memcontrol.c revision 9fb4b7cc0724f178d4b24a2a566ea1e7cb120b82
1/* memcontrol.c - Memory Controller 2 * 3 * Copyright IBM Corporation, 2007 4 * Author Balbir Singh <balbir@linux.vnet.ibm.com> 5 * 6 * Copyright 2007 OpenVZ SWsoft Inc 7 * Author: Pavel Emelianov <xemul@openvz.org> 8 * 9 * Memory thresholds 10 * Copyright (C) 2009 Nokia Corporation 11 * Author: Kirill A. Shutemov 12 * 13 * This program is free software; you can redistribute it and/or modify 14 * it under the terms of the GNU General Public License as published by 15 * the Free Software Foundation; either version 2 of the License, or 16 * (at your option) any later version. 17 * 18 * This program is distributed in the hope that it will be useful, 19 * but WITHOUT ANY WARRANTY; without even the implied warranty of 20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 21 * GNU General Public License for more details. 22 */ 23 24#include <linux/res_counter.h> 25#include <linux/memcontrol.h> 26#include <linux/cgroup.h> 27#include <linux/mm.h> 28#include <linux/hugetlb.h> 29#include <linux/pagemap.h> 30#include <linux/smp.h> 31#include <linux/page-flags.h> 32#include <linux/backing-dev.h> 33#include <linux/bit_spinlock.h> 34#include <linux/rcupdate.h> 35#include <linux/limits.h> 36#include <linux/export.h> 37#include <linux/mutex.h> 38#include <linux/rbtree.h> 39#include <linux/slab.h> 40#include <linux/swap.h> 41#include <linux/swapops.h> 42#include <linux/spinlock.h> 43#include <linux/eventfd.h> 44#include <linux/sort.h> 45#include <linux/fs.h> 46#include <linux/seq_file.h> 47#include <linux/vmalloc.h> 48#include <linux/mm_inline.h> 49#include <linux/page_cgroup.h> 50#include <linux/cpu.h> 51#include <linux/oom.h> 52#include "internal.h" 53#include <net/sock.h> 54#include <net/tcp_memcontrol.h> 55 56#include <asm/uaccess.h> 57 58#include <trace/events/vmscan.h> 59 60struct cgroup_subsys mem_cgroup_subsys __read_mostly; 61#define MEM_CGROUP_RECLAIM_RETRIES 5 62struct mem_cgroup *root_mem_cgroup __read_mostly; 63 64#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 65/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ 66int do_swap_account __read_mostly; 67 68/* for remember boot option*/ 69#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED 70static int really_do_swap_account __initdata = 1; 71#else 72static int really_do_swap_account __initdata = 0; 73#endif 74 75#else 76#define do_swap_account (0) 77#endif 78 79 80/* 81 * Statistics for memory cgroup. 82 */ 83enum mem_cgroup_stat_index { 84 /* 85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 86 */ 87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ 89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ 90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ 91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ 92 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ 93 MEM_CGROUP_STAT_NSTATS, 94}; 95 96enum mem_cgroup_events_index { 97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ 98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ 99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ 100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ 101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ 102 MEM_CGROUP_EVENTS_NSTATS, 103}; 104/* 105 * Per memcg event counter is incremented at every pagein/pageout. With THP, 106 * it will be incremated by the number of pages. This counter is used for 107 * for trigger some periodic events. This is straightforward and better 108 * than using jiffies etc. to handle periodic memcg event. 109 */ 110enum mem_cgroup_events_target { 111 MEM_CGROUP_TARGET_THRESH, 112 MEM_CGROUP_TARGET_SOFTLIMIT, 113 MEM_CGROUP_TARGET_NUMAINFO, 114 MEM_CGROUP_NTARGETS, 115}; 116#define THRESHOLDS_EVENTS_TARGET (128) 117#define SOFTLIMIT_EVENTS_TARGET (1024) 118#define NUMAINFO_EVENTS_TARGET (1024) 119 120struct mem_cgroup_stat_cpu { 121 long count[MEM_CGROUP_STAT_NSTATS]; 122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; 123 unsigned long targets[MEM_CGROUP_NTARGETS]; 124}; 125 126struct mem_cgroup_reclaim_iter { 127 /* css_id of the last scanned hierarchy member */ 128 int position; 129 /* scan generation, increased every round-trip */ 130 unsigned int generation; 131}; 132 133/* 134 * per-zone information in memory controller. 135 */ 136struct mem_cgroup_per_zone { 137 struct lruvec lruvec; 138 unsigned long count[NR_LRU_LISTS]; 139 140 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1]; 141 142 struct zone_reclaim_stat reclaim_stat; 143 struct rb_node tree_node; /* RB tree node */ 144 unsigned long long usage_in_excess;/* Set to the value by which */ 145 /* the soft limit is exceeded*/ 146 bool on_tree; 147 struct mem_cgroup *mem; /* Back pointer, we cannot */ 148 /* use container_of */ 149}; 150/* Macro for accessing counter */ 151#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 152 153struct mem_cgroup_per_node { 154 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 155}; 156 157struct mem_cgroup_lru_info { 158 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 159}; 160 161/* 162 * Cgroups above their limits are maintained in a RB-Tree, independent of 163 * their hierarchy representation 164 */ 165 166struct mem_cgroup_tree_per_zone { 167 struct rb_root rb_root; 168 spinlock_t lock; 169}; 170 171struct mem_cgroup_tree_per_node { 172 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; 173}; 174 175struct mem_cgroup_tree { 176 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; 177}; 178 179static struct mem_cgroup_tree soft_limit_tree __read_mostly; 180 181struct mem_cgroup_threshold { 182 struct eventfd_ctx *eventfd; 183 u64 threshold; 184}; 185 186/* For threshold */ 187struct mem_cgroup_threshold_ary { 188 /* An array index points to threshold just below usage. */ 189 int current_threshold; 190 /* Size of entries[] */ 191 unsigned int size; 192 /* Array of thresholds */ 193 struct mem_cgroup_threshold entries[0]; 194}; 195 196struct mem_cgroup_thresholds { 197 /* Primary thresholds array */ 198 struct mem_cgroup_threshold_ary *primary; 199 /* 200 * Spare threshold array. 201 * This is needed to make mem_cgroup_unregister_event() "never fail". 202 * It must be able to store at least primary->size - 1 entries. 203 */ 204 struct mem_cgroup_threshold_ary *spare; 205}; 206 207/* for OOM */ 208struct mem_cgroup_eventfd_list { 209 struct list_head list; 210 struct eventfd_ctx *eventfd; 211}; 212 213static void mem_cgroup_threshold(struct mem_cgroup *memcg); 214static void mem_cgroup_oom_notify(struct mem_cgroup *memcg); 215 216/* 217 * The memory controller data structure. The memory controller controls both 218 * page cache and RSS per cgroup. We would eventually like to provide 219 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 220 * to help the administrator determine what knobs to tune. 221 * 222 * TODO: Add a water mark for the memory controller. Reclaim will begin when 223 * we hit the water mark. May be even add a low water mark, such that 224 * no reclaim occurs from a cgroup at it's low water mark, this is 225 * a feature that will be implemented much later in the future. 226 */ 227struct mem_cgroup { 228 struct cgroup_subsys_state css; 229 /* 230 * the counter to account for memory usage 231 */ 232 struct res_counter res; 233 /* 234 * the counter to account for mem+swap usage. 235 */ 236 struct res_counter memsw; 237 /* 238 * Per cgroup active and inactive list, similar to the 239 * per zone LRU lists. 240 */ 241 struct mem_cgroup_lru_info info; 242 int last_scanned_node; 243#if MAX_NUMNODES > 1 244 nodemask_t scan_nodes; 245 atomic_t numainfo_events; 246 atomic_t numainfo_updating; 247#endif 248 /* 249 * Should the accounting and control be hierarchical, per subtree? 250 */ 251 bool use_hierarchy; 252 253 bool oom_lock; 254 atomic_t under_oom; 255 256 atomic_t refcnt; 257 258 int swappiness; 259 /* OOM-Killer disable */ 260 int oom_kill_disable; 261 262 /* set when res.limit == memsw.limit */ 263 bool memsw_is_minimum; 264 265 /* protect arrays of thresholds */ 266 struct mutex thresholds_lock; 267 268 /* thresholds for memory usage. RCU-protected */ 269 struct mem_cgroup_thresholds thresholds; 270 271 /* thresholds for mem+swap usage. RCU-protected */ 272 struct mem_cgroup_thresholds memsw_thresholds; 273 274 /* For oom notifier event fd */ 275 struct list_head oom_notify; 276 277 /* 278 * Should we move charges of a task when a task is moved into this 279 * mem_cgroup ? And what type of charges should we move ? 280 */ 281 unsigned long move_charge_at_immigrate; 282 /* 283 * percpu counter. 284 */ 285 struct mem_cgroup_stat_cpu *stat; 286 /* 287 * used when a cpu is offlined or other synchronizations 288 * See mem_cgroup_read_stat(). 289 */ 290 struct mem_cgroup_stat_cpu nocpu_base; 291 spinlock_t pcp_counter_lock; 292 293#ifdef CONFIG_INET 294 struct tcp_memcontrol tcp_mem; 295#endif 296}; 297 298/* Stuffs for move charges at task migration. */ 299/* 300 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a 301 * left-shifted bitmap of these types. 302 */ 303enum move_type { 304 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ 305 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ 306 NR_MOVE_TYPE, 307}; 308 309/* "mc" and its members are protected by cgroup_mutex */ 310static struct move_charge_struct { 311 spinlock_t lock; /* for from, to */ 312 struct mem_cgroup *from; 313 struct mem_cgroup *to; 314 unsigned long precharge; 315 unsigned long moved_charge; 316 unsigned long moved_swap; 317 struct task_struct *moving_task; /* a task moving charges */ 318 wait_queue_head_t waitq; /* a waitq for other context */ 319} mc = { 320 .lock = __SPIN_LOCK_UNLOCKED(mc.lock), 321 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), 322}; 323 324static bool move_anon(void) 325{ 326 return test_bit(MOVE_CHARGE_TYPE_ANON, 327 &mc.to->move_charge_at_immigrate); 328} 329 330static bool move_file(void) 331{ 332 return test_bit(MOVE_CHARGE_TYPE_FILE, 333 &mc.to->move_charge_at_immigrate); 334} 335 336/* 337 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft 338 * limit reclaim to prevent infinite loops, if they ever occur. 339 */ 340#define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) 341#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) 342 343enum charge_type { 344 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 345 MEM_CGROUP_CHARGE_TYPE_MAPPED, 346 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 347 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 348 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 349 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ 350 NR_CHARGE_TYPE, 351}; 352 353/* for encoding cft->private value on file */ 354#define _MEM (0) 355#define _MEMSWAP (1) 356#define _OOM_TYPE (2) 357#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 358#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 359#define MEMFILE_ATTR(val) ((val) & 0xffff) 360/* Used for OOM nofiier */ 361#define OOM_CONTROL (0) 362 363/* 364 * Reclaim flags for mem_cgroup_hierarchical_reclaim 365 */ 366#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 367#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) 368#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 369#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) 370 371static void mem_cgroup_get(struct mem_cgroup *memcg); 372static void mem_cgroup_put(struct mem_cgroup *memcg); 373 374/* Writing them here to avoid exposing memcg's inner layout */ 375#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 376#ifdef CONFIG_INET 377#include <net/sock.h> 378#include <net/ip.h> 379 380static bool mem_cgroup_is_root(struct mem_cgroup *memcg); 381void sock_update_memcg(struct sock *sk) 382{ 383 if (static_branch(&memcg_socket_limit_enabled)) { 384 struct mem_cgroup *memcg; 385 386 BUG_ON(!sk->sk_prot->proto_cgroup); 387 388 /* Socket cloning can throw us here with sk_cgrp already 389 * filled. It won't however, necessarily happen from 390 * process context. So the test for root memcg given 391 * the current task's memcg won't help us in this case. 392 * 393 * Respecting the original socket's memcg is a better 394 * decision in this case. 395 */ 396 if (sk->sk_cgrp) { 397 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg)); 398 mem_cgroup_get(sk->sk_cgrp->memcg); 399 return; 400 } 401 402 rcu_read_lock(); 403 memcg = mem_cgroup_from_task(current); 404 if (!mem_cgroup_is_root(memcg)) { 405 mem_cgroup_get(memcg); 406 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg); 407 } 408 rcu_read_unlock(); 409 } 410} 411EXPORT_SYMBOL(sock_update_memcg); 412 413void sock_release_memcg(struct sock *sk) 414{ 415 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) { 416 struct mem_cgroup *memcg; 417 WARN_ON(!sk->sk_cgrp->memcg); 418 memcg = sk->sk_cgrp->memcg; 419 mem_cgroup_put(memcg); 420 } 421} 422 423struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg) 424{ 425 if (!memcg || mem_cgroup_is_root(memcg)) 426 return NULL; 427 428 return &memcg->tcp_mem.cg_proto; 429} 430EXPORT_SYMBOL(tcp_proto_cgroup); 431#endif /* CONFIG_INET */ 432#endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */ 433 434static void drain_all_stock_async(struct mem_cgroup *memcg); 435 436static struct mem_cgroup_per_zone * 437mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid) 438{ 439 return &memcg->info.nodeinfo[nid]->zoneinfo[zid]; 440} 441 442struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg) 443{ 444 return &memcg->css; 445} 446 447static struct mem_cgroup_per_zone * 448page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page) 449{ 450 int nid = page_to_nid(page); 451 int zid = page_zonenum(page); 452 453 return mem_cgroup_zoneinfo(memcg, nid, zid); 454} 455 456static struct mem_cgroup_tree_per_zone * 457soft_limit_tree_node_zone(int nid, int zid) 458{ 459 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 460} 461 462static struct mem_cgroup_tree_per_zone * 463soft_limit_tree_from_page(struct page *page) 464{ 465 int nid = page_to_nid(page); 466 int zid = page_zonenum(page); 467 468 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 469} 470 471static void 472__mem_cgroup_insert_exceeded(struct mem_cgroup *memcg, 473 struct mem_cgroup_per_zone *mz, 474 struct mem_cgroup_tree_per_zone *mctz, 475 unsigned long long new_usage_in_excess) 476{ 477 struct rb_node **p = &mctz->rb_root.rb_node; 478 struct rb_node *parent = NULL; 479 struct mem_cgroup_per_zone *mz_node; 480 481 if (mz->on_tree) 482 return; 483 484 mz->usage_in_excess = new_usage_in_excess; 485 if (!mz->usage_in_excess) 486 return; 487 while (*p) { 488 parent = *p; 489 mz_node = rb_entry(parent, struct mem_cgroup_per_zone, 490 tree_node); 491 if (mz->usage_in_excess < mz_node->usage_in_excess) 492 p = &(*p)->rb_left; 493 /* 494 * We can't avoid mem cgroups that are over their soft 495 * limit by the same amount 496 */ 497 else if (mz->usage_in_excess >= mz_node->usage_in_excess) 498 p = &(*p)->rb_right; 499 } 500 rb_link_node(&mz->tree_node, parent, p); 501 rb_insert_color(&mz->tree_node, &mctz->rb_root); 502 mz->on_tree = true; 503} 504 505static void 506__mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, 507 struct mem_cgroup_per_zone *mz, 508 struct mem_cgroup_tree_per_zone *mctz) 509{ 510 if (!mz->on_tree) 511 return; 512 rb_erase(&mz->tree_node, &mctz->rb_root); 513 mz->on_tree = false; 514} 515 516static void 517mem_cgroup_remove_exceeded(struct mem_cgroup *memcg, 518 struct mem_cgroup_per_zone *mz, 519 struct mem_cgroup_tree_per_zone *mctz) 520{ 521 spin_lock(&mctz->lock); 522 __mem_cgroup_remove_exceeded(memcg, mz, mctz); 523 spin_unlock(&mctz->lock); 524} 525 526 527static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page) 528{ 529 unsigned long long excess; 530 struct mem_cgroup_per_zone *mz; 531 struct mem_cgroup_tree_per_zone *mctz; 532 int nid = page_to_nid(page); 533 int zid = page_zonenum(page); 534 mctz = soft_limit_tree_from_page(page); 535 536 /* 537 * Necessary to update all ancestors when hierarchy is used. 538 * because their event counter is not touched. 539 */ 540 for (; memcg; memcg = parent_mem_cgroup(memcg)) { 541 mz = mem_cgroup_zoneinfo(memcg, nid, zid); 542 excess = res_counter_soft_limit_excess(&memcg->res); 543 /* 544 * We have to update the tree if mz is on RB-tree or 545 * mem is over its softlimit. 546 */ 547 if (excess || mz->on_tree) { 548 spin_lock(&mctz->lock); 549 /* if on-tree, remove it */ 550 if (mz->on_tree) 551 __mem_cgroup_remove_exceeded(memcg, mz, mctz); 552 /* 553 * Insert again. mz->usage_in_excess will be updated. 554 * If excess is 0, no tree ops. 555 */ 556 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess); 557 spin_unlock(&mctz->lock); 558 } 559 } 560} 561 562static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg) 563{ 564 int node, zone; 565 struct mem_cgroup_per_zone *mz; 566 struct mem_cgroup_tree_per_zone *mctz; 567 568 for_each_node_state(node, N_POSSIBLE) { 569 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 570 mz = mem_cgroup_zoneinfo(memcg, node, zone); 571 mctz = soft_limit_tree_node_zone(node, zone); 572 mem_cgroup_remove_exceeded(memcg, mz, mctz); 573 } 574 } 575} 576 577static struct mem_cgroup_per_zone * 578__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 579{ 580 struct rb_node *rightmost = NULL; 581 struct mem_cgroup_per_zone *mz; 582 583retry: 584 mz = NULL; 585 rightmost = rb_last(&mctz->rb_root); 586 if (!rightmost) 587 goto done; /* Nothing to reclaim from */ 588 589 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); 590 /* 591 * Remove the node now but someone else can add it back, 592 * we will to add it back at the end of reclaim to its correct 593 * position in the tree. 594 */ 595 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 596 if (!res_counter_soft_limit_excess(&mz->mem->res) || 597 !css_tryget(&mz->mem->css)) 598 goto retry; 599done: 600 return mz; 601} 602 603static struct mem_cgroup_per_zone * 604mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 605{ 606 struct mem_cgroup_per_zone *mz; 607 608 spin_lock(&mctz->lock); 609 mz = __mem_cgroup_largest_soft_limit_node(mctz); 610 spin_unlock(&mctz->lock); 611 return mz; 612} 613 614/* 615 * Implementation Note: reading percpu statistics for memcg. 616 * 617 * Both of vmstat[] and percpu_counter has threshold and do periodic 618 * synchronization to implement "quick" read. There are trade-off between 619 * reading cost and precision of value. Then, we may have a chance to implement 620 * a periodic synchronizion of counter in memcg's counter. 621 * 622 * But this _read() function is used for user interface now. The user accounts 623 * memory usage by memory cgroup and he _always_ requires exact value because 624 * he accounts memory. Even if we provide quick-and-fuzzy read, we always 625 * have to visit all online cpus and make sum. So, for now, unnecessary 626 * synchronization is not implemented. (just implemented for cpu hotplug) 627 * 628 * If there are kernel internal actions which can make use of some not-exact 629 * value, and reading all cpu value can be performance bottleneck in some 630 * common workload, threashold and synchonization as vmstat[] should be 631 * implemented. 632 */ 633static long mem_cgroup_read_stat(struct mem_cgroup *memcg, 634 enum mem_cgroup_stat_index idx) 635{ 636 long val = 0; 637 int cpu; 638 639 get_online_cpus(); 640 for_each_online_cpu(cpu) 641 val += per_cpu(memcg->stat->count[idx], cpu); 642#ifdef CONFIG_HOTPLUG_CPU 643 spin_lock(&memcg->pcp_counter_lock); 644 val += memcg->nocpu_base.count[idx]; 645 spin_unlock(&memcg->pcp_counter_lock); 646#endif 647 put_online_cpus(); 648 return val; 649} 650 651static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg, 652 bool charge) 653{ 654 int val = (charge) ? 