memcontrol.c revision 7ffd4ca7a2cdd7a18f0b499a4e9e0e7cf36ba018
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/mutex.h> 37#include <linux/rbtree.h> 38#include <linux/slab.h> 39#include <linux/swap.h> 40#include <linux/swapops.h> 41#include <linux/spinlock.h> 42#include <linux/eventfd.h> 43#include <linux/sort.h> 44#include <linux/fs.h> 45#include <linux/seq_file.h> 46#include <linux/vmalloc.h> 47#include <linux/mm_inline.h> 48#include <linux/page_cgroup.h> 49#include <linux/cpu.h> 50#include <linux/oom.h> 51#include "internal.h" 52 53#include <asm/uaccess.h> 54 55#include <trace/events/vmscan.h> 56 57struct cgroup_subsys mem_cgroup_subsys __read_mostly; 58#define MEM_CGROUP_RECLAIM_RETRIES 5 59struct mem_cgroup *root_mem_cgroup __read_mostly; 60 61#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 62/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ 63int do_swap_account __read_mostly; 64 65/* for remember boot option*/ 66#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED 67static int really_do_swap_account __initdata = 1; 68#else 69static int really_do_swap_account __initdata = 0; 70#endif 71 72#else 73#define do_swap_account (0) 74#endif 75 76/* 77 * Per memcg event counter is incremented at every pagein/pageout. This counter 78 * is used for trigger some periodic events. This is straightforward and better 79 * than using jiffies etc. to handle periodic memcg event. 80 * 81 * These values will be used as !((event) & ((1 <<(thresh)) - 1)) 82 */ 83#define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */ 84#define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */ 85 86/* 87 * Statistics for memory cgroup. 88 */ 89enum mem_cgroup_stat_index { 90 /* 91 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 92 */ 93 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 94 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ 95 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ 96 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ 97 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ 98 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ 99 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ 100 /* incremented at every pagein/pageout */ 101 MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA, 102 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ 103 104 MEM_CGROUP_STAT_NSTATS, 105}; 106 107struct mem_cgroup_stat_cpu { 108 s64 count[MEM_CGROUP_STAT_NSTATS]; 109}; 110 111/* 112 * per-zone information in memory controller. 113 */ 114struct mem_cgroup_per_zone { 115 /* 116 * spin_lock to protect the per cgroup LRU 117 */ 118 struct list_head lists[NR_LRU_LISTS]; 119 unsigned long count[NR_LRU_LISTS]; 120 121 struct zone_reclaim_stat reclaim_stat; 122 struct rb_node tree_node; /* RB tree node */ 123 unsigned long long usage_in_excess;/* Set to the value by which */ 124 /* the soft limit is exceeded*/ 125 bool on_tree; 126 struct mem_cgroup *mem; /* Back pointer, we cannot */ 127 /* use container_of */ 128}; 129/* Macro for accessing counter */ 130#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 131 132struct mem_cgroup_per_node { 133 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 134}; 135 136struct mem_cgroup_lru_info { 137 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 138}; 139 140/* 141 * Cgroups above their limits are maintained in a RB-Tree, independent of 142 * their hierarchy representation 143 */ 144 145struct mem_cgroup_tree_per_zone { 146 struct rb_root rb_root; 147 spinlock_t lock; 148}; 149 150struct mem_cgroup_tree_per_node { 151 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; 152}; 153 154struct mem_cgroup_tree { 155 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; 156}; 157 158static struct mem_cgroup_tree soft_limit_tree __read_mostly; 159 160struct mem_cgroup_threshold { 161 struct eventfd_ctx *eventfd; 162 u64 threshold; 163}; 164 165/* For threshold */ 166struct mem_cgroup_threshold_ary { 167 /* An array index points to threshold just below usage. */ 168 int current_threshold; 169 /* Size of entries[] */ 170 unsigned int size; 171 /* Array of thresholds */ 172 struct mem_cgroup_threshold entries[0]; 173}; 174 175struct mem_cgroup_thresholds { 176 /* Primary thresholds array */ 177 struct mem_cgroup_threshold_ary *primary; 178 /* 179 * Spare threshold array. 180 * This is needed to make mem_cgroup_unregister_event() "never fail". 181 * It must be able to store at least primary->size - 1 entries. 182 */ 183 struct mem_cgroup_threshold_ary *spare; 184}; 185 186/* for OOM */ 187struct mem_cgroup_eventfd_list { 188 struct list_head list; 189 struct eventfd_ctx *eventfd; 190}; 191 192static void mem_cgroup_threshold(struct mem_cgroup *mem); 193static void mem_cgroup_oom_notify(struct mem_cgroup *mem); 194 195/* 196 * The memory controller data structure. The memory controller controls both 197 * page cache and RSS per cgroup. We would eventually like to provide 198 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 199 * to help the administrator determine what knobs to tune. 200 * 201 * TODO: Add a water mark for the memory controller. Reclaim will begin when 202 * we hit the water mark. May be even add a low water mark, such that 203 * no reclaim occurs from a cgroup at it's low water mark, this is 204 * a feature that will be implemented much later in the future. 205 */ 206struct mem_cgroup { 207 struct cgroup_subsys_state css; 208 /* 209 * the counter to account for memory usage 210 */ 211 struct res_counter res; 212 /* 213 * the counter to account for mem+swap usage. 214 */ 215 struct res_counter memsw; 216 /* 217 * Per cgroup active and inactive list, similar to the 218 * per zone LRU lists. 219 */ 220 struct mem_cgroup_lru_info info; 221 /* 222 * While reclaiming in a hierarchy, we cache the last child we 223 * reclaimed from. 224 */ 225 int last_scanned_child; 226 /* 227 * Should the accounting and control be hierarchical, per subtree? 228 */ 229 bool use_hierarchy; 230 atomic_t oom_lock; 231 atomic_t refcnt; 232 233 unsigned int swappiness; 234 /* OOM-Killer disable */ 235 int oom_kill_disable; 236 237 /* set when res.limit == memsw.limit */ 238 bool memsw_is_minimum; 239 240 /* protect arrays of thresholds */ 241 struct mutex thresholds_lock; 242 243 /* thresholds for memory usage. RCU-protected */ 244 struct mem_cgroup_thresholds thresholds; 245 246 /* thresholds for mem+swap usage. RCU-protected */ 247 struct mem_cgroup_thresholds memsw_thresholds; 248 249 /* For oom notifier event fd */ 250 struct list_head oom_notify; 251 252 /* 253 * Should we move charges of a task when a task is moved into this 254 * mem_cgroup ? And what type of charges should we move ? 255 */ 256 unsigned long move_charge_at_immigrate; 257 /* 258 * percpu counter. 259 */ 260 struct mem_cgroup_stat_cpu *stat; 261 /* 262 * used when a cpu is offlined or other synchronizations 263 * See mem_cgroup_read_stat(). 264 */ 265 struct mem_cgroup_stat_cpu nocpu_base; 266 spinlock_t pcp_counter_lock; 267}; 268 269/* Stuffs for move charges at task migration. */ 270/* 271 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a 272 * left-shifted bitmap of these types. 273 */ 274enum move_type { 275 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ 276 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ 277 NR_MOVE_TYPE, 278}; 279 280/* "mc" and its members are protected by cgroup_mutex */ 281static struct move_charge_struct { 282 spinlock_t lock; /* for from, to */ 283 struct mem_cgroup *from; 284 struct mem_cgroup *to; 285 unsigned long precharge; 286 unsigned long moved_charge; 287 unsigned long moved_swap; 288 struct task_struct *moving_task; /* a task moving charges */ 289 wait_queue_head_t waitq; /* a waitq for other context */ 290} mc = { 291 .lock = __SPIN_LOCK_UNLOCKED(mc.lock), 292 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), 293}; 294 295static bool move_anon(void) 296{ 297 return test_bit(MOVE_CHARGE_TYPE_ANON, 298 &mc.to->move_charge_at_immigrate); 299} 300 301static bool move_file(void) 302{ 303 return test_bit(MOVE_CHARGE_TYPE_FILE, 304 &mc.to->move_charge_at_immigrate); 305} 306 307/* 308 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft 309 * limit reclaim to prevent infinite loops, if they ever occur. 310 */ 311#define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) 312#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) 313 314enum charge_type { 315 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 316 MEM_CGROUP_CHARGE_TYPE_MAPPED, 317 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 318 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 319 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 320 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ 321 NR_CHARGE_TYPE, 322}; 323 324/* for encoding cft->private value on file */ 325#define _MEM (0) 326#define _MEMSWAP (1) 327#define _OOM_TYPE (2) 328#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 329#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 330#define MEMFILE_ATTR(val) ((val) & 0xffff) 331/* Used for OOM nofiier */ 332#define OOM_CONTROL (0) 333 334/* 335 * Reclaim flags for mem_cgroup_hierarchical_reclaim 336 */ 337#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 338#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) 339#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 340#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) 341#define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2 342#define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT) 343 344static void mem_cgroup_get(struct mem_cgroup *mem); 345static void mem_cgroup_put(struct mem_cgroup *mem); 346static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); 347static void drain_all_stock_async(void); 348 349static struct mem_cgroup_per_zone * 350mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) 351{ 352 return &mem->info.nodeinfo[nid]->zoneinfo[zid]; 353} 354 355struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem) 356{ 357 return &mem->css; 358} 359 360static struct mem_cgroup_per_zone * 361page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page) 362{ 363 int nid = page_to_nid(page); 364 int zid = page_zonenum(page); 365 366 return mem_cgroup_zoneinfo(mem, nid, zid); 367} 368 369static struct mem_cgroup_tree_per_zone * 370soft_limit_tree_node_zone(int nid, int zid) 371{ 372 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 373} 374 375static struct mem_cgroup_tree_per_zone * 376soft_limit_tree_from_page(struct page *page) 377{ 378 int nid = page_to_nid(page); 379 int zid = page_zonenum(page); 380 381 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 382} 383 384static void 385__mem_cgroup_insert_exceeded(struct mem_cgroup *mem, 386 struct mem_cgroup_per_zone *mz, 387 struct mem_cgroup_tree_per_zone *mctz, 388 unsigned long long new_usage_in_excess) 389{ 390 struct rb_node **p = &mctz->rb_root.rb_node; 391 struct rb_node *parent = NULL; 392 struct mem_cgroup_per_zone *mz_node; 393 394 if (mz->on_tree) 395 return; 396 397 mz->usage_in_excess = new_usage_in_excess; 398 if (!mz->usage_in_excess) 399 return; 400 while (*p) { 401 parent = *p; 402 mz_node = rb_entry(parent, struct mem_cgroup_per_zone, 403 tree_node); 404 if (mz->usage_in_excess < mz_node->usage_in_excess) 405 p = &(*p)->rb_left; 406 /* 407 * We can't avoid mem cgroups that are over their soft 408 * limit by the same amount 409 */ 410 else if (mz->usage_in_excess >= mz_node->usage_in_excess) 411 p = &(*p)->rb_right; 412 } 413 rb_link_node(&mz->tree_node, parent, p); 414 rb_insert_color(&mz->tree_node, &mctz->rb_root); 415 mz->on_tree = true; 416} 417 418static void 419__mem_cgroup_remove_exceeded(struct mem_cgroup *mem, 420 struct mem_cgroup_per_zone *mz, 421 struct mem_cgroup_tree_per_zone *mctz) 422{ 423 if (!mz->on_tree) 424 return; 425 rb_erase(&mz->tree_node, &mctz->rb_root); 426 mz->on_tree = false; 427} 428 429static void 430mem_cgroup_remove_exceeded(struct mem_cgroup *mem, 431 struct mem_cgroup_per_zone *mz, 432 struct mem_cgroup_tree_per_zone *mctz) 433{ 434 spin_lock(&mctz->lock); 435 __mem_cgroup_remove_exceeded(mem, mz, mctz); 436 spin_unlock(&mctz->lock); 437} 438 439 440static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) 441{ 442 unsigned long long excess; 443 struct mem_cgroup_per_zone *mz; 444 struct mem_cgroup_tree_per_zone *mctz; 445 int nid = page_to_nid(page); 446 int zid = page_zonenum(page); 447 mctz = soft_limit_tree_from_page(page); 448 449 /* 450 * Necessary to update all ancestors when hierarchy is used. 451 * because their event counter is not touched. 452 */ 453 for (; mem; mem = parent_mem_cgroup(mem)) { 454 mz = mem_cgroup_zoneinfo(mem, nid, zid); 455 excess = res_counter_soft_limit_excess(&mem->res); 456 /* 457 * We have to update the tree if mz is on RB-tree or 458 * mem is over its softlimit. 459 */ 460 if (excess || mz->on_tree) { 461 spin_lock(&mctz->lock); 462 /* if on-tree, remove it */ 463 if (mz->on_tree) 464 __mem_cgroup_remove_exceeded(mem, mz, mctz); 465 /* 466 * Insert again. mz->usage_in_excess will be updated. 467 * If excess is 0, no tree ops. 468 */ 469 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess); 470 spin_unlock(&mctz->lock); 471 } 472 } 473} 474 475static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem) 476{ 477 int node, zone; 478 struct mem_cgroup_per_zone *mz; 479 struct mem_cgroup_tree_per_zone *mctz; 480 481 for_each_node_state(node, N_POSSIBLE) { 482 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 483 mz = mem_cgroup_zoneinfo(mem, node, zone); 484 mctz = soft_limit_tree_node_zone(node, zone); 485 mem_cgroup_remove_exceeded(mem, mz, mctz); 486 } 487 } 488} 489 490static struct mem_cgroup_per_zone * 491__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 492{ 493 struct rb_node *rightmost = NULL; 494 struct mem_cgroup_per_zone *mz; 495 496retry: 497 mz = NULL; 498 rightmost = rb_last(&mctz->rb_root); 499 if (!rightmost) 500 goto done; /* Nothing to reclaim from */ 501 502 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); 503 /* 504 * Remove the node now but someone else can add it back, 505 * we will to add it back at the end of reclaim to its correct 506 * position in the tree. 507 */ 508 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 509 if (!res_counter_soft_limit_excess(&mz->mem->res) || 510 !css_tryget(&mz->mem->css)) 511 goto retry; 512done: 513 return mz; 514} 515 516static struct mem_cgroup_per_zone * 517mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 518{ 519 struct mem_cgroup_per_zone *mz; 520 521 spin_lock(&mctz->lock); 522 mz = __mem_cgroup_largest_soft_limit_node(mctz); 523 spin_unlock(&mctz->lock); 524 return mz; 525} 526 527/* 528 * Implementation Note: reading percpu statistics for memcg. 529 * 530 * Both of vmstat[] and percpu_counter has threshold and do periodic 531 * synchronization to implement "quick" read. There are trade-off between 532 * reading cost and precision of value. Then, we may have a chance to implement 533 * a periodic synchronizion of counter in memcg's counter. 534 * 535 * But this _read() function is used for user interface now. The user accounts 536 * memory usage by memory cgroup and he _always_ requires exact value because 537 * he accounts memory. Even if we provide quick-and-fuzzy read, we always 538 * have to visit all online cpus and make sum. So, for now, unnecessary 539 * synchronization is not implemented. (just implemented for cpu hotplug) 540 * 541 * If there are kernel internal actions which can make use of some not-exact 542 * value, and reading all cpu value can be performance bottleneck in some 543 * common workload, threashold and synchonization as vmstat[] should be 544 * implemented. 545 */ 546static s64 mem_cgroup_read_stat(struct mem_cgroup *mem, 547 enum mem_cgroup_stat_index idx) 548{ 549 int cpu; 550 s64 val = 0; 551 552 get_online_cpus(); 553 for_each_online_cpu(cpu) 554 val += per_cpu(mem->stat->count[idx], cpu); 555#ifdef CONFIG_HOTPLUG_CPU 556 spin_lock(&mem->pcp_counter_lock); 557 val += mem->nocpu_base.count[idx]; 558 spin_unlock(&mem->pcp_counter_lock); 559#endif 560 put_online_cpus(); 561 return val; 562} 563 564static s64 mem_cgroup_local_usage(struct mem_cgroup *mem) 565{ 566 s64 ret; 567 568 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); 569 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); 570 return ret; 571} 572 573static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, 574 bool charge) 575{ 576 int val = (charge) ? 