > 0) 3630 return 0; 3631 3632 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3633 /* 3634 * This loop can run a while, specially if mem_cgroup's continuously 3635 * keep exceeding their soft limit and putting the system under 3636 * pressure 3637 */ 3638 do { 3639 if (next_mz) 3640 mz = next_mz; 3641 else 3642 mz = mem_cgroup_largest_soft_limit_node(mctz); 3643 if (!mz) 3644 break; 3645 3646 nr_scanned = 0; 3647 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone, 3648 gfp_mask, &nr_scanned); 3649 nr_reclaimed += reclaimed; 3650 *total_scanned += nr_scanned; 3651 spin_lock(&mctz->lock); 3652 3653 /* 3654 * If we failed to reclaim anything from this memory cgroup 3655 * it is time to move on to the next cgroup 3656 */ 3657 next_mz = NULL; 3658 if (!reclaimed) { 3659 do { 3660 /* 3661 * Loop until we find yet another one. 3662 * 3663 * By the time we get the soft_limit lock 3664 * again, someone might have aded the 3665 * group back on the RB tree. Iterate to 3666 * make sure we get a different mem. 3667 * mem_cgroup_largest_soft_limit_node returns 3668 * NULL if no other cgroup is present on 3669 * the tree 3670 */ 3671 next_mz = 3672 __mem_cgroup_largest_soft_limit_node(mctz); 3673 if (next_mz == mz) 3674 css_put(&next_mz->mem->css); 3675 else /* next_mz == NULL or other memcg */ 3676 break; 3677 } while (1); 3678 } 3679 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3680 excess = res_counter_soft_limit_excess(&mz->mem->res); 3681 /* 3682 * One school of thought says that we should not add 3683 * back the node to the tree if reclaim returns 0. 3684 * But our reclaim could return 0, simply because due 3685 * to priority we are exposing a smaller subset of 3686 * memory to reclaim from. Consider this as a longer 3687 * term TODO. 3688 */ 3689 /* If excess == 0, no tree ops */ 3690 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3691 spin_unlock(&mctz->lock); 3692 css_put(&mz->mem->css); 3693 loop++; 3694 /* 3695 * Could not reclaim anything and there are no more 3696 * mem cgroups to try or we seem to be looping without 3697 * reclaiming anything. 3698 */ 3699 if (!nr_reclaimed && 3700 (next_mz == NULL || 3701 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3702 break; 3703 } while (!nr_reclaimed); 3704 if (next_mz) 3705 css_put(&next_mz->mem->css); 3706 return nr_reclaimed; 3707} 3708 3709/* 3710 * This routine traverse page_cgroup in given list and drop them all. 3711 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3712 */ 3713static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, 3714 int node, int zid, enum lru_list lru) 3715{ 3716 struct mem_cgroup_per_zone *mz; 3717 unsigned long flags, loop; 3718 struct list_head *list; 3719 struct page *busy; 3720 struct zone *zone; 3721 int ret = 0; 3722 3723 zone = &NODE_DATA(node)->node_zones[zid]; 3724 mz = mem_cgroup_zoneinfo(memcg, node, zid); 3725 list = &mz->lruvec.lists[lru]; 3726 3727 loop = MEM_CGROUP_ZSTAT(mz, lru); 3728 /* give some margin against EBUSY etc...*/ 3729 loop += 256; 3730 busy = NULL; 3731 while (loop--) { 3732 struct page_cgroup *pc; 3733 struct page *page; 3734 3735 ret = 0; 3736 spin_lock_irqsave(&zone->lru_lock, flags); 3737 if (list_empty(list)) { 3738 spin_unlock_irqrestore(&zone->lru_lock, flags); 3739 break; 3740 } 3741 page = list_entry(list->prev, struct page, lru); 3742 if (busy == page) { 3743 list_move(&page->lru, list); 3744 busy = NULL; 3745 spin_unlock_irqrestore(&zone->lru_lock, flags); 3746 continue; 3747 } 3748 spin_unlock_irqrestore(&zone->lru_lock, flags); 3749 3750 pc = lookup_page_cgroup(page); 3751 3752 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); 3753 if (ret == -ENOMEM) 3754 break; 3755 3756 if (ret == -EBUSY || ret == -EINVAL) { 3757 /* found lock contention or "pc" is obsolete. */ 3758 busy = page; 3759 cond_resched(); 3760 } else 3761 busy = NULL; 3762 } 3763 3764 if (!ret && !list_empty(list)) 3765 return -EBUSY; 3766 return ret; 3767} 3768 3769/* 3770 * make mem_cgroup's charge to be 0 if there is no task. 3771 * This enables deleting this mem_cgroup. 3772 */ 3773static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) 3774{ 3775 int ret; 3776 int node, zid, shrink; 3777 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3778 struct cgroup *cgrp = memcg->css.cgroup; 3779 3780 css_get(&memcg->css); 3781 3782 shrink = 0; 3783 /* should free all ? */ 3784 if (free_all) 3785 goto try_to_free; 3786move_account: 3787 do { 3788 ret = -EBUSY; 3789 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3790 goto out; 3791 ret = -EINTR; 3792 if (signal_pending(current)) 3793 goto out; 3794 /* This is for making all *used* pages to be on LRU. */ 3795 lru_add_drain_all(); 3796 drain_all_stock_sync(memcg); 3797 ret = 0; 3798 mem_cgroup_start_move(memcg); 3799 for_each_node_state(node, N_HIGH_MEMORY) { 3800 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3801 enum lru_list l; 3802 for_each_lru(l) { 3803 ret = mem_cgroup_force_empty_list(memcg, 3804 node, zid, l); 3805 if (ret) 3806 break; 3807 } 3808 } 3809 if (ret) 3810 break; 3811 } 3812 mem_cgroup_end_move(memcg); 3813 memcg_oom_recover(memcg); 3814 /* it seems parent cgroup doesn't have enough mem */ 3815 if (ret == -ENOMEM) 3816 goto try_to_free; 3817 cond_resched(); 3818 /* "ret" should also be checked to ensure all lists are empty. */ 3819 } while (memcg->res.usage > 0 || ret); 3820out: 3821 css_put(&memcg->css); 3822 return ret; 3823 3824try_to_free: 3825 /* returns EBUSY if there is a task or if we come here twice. */ 3826 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3827 ret = -EBUSY; 3828 goto out; 3829 } 3830 /* we call try-to-free pages for make this cgroup empty */ 3831 lru_add_drain_all(); 3832 /* try to free all pages in this cgroup */ 3833 shrink = 1; 3834 while (nr_retries && memcg->res.usage > 0) { 3835 int progress; 3836 3837 if (signal_pending(current)) { 3838 ret = -EINTR; 3839 goto out; 3840 } 3841 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, 3842 false); 3843 if (!progress) { 3844 nr_retries--; 3845 /* maybe some writeback is necessary */ 3846 congestion_wait(BLK_RW_ASYNC, HZ/10); 3847 } 3848 3849 } 3850 lru_add_drain(); 3851 /* try move_account...there may be some *locked* pages. */ 3852 goto move_account; 3853} 3854 3855int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3856{ 3857 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3858} 3859 3860 3861static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3862{ 3863 return mem_cgroup_from_cont(cont)->use_hierarchy; 3864} 3865 3866static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3867 u64 val) 3868{ 3869 int retval = 0; 3870 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3871 struct cgroup *parent = cont->parent; 3872 struct mem_cgroup *parent_memcg = NULL; 3873 3874 if (parent) 3875 parent_memcg = mem_cgroup_from_cont(parent); 3876 3877 cgroup_lock(); 3878 /* 3879 * If parent's use_hierarchy is set, we can't make any modifications 3880 * in the child subtrees. If it is unset, then the change can 3881 * occur, provided the current cgroup has no children. 