1 : -1; 655 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); 656} 657 658static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg, 659 enum mem_cgroup_events_index idx) 660{ 661 unsigned long val = 0; 662 int cpu; 663 664 for_each_online_cpu(cpu) 665 val += per_cpu(memcg->stat->events[idx], cpu); 666#ifdef CONFIG_HOTPLUG_CPU 667 spin_lock(&memcg->pcp_counter_lock); 668 val += memcg->nocpu_base.events[idx]; 669 spin_unlock(&memcg->pcp_counter_lock); 670#endif 671 return val; 672} 673 674static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg, 675 bool file, int nr_pages) 676{ 677 preempt_disable(); 678 679 if (file) 680 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE], 681 nr_pages); 682 else 683 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS], 684 nr_pages); 685 686 /* pagein of a big page is an event. So, ignore page size */ 687 if (nr_pages > 0) 688 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); 689 else { 690 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); 691 nr_pages = -nr_pages; /* for event */ 692 } 693 694 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); 695 696 preempt_enable(); 697} 698 699unsigned long 700mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid, 701 unsigned int lru_mask) 702{ 703 struct mem_cgroup_per_zone *mz; 704 enum lru_list l; 705 unsigned long ret = 0; 706 707 mz = mem_cgroup_zoneinfo(memcg, nid, zid); 708 709 for_each_lru(l) { 710 if (BIT(l) & lru_mask) 711 ret += MEM_CGROUP_ZSTAT(mz, l); 712 } 713 return ret; 714} 715 716static unsigned long 717mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, 718 int nid, unsigned int lru_mask) 719{ 720 u64 total = 0; 721 int zid; 722 723 for (zid = 0; zid < MAX_NR_ZONES; zid++) 724 total += mem_cgroup_zone_nr_lru_pages(memcg, 725 nid, zid, lru_mask); 726 727 return total; 728} 729 730static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg, 731 unsigned int lru_mask) 732{ 733 int nid; 734 u64 total = 0; 735 736 for_each_node_state(nid, N_HIGH_MEMORY) 737 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask); 738 return total; 739} 740 741static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg, 742 enum mem_cgroup_events_target target) 743{ 744 unsigned long val, next; 745 746 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]); 747 next = __this_cpu_read(memcg->stat->targets[target]); 748 /* from time_after() in jiffies.h */ 749 if ((long)next - (long)val < 0) { 750 switch (target) { 751 case MEM_CGROUP_TARGET_THRESH: 752 next = val + THRESHOLDS_EVENTS_TARGET; 753 break; 754 case MEM_CGROUP_TARGET_SOFTLIMIT: 755 next = val + SOFTLIMIT_EVENTS_TARGET; 756 break; 757 case MEM_CGROUP_TARGET_NUMAINFO: 758 next = val + NUMAINFO_EVENTS_TARGET; 759 break; 760 default: 761 break; 762 } 763 __this_cpu_write(memcg->stat->targets[target], next); 764 return true; 765 } 766 return false; 767} 768 769/* 770 * Check events in order. 771 * 772 */ 773static void memcg_check_events(struct mem_cgroup *memcg, struct page *page) 774{ 775 preempt_disable(); 776 /* threshold event is triggered in finer grain than soft limit */ 777 if (unlikely(mem_cgroup_event_ratelimit(memcg, 778 MEM_CGROUP_TARGET_THRESH))) { 779 bool do_softlimit, do_numainfo; 780 781 do_softlimit = mem_cgroup_event_ratelimit(memcg, 782 MEM_CGROUP_TARGET_SOFTLIMIT); 783#if MAX_NUMNODES > 1 784 do_numainfo = mem_cgroup_event_ratelimit(memcg, 785 MEM_CGROUP_TARGET_NUMAINFO); 786#endif 787 preempt_enable(); 788 789 mem_cgroup_threshold(memcg); 790 if (unlikely(do_softlimit)) 791 mem_cgroup_update_tree(memcg, page); 792#if MAX_NUMNODES > 1 793 if (unlikely(do_numainfo)) 794 atomic_inc(&memcg->numainfo_events); 795#endif 796 } else 797 preempt_enable(); 798} 799 800struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 801{ 802 return container_of(cgroup_subsys_state(cont, 803 mem_cgroup_subsys_id), struct mem_cgroup, 804 css); 805} 806 807struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 808{ 809 /* 810 * mm_update_next_owner() may clear mm->owner to NULL 811 * if it races with swapoff, page migration, etc. 812 * So this can be called with p == NULL. 813 */ 814 if (unlikely(!p)) 815 return NULL; 816 817 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 818 struct mem_cgroup, css); 819} 820 821struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 822{ 823 struct mem_cgroup *memcg = NULL; 824 825 if (!mm) 826 return NULL; 827 /* 828 * Because we have no locks, mm->owner's may be being moved to other 829 * cgroup. We use css_tryget() here even if this looks 830 * pessimistic (rather than adding locks here). 831 */ 832 rcu_read_lock(); 833 do { 834 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 835 if (unlikely(!memcg)) 836 break; 837 } while (!css_tryget(&memcg->css)); 838 rcu_read_unlock(); 839 return memcg; 840} 841 842/** 843 * mem_cgroup_iter - iterate over memory cgroup hierarchy 844 * @root: hierarchy root 845 * @prev: previously returned memcg, NULL on first invocation 846 * @reclaim: cookie for shared reclaim walks, NULL for full walks 847 * 848 * Returns references to children of the hierarchy below @root, or 849 * @root itself, or %NULL after a full round-trip. 850 * 851 * Caller must pass the return value in @prev on subsequent 852 * invocations for reference counting, or use mem_cgroup_iter_break() 853 * to cancel a hierarchy walk before the round-trip is complete. 854 * 855 * Reclaimers can specify a zone and a priority level in @reclaim to 856 * divide up the memcgs in the hierarchy among all concurrent 857 * reclaimers operating on the same zone and priority. 858 */ 859struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root, 860 struct mem_cgroup *prev, 861 struct mem_cgroup_reclaim_cookie *reclaim) 862{ 863 struct mem_cgroup *memcg = NULL; 864 int id = 0; 865 866 if (mem_cgroup_disabled()) 867 return NULL; 868 869 if (!root) 870 root = root_mem_cgroup; 871 872 if (prev && !reclaim) 873 id = css_id(&prev->css); 874 875 if (prev && prev != root) 876 css_put(&prev->css); 877 878 if (!root->use_hierarchy && root != root_mem_cgroup) { 879 if (prev) 880 return NULL; 881 return root; 882 } 883 884 while (!memcg) { 885 struct mem_cgroup_reclaim_iter *uninitialized_var(iter); 886 struct cgroup_subsys_state *css; 887 888 if (reclaim) { 889 int nid = zone_to_nid(reclaim->zone); 890 int zid = zone_idx(reclaim->zone); 891 struct mem_cgroup_per_zone *mz; 892 893 mz = mem_cgroup_zoneinfo(root, nid, zid); 894 iter = &mz->reclaim_iter[reclaim->priority]; 895 if (prev && reclaim->generation != iter->generation) 896 return NULL; 897 id = iter->position; 898 } 899 900 rcu_read_lock(); 901 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id); 902 if (css) { 903 if (css == &root->css || css_tryget(css)) 904 memcg = container_of(css, 905 struct mem_cgroup, css); 906 } else 907 id = 0; 908 rcu_read_unlock(); 909 910 if (reclaim) { 911 iter->position = id; 912 if (!css) 913 iter->generation++; 914 else if (!prev && memcg) 915 reclaim->generation = iter->generation; 916 } 917 918 if (prev && !css) 919 return NULL; 920 } 921 return memcg; 922} 923 924/** 925 * mem_cgroup_iter_break - abort a hierarchy walk prematurely 926 * @root: hierarchy root 927 * @prev: last visited hierarchy member as returned by mem_cgroup_iter() 928 */ 929void mem_cgroup_iter_break(struct mem_cgroup *root, 930 struct mem_cgroup *prev) 931{ 932 if (!root) 933 root = root_mem_cgroup; 934 if (prev && prev != root) 935 css_put(&prev->css); 936} 937 938/* 939 * Iteration constructs for visiting all cgroups (under a tree). If 940 * loops are exited prematurely (break), mem_cgroup_iter_break() must 941 * be used for reference counting. 942 */ 943#define for_each_mem_cgroup_tree(iter, root) \ 944 for (iter = mem_cgroup_iter(root, NULL, NULL); \ 945 iter != NULL; \ 946 iter = mem_cgroup_iter(root, iter, NULL)) 947 948#define for_each_mem_cgroup(iter) \ 949 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \ 950 iter != NULL; \ 951 iter = mem_cgroup_iter(NULL, iter, NULL)) 952 953static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg) 954{ 955 return (memcg == root_mem_cgroup); 956} 957 958void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) 959{ 960 struct mem_cgroup *memcg; 961 962 if (!mm) 963 return; 964 965 rcu_read_lock(); 966 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner)); 967 if (unlikely(!memcg)) 968 goto out; 969 970 switch (idx) { 971 case PGFAULT: 972 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]); 973 break; 974 case PGMAJFAULT: 975 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]); 976 break; 977 default: 978 BUG(); 979 } 980out: 981 rcu_read_unlock(); 982} 983EXPORT_SYMBOL(mem_cgroup_count_vm_event); 984 985/** 986 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg 987 * @zone: zone of the wanted lruvec 988 * @mem: memcg of the wanted lruvec 989 * 990 * Returns the lru list vector holding pages for the given @zone and 991 * @mem. This can be the global zone lruvec, if the memory controller 992 * is disabled. 993 */ 994struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone, 995 struct mem_cgroup *memcg) 996{ 997 struct mem_cgroup_per_zone *mz; 998 999 if (mem_cgroup_disabled()) 1000 return &zone->lruvec; 1001 1002 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone)); 1003 return &mz->lruvec; 1004} 1005 1006/* 1007 * Following LRU functions are allowed to be used without PCG_LOCK. 1008 * Operations are called by routine of global LRU independently from memcg. 1009 * What we have to take care of here is validness of pc->mem_cgroup. 1010 * 1011 * Changes to pc->mem_cgroup happens when 1012 * 1. charge 1013 * 2. moving account 1014 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 1015 * It is added to LRU before charge. 1016 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 1017 * When moving account, the page is not on LRU. It's isolated. 1018 */ 1019 1020/** 1021 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec 1022 * @zone: zone of the page 1023 * @page: the page 1024 * @lru: current lru 1025 * 1026 * This function accounts for @page being added to @lru, and returns 1027 * the lruvec for the given @zone and the memcg @page is charged to. 1028 * 1029 * The callsite is then responsible for physically linking the page to 1030 * the returned lruvec->lists[@lru]. 1031 */ 1032struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page, 1033 enum lru_list lru) 1034{ 1035 struct mem_cgroup_per_zone *mz; 1036 struct mem_cgroup *memcg; 1037 struct page_cgroup *pc; 1038 1039 if (mem_cgroup_disabled()) 1040 return &zone->lruvec; 1041 1042 pc = lookup_page_cgroup(page); 1043 VM_BUG_ON(PageCgroupAcctLRU(pc)); 1044 /* 1045 * putback: charge: 1046 * SetPageLRU SetPageCgroupUsed 1047 * smp_mb smp_mb 1048 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1049 * 1050 * Ensure that one of the two sides adds the page to the memcg 1051 * LRU during a race. 1052 */ 1053 smp_mb(); 1054 /* 1055 * If the page is uncharged, it may be freed soon, but it 1056 * could also be swap cache (readahead, swapoff) that needs to 1057 * be reclaimable in the future. root_mem_cgroup will babysit 1058 * it for the time being. 1059 */ 1060 if (PageCgroupUsed(pc)) { 1061 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1062 smp_rmb(); 1063 memcg = pc->mem_cgroup; 1064 SetPageCgroupAcctLRU(pc); 1065 } else 1066 memcg = root_mem_cgroup; 1067 mz = page_cgroup_zoneinfo(memcg, page); 1068 /* compound_order() is stabilized through lru_lock */ 1069 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 1070 return &mz->lruvec; 1071} 1072 1073/** 1074 * mem_cgroup_lru_del_list - account for removing an lru page 1075 * @page: the page 1076 * @lru: target lru 1077 * 1078 * This function accounts for @page being removed from @lru. 1079 * 1080 * The callsite is then responsible for physically unlinking 1081 * @page->lru. 1082 */ 1083void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru) 1084{ 1085 struct mem_cgroup_per_zone *mz; 1086 struct mem_cgroup *memcg; 1087 struct page_cgroup *pc; 1088 1089 if (mem_cgroup_disabled()) 1090 return; 1091 1092 pc = lookup_page_cgroup(page); 1093 /* 1094 * root_mem_cgroup babysits uncharged LRU pages, but 1095 * PageCgroupUsed is cleared when the page is about to get 1096 * freed. PageCgroupAcctLRU remembers whether the 1097 * LRU-accounting happened against pc->mem_cgroup or 1098 * root_mem_cgroup. 1099 */ 1100 if (TestClearPageCgroupAcctLRU(pc)) { 1101 VM_BUG_ON(!pc->mem_cgroup); 1102 memcg = pc->mem_cgroup; 1103 } else 1104 memcg = root_mem_cgroup; 1105 mz = page_cgroup_zoneinfo(memcg, page); 1106 /* huge page split is done under lru_lock. so, we have no races. */ 1107 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 1108} 1109 1110void mem_cgroup_lru_del(struct page *page) 1111{ 1112 mem_cgroup_lru_del_list(page, page_lru(page)); 1113} 1114 1115/** 1116 * mem_cgroup_lru_move_lists - account for moving a page between lrus 1117 * @zone: zone of the page 1118 * @page: the page 1119 * @from: current lru 1120 * @to: target lru 1121 * 1122 * This function accounts for @page being moved between the lrus @from 1123 * and @to, and returns the lruvec for the given @zone and the memcg 1124 * @page is charged to. 1125 * 1126 * The callsite is then responsible for physically relinking 1127 * @page->lru to the returned lruvec->lists[@to]. 1128 */ 1129struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone, 1130 struct page *page, 1131 enum lru_list from, 1132 enum lru_list to) 1133{ 1134 /* XXX: Optimize this, especially for @from == @to */ 1135 mem_cgroup_lru_del_list(page, from); 1136 return mem_cgroup_lru_add_list(zone, page, to); 1137} 1138 1139/* 1140 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed 1141 * while it's linked to lru because the page may be reused after it's fully 1142 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again. 1143 * It's done under lock_page and expected that zone->lru_lock isnever held. 1144 */ 1145static void mem_cgroup_lru_del_before_commit(struct page *page) 1146{ 1147 enum lru_list lru; 1148 unsigned long flags; 1149 struct zone *zone = page_zone(page); 1150 struct page_cgroup *pc = lookup_page_cgroup(page); 1151 1152 /* 1153 * Doing this check without taking ->lru_lock seems wrong but this 1154 * is safe. Because if page_cgroup's USED bit is unset, the page 1155 * will not be added to any memcg's LRU. If page_cgroup's USED bit is 1156 * set, the commit after this will fail, anyway. 1157 * This all charge/uncharge is done under some mutual execustion. 1158 * So, we don't need to taking care of changes in USED bit. 1159 */ 1160 if (likely(!PageLRU(page))) 1161 return; 1162 1163 spin_lock_irqsave(&zone->lru_lock, flags); 1164 lru = page_lru(page); 1165 /* 1166 * The uncharged page could still be registered to the LRU of 1167 * the stale pc->mem_cgroup. 1168 * 1169 * As pc->mem_cgroup is about to get overwritten, the old LRU 1170 * accounting needs to be taken care of. Let root_mem_cgroup 1171 * babysit the page until the new memcg is responsible for it. 1172 * 1173 * The PCG_USED bit is guarded by lock_page() as the page is 1174 * swapcache/pagecache. 1175 */ 1176 if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) { 1177 del_page_from_lru_list(zone, page, lru); 1178 add_page_to_lru_list(zone, page, lru); 1179 } 1180 spin_unlock_irqrestore(&zone->lru_lock, flags); 1181} 1182 1183static void mem_cgroup_lru_add_after_commit(struct page *page) 1184{ 1185 enum lru_list lru; 1186 unsigned long flags; 1187 struct zone *zone = page_zone(page); 1188 struct page_cgroup *pc = lookup_page_cgroup(page); 1189 /* 1190 * putback: charge: 1191 * SetPageLRU SetPageCgroupUsed 1192 * smp_mb smp_mb 1193 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU 1194 * 1195 * Ensure that one of the two sides adds the page to the memcg 1196 * LRU during a race. 1197 */ 1198 smp_mb(); 1199 /* taking care of that the page is added to LRU while we commit it */ 1200 if (likely(!PageLRU(page))) 1201 return; 1202 spin_lock_irqsave(&zone->lru_lock, flags); 1203 lru = page_lru(page); 1204 /* 1205 * If the page is not on the LRU, someone will soon put it 1206 * there. If it is, and also already accounted for on the 1207 * memcg-side, it must be on the right lruvec as setting 1208 * pc->mem_cgroup and PageCgroupUsed is properly ordered. 1209 * Otherwise, root_mem_cgroup has been babysitting the page 1210 * during the charge. Move it to the new memcg now. 1211 */ 1212 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) { 1213 del_page_from_lru_list(zone, page, lru); 1214 add_page_to_lru_list(zone, page, lru); 1215 } 1216 spin_unlock_irqrestore(&zone->lru_lock, flags); 1217} 1218 1219/* 1220 * Checks whether given mem is same or in the root_mem_cgroup's 1221 * hierarchy subtree 1222 */ 1223static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg, 1224 struct mem_cgroup *memcg) 1225{ 1226 if (root_memcg != memcg) { 1227 return (root_memcg->use_hierarchy && 1228 css_is_ancestor(&memcg->css, &root_memcg->css)); 1229 } 1230 1231 return true; 1232} 1233 1234int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg) 1235{ 1236 int ret; 1237 struct mem_cgroup *curr = NULL; 1238 struct task_struct *p; 1239 1240 p = find_lock_task_mm(task); 1241 if (!p) 1242 return 0; 1243 curr = try_get_mem_cgroup_from_mm(p->mm); 1244 task_unlock(p); 1245 if (!curr) 1246 return 0; 1247 /* 1248 * We should check use_hierarchy of "memcg" not "curr". Because checking 1249 * use_hierarchy of "curr" here make this function true if hierarchy is 1250 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup* 1251 * hierarchy(even if use_hierarchy is disabled in "memcg"). 1252 */ 1253 ret = mem_cgroup_same_or_subtree(memcg, curr); 1254 css_put(&curr->css); 1255 return ret; 1256} 1257 1258int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone) 1259{ 1260 unsigned long inactive_ratio; 1261 int nid = zone_to_nid(zone); 1262 int zid = zone_idx(zone); 1263 unsigned long inactive; 1264 unsigned long active; 1265 unsigned long gb; 1266 1267 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1268 BIT(LRU_INACTIVE_ANON)); 1269 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1270 BIT(LRU_ACTIVE_ANON)); 1271 1272 gb = (inactive + active) >> (30 - PAGE_SHIFT); 1273 if (gb) 1274 inactive_ratio = int_sqrt(10 * gb); 1275 else 1276 inactive_ratio = 1; 1277 1278 return inactive * inactive_ratio < active; 1279} 1280 1281int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone) 1282{ 1283 unsigned long active; 1284 unsigned long inactive; 1285 int zid = zone_idx(zone); 1286 int nid = zone_to_nid(zone); 1287 1288 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1289 BIT(LRU_INACTIVE_FILE)); 1290 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid, 1291 BIT(LRU_ACTIVE_FILE)); 1292 1293 return (active > inactive); 1294} 1295 1296struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1297 struct zone *zone) 1298{ 1299 int nid = zone_to_nid(zone); 1300 int zid = zone_idx(zone); 1301 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1302 1303 return &mz->reclaim_stat; 1304} 1305 1306struct zone_reclaim_stat * 1307mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1308{ 1309 struct page_cgroup *pc; 1310 struct mem_cgroup_per_zone *mz; 1311 1312 if (mem_cgroup_disabled()) 1313 return NULL; 1314 1315 pc = lookup_page_cgroup(page); 1316 if (!PageCgroupUsed(pc)) 1317 return NULL; 1318 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1319 smp_rmb(); 1320 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1321 return &mz->reclaim_stat; 1322} 1323 1324#define mem_cgroup_from_res_counter(counter, member) \ 1325 container_of(counter, struct mem_cgroup, member) 1326 1327/** 1328 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1329 * @mem: the memory cgroup 1330 * 1331 * Returns the maximum amount of memory @mem can be charged with, in 1332 * pages. 1333 */ 1334static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg) 1335{ 1336 unsigned long long margin; 1337 1338 margin = res_counter_margin(&memcg->res); 1339 if (do_swap_account) 1340 margin = min(margin, res_counter_margin(&memcg->memsw)); 1341 return margin >> PAGE_SHIFT; 1342} 1343 1344int mem_cgroup_swappiness(struct mem_cgroup *memcg) 1345{ 1346 struct cgroup *cgrp = memcg->css.cgroup; 1347 1348 /* root ? */ 1349 if (cgrp->parent == NULL) 1350 return vm_swappiness; 1351 1352 return memcg->swappiness; 1353} 1354 1355static void mem_cgroup_start_move(struct mem_cgroup *memcg) 1356{ 1357 int cpu; 1358 1359 get_online_cpus(); 1360 spin_lock(&memcg->pcp_counter_lock); 1361 for_each_online_cpu(cpu) 1362 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1363 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1364 spin_unlock(&memcg->pcp_counter_lock); 1365 put_online_cpus(); 1366 1367 synchronize_rcu(); 1368} 1369 1370static void mem_cgroup_end_move(struct mem_cgroup *memcg) 1371{ 1372 int cpu; 1373 1374 if (!memcg) 1375 return; 1376 get_online_cpus(); 1377 spin_lock(&memcg->pcp_counter_lock); 1378 for_each_online_cpu(cpu) 1379 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1380 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1381 spin_unlock(&memcg->pcp_counter_lock); 1382 put_online_cpus(); 1383} 1384/* 1385 * 2 routines for checking "mem" is under move_account() or not. 1386 * 1387 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1388 * for avoiding race in accounting. If true, 1389 * pc->mem_cgroup may be overwritten. 