1 : -1; 577 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); 578} 579 580static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, 581 bool file, int nr_pages) 582{ 583 preempt_disable(); 584 585 if (file) 586 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages); 587 else 588 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages); 589 590 /* pagein of a big page is an event. So, ignore page size */ 591 if (nr_pages > 0) 592 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]); 593 else { 594 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]); 595 nr_pages = -nr_pages; /* for event */ 596 } 597 598 __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages); 599 600 preempt_enable(); 601} 602 603static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, 604 enum lru_list idx) 605{ 606 int nid, zid; 607 struct mem_cgroup_per_zone *mz; 608 u64 total = 0; 609 610 for_each_online_node(nid) 611 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 612 mz = mem_cgroup_zoneinfo(mem, nid, zid); 613 total += MEM_CGROUP_ZSTAT(mz, idx); 614 } 615 return total; 616} 617 618static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift) 619{ 620 s64 val; 621 622 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]); 623 624 return !(val & ((1 << event_mask_shift) - 1)); 625} 626 627/* 628 * Check events in order. 629 * 630 */ 631static void memcg_check_events(struct mem_cgroup *mem, struct page *page) 632{ 633 /* threshold event is triggered in finer grain than soft limit */ 634 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) { 635 mem_cgroup_threshold(mem); 636 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH))) 637 mem_cgroup_update_tree(mem, page); 638 } 639} 640 641static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 642{ 643 return container_of(cgroup_subsys_state(cont, 644 mem_cgroup_subsys_id), struct mem_cgroup, 645 css); 646} 647 648struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 649{ 650 /* 651 * mm_update_next_owner() may clear mm->owner to NULL 652 * if it races with swapoff, page migration, etc. 653 * So this can be called with p == NULL. 654 */ 655 if (unlikely(!p)) 656 return NULL; 657 658 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 659 struct mem_cgroup, css); 660} 661 662static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 663{ 664 struct mem_cgroup *mem = NULL; 665 666 if (!mm) 667 return NULL; 668 /* 669 * Because we have no locks, mm->owner's may be being moved to other 670 * cgroup. We use css_tryget() here even if this looks 671 * pessimistic (rather than adding locks here). 672 */ 673 rcu_read_lock(); 674 do { 675 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 676 if (unlikely(!mem)) 677 break; 678 } while (!css_tryget(&mem->css)); 679 rcu_read_unlock(); 680 return mem; 681} 682 683/* The caller has to guarantee "mem" exists before calling this */ 684static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem) 685{ 686 struct cgroup_subsys_state *css; 687 int found; 688 689 if (!mem) /* ROOT cgroup has the smallest ID */ 690 return root_mem_cgroup; /*css_put/get against root is ignored*/ 691 if (!mem->use_hierarchy) { 692 if (css_tryget(&mem->css)) 693 return mem; 694 return NULL; 695 } 696 rcu_read_lock(); 697 /* 698 * searching a memory cgroup which has the smallest ID under given 699 * ROOT cgroup. (ID >= 1) 700 */ 701 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found); 702 if (css && css_tryget(css)) 703 mem = container_of(css, struct mem_cgroup, css); 704 else 705 mem = NULL; 706 rcu_read_unlock(); 707 return mem; 708} 709 710static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter, 711 struct mem_cgroup *root, 712 bool cond) 713{ 714 int nextid = css_id(&iter->css) + 1; 715 int found; 716 int hierarchy_used; 717 struct cgroup_subsys_state *css; 718 719 hierarchy_used = iter->use_hierarchy; 720 721 css_put(&iter->css); 722 /* If no ROOT, walk all, ignore hierarchy */ 723 if (!cond || (root && !hierarchy_used)) 724 return NULL; 725 726 if (!root) 727 root = root_mem_cgroup; 728 729 do { 730 iter = NULL; 731 rcu_read_lock(); 732 733 css = css_get_next(&mem_cgroup_subsys, nextid, 734 &root->css, &found); 735 if (css && css_tryget(css)) 736 iter = container_of(css, struct mem_cgroup, css); 737 rcu_read_unlock(); 738 /* If css is NULL, no more cgroups will be found */ 739 nextid = found + 1; 740 } while (css && !iter); 741 742 return iter; 743} 744/* 745 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please 746 * be careful that "break" loop is not allowed. We have reference count. 747 * Instead of that modify "cond" to be false and "continue" to exit the loop. 748 */ 749#define for_each_mem_cgroup_tree_cond(iter, root, cond) \ 750 for (iter = mem_cgroup_start_loop(root);\ 751 iter != NULL;\ 752 iter = mem_cgroup_get_next(iter, root, cond)) 753 754#define for_each_mem_cgroup_tree(iter, root) \ 755 for_each_mem_cgroup_tree_cond(iter, root, true) 756 757#define for_each_mem_cgroup_all(iter) \ 758 for_each_mem_cgroup_tree_cond(iter, NULL, true) 759 760 761static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) 762{ 763 return (mem == root_mem_cgroup); 764} 765 766/* 767 * Following LRU functions are allowed to be used without PCG_LOCK. 768 * Operations are called by routine of global LRU independently from memcg. 769 * What we have to take care of here is validness of pc->mem_cgroup. 770 * 771 * Changes to pc->mem_cgroup happens when 772 * 1. charge 773 * 2. moving account 774 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 775 * It is added to LRU before charge. 776 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 777 * When moving account, the page is not on LRU. It's isolated. 778 */ 779 780void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) 781{ 782 struct page_cgroup *pc; 783 struct mem_cgroup_per_zone *mz; 784 785 if (mem_cgroup_disabled()) 786 return; 787 pc = lookup_page_cgroup(page); 788 /* can happen while we handle swapcache. */ 789 if (!TestClearPageCgroupAcctLRU(pc)) 790 return; 791 VM_BUG_ON(!pc->mem_cgroup); 792 /* 793 * We don't check PCG_USED bit. It's cleared when the "page" is finally 794 * removed from global LRU. 795 */ 796 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 797 /* huge page split is done under lru_lock. so, we have no races. */ 798 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 799 if (mem_cgroup_is_root(pc->mem_cgroup)) 800 return; 801 VM_BUG_ON(list_empty(&pc->lru)); 802 list_del_init(&pc->lru); 803} 804 805void mem_cgroup_del_lru(struct page *page) 806{ 807 mem_cgroup_del_lru_list(page, page_lru(page)); 808} 809 810/* 811 * Writeback is about to end against a page which has been marked for immediate 812 * reclaim. If it still appears to be reclaimable, move it to the tail of the 813 * inactive list. 814 */ 815void mem_cgroup_rotate_reclaimable_page(struct page *page) 816{ 817 struct mem_cgroup_per_zone *mz; 818 struct page_cgroup *pc; 819 enum lru_list lru = page_lru(page); 820 821 if (mem_cgroup_disabled()) 822 return; 823 824 pc = lookup_page_cgroup(page); 825 /* unused or root page is not rotated. */ 826 if (!PageCgroupUsed(pc)) 827 return; 828 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 829 smp_rmb(); 830 if (mem_cgroup_is_root(pc->mem_cgroup)) 831 return; 832 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 833 list_move_tail(&pc->lru, &mz->lists[lru]); 834} 835 836void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) 837{ 838 struct mem_cgroup_per_zone *mz; 839 struct page_cgroup *pc; 840 841 if (mem_cgroup_disabled()) 842 return; 843 844 pc = lookup_page_cgroup(page); 845 /* unused or root page is not rotated. */ 846 if (!PageCgroupUsed(pc)) 847 return; 848 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 849 smp_rmb(); 850 if (mem_cgroup_is_root(pc->mem_cgroup)) 851 return; 852 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 853 list_move(&pc->lru, &mz->lists[lru]); 854} 855 856void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) 857{ 858 struct page_cgroup *pc; 859 struct mem_cgroup_per_zone *mz; 860 861 if (mem_cgroup_disabled()) 862 return; 863 pc = lookup_page_cgroup(page); 864 VM_BUG_ON(PageCgroupAcctLRU(pc)); 865 if (!PageCgroupUsed(pc)) 866 return; 867 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 868 smp_rmb(); 869 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 870 /* huge page split is done under lru_lock. so, we have no races. */ 871 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 872 SetPageCgroupAcctLRU(pc); 873 if (mem_cgroup_is_root(pc->mem_cgroup)) 874 return; 875 list_add(&pc->lru, &mz->lists[lru]); 876} 877 878/* 879 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to 880 * lru because the page may.be reused after it's fully uncharged (because of 881 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge 882 * it again. This function is only used to charge SwapCache. It's done under 883 * lock_page and expected that zone->lru_lock is never held. 884 */ 885static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) 886{ 887 unsigned long flags; 888 struct zone *zone = page_zone(page); 889 struct page_cgroup *pc = lookup_page_cgroup(page); 890 891 spin_lock_irqsave(&zone->lru_lock, flags); 892 /* 893 * Forget old LRU when this page_cgroup is *not* used. This Used bit 894 * is guarded by lock_page() because the page is SwapCache. 895 */ 896 if (!PageCgroupUsed(pc)) 897 mem_cgroup_del_lru_list(page, page_lru(page)); 898 spin_unlock_irqrestore(&zone->lru_lock, flags); 899} 900 901static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) 902{ 903 unsigned long flags; 904 struct zone *zone = page_zone(page); 905 struct page_cgroup *pc = lookup_page_cgroup(page); 906 907 spin_lock_irqsave(&zone->lru_lock, flags); 908 /* link when the page is linked to LRU but page_cgroup isn't */ 909 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) 910 mem_cgroup_add_lru_list(page, page_lru(page)); 911 spin_unlock_irqrestore(&zone->lru_lock, flags); 912} 913 914 915void mem_cgroup_move_lists(struct page *page, 916 enum lru_list from, enum lru_list to) 917{ 918 if (mem_cgroup_disabled()) 919 return; 920 mem_cgroup_del_lru_list(page, from); 921 mem_cgroup_add_lru_list(page, to); 922} 923 924int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) 925{ 926 int ret; 927 struct mem_cgroup *curr = NULL; 928 struct task_struct *p; 929 930 p = find_lock_task_mm(task); 931 if (!p) 932 return 0; 933 curr = try_get_mem_cgroup_from_mm(p->mm); 934 task_unlock(p); 935 if (!curr) 936 return 0; 937 /* 938 * We should check use_hierarchy of "mem" not "curr". Because checking 939 * use_hierarchy of "curr" here make this function true if hierarchy is 940 * enabled in "curr" and "curr" is a child of "mem" in *cgroup* 941 * hierarchy(even if use_hierarchy is disabled in "mem"). 942 */ 943 if (mem->use_hierarchy) 944 ret = css_is_ancestor(&curr->css, &mem->css); 945 else 946 ret = (curr == mem); 947 css_put(&curr->css); 948 return ret; 949} 950 951static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) 952{ 953 unsigned long active; 954 unsigned long inactive; 955 unsigned long gb; 956 unsigned long inactive_ratio; 957 958 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); 959 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); 960 961 gb = (inactive + active) >> (30 - PAGE_SHIFT); 962 if (gb) 963 inactive_ratio = int_sqrt(10 * gb); 964 else 965 inactive_ratio = 1; 966 967 if (present_pages) { 968 present_pages[0] = inactive; 969 present_pages[1] = active; 970 } 971 972 return inactive_ratio; 973} 974 975int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) 976{ 977 unsigned long active; 978 unsigned long inactive; 979 unsigned long present_pages[2]; 980 unsigned long inactive_ratio; 981 982 inactive_ratio = calc_inactive_ratio(memcg, present_pages); 983 984 inactive = present_pages[0]; 985 active = present_pages[1]; 986 987 if (inactive * inactive_ratio < active) 988 return 1; 989 990 return 0; 991} 992 993int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) 994{ 995 unsigned long active; 996 unsigned long inactive; 997 998 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); 999 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); 1000 1001 return (active > inactive); 1002} 1003 1004unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, 1005 struct zone *zone, 1006 enum lru_list lru) 1007{ 1008 int nid = zone_to_nid(zone); 1009 int zid = zone_idx(zone); 1010 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1011 1012 return MEM_CGROUP_ZSTAT(mz, lru); 1013} 1014 1015struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1016 struct zone *zone) 1017{ 1018 int nid = zone_to_nid(zone); 1019 int zid = zone_idx(zone); 1020 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1021 1022 return &mz->reclaim_stat; 1023} 1024 1025struct zone_reclaim_stat * 1026mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1027{ 1028 struct page_cgroup *pc; 1029 struct mem_cgroup_per_zone *mz; 1030 1031 if (mem_cgroup_disabled()) 1032 return NULL; 1033 1034 pc = lookup_page_cgroup(page); 1035 if (!PageCgroupUsed(pc)) 1036 return NULL; 1037 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1038 smp_rmb(); 1039 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1040 return &mz->reclaim_stat; 1041} 1042 1043unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, 1044 struct list_head *dst, 1045 unsigned long *scanned, int order, 1046 int mode, struct zone *z, 1047 struct mem_cgroup *mem_cont, 1048 int active, int file) 1049{ 1050 unsigned long nr_taken = 0; 1051 struct page *page; 1052 unsigned long scan; 1053 LIST_HEAD(pc_list); 1054 struct list_head *src; 1055 struct page_cgroup *pc, *tmp; 1056 int nid = zone_to_nid(z); 1057 int zid = zone_idx(z); 1058 struct mem_cgroup_per_zone *mz; 1059 int lru = LRU_FILE * file + active; 1060 int ret; 1061 1062 BUG_ON(!mem_cont); 1063 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 1064 src = &mz->lists[lru]; 1065 1066 scan = 0; 1067 list_for_each_entry_safe_reverse(pc, tmp, src, lru) { 1068 if (scan >= nr_to_scan) 1069 break; 1070 1071 if (unlikely(!PageCgroupUsed(pc))) 1072 continue; 1073 1074 page = lookup_cgroup_page(pc); 1075 1076 if (unlikely(!PageLRU(page))) 1077 continue; 1078 1079 scan++; 1080 ret = __isolate_lru_page(page, mode, file); 1081 switch (ret) { 1082 case 0: 1083 list_move(&page->lru, dst); 1084 mem_cgroup_del_lru(page); 1085 nr_taken += hpage_nr_pages(page); 1086 break; 1087 case -EBUSY: 1088 /* we don't affect global LRU but rotate in our LRU */ 1089 mem_cgroup_rotate_lru_list(page, page_lru(page)); 1090 break; 1091 default: 1092 break; 1093 } 1094 } 1095 1096 *scanned = scan; 1097 1098 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken, 1099 0, 0, 0, mode); 1100 1101 return nr_taken; 1102} 1103 1104#define mem_cgroup_from_res_counter(counter, member) \ 1105 container_of(counter, struct mem_cgroup, member) 1106 1107/** 1108 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1109 * @mem: the memory cgroup 1110 * 1111 * Returns the maximum amount of memory @mem can be charged with, in 1112 * bytes. 1113 */ 1114static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem) 1115{ 1116 unsigned long long margin; 1117 1118 margin = res_counter_margin(&mem->res); 1119 if (do_swap_account) 1120 margin = min(margin, res_counter_margin(&mem->memsw)); 1121 return margin; 1122} 1123 1124static unsigned int get_swappiness(struct mem_cgroup *memcg) 1125{ 1126 struct cgroup *cgrp = memcg->css.cgroup; 1127 1128 /* root ? */ 1129 if (cgrp->parent == NULL) 1130 return vm_swappiness; 1131 1132 return memcg->swappiness; 1133} 1134 1135static void mem_cgroup_start_move(struct mem_cgroup *mem) 1136{ 1137 int cpu; 1138 1139 get_online_cpus(); 1140 spin_lock(&mem->pcp_counter_lock); 1141 for_each_online_cpu(cpu) 1142 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1143 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1144 spin_unlock(&mem->pcp_counter_lock); 1145 put_online_cpus(); 1146 1147 synchronize_rcu(); 1148} 1149 1150static void mem_cgroup_end_move(struct mem_cgroup *mem) 1151{ 1152 int cpu; 1153 1154 if (!mem) 1155 return; 1156 get_online_cpus(); 1157 spin_lock(&mem->pcp_counter_lock); 1158 for_each_online_cpu(cpu) 1159 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1160 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1161 spin_unlock(&mem->pcp_counter_lock); 1162 put_online_cpus(); 1163} 1164/* 1165 * 2 routines for checking "mem" is under move_account() or not. 1166 * 1167 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1168 * for avoiding race in accounting. If true, 1169 * pc->mem_cgroup may be overwritten. 1170 * 1171 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1172 * under hierarchy of moving cgroups. This is for 1173 * waiting at hith-memory prressure caused by "move". 