3882 * 3883 * For the root cgroup, parent_mem is NULL, we allow value to be 3884 * set if there are no children. 3885 */ 3886 if ((!parent_memcg || !parent_memcg->use_hierarchy) && 3887 (val == 1 || val == 0)) { 3888 if (list_empty(&cont->children)) 3889 memcg->use_hierarchy = val; 3890 else 3891 retval = -EBUSY; 3892 } else 3893 retval = -EINVAL; 3894 cgroup_unlock(); 3895 3896 return retval; 3897} 3898 3899 3900static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, 3901 enum mem_cgroup_stat_index idx) 3902{ 3903 struct mem_cgroup *iter; 3904 long val = 0; 3905 3906 /* Per-cpu values can be negative, use a signed accumulator */ 3907 for_each_mem_cgroup_tree(iter, memcg) 3908 val += mem_cgroup_read_stat(iter, idx); 3909 3910 if (val < 0) /* race ? */ 3911 val = 0; 3912 return val; 3913} 3914 3915static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) 3916{ 3917 u64 val; 3918 3919 if (!mem_cgroup_is_root(memcg)) { 3920 if (!swap) 3921 return res_counter_read_u64(&memcg->res, RES_USAGE); 3922 else 3923 return res_counter_read_u64(&memcg->memsw, RES_USAGE); 3924 } 3925 3926 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); 3927 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); 3928 3929 if (swap) 3930 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 3931 3932 return val << PAGE_SHIFT; 3933} 3934 3935static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3936{ 3937 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3938 u64 val; 3939 int type, name; 3940 3941 type = MEMFILE_TYPE(cft->private); 3942 name = MEMFILE_ATTR(cft->private); 3943 switch (type) { 3944 case _MEM: 3945 if (name == RES_USAGE) 3946 val = mem_cgroup_usage(memcg, false); 3947 else 3948 val = res_counter_read_u64(&memcg->res, name); 3949 break; 3950 case _MEMSWAP: 3951 if (name == RES_USAGE) 3952 val = mem_cgroup_usage(memcg, true); 3953 else 3954 val = res_counter_read_u64(&memcg->memsw, name); 3955 break; 3956 default: 3957 BUG(); 3958 break; 3959 } 3960 return val; 3961} 3962/* 3963 * The user of this function is... 3964 * RES_LIMIT. 3965 */ 3966static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3967 const char *buffer) 3968{ 3969 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3970 int type, name; 3971 unsigned long long val; 3972 int ret; 3973 3974 type = MEMFILE_TYPE(cft->private); 3975 name = MEMFILE_ATTR(cft->private); 3976 switch (name) { 3977 case RES_LIMIT: 3978 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3979 ret = -EINVAL; 3980 break; 3981 } 3982 /* This function does all necessary parse...reuse it */ 3983 ret = res_counter_memparse_write_strategy(buffer, &val); 3984 if (ret) 3985 break; 3986 if (type == _MEM) 3987 ret = mem_cgroup_resize_limit(memcg, val); 3988 else 3989 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3990 break; 3991 case RES_SOFT_LIMIT: 3992 ret = res_counter_memparse_write_strategy(buffer, &val); 3993 if (ret) 3994 break; 3995 /* 3996 * For memsw, soft limits are hard to implement in terms 3997 * of semantics, for now, we support soft limits for 3998 * control without swap 3999 */ 4000 if (type == _MEM) 4001 ret = res_counter_set_soft_limit(&memcg->res, val); 4002 else 4003 ret = -EINVAL; 4004 break; 4005 default: 4006 ret = -EINVAL; /* should be BUG() ? */ 4007 break; 4008 } 4009 return ret; 4010} 4011 4012static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 4013 unsigned long long *mem_limit, unsigned long long *memsw_limit) 4014{ 4015 struct cgroup *cgroup; 4016 unsigned long long min_limit, min_memsw_limit, tmp; 4017 4018 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 4019 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4020 cgroup = memcg->css.cgroup; 4021 if (!memcg->use_hierarchy) 4022 goto out; 4023 4024 while (cgroup->parent) { 4025 cgroup = cgroup->parent; 4026 memcg = mem_cgroup_from_cont(cgroup); 4027 if (!memcg->use_hierarchy) 4028 break; 4029 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 4030 min_limit = min(min_limit, tmp); 4031 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4032 min_memsw_limit = min(min_memsw_limit, tmp); 4033 } 4034out: 4035 *mem_limit = min_limit; 4036 *memsw_limit = min_memsw_limit; 4037 return; 4038} 4039 4040static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 4041{ 4042 struct mem_cgroup *memcg; 4043 int type, name; 4044 4045 memcg = mem_cgroup_from_cont(cont); 4046 type = MEMFILE_TYPE(event); 4047 name = MEMFILE_ATTR(event); 4048 switch (name) { 4049 case RES_MAX_USAGE: 4050 if (type == _MEM) 4051 res_counter_reset_max(&memcg->res); 4052 else 4053 res_counter_reset_max(&memcg->memsw); 4054 break; 4055 case RES_FAILCNT: 4056 if (type == _MEM) 4057 res_counter_reset_failcnt(&memcg->res); 4058 else 4059 res_counter_reset_failcnt(&memcg->memsw); 4060 break; 4061 } 4062 4063 return 0; 4064} 4065 4066static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 4067 struct cftype *cft) 4068{ 4069 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 4070} 4071 4072#ifdef CONFIG_MMU 4073static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4074 struct cftype *cft, u64 val) 4075{ 4076 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4077 4078 if (val >= (1 << NR_MOVE_TYPE)) 4079 return -EINVAL; 4080 /* 4081 * We check this value several times in both in can_attach() and 4082 * attach(), so we need cgroup lock to prevent this value from being 4083 * inconsistent. 4084 */ 4085 cgroup_lock(); 4086 memcg->move_charge_at_immigrate = val; 4087 cgroup_unlock(); 4088 4089 return 0; 4090} 4091#else 4092static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4093 struct cftype *cft, u64 val) 4094{ 4095 return -ENOSYS; 4096} 4097#endif 4098 4099 4100/* For read statistics */ 4101enum { 4102 MCS_CACHE, 4103 MCS_RSS, 4104 MCS_FILE_MAPPED, 4105 MCS_PGPGIN, 4106 MCS_PGPGOUT, 4107 MCS_SWAP, 4108 MCS_PGFAULT, 4109 MCS_PGMAJFAULT, 4110 MCS_INACTIVE_ANON, 4111 MCS_ACTIVE_ANON, 4112 MCS_INACTIVE_FILE, 4113 MCS_ACTIVE_FILE, 4114 MCS_UNEVICTABLE, 4115 NR_MCS_STAT, 4116}; 4117 4118struct mcs_total_stat { 4119 s64 stat[NR_MCS_STAT]; 4120}; 4121 4122struct { 4123 char *local_name; 4124 char *total_name; 4125} memcg_stat_strings[NR_MCS_STAT] = { 4126 {"cache", "total_cache"}, 4127 {"rss", "total_rss"}, 4128 {"mapped_file", "total_mapped_file"}, 4129 {"pgpgin", "total_pgpgin"}, 4130 {"pgpgout", "total_pgpgout"}, 4131 {"swap", "total_swap"}, 4132 {"pgfault", "total_pgfault"}, 4133 {"pgmajfault", "total_pgmajfault"}, 4134 {"inactive_anon", "total_inactive_anon"}, 4135 {"active_anon", "total_active_anon"}, 4136 {"inactive_file", "total_inactive_file"}, 4137 {"active_file", "total_active_file"}, 4138 {"unevictable", "total_unevictable"} 4139}; 4140 4141 4142static void 4143mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4144{ 4145 s64 val; 4146 4147 /* per cpu stat */ 4148 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); 4149 s->stat[MCS_CACHE] += val * PAGE_SIZE; 4150 