1390 * 1391 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1392 * under hierarchy of moving cgroups. This is for 1393 * waiting at hith-memory prressure caused by "move". 1394 */ 1395 1396static bool mem_cgroup_stealed(struct mem_cgroup *memcg) 1397{ 1398 VM_BUG_ON(!rcu_read_lock_held()); 1399 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1400} 1401 1402static bool mem_cgroup_under_move(struct mem_cgroup *memcg) 1403{ 1404 struct mem_cgroup *from; 1405 struct mem_cgroup *to; 1406 bool ret = false; 1407 /* 1408 * Unlike task_move routines, we access mc.to, mc.from not under 1409 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1410 */ 1411 spin_lock(&mc.lock); 1412 from = mc.from; 1413 to = mc.to; 1414 if (!from) 1415 goto unlock; 1416 1417 ret = mem_cgroup_same_or_subtree(memcg, from) 1418 || mem_cgroup_same_or_subtree(memcg, to); 1419unlock: 1420 spin_unlock(&mc.lock); 1421 return ret; 1422} 1423 1424static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg) 1425{ 1426 if (mc.moving_task && current != mc.moving_task) { 1427 if (mem_cgroup_under_move(memcg)) { 1428 DEFINE_WAIT(wait); 1429 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1430 /* moving charge context might have finished. */ 1431 if (mc.moving_task) 1432 schedule(); 1433 finish_wait(&mc.waitq, &wait); 1434 return true; 1435 } 1436 } 1437 return false; 1438} 1439 1440/** 1441 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1442 * @memcg: The memory cgroup that went over limit 1443 * @p: Task that is going to be killed 1444 * 1445 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1446 * enabled 1447 */ 1448void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1449{ 1450 struct cgroup *task_cgrp; 1451 struct cgroup *mem_cgrp; 1452 /* 1453 * Need a buffer in BSS, can't rely on allocations. The code relies 1454 * on the assumption that OOM is serialized for memory controller. 1455 * If this assumption is broken, revisit this code. 1456 */ 1457 static char memcg_name[PATH_MAX]; 1458 int ret; 1459 1460 if (!memcg || !p) 1461 return; 1462 1463 1464 rcu_read_lock(); 1465 1466 mem_cgrp = memcg->css.cgroup; 1467 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1468 1469 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1470 if (ret < 0) { 1471 /* 1472 * Unfortunately, we are unable to convert to a useful name 1473 * But we'll still print out the usage information 1474 */ 1475 rcu_read_unlock(); 1476 goto done; 1477 } 1478 rcu_read_unlock(); 1479 1480 printk(KERN_INFO "Task in %s killed", memcg_name); 1481 1482 rcu_read_lock(); 1483 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1484 if (ret < 0) { 1485 rcu_read_unlock(); 1486 goto done; 1487 } 1488 rcu_read_unlock(); 1489 1490 /* 1491 * Continues from above, so we don't need an KERN_ level 1492 */ 1493 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1494done: 1495 1496 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1497 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1498 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1499 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1500 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1501 "failcnt %llu\n", 1502 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1503 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1504 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1505} 1506 1507/* 1508 * This function returns the number of memcg under hierarchy tree. Returns 1509 * 1(self count) if no children. 1510 */ 1511static int mem_cgroup_count_children(struct mem_cgroup *memcg) 1512{ 1513 int num = 0; 1514 struct mem_cgroup *iter; 1515 1516 for_each_mem_cgroup_tree(iter, memcg) 1517 num++; 1518 return num; 1519} 1520 1521/* 1522 * Return the memory (and swap, if configured) limit for a memcg. 1523 */ 1524u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1525{ 1526 u64 limit; 1527 u64 memsw; 1528 1529 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1530 limit += total_swap_pages << PAGE_SHIFT; 1531 1532 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1533 /* 1534 * If memsw is finite and limits the amount of swap space available 1535 * to this memcg, return that limit. 1536 */ 1537 return min(limit, memsw); 1538} 1539 1540static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg, 1541 gfp_t gfp_mask, 1542 unsigned long flags) 1543{ 1544 unsigned long total = 0; 1545 bool noswap = false; 1546 int loop; 1547 1548 if (flags & MEM_CGROUP_RECLAIM_NOSWAP) 1549 noswap = true; 1550 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum) 1551 noswap = true; 1552 1553 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) { 1554 if (loop) 1555 drain_all_stock_async(memcg); 1556 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap); 1557 /* 1558 * Allow limit shrinkers, which are triggered directly 1559 * by userspace, to catch signals and stop reclaim 1560 * after minimal progress, regardless of the margin. 1561 */ 1562 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK)) 1563 break; 1564 if (mem_cgroup_margin(memcg)) 1565 break; 1566 /* 1567 * If nothing was reclaimed after two attempts, there 1568 * may be no reclaimable pages in this hierarchy. 1569 */ 1570 if (loop && !total) 1571 break; 1572 } 1573 return total; 1574} 1575 1576/** 1577 * test_mem_cgroup_node_reclaimable 1578 * @mem: the target memcg 1579 * @nid: the node ID to be checked. 1580 * @noswap : specify true here if the user wants flle only information. 1581 * 1582 * This function returns whether the specified memcg contains any 1583 * reclaimable pages on a node. Returns true if there are any reclaimable 1584 * pages in the node. 1585 */ 1586static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg, 1587 int nid, bool noswap) 1588{ 1589 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE)) 1590 return true; 1591 if (noswap || !total_swap_pages) 1592 return false; 1593 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON)) 1594 return true; 1595 return false; 1596 1597} 1598#if MAX_NUMNODES > 1 1599 1600/* 1601 * Always updating the nodemask is not very good - even if we have an empty 1602 * list or the wrong list here, we can start from some node and traverse all 1603 * nodes based on the zonelist. So update the list loosely once per 10 secs. 1604 * 1605 */ 1606static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg) 1607{ 1608 int nid; 1609 /* 1610 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET 1611 * pagein/pageout changes since the last update. 1612 */ 1613 if (!atomic_read(&memcg->numainfo_events)) 1614 return; 1615 if (atomic_inc_return(&memcg->numainfo_updating) > 1) 1616 return; 1617 1618 /* make a nodemask where this memcg uses memory from */ 1619 memcg->scan_nodes = node_states[N_HIGH_MEMORY]; 1620 1621 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { 1622 1623 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false)) 1624 node_clear(nid, memcg->scan_nodes); 1625 } 1626 1627 atomic_set(&memcg->numainfo_events, 0); 1628 atomic_set(&memcg->numainfo_updating, 0); 1629} 1630 1631/* 1632 * Selecting a node where we start reclaim from. Because what we need is just 1633 * reducing usage counter, start from anywhere is O,K. Considering 1634 * memory reclaim from current node, there are pros. and cons. 1635 * 1636 * Freeing memory from current node means freeing memory from a node which 1637 * we'll use or we've used. So, it may make LRU bad. And if several threads 1638 * hit limits, it will see a contention on a node. But freeing from remote 1639 * node means more costs for memory reclaim because of memory latency. 1640 * 1641 * Now, we use round-robin. Better algorithm is welcomed. 1642 */ 1643int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1644{ 1645 int node; 1646 1647 mem_cgroup_may_update_nodemask(memcg); 1648 node = memcg->last_scanned_node; 1649 1650 node = next_node(node, memcg->scan_nodes); 1651 if (node == MAX_NUMNODES) 1652 node = first_node(memcg->scan_nodes); 1653 /* 1654 * We call this when we hit limit, not when pages are added to LRU. 1655 * No LRU may hold pages because all pages are UNEVICTABLE or 1656 * memcg is too small and all pages are not on LRU. In that case, 1657 * we use curret node. 1658 */ 1659 if (unlikely(node == MAX_NUMNODES)) 1660 node = numa_node_id(); 1661 1662 memcg->last_scanned_node = node; 1663 return node; 1664} 1665 1666/* 1667 * Check all nodes whether it contains reclaimable pages or not. 1668 * For quick scan, we make use of scan_nodes. This will allow us to skip 1669 * unused nodes. But scan_nodes is lazily updated and may not cotain 1670 * enough new information. We need to do double check. 1671 */ 1672bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1673{ 1674 int nid; 1675 1676 /* 1677 * quick check...making use of scan_node. 1678 * We can skip unused nodes. 1679 */ 1680 if (!nodes_empty(memcg->scan_nodes)) { 1681 for (nid = first_node(memcg->scan_nodes); 1682 nid < MAX_NUMNODES; 1683 nid = next_node(nid, memcg->scan_nodes)) { 1684 1685 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1686 return true; 1687 } 1688 } 1689 /* 1690 * Check rest of nodes. 1691 */ 1692 for_each_node_state(nid, N_HIGH_MEMORY) { 1693 if (node_isset(nid, memcg->scan_nodes)) 1694 continue; 1695 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap)) 1696 return true; 1697 } 1698 return false; 1699} 1700 1701#else 1702int mem_cgroup_select_victim_node(struct mem_cgroup *memcg) 1703{ 1704 return 0; 1705} 1706 1707bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap) 1708{ 1709 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap); 1710} 1711#endif 1712 1713static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg, 1714 struct zone *zone, 1715 gfp_t gfp_mask, 1716 unsigned long *total_scanned) 1717{ 1718 struct mem_cgroup *victim = NULL; 1719 int total = 0; 1720 int loop = 0; 1721 unsigned long excess; 1722 unsigned long nr_scanned; 1723 struct mem_cgroup_reclaim_cookie reclaim = { 1724 .zone = zone, 1725 .priority = 0, 1726 }; 1727 1728 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT; 1729 1730 while (1) { 1731 victim = mem_cgroup_iter(root_memcg, victim, &reclaim); 1732 if (!victim) { 1733 loop++; 1734 if (loop >= 2) { 1735 /* 1736 * If we have not been able to reclaim 1737 * anything, it might because there are 1738 * no reclaimable pages under this hierarchy 1739 */ 1740 if (!total) 1741 break; 1742 /* 1743 * We want to do more targeted reclaim. 1744 * excess >> 2 is not to excessive so as to 1745 * reclaim too much, nor too less that we keep 1746 * coming back to reclaim from this cgroup 1747 */ 1748 if (total >= (excess >> 2) || 1749 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) 1750 break; 1751 } 1752 continue; 1753 } 1754 if (!mem_cgroup_reclaimable(victim, false)) 1755 continue; 1756 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false, 1757 zone, &nr_scanned); 1758 *total_scanned += nr_scanned; 1759 if (!res_counter_soft_limit_excess(&root_memcg->res)) 1760 break; 1761 } 1762 mem_cgroup_iter_break(root_memcg, victim); 1763 return total; 1764} 1765 1766/* 1767 * Check OOM-Killer is already running under our hierarchy. 1768 * If someone is running, return false. 1769 * Has to be called with memcg_oom_lock 1770 */ 1771static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg) 1772{ 1773 struct mem_cgroup *iter, *failed = NULL; 1774 1775 for_each_mem_cgroup_tree(iter, memcg) { 1776 if (iter->oom_lock) { 1777 /* 1778 * this subtree of our hierarchy is already locked 1779 * so we cannot give a lock. 1780 */ 1781 failed = iter; 1782 mem_cgroup_iter_break(memcg, iter); 1783 break; 1784 } else 1785 iter->oom_lock = true; 1786 } 1787 1788 if (!failed) 1789 return true; 1790 1791 /* 1792 * OK, we failed to lock the whole subtree so we have to clean up 1793 * what we set up to the failing subtree 1794 */ 1795 for_each_mem_cgroup_tree(iter, memcg) { 1796 if (iter == failed) { 1797 mem_cgroup_iter_break(memcg, iter); 1798 break; 1799 } 1800 iter->oom_lock = false; 1801 } 1802 return false; 1803} 1804 1805/* 1806 * Has to be called with memcg_oom_lock 1807 */ 1808static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg) 1809{ 1810 struct mem_cgroup *iter; 1811 1812 for_each_mem_cgroup_tree(iter, memcg) 1813 iter->oom_lock = false; 1814 return 0; 1815} 1816 1817static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg) 1818{ 1819 struct mem_cgroup *iter; 1820 1821 for_each_mem_cgroup_tree(iter, memcg) 1822 atomic_inc(&iter->under_oom); 1823} 1824 1825static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg) 1826{ 1827 struct mem_cgroup *iter; 1828 1829 /* 1830 * When a new child is created while the hierarchy is under oom, 1831 * mem_cgroup_oom_lock() may not be called. We have to use 1832 * atomic_add_unless() here. 1833 */ 1834 for_each_mem_cgroup_tree(iter, memcg) 1835 atomic_add_unless(&iter->under_oom, -1, 0); 1836} 1837 1838static DEFINE_SPINLOCK(memcg_oom_lock); 1839static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1840 1841struct oom_wait_info { 1842 struct mem_cgroup *mem; 1843 wait_queue_t wait; 1844}; 1845 1846static int memcg_oom_wake_function(wait_queue_t *wait, 1847 unsigned mode, int sync, void *arg) 1848{ 1849 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg, 1850 *oom_wait_memcg; 1851 struct oom_wait_info *oom_wait_info; 1852 1853 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1854 oom_wait_memcg = oom_wait_info->mem; 1855 1856 /* 1857 * Both of oom_wait_info->mem and wake_mem are stable under us. 1858 * Then we can use css_is_ancestor without taking care of RCU. 1859 */ 1860 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg) 1861 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg)) 1862 return 0; 1863 return autoremove_wake_function(wait, mode, sync, arg); 1864} 1865 1866static void memcg_wakeup_oom(struct mem_cgroup *memcg) 1867{ 1868 /* for filtering, pass "memcg" as argument. */ 1869 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg); 1870} 1871 1872static void memcg_oom_recover(struct mem_cgroup *memcg) 1873{ 1874 if (memcg && atomic_read(&memcg->under_oom)) 1875 memcg_wakeup_oom(memcg); 1876} 1877 1878/* 1879 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1880 */ 1881bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask) 1882{ 1883 struct oom_wait_info owait; 1884 bool locked, need_to_kill; 1885 1886 owait.mem = memcg; 1887 owait.wait.flags = 0; 1888 owait.wait.func = memcg_oom_wake_function; 1889 owait.wait.private = current; 1890 INIT_LIST_HEAD(&owait.wait.task_list); 1891 need_to_kill = true; 1892 mem_cgroup_mark_under_oom(memcg); 1893 1894 /* At first, try to OOM lock hierarchy under memcg.*/ 1895 spin_lock(&memcg_oom_lock); 1896 locked = mem_cgroup_oom_lock(memcg); 1897 /* 1898 * Even if signal_pending(), we can't quit charge() loop without 1899 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1900 * under OOM is always welcomed, use TASK_KILLABLE here. 1901 */ 1902 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1903 if (!locked || memcg->oom_kill_disable) 1904 need_to_kill = false; 1905 if (locked) 1906 mem_cgroup_oom_notify(memcg); 1907 spin_unlock(&memcg_oom_lock); 1908 1909 if (need_to_kill) { 1910 finish_wait(&memcg_oom_waitq, &owait.wait); 1911 mem_cgroup_out_of_memory(memcg, mask); 1912 } else { 1913 schedule(); 1914 finish_wait(&memcg_oom_waitq, &owait.wait); 1915 } 1916 spin_lock(&memcg_oom_lock); 1917 if (locked) 1918 mem_cgroup_oom_unlock(memcg); 1919 memcg_wakeup_oom(memcg); 1920 spin_unlock(&memcg_oom_lock); 1921 1922 mem_cgroup_unmark_under_oom(memcg); 1923 1924 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1925 return false; 1926 /* Give chance to dying process */ 1927 schedule_timeout_uninterruptible(1); 1928 return true; 1929} 1930 1931/* 1932 * Currently used to update mapped file statistics, but the routine can be 1933 * generalized to update other statistics as well. 1934 * 1935 * Notes: Race condition 1936 * 1937 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1938 * it tends to be costly. But considering some conditions, we doesn't need 1939 * to do so _always_. 1940 * 1941 * Considering "charge", lock_page_cgroup() is not required because all 1942 * file-stat operations happen after a page is attached to radix-tree. There 1943 * are no race with "charge". 1944 * 1945 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1946 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1947 * if there are race with "uncharge". Statistics itself is properly handled 1948 * by flags. 1949 * 1950 * Considering "move", this is an only case we see a race. To make the race 1951 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1952 * possibility of race condition. If there is, we take a lock. 1953 */ 1954 1955void mem_cgroup_update_page_stat(struct page *page, 1956 enum mem_cgroup_page_stat_item idx, int val) 1957{ 1958 struct mem_cgroup *memcg; 1959 struct page_cgroup *pc = lookup_page_cgroup(page); 1960 bool need_unlock = false; 1961 unsigned long uninitialized_var(flags); 1962 1963 if (mem_cgroup_disabled()) 1964 return; 1965 1966 rcu_read_lock(); 1967 memcg = pc->mem_cgroup; 1968 if (unlikely(!memcg || !PageCgroupUsed(pc))) 1969 goto out; 1970 /* pc->mem_cgroup is unstable ? */ 1971 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) { 1972 /* take a lock against to access pc->mem_cgroup */ 1973 move_lock_page_cgroup(pc, &flags); 1974 need_unlock = true; 1975 memcg = pc->mem_cgroup; 1976 if (!memcg || !PageCgroupUsed(pc)) 1977 goto out; 1978 } 1979 1980 switch (idx) { 1981 case MEMCG_NR_FILE_MAPPED: 1982 if (val > 0) 1983 SetPageCgroupFileMapped(pc); 1984 else if (!page_mapped(page)) 1985 ClearPageCgroupFileMapped(pc); 1986 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1987 break; 1988 default: 1989 BUG(); 1990 } 1991 1992 this_cpu_add(memcg->stat->count[idx], val); 1993 1994out: 1995 if (unlikely(need_unlock)) 1996 move_unlock_page_cgroup(pc, &flags); 1997 rcu_read_unlock(); 1998 return; 1999} 2000EXPORT_SYMBOL(mem_cgroup_update_page_stat); 2001 2002/* 2003 * size of first charge trial. "32" comes from vmscan.c's magic value. 2004 * TODO: maybe necessary to use big numbers in big irons. 2005 */ 2006#define CHARGE_BATCH 32U 2007struct memcg_stock_pcp { 2008 struct mem_cgroup *cached; /* this never be root cgroup */ 2009 unsigned int nr_pages; 2010 struct work_struct work; 2011 unsigned long flags; 2012#define FLUSHING_CACHED_CHARGE (0) 2013}; 2014static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 2015static DEFINE_MUTEX(percpu_charge_mutex); 2016 2017/* 2018 * Try to consume stocked charge on this cpu. If success, one page is consumed 2019 * from local stock and true is returned. If the stock is 0 or charges from a 2020 * cgroup which is not current target, returns false. This stock will be 2021 * refilled. 2022 */ 2023static bool consume_stock(struct mem_cgroup *memcg) 2024{ 2025 struct memcg_stock_pcp *stock; 2026 bool ret = true; 2027 2028 stock = &get_cpu_var(memcg_stock); 2029 if (memcg == stock->cached && stock->nr_pages) 2030 stock->nr_pages--; 2031 else /* need to call res_counter_charge */ 2032 ret = false; 2033 put_cpu_var(memcg_stock); 2034 return ret; 2035} 2036 2037/* 2038 * Returns stocks cached in percpu to res_counter and reset cached information. 2039 */ 2040static void drain_stock(struct memcg_stock_pcp *stock) 2041{ 2042 struct mem_cgroup *old = stock->cached; 2043 2044 if (stock->nr_pages) { 2045 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 2046 2047 res_counter_uncharge(&old->res, bytes); 2048 if (do_swap_account) 2049 res_counter_uncharge(&old->memsw, bytes); 2050 stock->nr_pages = 0; 2051 } 2052 stock->cached = NULL; 2053} 2054 2055/* 2056 * This must be called under preempt disabled or must be called by 2057 * a thread which is pinned to local cpu. 2058 */ 2059static void drain_local_stock(struct work_struct *dummy) 2060{ 2061 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 2062 drain_stock(stock); 2063 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); 2064} 2065 2066/* 2067 * Cache charges(val) which is from res_counter, to local per_cpu area. 2068 * This will be consumed by consume_stock() function, later. 