1174 */ 1175 1176static bool mem_cgroup_stealed(struct mem_cgroup *mem) 1177{ 1178 VM_BUG_ON(!rcu_read_lock_held()); 1179 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1180} 1181 1182static bool mem_cgroup_under_move(struct mem_cgroup *mem) 1183{ 1184 struct mem_cgroup *from; 1185 struct mem_cgroup *to; 1186 bool ret = false; 1187 /* 1188 * Unlike task_move routines, we access mc.to, mc.from not under 1189 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1190 */ 1191 spin_lock(&mc.lock); 1192 from = mc.from; 1193 to = mc.to; 1194 if (!from) 1195 goto unlock; 1196 if (from == mem || to == mem 1197 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css)) 1198 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css))) 1199 ret = true; 1200unlock: 1201 spin_unlock(&mc.lock); 1202 return ret; 1203} 1204 1205static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem) 1206{ 1207 if (mc.moving_task && current != mc.moving_task) { 1208 if (mem_cgroup_under_move(mem)) { 1209 DEFINE_WAIT(wait); 1210 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1211 /* moving charge context might have finished. */ 1212 if (mc.moving_task) 1213 schedule(); 1214 finish_wait(&mc.waitq, &wait); 1215 return true; 1216 } 1217 } 1218 return false; 1219} 1220 1221/** 1222 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1223 * @memcg: The memory cgroup that went over limit 1224 * @p: Task that is going to be killed 1225 * 1226 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1227 * enabled 1228 */ 1229void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1230{ 1231 struct cgroup *task_cgrp; 1232 struct cgroup *mem_cgrp; 1233 /* 1234 * Need a buffer in BSS, can't rely on allocations. The code relies 1235 * on the assumption that OOM is serialized for memory controller. 1236 * If this assumption is broken, revisit this code. 1237 */ 1238 static char memcg_name[PATH_MAX]; 1239 int ret; 1240 1241 if (!memcg || !p) 1242 return; 1243 1244 1245 rcu_read_lock(); 1246 1247 mem_cgrp = memcg->css.cgroup; 1248 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1249 1250 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1251 if (ret < 0) { 1252 /* 1253 * Unfortunately, we are unable to convert to a useful name 1254 * But we'll still print out the usage information 1255 */ 1256 rcu_read_unlock(); 1257 goto done; 1258 } 1259 rcu_read_unlock(); 1260 1261 printk(KERN_INFO "Task in %s killed", memcg_name); 1262 1263 rcu_read_lock(); 1264 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1265 if (ret < 0) { 1266 rcu_read_unlock(); 1267 goto done; 1268 } 1269 rcu_read_unlock(); 1270 1271 /* 1272 * Continues from above, so we don't need an KERN_ level 1273 */ 1274 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1275done: 1276 1277 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1278 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1279 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1280 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1281 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1282 "failcnt %llu\n", 1283 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1284 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1285 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1286} 1287 1288/* 1289 * This function returns the number of memcg under hierarchy tree. Returns 1290 * 1(self count) if no children. 1291 */ 1292static int mem_cgroup_count_children(struct mem_cgroup *mem) 1293{ 1294 int num = 0; 1295 struct mem_cgroup *iter; 1296 1297 for_each_mem_cgroup_tree(iter, mem) 1298 num++; 1299 return num; 1300} 1301 1302/* 1303 * Return the memory (and swap, if configured) limit for a memcg. 1304 */ 1305u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1306{ 1307 u64 limit; 1308 u64 memsw; 1309 1310 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1311 limit += total_swap_pages << PAGE_SHIFT; 1312 1313 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1314 /* 1315 * If memsw is finite and limits the amount of swap space available 1316 * to this memcg, return that limit. 1317 */ 1318 return min(limit, memsw); 1319} 1320 1321/* 1322 * Visit the first child (need not be the first child as per the ordering 1323 * of the cgroup list, since we track last_scanned_child) of @mem and use 1324 * that to reclaim free pages from. 1325 */ 1326static struct mem_cgroup * 1327mem_cgroup_select_victim(struct mem_cgroup *root_mem) 1328{ 1329 struct mem_cgroup *ret = NULL; 1330 struct cgroup_subsys_state *css; 1331 int nextid, found; 1332 1333 if (!root_mem->use_hierarchy) { 1334 css_get(&root_mem->css); 1335 ret = root_mem; 1336 } 1337 1338 while (!ret) { 1339 rcu_read_lock(); 1340 nextid = root_mem->last_scanned_child + 1; 1341 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, 1342 &found); 1343 if (css && css_tryget(css)) 1344 ret = container_of(css, struct mem_cgroup, css); 1345 1346 rcu_read_unlock(); 1347 /* Updates scanning parameter */ 1348 if (!css) { 1349 /* this means start scan from ID:1 */ 1350 root_mem->last_scanned_child = 0; 1351 } else 1352 root_mem->last_scanned_child = found; 1353 } 1354 1355 return ret; 1356} 1357 1358/* 1359 * Scan the hierarchy if needed to reclaim memory. We remember the last child 1360 * we reclaimed from, so that we don't end up penalizing one child extensively 1361 * based on its position in the children list. 1362 * 1363 * root_mem is the original ancestor that we've been reclaim from. 1364 * 1365 * We give up and return to the caller when we visit root_mem twice. 1366 * (other groups can be removed while we're walking....) 1367 * 1368 * If shrink==true, for avoiding to free too much, this returns immedieately. 1369 */ 1370static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, 1371 struct zone *zone, 1372 gfp_t gfp_mask, 1373 unsigned long reclaim_options) 1374{ 1375 struct mem_cgroup *victim; 1376 int ret, total = 0; 1377 int loop = 0; 1378 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP; 1379 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK; 1380 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT; 1381 unsigned long excess; 1382 1383 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT; 1384 1385 /* If memsw_is_minimum==1, swap-out is of-no-use. */ 1386 if (root_mem->memsw_is_minimum) 1387 noswap = true; 1388 1389 while (1) { 1390 victim = mem_cgroup_select_victim(root_mem); 1391 if (victim == root_mem) { 1392 loop++; 1393 if (loop >= 1) 1394 drain_all_stock_async(); 1395 if (loop >= 2) { 1396 /* 1397 * If we have not been able to reclaim 1398 * anything, it might because there are 1399 * no reclaimable pages under this hierarchy 1400 */ 1401 if (!check_soft || !total) { 1402 css_put(&victim->css); 1403 break; 1404 } 1405 /* 1406 * We want to do more targetted reclaim. 1407 * excess >> 2 is not to excessive so as to 1408 * reclaim too much, nor too less that we keep 1409 * coming back to reclaim from this cgroup 1410 */ 1411 if (total >= (excess >> 2) || 1412 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) { 1413 css_put(&victim->css); 1414 break; 1415 } 1416 } 1417 } 1418 if (!mem_cgroup_local_usage(victim)) { 1419 /* this cgroup's local usage == 0 */ 1420 css_put(&victim->css); 1421 continue; 1422 } 1423 /* we use swappiness of local cgroup */ 1424 if (check_soft) 1425 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask, 1426 noswap, get_swappiness(victim), zone); 1427 else 1428 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, 1429 noswap, get_swappiness(victim)); 1430 css_put(&victim->css); 1431 /* 1432 * At shrinking usage, we can't check we should stop here or 1433 * reclaim more. It's depends on callers. last_scanned_child 1434 * will work enough for keeping fairness under tree. 1435 */ 1436 if (shrink) 1437 return ret; 1438 total += ret; 1439 if (check_soft) { 1440 if (!res_counter_soft_limit_excess(&root_mem->res)) 1441 return total; 1442 } else if (mem_cgroup_margin(root_mem)) 1443 return 1 + total; 1444 } 1445 return total; 1446} 1447 1448/* 1449 * Check OOM-Killer is already running under our hierarchy. 1450 * If someone is running, return false. 1451 */ 1452static bool mem_cgroup_oom_lock(struct mem_cgroup *mem) 1453{ 1454 int x, lock_count = 0; 1455 struct mem_cgroup *iter; 1456 1457 for_each_mem_cgroup_tree(iter, mem) { 1458 x = atomic_inc_return(&iter->oom_lock); 1459 lock_count = max(x, lock_count); 1460 } 1461 1462 if (lock_count == 1) 1463 return true; 1464 return false; 1465} 1466 1467static int mem_cgroup_oom_unlock(struct mem_cgroup *mem) 1468{ 1469 struct mem_cgroup *iter; 1470 1471 /* 1472 * When a new child is created while the hierarchy is under oom, 1473 * mem_cgroup_oom_lock() may not be called. We have to use 1474 * atomic_add_unless() here. 1475 */ 1476 for_each_mem_cgroup_tree(iter, mem) 1477 atomic_add_unless(&iter->oom_lock, -1, 0); 1478 return 0; 1479} 1480 1481 1482static DEFINE_MUTEX(memcg_oom_mutex); 1483static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1484 1485struct oom_wait_info { 1486 struct mem_cgroup *mem; 1487 wait_queue_t wait; 1488}; 1489 1490static int memcg_oom_wake_function(wait_queue_t *wait, 1491 unsigned mode, int sync, void *arg) 1492{ 1493 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg; 1494 struct oom_wait_info *oom_wait_info; 1495 1496 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1497 1498 if (oom_wait_info->mem == wake_mem) 1499 goto wakeup; 1500 /* if no hierarchy, no match */ 1501 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy) 1502 return 0; 1503 /* 1504 * Both of oom_wait_info->mem and wake_mem are stable under us. 1505 * Then we can use css_is_ancestor without taking care of RCU. 1506 */ 1507 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) && 1508 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css)) 1509 return 0; 1510 1511wakeup: 1512 return autoremove_wake_function(wait, mode, sync, arg); 1513} 1514 1515static void memcg_wakeup_oom(struct mem_cgroup *mem) 1516{ 1517 /* for filtering, pass "mem" as argument. */ 1518 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem); 1519} 1520 1521static void memcg_oom_recover(struct mem_cgroup *mem) 1522{ 1523 if (mem && atomic_read(&mem->oom_lock)) 1524 memcg_wakeup_oom(mem); 1525} 1526 1527/* 1528 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1529 */ 1530bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask) 1531{ 1532 struct oom_wait_info owait; 1533 bool locked, need_to_kill; 1534 1535 owait.mem = mem; 1536 owait.wait.flags = 0; 1537 owait.wait.func = memcg_oom_wake_function; 1538 owait.wait.private = current; 1539 INIT_LIST_HEAD(&owait.wait.task_list); 1540 need_to_kill = true; 1541 /* At first, try to OOM lock hierarchy under mem.*/ 1542 mutex_lock(&memcg_oom_mutex); 1543 locked = mem_cgroup_oom_lock(mem); 1544 /* 1545 * Even if signal_pending(), we can't quit charge() loop without 1546 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1547 * under OOM is always welcomed, use TASK_KILLABLE here. 1548 */ 1549 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1550 if (!locked || mem->oom_kill_disable) 1551 need_to_kill = false; 1552 if (locked) 1553 mem_cgroup_oom_notify(mem); 1554 mutex_unlock(&memcg_oom_mutex); 1555 1556 if (need_to_kill) { 1557 finish_wait(&memcg_oom_waitq, &owait.wait); 1558 mem_cgroup_out_of_memory(mem, mask); 1559 } else { 1560 schedule(); 1561 finish_wait(&memcg_oom_waitq, &owait.wait); 1562 } 1563 mutex_lock(&memcg_oom_mutex); 1564 mem_cgroup_oom_unlock(mem); 1565 memcg_wakeup_oom(mem); 1566 mutex_unlock(&memcg_oom_mutex); 1567 1568 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1569 return false; 1570 /* Give chance to dying process */ 1571 schedule_timeout(1); 1572 return true; 1573} 1574 1575/* 1576 * Currently used to update mapped file statistics, but the routine can be 1577 * generalized to update other statistics as well. 1578 * 1579 * Notes: Race condition 1580 * 1581 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1582 * it tends to be costly. But considering some conditions, we doesn't need 1583 * to do so _always_. 1584 * 1585 * Considering "charge", lock_page_cgroup() is not required because all 1586 * file-stat operations happen after a page is attached to radix-tree. There 1587 * are no race with "charge". 1588 * 1589 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1590 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1591 * if there are race with "uncharge". Statistics itself is properly handled 1592 * by flags. 1593 * 1594 * Considering "move", this is an only case we see a race. To make the race 1595 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1596 * possibility of race condition. If there is, we take a lock. 1597 */ 1598 1599void mem_cgroup_update_page_stat(struct page *page, 1600 enum mem_cgroup_page_stat_item idx, int val) 1601{ 1602 struct mem_cgroup *mem; 1603 struct page_cgroup *pc = lookup_page_cgroup(page); 1604 bool need_unlock = false; 1605 unsigned long uninitialized_var(flags); 1606 1607 if (unlikely(!pc)) 1608 return; 1609 1610 rcu_read_lock(); 1611 mem = pc->mem_cgroup; 1612 if (unlikely(!mem || !PageCgroupUsed(pc))) 1613 goto out; 1614 /* pc->mem_cgroup is unstable ? */ 1615 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) { 1616 /* take a lock against to access pc->mem_cgroup */ 1617 move_lock_page_cgroup(pc, &flags); 1618 need_unlock = true; 1619 mem = pc->mem_cgroup; 1620 if (!mem || !PageCgroupUsed(pc)) 1621 goto out; 1622 } 1623 1624 switch (idx) { 1625 case MEMCG_NR_FILE_MAPPED: 1626 if (val > 0) 1627 SetPageCgroupFileMapped(pc); 1628 else if (!page_mapped(page)) 1629 ClearPageCgroupFileMapped(pc); 1630 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1631 break; 1632 default: 1633 BUG(); 1634 } 1635 1636 this_cpu_add(mem->stat->count[idx], val); 1637 1638out: 1639 if (unlikely(need_unlock)) 1640 move_unlock_page_cgroup(pc, &flags); 1641 rcu_read_unlock(); 1642 return; 1643} 1644EXPORT_SYMBOL(mem_cgroup_update_page_stat); 1645 1646/* 1647 * size of first charge trial. "32" comes from vmscan.c's magic value. 1648 * TODO: maybe necessary to use big numbers in big irons. 1649 */ 1650#define CHARGE_SIZE (32 * PAGE_SIZE) 1651struct memcg_stock_pcp { 1652 struct mem_cgroup *cached; /* this never be root cgroup */ 1653 unsigned int nr_pages; 1654 struct work_struct work; 1655}; 1656static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 1657static atomic_t memcg_drain_count; 1658 1659/* 1660 * Try to consume stocked charge on this cpu. If success, one page is consumed 1661 * from local stock and true is returned. If the stock is 0 or charges from a 1662 * cgroup which is not current target, returns false. This stock will be 1663 * refilled. 1664 */ 1665static bool consume_stock(struct mem_cgroup *mem) 1666{ 1667 struct memcg_stock_pcp *stock; 1668 bool ret = true; 1669 1670 stock = &get_cpu_var(memcg_stock); 1671 if (mem == stock->cached && stock->nr_pages) 1672 stock->nr_pages--; 1673 else /* need to call res_counter_charge */ 1674 ret = false; 1675 put_cpu_var(memcg_stock); 1676 return ret; 1677} 1678 1679/* 1680 * Returns stocks cached in percpu to res_counter and reset cached information. 1681 */ 1682static void drain_stock(struct memcg_stock_pcp *stock) 1683{ 1684 struct mem_cgroup *old = stock->cached; 1685 1686 if (stock->nr_pages) { 1687 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 1688 1689 res_counter_uncharge(&old->res, bytes); 1690 if (do_swap_account) 1691 res_counter_uncharge(&old->memsw, bytes); 1692 stock->nr_pages = 0; 1693 } 1694 stock->cached = NULL; 1695} 1696 1697/* 1698 * This must be called under preempt disabled or must be called by 1699 * a thread which is pinned to local cpu. 1700 */ 1701static void drain_local_stock(struct work_struct *dummy) 1702{ 1703 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 1704 drain_stock(stock); 1705} 1706 1707/* 1708 * Cache charges(val) which is from res_counter, to local per_cpu area. 1709 * This will be consumed by consume_stock() function, later. 1710 */ 1711static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages) 1712{ 1713 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 1714 1715 if (stock->cached != mem) { /* reset if necessary */ 1716 drain_stock(stock); 1717 stock->cached = mem; 1718 } 1719 stock->nr_pages += nr_pages; 1720 put_cpu_var(memcg_stock); 1721} 1722 1723/* 1724 * Tries to drain stocked charges in other cpus. This function is asynchronous 1725 * and just put a work per cpu for draining localy on each cpu. Caller can 1726 * expects some charges will be back to res_counter later but cannot wait for 1727 * it. 1728 */ 1729static void drain_all_stock_async(void) 1730{ 1731 int cpu; 1732 /* This function is for scheduling "drain" in asynchronous way. 1733 * The result of "drain" is not directly handled by callers. Then, 1734 * if someone is calling drain, we don't have to call drain more. 1735 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if 1736 * there is a race. We just do loose check here. 1737 */ 1738 if (atomic_read(&memcg_drain_count)) 1739 return; 1740 /* Notify other cpus that system-wide "drain" is running */ 1741 atomic_inc(&memcg_drain_count); 1742 get_online_cpus(); 1743 for_each_online_cpu(cpu) { 1744 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 1745 schedule_work_on(cpu, &stock->work); 1746 } 1747 put_online_cpus(); 1748 atomic_dec(&memcg_drain_count); 1749 /* We don't wait for flush_work */ 1750} 1751 1752/* This is a synchronous drain interface. */ 1753static void drain_all_stock_sync(void) 1754{ 1755 /* called when force_empty is called */ 1756 atomic_inc(&memcg_drain_count); 1757 schedule_on_each_cpu(drain_local_stock); 1758 atomic_dec(&memcg_drain_count); 1759} 1760 1761/* 1762 * This function drains percpu counter value from DEAD cpu and 1763 * move it to local cpu. Note that this function can be preempted. 