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); 4151 s->stat[MCS_RSS] += val * PAGE_SIZE; 4152 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); 4153 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 4154 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); 4155 s->stat[MCS_PGPGIN] += val; 4156 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); 4157 s->stat[MCS_PGPGOUT] += val; 4158 if (do_swap_account) { 4159 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 4160 s->stat[MCS_SWAP] += val * PAGE_SIZE; 4161 } 4162 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); 4163 s->stat[MCS_PGFAULT] += val; 4164 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); 4165 s->stat[MCS_PGMAJFAULT] += val; 4166 4167 /* per zone stat */ 4168 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); 4169 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 4170 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); 4171 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 4172 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); 4173 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 4174 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); 4175 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 4176 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); 4177 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 4178} 4179 4180static void 4181mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4182{ 4183 struct mem_cgroup *iter; 4184 4185 for_each_mem_cgroup_tree(iter, memcg) 4186 mem_cgroup_get_local_stat(iter, s); 4187} 4188 4189#ifdef CONFIG_NUMA 4190static int mem_control_numa_stat_show(struct seq_file *m, void *arg) 4191{ 4192 int nid; 4193 unsigned long total_nr, file_nr, anon_nr, unevictable_nr; 4194 unsigned long node_nr; 4195 struct cgroup *cont = m->private; 4196 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4197 4198 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL); 4199 seq_printf(m, "total=%lu", total_nr); 4200 for_each_node_state(nid, N_HIGH_MEMORY) { 4201 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL); 4202 seq_printf(m, " N%d=%lu", nid, node_nr); 4203 } 4204 seq_putc(m, '\n'); 4205 4206 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE); 4207 seq_printf(m, "file=%lu", file_nr); 4208 for_each_node_state(nid, N_HIGH_MEMORY) { 4209 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4210 LRU_ALL_FILE); 4211 seq_printf(m, " N%d=%lu", nid, node_nr); 4212 } 4213 seq_putc(m, '\n'); 4214 4215 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON); 4216 seq_printf(m, "anon=%lu", anon_nr); 4217 for_each_node_state(nid, N_HIGH_MEMORY) { 4218 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4219 LRU_ALL_ANON); 4220 seq_printf(m, " N%d=%lu", nid, node_nr); 4221 } 4222 seq_putc(m, '\n'); 4223 4224 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE)); 4225 seq_printf(m, "unevictable=%lu", unevictable_nr); 4226 for_each_node_state(nid, N_HIGH_MEMORY) { 4227 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4228 BIT(LRU_UNEVICTABLE)); 4229 seq_printf(m, " N%d=%lu", nid, node_nr); 4230 } 4231 seq_putc(m, '\n'); 4232 return 0; 4233} 4234#endif /* CONFIG_NUMA */ 4235 4236static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 4237 struct cgroup_map_cb *cb) 4238{ 4239 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4240 struct mcs_total_stat mystat; 4241 int i; 4242 4243 memset(&mystat, 0, sizeof(mystat)); 4244 mem_cgroup_get_local_stat(mem_cont, &mystat); 4245 4246 4247 for (i = 0; i < NR_MCS_STAT; i++) { 4248 if (i == MCS_SWAP && !do_swap_account) 4249 continue; 4250 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 4251 } 4252 4253 /* Hierarchical information */ 4254 { 4255 unsigned long long limit, memsw_limit; 4256 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 4257 cb->fill(cb, "hierarchical_memory_limit", limit); 4258 if (do_swap_account) 4259 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 4260 } 4261 4262 memset(&mystat, 0, sizeof(mystat)); 4263 mem_cgroup_get_total_stat(mem_cont, &mystat); 4264 for (i = 0; i < NR_MCS_STAT; i++) { 4265 if (i == MCS_SWAP && !do_swap_account) 4266 continue; 4267 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 4268 } 4269 4270#ifdef CONFIG_DEBUG_VM 4271 { 4272 int nid, zid; 4273 struct mem_cgroup_per_zone *mz; 4274 unsigned long recent_rotated[2] = {0, 0}; 4275 unsigned long recent_scanned[2] = {0, 0}; 4276 4277 for_each_online_node(nid) 4278 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 4279 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 4280 4281 recent_rotated[0] += 4282 mz->reclaim_stat.recent_rotated[0]; 4283 recent_rotated[1] += 4284 mz->reclaim_stat.recent_rotated[1]; 4285 recent_scanned[0] += 4286 mz->reclaim_stat.recent_scanned[0]; 4287 recent_scanned[1] += 4288 mz->reclaim_stat.recent_scanned[1]; 4289 } 4290 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 4291 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 4292 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 4293 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 4294 } 4295#endif 4296 4297 return 0; 4298} 4299 4300static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 4301{ 4302 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4303 4304 return mem_cgroup_swappiness(memcg); 4305} 4306 4307static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 4308 u64 val) 4309{ 4310 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4311 struct mem_cgroup *parent; 4312 4313 if (val > 100) 4314 return -EINVAL; 4315 4316 if (cgrp->parent == NULL) 4317 return -EINVAL; 4318 4319 parent = mem_cgroup_from_cont(cgrp->parent); 4320 4321 cgroup_lock(); 4322 4323 /* If under hierarchy, only empty-root can set this value */ 4324 if ((parent->use_hierarchy) || 4325 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4326 cgroup_unlock(); 4327 return -EINVAL; 4328 } 4329 4330 memcg->swappiness = val; 4331 4332 cgroup_unlock(); 4333 4334 return 0; 4335} 4336 4337static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 4338{ 4339 struct mem_cgroup_threshold_ary *t; 4340 u64 usage; 4341 int i; 4342 4343 rcu_read_lock(); 4344 if (!swap) 4345 t = rcu_dereference(memcg->thresholds.primary); 4346 else 4347 t = rcu_dereference(memcg->memsw_thresholds.primary); 4348 4349 if (!t) 4350 goto unlock; 4351 4352 usage = mem_cgroup_usage(memcg, swap); 4353 4354 /* 4355 * current_threshold points to threshold just below usage. 4356 * If it's not true, a threshold was crossed after last 4357 * call of __mem_cgroup_threshold(). 4358 */ 4359 i = t->current_threshold; 4360 4361 /* 4362 * Iterate backward over array of thresholds starting from 4363 * current_threshold and check if a threshold is crossed. 4364 * If none of thresholds below usage is crossed, we read 4365 * only one element of the array here. 4366 */ 4367 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 4368 eventfd_signal(t->entries[i].eventfd, 1); 4369 4370 /* i = current_threshold + 1 */ 4371 i++; 4372 4373 /* 4374 * Iterate forward over array of thresholds starting from 4375 * current_threshold+1 and check if a threshold is crossed. 