2069 */ 2070static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) 2071{ 2072 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 2073 2074 if (stock->cached != memcg) { /* reset if necessary */ 2075 drain_stock(stock); 2076 stock->cached = memcg; 2077 } 2078 stock->nr_pages += nr_pages; 2079 put_cpu_var(memcg_stock); 2080} 2081 2082/* 2083 * Drains all per-CPU charge caches for given root_memcg resp. subtree 2084 * of the hierarchy under it. sync flag says whether we should block 2085 * until the work is done. 2086 */ 2087static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync) 2088{ 2089 int cpu, curcpu; 2090 2091 /* Notify other cpus that system-wide "drain" is running */ 2092 get_online_cpus(); 2093 curcpu = get_cpu(); 2094 for_each_online_cpu(cpu) { 2095 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2096 struct mem_cgroup *memcg; 2097 2098 memcg = stock->cached; 2099 if (!memcg || !stock->nr_pages) 2100 continue; 2101 if (!mem_cgroup_same_or_subtree(root_memcg, memcg)) 2102 continue; 2103 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) { 2104 if (cpu == curcpu) 2105 drain_local_stock(&stock->work); 2106 else 2107 schedule_work_on(cpu, &stock->work); 2108 } 2109 } 2110 put_cpu(); 2111 2112 if (!sync) 2113 goto out; 2114 2115 for_each_online_cpu(cpu) { 2116 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2117 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) 2118 flush_work(&stock->work); 2119 } 2120out: 2121 put_online_cpus(); 2122} 2123 2124/* 2125 * Tries to drain stocked charges in other cpus. This function is asynchronous 2126 * and just put a work per cpu for draining localy on each cpu. Caller can 2127 * expects some charges will be back to res_counter later but cannot wait for 2128 * it. 2129 */ 2130static void drain_all_stock_async(struct mem_cgroup *root_memcg) 2131{ 2132 /* 2133 * If someone calls draining, avoid adding more kworker runs. 2134 */ 2135 if (!mutex_trylock(&percpu_charge_mutex)) 2136 return; 2137 drain_all_stock(root_memcg, false); 2138 mutex_unlock(&percpu_charge_mutex); 2139} 2140 2141/* This is a synchronous drain interface. */ 2142static void drain_all_stock_sync(struct mem_cgroup *root_memcg) 2143{ 2144 /* called when force_empty is called */ 2145 mutex_lock(&percpu_charge_mutex); 2146 drain_all_stock(root_memcg, true); 2147 mutex_unlock(&percpu_charge_mutex); 2148} 2149 2150/* 2151 * This function drains percpu counter value from DEAD cpu and 2152 * move it to local cpu. Note that this function can be preempted. 2153 */ 2154static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu) 2155{ 2156 int i; 2157 2158 spin_lock(&memcg->pcp_counter_lock); 2159 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 2160 long x = per_cpu(memcg->stat->count[i], cpu); 2161 2162 per_cpu(memcg->stat->count[i], cpu) = 0; 2163 memcg->nocpu_base.count[i] += x; 2164 } 2165 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { 2166 unsigned long x = per_cpu(memcg->stat->events[i], cpu); 2167 2168 per_cpu(memcg->stat->events[i], cpu) = 0; 2169 memcg->nocpu_base.events[i] += x; 2170 } 2171 /* need to clear ON_MOVE value, works as a kind of lock. */ 2172 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 2173 spin_unlock(&memcg->pcp_counter_lock); 2174} 2175 2176static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu) 2177{ 2178 int idx = MEM_CGROUP_ON_MOVE; 2179 2180 spin_lock(&memcg->pcp_counter_lock); 2181 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx]; 2182 spin_unlock(&memcg->pcp_counter_lock); 2183} 2184 2185static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 2186 unsigned long action, 2187 void *hcpu) 2188{ 2189 int cpu = (unsigned long)hcpu; 2190 struct memcg_stock_pcp *stock; 2191 struct mem_cgroup *iter; 2192 2193 if ((action == CPU_ONLINE)) { 2194 for_each_mem_cgroup(iter) 2195 synchronize_mem_cgroup_on_move(iter, cpu); 2196 return NOTIFY_OK; 2197 } 2198 2199 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 2200 return NOTIFY_OK; 2201 2202 for_each_mem_cgroup(iter) 2203 mem_cgroup_drain_pcp_counter(iter, cpu); 2204 2205 stock = &per_cpu(memcg_stock, cpu); 2206 drain_stock(stock); 2207 return NOTIFY_OK; 2208} 2209 2210 2211/* See __mem_cgroup_try_charge() for details */ 2212enum { 2213 CHARGE_OK, /* success */ 2214 CHARGE_RETRY, /* need to retry but retry is not bad */ 2215 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 2216 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 2217 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 2218}; 2219 2220static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask, 2221 unsigned int nr_pages, bool oom_check) 2222{ 2223 unsigned long csize = nr_pages * PAGE_SIZE; 2224 struct mem_cgroup *mem_over_limit; 2225 struct res_counter *fail_res; 2226 unsigned long flags = 0; 2227 int ret; 2228 2229 ret = res_counter_charge(&memcg->res, csize, &fail_res); 2230 2231 if (likely(!ret)) { 2232 if (!do_swap_account) 2233 return CHARGE_OK; 2234 ret = res_counter_charge(&memcg->memsw, csize, &fail_res); 2235 if (likely(!ret)) 2236 return CHARGE_OK; 2237 2238 res_counter_uncharge(&memcg->res, csize); 2239 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 2240 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 2241 } else 2242 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 2243 /* 2244 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch 2245 * of regular pages (CHARGE_BATCH), or a single regular page (1). 2246 * 2247 * Never reclaim on behalf of optional batching, retry with a 2248 * single page instead. 2249 */ 2250 if (nr_pages == CHARGE_BATCH) 2251 return CHARGE_RETRY; 2252 2253 if (!(gfp_mask & __GFP_WAIT)) 2254 return CHARGE_WOULDBLOCK; 2255 2256 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags); 2257 if (mem_cgroup_margin(mem_over_limit) >= nr_pages) 2258 return CHARGE_RETRY; 2259 /* 2260 * Even though the limit is exceeded at this point, reclaim 2261 * may have been able to free some pages. Retry the charge 2262 * before killing the task. 2263 * 2264 * Only for regular pages, though: huge pages are rather 2265 * unlikely to succeed so close to the limit, and we fall back 2266 * to regular pages anyway in case of failure. 2267 */ 2268 if (nr_pages == 1 && ret) 2269 return CHARGE_RETRY; 2270 2271 /* 2272 * At task move, charge accounts can be doubly counted. So, it's 2273 * better to wait until the end of task_move if something is going on. 2274 */ 2275 if (mem_cgroup_wait_acct_move(mem_over_limit)) 2276 return CHARGE_RETRY; 2277 2278 /* If we don't need to call oom-killer at el, return immediately */ 2279 if (!oom_check) 2280 return CHARGE_NOMEM; 2281 /* check OOM */ 2282 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 2283 return CHARGE_OOM_DIE; 2284 2285 return CHARGE_RETRY; 2286} 2287 2288/* 2289 * Unlike exported interface, "oom" parameter is added. if oom==true, 2290 * oom-killer can be invoked. 2291 */ 2292static int __mem_cgroup_try_charge(struct mm_struct *mm, 2293 gfp_t gfp_mask, 2294 unsigned int nr_pages, 2295 struct mem_cgroup **ptr, 2296 bool oom) 2297{ 2298 unsigned int batch = max(CHARGE_BATCH, nr_pages); 2299 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2300 struct mem_cgroup *memcg = NULL; 2301 int ret; 2302 2303 /* 2304 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 2305 * in system level. So, allow to go ahead dying process in addition to 2306 * MEMDIE process. 2307 */ 2308 if (unlikely(test_thread_flag(TIF_MEMDIE) 2309 || fatal_signal_pending(current))) 2310 goto bypass; 2311 2312 /* 2313 * We always charge the cgroup the mm_struct belongs to. 2314 * The mm_struct's mem_cgroup changes on task migration if the 2315 * thread group leader migrates. It's possible that mm is not 2316 * set, if so charge the init_mm (happens for pagecache usage). 2317 */ 2318 if (!*ptr && !mm) 2319 goto bypass; 2320again: 2321 if (*ptr) { /* css should be a valid one */ 2322 memcg = *ptr; 2323 VM_BUG_ON(css_is_removed(&memcg->css)); 2324 if (mem_cgroup_is_root(memcg)) 2325 goto done; 2326 if (nr_pages == 1 && consume_stock(memcg)) 2327 goto done; 2328 css_get(&memcg->css); 2329 } else { 2330 struct task_struct *p; 2331 2332 rcu_read_lock(); 2333 p = rcu_dereference(mm->owner); 2334 /* 2335 * Because we don't have task_lock(), "p" can exit. 2336 * In that case, "memcg" can point to root or p can be NULL with 2337 * race with swapoff. Then, we have small risk of mis-accouning. 2338 * But such kind of mis-account by race always happens because 2339 * we don't have cgroup_mutex(). It's overkill and we allo that 2340 * small race, here. 2341 * (*) swapoff at el will charge against mm-struct not against 2342 * task-struct. So, mm->owner can be NULL. 2343 */ 2344 memcg = mem_cgroup_from_task(p); 2345 if (!memcg || mem_cgroup_is_root(memcg)) { 2346 rcu_read_unlock(); 2347 goto done; 2348 } 2349 if (nr_pages == 1 && consume_stock(memcg)) { 2350 /* 2351 * It seems dagerous to access memcg without css_get(). 2352 * But considering how consume_stok works, it's not 2353 * necessary. If consume_stock success, some charges 2354 * from this memcg are cached on this cpu. So, we 2355 * don't need to call css_get()/css_tryget() before 2356 * calling consume_stock(). 2357 */ 2358 rcu_read_unlock(); 2359 goto done; 2360 } 2361 /* after here, we may be blocked. we need to get refcnt */ 2362 if (!css_tryget(&memcg->css)) { 2363 rcu_read_unlock(); 2364 goto again; 2365 } 2366 rcu_read_unlock(); 2367 } 2368 2369 do { 2370 bool oom_check; 2371 2372 /* If killed, bypass charge */ 2373 if (fatal_signal_pending(current)) { 2374 css_put(&memcg->css); 2375 goto bypass; 2376 } 2377 2378 oom_check = false; 2379 if (oom && !nr_oom_retries) { 2380 oom_check = true; 2381 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2382 } 2383 2384 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check); 2385 switch (ret) { 2386 case CHARGE_OK: 2387 break; 2388 case CHARGE_RETRY: /* not in OOM situation but retry */ 2389 batch = nr_pages; 2390 css_put(&memcg->css); 2391 memcg = NULL; 2392 goto again; 2393 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2394 css_put(&memcg->css); 2395 goto nomem; 2396 case CHARGE_NOMEM: /* OOM routine works */ 2397 if (!oom) { 2398 css_put(&memcg->css); 2399 goto nomem; 2400 } 2401 /* If oom, we never return -ENOMEM */ 2402 nr_oom_retries--; 2403 break; 2404 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2405 css_put(&memcg->css); 2406 goto bypass; 2407 } 2408 } while (ret != CHARGE_OK); 2409 2410 if (batch > nr_pages) 2411 refill_stock(memcg, batch - nr_pages); 2412 css_put(&memcg->css); 2413done: 2414 *ptr = memcg; 2415 return 0; 2416nomem: 2417 *ptr = NULL; 2418 return -ENOMEM; 2419bypass: 2420 *ptr = NULL; 2421 return 0; 2422} 2423 2424/* 2425 * Somemtimes we have to undo a charge we got by try_charge(). 2426 * This function is for that and do uncharge, put css's refcnt. 2427 * gotten by try_charge(). 2428 */ 2429static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg, 2430 unsigned int nr_pages) 2431{ 2432 if (!mem_cgroup_is_root(memcg)) { 2433 unsigned long bytes = nr_pages * PAGE_SIZE; 2434 2435 res_counter_uncharge(&memcg->res, bytes); 2436 if (do_swap_account) 2437 res_counter_uncharge(&memcg->memsw, bytes); 2438 } 2439} 2440 2441/* 2442 * A helper function to get mem_cgroup from ID. must be called under 2443 * rcu_read_lock(). The caller must check css_is_removed() or some if 2444 * it's concern. (dropping refcnt from swap can be called against removed 2445 * memcg.) 2446 */ 2447static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2448{ 2449 struct cgroup_subsys_state *css; 2450 2451 /* ID 0 is unused ID */ 2452 if (!id) 2453 return NULL; 2454 css = css_lookup(&mem_cgroup_subsys, id); 2455 if (!css) 2456 return NULL; 2457 return container_of(css, struct mem_cgroup, css); 2458} 2459 2460struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2461{ 2462 struct mem_cgroup *memcg = NULL; 2463 struct page_cgroup *pc; 2464 unsigned short id; 2465 swp_entry_t ent; 2466 2467 VM_BUG_ON(!PageLocked(page)); 2468 2469 pc = lookup_page_cgroup(page); 2470 lock_page_cgroup(pc); 2471 if (PageCgroupUsed(pc)) { 2472 memcg = pc->mem_cgroup; 2473 if (memcg && !css_tryget(&memcg->css)) 2474 memcg = NULL; 2475 } else if (PageSwapCache(page)) { 2476 ent.val = page_private(page); 2477 id = lookup_swap_cgroup_id(ent); 2478 rcu_read_lock(); 2479 memcg = mem_cgroup_lookup(id); 2480 if (memcg && !css_tryget(&memcg->css)) 2481 memcg = NULL; 2482 rcu_read_unlock(); 2483 } 2484 unlock_page_cgroup(pc); 2485 return memcg; 2486} 2487 2488static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg, 2489 struct page *page, 2490 unsigned int nr_pages, 2491 struct page_cgroup *pc, 2492 enum charge_type ctype) 2493{ 2494 lock_page_cgroup(pc); 2495 if (unlikely(PageCgroupUsed(pc))) { 2496 unlock_page_cgroup(pc); 2497 __mem_cgroup_cancel_charge(memcg, nr_pages); 2498 return; 2499 } 2500 /* 2501 * we don't need page_cgroup_lock about tail pages, becase they are not 2502 * accessed by any other context at this point. 2503 */ 2504 pc->mem_cgroup = memcg; 2505 /* 2506 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2507 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2508 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2509 * before USED bit, we need memory barrier here. 2510 * See mem_cgroup_add_lru_list(), etc. 2511 */ 2512 smp_wmb(); 2513 switch (ctype) { 2514 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2515 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2516 SetPageCgroupCache(pc); 2517 SetPageCgroupUsed(pc); 2518 break; 2519 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2520 ClearPageCgroupCache(pc); 2521 SetPageCgroupUsed(pc); 2522 break; 2523 default: 2524 break; 2525 } 2526 2527 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages); 2528 unlock_page_cgroup(pc); 2529 /* 2530 * "charge_statistics" updated event counter. Then, check it. 2531 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2532 * if they exceeds softlimit. 2533 */ 2534 memcg_check_events(memcg, page); 2535} 2536 2537#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2538 2539#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2540 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) 2541/* 2542 * Because tail pages are not marked as "used", set it. We're under 2543 * zone->lru_lock, 'splitting on pmd' and compound_lock. 2544 * charge/uncharge will be never happen and move_account() is done under 2545 * compound_lock(), so we don't have to take care of races. 2546 */ 2547void mem_cgroup_split_huge_fixup(struct page *head) 2548{ 2549 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2550 struct page_cgroup *pc; 2551 int i; 2552 2553 if (mem_cgroup_disabled()) 2554 return; 2555 for (i = 1; i < HPAGE_PMD_NR; i++) { 2556 pc = head_pc + i; 2557 pc->mem_cgroup = head_pc->mem_cgroup; 2558 smp_wmb();/* see __commit_charge() */ 2559 /* 2560 * LRU flags cannot be copied because we need to add tail 2561 * page to LRU by generic call and our hooks will be called. 2562 */ 2563 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2564 } 2565 2566 if (PageCgroupAcctLRU(head_pc)) { 2567 enum lru_list lru; 2568 struct mem_cgroup_per_zone *mz; 2569 /* 2570 * We hold lru_lock, then, reduce counter directly. 2571 */ 2572 lru = page_lru(head); 2573 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); 2574 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1; 2575 } 2576} 2577#endif 2578 2579/** 2580 * mem_cgroup_move_account - move account of the page 2581 * @page: the page 2582 * @nr_pages: number of regular pages (>1 for huge pages) 2583 * @pc: page_cgroup of the page. 2584 * @from: mem_cgroup which the page is moved from. 2585 * @to: mem_cgroup which the page is moved to. @from != @to. 2586 * @uncharge: whether we should call uncharge and css_put against @from. 2587 * 2588 * The caller must confirm following. 2589 * - page is not on LRU (isolate_page() is useful.) 2590 * - compound_lock is held when nr_pages > 1 2591 * 2592 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2593 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is 2594 * true, this function does "uncharge" from old cgroup, but it doesn't if 2595 * @uncharge is false, so a caller should do "uncharge". 2596 */ 2597static int mem_cgroup_move_account(struct page *page, 2598 unsigned int nr_pages, 2599 struct page_cgroup *pc, 2600 struct mem_cgroup *from, 2601 struct mem_cgroup *to, 2602 bool uncharge) 2603{ 2604 unsigned long flags; 2605 int ret; 2606 2607 VM_BUG_ON(from == to); 2608 VM_BUG_ON(PageLRU(page)); 2609 /* 2610 * The page is isolated from LRU. So, collapse function 2611 * will not handle this page. But page splitting can happen. 2612 * Do this check under compound_page_lock(). The caller should 2613 * hold it. 2614 */ 2615 ret = -EBUSY; 2616 if (nr_pages > 1 && !PageTransHuge(page)) 2617 goto out; 2618 2619 lock_page_cgroup(pc); 2620 2621 ret = -EINVAL; 2622 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2623 goto unlock; 2624 2625 move_lock_page_cgroup(pc, &flags); 2626 2627 if (PageCgroupFileMapped(pc)) { 2628 /* Update mapped_file data for mem_cgroup */ 2629 preempt_disable(); 2630 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2631 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2632 preempt_enable(); 2633 } 2634 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2635 if (uncharge) 2636 /* This is not "cancel", but cancel_charge does all we need. */ 2637 __mem_cgroup_cancel_charge(from, nr_pages); 2638 2639 /* caller should have done css_get */ 2640 pc->mem_cgroup = to; 2641 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2642 /* 2643 * We charges against "to" which may not have any tasks. Then, "to" 2644 * can be under rmdir(). But in current implementation, caller of 2645 * this function is just force_empty() and move charge, so it's 2646 * guaranteed that "to" is never removed. So, we don't check rmdir 2647 * status here. 2648 */ 2649 move_unlock_page_cgroup(pc, &flags); 2650 ret = 0; 2651unlock: 2652 unlock_page_cgroup(pc); 2653 /* 2654 * check events 2655 */ 2656 memcg_check_events(to, page); 2657 memcg_check_events(from, page); 2658out: 2659 return ret; 2660} 2661 2662/* 2663 * move charges to its parent. 2664 */ 2665 2666static int mem_cgroup_move_parent(struct page *page, 2667 struct page_cgroup *pc, 2668 struct mem_cgroup *child, 2669 gfp_t gfp_mask) 2670{ 2671 struct cgroup *cg = child->css.cgroup; 2672 struct cgroup *pcg = cg->parent; 2673 struct mem_cgroup *parent; 2674 unsigned int nr_pages; 2675 unsigned long uninitialized_var(flags); 2676 int ret; 2677 2678 /* Is ROOT ? */ 2679 if (!pcg) 2680 return -EINVAL; 2681 2682 ret = -EBUSY; 2683 if (!get_page_unless_zero(page)) 2684 goto out; 2685 if (isolate_lru_page(page)) 2686 goto put; 2687 2688 nr_pages = hpage_nr_pages(page); 2689 2690 parent = mem_cgroup_from_cont(pcg); 2691 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); 2692 if (ret || !parent) 2693 goto put_back; 2694 2695 if (nr_pages > 1) 2696 flags = compound_lock_irqsave(page); 2697 2698 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); 2699 if (ret) 2700 __mem_cgroup_cancel_charge(parent, nr_pages); 2701 2702 if (nr_pages > 1) 2703 compound_unlock_irqrestore(page, flags); 2704put_back: 2705 putback_lru_page(page); 2706put: 2707 put_page(page); 2708out: 2709 return ret; 2710} 2711 2712/* 2713 * Charge the memory controller for page usage. 2714 * Return 2715 * 0 if the charge was successful 2716 * < 0 if the cgroup is over its limit 2717 */ 2718static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2719 gfp_t gfp_mask, enum charge_type ctype) 2720{ 2721 struct mem_cgroup *memcg = NULL; 2722 unsigned int nr_pages = 1; 2723 struct page_cgroup *pc; 2724 bool oom = true; 2725 int ret; 2726 2727 if (PageTransHuge(page)) { 2728 nr_pages <<= compound_order(page); 2729 VM_BUG_ON(!PageTransHuge(page)); 2730 /* 2731 * Never OOM-kill a process for a huge page. The 2732 * fault handler will fall back to regular pages. 