1764 */ 1765static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu) 1766{ 1767 int i; 1768 1769 spin_lock(&mem->pcp_counter_lock); 1770 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 1771 s64 x = per_cpu(mem->stat->count[i], cpu); 1772 1773 per_cpu(mem->stat->count[i], cpu) = 0; 1774 mem->nocpu_base.count[i] += x; 1775 } 1776 /* need to clear ON_MOVE value, works as a kind of lock. */ 1777 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 1778 spin_unlock(&mem->pcp_counter_lock); 1779} 1780 1781static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu) 1782{ 1783 int idx = MEM_CGROUP_ON_MOVE; 1784 1785 spin_lock(&mem->pcp_counter_lock); 1786 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx]; 1787 spin_unlock(&mem->pcp_counter_lock); 1788} 1789 1790static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 1791 unsigned long action, 1792 void *hcpu) 1793{ 1794 int cpu = (unsigned long)hcpu; 1795 struct memcg_stock_pcp *stock; 1796 struct mem_cgroup *iter; 1797 1798 if ((action == CPU_ONLINE)) { 1799 for_each_mem_cgroup_all(iter) 1800 synchronize_mem_cgroup_on_move(iter, cpu); 1801 return NOTIFY_OK; 1802 } 1803 1804 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 1805 return NOTIFY_OK; 1806 1807 for_each_mem_cgroup_all(iter) 1808 mem_cgroup_drain_pcp_counter(iter, cpu); 1809 1810 stock = &per_cpu(memcg_stock, cpu); 1811 drain_stock(stock); 1812 return NOTIFY_OK; 1813} 1814 1815 1816/* See __mem_cgroup_try_charge() for details */ 1817enum { 1818 CHARGE_OK, /* success */ 1819 CHARGE_RETRY, /* need to retry but retry is not bad */ 1820 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 1821 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 1822 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 1823}; 1824 1825static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask, 1826 int csize, bool oom_check) 1827{ 1828 struct mem_cgroup *mem_over_limit; 1829 struct res_counter *fail_res; 1830 unsigned long flags = 0; 1831 int ret; 1832 1833 ret = res_counter_charge(&mem->res, csize, &fail_res); 1834 1835 if (likely(!ret)) { 1836 if (!do_swap_account) 1837 return CHARGE_OK; 1838 ret = res_counter_charge(&mem->memsw, csize, &fail_res); 1839 if (likely(!ret)) 1840 return CHARGE_OK; 1841 1842 res_counter_uncharge(&mem->res, csize); 1843 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 1844 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 1845 } else 1846 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 1847 /* 1848 * csize can be either a huge page (HPAGE_SIZE), a batch of 1849 * regular pages (CHARGE_SIZE), or a single regular page 1850 * (PAGE_SIZE). 1851 * 1852 * Never reclaim on behalf of optional batching, retry with a 1853 * single page instead. 1854 */ 1855 if (csize == CHARGE_SIZE) 1856 return CHARGE_RETRY; 1857 1858 if (!(gfp_mask & __GFP_WAIT)) 1859 return CHARGE_WOULDBLOCK; 1860 1861 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL, 1862 gfp_mask, flags); 1863 if (mem_cgroup_margin(mem_over_limit) >= csize) 1864 return CHARGE_RETRY; 1865 /* 1866 * Even though the limit is exceeded at this point, reclaim 1867 * may have been able to free some pages. Retry the charge 1868 * before killing the task. 1869 * 1870 * Only for regular pages, though: huge pages are rather 1871 * unlikely to succeed so close to the limit, and we fall back 1872 * to regular pages anyway in case of failure. 1873 */ 1874 if (csize == PAGE_SIZE && ret) 1875 return CHARGE_RETRY; 1876 1877 /* 1878 * At task move, charge accounts can be doubly counted. So, it's 1879 * better to wait until the end of task_move if something is going on. 1880 */ 1881 if (mem_cgroup_wait_acct_move(mem_over_limit)) 1882 return CHARGE_RETRY; 1883 1884 /* If we don't need to call oom-killer at el, return immediately */ 1885 if (!oom_check) 1886 return CHARGE_NOMEM; 1887 /* check OOM */ 1888 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 1889 return CHARGE_OOM_DIE; 1890 1891 return CHARGE_RETRY; 1892} 1893 1894/* 1895 * Unlike exported interface, "oom" parameter is added. if oom==true, 1896 * oom-killer can be invoked. 1897 */ 1898static int __mem_cgroup_try_charge(struct mm_struct *mm, 1899 gfp_t gfp_mask, 1900 struct mem_cgroup **memcg, bool oom, 1901 int page_size) 1902{ 1903 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 1904 struct mem_cgroup *mem = NULL; 1905 int ret; 1906 int csize = max(CHARGE_SIZE, (unsigned long) page_size); 1907 1908 /* 1909 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 1910 * in system level. So, allow to go ahead dying process in addition to 1911 * MEMDIE process. 1912 */ 1913 if (unlikely(test_thread_flag(TIF_MEMDIE) 1914 || fatal_signal_pending(current))) 1915 goto bypass; 1916 1917 /* 1918 * We always charge the cgroup the mm_struct belongs to. 1919 * The mm_struct's mem_cgroup changes on task migration if the 1920 * thread group leader migrates. It's possible that mm is not 1921 * set, if so charge the init_mm (happens for pagecache usage). 1922 */ 1923 if (!*memcg && !mm) 1924 goto bypass; 1925again: 1926 if (*memcg) { /* css should be a valid one */ 1927 mem = *memcg; 1928 VM_BUG_ON(css_is_removed(&mem->css)); 1929 if (mem_cgroup_is_root(mem)) 1930 goto done; 1931 if (page_size == PAGE_SIZE && consume_stock(mem)) 1932 goto done; 1933 css_get(&mem->css); 1934 } else { 1935 struct task_struct *p; 1936 1937 rcu_read_lock(); 1938 p = rcu_dereference(mm->owner); 1939 /* 1940 * Because we don't have task_lock(), "p" can exit. 1941 * In that case, "mem" can point to root or p can be NULL with 1942 * race with swapoff. Then, we have small risk of mis-accouning. 1943 * But such kind of mis-account by race always happens because 1944 * we don't have cgroup_mutex(). It's overkill and we allo that 1945 * small race, here. 1946 * (*) swapoff at el will charge against mm-struct not against 1947 * task-struct. So, mm->owner can be NULL. 1948 */ 1949 mem = mem_cgroup_from_task(p); 1950 if (!mem || mem_cgroup_is_root(mem)) { 1951 rcu_read_unlock(); 1952 goto done; 1953 } 1954 if (page_size == PAGE_SIZE && consume_stock(mem)) { 1955 /* 1956 * It seems dagerous to access memcg without css_get(). 1957 * But considering how consume_stok works, it's not 1958 * necessary. If consume_stock success, some charges 1959 * from this memcg are cached on this cpu. So, we 1960 * don't need to call css_get()/css_tryget() before 1961 * calling consume_stock(). 1962 */ 1963 rcu_read_unlock(); 1964 goto done; 1965 } 1966 /* after here, we may be blocked. we need to get refcnt */ 1967 if (!css_tryget(&mem->css)) { 1968 rcu_read_unlock(); 1969 goto again; 1970 } 1971 rcu_read_unlock(); 1972 } 1973 1974 do { 1975 bool oom_check; 1976 1977 /* If killed, bypass charge */ 1978 if (fatal_signal_pending(current)) { 1979 css_put(&mem->css); 1980 goto bypass; 1981 } 1982 1983 oom_check = false; 1984 if (oom && !nr_oom_retries) { 1985 oom_check = true; 1986 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 1987 } 1988 1989 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check); 1990 1991 switch (ret) { 1992 case CHARGE_OK: 1993 break; 1994 case CHARGE_RETRY: /* not in OOM situation but retry */ 1995 csize = page_size; 1996 css_put(&mem->css); 1997 mem = NULL; 1998 goto again; 1999 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2000 css_put(&mem->css); 2001 goto nomem; 2002 case CHARGE_NOMEM: /* OOM routine works */ 2003 if (!oom) { 2004 css_put(&mem->css); 2005 goto nomem; 2006 } 2007 /* If oom, we never return -ENOMEM */ 2008 nr_oom_retries--; 2009 break; 2010 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2011 css_put(&mem->css); 2012 goto bypass; 2013 } 2014 } while (ret != CHARGE_OK); 2015 2016 if (csize > page_size) 2017 refill_stock(mem, (csize - page_size) >> PAGE_SHIFT); 2018 css_put(&mem->css); 2019done: 2020 *memcg = mem; 2021 return 0; 2022nomem: 2023 *memcg = NULL; 2024 return -ENOMEM; 2025bypass: 2026 *memcg = NULL; 2027 return 0; 2028} 2029 2030/* 2031 * Somemtimes we have to undo a charge we got by try_charge(). 2032 * This function is for that and do uncharge, put css's refcnt. 2033 * gotten by try_charge(). 2034 */ 2035static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem, 2036 unsigned int nr_pages) 2037{ 2038 if (!mem_cgroup_is_root(mem)) { 2039 unsigned long bytes = nr_pages * PAGE_SIZE; 2040 2041 res_counter_uncharge(&mem->res, bytes); 2042 if (do_swap_account) 2043 res_counter_uncharge(&mem->memsw, bytes); 2044 } 2045} 2046 2047/* 2048 * A helper function to get mem_cgroup from ID. must be called under 2049 * rcu_read_lock(). The caller must check css_is_removed() or some if 2050 * it's concern. (dropping refcnt from swap can be called against removed 2051 * memcg.) 2052 */ 2053static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2054{ 2055 struct cgroup_subsys_state *css; 2056 2057 /* ID 0 is unused ID */ 2058 if (!id) 2059 return NULL; 2060 css = css_lookup(&mem_cgroup_subsys, id); 2061 if (!css) 2062 return NULL; 2063 return container_of(css, struct mem_cgroup, css); 2064} 2065 2066struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2067{ 2068 struct mem_cgroup *mem = NULL; 2069 struct page_cgroup *pc; 2070 unsigned short id; 2071 swp_entry_t ent; 2072 2073 VM_BUG_ON(!PageLocked(page)); 2074 2075 pc = lookup_page_cgroup(page); 2076 lock_page_cgroup(pc); 2077 if (PageCgroupUsed(pc)) { 2078 mem = pc->mem_cgroup; 2079 if (mem && !css_tryget(&mem->css)) 2080 mem = NULL; 2081 } else if (PageSwapCache(page)) { 2082 ent.val = page_private(page); 2083 id = lookup_swap_cgroup(ent); 2084 rcu_read_lock(); 2085 mem = mem_cgroup_lookup(id); 2086 if (mem && !css_tryget(&mem->css)) 2087 mem = NULL; 2088 rcu_read_unlock(); 2089 } 2090 unlock_page_cgroup(pc); 2091 return mem; 2092} 2093 2094static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, 2095 struct page *page, 2096 struct page_cgroup *pc, 2097 enum charge_type ctype, 2098 int page_size) 2099{ 2100 int nr_pages = page_size >> PAGE_SHIFT; 2101 2102 lock_page_cgroup(pc); 2103 if (unlikely(PageCgroupUsed(pc))) { 2104 unlock_page_cgroup(pc); 2105 __mem_cgroup_cancel_charge(mem, nr_pages); 2106 return; 2107 } 2108 /* 2109 * we don't need page_cgroup_lock about tail pages, becase they are not 2110 * accessed by any other context at this point. 2111 */ 2112 pc->mem_cgroup = mem; 2113 /* 2114 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2115 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2116 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2117 * before USED bit, we need memory barrier here. 2118 * See mem_cgroup_add_lru_list(), etc. 2119 */ 2120 smp_wmb(); 2121 switch (ctype) { 2122 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2123 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2124 SetPageCgroupCache(pc); 2125 SetPageCgroupUsed(pc); 2126 break; 2127 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2128 ClearPageCgroupCache(pc); 2129 SetPageCgroupUsed(pc); 2130 break; 2131 default: 2132 break; 2133 } 2134 2135 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages); 2136 unlock_page_cgroup(pc); 2137 /* 2138 * "charge_statistics" updated event counter. Then, check it. 2139 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2140 * if they exceeds softlimit. 2141 */ 2142 memcg_check_events(mem, page); 2143} 2144 2145#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2146 2147#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2148 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) 2149/* 2150 * Because tail pages are not marked as "used", set it. We're under 2151 * zone->lru_lock, 'splitting on pmd' and compund_lock. 2152 */ 2153void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail) 2154{ 2155 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2156 struct page_cgroup *tail_pc = lookup_page_cgroup(tail); 2157 unsigned long flags; 2158 2159 if (mem_cgroup_disabled()) 2160 return; 2161 /* 2162 * We have no races with charge/uncharge but will have races with 2163 * page state accounting. 2164 */ 2165 move_lock_page_cgroup(head_pc, &flags); 2166 2167 tail_pc->mem_cgroup = head_pc->mem_cgroup; 2168 smp_wmb(); /* see __commit_charge() */ 2169 if (PageCgroupAcctLRU(head_pc)) { 2170 enum lru_list lru; 2171 struct mem_cgroup_per_zone *mz; 2172 2173 /* 2174 * LRU flags cannot be copied because we need to add tail 2175 *.page to LRU by generic call and our hook will be called. 2176 * We hold lru_lock, then, reduce counter directly. 2177 */ 2178 lru = page_lru(head); 2179 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); 2180 MEM_CGROUP_ZSTAT(mz, lru) -= 1; 2181 } 2182 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2183 move_unlock_page_cgroup(head_pc, &flags); 2184} 2185#endif 2186 2187/** 2188 * mem_cgroup_move_account - move account of the page 2189 * @page: the page 2190 * @pc: page_cgroup of the page. 2191 * @from: mem_cgroup which the page is moved from. 2192 * @to: mem_cgroup which the page is moved to. @from != @to. 2193 * @uncharge: whether we should call uncharge and css_put against @from. 2194 * @charge_size: number of bytes to charge (regular or huge page) 2195 * 2196 * The caller must confirm following. 2197 * - page is not on LRU (isolate_page() is useful.) 2198 * - compound_lock is held when charge_size > PAGE_SIZE 2199 * 2200 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2201 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is 2202 * true, this function does "uncharge" from old cgroup, but it doesn't if 2203 * @uncharge is false, so a caller should do "uncharge". 2204 */ 2205static int mem_cgroup_move_account(struct page *page, struct page_cgroup *pc, 2206 struct mem_cgroup *from, struct mem_cgroup *to, 2207 bool uncharge, int charge_size) 2208{ 2209 int nr_pages = charge_size >> PAGE_SHIFT; 2210 unsigned long flags; 2211 int ret; 2212 2213 VM_BUG_ON(from == to); 2214 VM_BUG_ON(PageLRU(page)); 2215 /* 2216 * The page is isolated from LRU. So, collapse function 2217 * will not handle this page. But page splitting can happen. 2218 * Do this check under compound_page_lock(). The caller should 2219 * hold it. 2220 */ 2221 ret = -EBUSY; 2222 if (charge_size > PAGE_SIZE && !PageTransHuge(page)) 2223 goto out; 2224 2225 lock_page_cgroup(pc); 2226 2227 ret = -EINVAL; 2228 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2229 goto unlock; 2230 2231 move_lock_page_cgroup(pc, &flags); 2232 2233 if (PageCgroupFileMapped(pc)) { 2234 /* Update mapped_file data for mem_cgroup */ 2235 preempt_disable(); 2236 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2237 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2238 preempt_enable(); 2239 } 2240 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2241 if (uncharge) 2242 /* This is not "cancel", but cancel_charge does all we need. */ 2243 __mem_cgroup_cancel_charge(from, nr_pages); 2244 2245 /* caller should have done css_get */ 2246 pc->mem_cgroup = to; 2247 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2248 /* 2249 * We charges against "to" which may not have any tasks. Then, "to" 2250 * can be under rmdir(). But in current implementation, caller of 2251 * this function is just force_empty() and move charge, so it's 2252 * garanteed that "to" is never removed. So, we don't check rmdir 2253 * status here. 2254 */ 2255 move_unlock_page_cgroup(pc, &flags); 2256 ret = 0; 2257unlock: 2258 unlock_page_cgroup(pc); 2259 /* 2260 * check events 2261 */ 2262 memcg_check_events(to, page); 2263 memcg_check_events(from, page); 2264out: 2265 return ret; 2266} 2267 2268/* 2269 * move charges to its parent. 2270 */ 2271 2272static int mem_cgroup_move_parent(struct page *page, 2273 struct page_cgroup *pc, 2274 struct mem_cgroup *child, 2275 gfp_t gfp_mask) 2276{ 2277 struct cgroup *cg = child->css.cgroup; 2278 struct cgroup *pcg = cg->parent; 2279 struct mem_cgroup *parent; 2280 int page_size = PAGE_SIZE; 2281 unsigned long flags; 2282 int ret; 2283 2284 /* Is ROOT ? */ 2285 if (!pcg) 2286 return -EINVAL; 2287 2288 ret = -EBUSY; 2289 if (!get_page_unless_zero(page)) 2290 goto out; 2291 if (isolate_lru_page(page)) 2292 goto put; 2293 2294 if (PageTransHuge(page)) 2295 page_size = HPAGE_SIZE; 2296 2297 parent = mem_cgroup_from_cont(pcg); 2298 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 2299 &parent, false, page_size); 2300 if (ret || !parent) 2301 goto put_back; 2302 2303 if (page_size > PAGE_SIZE) 2304 flags = compound_lock_irqsave(page); 2305 2306 ret = mem_cgroup_move_account(page, pc, child, parent, true, page_size); 2307 if (ret) 2308 __mem_cgroup_cancel_charge(parent, page_size >> PAGE_SHIFT); 2309 2310 if (page_size > PAGE_SIZE) 2311 compound_unlock_irqrestore(page, flags); 2312put_back: 2313 putback_lru_page(page); 2314put: 2315 put_page(page); 2316out: 2317 return ret; 2318} 2319 2320/* 2321 * Charge the memory controller for page usage. 