4376 * If none of thresholds above usage is crossed, we read 4377 * only one element of the array here. 4378 */ 4379 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 4380 eventfd_signal(t->entries[i].eventfd, 1); 4381 4382 /* Update current_threshold */ 4383 t->current_threshold = i - 1; 4384unlock: 4385 rcu_read_unlock(); 4386} 4387 4388static void mem_cgroup_threshold(struct mem_cgroup *memcg) 4389{ 4390 while (memcg) { 4391 __mem_cgroup_threshold(memcg, false); 4392 if (do_swap_account) 4393 __mem_cgroup_threshold(memcg, true); 4394 4395 memcg = parent_mem_cgroup(memcg); 4396 } 4397} 4398 4399static int compare_thresholds(const void *a, const void *b) 4400{ 4401 const struct mem_cgroup_threshold *_a = a; 4402 const struct mem_cgroup_threshold *_b = b; 4403 4404 return _a->threshold - _b->threshold; 4405} 4406 4407static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) 4408{ 4409 struct mem_cgroup_eventfd_list *ev; 4410 4411 list_for_each_entry(ev, &memcg->oom_notify, list) 4412 eventfd_signal(ev->eventfd, 1); 4413 return 0; 4414} 4415 4416static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) 4417{ 4418 struct mem_cgroup *iter; 4419 4420 for_each_mem_cgroup_tree(iter, memcg) 4421 mem_cgroup_oom_notify_cb(iter); 4422} 4423 4424static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 4425 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4426{ 4427 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4428 struct mem_cgroup_thresholds *thresholds; 4429 struct mem_cgroup_threshold_ary *new; 4430 int type = MEMFILE_TYPE(cft->private); 4431 u64 threshold, usage; 4432 int i, size, ret; 4433 4434 ret = res_counter_memparse_write_strategy(args, &threshold); 4435 if (ret) 4436 return ret; 4437 4438 mutex_lock(&memcg->thresholds_lock); 4439 4440 if (type == _MEM) 4441 thresholds = &memcg->thresholds; 4442 else if (type == _MEMSWAP) 4443 thresholds = &memcg->memsw_thresholds; 4444 else 4445 BUG(); 4446 4447 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4448 4449 /* Check if a threshold crossed before adding a new one */ 4450 if (thresholds->primary) 4451 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4452 4453 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 4454 4455 /* Allocate memory for new array of thresholds */ 4456 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 4457 GFP_KERNEL); 4458 if (!new) { 4459 ret = -ENOMEM; 4460 goto unlock; 4461 } 4462 new->size = size; 4463 4464 /* Copy thresholds (if any) to new array */ 4465 if (thresholds->primary) { 4466 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 4467 sizeof(struct mem_cgroup_threshold)); 4468 } 4469 4470 /* Add new threshold */ 4471 new->entries[size - 1].eventfd = eventfd; 4472 new->entries[size - 1].threshold = threshold; 4473 4474 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4475 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4476 compare_thresholds, NULL); 4477 4478 /* Find current threshold */ 4479 new->current_threshold = -1; 4480 for (i = 0; i < size; i++) { 4481 if (new->entries[i].threshold < usage) { 4482 /* 4483 * new->current_threshold will not be used until 4484 * rcu_assign_pointer(), so it's safe to increment 4485 * it here. 4486 */ 4487 ++new->current_threshold; 4488 } 4489 } 4490 4491 /* Free old spare buffer and save old primary buffer as spare */ 4492 kfree(thresholds->spare); 4493 thresholds->spare = thresholds->primary; 4494 4495 rcu_assign_pointer(thresholds->primary, new); 4496 4497 /* To be sure that nobody uses thresholds */ 4498 synchronize_rcu(); 4499 4500unlock: 4501 mutex_unlock(&memcg->thresholds_lock); 4502 4503 return ret; 4504} 4505 4506static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4507 struct cftype *cft, struct eventfd_ctx *eventfd) 4508{ 4509 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4510 struct mem_cgroup_thresholds *thresholds; 4511 struct mem_cgroup_threshold_ary *new; 4512 int type = MEMFILE_TYPE(cft->private); 4513 u64 usage; 4514 int i, j, size; 4515 4516 mutex_lock(&memcg->thresholds_lock); 4517 if (type == _MEM) 4518 thresholds = &memcg->thresholds; 4519 else if (type == _MEMSWAP) 4520 thresholds = &memcg->memsw_thresholds; 4521 else 4522 BUG(); 4523 4524 /* 4525 * Something went wrong if we trying to unregister a threshold 4526 * if we don't have thresholds 4527 */ 4528 BUG_ON(!thresholds); 4529 4530 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4531 4532 /* Check if a threshold crossed before removing */ 4533 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4534 4535 /* Calculate new number of threshold */ 4536 size = 0; 4537 for (i = 0; i < thresholds->primary->size; i++) { 4538 if (thresholds->primary->entries[i].eventfd != eventfd) 4539 size++; 4540 } 4541 4542 new = thresholds->spare; 4543 4544 /* Set thresholds array to NULL if we don't have thresholds */ 4545 if (!size) { 4546 kfree(new); 4547 new = NULL; 4548 goto swap_buffers; 4549 } 4550 4551 new->size = size; 4552 4553 /* Copy thresholds and find current threshold */ 4554 new->current_threshold = -1; 4555 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4556 if (thresholds->primary->entries[i].eventfd == eventfd) 4557 continue; 4558 4559 new->entries[j] = thresholds->primary->entries[i]; 4560 if (new->entries[j].threshold < usage) { 4561 /* 4562 * new->current_threshold will not be used 4563 * until rcu_assign_pointer(), so it's safe to increment 4564 * it here. 4565 */ 4566 ++new->current_threshold; 4567 } 4568 j++; 4569 } 4570 4571swap_buffers: 4572 /* Swap primary and spare array */ 4573 thresholds->spare = thresholds->primary; 4574 rcu_assign_pointer(thresholds->primary, new); 4575 4576 /* To be sure that nobody uses thresholds */ 4577 synchronize_rcu(); 4578 4579 mutex_unlock(&memcg->thresholds_lock); 4580} 4581 4582static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4583 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4584{ 4585 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4586 struct mem_cgroup_eventfd_list *event; 4587 int type = MEMFILE_TYPE(cft->private); 4588 4589 BUG_ON(type != _OOM_TYPE); 4590 event = kmalloc(sizeof(*event), GFP_KERNEL); 4591 if (!event) 4592 return -ENOMEM; 4593 4594 spin_lock(&memcg_oom_lock); 4595 4596 event->eventfd = eventfd; 4597 list_add(&event->list, &memcg->oom_notify); 4598 4599 /* already in OOM ? */ 4600 if (atomic_read(&memcg->under_oom)) 4601 eventfd_signal(eventfd, 1); 4602 spin_unlock(&memcg_oom_lock); 4603 4604 return 0; 4605} 4606 4607static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4608 struct cftype *cft, struct eventfd_ctx *eventfd) 4609{ 4610 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4611 struct mem_cgroup_eventfd_list *ev, *tmp; 4612 int type = MEMFILE_TYPE(cft->private); 4613 4614 BUG_ON(type != _OOM_TYPE); 4615 4616 spin_lock(&memcg_oom_lock); 4617 4618 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { 4619 if (ev->eventfd == eventfd) { 4620 list_del(&ev->list); 4621 kfree(ev); 4622 } 4623 } 4624 4625 spin_unlock(&memcg_oom_lock); 4626} 4627 4628static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4629 struct cftype *cft, struct cgroup_map_cb *cb) 4630{ 4631 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4632 4633 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); 4634 4635 if (atomic_read(&memcg->under_oom)) 4636 cb->fill(cb, "under_oom", 1); 4637 else 4638 cb->fill(cb, "under_oom", 0); 4639 return 0; 4640} 4641 4642static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4643 struct cftype *cft, u64 val) 4644{ 4645 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4646 struct mem_cgroup *parent; 4647 4648 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4649 if (!cgrp->parent || !