2733 */ 2734 oom = false; 2735 } 2736 2737 pc = lookup_page_cgroup(page); 2738 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom); 2739 if (ret || !memcg) 2740 return ret; 2741 2742 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype); 2743 return 0; 2744} 2745 2746int mem_cgroup_newpage_charge(struct page *page, 2747 struct mm_struct *mm, gfp_t gfp_mask) 2748{ 2749 if (mem_cgroup_disabled()) 2750 return 0; 2751 VM_BUG_ON(page_mapped(page)); 2752 VM_BUG_ON(page->mapping && !PageAnon(page)); 2753 VM_BUG_ON(!mm); 2754 return mem_cgroup_charge_common(page, mm, gfp_mask, 2755 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2756} 2757 2758static void 2759__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2760 enum charge_type ctype); 2761 2762static void 2763__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg, 2764 enum charge_type ctype) 2765{ 2766 struct page_cgroup *pc = lookup_page_cgroup(page); 2767 /* 2768 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page 2769 * is already on LRU. It means the page may on some other page_cgroup's 2770 * LRU. Take care of it. 2771 */ 2772 mem_cgroup_lru_del_before_commit(page); 2773 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 2774 mem_cgroup_lru_add_after_commit(page); 2775 return; 2776} 2777 2778int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2779 gfp_t gfp_mask) 2780{ 2781 struct mem_cgroup *memcg = NULL; 2782 int ret; 2783 2784 if (mem_cgroup_disabled()) 2785 return 0; 2786 if (PageCompound(page)) 2787 return 0; 2788 2789 if (unlikely(!mm)) 2790 mm = &init_mm; 2791 2792 if (page_is_file_cache(page)) { 2793 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true); 2794 if (ret || !memcg) 2795 return ret; 2796 2797 /* 2798 * FUSE reuses pages without going through the final 2799 * put that would remove them from the LRU list, make 2800 * sure that they get relinked properly. 2801 */ 2802 __mem_cgroup_commit_charge_lrucare(page, memcg, 2803 MEM_CGROUP_CHARGE_TYPE_CACHE); 2804 return ret; 2805 } 2806 /* shmem */ 2807 if (PageSwapCache(page)) { 2808 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg); 2809 if (!ret) 2810 __mem_cgroup_commit_charge_swapin(page, memcg, 2811 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2812 } else 2813 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 2814 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2815 2816 return ret; 2817} 2818 2819/* 2820 * While swap-in, try_charge -> commit or cancel, the page is locked. 2821 * And when try_charge() successfully returns, one refcnt to memcg without 2822 * struct page_cgroup is acquired. This refcnt will be consumed by 2823 * "commit()" or removed by "cancel()" 2824 */ 2825int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2826 struct page *page, 2827 gfp_t mask, struct mem_cgroup **memcgp) 2828{ 2829 struct mem_cgroup *memcg; 2830 int ret; 2831 2832 *memcgp = NULL; 2833 2834 if (mem_cgroup_disabled()) 2835 return 0; 2836 2837 if (!do_swap_account) 2838 goto charge_cur_mm; 2839 /* 2840 * A racing thread's fault, or swapoff, may have already updated 2841 * the pte, and even removed page from swap cache: in those cases 2842 * do_swap_page()'s pte_same() test will fail; but there's also a 2843 * KSM case which does need to charge the page. 2844 */ 2845 if (!PageSwapCache(page)) 2846 goto charge_cur_mm; 2847 memcg = try_get_mem_cgroup_from_page(page); 2848 if (!memcg) 2849 goto charge_cur_mm; 2850 *memcgp = memcg; 2851 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true); 2852 css_put(&memcg->css); 2853 return ret; 2854charge_cur_mm: 2855 if (unlikely(!mm)) 2856 mm = &init_mm; 2857 return __mem_cgroup_try_charge(mm, mask, 1, memcgp, true); 2858} 2859 2860static void 2861__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg, 2862 enum charge_type ctype) 2863{ 2864 if (mem_cgroup_disabled()) 2865 return; 2866 if (!memcg) 2867 return; 2868 cgroup_exclude_rmdir(&memcg->css); 2869 2870 __mem_cgroup_commit_charge_lrucare(page, memcg, ctype); 2871 /* 2872 * Now swap is on-memory. This means this page may be 2873 * counted both as mem and swap....double count. 2874 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2875 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2876 * may call delete_from_swap_cache() before reach here. 2877 */ 2878 if (do_swap_account && PageSwapCache(page)) { 2879 swp_entry_t ent = {.val = page_private(page)}; 2880 struct mem_cgroup *swap_memcg; 2881 unsigned short id; 2882 2883 id = swap_cgroup_record(ent, 0); 2884 rcu_read_lock(); 2885 swap_memcg = mem_cgroup_lookup(id); 2886 if (swap_memcg) { 2887 /* 2888 * This recorded memcg can be obsolete one. So, avoid 2889 * calling css_tryget 2890 */ 2891 if (!mem_cgroup_is_root(swap_memcg)) 2892 res_counter_uncharge(&swap_memcg->memsw, 2893 PAGE_SIZE); 2894 mem_cgroup_swap_statistics(swap_memcg, false); 2895 mem_cgroup_put(swap_memcg); 2896 } 2897 rcu_read_unlock(); 2898 } 2899 /* 2900 * At swapin, we may charge account against cgroup which has no tasks. 2901 * So, rmdir()->pre_destroy() can be called while we do this charge. 2902 * In that case, we need to call pre_destroy() again. check it here. 2903 */ 2904 cgroup_release_and_wakeup_rmdir(&memcg->css); 2905} 2906 2907void mem_cgroup_commit_charge_swapin(struct page *page, 2908 struct mem_cgroup *memcg) 2909{ 2910 __mem_cgroup_commit_charge_swapin(page, memcg, 2911 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2912} 2913 2914void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg) 2915{ 2916 if (mem_cgroup_disabled()) 2917 return; 2918 if (!memcg) 2919 return; 2920 __mem_cgroup_cancel_charge(memcg, 1); 2921} 2922 2923static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg, 2924 unsigned int nr_pages, 2925 const enum charge_type ctype) 2926{ 2927 struct memcg_batch_info *batch = NULL; 2928 bool uncharge_memsw = true; 2929 2930 /* If swapout, usage of swap doesn't decrease */ 2931 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2932 uncharge_memsw = false; 2933 2934 batch = ¤t->memcg_batch; 2935 /* 2936 * In usual, we do css_get() when we remember memcg pointer. 2937 * But in this case, we keep res->usage until end of a series of 2938 * uncharges. Then, it's ok to ignore memcg's refcnt. 2939 */ 2940 if (!batch->memcg) 2941 batch->memcg = memcg; 2942 /* 2943 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2944 * In those cases, all pages freed continuously can be expected to be in 2945 * the same cgroup and we have chance to coalesce uncharges. 2946 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2947 * because we want to do uncharge as soon as possible. 2948 */ 2949 2950 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2951 goto direct_uncharge; 2952 2953 if (nr_pages > 1) 2954 goto direct_uncharge; 2955 2956 /* 2957 * In typical case, batch->memcg == mem. This means we can 2958 * merge a series of uncharges to an uncharge of res_counter. 2959 * If not, we uncharge res_counter ony by one. 2960 */ 2961 if (batch->memcg != memcg) 2962 goto direct_uncharge; 2963 /* remember freed charge and uncharge it later */ 2964 batch->nr_pages++; 2965 if (uncharge_memsw) 2966 batch->memsw_nr_pages++; 2967 return; 2968direct_uncharge: 2969 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE); 2970 if (uncharge_memsw) 2971 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE); 2972 if (unlikely(batch->memcg != memcg)) 2973 memcg_oom_recover(memcg); 2974 return; 2975} 2976 2977/* 2978 * uncharge if !page_mapped(page) 2979 */ 2980static struct mem_cgroup * 2981__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 2982{ 2983 struct mem_cgroup *memcg = NULL; 2984 unsigned int nr_pages = 1; 2985 struct page_cgroup *pc; 2986 2987 if (mem_cgroup_disabled()) 2988 return NULL; 2989 2990 if (PageSwapCache(page)) 2991 return NULL; 2992 2993 if (PageTransHuge(page)) { 2994 nr_pages <<= compound_order(page); 2995 VM_BUG_ON(!PageTransHuge(page)); 2996 } 2997 /* 2998 * Check if our page_cgroup is valid 2999 */ 3000 pc = lookup_page_cgroup(page); 3001 if (unlikely(!PageCgroupUsed(pc))) 3002 return NULL; 3003 3004 lock_page_cgroup(pc); 3005 3006 memcg = pc->mem_cgroup; 3007 3008 if (!PageCgroupUsed(pc)) 3009 goto unlock_out; 3010 3011 switch (ctype) { 3012 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 3013 case MEM_CGROUP_CHARGE_TYPE_DROP: 3014 /* See mem_cgroup_prepare_migration() */ 3015 if (page_mapped(page) || PageCgroupMigration(pc)) 3016 goto unlock_out; 3017 break; 3018 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 3019 if (!PageAnon(page)) { /* Shared memory */ 3020 if (page->mapping && !page_is_file_cache(page)) 3021 goto unlock_out; 3022 } else if (page_mapped(page)) /* Anon */ 3023 goto unlock_out; 3024 break; 3025 default: 3026 break; 3027 } 3028 3029 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages); 3030 3031 ClearPageCgroupUsed(pc); 3032 /* 3033 * pc->mem_cgroup is not cleared here. It will be accessed when it's 3034 * freed from LRU. This is safe because uncharged page is expected not 3035 * to be reused (freed soon). Exception is SwapCache, it's handled by 3036 * special functions. 3037 */ 3038 3039 unlock_page_cgroup(pc); 3040 /* 3041 * even after unlock, we have memcg->res.usage here and this memcg 3042 * will never be freed. 3043 */ 3044 memcg_check_events(memcg, page); 3045 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 3046 mem_cgroup_swap_statistics(memcg, true); 3047 mem_cgroup_get(memcg); 3048 } 3049 if (!mem_cgroup_is_root(memcg)) 3050 mem_cgroup_do_uncharge(memcg, nr_pages, ctype); 3051 3052 return memcg; 3053 3054unlock_out: 3055 unlock_page_cgroup(pc); 3056 return NULL; 3057} 3058 3059void mem_cgroup_uncharge_page(struct page *page) 3060{ 3061 /* early check. */ 3062 if (page_mapped(page)) 3063 return; 3064 VM_BUG_ON(page->mapping && !PageAnon(page)); 3065 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 3066} 3067 3068void mem_cgroup_uncharge_cache_page(struct page *page) 3069{ 3070 VM_BUG_ON(page_mapped(page)); 3071 VM_BUG_ON(page->mapping); 3072 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 3073} 3074 3075/* 3076 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 3077 * In that cases, pages are freed continuously and we can expect pages 3078 * are in the same memcg. All these calls itself limits the number of 3079 * pages freed at once, then uncharge_start/end() is called properly. 3080 * This may be called prural(2) times in a context, 3081 */ 3082 3083void mem_cgroup_uncharge_start(void) 3084{ 3085 current->memcg_batch.do_batch++; 3086 /* We can do nest. */ 3087 if (current->memcg_batch.do_batch == 1) { 3088 current->memcg_batch.memcg = NULL; 3089 current->memcg_batch.nr_pages = 0; 3090 current->memcg_batch.memsw_nr_pages = 0; 3091 } 3092} 3093 3094void mem_cgroup_uncharge_end(void) 3095{ 3096 struct memcg_batch_info *batch = ¤t->memcg_batch; 3097 3098 if (!batch->do_batch) 3099 return; 3100 3101 batch->do_batch--; 3102 if (batch->do_batch) /* If stacked, do nothing. */ 3103 return; 3104 3105 if (!batch->memcg) 3106 return; 3107 /* 3108 * This "batch->memcg" is valid without any css_get/put etc... 3109 * bacause we hide charges behind us. 3110 */ 3111 if (batch->nr_pages) 3112 res_counter_uncharge(&batch->memcg->res, 3113 batch->nr_pages * PAGE_SIZE); 3114 if (batch->memsw_nr_pages) 3115 res_counter_uncharge(&batch->memcg->memsw, 3116 batch->memsw_nr_pages * PAGE_SIZE); 3117 memcg_oom_recover(batch->memcg); 3118 /* forget this pointer (for sanity check) */ 3119 batch->memcg = NULL; 3120} 3121 3122#ifdef CONFIG_SWAP 3123/* 3124 * called after __delete_from_swap_cache() and drop "page" account. 3125 * memcg information is recorded to swap_cgroup of "ent" 3126 */ 3127void 3128mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 3129{ 3130 struct mem_cgroup *memcg; 3131 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 3132 3133 if (!swapout) /* this was a swap cache but the swap is unused ! */ 3134 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 3135 3136 memcg = __mem_cgroup_uncharge_common(page, ctype); 3137 3138 /* 3139 * record memcg information, if swapout && memcg != NULL, 3140 * mem_cgroup_get() was called in uncharge(). 3141 */ 3142 if (do_swap_account && swapout && memcg) 3143 swap_cgroup_record(ent, css_id(&memcg->css)); 3144} 3145#endif 3146 3147#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3148/* 3149 * called from swap_entry_free(). remove record in swap_cgroup and 3150 * uncharge "memsw" account. 3151 */ 3152void mem_cgroup_uncharge_swap(swp_entry_t ent) 3153{ 3154 struct mem_cgroup *memcg; 3155 unsigned short id; 3156 3157 if (!do_swap_account) 3158 return; 3159 3160 id = swap_cgroup_record(ent, 0); 3161 rcu_read_lock(); 3162 memcg = mem_cgroup_lookup(id); 3163 if (memcg) { 3164 /* 3165 * We uncharge this because swap is freed. 3166 * This memcg can be obsolete one. We avoid calling css_tryget 3167 */ 3168 if (!mem_cgroup_is_root(memcg)) 3169 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 3170 mem_cgroup_swap_statistics(memcg, false); 3171 mem_cgroup_put(memcg); 3172 } 3173 rcu_read_unlock(); 3174} 3175 3176/** 3177 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 3178 * @entry: swap entry to be moved 3179 * @from: mem_cgroup which the entry is moved from 3180 * @to: mem_cgroup which the entry is moved to 3181 * @need_fixup: whether we should fixup res_counters and refcounts. 3182 * 3183 * It succeeds only when the swap_cgroup's record for this entry is the same 3184 * as the mem_cgroup's id of @from. 3185 * 3186 * Returns 0 on success, -EINVAL on failure. 3187 * 3188 * The caller must have charged to @to, IOW, called res_counter_charge() about 3189 * both res and memsw, and called css_get(). 3190 */ 3191static int mem_cgroup_move_swap_account(swp_entry_t entry, 3192 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3193{ 3194 unsigned short old_id, new_id; 3195 3196 old_id = css_id(&from->css); 3197 new_id = css_id(&to->css); 3198 3199 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 3200 mem_cgroup_swap_statistics(from, false); 3201 mem_cgroup_swap_statistics(to, true); 3202 /* 3203 * This function is only called from task migration context now. 3204 * It postpones res_counter and refcount handling till the end 3205 * of task migration(mem_cgroup_clear_mc()) for performance 3206 * improvement. But we cannot postpone mem_cgroup_get(to) 3207 * because if the process that has been moved to @to does 3208 * swap-in, the refcount of @to might be decreased to 0. 3209 */ 3210 mem_cgroup_get(to); 3211 if (need_fixup) { 3212 if (!mem_cgroup_is_root(from)) 3213 res_counter_uncharge(&from->memsw, PAGE_SIZE); 3214 mem_cgroup_put(from); 3215 /* 3216 * we charged both to->res and to->memsw, so we should 3217 * uncharge to->res. 3218 */ 3219 if (!mem_cgroup_is_root(to)) 3220 res_counter_uncharge(&to->res, PAGE_SIZE); 3221 } 3222 return 0; 3223 } 3224 return -EINVAL; 3225} 3226#else 3227static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 3228 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3229{ 3230 return -EINVAL; 3231} 3232#endif 3233 3234/* 3235 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 3236 * page belongs to. 3237 */ 3238int mem_cgroup_prepare_migration(struct page *page, 3239 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask) 3240{ 3241 struct mem_cgroup *memcg = NULL; 3242 struct page_cgroup *pc; 3243 enum charge_type ctype; 3244 int ret = 0; 3245 3246 *memcgp = NULL; 3247 3248 VM_BUG_ON(PageTransHuge(page)); 3249 if (mem_cgroup_disabled()) 3250 return 0; 3251 3252 pc = lookup_page_cgroup(page); 3253 lock_page_cgroup(pc); 3254 if (PageCgroupUsed(pc)) { 3255 memcg = pc->mem_cgroup; 3256 css_get(&memcg->css); 3257 /* 3258 * At migrating an anonymous page, its mapcount goes down 3259 * to 0 and uncharge() will be called. But, even if it's fully 3260 * unmapped, migration may fail and this page has to be 3261 * charged again. We set MIGRATION flag here and delay uncharge 3262 * until end_migration() is called 3263 * 3264 * Corner Case Thinking 3265 * A) 3266 * When the old page was mapped as Anon and it's unmap-and-freed 3267 * while migration was ongoing. 3268 * If unmap finds the old page, uncharge() of it will be delayed 3269 * until end_migration(). If unmap finds a new page, it's 3270 * uncharged when it make mapcount to be 1->0. If unmap code 3271 * finds swap_migration_entry, the new page will not be mapped 3272 * and end_migration() will find it(mapcount==0). 3273 * 3274 * B) 3275 * When the old page was mapped but migraion fails, the kernel 3276 * remaps it. A charge for it is kept by MIGRATION flag even 3277 * if mapcount goes down to 0. We can do remap successfully 3278 * without charging it again. 3279 * 3280 * C) 3281 * The "old" page is under lock_page() until the end of 3282 * migration, so, the old page itself will not be swapped-out. 3283 * If the new page is swapped out before end_migraton, our 3284 * hook to usual swap-out path will catch the event. 3285 */ 3286 if (PageAnon(page)) 3287 SetPageCgroupMigration(pc); 3288 } 3289 unlock_page_cgroup(pc); 3290 /* 3291 * If the page is not charged at this point, 3292 * we return here. 3293 */ 3294 if (!memcg) 3295 return 0; 3296 3297 *memcgp = memcg; 3298 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false); 3299 css_put(&memcg->css);/* drop extra refcnt */ 3300 if (ret || *memcgp == NULL) { 3301 if (PageAnon(page)) { 3302 lock_page_cgroup(pc); 3303 ClearPageCgroupMigration(pc); 3304 unlock_page_cgroup(pc); 3305 /* 3306 * The old page may be fully unmapped while we kept it. 3307 */ 3308 mem_cgroup_uncharge_page(page); 3309 } 3310 return -ENOMEM; 3311 } 3312 /* 3313 * We charge new page before it's used/mapped. So, even if unlock_page() 3314 * is called before end_migration, we can catch all events on this new 3315 * page. In the case new page is migrated but not remapped, new page's 3316 * mapcount will be finally 0 and we call uncharge in end_migration(). 3317 */ 3318 pc = lookup_page_cgroup(newpage); 3319 if (PageAnon(page)) 3320 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 3321 else if (page_is_file_cache(page)) 3322 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 3323 else 3324 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3325 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype); 3326 return ret; 3327} 3328 3329/* remove redundant charge if migration failed*/ 3330void mem_cgroup_end_migration(struct mem_cgroup *memcg, 3331 struct page *oldpage, struct page *newpage, bool migration_ok) 3332{ 3333 struct page *used, *unused; 3334 struct page_cgroup *pc; 3335 3336 if (!memcg) 3337 return; 3338 /* blocks rmdir() */ 3339 cgroup_exclude_rmdir(&memcg->css); 3340 if (!migration_ok) { 3341 used = oldpage; 3342 unused = newpage; 3343 } else { 3344 used = newpage; 3345 unused = oldpage; 3346 } 3347 /* 3348 * We disallowed uncharge of pages under migration because mapcount 3349 * of the page goes down to zero, temporarly. 3350 * Clear the flag and check the page should be charged. 3351 */ 3352 pc = lookup_page_cgroup(oldpage); 3353 lock_page_cgroup(pc); 3354 ClearPageCgroupMigration(pc); 3355 unlock_page_cgroup(pc); 3356 3357 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 3358 3359 /* 3360 * If a page is a file cache, radix-tree replacement is very atomic 3361 * and we can skip this check. When it was an Anon page, its mapcount 3362 * goes down to 0. But because we added MIGRATION flage, it's not 3363 * uncharged yet. There are several case but page->mapcount check 3364 * and USED bit check in mem_cgroup_uncharge_page() will do enough 3365 * check. (see prepare_charge() also) 3366 */ 3367 if (PageAnon(used)) 3368 mem_cgroup_uncharge_page(used); 3369 /* 3370 * At migration, we may charge account against cgroup which has no 3371 * tasks. 3372 * So, rmdir()->pre_destroy() can be called while we do this charge. 3373 * In that case, we need to call pre_destroy() again. check it here. 3374 */ 3375 cgroup_release_and_wakeup_rmdir(&memcg->css); 3376} 3377 3378/* 3379 * At replace page cache, newpage is not under any memcg but it's on 3380 * LRU. So, this function doesn't touch res_counter but handles LRU 3381 * in correct way. Both pages are locked so we cannot race with uncharge. 3382 */ 3383void mem_cgroup_replace_page_cache(struct page *oldpage, 3384 struct page *newpage) 3385{ 3386 struct mem_cgroup *memcg; 3387 struct page_cgroup *pc; 3388 struct zone *zone; 3389 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE; 3390 unsigned long flags; 3391 3392 if (mem_cgroup_disabled()) 3393 return; 3394 3395 pc = lookup_page_cgroup(oldpage); 3396 /* fix accounting on old pages */ 3397 lock_page_cgroup(pc); 3398 memcg = pc->mem_cgroup; 3399 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1); 3400 ClearPageCgroupUsed(pc); 3401 unlock_page_cgroup(pc); 3402 3403 if (PageSwapBacked(oldpage)) 3404 type = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3405 3406 zone = page_zone(newpage); 3407 pc = lookup_page_cgroup(newpage); 3408 /* 3409 * Even if newpage->mapping was NULL before starting replacement, 3410 * the newpage may be on LRU(or pagevec for LRU) already. We lock 3411 * LRU while we overwrite pc->mem_cgroup. 