2322 * Return 2323 * 0 if the charge was successful 2324 * < 0 if the cgroup is over its limit 2325 */ 2326static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2327 gfp_t gfp_mask, enum charge_type ctype) 2328{ 2329 struct mem_cgroup *mem = NULL; 2330 int page_size = PAGE_SIZE; 2331 struct page_cgroup *pc; 2332 bool oom = true; 2333 int ret; 2334 2335 if (PageTransHuge(page)) { 2336 page_size <<= compound_order(page); 2337 VM_BUG_ON(!PageTransHuge(page)); 2338 /* 2339 * Never OOM-kill a process for a huge page. The 2340 * fault handler will fall back to regular pages. 2341 */ 2342 oom = false; 2343 } 2344 2345 pc = lookup_page_cgroup(page); 2346 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */ 2347 2348 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size); 2349 if (ret || !mem) 2350 return ret; 2351 2352 __mem_cgroup_commit_charge(mem, page, pc, ctype, page_size); 2353 return 0; 2354} 2355 2356int mem_cgroup_newpage_charge(struct page *page, 2357 struct mm_struct *mm, gfp_t gfp_mask) 2358{ 2359 if (mem_cgroup_disabled()) 2360 return 0; 2361 /* 2362 * If already mapped, we don't have to account. 2363 * If page cache, page->mapping has address_space. 2364 * But page->mapping may have out-of-use anon_vma pointer, 2365 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 2366 * is NULL. 2367 */ 2368 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 2369 return 0; 2370 if (unlikely(!mm)) 2371 mm = &init_mm; 2372 return mem_cgroup_charge_common(page, mm, gfp_mask, 2373 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2374} 2375 2376static void 2377__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2378 enum charge_type ctype); 2379 2380int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2381 gfp_t gfp_mask) 2382{ 2383 int ret; 2384 2385 if (mem_cgroup_disabled()) 2386 return 0; 2387 if (PageCompound(page)) 2388 return 0; 2389 /* 2390 * Corner case handling. This is called from add_to_page_cache() 2391 * in usual. But some FS (shmem) precharges this page before calling it 2392 * and call add_to_page_cache() with GFP_NOWAIT. 2393 * 2394 * For GFP_NOWAIT case, the page may be pre-charged before calling 2395 * add_to_page_cache(). (See shmem.c) check it here and avoid to call 2396 * charge twice. (It works but has to pay a bit larger cost.) 2397 * And when the page is SwapCache, it should take swap information 2398 * into account. This is under lock_page() now. 2399 */ 2400 if (!(gfp_mask & __GFP_WAIT)) { 2401 struct page_cgroup *pc; 2402 2403 pc = lookup_page_cgroup(page); 2404 if (!pc) 2405 return 0; 2406 lock_page_cgroup(pc); 2407 if (PageCgroupUsed(pc)) { 2408 unlock_page_cgroup(pc); 2409 return 0; 2410 } 2411 unlock_page_cgroup(pc); 2412 } 2413 2414 if (unlikely(!mm)) 2415 mm = &init_mm; 2416 2417 if (page_is_file_cache(page)) 2418 return mem_cgroup_charge_common(page, mm, gfp_mask, 2419 MEM_CGROUP_CHARGE_TYPE_CACHE); 2420 2421 /* shmem */ 2422 if (PageSwapCache(page)) { 2423 struct mem_cgroup *mem; 2424 2425 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 2426 if (!ret) 2427 __mem_cgroup_commit_charge_swapin(page, mem, 2428 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2429 } else 2430 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 2431 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2432 2433 return ret; 2434} 2435 2436/* 2437 * While swap-in, try_charge -> commit or cancel, the page is locked. 2438 * And when try_charge() successfully returns, one refcnt to memcg without 2439 * struct page_cgroup is acquired. This refcnt will be consumed by 2440 * "commit()" or removed by "cancel()" 2441 */ 2442int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2443 struct page *page, 2444 gfp_t mask, struct mem_cgroup **ptr) 2445{ 2446 struct mem_cgroup *mem; 2447 int ret; 2448 2449 *ptr = NULL; 2450 2451 if (mem_cgroup_disabled()) 2452 return 0; 2453 2454 if (!do_swap_account) 2455 goto charge_cur_mm; 2456 /* 2457 * A racing thread's fault, or swapoff, may have already updated 2458 * the pte, and even removed page from swap cache: in those cases 2459 * do_swap_page()'s pte_same() test will fail; but there's also a 2460 * KSM case which does need to charge the page. 2461 */ 2462 if (!PageSwapCache(page)) 2463 goto charge_cur_mm; 2464 mem = try_get_mem_cgroup_from_page(page); 2465 if (!mem) 2466 goto charge_cur_mm; 2467 *ptr = mem; 2468 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE); 2469 css_put(&mem->css); 2470 return ret; 2471charge_cur_mm: 2472 if (unlikely(!mm)) 2473 mm = &init_mm; 2474 return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE); 2475} 2476 2477static void 2478__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2479 enum charge_type ctype) 2480{ 2481 struct page_cgroup *pc; 2482 2483 if (mem_cgroup_disabled()) 2484 return; 2485 if (!ptr) 2486 return; 2487 cgroup_exclude_rmdir(&ptr->css); 2488 pc = lookup_page_cgroup(page); 2489 mem_cgroup_lru_del_before_commit_swapcache(page); 2490 __mem_cgroup_commit_charge(ptr, page, pc, ctype, PAGE_SIZE); 2491 mem_cgroup_lru_add_after_commit_swapcache(page); 2492 /* 2493 * Now swap is on-memory. This means this page may be 2494 * counted both as mem and swap....double count. 2495 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2496 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2497 * may call delete_from_swap_cache() before reach here. 2498 */ 2499 if (do_swap_account && PageSwapCache(page)) { 2500 swp_entry_t ent = {.val = page_private(page)}; 2501 unsigned short id; 2502 struct mem_cgroup *memcg; 2503 2504 id = swap_cgroup_record(ent, 0); 2505 rcu_read_lock(); 2506 memcg = mem_cgroup_lookup(id); 2507 if (memcg) { 2508 /* 2509 * This recorded memcg can be obsolete one. So, avoid 2510 * calling css_tryget 2511 */ 2512 if (!mem_cgroup_is_root(memcg)) 2513 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 2514 mem_cgroup_swap_statistics(memcg, false); 2515 mem_cgroup_put(memcg); 2516 } 2517 rcu_read_unlock(); 2518 } 2519 /* 2520 * At swapin, we may charge account against cgroup which has no tasks. 2521 * So, rmdir()->pre_destroy() can be called while we do this charge. 2522 * In that case, we need to call pre_destroy() again. check it here. 2523 */ 2524 cgroup_release_and_wakeup_rmdir(&ptr->css); 2525} 2526 2527void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) 2528{ 2529 __mem_cgroup_commit_charge_swapin(page, ptr, 2530 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2531} 2532 2533void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) 2534{ 2535 if (mem_cgroup_disabled()) 2536 return; 2537 if (!mem) 2538 return; 2539 __mem_cgroup_cancel_charge(mem, 1); 2540} 2541 2542static void 2543__do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype, 2544 int page_size) 2545{ 2546 struct memcg_batch_info *batch = NULL; 2547 bool uncharge_memsw = true; 2548 /* If swapout, usage of swap doesn't decrease */ 2549 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2550 uncharge_memsw = false; 2551 2552 batch = ¤t->memcg_batch; 2553 /* 2554 * In usual, we do css_get() when we remember memcg pointer. 2555 * But in this case, we keep res->usage until end of a series of 2556 * uncharges. Then, it's ok to ignore memcg's refcnt. 2557 */ 2558 if (!batch->memcg) 2559 batch->memcg = mem; 2560 /* 2561 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2562 * In those cases, all pages freed continously can be expected to be in 2563 * the same cgroup and we have chance to coalesce uncharges. 2564 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2565 * because we want to do uncharge as soon as possible. 2566 */ 2567 2568 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2569 goto direct_uncharge; 2570 2571 if (page_size != PAGE_SIZE) 2572 goto direct_uncharge; 2573 2574 /* 2575 * In typical case, batch->memcg == mem. This means we can 2576 * merge a series of uncharges to an uncharge of res_counter. 2577 * If not, we uncharge res_counter ony by one. 2578 */ 2579 if (batch->memcg != mem) 2580 goto direct_uncharge; 2581 /* remember freed charge and uncharge it later */ 2582 batch->nr_pages++; 2583 if (uncharge_memsw) 2584 batch->memsw_nr_pages++; 2585 return; 2586direct_uncharge: 2587 res_counter_uncharge(&mem->res, page_size); 2588 if (uncharge_memsw) 2589 res_counter_uncharge(&mem->memsw, page_size); 2590 if (unlikely(batch->memcg != mem)) 2591 memcg_oom_recover(mem); 2592 return; 2593} 2594 2595/* 2596 * uncharge if !page_mapped(page) 2597 */ 2598static struct mem_cgroup * 2599__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 2600{ 2601 int count; 2602 struct page_cgroup *pc; 2603 struct mem_cgroup *mem = NULL; 2604 int page_size = PAGE_SIZE; 2605 2606 if (mem_cgroup_disabled()) 2607 return NULL; 2608 2609 if (PageSwapCache(page)) 2610 return NULL; 2611 2612 if (PageTransHuge(page)) { 2613 page_size <<= compound_order(page); 2614 VM_BUG_ON(!PageTransHuge(page)); 2615 } 2616 2617 count = page_size >> PAGE_SHIFT; 2618 /* 2619 * Check if our page_cgroup is valid 2620 */ 2621 pc = lookup_page_cgroup(page); 2622 if (unlikely(!pc || !PageCgroupUsed(pc))) 2623 return NULL; 2624 2625 lock_page_cgroup(pc); 2626 2627 mem = pc->mem_cgroup; 2628 2629 if (!PageCgroupUsed(pc)) 2630 goto unlock_out; 2631 2632 switch (ctype) { 2633 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2634 case MEM_CGROUP_CHARGE_TYPE_DROP: 2635 /* See mem_cgroup_prepare_migration() */ 2636 if (page_mapped(page) || PageCgroupMigration(pc)) 2637 goto unlock_out; 2638 break; 2639 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 2640 if (!PageAnon(page)) { /* Shared memory */ 2641 if (page->mapping && !page_is_file_cache(page)) 2642 goto unlock_out; 2643 } else if (page_mapped(page)) /* Anon */ 2644 goto unlock_out; 2645 break; 2646 default: 2647 break; 2648 } 2649 2650 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count); 2651 2652 ClearPageCgroupUsed(pc); 2653 /* 2654 * pc->mem_cgroup is not cleared here. It will be accessed when it's 2655 * freed from LRU. This is safe because uncharged page is expected not 2656 * to be reused (freed soon). Exception is SwapCache, it's handled by 2657 * special functions. 2658 */ 2659 2660 unlock_page_cgroup(pc); 2661 /* 2662 * even after unlock, we have mem->res.usage here and this memcg 2663 * will never be freed. 2664 */ 2665 memcg_check_events(mem, page); 2666 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 2667 mem_cgroup_swap_statistics(mem, true); 2668 mem_cgroup_get(mem); 2669 } 2670 if (!mem_cgroup_is_root(mem)) 2671 __do_uncharge(mem, ctype, page_size); 2672 2673 return mem; 2674 2675unlock_out: 2676 unlock_page_cgroup(pc); 2677 return NULL; 2678} 2679 2680void mem_cgroup_uncharge_page(struct page *page) 2681{ 2682 /* early check. */ 2683 if (page_mapped(page)) 2684 return; 2685 if (page->mapping && !PageAnon(page)) 2686 return; 2687 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 2688} 2689 2690void mem_cgroup_uncharge_cache_page(struct page *page) 2691{ 2692 VM_BUG_ON(page_mapped(page)); 2693 VM_BUG_ON(page->mapping); 2694 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 2695} 2696 2697/* 2698 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 2699 * In that cases, pages are freed continuously and we can expect pages 2700 * are in the same memcg. All these calls itself limits the number of 2701 * pages freed at once, then uncharge_start/end() is called properly. 2702 * This may be called prural(2) times in a context, 2703 */ 2704 2705void mem_cgroup_uncharge_start(void) 2706{ 2707 current->memcg_batch.do_batch++; 2708 /* We can do nest. */ 2709 if (current->memcg_batch.do_batch == 1) { 2710 current->memcg_batch.memcg = NULL; 2711 current->memcg_batch.nr_pages = 0; 2712 current->memcg_batch.memsw_nr_pages = 0; 2713 } 2714} 2715 2716void mem_cgroup_uncharge_end(void) 2717{ 2718 struct memcg_batch_info *batch = ¤t->memcg_batch; 2719 2720 if (!batch->do_batch) 2721 return; 2722 2723 batch->do_batch--; 2724 if (batch->do_batch) /* If stacked, do nothing. */ 2725 return; 2726 2727 if (!batch->memcg) 2728 return; 2729 /* 2730 * This "batch->memcg" is valid without any css_get/put etc... 2731 * bacause we hide charges behind us. 2732 */ 2733 if (batch->nr_pages) 2734 res_counter_uncharge(&batch->memcg->res, 2735 batch->nr_pages * PAGE_SIZE); 2736 if (batch->memsw_nr_pages) 2737 res_counter_uncharge(&batch->memcg->memsw, 2738 batch->memsw_nr_pages * PAGE_SIZE); 2739 memcg_oom_recover(batch->memcg); 2740 /* forget this pointer (for sanity check) */ 2741 batch->memcg = NULL; 2742} 2743 2744#ifdef CONFIG_SWAP 2745/* 2746 * called after __delete_from_swap_cache() and drop "page" account. 2747 * memcg information is recorded to swap_cgroup of "ent" 2748 */ 2749void 2750mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 2751{ 2752 struct mem_cgroup *memcg; 2753 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 2754 2755 if (!swapout) /* this was a swap cache but the swap is unused ! */ 2756 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 2757 2758 memcg = __mem_cgroup_uncharge_common(page, ctype); 2759 2760 /* 2761 * record memcg information, if swapout && memcg != NULL, 2762 * mem_cgroup_get() was called in uncharge(). 2763 */ 2764 if (do_swap_account && swapout && memcg) 2765 swap_cgroup_record(ent, css_id(&memcg->css)); 2766} 2767#endif 2768 2769#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2770/* 2771 * called from swap_entry_free(). remove record in swap_cgroup and 2772 * uncharge "memsw" account. 2773 */ 2774void mem_cgroup_uncharge_swap(swp_entry_t ent) 2775{ 2776 struct mem_cgroup *memcg; 2777 unsigned short id; 2778 2779 if (!do_swap_account) 2780 return; 2781 2782 id = swap_cgroup_record(ent, 0); 2783 rcu_read_lock(); 2784 memcg = mem_cgroup_lookup(id); 2785 if (memcg) { 2786 /* 2787 * We uncharge this because swap is freed. 2788 * This memcg can be obsolete one. We avoid calling css_tryget 2789 */ 2790 if (!mem_cgroup_is_root(memcg)) 2791 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 2792 mem_cgroup_swap_statistics(memcg, false); 2793 mem_cgroup_put(memcg); 2794 } 2795 rcu_read_unlock(); 2796} 2797 2798/** 2799 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 2800 * @entry: swap entry to be moved 2801 * @from: mem_cgroup which the entry is moved from 2802 * @to: mem_cgroup which the entry is moved to 2803 * @need_fixup: whether we should fixup res_counters and refcounts. 2804 * 2805 * It succeeds only when the swap_cgroup's record for this entry is the same 2806 * as the mem_cgroup's id of @from. 2807 * 2808 * Returns 0 on success, -EINVAL on failure. 2809 * 2810 * The caller must have charged to @to, IOW, called res_counter_charge() about 2811 * both res and memsw, and called css_get(). 2812 */ 2813static int mem_cgroup_move_swap_account(swp_entry_t entry, 2814 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 2815{ 2816 unsigned short old_id, new_id; 2817 2818 old_id = css_id(&from->css); 2819 new_id = css_id(&to->css); 2820 2821 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 2822 mem_cgroup_swap_statistics(from, false); 2823 mem_cgroup_swap_statistics(to, true); 2824 /* 2825 * This function is only called from task migration context now. 2826 * It postpones res_counter and refcount handling till the end 2827 * of task migration(mem_cgroup_clear_mc()) for performance 2828 * improvement. But we cannot postpone mem_cgroup_get(to) 2829 * because if the process that has been moved to @to does 2830 * swap-in, the refcount of @to might be decreased to 0. 2831 */ 2832 mem_cgroup_get(to); 2833 if (need_fixup) { 2834 if (!mem_cgroup_is_root(from)) 2835 res_counter_uncharge(&from->memsw, PAGE_SIZE); 2836 mem_cgroup_put(from); 2837 /* 2838 * we charged both to->res and to->memsw, so we should 2839 * uncharge to->res. 2840 */ 2841 if (!mem_cgroup_is_root(to)) 2842 res_counter_uncharge(&to->res, PAGE_SIZE); 2843 } 2844 return 0; 2845 } 2846 return -EINVAL; 2847} 2848#else 2849static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 2850 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 2851{ 2852 return -EINVAL; 2853} 2854#endif 2855 2856/* 2857 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 2858 * page belongs to. 2859 */ 2860int mem_cgroup_prepare_migration(struct page *page, 2861 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask) 2862{ 2863 struct page_cgroup *pc; 2864 struct mem_cgroup *mem = NULL; 2865 enum charge_type ctype; 2866 int ret = 0; 2867 2868 *ptr = NULL; 2869 2870 VM_BUG_ON(PageTransHuge(page)); 2871 if (mem_cgroup_disabled()) 2872 return 0; 2873 2874 pc = lookup_page_cgroup(page); 2875 lock_page_cgroup(pc); 2876 if (PageCgroupUsed(pc)) { 2877 mem = pc->mem_cgroup; 2878 css_get(&mem->css); 2879 /* 2880 * At migrating an anonymous page, its mapcount goes down 2881 * to 0 and uncharge() will be called. But, even if it's fully 2882 * unmapped, migration may fail and this page has to be 2883 * charged again. We set MIGRATION flag here and delay uncharge 2884 * until end_migration() is called 2885 * 2886 * Corner Case Thinking 2887 * A) 2888 * When the old page was mapped as Anon and it's unmap-and-freed 2889 * while migration was ongoing. 