((val == 0) || (val == 1))) 4650 return -EINVAL; 4651 4652 parent = mem_cgroup_from_cont(cgrp->parent); 4653 4654 cgroup_lock(); 4655 /* oom-kill-disable is a flag for subhierarchy. */ 4656 if ((parent->use_hierarchy) || 4657 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4658 cgroup_unlock(); 4659 return -EINVAL; 4660 } 4661 memcg->oom_kill_disable = val; 4662 if (!val) 4663 memcg_oom_recover(memcg); 4664 cgroup_unlock(); 4665 return 0; 4666} 4667 4668#ifdef CONFIG_NUMA 4669static const struct file_operations mem_control_numa_stat_file_operations = { 4670 .read = seq_read, 4671 .llseek = seq_lseek, 4672 .release = single_release, 4673}; 4674 4675static int mem_control_numa_stat_open(struct inode *unused, struct file *file) 4676{ 4677 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; 4678 4679 file->f_op = &mem_control_numa_stat_file_operations; 4680 return single_open(file, mem_control_numa_stat_show, cont); 4681} 4682#endif /* CONFIG_NUMA */ 4683 4684#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 4685static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4686{ 4687 /* 4688 * Part of this would be better living in a separate allocation 4689 * function, leaving us with just the cgroup tree population work. 4690 * We, however, depend on state such as network's proto_list that 4691 * is only initialized after cgroup creation. I found the less 4692 * cumbersome way to deal with it to defer it all to populate time 4693 */ 4694 return mem_cgroup_sockets_init(cont, ss); 4695}; 4696 4697static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4698 struct cgroup *cont) 4699{ 4700 mem_cgroup_sockets_destroy(cont, ss); 4701} 4702#else 4703static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4704{ 4705 return 0; 4706} 4707 4708static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4709 struct cgroup *cont) 4710{ 4711} 4712#endif 4713 4714static struct cftype mem_cgroup_files[] = { 4715 { 4716 .name = "usage_in_bytes", 4717 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4718 .read_u64 = mem_cgroup_read, 4719 .register_event = mem_cgroup_usage_register_event, 4720 .unregister_event = mem_cgroup_usage_unregister_event, 4721 }, 4722 { 4723 .name = "max_usage_in_bytes", 4724 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4725 .trigger = mem_cgroup_reset, 4726 .read_u64 = mem_cgroup_read, 4727 }, 4728 { 4729 .name = "limit_in_bytes", 4730 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4731 .write_string = mem_cgroup_write, 4732 .read_u64 = mem_cgroup_read, 4733 }, 4734 { 4735 .name = "soft_limit_in_bytes", 4736 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4737 .write_string = mem_cgroup_write, 4738 .read_u64 = mem_cgroup_read, 4739 }, 4740 { 4741 .name = "failcnt", 4742 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4743 .trigger = mem_cgroup_reset, 4744 .read_u64 = mem_cgroup_read, 4745 }, 4746 { 4747 .name = "stat", 4748 .read_map = mem_control_stat_show, 4749 }, 4750 { 4751 .name = "force_empty", 4752 .trigger = mem_cgroup_force_empty_write, 4753 }, 4754 { 4755 .name = "use_hierarchy", 4756 .write_u64 = mem_cgroup_hierarchy_write, 4757 .read_u64 = mem_cgroup_hierarchy_read, 4758 }, 4759 { 4760 .name = "swappiness", 4761 .read_u64 = mem_cgroup_swappiness_read, 4762 .write_u64 = mem_cgroup_swappiness_write, 4763 }, 4764 { 4765 .name = "move_charge_at_immigrate", 4766 .read_u64 = mem_cgroup_move_charge_read, 4767 .write_u64 = mem_cgroup_move_charge_write, 4768 }, 4769 { 4770 .name = "oom_control", 4771 .read_map = mem_cgroup_oom_control_read, 4772 .write_u64 = mem_cgroup_oom_control_write, 4773 .register_event = mem_cgroup_oom_register_event, 4774 .unregister_event = mem_cgroup_oom_unregister_event, 4775 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4776 }, 4777#ifdef CONFIG_NUMA 4778 { 4779 .name = "numa_stat", 4780 .open = mem_control_numa_stat_open, 4781 .mode = S_IRUGO, 4782 }, 4783#endif 4784}; 4785 4786#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4787static struct cftype memsw_cgroup_files[] = { 4788 { 4789 .name = "memsw.usage_in_bytes", 4790 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4791 .read_u64 = mem_cgroup_read, 4792 .register_event = mem_cgroup_usage_register_event, 4793 .unregister_event = mem_cgroup_usage_unregister_event, 4794 }, 4795 { 4796 .name = "memsw.max_usage_in_bytes", 4797 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4798 .trigger = mem_cgroup_reset, 4799 .read_u64 = mem_cgroup_read, 4800 }, 4801 { 4802 .name = "memsw.limit_in_bytes", 4803 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4804 .write_string = mem_cgroup_write, 4805 .read_u64 = mem_cgroup_read, 4806 }, 4807 { 4808 .name = "memsw.failcnt", 4809 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4810 .trigger = mem_cgroup_reset, 4811 .read_u64 = mem_cgroup_read, 4812 }, 4813}; 4814 4815static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4816{ 4817 if (!do_swap_account) 4818 return 0; 4819 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4820 ARRAY_SIZE(memsw_cgroup_files)); 4821}; 4822#else 4823static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4824{ 4825 return 0; 4826} 4827#endif 4828 4829static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4830{ 4831 struct mem_cgroup_per_node *pn; 4832 struct mem_cgroup_per_zone *mz; 4833 enum lru_list l; 4834 int zone, tmp = node; 4835 /* 4836 * This routine is called against possible nodes. 4837 * But it's BUG to call kmalloc() against offline node. 4838 * 4839 * TODO: this routine can waste much memory for nodes which will 4840 * never be onlined. It's better to use memory hotplug callback 4841 * function. 