3412 */ 3413 spin_lock_irqsave(&zone->lru_lock, flags); 3414 if (PageLRU(newpage)) 3415 del_page_from_lru_list(zone, newpage, page_lru(newpage)); 3416 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type); 3417 if (PageLRU(newpage)) 3418 add_page_to_lru_list(zone, newpage, page_lru(newpage)); 3419 spin_unlock_irqrestore(&zone->lru_lock, flags); 3420} 3421 3422#ifdef CONFIG_DEBUG_VM 3423static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3424{ 3425 struct page_cgroup *pc; 3426 3427 pc = lookup_page_cgroup(page); 3428 /* 3429 * Can be NULL while feeding pages into the page allocator for 3430 * the first time, i.e. during boot or memory hotplug; 3431 * or when mem_cgroup_disabled(). 3432 */ 3433 if (likely(pc) && PageCgroupUsed(pc)) 3434 return pc; 3435 return NULL; 3436} 3437 3438bool mem_cgroup_bad_page_check(struct page *page) 3439{ 3440 if (mem_cgroup_disabled()) 3441 return false; 3442 3443 return lookup_page_cgroup_used(page) != NULL; 3444} 3445 3446void mem_cgroup_print_bad_page(struct page *page) 3447{ 3448 struct page_cgroup *pc; 3449 3450 pc = lookup_page_cgroup_used(page); 3451 if (pc) { 3452 int ret = -1; 3453 char *path; 3454 3455 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", 3456 pc, pc->flags, pc->mem_cgroup); 3457 3458 path = kmalloc(PATH_MAX, GFP_KERNEL); 3459 if (path) { 3460 rcu_read_lock(); 3461 ret = cgroup_path(pc->mem_cgroup->css.cgroup, 3462 path, PATH_MAX); 3463 rcu_read_unlock(); 3464 } 3465 3466 printk(KERN_CONT "(%s)\n", 3467 (ret < 0) ? "cannot get the path" : path); 3468 kfree(path); 3469 } 3470} 3471#endif 3472 3473static DEFINE_MUTEX(set_limit_mutex); 3474 3475static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3476 unsigned long long val) 3477{ 3478 int retry_count; 3479 u64 memswlimit, memlimit; 3480 int ret = 0; 3481 int children = mem_cgroup_count_children(memcg); 3482 u64 curusage, oldusage; 3483 int enlarge; 3484 3485 /* 3486 * For keeping hierarchical_reclaim simple, how long we should retry 3487 * is depends on callers. We set our retry-count to be function 3488 * of # of children which we should visit in this loop. 3489 */ 3490 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3491 3492 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3493 3494 enlarge = 0; 3495 while (retry_count) { 3496 if (signal_pending(current)) { 3497 ret = -EINTR; 3498 break; 3499 } 3500 /* 3501 * Rather than hide all in some function, I do this in 3502 * open coded manner. You see what this really does. 3503 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3504 */ 3505 mutex_lock(&set_limit_mutex); 3506 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3507 if (memswlimit < val) { 3508 ret = -EINVAL; 3509 mutex_unlock(&set_limit_mutex); 3510 break; 3511 } 3512 3513 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3514 if (memlimit < val) 3515 enlarge = 1; 3516 3517 ret = res_counter_set_limit(&memcg->res, val); 3518 if (!ret) { 3519 if (memswlimit == val) 3520 memcg->memsw_is_minimum = true; 3521 else 3522 memcg->memsw_is_minimum = false; 3523 } 3524 mutex_unlock(&set_limit_mutex); 3525 3526 if (!ret) 3527 break; 3528 3529 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3530 MEM_CGROUP_RECLAIM_SHRINK); 3531 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3532 /* Usage is reduced ? */ 3533 if (curusage >= oldusage) 3534 retry_count--; 3535 else 3536 oldusage = curusage; 3537 } 3538 if (!ret && enlarge) 3539 memcg_oom_recover(memcg); 3540 3541 return ret; 3542} 3543 3544static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3545 unsigned long long val) 3546{ 3547 int retry_count; 3548 u64 memlimit, memswlimit, oldusage, curusage; 3549 int children = mem_cgroup_count_children(memcg); 3550 int ret = -EBUSY; 3551 int enlarge = 0; 3552 3553 /* see mem_cgroup_resize_res_limit */ 3554 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3555 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3556 while (retry_count) { 3557 if (signal_pending(current)) { 3558 ret = -EINTR; 3559 break; 3560 } 3561 /* 3562 * Rather than hide all in some function, I do this in 3563 * open coded manner. You see what this really does. 3564 * We have to guarantee memcg->res.limit < memcg->memsw.limit. 3565 */ 3566 mutex_lock(&set_limit_mutex); 3567 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3568 if (memlimit > val) { 3569 ret = -EINVAL; 3570 mutex_unlock(&set_limit_mutex); 3571 break; 3572 } 3573 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3574 if (memswlimit < val) 3575 enlarge = 1; 3576 ret = res_counter_set_limit(&memcg->memsw, val); 3577 if (!ret) { 3578 if (memlimit == val) 3579 memcg->memsw_is_minimum = true; 3580 else 3581 memcg->memsw_is_minimum = false; 3582 } 3583 mutex_unlock(&set_limit_mutex); 3584 3585 if (!ret) 3586 break; 3587 3588 mem_cgroup_reclaim(memcg, GFP_KERNEL, 3589 MEM_CGROUP_RECLAIM_NOSWAP | 3590 MEM_CGROUP_RECLAIM_SHRINK); 3591 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3592 /* Usage is reduced ? */ 3593 if (curusage >= oldusage) 3594 retry_count--; 3595 else 3596 oldusage = curusage; 3597 } 3598 if (!ret && enlarge) 3599 memcg_oom_recover(memcg); 3600 return ret; 3601} 3602 3603unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3604 gfp_t gfp_mask, 3605 unsigned long *total_scanned) 3606{ 3607 unsigned long nr_reclaimed = 0; 3608 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3609 unsigned long reclaimed; 3610 int loop = 0; 3611 struct mem_cgroup_tree_per_zone *mctz; 3612 unsigned long long excess; 3613 unsigned long nr_scanned; 3614 3615 if (order > 0) 3616 return 0; 3617 3618 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3619 /* 3620 * This loop can run a while, specially if mem_cgroup's continuously 3621 * keep exceeding their soft limit and putting the system under 3622 * pressure 3623 */ 3624 do { 3625 if (next_mz) 3626 mz = next_mz; 3627 else 3628 mz = mem_cgroup_largest_soft_limit_node(mctz); 3629 if (!mz) 3630 break; 3631 3632 nr_scanned = 0; 3633 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone, 3634 gfp_mask, &nr_scanned); 3635 nr_reclaimed += reclaimed; 3636 *total_scanned += nr_scanned; 3637 spin_lock(&mctz->lock); 3638 3639 /* 3640 * If we failed to reclaim anything from this memory cgroup 3641 * it is time to move on to the next cgroup 3642 */ 3643 next_mz = NULL; 3644 if (!reclaimed) { 3645 do { 3646 /* 3647 * Loop until we find yet another one. 3648 * 3649 * By the time we get the soft_limit lock 3650 * again, someone might have aded the 3651 * group back on the RB tree. Iterate to 3652 * make sure we get a different mem. 3653 * mem_cgroup_largest_soft_limit_node returns 3654 * NULL if no other cgroup is present on 3655 * the tree 3656 */ 3657 next_mz = 3658 __mem_cgroup_largest_soft_limit_node(mctz); 3659 if (next_mz == mz) 3660 css_put(&next_mz->mem->css); 3661 else /* next_mz == NULL or other memcg */ 3662 break; 3663 } while (1); 3664 } 3665 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3666 excess = res_counter_soft_limit_excess(&mz->mem->res); 3667 /* 3668 * One school of thought says that we should not add 3669 * back the node to the tree if reclaim returns 0. 3670 * But our reclaim could return 0, simply because due 3671 * to priority we are exposing a smaller subset of 3672 * memory to reclaim from. Consider this as a longer 3673 * term TODO. 3674 */ 3675 /* If excess == 0, no tree ops */ 3676 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3677 spin_unlock(&mctz->lock); 3678 css_put(&mz->mem->css); 3679 loop++; 3680 /* 3681 * Could not reclaim anything and there are no more 3682 * mem cgroups to try or we seem to be looping without 3683 * reclaiming anything. 3684 */ 3685 if (!nr_reclaimed && 3686 (next_mz == NULL || 3687 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3688 break; 3689 } while (!nr_reclaimed); 3690 if (next_mz) 3691 css_put(&next_mz->mem->css); 3692 return nr_reclaimed; 3693} 3694 3695/* 3696 * This routine traverse page_cgroup in given list and drop them all. 3697 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3698 */ 3699static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg, 3700 int node, int zid, enum lru_list lru) 3701{ 3702 struct mem_cgroup_per_zone *mz; 3703 unsigned long flags, loop; 3704 struct list_head *list; 3705 struct page *busy; 3706 struct zone *zone; 3707 int ret = 0; 3708 3709 zone = &NODE_DATA(node)->node_zones[zid]; 3710 mz = mem_cgroup_zoneinfo(memcg, node, zid); 3711 list = &mz->lruvec.lists[lru]; 3712 3713 loop = MEM_CGROUP_ZSTAT(mz, lru); 3714 /* give some margin against EBUSY etc...*/ 3715 loop += 256; 3716 busy = NULL; 3717 while (loop--) { 3718 struct page_cgroup *pc; 3719 struct page *page; 3720 3721 ret = 0; 3722 spin_lock_irqsave(&zone->lru_lock, flags); 3723 if (list_empty(list)) { 3724 spin_unlock_irqrestore(&zone->lru_lock, flags); 3725 break; 3726 } 3727 page = list_entry(list->prev, struct page, lru); 3728 if (busy == page) { 3729 list_move(&page->lru, list); 3730 busy = NULL; 3731 spin_unlock_irqrestore(&zone->lru_lock, flags); 3732 continue; 3733 } 3734 spin_unlock_irqrestore(&zone->lru_lock, flags); 3735 3736 pc = lookup_page_cgroup(page); 3737 3738 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL); 3739 if (ret == -ENOMEM) 3740 break; 3741 3742 if (ret == -EBUSY || ret == -EINVAL) { 3743 /* found lock contention or "pc" is obsolete. */ 3744 busy = page; 3745 cond_resched(); 3746 } else 3747 busy = NULL; 3748 } 3749 3750 if (!ret && !list_empty(list)) 3751 return -EBUSY; 3752 return ret; 3753} 3754 3755/* 3756 * make mem_cgroup's charge to be 0 if there is no task. 3757 * This enables deleting this mem_cgroup. 3758 */ 3759static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all) 3760{ 3761 int ret; 3762 int node, zid, shrink; 3763 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3764 struct cgroup *cgrp = memcg->css.cgroup; 3765 3766 css_get(&memcg->css); 3767 3768 shrink = 0; 3769 /* should free all ? */ 3770 if (free_all) 3771 goto try_to_free; 3772move_account: 3773 do { 3774 ret = -EBUSY; 3775 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3776 goto out; 3777 ret = -EINTR; 3778 if (signal_pending(current)) 3779 goto out; 3780 /* This is for making all *used* pages to be on LRU. */ 3781 lru_add_drain_all(); 3782 drain_all_stock_sync(memcg); 3783 ret = 0; 3784 mem_cgroup_start_move(memcg); 3785 for_each_node_state(node, N_HIGH_MEMORY) { 3786 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3787 enum lru_list l; 3788 for_each_lru(l) { 3789 ret = mem_cgroup_force_empty_list(memcg, 3790 node, zid, l); 3791 if (ret) 3792 break; 3793 } 3794 } 3795 if (ret) 3796 break; 3797 } 3798 mem_cgroup_end_move(memcg); 3799 memcg_oom_recover(memcg); 3800 /* it seems parent cgroup doesn't have enough mem */ 3801 if (ret == -ENOMEM) 3802 goto try_to_free; 3803 cond_resched(); 3804 /* "ret" should also be checked to ensure all lists are empty. */ 3805 } while (memcg->res.usage > 0 || ret); 3806out: 3807 css_put(&memcg->css); 3808 return ret; 3809 3810try_to_free: 3811 /* returns EBUSY if there is a task or if we come here twice. */ 3812 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3813 ret = -EBUSY; 3814 goto out; 3815 } 3816 /* we call try-to-free pages for make this cgroup empty */ 3817 lru_add_drain_all(); 3818 /* try to free all pages in this cgroup */ 3819 shrink = 1; 3820 while (nr_retries && memcg->res.usage > 0) { 3821 int progress; 3822 3823 if (signal_pending(current)) { 3824 ret = -EINTR; 3825 goto out; 3826 } 3827 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL, 3828 false); 3829 if (!progress) { 3830 nr_retries--; 3831 /* maybe some writeback is necessary */ 3832 congestion_wait(BLK_RW_ASYNC, HZ/10); 3833 } 3834 3835 } 3836 lru_add_drain(); 3837 /* try move_account...there may be some *locked* pages. */ 3838 goto move_account; 3839} 3840 3841int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3842{ 3843 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3844} 3845 3846 3847static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3848{ 3849 return mem_cgroup_from_cont(cont)->use_hierarchy; 3850} 3851 3852static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3853 u64 val) 3854{ 3855 int retval = 0; 3856 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3857 struct cgroup *parent = cont->parent; 3858 struct mem_cgroup *parent_memcg = NULL; 3859 3860 if (parent) 3861 parent_memcg = mem_cgroup_from_cont(parent); 3862 3863 cgroup_lock(); 3864 /* 3865 * If parent's use_hierarchy is set, we can't make any modifications 3866 * in the child subtrees. If it is unset, then the change can 3867 * occur, provided the current cgroup has no children. 3868 * 3869 * For the root cgroup, parent_mem is NULL, we allow value to be 3870 * set if there are no children. 3871 */ 3872 if ((!parent_memcg || !parent_memcg->use_hierarchy) && 3873 (val == 1 || val == 0)) { 3874 if (list_empty(&cont->children)) 3875 memcg->use_hierarchy = val; 3876 else 3877 retval = -EBUSY; 3878 } else 3879 retval = -EINVAL; 3880 cgroup_unlock(); 3881 3882 return retval; 3883} 3884 3885 3886static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg, 3887 enum mem_cgroup_stat_index idx) 3888{ 3889 struct mem_cgroup *iter; 3890 long val = 0; 3891 3892 /* Per-cpu values can be negative, use a signed accumulator */ 3893 for_each_mem_cgroup_tree(iter, memcg) 3894 val += mem_cgroup_read_stat(iter, idx); 3895 3896 if (val < 0) /* race ? */ 3897 val = 0; 3898 return val; 3899} 3900 3901static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap) 3902{ 3903 u64 val; 3904 3905 if (!mem_cgroup_is_root(memcg)) { 3906 if (!swap) 3907 return res_counter_read_u64(&memcg->res, RES_USAGE); 3908 else 3909 return res_counter_read_u64(&memcg->memsw, RES_USAGE); 3910 } 3911 3912 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE); 3913 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS); 3914 3915 if (swap) 3916 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 3917 3918 return val << PAGE_SHIFT; 3919} 3920 3921static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3922{ 3923 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3924 u64 val; 3925 int type, name; 3926 3927 type = MEMFILE_TYPE(cft->private); 3928 name = MEMFILE_ATTR(cft->private); 3929 switch (type) { 3930 case _MEM: 3931 if (name == RES_USAGE) 3932 val = mem_cgroup_usage(memcg, false); 3933 else 3934 val = res_counter_read_u64(&memcg->res, name); 3935 break; 3936 case _MEMSWAP: 3937 if (name == RES_USAGE) 3938 val = mem_cgroup_usage(memcg, true); 3939 else 3940 val = res_counter_read_u64(&memcg->memsw, name); 3941 break; 3942 default: 3943 BUG(); 3944 break; 3945 } 3946 return val; 3947} 3948/* 3949 * The user of this function is... 3950 * RES_LIMIT. 3951 */ 3952static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3953 const char *buffer) 3954{ 3955 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3956 int type, name; 3957 unsigned long long val; 3958 int ret; 3959 3960 type = MEMFILE_TYPE(cft->private); 3961 name = MEMFILE_ATTR(cft->private); 3962 switch (name) { 3963 case RES_LIMIT: 3964 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3965 ret = -EINVAL; 3966 break; 3967 } 3968 /* This function does all necessary parse...reuse it */ 3969 ret = res_counter_memparse_write_strategy(buffer, &val); 3970 if (ret) 3971 break; 3972 if (type == _MEM) 3973 ret = mem_cgroup_resize_limit(memcg, val); 3974 else 3975 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3976 break; 3977 case RES_SOFT_LIMIT: 3978 ret = res_counter_memparse_write_strategy(buffer, &val); 3979 if (ret) 3980 break; 3981 /* 3982 * For memsw, soft limits are hard to implement in terms 3983 * of semantics, for now, we support soft limits for 3984 * control without swap 3985 */ 3986 if (type == _MEM) 3987 ret = res_counter_set_soft_limit(&memcg->res, val); 3988 else 3989 ret = -EINVAL; 3990 break; 3991 default: 3992 ret = -EINVAL; /* should be BUG() ? */ 3993 break; 3994 } 3995 return ret; 3996} 3997 3998static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 3999 unsigned long long *mem_limit, unsigned long long *memsw_limit) 4000{ 4001 struct cgroup *cgroup; 4002 unsigned long long min_limit, min_memsw_limit, tmp; 4003 4004 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 4005 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4006 cgroup = memcg->css.cgroup; 4007 if (!memcg->use_hierarchy) 4008 goto out; 4009 4010 while (cgroup->parent) { 4011 cgroup = cgroup->parent; 4012 memcg = mem_cgroup_from_cont(cgroup); 4013 if (!memcg->use_hierarchy) 4014 break; 4015 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 4016 min_limit = min(min_limit, tmp); 4017 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 4018 min_memsw_limit = min(min_memsw_limit, tmp); 4019 } 4020out: 4021 *mem_limit = min_limit; 4022 *memsw_limit = min_memsw_limit; 4023 return; 4024} 4025 4026static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 4027{ 4028 struct mem_cgroup *memcg; 4029 int type, name; 4030 4031 memcg = mem_cgroup_from_cont(cont); 4032 type = MEMFILE_TYPE(event); 4033 name = MEMFILE_ATTR(event); 4034 switch (name) { 4035 case RES_MAX_USAGE: 4036 if (type == _MEM) 4037 res_counter_reset_max(&memcg->res); 4038 else 4039 res_counter_reset_max(&memcg->memsw); 4040 break; 4041 case RES_FAILCNT: 4042 if (type == _MEM) 4043 res_counter_reset_failcnt(&memcg->res); 4044 else 4045 res_counter_reset_failcnt(&memcg->memsw); 4046 break; 4047 } 4048 4049 return 0; 4050} 4051 4052static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 4053 struct cftype *cft) 4054{ 4055 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 4056} 4057 4058#ifdef CONFIG_MMU 4059static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4060 struct cftype *cft, u64 val) 4061{ 4062 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4063 4064 if (val >= (1 << NR_MOVE_TYPE)) 4065 return -EINVAL; 4066 /* 4067 * We check this value several times in both in can_attach() and 4068 * attach(), so we need cgroup lock to prevent this value from being 4069 * inconsistent. 