2890 * If unmap finds the old page, uncharge() of it will be delayed 2891 * until end_migration(). If unmap finds a new page, it's 2892 * uncharged when it make mapcount to be 1->0. If unmap code 2893 * finds swap_migration_entry, the new page will not be mapped 2894 * and end_migration() will find it(mapcount==0). 2895 * 2896 * B) 2897 * When the old page was mapped but migraion fails, the kernel 2898 * remaps it. A charge for it is kept by MIGRATION flag even 2899 * if mapcount goes down to 0. We can do remap successfully 2900 * without charging it again. 2901 * 2902 * C) 2903 * The "old" page is under lock_page() until the end of 2904 * migration, so, the old page itself will not be swapped-out. 2905 * If the new page is swapped out before end_migraton, our 2906 * hook to usual swap-out path will catch the event. 2907 */ 2908 if (PageAnon(page)) 2909 SetPageCgroupMigration(pc); 2910 } 2911 unlock_page_cgroup(pc); 2912 /* 2913 * If the page is not charged at this point, 2914 * we return here. 2915 */ 2916 if (!mem) 2917 return 0; 2918 2919 *ptr = mem; 2920 ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE); 2921 css_put(&mem->css);/* drop extra refcnt */ 2922 if (ret || *ptr == NULL) { 2923 if (PageAnon(page)) { 2924 lock_page_cgroup(pc); 2925 ClearPageCgroupMigration(pc); 2926 unlock_page_cgroup(pc); 2927 /* 2928 * The old page may be fully unmapped while we kept it. 2929 */ 2930 mem_cgroup_uncharge_page(page); 2931 } 2932 return -ENOMEM; 2933 } 2934 /* 2935 * We charge new page before it's used/mapped. So, even if unlock_page() 2936 * is called before end_migration, we can catch all events on this new 2937 * page. In the case new page is migrated but not remapped, new page's 2938 * mapcount will be finally 0 and we call uncharge in end_migration(). 2939 */ 2940 pc = lookup_page_cgroup(newpage); 2941 if (PageAnon(page)) 2942 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 2943 else if (page_is_file_cache(page)) 2944 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 2945 else 2946 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 2947 __mem_cgroup_commit_charge(mem, page, pc, ctype, PAGE_SIZE); 2948 return ret; 2949} 2950 2951/* remove redundant charge if migration failed*/ 2952void mem_cgroup_end_migration(struct mem_cgroup *mem, 2953 struct page *oldpage, struct page *newpage, bool migration_ok) 2954{ 2955 struct page *used, *unused; 2956 struct page_cgroup *pc; 2957 2958 if (!mem) 2959 return; 2960 /* blocks rmdir() */ 2961 cgroup_exclude_rmdir(&mem->css); 2962 if (!migration_ok) { 2963 used = oldpage; 2964 unused = newpage; 2965 } else { 2966 used = newpage; 2967 unused = oldpage; 2968 } 2969 /* 2970 * We disallowed uncharge of pages under migration because mapcount 2971 * of the page goes down to zero, temporarly. 2972 * Clear the flag and check the page should be charged. 2973 */ 2974 pc = lookup_page_cgroup(oldpage); 2975 lock_page_cgroup(pc); 2976 ClearPageCgroupMigration(pc); 2977 unlock_page_cgroup(pc); 2978 2979 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 2980 2981 /* 2982 * If a page is a file cache, radix-tree replacement is very atomic 2983 * and we can skip this check. When it was an Anon page, its mapcount 2984 * goes down to 0. But because we added MIGRATION flage, it's not 2985 * uncharged yet. There are several case but page->mapcount check 2986 * and USED bit check in mem_cgroup_uncharge_page() will do enough 2987 * check. (see prepare_charge() also) 2988 */ 2989 if (PageAnon(used)) 2990 mem_cgroup_uncharge_page(used); 2991 /* 2992 * At migration, we may charge account against cgroup which has no 2993 * tasks. 2994 * So, rmdir()->pre_destroy() can be called while we do this charge. 2995 * In that case, we need to call pre_destroy() again. check it here. 2996 */ 2997 cgroup_release_and_wakeup_rmdir(&mem->css); 2998} 2999 3000/* 3001 * A call to try to shrink memory usage on charge failure at shmem's swapin. 3002 * Calling hierarchical_reclaim is not enough because we should update 3003 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. 3004 * Moreover considering hierarchy, we should reclaim from the mem_over_limit, 3005 * not from the memcg which this page would be charged to. 3006 * try_charge_swapin does all of these works properly. 3007 */ 3008int mem_cgroup_shmem_charge_fallback(struct page *page, 3009 struct mm_struct *mm, 3010 gfp_t gfp_mask) 3011{ 3012 struct mem_cgroup *mem; 3013 int ret; 3014 3015 if (mem_cgroup_disabled()) 3016 return 0; 3017 3018 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 3019 if (!ret) 3020 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ 3021 3022 return ret; 3023} 3024 3025#ifdef CONFIG_DEBUG_VM 3026static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3027{ 3028 struct page_cgroup *pc; 3029 3030 pc = lookup_page_cgroup(page); 3031 if (likely(pc) && PageCgroupUsed(pc)) 3032 return pc; 3033 return NULL; 3034} 3035 3036bool mem_cgroup_bad_page_check(struct page *page) 3037{ 3038 if (mem_cgroup_disabled()) 3039 return false; 3040 3041 return lookup_page_cgroup_used(page) != NULL; 3042} 3043 3044void mem_cgroup_print_bad_page(struct page *page) 3045{ 3046 struct page_cgroup *pc; 3047 3048 pc = lookup_page_cgroup_used(page); 3049 if (pc) { 3050 int ret = -1; 3051 char *path; 3052 3053 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", 3054 pc, pc->flags, pc->mem_cgroup); 3055 3056 path = kmalloc(PATH_MAX, GFP_KERNEL); 3057 if (path) { 3058 rcu_read_lock(); 3059 ret = cgroup_path(pc->mem_cgroup->css.cgroup, 3060 path, PATH_MAX); 3061 rcu_read_unlock(); 3062 } 3063 3064 printk(KERN_CONT "(%s)\n", 3065 (ret < 0) ? "cannot get the path" : path); 3066 kfree(path); 3067 } 3068} 3069#endif 3070 3071static DEFINE_MUTEX(set_limit_mutex); 3072 3073static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3074 unsigned long long val) 3075{ 3076 int retry_count; 3077 u64 memswlimit, memlimit; 3078 int ret = 0; 3079 int children = mem_cgroup_count_children(memcg); 3080 u64 curusage, oldusage; 3081 int enlarge; 3082 3083 /* 3084 * For keeping hierarchical_reclaim simple, how long we should retry 3085 * is depends on callers. We set our retry-count to be function 3086 * of # of children which we should visit in this loop. 3087 */ 3088 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3089 3090 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3091 3092 enlarge = 0; 3093 while (retry_count) { 3094 if (signal_pending(current)) { 3095 ret = -EINTR; 3096 break; 3097 } 3098 /* 3099 * Rather than hide all in some function, I do this in 3100 * open coded manner. You see what this really does. 3101 * We have to guarantee mem->res.limit < mem->memsw.limit. 3102 */ 3103 mutex_lock(&set_limit_mutex); 3104 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3105 if (memswlimit < val) { 3106 ret = -EINVAL; 3107 mutex_unlock(&set_limit_mutex); 3108 break; 3109 } 3110 3111 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3112 if (memlimit < val) 3113 enlarge = 1; 3114 3115 ret = res_counter_set_limit(&memcg->res, val); 3116 if (!ret) { 3117 if (memswlimit == val) 3118 memcg->memsw_is_minimum = true; 3119 else 3120 memcg->memsw_is_minimum = false; 3121 } 3122 mutex_unlock(&set_limit_mutex); 3123 3124 if (!ret) 3125 break; 3126 3127 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, 3128 MEM_CGROUP_RECLAIM_SHRINK); 3129 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3130 /* Usage is reduced ? */ 3131 if (curusage >= oldusage) 3132 retry_count--; 3133 else 3134 oldusage = curusage; 3135 } 3136 if (!ret && enlarge) 3137 memcg_oom_recover(memcg); 3138 3139 return ret; 3140} 3141 3142static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3143 unsigned long long val) 3144{ 3145 int retry_count; 3146 u64 memlimit, memswlimit, oldusage, curusage; 3147 int children = mem_cgroup_count_children(memcg); 3148 int ret = -EBUSY; 3149 int enlarge = 0; 3150 3151 /* see mem_cgroup_resize_res_limit */ 3152 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3153 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3154 while (retry_count) { 3155 if (signal_pending(current)) { 3156 ret = -EINTR; 3157 break; 3158 } 3159 /* 3160 * Rather than hide all in some function, I do this in 3161 * open coded manner. You see what this really does. 3162 * We have to guarantee mem->res.limit < mem->memsw.limit. 3163 */ 3164 mutex_lock(&set_limit_mutex); 3165 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3166 if (memlimit > val) { 3167 ret = -EINVAL; 3168 mutex_unlock(&set_limit_mutex); 3169 break; 3170 } 3171 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3172 if (memswlimit < val) 3173 enlarge = 1; 3174 ret = res_counter_set_limit(&memcg->memsw, val); 3175 if (!ret) { 3176 if (memlimit == val) 3177 memcg->memsw_is_minimum = true; 3178 else 3179 memcg->memsw_is_minimum = false; 3180 } 3181 mutex_unlock(&set_limit_mutex); 3182 3183 if (!ret) 3184 break; 3185 3186 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, 3187 MEM_CGROUP_RECLAIM_NOSWAP | 3188 MEM_CGROUP_RECLAIM_SHRINK); 3189 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3190 /* Usage is reduced ? */ 3191 if (curusage >= oldusage) 3192 retry_count--; 3193 else 3194 oldusage = curusage; 3195 } 3196 if (!ret && enlarge) 3197 memcg_oom_recover(memcg); 3198 return ret; 3199} 3200 3201unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3202 gfp_t gfp_mask) 3203{ 3204 unsigned long nr_reclaimed = 0; 3205 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3206 unsigned long reclaimed; 3207 int loop = 0; 3208 struct mem_cgroup_tree_per_zone *mctz; 3209 unsigned long long excess; 3210 3211 if (order > 0) 3212 return 0; 3213 3214 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3215 /* 3216 * This loop can run a while, specially if mem_cgroup's continuously 3217 * keep exceeding their soft limit and putting the system under 3218 * pressure 3219 */ 3220 do { 3221 if (next_mz) 3222 mz = next_mz; 3223 else 3224 mz = mem_cgroup_largest_soft_limit_node(mctz); 3225 if (!mz) 3226 break; 3227 3228 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone, 3229 gfp_mask, 3230 MEM_CGROUP_RECLAIM_SOFT); 3231 nr_reclaimed += reclaimed; 3232 spin_lock(&mctz->lock); 3233 3234 /* 3235 * If we failed to reclaim anything from this memory cgroup 3236 * it is time to move on to the next cgroup 3237 */ 3238 next_mz = NULL; 3239 if (!reclaimed) { 3240 do { 3241 /* 3242 * Loop until we find yet another one. 3243 * 3244 * By the time we get the soft_limit lock 3245 * again, someone might have aded the 3246 * group back on the RB tree. Iterate to 3247 * make sure we get a different mem. 3248 * mem_cgroup_largest_soft_limit_node returns 3249 * NULL if no other cgroup is present on 3250 * the tree 3251 */ 3252 next_mz = 3253 __mem_cgroup_largest_soft_limit_node(mctz); 3254 if (next_mz == mz) { 3255 css_put(&next_mz->mem->css); 3256 next_mz = NULL; 3257 } else /* next_mz == NULL or other memcg */ 3258 break; 3259 } while (1); 3260 } 3261 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3262 excess = res_counter_soft_limit_excess(&mz->mem->res); 3263 /* 3264 * One school of thought says that we should not add 3265 * back the node to the tree if reclaim returns 0. 3266 * But our reclaim could return 0, simply because due 3267 * to priority we are exposing a smaller subset of 3268 * memory to reclaim from. Consider this as a longer 3269 * term TODO. 3270 */ 3271 /* If excess == 0, no tree ops */ 3272 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3273 spin_unlock(&mctz->lock); 3274 css_put(&mz->mem->css); 3275 loop++; 3276 /* 3277 * Could not reclaim anything and there are no more 3278 * mem cgroups to try or we seem to be looping without 3279 * reclaiming anything. 3280 */ 3281 if (!nr_reclaimed && 3282 (next_mz == NULL || 3283 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3284 break; 3285 } while (!nr_reclaimed); 3286 if (next_mz) 3287 css_put(&next_mz->mem->css); 3288 return nr_reclaimed; 3289} 3290 3291/* 3292 * This routine traverse page_cgroup in given list and drop them all. 3293 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3294 */ 3295static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, 3296 int node, int zid, enum lru_list lru) 3297{ 3298 struct zone *zone; 3299 struct mem_cgroup_per_zone *mz; 3300 struct page_cgroup *pc, *busy; 3301 unsigned long flags, loop; 3302 struct list_head *list; 3303 int ret = 0; 3304 3305 zone = &NODE_DATA(node)->node_zones[zid]; 3306 mz = mem_cgroup_zoneinfo(mem, node, zid); 3307 list = &mz->lists[lru]; 3308 3309 loop = MEM_CGROUP_ZSTAT(mz, lru); 3310 /* give some margin against EBUSY etc...*/ 3311 loop += 256; 3312 busy = NULL; 3313 while (loop--) { 3314 struct page *page; 3315 3316 ret = 0; 3317 spin_lock_irqsave(&zone->lru_lock, flags); 3318 if (list_empty(list)) { 3319 spin_unlock_irqrestore(&zone->lru_lock, flags); 3320 break; 3321 } 3322 pc = list_entry(list->prev, struct page_cgroup, lru); 3323 if (busy == pc) { 3324 list_move(&pc->lru, list); 3325 busy = NULL; 3326 spin_unlock_irqrestore(&zone->lru_lock, flags); 3327 continue; 3328 } 3329 spin_unlock_irqrestore(&zone->lru_lock, flags); 3330 3331 page = lookup_cgroup_page(pc); 3332 3333 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL); 3334 if (ret == -ENOMEM) 3335 break; 3336 3337 if (ret == -EBUSY || ret == -EINVAL) { 3338 /* found lock contention or "pc" is obsolete. */ 3339 busy = pc; 3340 cond_resched(); 3341 } else 3342 busy = NULL; 3343 } 3344 3345 if (!ret && !list_empty(list)) 3346 return -EBUSY; 3347 return ret; 3348} 3349 3350/* 3351 * make mem_cgroup's charge to be 0 if there is no task. 3352 * This enables deleting this mem_cgroup. 3353 */ 3354static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) 3355{ 3356 int ret; 3357 int node, zid, shrink; 3358 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3359 struct cgroup *cgrp = mem->css.cgroup; 3360 3361 css_get(&mem->css); 3362 3363 shrink = 0; 3364 /* should free all ? */ 3365 if (free_all) 3366 goto try_to_free; 3367move_account: 3368 do { 3369 ret = -EBUSY; 3370 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3371 goto out; 3372 ret = -EINTR; 3373 if (signal_pending(current)) 3374 goto out; 3375 /* This is for making all *used* pages to be on LRU. */ 3376 lru_add_drain_all(); 3377 drain_all_stock_sync(); 3378 ret = 0; 3379 mem_cgroup_start_move(mem); 3380 for_each_node_state(node, N_HIGH_MEMORY) { 3381 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3382 enum lru_list l; 3383 for_each_lru(l) { 3384 ret = mem_cgroup_force_empty_list(mem, 3385 node, zid, l); 3386 if (ret) 3387 break; 3388 } 3389 } 3390 if (ret) 3391 break; 3392 } 3393 mem_cgroup_end_move(mem); 3394 memcg_oom_recover(mem); 3395 /* it seems parent cgroup doesn't have enough mem */ 3396 if (ret == -ENOMEM) 3397 goto try_to_free; 3398 cond_resched(); 3399 /* "ret" should also be checked to ensure all lists are empty. */ 3400 } while (mem->res.usage > 0 || ret); 3401out: 3402 css_put(&mem->css); 3403 return ret; 3404 3405try_to_free: 3406 /* returns EBUSY if there is a task or if we come here twice. */ 3407 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3408 ret = -EBUSY; 3409 goto out; 3410 } 3411 /* we call try-to-free pages for make this cgroup empty */ 3412 lru_add_drain_all(); 3413 /* try to free all pages in this cgroup */ 3414 shrink = 1; 3415 while (nr_retries && mem->res.usage > 0) { 3416 int progress; 3417 3418 if (signal_pending(current)) { 3419 ret = -EINTR; 3420 goto out; 3421 } 3422 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, 3423 false, get_swappiness(mem)); 3424 if (!progress) { 3425 nr_retries--; 3426 /* maybe some writeback is necessary */ 3427 congestion_wait(BLK_RW_ASYNC, HZ/10); 3428 } 3429 3430 } 3431 lru_add_drain(); 3432 /* try move_account...there may be some *locked* pages. */ 3433 goto move_account; 3434} 3435 3436int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3437{ 3438 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3439} 3440 3441 3442static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3443{ 3444 return mem_cgroup_from_cont(cont)->use_hierarchy; 3445} 3446 3447static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3448 u64 val) 3449{ 3450 int retval = 0; 3451 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 3452 struct cgroup *parent = cont->parent; 3453 struct mem_cgroup *parent_mem = NULL; 3454 3455 if (parent) 3456 parent_mem = mem_cgroup_from_cont(parent); 3457 3458 cgroup_lock(); 3459 /* 3460 * If parent's use_hierarchy is set, we can't make any modifications 3461 * in the child subtrees. If it is unset, then the change can 3462 * occur, provided the current cgroup has no children. 