4842 */ 4843 if (!node_state(node, N_NORMAL_MEMORY)) 4844 tmp = -1; 4845 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4846 if (!pn) 4847 return 1; 4848 4849 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4850 mz = &pn->zoneinfo[zone]; 4851 for_each_lru(l) 4852 INIT_LIST_HEAD(&mz->lruvec.lists[l]); 4853 mz->usage_in_excess = 0; 4854 mz->on_tree = false; 4855 mz->mem = memcg; 4856 } 4857 memcg->info.nodeinfo[node] = pn; 4858 return 0; 4859} 4860 4861static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4862{ 4863 kfree(memcg->info.nodeinfo[node]); 4864} 4865 4866static struct mem_cgroup *mem_cgroup_alloc(void) 4867{ 4868 struct mem_cgroup *mem; 4869 int size = sizeof(struct mem_cgroup); 4870 4871 /* Can be very big if MAX_NUMNODES is very big */ 4872 if (size < PAGE_SIZE) 4873 mem = kzalloc(size, GFP_KERNEL); 4874 else 4875 mem = vzalloc(size); 4876 4877 if (!mem) 4878 return NULL; 4879 4880 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4881 if (!mem->stat) 4882 goto out_free; 4883 spin_lock_init(&mem->pcp_counter_lock); 4884 return mem; 4885 4886out_free: 4887 if (size < PAGE_SIZE) 4888 kfree(mem); 4889 else 4890 vfree(mem); 4891 return NULL; 4892} 4893 4894/* 4895 * At destroying mem_cgroup, references from swap_cgroup can remain. 4896 * (scanning all at force_empty is too costly...) 4897 * 4898 * Instead of clearing all references at force_empty, we remember 4899 * the number of reference from swap_cgroup and free mem_cgroup when 4900 * it goes down to 0. 4901 * 4902 * Removal of cgroup itself succeeds regardless of refs from swap. 4903 */ 4904 4905static void __mem_cgroup_free(struct mem_cgroup *memcg) 4906{ 4907 int node; 4908 4909 mem_cgroup_remove_from_trees(memcg); 4910 free_css_id(&mem_cgroup_subsys, &memcg->css); 4911 4912 for_each_node_state(node, N_POSSIBLE) 4913 free_mem_cgroup_per_zone_info(memcg, node); 4914 4915 free_percpu(memcg->stat); 4916 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4917 kfree(memcg); 4918 else 4919 vfree(memcg); 4920} 4921 4922static void mem_cgroup_get(struct mem_cgroup *memcg) 4923{ 4924 atomic_inc(&memcg->refcnt); 4925} 4926 4927static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) 4928{ 4929 if (atomic_sub_and_test(count, &memcg->refcnt)) { 4930 struct mem_cgroup *parent = parent_mem_cgroup(memcg); 4931 __mem_cgroup_free(memcg); 4932 if (parent) 4933 mem_cgroup_put(parent); 4934 } 4935} 4936 4937static void mem_cgroup_put(struct mem_cgroup *memcg) 4938{ 4939 __mem_cgroup_put(memcg, 1); 4940} 4941 4942/* 4943 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4944 */ 4945struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) 4946{ 4947 if (!memcg->res.parent) 4948 return NULL; 4949 return mem_cgroup_from_res_counter(memcg->res.parent, res); 4950} 4951EXPORT_SYMBOL(parent_mem_cgroup); 4952 4953#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4954static void __init enable_swap_cgroup(void) 4955{ 4956 if (!mem_cgroup_disabled() && really_do_swap_account) 4957 do_swap_account = 1; 4958} 4959#else 4960static void __init enable_swap_cgroup(void) 4961{ 4962} 4963#endif 4964 4965static int mem_cgroup_soft_limit_tree_init(void) 4966{ 4967 struct mem_cgroup_tree_per_node *rtpn; 4968 struct mem_cgroup_tree_per_zone *rtpz; 4969 int tmp, node, zone; 4970 4971 for_each_node_state(node, N_POSSIBLE) { 4972 tmp = node; 4973 if (!node_state(node, N_NORMAL_MEMORY)) 4974 tmp = -1; 4975 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4976 if (!rtpn) 4977 return 1; 4978 4979 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4980 4981 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4982 rtpz = &rtpn->rb_tree_per_zone[zone]; 4983 rtpz->rb_root = RB_ROOT; 4984 spin_lock_init(&rtpz->lock); 4985 } 4986 } 4987 return 0; 4988} 4989 4990static struct cgroup_subsys_state * __ref 4991mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4992{ 4993 struct mem_cgroup *memcg, *parent; 4994 long error = -ENOMEM; 4995 int node; 4996 4997 memcg = mem_cgroup_alloc(); 4998 if (!memcg) 4999 return ERR_PTR(error); 5000 5001 for_each_node_state(node, N_POSSIBLE) 5002 if (alloc_mem_cgroup_per_zone_info(memcg, node)) 5003 goto free_out; 5004 5005 /* root ? */ 5006 if (cont->parent == NULL) { 5007 int cpu; 5008 enable_swap_cgroup(); 5009 parent = NULL; 5010 if (mem_cgroup_soft_limit_tree_init()) 5011 goto free_out; 5012 root_mem_cgroup = memcg; 5013 for_each_possible_cpu(cpu) { 5014 struct memcg_stock_pcp *stock = 5015 &per_cpu(memcg_stock, cpu); 5016 INIT_WORK(&stock->work, drain_local_stock); 5017 } 5018 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 5019 } else { 5020 parent = mem_cgroup_from_cont(cont->parent); 5021 memcg->use_hierarchy = parent->use_hierarchy; 5022 memcg->oom_kill_disable = parent->oom_kill_disable; 5023 } 5024 5025 if (parent && parent->use_hierarchy) { 5026 res_counter_init(&memcg->res, &parent->res); 5027 res_counter_init(&memcg->memsw, &parent->memsw); 5028 /* 5029 * We increment refcnt of the parent to ensure that we can 5030 * safely access it on res_counter_charge/uncharge. 5031 * This refcnt will be decremented when freeing this 5032 * mem_cgroup(see mem_cgroup_put). 5033 */ 5034 mem_cgroup_get(parent); 5035 } else { 5036 res_counter_init(&memcg->res, NULL); 5037 res_counter_init(&memcg->memsw, NULL); 5038 } 5039 memcg->last_scanned_node = MAX_NUMNODES; 5040 INIT_LIST_HEAD(&memcg->oom_notify); 5041 5042 if (parent) 5043 memcg->swappiness = mem_cgroup_swappiness(parent); 5044 atomic_set(&memcg->refcnt, 1); 5045 memcg->move_charge_at_immigrate = 0; 5046 mutex_init(&memcg->thresholds_lock); 5047 return &memcg->css; 5048free_out: 5049 __mem_cgroup_free(memcg); 5050 return ERR_PTR(error); 5051} 5052 5053static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 5054 struct cgroup *cont) 5055{ 5056 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5057 5058 return mem_cgroup_force_empty(memcg, false); 5059} 5060 5061static void mem_cgroup_destroy(struct cgroup_subsys *ss, 5062 struct cgroup *cont) 5063{ 5064 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5065 5066 kmem_cgroup_destroy(ss, cont); 5067 5068 mem_cgroup_put(memcg); 5069} 5070 5071static int mem_cgroup_populate(struct cgroup_subsys *ss, 5072 struct cgroup *cont) 5073{ 5074 int ret; 5075 5076 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 5077 ARRAY_SIZE(mem_cgroup_files)); 5078 5079 if (!ret) 5080 ret = register_memsw_files(cont, ss); 5081 5082 if (!ret) 5083 ret = register_kmem_files(cont, ss); 5084 5085 return ret; 5086} 5087 5088#ifdef CONFIG_MMU 5089/* Handlers for move charge at task migration. */ 5090#define PRECHARGE_COUNT_AT_ONCE 256 5091static int mem_cgroup_do_precharge(unsigned long count) 5092{ 5093 int ret = 0; 5094 int batch_count = PRECHARGE_COUNT_AT_ONCE; 5095 struct mem_cgroup *memcg = mc.to; 5096 5097 if (mem_cgroup_is_root(memcg)) { 5098 mc.precharge += count; 5099 /* we don't need css_get for root */ 5100 return ret; 5101 } 5102 /* try to charge at once */ 5103 if (count > 1) { 5104 struct res_counter *dummy; 5105 /* 5106 * "memcg" cannot be under rmdir() because we've already checked 5107 * by cgroup_lock_live_cgroup() that it is not removed and we 5108 * are still under the same cgroup_mutex. So we can postpone 5109 * css_get(). 