4070 */ 4071 cgroup_lock(); 4072 memcg->move_charge_at_immigrate = val; 4073 cgroup_unlock(); 4074 4075 return 0; 4076} 4077#else 4078static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 4079 struct cftype *cft, u64 val) 4080{ 4081 return -ENOSYS; 4082} 4083#endif 4084 4085 4086/* For read statistics */ 4087enum { 4088 MCS_CACHE, 4089 MCS_RSS, 4090 MCS_FILE_MAPPED, 4091 MCS_PGPGIN, 4092 MCS_PGPGOUT, 4093 MCS_SWAP, 4094 MCS_PGFAULT, 4095 MCS_PGMAJFAULT, 4096 MCS_INACTIVE_ANON, 4097 MCS_ACTIVE_ANON, 4098 MCS_INACTIVE_FILE, 4099 MCS_ACTIVE_FILE, 4100 MCS_UNEVICTABLE, 4101 NR_MCS_STAT, 4102}; 4103 4104struct mcs_total_stat { 4105 s64 stat[NR_MCS_STAT]; 4106}; 4107 4108struct { 4109 char *local_name; 4110 char *total_name; 4111} memcg_stat_strings[NR_MCS_STAT] = { 4112 {"cache", "total_cache"}, 4113 {"rss", "total_rss"}, 4114 {"mapped_file", "total_mapped_file"}, 4115 {"pgpgin", "total_pgpgin"}, 4116 {"pgpgout", "total_pgpgout"}, 4117 {"swap", "total_swap"}, 4118 {"pgfault", "total_pgfault"}, 4119 {"pgmajfault", "total_pgmajfault"}, 4120 {"inactive_anon", "total_inactive_anon"}, 4121 {"active_anon", "total_active_anon"}, 4122 {"inactive_file", "total_inactive_file"}, 4123 {"active_file", "total_active_file"}, 4124 {"unevictable", "total_unevictable"} 4125}; 4126 4127 4128static void 4129mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4130{ 4131 s64 val; 4132 4133 /* per cpu stat */ 4134 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE); 4135 s->stat[MCS_CACHE] += val * PAGE_SIZE; 4136 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS); 4137 s->stat[MCS_RSS] += val * PAGE_SIZE; 4138 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED); 4139 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 4140 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN); 4141 s->stat[MCS_PGPGIN] += val; 4142 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT); 4143 s->stat[MCS_PGPGOUT] += val; 4144 if (do_swap_account) { 4145 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT); 4146 s->stat[MCS_SWAP] += val * PAGE_SIZE; 4147 } 4148 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT); 4149 s->stat[MCS_PGFAULT] += val; 4150 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT); 4151 s->stat[MCS_PGMAJFAULT] += val; 4152 4153 /* per zone stat */ 4154 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON)); 4155 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 4156 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON)); 4157 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 4158 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE)); 4159 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 4160 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE)); 4161 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 4162 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE)); 4163 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 4164} 4165 4166static void 4167mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s) 4168{ 4169 struct mem_cgroup *iter; 4170 4171 for_each_mem_cgroup_tree(iter, memcg) 4172 mem_cgroup_get_local_stat(iter, s); 4173} 4174 4175#ifdef CONFIG_NUMA 4176static int mem_control_numa_stat_show(struct seq_file *m, void *arg) 4177{ 4178 int nid; 4179 unsigned long total_nr, file_nr, anon_nr, unevictable_nr; 4180 unsigned long node_nr; 4181 struct cgroup *cont = m->private; 4182 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4183 4184 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL); 4185 seq_printf(m, "total=%lu", total_nr); 4186 for_each_node_state(nid, N_HIGH_MEMORY) { 4187 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL); 4188 seq_printf(m, " N%d=%lu", nid, node_nr); 4189 } 4190 seq_putc(m, '\n'); 4191 4192 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE); 4193 seq_printf(m, "file=%lu", file_nr); 4194 for_each_node_state(nid, N_HIGH_MEMORY) { 4195 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4196 LRU_ALL_FILE); 4197 seq_printf(m, " N%d=%lu", nid, node_nr); 4198 } 4199 seq_putc(m, '\n'); 4200 4201 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON); 4202 seq_printf(m, "anon=%lu", anon_nr); 4203 for_each_node_state(nid, N_HIGH_MEMORY) { 4204 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4205 LRU_ALL_ANON); 4206 seq_printf(m, " N%d=%lu", nid, node_nr); 4207 } 4208 seq_putc(m, '\n'); 4209 4210 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE)); 4211 seq_printf(m, "unevictable=%lu", unevictable_nr); 4212 for_each_node_state(nid, N_HIGH_MEMORY) { 4213 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, 4214 BIT(LRU_UNEVICTABLE)); 4215 seq_printf(m, " N%d=%lu", nid, node_nr); 4216 } 4217 seq_putc(m, '\n'); 4218 return 0; 4219} 4220#endif /* CONFIG_NUMA */ 4221 4222static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 4223 struct cgroup_map_cb *cb) 4224{ 4225 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4226 struct mcs_total_stat mystat; 4227 int i; 4228 4229 memset(&mystat, 0, sizeof(mystat)); 4230 mem_cgroup_get_local_stat(mem_cont, &mystat); 4231 4232 4233 for (i = 0; i < NR_MCS_STAT; i++) { 4234 if (i == MCS_SWAP && !do_swap_account) 4235 continue; 4236 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 4237 } 4238 4239 /* Hierarchical information */ 4240 { 4241 unsigned long long limit, memsw_limit; 4242 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 4243 cb->fill(cb, "hierarchical_memory_limit", limit); 4244 if (do_swap_account) 4245 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 4246 } 4247 4248 memset(&mystat, 0, sizeof(mystat)); 4249 mem_cgroup_get_total_stat(mem_cont, &mystat); 4250 for (i = 0; i < NR_MCS_STAT; i++) { 4251 if (i == MCS_SWAP && !do_swap_account) 4252 continue; 4253 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 4254 } 4255 4256#ifdef CONFIG_DEBUG_VM 4257 { 4258 int nid, zid; 4259 struct mem_cgroup_per_zone *mz; 4260 unsigned long recent_rotated[2] = {0, 0}; 4261 unsigned long recent_scanned[2] = {0, 0}; 4262 4263 for_each_online_node(nid) 4264 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 4265 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 4266 4267 recent_rotated[0] += 4268 mz->reclaim_stat.recent_rotated[0]; 4269 recent_rotated[1] += 4270 mz->reclaim_stat.recent_rotated[1]; 4271 recent_scanned[0] += 4272 mz->reclaim_stat.recent_scanned[0]; 4273 recent_scanned[1] += 4274 mz->reclaim_stat.recent_scanned[1]; 4275 } 4276 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 4277 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 4278 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 4279 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 4280 } 4281#endif 4282 4283 return 0; 4284} 4285 4286static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 4287{ 4288 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4289 4290 return mem_cgroup_swappiness(memcg); 4291} 4292 4293static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 4294 u64 val) 4295{ 4296 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4297 struct mem_cgroup *parent; 4298 4299 if (val > 100) 4300 return -EINVAL; 4301 4302 if (cgrp->parent == NULL) 4303 return -EINVAL; 4304 4305 parent = mem_cgroup_from_cont(cgrp->parent); 4306 4307 cgroup_lock(); 4308 4309 /* If under hierarchy, only empty-root can set this value */ 4310 if ((parent->use_hierarchy) || 4311 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4312 cgroup_unlock(); 4313 return -EINVAL; 4314 } 4315 4316 memcg->swappiness = val; 4317 4318 cgroup_unlock(); 4319 4320 return 0; 4321} 4322 4323static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 4324{ 4325 struct mem_cgroup_threshold_ary *t; 4326 u64 usage; 4327 int i; 4328 4329 rcu_read_lock(); 4330 if (!swap) 4331 t = rcu_dereference(memcg->thresholds.primary); 4332 else 4333 t = rcu_dereference(memcg->memsw_thresholds.primary); 4334 4335 if (!t) 4336 goto unlock; 4337 4338 usage = mem_cgroup_usage(memcg, swap); 4339 4340 /* 4341 * current_threshold points to threshold just below usage. 4342 * If it's not true, a threshold was crossed after last 4343 * call of __mem_cgroup_threshold(). 4344 */ 4345 i = t->current_threshold; 4346 4347 /* 4348 * Iterate backward over array of thresholds starting from 4349 * current_threshold and check if a threshold is crossed. 4350 * If none of thresholds below usage is crossed, we read 4351 * only one element of the array here. 4352 */ 4353 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 4354 eventfd_signal(t->entries[i].eventfd, 1); 4355 4356 /* i = current_threshold + 1 */ 4357 i++; 4358 4359 /* 4360 * Iterate forward over array of thresholds starting from 4361 * current_threshold+1 and check if a threshold is crossed. 4362 * If none of thresholds above usage is crossed, we read 4363 * only one element of the array here. 4364 */ 4365 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 4366 eventfd_signal(t->entries[i].eventfd, 1); 4367 4368 /* Update current_threshold */ 4369 t->current_threshold = i - 1; 4370unlock: 4371 rcu_read_unlock(); 4372} 4373 4374static void mem_cgroup_threshold(struct mem_cgroup *memcg) 4375{ 4376 while (memcg) { 4377 __mem_cgroup_threshold(memcg, false); 4378 if (do_swap_account) 4379 __mem_cgroup_threshold(memcg, true); 4380 4381 memcg = parent_mem_cgroup(memcg); 4382 } 4383} 4384 4385static int compare_thresholds(const void *a, const void *b) 4386{ 4387 const struct mem_cgroup_threshold *_a = a; 4388 const struct mem_cgroup_threshold *_b = b; 4389 4390 return _a->threshold - _b->threshold; 4391} 4392 4393static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg) 4394{ 4395 struct mem_cgroup_eventfd_list *ev; 4396 4397 list_for_each_entry(ev, &memcg->oom_notify, list) 4398 eventfd_signal(ev->eventfd, 1); 4399 return 0; 4400} 4401 4402static void mem_cgroup_oom_notify(struct mem_cgroup *memcg) 4403{ 4404 struct mem_cgroup *iter; 4405 4406 for_each_mem_cgroup_tree(iter, memcg) 4407 mem_cgroup_oom_notify_cb(iter); 4408} 4409 4410static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 4411 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4412{ 4413 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4414 struct mem_cgroup_thresholds *thresholds; 4415 struct mem_cgroup_threshold_ary *new; 4416 int type = MEMFILE_TYPE(cft->private); 4417 u64 threshold, usage; 4418 int i, size, ret; 4419 4420 ret = res_counter_memparse_write_strategy(args, &threshold); 4421 if (ret) 4422 return ret; 4423 4424 mutex_lock(&memcg->thresholds_lock); 4425 4426 if (type == _MEM) 4427 thresholds = &memcg->thresholds; 4428 else if (type == _MEMSWAP) 4429 thresholds = &memcg->memsw_thresholds; 4430 else 4431 BUG(); 4432 4433 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4434 4435 /* Check if a threshold crossed before adding a new one */ 4436 if (thresholds->primary) 4437 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4438 4439 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 4440 4441 /* Allocate memory for new array of thresholds */ 4442 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 4443 GFP_KERNEL); 4444 if (!new) { 4445 ret = -ENOMEM; 4446 goto unlock; 4447 } 4448 new->size = size; 4449 4450 /* Copy thresholds (if any) to new array */ 4451 if (thresholds->primary) { 4452 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 4453 sizeof(struct mem_cgroup_threshold)); 4454 } 4455 4456 /* Add new threshold */ 4457 new->entries[size - 1].eventfd = eventfd; 4458 new->entries[size - 1].threshold = threshold; 4459 4460 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4461 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4462 compare_thresholds, NULL); 4463 4464 /* Find current threshold */ 4465 new->current_threshold = -1; 4466 for (i = 0; i < size; i++) { 4467 if (new->entries[i].threshold < usage) { 4468 /* 4469 * new->current_threshold will not be used until 4470 * rcu_assign_pointer(), so it's safe to increment 4471 * it here. 4472 */ 4473 ++new->current_threshold; 4474 } 4475 } 4476 4477 /* Free old spare buffer and save old primary buffer as spare */ 4478 kfree(thresholds->spare); 4479 thresholds->spare = thresholds->primary; 4480 4481 rcu_assign_pointer(thresholds->primary, new); 4482 4483 /* To be sure that nobody uses thresholds */ 4484 synchronize_rcu(); 4485 4486unlock: 4487 mutex_unlock(&memcg->thresholds_lock); 4488 4489 return ret; 4490} 4491 4492static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4493 struct cftype *cft, struct eventfd_ctx *eventfd) 4494{ 4495 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4496 struct mem_cgroup_thresholds *thresholds; 4497 struct mem_cgroup_threshold_ary *new; 4498 int type = MEMFILE_TYPE(cft->private); 4499 u64 usage; 4500 int i, j, size; 4501 4502 mutex_lock(&memcg->thresholds_lock); 4503 if (type == _MEM) 4504 thresholds = &memcg->thresholds; 4505 else if (type == _MEMSWAP) 4506 thresholds = &memcg->memsw_thresholds; 4507 else 4508 BUG(); 4509 4510 /* 4511 * Something went wrong if we trying to unregister a threshold 4512 * if we don't have thresholds 4513 */ 4514 BUG_ON(!thresholds); 4515 4516 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4517 4518 /* Check if a threshold crossed before removing */ 4519 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4520 4521 /* Calculate new number of threshold */ 4522 size = 0; 4523 for (i = 0; i < thresholds->primary->size; i++) { 4524 if (thresholds->primary->entries[i].eventfd != eventfd) 4525 size++; 4526 } 4527 4528 new = thresholds->spare; 4529 4530 /* Set thresholds array to NULL if we don't have thresholds */ 4531 if (!size) { 4532 kfree(new); 4533 new = NULL; 4534 goto swap_buffers; 4535 } 4536 4537 new->size = size; 4538 4539 /* Copy thresholds and find current threshold */ 4540 new->current_threshold = -1; 4541 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4542 if (thresholds->primary->entries[i].eventfd == eventfd) 4543 continue; 4544 4545 new->entries[j] = thresholds->primary->entries[i]; 4546 if (new->entries[j].threshold < usage) { 4547 /* 4548 * new->current_threshold will not be used 4549 * until rcu_assign_pointer(), so it's safe to increment 4550 * it here. 4551 */ 4552 ++new->current_threshold; 4553 } 4554 j++; 4555 } 4556 4557swap_buffers: 4558 /* Swap primary and spare array */ 4559 thresholds->spare = thresholds->primary; 4560 rcu_assign_pointer(thresholds->primary, new); 4561 4562 /* To be sure that nobody uses thresholds */ 4563 synchronize_rcu(); 4564 4565 mutex_unlock(&memcg->thresholds_lock); 4566} 4567 4568static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4569 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4570{ 4571 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4572 struct mem_cgroup_eventfd_list *event; 4573 int type = MEMFILE_TYPE(cft->private); 4574 4575 BUG_ON(type != _OOM_TYPE); 4576 event = kmalloc(sizeof(*event), GFP_KERNEL); 4577 if (!event) 4578 return -ENOMEM; 4579 4580 spin_lock(&memcg_oom_lock); 4581 4582 event->eventfd = eventfd; 4583 list_add(&event->list, &memcg->oom_notify); 4584 4585 /* already in OOM ? */ 4586 if (atomic_read(&memcg->under_oom)) 4587 eventfd_signal(eventfd, 1); 4588 spin_unlock(&memcg_oom_lock); 4589 4590 return 0; 4591} 4592 4593static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4594 struct cftype *cft, struct eventfd_ctx *eventfd) 4595{ 4596 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4597 struct mem_cgroup_eventfd_list *ev, *tmp; 4598 int type = MEMFILE_TYPE(cft->private); 4599 4600 BUG_ON(type != _OOM_TYPE); 4601 4602 spin_lock(&memcg_oom_lock); 4603 4604 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) { 4605 if (ev->eventfd == eventfd) { 4606 list_del(&ev->list); 4607 kfree(ev); 4608 } 4609 } 4610 4611 spin_unlock(&memcg_oom_lock); 4612} 4613 4614static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4615 struct cftype *cft, struct cgroup_map_cb *cb) 4616{ 4617 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4618 4619 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable); 4620 4621 if (atomic_read(&memcg->under_oom)) 4622 cb->fill(cb, "under_oom", 1); 4623 else 4624 cb->fill(cb, "under_oom", 0); 4625 return 0; 4626} 4627 4628static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4629 struct cftype *cft, u64 val) 4630{ 4631 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4632 struct mem_cgroup *parent; 4633 4634 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4635 if (!cgrp->parent || !((val == 0) || (val == 1))) 4636 return -EINVAL; 4637 4638 parent = mem_cgroup_from_cont(cgrp->parent); 4639 4640 cgroup_lock(); 4641 /* oom-kill-disable is a flag for subhierarchy. */ 4642 if ((parent->use_hierarchy) || 4643 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4644 cgroup_unlock(); 4645 return -EINVAL; 4646 } 4647 memcg->oom_kill_disable = val; 4648 if (!val) 4649 memcg_oom_recover(memcg); 4650 cgroup_unlock(); 4651 return 0; 4652} 4653 4654#ifdef CONFIG_NUMA 4655static const struct file_operations mem_control_numa_stat_file_operations = { 4656 .read = seq_read, 4657 .llseek = seq_lseek, 4658 .release = single_release, 4659}; 4660 4661static int mem_control_numa_stat_open(struct inode *unused, struct file *file) 4662{ 4663 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; 4664 4665 file->f_op = &mem_control_numa_stat_file_operations; 4666 return single_open(file, mem_control_numa_stat_show, cont); 4667} 4668#endif /* CONFIG_NUMA */ 4669 4670#ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM 4671static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4672{ 4673 /* 4674 * Part of this would be better living in a separate allocation 4675 * function, leaving us with just the cgroup tree population work. 4676 * We, however, depend on state such as network's proto_list that 4677 * is only initialized after cgroup creation. I found the less 4678 * cumbersome way to deal with it to defer it all to populate time 4679 */ 4680 return mem_cgroup_sockets_init(cont, ss); 4681}; 4682 4683static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4684 struct cgroup *cont) 4685{ 4686 mem_cgroup_sockets_destroy(cont, ss); 4687} 4688#else 4689static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss) 4690{ 4691 return 0; 4692} 4693 4694static void kmem_cgroup_destroy(struct cgroup_subsys *ss, 4695 struct cgroup *cont) 4696{ 4697} 4698#endif 4699 4700static struct cftype mem_cgroup_files[] = { 4701 { 4702 .name = "usage_in_bytes", 4703 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4704 .read_u64 = mem_cgroup_read, 4705 .register_event = mem_cgroup_usage_register_event, 4706 .unregister_event = mem_cgroup_usage_unregister_event, 4707 }, 4708 { 4709 .name = "max_usage_in_bytes", 4710 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4711 .trigger = mem_cgroup_reset, 4712 .read_u64 = mem_cgroup_read, 4713 }, 4714 { 4715 .name = "limit_in_bytes", 4716 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4717 .write_string = mem_cgroup_write, 4718 .read_u64 = mem_cgroup_read, 4719 }, 4720 { 4721 .name = "soft_limit_in_bytes", 4722 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4723 .write_string = mem_cgroup_write, 4724 .read_u64 = mem_cgroup_read, 4725 }, 4726 { 4727 .name = "failcnt", 4728 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4729 .trigger = mem_cgroup_reset, 4730 .read_u64 = mem_cgroup_read, 4731 }, 4732 { 4733 .name = "stat", 4734 .read_map = mem_control_stat_show, 4735 }, 4736 { 4737 .name = "force_empty", 4738 .trigger = mem_cgroup_force_empty_write, 4739 }, 4740 { 4741 .name = "use_hierarchy", 4742 .write_u64 = mem_cgroup_hierarchy_write, 4743 .read_u64 = mem_cgroup_hierarchy_read, 4744 }, 4745 { 4746 .name = "swappiness", 4747 .read_u64 = mem_cgroup_swappiness_read, 4748 .write_u64 = mem_cgroup_swappiness_write, 4749 }, 4750 { 4751 .name = "move_charge_at_immigrate", 4752 .read_u64 = mem_cgroup_move_charge_read, 4753 .write_u64 = mem_cgroup_move_charge_write, 4754 }, 4755 { 4756 .name = "oom_control", 4757 .read_map = mem_cgroup_oom_control_read, 4758 .write_u64 = mem_cgroup_oom_control_write, 4759 .register_event = mem_cgroup_oom_register_event, 4760 .unregister_event = mem_cgroup_oom_unregister_event, 4761 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4762 }, 4763#ifdef CONFIG_NUMA 4764 { 4765 .name = "numa_stat", 4766 .open = mem_control_numa_stat_open, 4767 .mode = S_IRUGO, 4768 }, 4769#endif 4770}; 4771 4772#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4773static struct cftype memsw_cgroup_files[] = { 4774 { 4775 .name = "memsw.usage_in_bytes", 4776 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4777 .read_u64 = mem_cgroup_read, 4778 .register_event = mem_cgroup_usage_register_event, 4779 .unregister_event = mem_cgroup_usage_unregister_event, 4780 }, 4781 { 4782 .name = "memsw.max_usage_in_bytes", 4783 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4784 .trigger = mem_cgroup_reset, 4785 .read_u64 = mem_cgroup_read, 4786 }, 4787 { 4788 .name = "memsw.limit_in_bytes", 4789 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4790 .write_string = mem_cgroup_write, 4791 .read_u64 = mem_cgroup_read, 4792 }, 4793 { 4794 .name = "memsw.failcnt", 4795 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4796 .trigger = mem_cgroup_reset, 4797 .read_u64 = mem_cgroup_read, 4798 }, 4799}; 4800 4801static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4802{ 4803 if (!do_swap_account) 4804 return 0; 4805 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4806 ARRAY_SIZE(memsw_cgroup_files)); 4807}; 4808#else 4809static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4810{ 4811 return 0; 4812} 4813#endif 4814 4815static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4816{ 4817 struct mem_cgroup_per_node *pn; 4818 struct mem_cgroup_per_zone *mz; 4819 enum lru_list l; 4820 int zone, tmp = node; 4821 /* 4822 * This routine is called against possible nodes. 4823 * But it's BUG to call kmalloc() against offline node. 4824 * 4825 * TODO: this routine can waste much memory for nodes which will 4826 * never be onlined. It's better to use memory hotplug callback 4827 * function. 4828 */ 4829 if (!node_state(node, N_NORMAL_MEMORY)) 4830 tmp = -1; 4831 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4832 if (!pn) 4833 return 1; 4834 4835 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4836 mz = &pn->zoneinfo[zone]; 4837 for_each_lru(l) 4838 INIT_LIST_HEAD(&mz->lruvec.lists[l]); 4839 mz->usage_in_excess = 0; 4840 mz->on_tree = false; 4841 mz->mem = memcg; 4842 } 4843 memcg->info.nodeinfo[node] = pn; 4844 return 0; 4845} 4846 4847static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node) 4848{ 4849 kfree(memcg->info.nodeinfo[node]); 4850} 4851 4852static struct mem_cgroup *mem_cgroup_alloc(void) 4853{ 4854 struct mem_cgroup *mem; 4855 int size = sizeof(struct mem_cgroup); 4856 4857 /* Can be very big if MAX_NUMNODES is very big */ 4858 if (size < PAGE_SIZE) 4859 mem = kzalloc(size, GFP_KERNEL); 4860 else 4861 mem = vzalloc(size); 4862 4863 if (!mem) 4864 return NULL; 4865 4866 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4867 if (!mem->stat) 4868 goto out_free; 4869 spin_lock_init(&mem->pcp_counter_lock); 4870 return mem; 4871 4872out_free: 4873 if (size < PAGE_SIZE) 4874 kfree(mem); 4875 else 4876 vfree(mem); 4877 return NULL; 4878} 4879 4880/* 4881 * At destroying mem_cgroup, references from swap_cgroup can remain. 