3463 * 3464 * For the root cgroup, parent_mem is NULL, we allow value to be 3465 * set if there are no children. 3466 */ 3467 if ((!parent_mem || !parent_mem->use_hierarchy) && 3468 (val == 1 || val == 0)) { 3469 if (list_empty(&cont->children)) 3470 mem->use_hierarchy = val; 3471 else 3472 retval = -EBUSY; 3473 } else 3474 retval = -EINVAL; 3475 cgroup_unlock(); 3476 3477 return retval; 3478} 3479 3480 3481static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem, 3482 enum mem_cgroup_stat_index idx) 3483{ 3484 struct mem_cgroup *iter; 3485 s64 val = 0; 3486 3487 /* each per cpu's value can be minus.Then, use s64 */ 3488 for_each_mem_cgroup_tree(iter, mem) 3489 val += mem_cgroup_read_stat(iter, idx); 3490 3491 if (val < 0) /* race ? */ 3492 val = 0; 3493 return val; 3494} 3495 3496static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap) 3497{ 3498 u64 val; 3499 3500 if (!mem_cgroup_is_root(mem)) { 3501 if (!swap) 3502 return res_counter_read_u64(&mem->res, RES_USAGE); 3503 else 3504 return res_counter_read_u64(&mem->memsw, RES_USAGE); 3505 } 3506 3507 val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE); 3508 val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS); 3509 3510 if (swap) 3511 val += mem_cgroup_get_recursive_idx_stat(mem, 3512 MEM_CGROUP_STAT_SWAPOUT); 3513 3514 return val << PAGE_SHIFT; 3515} 3516 3517static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3518{ 3519 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 3520 u64 val; 3521 int type, name; 3522 3523 type = MEMFILE_TYPE(cft->private); 3524 name = MEMFILE_ATTR(cft->private); 3525 switch (type) { 3526 case _MEM: 3527 if (name == RES_USAGE) 3528 val = mem_cgroup_usage(mem, false); 3529 else 3530 val = res_counter_read_u64(&mem->res, name); 3531 break; 3532 case _MEMSWAP: 3533 if (name == RES_USAGE) 3534 val = mem_cgroup_usage(mem, true); 3535 else 3536 val = res_counter_read_u64(&mem->memsw, name); 3537 break; 3538 default: 3539 BUG(); 3540 break; 3541 } 3542 return val; 3543} 3544/* 3545 * The user of this function is... 3546 * RES_LIMIT. 3547 */ 3548static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3549 const char *buffer) 3550{ 3551 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3552 int type, name; 3553 unsigned long long val; 3554 int ret; 3555 3556 type = MEMFILE_TYPE(cft->private); 3557 name = MEMFILE_ATTR(cft->private); 3558 switch (name) { 3559 case RES_LIMIT: 3560 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3561 ret = -EINVAL; 3562 break; 3563 } 3564 /* This function does all necessary parse...reuse it */ 3565 ret = res_counter_memparse_write_strategy(buffer, &val); 3566 if (ret) 3567 break; 3568 if (type == _MEM) 3569 ret = mem_cgroup_resize_limit(memcg, val); 3570 else 3571 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3572 break; 3573 case RES_SOFT_LIMIT: 3574 ret = res_counter_memparse_write_strategy(buffer, &val); 3575 if (ret) 3576 break; 3577 /* 3578 * For memsw, soft limits are hard to implement in terms 3579 * of semantics, for now, we support soft limits for 3580 * control without swap 3581 */ 3582 if (type == _MEM) 3583 ret = res_counter_set_soft_limit(&memcg->res, val); 3584 else 3585 ret = -EINVAL; 3586 break; 3587 default: 3588 ret = -EINVAL; /* should be BUG() ? */ 3589 break; 3590 } 3591 return ret; 3592} 3593 3594static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 3595 unsigned long long *mem_limit, unsigned long long *memsw_limit) 3596{ 3597 struct cgroup *cgroup; 3598 unsigned long long min_limit, min_memsw_limit, tmp; 3599 3600 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3601 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3602 cgroup = memcg->css.cgroup; 3603 if (!memcg->use_hierarchy) 3604 goto out; 3605 3606 while (cgroup->parent) { 3607 cgroup = cgroup->parent; 3608 memcg = mem_cgroup_from_cont(cgroup); 3609 if (!memcg->use_hierarchy) 3610 break; 3611 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 3612 min_limit = min(min_limit, tmp); 3613 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3614 min_memsw_limit = min(min_memsw_limit, tmp); 3615 } 3616out: 3617 *mem_limit = min_limit; 3618 *memsw_limit = min_memsw_limit; 3619 return; 3620} 3621 3622static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 3623{ 3624 struct mem_cgroup *mem; 3625 int type, name; 3626 3627 mem = mem_cgroup_from_cont(cont); 3628 type = MEMFILE_TYPE(event); 3629 name = MEMFILE_ATTR(event); 3630 switch (name) { 3631 case RES_MAX_USAGE: 3632 if (type == _MEM) 3633 res_counter_reset_max(&mem->res); 3634 else 3635 res_counter_reset_max(&mem->memsw); 3636 break; 3637 case RES_FAILCNT: 3638 if (type == _MEM) 3639 res_counter_reset_failcnt(&mem->res); 3640 else 3641 res_counter_reset_failcnt(&mem->memsw); 3642 break; 3643 } 3644 3645 return 0; 3646} 3647 3648static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 3649 struct cftype *cft) 3650{ 3651 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 3652} 3653 3654#ifdef CONFIG_MMU 3655static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3656 struct cftype *cft, u64 val) 3657{ 3658 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 3659 3660 if (val >= (1 << NR_MOVE_TYPE)) 3661 return -EINVAL; 3662 /* 3663 * We check this value several times in both in can_attach() and 3664 * attach(), so we need cgroup lock to prevent this value from being 3665 * inconsistent. 3666 */ 3667 cgroup_lock(); 3668 mem->move_charge_at_immigrate = val; 3669 cgroup_unlock(); 3670 3671 return 0; 3672} 3673#else 3674static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3675 struct cftype *cft, u64 val) 3676{ 3677 return -ENOSYS; 3678} 3679#endif 3680 3681 3682/* For read statistics */ 3683enum { 3684 MCS_CACHE, 3685 MCS_RSS, 3686 MCS_FILE_MAPPED, 3687 MCS_PGPGIN, 3688 MCS_PGPGOUT, 3689 MCS_SWAP, 3690 MCS_INACTIVE_ANON, 3691 MCS_ACTIVE_ANON, 3692 MCS_INACTIVE_FILE, 3693 MCS_ACTIVE_FILE, 3694 MCS_UNEVICTABLE, 3695 NR_MCS_STAT, 3696}; 3697 3698struct mcs_total_stat { 3699 s64 stat[NR_MCS_STAT]; 3700}; 3701 3702struct { 3703 char *local_name; 3704 char *total_name; 3705} memcg_stat_strings[NR_MCS_STAT] = { 3706 {"cache", "total_cache"}, 3707 {"rss", "total_rss"}, 3708 {"mapped_file", "total_mapped_file"}, 3709 {"pgpgin", "total_pgpgin"}, 3710 {"pgpgout", "total_pgpgout"}, 3711 {"swap", "total_swap"}, 3712 {"inactive_anon", "total_inactive_anon"}, 3713 {"active_anon", "total_active_anon"}, 3714 {"inactive_file", "total_inactive_file"}, 3715 {"active_file", "total_active_file"}, 3716 {"unevictable", "total_unevictable"} 3717}; 3718 3719 3720static void 3721mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 3722{ 3723 s64 val; 3724 3725 /* per cpu stat */ 3726 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); 3727 s->stat[MCS_CACHE] += val * PAGE_SIZE; 3728 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); 3729 s->stat[MCS_RSS] += val * PAGE_SIZE; 3730 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED); 3731 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 3732 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT); 3733 s->stat[MCS_PGPGIN] += val; 3734 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT); 3735 s->stat[MCS_PGPGOUT] += val; 3736 if (do_swap_account) { 3737 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT); 3738 s->stat[MCS_SWAP] += val * PAGE_SIZE; 3739 } 3740 3741 /* per zone stat */ 3742 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); 3743 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 3744 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); 3745 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 3746 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); 3747 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 3748 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); 3749 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 3750 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); 3751 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 3752} 3753 3754static void 3755mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 3756{ 3757 struct mem_cgroup *iter; 3758 3759 for_each_mem_cgroup_tree(iter, mem) 3760 mem_cgroup_get_local_stat(iter, s); 3761} 3762 3763static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 3764 struct cgroup_map_cb *cb) 3765{ 3766 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 3767 struct mcs_total_stat mystat; 3768 int i; 3769 3770 memset(&mystat, 0, sizeof(mystat)); 3771 mem_cgroup_get_local_stat(mem_cont, &mystat); 3772 3773 for (i = 0; i < NR_MCS_STAT; i++) { 3774 if (i == MCS_SWAP && !do_swap_account) 3775 continue; 3776 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 3777 } 3778 3779 /* Hierarchical information */ 3780 { 3781 unsigned long long limit, memsw_limit; 3782 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 3783 cb->fill(cb, "hierarchical_memory_limit", limit); 3784 if (do_swap_account) 3785 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 3786 } 3787 3788 memset(&mystat, 0, sizeof(mystat)); 3789 mem_cgroup_get_total_stat(mem_cont, &mystat); 3790 for (i = 0; i < NR_MCS_STAT; i++) { 3791 if (i == MCS_SWAP && !do_swap_account) 3792 continue; 3793 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 3794 } 3795 3796#ifdef CONFIG_DEBUG_VM 3797 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); 3798 3799 { 3800 int nid, zid; 3801 struct mem_cgroup_per_zone *mz; 3802 unsigned long recent_rotated[2] = {0, 0}; 3803 unsigned long recent_scanned[2] = {0, 0}; 3804 3805 for_each_online_node(nid) 3806 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 3807 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 3808 3809 recent_rotated[0] += 3810 mz->reclaim_stat.recent_rotated[0]; 3811 recent_rotated[1] += 3812 mz->reclaim_stat.recent_rotated[1]; 3813 recent_scanned[0] += 3814 mz->reclaim_stat.recent_scanned[0]; 3815 recent_scanned[1] += 3816 mz->reclaim_stat.recent_scanned[1]; 3817 } 3818 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 3819 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 3820 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 3821 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 3822 } 3823#endif 3824 3825 return 0; 3826} 3827 3828static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 3829{ 3830 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 3831 3832 return get_swappiness(memcg); 3833} 3834 3835static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 3836 u64 val) 3837{ 3838 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 3839 struct mem_cgroup *parent; 3840 3841 if (val > 100) 3842 return -EINVAL; 3843 3844 if (cgrp->parent == NULL) 3845 return -EINVAL; 3846 3847 parent = mem_cgroup_from_cont(cgrp->parent); 3848 3849 cgroup_lock(); 3850 3851 /* If under hierarchy, only empty-root can set this value */ 3852 if ((parent->use_hierarchy) || 3853 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 3854 cgroup_unlock(); 3855 return -EINVAL; 3856 } 3857 3858 memcg->swappiness = val; 3859 3860 cgroup_unlock(); 3861 3862 return 0; 3863} 3864 3865static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 3866{ 3867 struct mem_cgroup_threshold_ary *t; 3868 u64 usage; 3869 int i; 3870 3871 rcu_read_lock(); 3872 if (!swap) 3873 t = rcu_dereference(memcg->thresholds.primary); 3874 else 3875 t = rcu_dereference(memcg->memsw_thresholds.primary); 3876 3877 if (!t) 3878 goto unlock; 3879 3880 usage = mem_cgroup_usage(memcg, swap); 3881 3882 /* 3883 * current_threshold points to threshold just below usage. 3884 * If it's not true, a threshold was crossed after last 3885 * call of __mem_cgroup_threshold(). 3886 */ 3887 i = t->current_threshold; 3888 3889 /* 3890 * Iterate backward over array of thresholds starting from 3891 * current_threshold and check if a threshold is crossed. 3892 * If none of thresholds below usage is crossed, we read 3893 * only one element of the array here. 3894 */ 3895 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 3896 eventfd_signal(t->entries[i].eventfd, 1); 3897 3898 /* i = current_threshold + 1 */ 3899 i++; 3900 3901 /* 3902 * Iterate forward over array of thresholds starting from 3903 * current_threshold+1 and check if a threshold is crossed. 3904 * If none of thresholds above usage is crossed, we read 3905 * only one element of the array here. 3906 */ 3907 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 3908 eventfd_signal(t->entries[i].eventfd, 1); 3909 3910 /* Update current_threshold */ 3911 t->current_threshold = i - 1; 3912unlock: 3913 rcu_read_unlock(); 3914} 3915 3916static void mem_cgroup_threshold(struct mem_cgroup *memcg) 3917{ 3918 while (memcg) { 3919 __mem_cgroup_threshold(memcg, false); 3920 if (do_swap_account) 3921 __mem_cgroup_threshold(memcg, true); 3922 3923 memcg = parent_mem_cgroup(memcg); 3924 } 3925} 3926 3927static int compare_thresholds(const void *a, const void *b) 3928{ 3929 const struct mem_cgroup_threshold *_a = a; 3930 const struct mem_cgroup_threshold *_b = b; 3931 3932 return _a->threshold - _b->threshold; 3933} 3934 3935static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem) 3936{ 3937 struct mem_cgroup_eventfd_list *ev; 3938 3939 list_for_each_entry(ev, &mem->oom_notify, list) 3940 eventfd_signal(ev->eventfd, 1); 3941 return 0; 3942} 3943 3944static void mem_cgroup_oom_notify(struct mem_cgroup *mem) 3945{ 3946 struct mem_cgroup *iter; 3947 3948 for_each_mem_cgroup_tree(iter, mem) 3949 mem_cgroup_oom_notify_cb(iter); 3950} 3951 3952static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 3953 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 3954{ 3955 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 3956 struct mem_cgroup_thresholds *thresholds; 3957 struct mem_cgroup_threshold_ary *new; 3958 int type = MEMFILE_TYPE(cft->private); 3959 u64 threshold, usage; 3960 int i, size, ret; 3961 3962 ret = res_counter_memparse_write_strategy(args, &threshold); 3963 if (ret) 3964 return ret; 3965 3966 mutex_lock(&memcg->thresholds_lock); 3967 3968 if (type == _MEM) 3969 thresholds = &memcg->thresholds; 3970 else if (type == _MEMSWAP) 3971 thresholds = &memcg->memsw_thresholds; 3972 else 3973 BUG(); 3974 3975 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 3976 3977 /* Check if a threshold crossed before adding a new one */ 3978 if (thresholds->primary) 3979 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 3980 3981 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 3982 3983 /* Allocate memory for new array of thresholds */ 3984 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 3985 GFP_KERNEL); 3986 if (!new) { 3987 ret = -ENOMEM; 3988 goto unlock; 3989 } 3990 new->size = size; 3991 3992 /* Copy thresholds (if any) to new array */ 3993 if (thresholds->primary) { 3994 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 3995 sizeof(struct mem_cgroup_threshold)); 3996 } 3997 3998 /* Add new threshold */ 3999 new->entries[size - 1].eventfd = eventfd; 4000 new->entries[size - 1].threshold = threshold; 4001 4002 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4003 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4004 compare_thresholds, NULL); 4005 4006 /* Find current threshold */ 4007 new->current_threshold = -1; 4008 for (i = 0; i < size; i++) { 4009 if (new->entries[i].threshold < usage) { 4010 /* 4011 * new->current_threshold will not be used until 4012 * rcu_assign_pointer(), so it's safe to increment 4013 * it here. 4014 */ 4015 ++new->current_threshold; 4016 } 4017 } 4018 4019 /* Free old spare buffer and save old primary buffer as spare */ 4020 kfree(thresholds->spare); 4021 thresholds->spare = thresholds->primary; 4022 4023 rcu_assign_pointer(thresholds->primary, new); 4024 4025 /* To be sure that nobody uses thresholds */ 4026 synchronize_rcu(); 4027 4028unlock: 4029 mutex_unlock(&memcg->thresholds_lock); 4030 4031 return ret; 4032} 4033 4034static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4035 struct cftype *cft, struct eventfd_ctx *eventfd) 4036{ 4037 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4038 struct mem_cgroup_thresholds *thresholds; 4039 struct mem_cgroup_threshold_ary *new; 4040 int type = MEMFILE_TYPE(cft->private); 4041 u64 usage; 4042 int i, j, size; 4043 4044 mutex_lock(&memcg->thresholds_lock); 4045 if (type == _MEM) 4046 thresholds = &memcg->thresholds; 4047 else if (type == _MEMSWAP) 4048 thresholds = &memcg->memsw_thresholds; 4049 else 4050 BUG(); 4051 4052 /* 4053 * Something went wrong if we trying to unregister a threshold 4054 * if we don't have thresholds 4055 */ 4056 BUG_ON(!thresholds); 4057 4058 