5110 */ 5111 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) 5112 goto one_by_one; 5113 if (do_swap_account && res_counter_charge(&memcg->memsw, 5114 PAGE_SIZE * count, &dummy)) { 5115 res_counter_uncharge(&memcg->res, PAGE_SIZE * count); 5116 goto one_by_one; 5117 } 5118 mc.precharge += count; 5119 return ret; 5120 } 5121one_by_one: 5122 /* fall back to one by one charge */ 5123 while (count--) { 5124 if (signal_pending(current)) { 5125 ret = -EINTR; 5126 break; 5127 } 5128 if (!batch_count--) { 5129 batch_count = PRECHARGE_COUNT_AT_ONCE; 5130 cond_resched(); 5131 } 5132 ret = __mem_cgroup_try_charge(NULL, 5133 GFP_KERNEL, 1, &memcg, false); 5134 if (ret || !memcg) 5135 /* mem_cgroup_clear_mc() will do uncharge later */ 5136 return -ENOMEM; 5137 mc.precharge++; 5138 } 5139 return ret; 5140} 5141 5142/** 5143 * is_target_pte_for_mc - check a pte whether it is valid for move charge 5144 * @vma: the vma the pte to be checked belongs 5145 * @addr: the address corresponding to the pte to be checked 5146 * @ptent: the pte to be checked 5147 * @target: the pointer the target page or swap ent will be stored(can be NULL) 5148 * 5149 * Returns 5150 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 5151 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 5152 * move charge. if @target is not NULL, the page is stored in target->page 5153 * with extra refcnt got(Callers should handle it). 5154 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 5155 * target for charge migration. if @target is not NULL, the entry is stored 5156 * in target->ent. 5157 * 5158 * Called with pte lock held. 5159 */ 5160union mc_target { 5161 struct page *page; 5162 swp_entry_t ent; 5163}; 5164 5165enum mc_target_type { 5166 MC_TARGET_NONE, /* not used */ 5167 MC_TARGET_PAGE, 5168 MC_TARGET_SWAP, 5169}; 5170 5171static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 5172 unsigned long addr, pte_t ptent) 5173{ 5174 struct page *page = vm_normal_page(vma, addr, ptent); 5175 5176 if (!page || !page_mapped(page)) 5177 return NULL; 5178 if (PageAnon(page)) { 5179 /* we don't move shared anon */ 5180 if (!move_anon() || page_mapcount(page) > 2) 5181 return NULL; 5182 } else if (!move_file()) 5183 /* we ignore mapcount for file pages */ 5184 return NULL; 5185 if (!get_page_unless_zero(page)) 5186 return NULL; 5187 5188 return page; 5189} 5190 5191static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 5192 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5193{ 5194 int usage_count; 5195 struct page *page = NULL; 5196 swp_entry_t ent = pte_to_swp_entry(ptent); 5197 5198 if (!move_anon() || non_swap_entry(ent)) 5199 return NULL; 5200 usage_count = mem_cgroup_count_swap_user(ent, &page); 5201 if (usage_count > 1) { /* we don't move shared anon */ 5202 if (page) 5203 put_page(page); 5204 return NULL; 5205 } 5206 if (do_swap_account) 5207 entry->val = ent.val; 5208 5209 return page; 5210} 5211 5212static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 5213 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5214{ 5215 struct page *page = NULL; 5216 struct inode *inode; 5217 struct address_space *mapping; 5218 pgoff_t pgoff; 5219 5220 if (!vma->vm_file) /* anonymous vma */ 5221 return NULL; 5222 if (!move_file()) 5223 return NULL; 5224 5225 inode = vma->vm_file->f_path.dentry->d_inode; 5226 mapping = vma->vm_file->f_mapping; 5227 if (pte_none(ptent)) 5228 pgoff = linear_page_index(vma, addr); 5229 else /* pte_file(ptent) is true */ 5230 pgoff = pte_to_pgoff(ptent); 5231 5232 /* page is moved even if it's not RSS of this task(page-faulted). */ 5233 page = find_get_page(mapping, pgoff); 5234 5235#ifdef CONFIG_SWAP 5236 /* shmem/tmpfs may report page out on swap: account for that too. */ 5237 if (radix_tree_exceptional_entry(page)) { 5238 swp_entry_t swap = radix_to_swp_entry(page); 5239 if (do_swap_account) 5240 *entry = swap; 5241 page = find_get_page(&swapper_space, swap.val); 5242 } 5243#endif 5244 return page; 5245} 5246 5247static int is_target_pte_for_mc(struct vm_area_struct *vma, 5248 unsigned long addr, pte_t ptent, union mc_target *target) 5249{ 5250 struct page *page = NULL; 5251 struct page_cgroup *pc; 5252 int ret = 0; 5253 swp_entry_t ent = { .val = 0 }; 5254 5255 if (pte_present(ptent)) 5256 page = mc_handle_present_pte(vma, addr, ptent); 5257 else if (is_swap_pte(ptent)) 5258 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 5259 else if (pte_none(ptent) || pte_file(ptent)) 5260 page = mc_handle_file_pte(vma, addr, ptent, &ent); 5261 5262 if (!page && !ent.val) 5263 return 0; 5264 if (page) { 5265 pc = lookup_page_cgroup(page); 5266 /* 5267 * Do only loose check w/o page_cgroup lock. 5268 * mem_cgroup_move_account() checks the pc is valid or not under 5269 * the lock. 5270 */ 5271 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 5272 ret = MC_TARGET_PAGE; 5273 if (target) 5274 target->page = page; 5275 } 5276 if (!ret || !target) 5277 put_page(page); 5278 } 5279 /* There is a swap entry and a page doesn't exist or isn't charged */ 5280 if (ent.val && !ret && 5281 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { 5282 ret = MC_TARGET_SWAP; 5283 if (target) 5284 target->ent = ent; 5285 } 5286 return ret; 5287} 5288 5289static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 5290 unsigned long addr, unsigned long end, 5291 struct mm_walk *walk) 5292{ 5293 struct vm_area_struct *vma = walk->private; 5294 pte_t *pte; 5295 spinlock_t *ptl; 5296 5297 split_huge_page_pmd(walk->mm, pmd); 5298 5299 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5300 for (; addr != end; pte++, addr += PAGE_SIZE) 5301 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 5302 mc.precharge++; /* increment precharge temporarily */ 5303 pte_unmap_unlock(pte - 1, ptl); 5304 cond_resched(); 5305 5306 return 0; 5307} 5308 5309static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 5310{ 5311 unsigned long precharge; 5312 struct vm_area_struct *vma; 5313 5314 down_read(&mm->mmap_sem); 5315 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5316 struct mm_walk mem_cgroup_count_precharge_walk = { 5317 .pmd_entry = mem_cgroup_count_precharge_pte_range, 5318 .mm = mm, 5319 .private = vma, 5320 }; 5321 if (is_vm_hugetlb_page(vma)) 5322 continue; 5323 walk_page_range(vma->vm_start, vma->vm_end, 5324 &mem_cgroup_count_precharge_walk); 5325 } 5326 up_read(&mm->mmap_sem); 5327 5328 precharge = mc.precharge; 5329 mc.precharge = 0; 5330 5331 return precharge; 5332} 5333 5334static int mem_cgroup_precharge_mc(struct mm_struct *mm) 5335{ 5336 unsigned long precharge = mem_cgroup_count_precharge(mm); 5337 5338 VM_BUG_ON(mc.moving_task); 5339 mc.moving_task = current; 5340 return mem_cgroup_do_precharge(precharge); 5341} 5342 5343/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 5344static void __mem_cgroup_clear_mc(void) 5345{ 5346 struct mem_cgroup *from = mc.from; 5347 struct mem_cgroup *to = mc.to; 5348 5349 /* we must uncharge all the leftover precharges from mc.to */ 5350 if (mc.precharge) { 5351 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 5352 mc.precharge = 0; 5353 } 5354 /* 5355 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 5356 * we must uncharge here. 5357 */ 5358 if (mc.moved_charge) { 5359 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 5360 mc.moved_charge = 0; 5361 } 5362 /* we must fixup refcnts and charges */ 5363 if (mc.moved_swap) { 5364 /* uncharge swap account from the old cgroup */ 5365 if (!mem_cgroup_is_root(mc.from)) 5366 res_counter_uncharge(&mc.from->memsw, 5367 PAGE_SIZE * mc.moved_swap); 5368 __mem_cgroup_put(mc.from, mc.moved_swap); 5369 5370 if (!mem_cgroup_is_root(mc.to)) { 5371 /* 5372 * we charged both to->res and to->memsw, so we should 5373 * uncharge to->res. 5374 */ 5375 res_counter_uncharge(&mc.to->res, 5376 PAGE_SIZE * mc.moved_swap); 5377 } 5378 /* we've already done mem_cgroup_get(mc.to) */ 5379 mc.moved_swap = 0; 5380 } 5381 memcg_oom_recover(from); 5382 memcg_oom_recover(to); 5383 wake_up_all(&mc.waitq); 5384} 5385 5386static void mem_cgroup_clear_mc(void) 5387{ 5388 struct mem_cgroup *from = mc.from; 5389 5390 /* 5391 * we must clear moving_task before waking up waiters at the end of 5392 * task migration. 