4882 * (scanning all at force_empty is too costly...) 4883 * 4884 * Instead of clearing all references at force_empty, we remember 4885 * the number of reference from swap_cgroup and free mem_cgroup when 4886 * it goes down to 0. 4887 * 4888 * Removal of cgroup itself succeeds regardless of refs from swap. 4889 */ 4890 4891static void __mem_cgroup_free(struct mem_cgroup *memcg) 4892{ 4893 int node; 4894 4895 mem_cgroup_remove_from_trees(memcg); 4896 free_css_id(&mem_cgroup_subsys, &memcg->css); 4897 4898 for_each_node_state(node, N_POSSIBLE) 4899 free_mem_cgroup_per_zone_info(memcg, node); 4900 4901 free_percpu(memcg->stat); 4902 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4903 kfree(memcg); 4904 else 4905 vfree(memcg); 4906} 4907 4908static void mem_cgroup_get(struct mem_cgroup *memcg) 4909{ 4910 atomic_inc(&memcg->refcnt); 4911} 4912 4913static void __mem_cgroup_put(struct mem_cgroup *memcg, int count) 4914{ 4915 if (atomic_sub_and_test(count, &memcg->refcnt)) { 4916 struct mem_cgroup *parent = parent_mem_cgroup(memcg); 4917 __mem_cgroup_free(memcg); 4918 if (parent) 4919 mem_cgroup_put(parent); 4920 } 4921} 4922 4923static void mem_cgroup_put(struct mem_cgroup *memcg) 4924{ 4925 __mem_cgroup_put(memcg, 1); 4926} 4927 4928/* 4929 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4930 */ 4931struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg) 4932{ 4933 if (!memcg->res.parent) 4934 return NULL; 4935 return mem_cgroup_from_res_counter(memcg->res.parent, res); 4936} 4937EXPORT_SYMBOL(parent_mem_cgroup); 4938 4939#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4940static void __init enable_swap_cgroup(void) 4941{ 4942 if (!mem_cgroup_disabled() && really_do_swap_account) 4943 do_swap_account = 1; 4944} 4945#else 4946static void __init enable_swap_cgroup(void) 4947{ 4948} 4949#endif 4950 4951static int mem_cgroup_soft_limit_tree_init(void) 4952{ 4953 struct mem_cgroup_tree_per_node *rtpn; 4954 struct mem_cgroup_tree_per_zone *rtpz; 4955 int tmp, node, zone; 4956 4957 for_each_node_state(node, N_POSSIBLE) { 4958 tmp = node; 4959 if (!node_state(node, N_NORMAL_MEMORY)) 4960 tmp = -1; 4961 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4962 if (!rtpn) 4963 return 1; 4964 4965 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4966 4967 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4968 rtpz = &rtpn->rb_tree_per_zone[zone]; 4969 rtpz->rb_root = RB_ROOT; 4970 spin_lock_init(&rtpz->lock); 4971 } 4972 } 4973 return 0; 4974} 4975 4976static struct cgroup_subsys_state * __ref 4977mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4978{ 4979 struct mem_cgroup *memcg, *parent; 4980 long error = -ENOMEM; 4981 int node; 4982 4983 memcg = mem_cgroup_alloc(); 4984 if (!memcg) 4985 return ERR_PTR(error); 4986 4987 for_each_node_state(node, N_POSSIBLE) 4988 if (alloc_mem_cgroup_per_zone_info(memcg, node)) 4989 goto free_out; 4990 4991 /* root ? */ 4992 if (cont->parent == NULL) { 4993 int cpu; 4994 enable_swap_cgroup(); 4995 parent = NULL; 4996 if (mem_cgroup_soft_limit_tree_init()) 4997 goto free_out; 4998 root_mem_cgroup = memcg; 4999 for_each_possible_cpu(cpu) { 5000 struct memcg_stock_pcp *stock = 5001 &per_cpu(memcg_stock, cpu); 5002 INIT_WORK(&stock->work, drain_local_stock); 5003 } 5004 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 5005 } else { 5006 parent = mem_cgroup_from_cont(cont->parent); 5007 memcg->use_hierarchy = parent->use_hierarchy; 5008 memcg->oom_kill_disable = parent->oom_kill_disable; 5009 } 5010 5011 if (parent && parent->use_hierarchy) { 5012 res_counter_init(&memcg->res, &parent->res); 5013 res_counter_init(&memcg->memsw, &parent->memsw); 5014 /* 5015 * We increment refcnt of the parent to ensure that we can 5016 * safely access it on res_counter_charge/uncharge. 5017 * This refcnt will be decremented when freeing this 5018 * mem_cgroup(see mem_cgroup_put). 5019 */ 5020 mem_cgroup_get(parent); 5021 } else { 5022 res_counter_init(&memcg->res, NULL); 5023 res_counter_init(&memcg->memsw, NULL); 5024 } 5025 memcg->last_scanned_node = MAX_NUMNODES; 5026 INIT_LIST_HEAD(&memcg->oom_notify); 5027 5028 if (parent) 5029 memcg->swappiness = mem_cgroup_swappiness(parent); 5030 atomic_set(&memcg->refcnt, 1); 5031 memcg->move_charge_at_immigrate = 0; 5032 mutex_init(&memcg->thresholds_lock); 5033 return &memcg->css; 5034free_out: 5035 __mem_cgroup_free(memcg); 5036 return ERR_PTR(error); 5037} 5038 5039static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 5040 struct cgroup *cont) 5041{ 5042 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5043 5044 return mem_cgroup_force_empty(memcg, false); 5045} 5046 5047static void mem_cgroup_destroy(struct cgroup_subsys *ss, 5048 struct cgroup *cont) 5049{ 5050 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 5051 5052 kmem_cgroup_destroy(ss, cont); 5053 5054 mem_cgroup_put(memcg); 5055} 5056 5057static int mem_cgroup_populate(struct cgroup_subsys *ss, 5058 struct cgroup *cont) 5059{ 5060 int ret; 5061 5062 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 5063 ARRAY_SIZE(mem_cgroup_files)); 5064 5065 if (!ret) 5066 ret = register_memsw_files(cont, ss); 5067 5068 if (!ret) 5069 ret = register_kmem_files(cont, ss); 5070 5071 return ret; 5072} 5073 5074#ifdef CONFIG_MMU 5075/* Handlers for move charge at task migration. */ 5076#define PRECHARGE_COUNT_AT_ONCE 256 5077static int mem_cgroup_do_precharge(unsigned long count) 5078{ 5079 int ret = 0; 5080 int batch_count = PRECHARGE_COUNT_AT_ONCE; 5081 struct mem_cgroup *memcg = mc.to; 5082 5083 if (mem_cgroup_is_root(memcg)) { 5084 mc.precharge += count; 5085 /* we don't need css_get for root */ 5086 return ret; 5087 } 5088 /* try to charge at once */ 5089 if (count > 1) { 5090 struct res_counter *dummy; 5091 /* 5092 * "memcg" cannot be under rmdir() because we've already checked 5093 * by cgroup_lock_live_cgroup() that it is not removed and we 5094 * are still under the same cgroup_mutex. So we can postpone 5095 * css_get(). 5096 */ 5097 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy)) 5098 goto one_by_one; 5099 if (do_swap_account && res_counter_charge(&memcg->memsw, 5100 PAGE_SIZE * count, &dummy)) { 5101 res_counter_uncharge(&memcg->res, PAGE_SIZE * count); 5102 goto one_by_one; 5103 } 5104 mc.precharge += count; 5105 return ret; 5106 } 5107one_by_one: 5108 /* fall back to one by one charge */ 5109 while (count--) { 5110 if (signal_pending(current)) { 5111 ret = -EINTR; 5112 break; 5113 } 5114 if (!batch_count--) { 5115 batch_count = PRECHARGE_COUNT_AT_ONCE; 5116 cond_resched(); 5117 } 5118 ret = __mem_cgroup_try_charge(NULL, 5119 GFP_KERNEL, 1, &memcg, false); 5120 if (ret || !memcg) 5121 /* mem_cgroup_clear_mc() will do uncharge later */ 5122 return -ENOMEM; 5123 mc.precharge++; 5124 } 5125 return ret; 5126} 5127 5128/** 5129 * is_target_pte_for_mc - check a pte whether it is valid for move charge 5130 * @vma: the vma the pte to be checked belongs 5131 * @addr: the address corresponding to the pte to be checked 5132 * @ptent: the pte to be checked 5133 * @target: the pointer the target page or swap ent will be stored(can be NULL) 5134 * 5135 * Returns 5136 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 5137 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 5138 * move charge. if @target is not NULL, the page is stored in target->page 5139 * with extra refcnt got(Callers should handle it). 5140 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 5141 * target for charge migration. if @target is not NULL, the entry is stored 5142 * in target->ent. 5143 * 5144 * Called with pte lock held. 5145 */ 5146union mc_target { 5147 struct page *page; 5148 swp_entry_t ent; 5149}; 5150 5151enum mc_target_type { 5152 MC_TARGET_NONE, /* not used */ 5153 MC_TARGET_PAGE, 5154 MC_TARGET_SWAP, 5155}; 5156 5157static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 5158 unsigned long addr, pte_t ptent) 5159{ 5160 struct page *page = vm_normal_page(vma, addr, ptent); 5161 5162 if (!page || !page_mapped(page)) 5163 return NULL; 5164 if (PageAnon(page)) { 5165 /* we don't move shared anon */ 5166 if (!move_anon() || page_mapcount(page) > 2) 5167 return NULL; 5168 } else if (!move_file()) 5169 /* we ignore mapcount for file pages */ 5170 return NULL; 5171 if (!get_page_unless_zero(page)) 5172 return NULL; 5173 5174 return page; 5175} 5176 5177static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 5178 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5179{ 5180 int usage_count; 5181 struct page *page = NULL; 5182 swp_entry_t ent = pte_to_swp_entry(ptent); 5183 5184 if (!move_anon() || non_swap_entry(ent)) 5185 return NULL; 5186 usage_count = mem_cgroup_count_swap_user(ent, &page); 5187 if (usage_count > 1) { /* we don't move shared anon */ 5188 if (page) 5189 put_page(page); 5190 return NULL; 5191 } 5192 if (do_swap_account) 5193 entry->val = ent.val; 5194 5195 return page; 5196} 5197 5198static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 5199 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5200{ 5201 struct page *page = NULL; 5202 struct inode *inode; 5203 struct address_space *mapping; 5204 pgoff_t pgoff; 5205 5206 if (!vma->vm_file) /* anonymous vma */ 5207 return NULL; 5208 if (!move_file()) 5209 return NULL; 5210 5211 inode = vma->vm_file->f_path.dentry->d_inode; 5212 mapping = vma->vm_file->f_mapping; 5213 if (pte_none(ptent)) 5214 pgoff = linear_page_index(vma, addr); 5215 else /* pte_file(ptent) is true */ 5216 pgoff = pte_to_pgoff(ptent); 5217 5218 /* page is moved even if it's not RSS of this task(page-faulted). */ 5219 page = find_get_page(mapping, pgoff); 5220 5221#ifdef CONFIG_SWAP 5222 /* shmem/tmpfs may report page out on swap: account for that too. */ 5223 if (radix_tree_exceptional_entry(page)) { 5224 swp_entry_t swap = radix_to_swp_entry(page); 5225 if (do_swap_account) 5226 *entry = swap; 5227 page = find_get_page(&swapper_space, swap.val); 5228 } 5229#endif 5230 return page; 5231} 5232 5233static int is_target_pte_for_mc(struct vm_area_struct *vma, 5234 unsigned long addr, pte_t ptent, union mc_target *target) 5235{ 5236 struct page *page = NULL; 5237 struct page_cgroup *pc; 5238 int ret = 0; 5239 swp_entry_t ent = { .val = 0 }; 5240 5241 if (pte_present(ptent)) 5242 page = mc_handle_present_pte(vma, addr, ptent); 5243 else if (is_swap_pte(ptent)) 5244 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 5245 else if (pte_none(ptent) || pte_file(ptent)) 5246 page = mc_handle_file_pte(vma, addr, ptent, &ent); 5247 5248 if (!page && !ent.val) 5249 return 0; 5250 if (page) { 5251 pc = lookup_page_cgroup(page); 5252 /* 5253 * Do only loose check w/o page_cgroup lock. 5254 * mem_cgroup_move_account() checks the pc is valid or not under 5255 * the lock. 5256 */ 5257 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 5258 ret = MC_TARGET_PAGE; 5259 if (target) 5260 target->page = page; 5261 } 5262 if (!ret || !target) 5263 put_page(page); 5264 } 5265 /* There is a swap entry and a page doesn't exist or isn't charged */ 5266 if (ent.val && !ret && 5267 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) { 5268 ret = MC_TARGET_SWAP; 5269 if (target) 5270 target->ent = ent; 5271 } 5272 return ret; 5273} 5274 5275static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 5276 unsigned long addr, unsigned long end, 5277 struct mm_walk *walk) 5278{ 5279 struct vm_area_struct *vma = walk->private; 5280 pte_t *pte; 5281 spinlock_t *ptl; 5282 5283 split_huge_page_pmd(walk->mm, pmd); 5284 5285 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5286 for (; addr != end; pte++, addr += PAGE_SIZE) 5287 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 5288 mc.precharge++; /* increment precharge temporarily */ 5289 pte_unmap_unlock(pte - 1, ptl); 5290 cond_resched(); 5291 5292 return 0; 5293} 5294 5295static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 5296{ 5297 unsigned long precharge; 5298 struct vm_area_struct *vma; 5299 5300 down_read(&mm->mmap_sem); 5301 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5302 struct mm_walk mem_cgroup_count_precharge_walk = { 5303 .pmd_entry = mem_cgroup_count_precharge_pte_range, 5304 .mm = mm, 5305 .private = vma, 5306 }; 5307 if (is_vm_hugetlb_page(vma)) 5308 continue; 5309 walk_page_range(vma->vm_start, vma->vm_end, 5310 &mem_cgroup_count_precharge_walk); 5311 } 5312 up_read(&mm->mmap_sem); 5313 5314 precharge = mc.precharge; 5315 mc.precharge = 0; 5316 5317 return precharge; 5318} 5319 5320static int mem_cgroup_precharge_mc(struct mm_struct *mm) 5321{ 5322 unsigned long precharge = mem_cgroup_count_precharge(mm); 5323 5324 VM_BUG_ON(mc.moving_task); 5325 mc.moving_task = current; 5326 return mem_cgroup_do_precharge(precharge); 5327} 5328 5329/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 5330static void __mem_cgroup_clear_mc(void) 5331{ 5332 struct mem_cgroup *from = mc.from; 5333 struct mem_cgroup *to = mc.to; 5334 5335 /* we must uncharge all the leftover precharges from mc.to */ 5336 if (mc.precharge) { 5337 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 5338 mc.precharge = 0; 5339 } 5340 /* 5341 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 5342 * we must uncharge here. 5343 */ 5344 if (mc.moved_charge) { 5345 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 5346 mc.moved_charge = 0; 5347 } 5348 /* we must fixup refcnts and charges */ 5349 if (mc.moved_swap) { 5350 /* uncharge swap account from the old cgroup */ 5351 if (!mem_cgroup_is_root(mc.from)) 5352 res_counter_uncharge(&mc.from->memsw, 5353 PAGE_SIZE * mc.moved_swap); 5354 __mem_cgroup_put(mc.from, mc.moved_swap); 5355 5356 if (!mem_cgroup_is_root(mc.to)) { 5357 /* 5358 * we charged both to->res and to->memsw, so we should 5359 * uncharge to->res. 5360 */ 5361 res_counter_uncharge(&mc.to->res, 5362 PAGE_SIZE * mc.moved_swap); 5363 } 5364 /* we've already done mem_cgroup_get(mc.to) */ 5365 mc.moved_swap = 0; 5366 } 5367 memcg_oom_recover(from); 5368 memcg_oom_recover(to); 5369 wake_up_all(&mc.waitq); 5370} 5371 5372static void mem_cgroup_clear_mc(void) 5373{ 5374 struct mem_cgroup *from = mc.from; 5375 5376 /* 5377 * we must clear moving_task before waking up waiters at the end of 5378 * task migration. 5379 */ 5380 mc.moving_task = NULL; 5381 __mem_cgroup_clear_mc(); 5382 spin_lock(&mc.lock); 5383 mc.from = NULL; 5384 mc.to = NULL; 5385 spin_unlock(&mc.lock); 5386 mem_cgroup_end_move(from); 5387} 5388 5389static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5390 struct cgroup *cgroup, 5391 struct cgroup_taskset *tset) 5392{ 5393 struct task_struct *p = cgroup_taskset_first(tset); 5394 int ret = 0; 5395 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup); 5396 5397 if (memcg->move_charge_at_immigrate) { 5398 struct mm_struct *mm; 5399 struct mem_cgroup *from = mem_cgroup_from_task(p); 5400 5401 VM_BUG_ON(from == memcg); 5402 5403 mm = get_task_mm(p); 5404 if (!mm) 5405 return 0; 5406 /* We move charges only when we move a owner of the mm */ 5407 if (mm->owner == p) { 5408 VM_BUG_ON(mc.from); 5409 VM_BUG_ON(mc.to); 5410 VM_BUG_ON(mc.precharge); 5411 VM_BUG_ON(mc.moved_charge); 5412 VM_BUG_ON(mc.moved_swap); 5413 mem_cgroup_start_move(from); 5414 spin_lock(&mc.lock); 5415 mc.from = from; 5416 mc.to = memcg; 5417 spin_unlock(&mc.lock); 5418 /* We set mc.moving_task later */ 5419 5420 ret = mem_cgroup_precharge_mc(mm); 5421 if (ret) 5422 mem_cgroup_clear_mc(); 5423 } 5424 mmput(mm); 5425 } 5426 return ret; 5427} 5428 5429static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5430 struct cgroup *cgroup, 5431 struct cgroup_taskset *tset) 5432{ 5433 mem_cgroup_clear_mc(); 5434} 5435 5436static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 5437 unsigned long addr, unsigned long end, 5438 struct mm_walk *walk) 5439{ 5440 int ret = 0; 5441 struct vm_area_struct *vma = walk->private; 5442 pte_t *pte; 5443 spinlock_t *ptl; 5444 5445 split_huge_page_pmd(walk->mm, pmd); 5446retry: 5447 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5448 for (; addr != end; addr += PAGE_SIZE) { 5449 pte_t ptent = *(pte++); 5450 union mc_target target; 5451 int type; 5452 struct page *page; 5453 struct page_cgroup *pc; 5454 swp_entry_t ent; 5455 5456 if (!mc.precharge) 5457 break; 5458 5459 type = is_target_pte_for_mc(vma, addr, ptent, &target); 5460 switch (type) { 5461 case MC_TARGET_PAGE: 5462 page = target.page; 5463 if (isolate_lru_page(page)) 5464 goto put; 5465 pc = lookup_page_cgroup(page); 5466 if (!mem_cgroup_move_account(page, 1, pc, 5467 mc.from, mc.to, false)) { 5468 mc.precharge--; 5469 /* we uncharge from mc.from later. */ 5470 mc.moved_charge++; 5471 } 5472 putback_lru_page(page); 5473put: /* is_target_pte_for_mc() gets the page */ 5474 put_page(page); 5475 break; 5476 case MC_TARGET_SWAP: 5477 ent = target.ent; 5478 if (!mem_cgroup_move_swap_account(ent, 5479 mc.from, mc.to, false)) { 5480 mc.precharge--; 5481 /* we fixup refcnts and charges later. */ 5482 mc.moved_swap++; 5483 } 5484 break; 5485 default: 5486 break; 5487 } 5488 } 5489 pte_unmap_unlock(pte - 1, ptl); 5490 cond_resched(); 5491 5492 if (addr != end) { 5493 /* 5494 * We have consumed all precharges we got in can_attach(). 5495 * We try charge one by one, but don't do any additional 5496 * charges to mc.to if we have failed in charge once in attach() 5497 * phase. 5498 */ 5499 ret = mem_cgroup_do_precharge(1); 5500 if (!ret) 5501 goto retry; 5502 } 5503 5504 return ret; 5505} 5506 5507static void mem_cgroup_move_charge(struct mm_struct *mm) 5508{ 5509 struct vm_area_struct *vma; 5510 5511 lru_add_drain_all(); 5512retry: 5513 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 5514 /* 5515 * Someone who are holding the mmap_sem might be waiting in 5516 * waitq. So we cancel all extra charges, wake up all waiters, 5517 * and retry. Because we cancel precharges, we might not be able 5518 * to move enough charges, but moving charge is a best-effort 5519 * feature anyway, so it wouldn't be a big problem. 5520 */ 5521 __mem_cgroup_clear_mc(); 5522 cond_resched(); 5523 goto retry; 5524 } 5525 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5526 int ret; 5527 struct mm_walk mem_cgroup_move_charge_walk = { 5528 .pmd_entry = mem_cgroup_move_charge_pte_range, 5529 .mm = mm, 5530 .private = vma, 5531 }; 5532 if (is_vm_hugetlb_page(vma)) 5533 continue; 5534 ret = walk_page_range(vma->vm_start, vma->vm_end, 5535 &mem_cgroup_move_charge_walk); 5536 if (ret) 5537 /* 5538 * means we have consumed all precharges and failed in 5539 * doing additional charge. Just abandon here. 5540 */ 5541 break; 5542 } 5543 up_read(&mm->mmap_sem); 5544} 5545 5546static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5547 struct cgroup *cont, 5548 struct cgroup_taskset *tset) 5549{ 5550 struct task_struct *p = cgroup_taskset_first(tset); 5551 struct mm_struct *mm = get_task_mm(p); 5552 5553 if (mm) { 5554 if (mc.to) 5555 mem_cgroup_move_charge(mm); 5556 put_swap_token(mm); 5557 mmput(mm); 5558 } 5559 if (mc.to) 5560 mem_cgroup_clear_mc(); 5561} 5562#else /* !CONFIG_MMU */ 5563static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5564 struct cgroup *cgroup, 5565 struct cgroup_taskset *tset) 5566{ 5567 return 0; 5568} 5569static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5570 struct cgroup *cgroup, 5571 struct cgroup_taskset *tset) 5572{ 5573} 5574static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5575 struct cgroup *cont, 5576 struct cgroup_taskset *tset) 5577{ 5578} 5579#endif 5580 5581struct cgroup_subsys mem_cgroup_subsys = { 5582 .name = "memory", 5583 .subsys_id = mem_cgroup_subsys_id, 5584 .create = mem_cgroup_create, 5585 .pre_destroy = mem_cgroup_pre_destroy, 5586 .destroy = mem_cgroup_destroy, 5587 .populate = mem_cgroup_populate, 5588 .can_attach = mem_cgroup_can_attach, 5589 .cancel_attach = mem_cgroup_cancel_attach, 5590 .attach = mem_cgroup_move_task, 5591 .early_init = 0, 5592 .use_id = 1, 5593}; 5594 5595#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5596static int __init enable_swap_account(char *s) 5597{ 5598 /* consider enabled if no parameter or 1 is given */ 5599 if (!strcmp(s, "1")) 5600 really_do_swap_account = 1; 5601 else if (!strcmp(s, "0")) 5602 really_do_swap_account = 0; 5603 return 1; 5604} 5605__setup("swapaccount=", enable_swap_account); 5606 5607#endif 5608