5393 */ 5394 mc.moving_task = NULL; 5395 __mem_cgroup_clear_mc(); 5396 spin_lock(&mc.lock); 5397 mc.from = NULL; 5398 mc.to = NULL; 5399 spin_unlock(&mc.lock); 5400 mem_cgroup_end_move(from); 5401} 5402 5403static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5404 struct cgroup *cgroup, 5405 struct cgroup_taskset *tset) 5406{ 5407 struct task_struct *p = cgroup_taskset_first(tset); 5408 int ret = 0; 5409 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); 5410 5411 if (memcg->move_charge_at_immigrate) { 5412 struct mm_struct *mm; 5413 struct mem_cgroup *from = mem_cgroup_from_task(p); 5414 5415 VM_BUG_ON(from == memcg); 5416 5417 mm = get_task_mm(p); 5418 if (!mm) 5419 return 0; 5420 /* We move charges only when we move a owner of the mm */ 5421 if (mm->owner == p) { 5422 VM_BUG_ON(mc.from); 5423 VM_BUG_ON(mc.to); 5424 VM_BUG_ON(mc.precharge); 5425 VM_BUG_ON(mc.moved_charge); 5426 VM_BUG_ON(mc.moved_swap); 5427 mem_cgroup_start_move(from); 5428 spin_lock(&mc.lock); 5429 mc.from = from; 5430 mc.to = memcg; 5431 spin_unlock(&mc.lock); 5432 /* We set mc.moving_task later */ 5433 5434 ret = mem_cgroup_precharge_mc(mm); 5435 if (ret) 5436 mem_cgroup_clear_mc(); 5437 } 5438 mmput(mm); 5439 } 5440 return ret; 5441} 5442 5443static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5444 struct cgroup *cgroup, 5445 struct cgroup_taskset *tset) 5446{ 5447 mem_cgroup_clear_mc(); 5448} 5449 5450static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 5451 unsigned long addr, unsigned long end, 5452 struct mm_walk *walk) 5453{ 5454 int ret = 0; 5455 struct vm_area_struct *vma = walk->private; 5456 pte_t *pte; 5457 spinlock_t *ptl; 5458 5459 split_huge_page_pmd(walk->mm, pmd); 5460retry: 5461 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5462 for (; addr != end; addr += PAGE_SIZE) { 5463 pte_t ptent = *(pte++); 5464 union mc_target target; 5465 int type; 5466 struct page *page; 5467 struct page_cgroup *pc; 5468 swp_entry_t ent; 5469 5470 if (!mc.precharge) 5471 break; 5472 5473 type = is_target_pte_for_mc(vma, addr, ptent, &target); 5474 switch (type) { 5475 case MC_TARGET_PAGE: 5476 page = target.page; 5477 if (isolate_lru_page(page)) 5478 goto put; 5479 pc = lookup_page_cgroup(page); 5480 if (!mem_cgroup_move_account(page, 1, pc, 5481 mc.from, mc.to, false)) { 5482 mc.precharge--; 5483 /* we uncharge from mc.from later. */ 5484 mc.moved_charge++; 5485 } 5486 putback_lru_page(page); 5487put: /* is_target_pte_for_mc() gets the page */ 5488 put_page(page); 5489 break; 5490 case MC_TARGET_SWAP: 5491 ent = target.ent; 5492 if (!mem_cgroup_move_swap_account(ent, 5493 mc.from, mc.to, false)) { 5494 mc.precharge--; 5495 /* we fixup refcnts and charges later. */ 5496 mc.moved_swap++; 5497 } 5498 break; 5499 default: 5500 break; 5501 } 5502 } 5503 pte_unmap_unlock(pte - 1, ptl); 5504 cond_resched(); 5505 5506 if (addr != end) { 5507 /* 5508 * We have consumed all precharges we got in can_attach(). 5509 * We try charge one by one, but don't do any additional 5510 * charges to mc.to if we have failed in charge once in attach() 5511 * phase. 5512 */ 5513 ret = mem_cgroup_do_precharge(1); 5514 if (!ret) 5515 goto retry; 5516 } 5517 5518 return ret; 5519} 5520 5521static void mem_cgroup_move_charge(struct mm_struct *mm) 5522{ 5523 struct vm_area_struct *vma; 5524 5525 lru_add_drain_all(); 5526retry: 5527 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 5528 /* 5529 * Someone who are holding the mmap_sem might be waiting in 5530 * waitq. So we cancel all extra charges, wake up all waiters, 5531 * and retry. Because we cancel precharges, we might not be able 5532 * to move enough charges, but moving charge is a best-effort 5533 * feature anyway, so it wouldn't be a big problem. 5534 */ 5535 __mem_cgroup_clear_mc(); 5536 cond_resched(); 5537 goto retry; 5538 } 5539 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5540 int ret; 5541 struct mm_walk mem_cgroup_move_charge_walk = { 5542 .pmd_entry = mem_cgroup_move_charge_pte_range, 5543 .mm = mm, 5544 .private = vma, 5545 }; 5546 if (is_vm_hugetlb_page(vma)) 5547 continue; 5548 ret = walk_page_range(vma->vm_start, vma->vm_end, 5549 &mem_cgroup_move_charge_walk); 5550 if (ret) 5551 /* 5552 * means we have consumed all precharges and failed in 5553 * doing additional charge. Just abandon here. 5554 */ 5555 break; 5556 } 5557 up_read(&mm->mmap_sem); 5558} 5559 5560static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5561 struct cgroup *cont, 5562 struct cgroup_taskset *tset) 5563{ 5564 struct task_struct *p = cgroup_taskset_first(tset); 5565 struct mm_struct *mm = get_task_mm(p); 5566 5567 if (mm) { 5568 if (mc.to) 5569 mem_cgroup_move_charge(mm); 5570 put_swap_token(mm); 5571 mmput(mm); 5572 } 5573 if (mc.to) 5574 mem_cgroup_clear_mc(); 5575} 5576#else /* !CONFIG_MMU */ 5577static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5578 struct cgroup *cgroup, 5579 struct cgroup_taskset *tset) 5580{ 5581 return 0; 5582} 5583static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5584 struct cgroup *cgroup, 5585 struct cgroup_taskset *tset) 5586{ 5587} 5588static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5589 struct cgroup *cont, 5590 struct cgroup_taskset *tset) 5591{ 5592} 5593#endif 5594 5595struct cgroup_subsys mem_cgroup_subsys = { 5596 .name = "memory", 5597 .subsys_id = mem_cgroup_subsys_id, 5598 .create = mem_cgroup_create, 5599 .pre_destroy = mem_cgroup_pre_destroy, 5600 .destroy = mem_cgroup_destroy, 5601 .populate = mem_cgroup_populate, 5602 .can_attach = mem_cgroup_can_attach, 5603 .cancel_attach = mem_cgroup_cancel_attach, 5604 .attach = mem_cgroup_move_task, 5605 .early_init = 0, 5606 .use_id = 1, 5607}; 5608 5609#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5610static int __init enable_swap_account(char *s) 5611{ 5612 /* consider enabled if no parameter or 1 is given */ 5613 if (!strcmp(s, "1")) 5614 really_do_swap_account = 1; 5615 else if (!strcmp(s, "0")) 5616 really_do_swap_account = 0; 5617 return 1; 5618} 5619__setup("swapaccount=", enable_swap_account); 5620 5621#endif 5622