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4059 4060 /* Check if a threshold crossed before removing */ 4061 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4062 4063 /* Calculate new number of threshold */ 4064 size = 0; 4065 for (i = 0; i < thresholds->primary->size; i++) { 4066 if (thresholds->primary->entries[i].eventfd != eventfd) 4067 size++; 4068 } 4069 4070 new = thresholds->spare; 4071 4072 /* Set thresholds array to NULL if we don't have thresholds */ 4073 if (!size) { 4074 kfree(new); 4075 new = NULL; 4076 goto swap_buffers; 4077 } 4078 4079 new->size = size; 4080 4081 /* Copy thresholds and find current threshold */ 4082 new->current_threshold = -1; 4083 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4084 if (thresholds->primary->entries[i].eventfd == eventfd) 4085 continue; 4086 4087 new->entries[j] = thresholds->primary->entries[i]; 4088 if (new->entries[j].threshold < usage) { 4089 /* 4090 * new->current_threshold will not be used 4091 * until rcu_assign_pointer(), so it's safe to increment 4092 * it here. 4093 */ 4094 ++new->current_threshold; 4095 } 4096 j++; 4097 } 4098 4099swap_buffers: 4100 /* Swap primary and spare array */ 4101 thresholds->spare = thresholds->primary; 4102 rcu_assign_pointer(thresholds->primary, new); 4103 4104 /* To be sure that nobody uses thresholds */ 4105 synchronize_rcu(); 4106 4107 mutex_unlock(&memcg->thresholds_lock); 4108} 4109 4110static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4111 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4112{ 4113 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4114 struct mem_cgroup_eventfd_list *event; 4115 int type = MEMFILE_TYPE(cft->private); 4116 4117 BUG_ON(type != _OOM_TYPE); 4118 event = kmalloc(sizeof(*event), GFP_KERNEL); 4119 if (!event) 4120 return -ENOMEM; 4121 4122 mutex_lock(&memcg_oom_mutex); 4123 4124 event->eventfd = eventfd; 4125 list_add(&event->list, &memcg->oom_notify); 4126 4127 /* already in OOM ? */ 4128 if (atomic_read(&memcg->oom_lock)) 4129 eventfd_signal(eventfd, 1); 4130 mutex_unlock(&memcg_oom_mutex); 4131 4132 return 0; 4133} 4134 4135static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4136 struct cftype *cft, struct eventfd_ctx *eventfd) 4137{ 4138 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4139 struct mem_cgroup_eventfd_list *ev, *tmp; 4140 int type = MEMFILE_TYPE(cft->private); 4141 4142 BUG_ON(type != _OOM_TYPE); 4143 4144 mutex_lock(&memcg_oom_mutex); 4145 4146 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) { 4147 if (ev->eventfd == eventfd) { 4148 list_del(&ev->list); 4149 kfree(ev); 4150 } 4151 } 4152 4153 mutex_unlock(&memcg_oom_mutex); 4154} 4155 4156static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4157 struct cftype *cft, struct cgroup_map_cb *cb) 4158{ 4159 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4160 4161 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable); 4162 4163 if (atomic_read(&mem->oom_lock)) 4164 cb->fill(cb, "under_oom", 1); 4165 else 4166 cb->fill(cb, "under_oom", 0); 4167 return 0; 4168} 4169 4170static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4171 struct cftype *cft, u64 val) 4172{ 4173 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4174 struct mem_cgroup *parent; 4175 4176 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4177 if (!cgrp->parent || !((val == 0) || (val == 1))) 4178 return -EINVAL; 4179 4180 parent = mem_cgroup_from_cont(cgrp->parent); 4181 4182 cgroup_lock(); 4183 /* oom-kill-disable is a flag for subhierarchy. */ 4184 if ((parent->use_hierarchy) || 4185 (mem->use_hierarchy && !list_empty(&cgrp->children))) { 4186 cgroup_unlock(); 4187 return -EINVAL; 4188 } 4189 mem->oom_kill_disable = val; 4190 if (!val) 4191 memcg_oom_recover(mem); 4192 cgroup_unlock(); 4193 return 0; 4194} 4195 4196static struct cftype mem_cgroup_files[] = { 4197 { 4198 .name = "usage_in_bytes", 4199 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4200 .read_u64 = mem_cgroup_read, 4201 .register_event = mem_cgroup_usage_register_event, 4202 .unregister_event = mem_cgroup_usage_unregister_event, 4203 }, 4204 { 4205 .name = "max_usage_in_bytes", 4206 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4207 .trigger = mem_cgroup_reset, 4208 .read_u64 = mem_cgroup_read, 4209 }, 4210 { 4211 .name = "limit_in_bytes", 4212 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4213 .write_string = mem_cgroup_write, 4214 .read_u64 = mem_cgroup_read, 4215 }, 4216 { 4217 .name = "soft_limit_in_bytes", 4218 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4219 .write_string = mem_cgroup_write, 4220 .read_u64 = mem_cgroup_read, 4221 }, 4222 { 4223 .name = "failcnt", 4224 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4225 .trigger = mem_cgroup_reset, 4226 .read_u64 = mem_cgroup_read, 4227 }, 4228 { 4229 .name = "stat", 4230 .read_map = mem_control_stat_show, 4231 }, 4232 { 4233 .name = "force_empty", 4234 .trigger = mem_cgroup_force_empty_write, 4235 }, 4236 { 4237 .name = "use_hierarchy", 4238 .write_u64 = mem_cgroup_hierarchy_write, 4239 .read_u64 = mem_cgroup_hierarchy_read, 4240 }, 4241 { 4242 .name = "swappiness", 4243 .read_u64 = mem_cgroup_swappiness_read, 4244 .write_u64 = mem_cgroup_swappiness_write, 4245 }, 4246 { 4247 .name = "move_charge_at_immigrate", 4248 .read_u64 = mem_cgroup_move_charge_read, 4249 .write_u64 = mem_cgroup_move_charge_write, 4250 }, 4251 { 4252 .name = "oom_control", 4253 .read_map = mem_cgroup_oom_control_read, 4254 .write_u64 = mem_cgroup_oom_control_write, 4255 .register_event = mem_cgroup_oom_register_event, 4256 .unregister_event = mem_cgroup_oom_unregister_event, 4257 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4258 }, 4259}; 4260 4261#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4262static struct cftype memsw_cgroup_files[] = { 4263 { 4264 .name = "memsw.usage_in_bytes", 4265 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4266 .read_u64 = mem_cgroup_read, 4267 .register_event = mem_cgroup_usage_register_event, 4268 .unregister_event = mem_cgroup_usage_unregister_event, 4269 }, 4270 { 4271 .name = "memsw.max_usage_in_bytes", 4272 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4273 .trigger = mem_cgroup_reset, 4274 .read_u64 = mem_cgroup_read, 4275 }, 4276 { 4277 .name = "memsw.limit_in_bytes", 4278 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4279 .write_string = mem_cgroup_write, 4280 .read_u64 = mem_cgroup_read, 4281 }, 4282 { 4283 .name = "memsw.failcnt", 4284 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4285 .trigger = mem_cgroup_reset, 4286 .read_u64 = mem_cgroup_read, 4287 }, 4288}; 4289 4290static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4291{ 4292 if (!do_swap_account) 4293 return 0; 4294 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4295 ARRAY_SIZE(memsw_cgroup_files)); 4296}; 4297#else 4298static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4299{ 4300 return 0; 4301} 4302#endif 4303 4304static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 4305{ 4306 struct mem_cgroup_per_node *pn; 4307 struct mem_cgroup_per_zone *mz; 4308 enum lru_list l; 4309 int zone, tmp = node; 4310 /* 4311 * This routine is called against possible nodes. 4312 * But it's BUG to call kmalloc() against offline node. 4313 * 4314 * TODO: this routine can waste much memory for nodes which will 4315 * never be onlined. It's better to use memory hotplug callback 4316 * function. 4317 */ 4318 if (!node_state(node, N_NORMAL_MEMORY)) 4319 tmp = -1; 4320 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4321 if (!pn) 4322 return 1; 4323 4324 mem->info.nodeinfo[node] = pn; 4325 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4326 mz = &pn->zoneinfo[zone]; 4327 for_each_lru(l) 4328 INIT_LIST_HEAD(&mz->lists[l]); 4329 mz->usage_in_excess = 0; 4330 mz->on_tree = false; 4331 mz->mem = mem; 4332 } 4333 return 0; 4334} 4335 4336static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 4337{ 4338 kfree(mem->info.nodeinfo[node]); 4339} 4340 4341static struct mem_cgroup *mem_cgroup_alloc(void) 4342{ 4343 struct mem_cgroup *mem; 4344 int size = sizeof(struct mem_cgroup); 4345 4346 /* Can be very big if MAX_NUMNODES is very big */ 4347 if (size < PAGE_SIZE) 4348 mem = kzalloc(size, GFP_KERNEL); 4349 else 4350 mem = vzalloc(size); 4351 4352 if (!mem) 4353 return NULL; 4354 4355 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4356 if (!mem->stat) 4357 goto out_free; 4358 spin_lock_init(&mem->pcp_counter_lock); 4359 return mem; 4360 4361out_free: 4362 if (size < PAGE_SIZE) 4363 kfree(mem); 4364 else 4365 vfree(mem); 4366 return NULL; 4367} 4368 4369/* 4370 * At destroying mem_cgroup, references from swap_cgroup can remain. 4371 * (scanning all at force_empty is too costly...) 4372 * 4373 * Instead of clearing all references at force_empty, we remember 4374 * the number of reference from swap_cgroup and free mem_cgroup when 4375 * it goes down to 0. 4376 * 4377 * Removal of cgroup itself succeeds regardless of refs from swap. 4378 */ 4379 4380static void __mem_cgroup_free(struct mem_cgroup *mem) 4381{ 4382 int node; 4383 4384 mem_cgroup_remove_from_trees(mem); 4385 free_css_id(&mem_cgroup_subsys, &mem->css); 4386 4387 for_each_node_state(node, N_POSSIBLE) 4388 free_mem_cgroup_per_zone_info(mem, node); 4389 4390 free_percpu(mem->stat); 4391 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4392 kfree(mem); 4393 else 4394 vfree(mem); 4395} 4396 4397static void mem_cgroup_get(struct mem_cgroup *mem) 4398{ 4399 atomic_inc(&mem->refcnt); 4400} 4401 4402static void __mem_cgroup_put(struct mem_cgroup *mem, int count) 4403{ 4404 if (atomic_sub_and_test(count, &mem->refcnt)) { 4405 struct mem_cgroup *parent = parent_mem_cgroup(mem); 4406 __mem_cgroup_free(mem); 4407 if (parent) 4408 mem_cgroup_put(parent); 4409 } 4410} 4411 4412static void mem_cgroup_put(struct mem_cgroup *mem) 4413{ 4414 __mem_cgroup_put(mem, 1); 4415} 4416 4417/* 4418 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4419 */ 4420static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) 4421{ 4422 if (!mem->res.parent) 4423 return NULL; 4424 return mem_cgroup_from_res_counter(mem->res.parent, res); 4425} 4426 4427#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4428static void __init enable_swap_cgroup(void) 4429{ 4430 if (!mem_cgroup_disabled() && really_do_swap_account) 4431 do_swap_account = 1; 4432} 4433#else 4434static void __init enable_swap_cgroup(void) 4435{ 4436} 4437#endif 4438 4439static int mem_cgroup_soft_limit_tree_init(void) 4440{ 4441 struct mem_cgroup_tree_per_node *rtpn; 4442 struct mem_cgroup_tree_per_zone *rtpz; 4443 int tmp, node, zone; 4444 4445 for_each_node_state(node, N_POSSIBLE) { 4446 tmp = node; 4447 if (!node_state(node, N_NORMAL_MEMORY)) 4448 tmp = -1; 4449 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4450 if (!rtpn) 4451 return 1; 4452 4453 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4454 4455 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4456 rtpz = &rtpn->rb_tree_per_zone[zone]; 4457 rtpz->rb_root = RB_ROOT; 4458 spin_lock_init(&rtpz->lock); 4459 } 4460 } 4461 return 0; 4462} 4463 4464static struct cgroup_subsys_state * __ref 4465mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4466{ 4467 struct mem_cgroup *mem, *parent; 4468 long error = -ENOMEM; 4469 int node; 4470 4471 mem = mem_cgroup_alloc(); 4472 if (!mem) 4473 return ERR_PTR(error); 4474 4475 for_each_node_state(node, N_POSSIBLE) 4476 if (alloc_mem_cgroup_per_zone_info(mem, node)) 4477 goto free_out; 4478 4479 /* root ? */ 4480 if (cont->parent == NULL) { 4481 int cpu; 4482 enable_swap_cgroup(); 4483 parent = NULL; 4484 root_mem_cgroup = mem; 4485 if (mem_cgroup_soft_limit_tree_init()) 4486 goto free_out; 4487 for_each_possible_cpu(cpu) { 4488 struct memcg_stock_pcp *stock = 4489 &per_cpu(memcg_stock, cpu); 4490 INIT_WORK(&stock->work, drain_local_stock); 4491 } 4492 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 4493 } else { 4494 parent = mem_cgroup_from_cont(cont->parent); 4495 mem->use_hierarchy = parent->use_hierarchy; 4496 mem->oom_kill_disable = parent->oom_kill_disable; 4497 } 4498 4499 if (parent && parent->use_hierarchy) { 4500 res_counter_init(&mem->res, &parent->res); 4501 res_counter_init(&mem->memsw, &parent->memsw); 4502 /* 4503 * We increment refcnt of the parent to ensure that we can 4504 * safely access it on res_counter_charge/uncharge. 4505 * This refcnt will be decremented when freeing this 4506 * mem_cgroup(see mem_cgroup_put). 4507 */ 4508 mem_cgroup_get(parent); 4509 } else { 4510 res_counter_init(&mem->res, NULL); 4511 res_counter_init(&mem->memsw, NULL); 4512 } 4513 mem->last_scanned_child = 0; 4514 INIT_LIST_HEAD(&mem->oom_notify); 4515 4516 if (parent) 4517 mem->swappiness = get_swappiness(parent); 4518 atomic_set(&mem->refcnt, 1); 4519 mem->move_charge_at_immigrate = 0; 4520 mutex_init(&mem->thresholds_lock); 4521 return &mem->css; 4522free_out: 4523 __mem_cgroup_free(mem); 4524 root_mem_cgroup = NULL; 4525 return ERR_PTR(error); 4526} 4527 4528static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 4529 struct cgroup *cont) 4530{ 4531 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 4532 4533 return mem_cgroup_force_empty(mem, false); 4534} 4535 4536static void mem_cgroup_destroy(struct cgroup_subsys *ss, 4537 struct cgroup *cont) 4538{ 4539 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 4540 4541 mem_cgroup_put(mem); 4542} 4543 4544static int mem_cgroup_populate(struct cgroup_subsys *ss, 4545 struct cgroup *cont) 4546{ 4547 int ret; 4548 4549 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 4550 ARRAY_SIZE(mem_cgroup_files)); 4551 4552 if (!ret) 4553 ret = register_memsw_files(cont, ss); 4554 return ret; 4555} 4556 4557#ifdef CONFIG_MMU 4558/* Handlers for move charge at task migration. */ 4559#define PRECHARGE_COUNT_AT_ONCE 256 4560static int mem_cgroup_do_precharge(unsigned long count) 4561{ 4562 int ret = 0; 4563 int batch_count = PRECHARGE_COUNT_AT_ONCE; 4564 struct mem_cgroup *mem = mc.to; 4565 4566 if (mem_cgroup_is_root(mem)) { 4567 mc.precharge += count; 4568 /* we don't need css_get for root */ 4569 return ret; 4570 } 4571 /* try to charge at once */ 4572 if (count > 1) { 4573 struct res_counter *dummy; 4574 /* 4575 * "mem" cannot be under rmdir() because we've already checked 4576 * by cgroup_lock_live_cgroup() that it is not removed and we 4577 * are still under the same cgroup_mutex. So we can postpone 4578 * css_get(). 4579 */ 4580 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy)) 4581 goto one_by_one; 4582 if (do_swap_account && res_counter_charge(&mem->memsw, 4583 PAGE_SIZE * count, &dummy)) { 4584 res_counter_uncharge(&mem->res, PAGE_SIZE * count); 4585 goto one_by_one; 4586 } 4587 mc.precharge += count; 4588 return ret; 4589 } 4590one_by_one: 4591 /* fall back to one by one charge */ 4592 while (count--) { 4593 if (signal_pending(current)) { 4594 ret = -EINTR; 4595 break; 4596 } 4597 if (!batch_count--) { 4598 batch_count = PRECHARGE_COUNT_AT_ONCE; 4599 cond_resched(); 4600 } 4601 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false, 4602 PAGE_SIZE); 4603 if (ret || !mem) 4604 /* mem_cgroup_clear_mc() will do uncharge later */ 4605 return -ENOMEM; 4606 mc.precharge++; 4607 } 4608 return ret; 4609} 4610 4611/** 4612 * is_target_pte_for_mc - check a pte whether it is valid for move charge 4613 * @vma: the vma the pte to be checked belongs 4614 * @addr: the address corresponding to the pte to be checked 4615 * @ptent: the pte to be checked 4616 * @target: the pointer the target page or swap ent will be stored(can be NULL) 4617 * 4618 * Returns 4619 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 4620 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 4621 * move charge. if @target is not NULL, the page is stored in target->page 4622 * with extra refcnt got(Callers should handle it). 4623 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 4624 * target for charge migration. if @target is not NULL, the entry is stored 4625 * in target->ent. 4626 * 4627 * Called with pte lock held. 4628 */ 4629union mc_target { 4630 struct page *page; 4631 swp_entry_t ent; 4632}; 4633 4634enum mc_target_type { 4635 MC_TARGET_NONE, /* not used */ 4636 MC_TARGET_PAGE, 4637 MC_TARGET_SWAP, 4638}; 4639 4640static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 4641 unsigned long addr, pte_t ptent) 4642{ 4643 struct page *page = vm_normal_page(vma, addr, ptent); 4644 4645 if (!page || !page_mapped(page)) 4646 return NULL; 4647 if (PageAnon(page)) { 4648 /* we don't move shared anon */ 4649 if (!move_anon() || page_mapcount(page) > 2) 4650 return NULL; 4651 } else if (!move_file()) 4652 /* we ignore mapcount for file pages */ 4653 return NULL; 4654 if (!get_page_unless_zero(page)) 4655 return NULL; 4656 4657 return page; 4658} 4659 4660static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 4661 unsigned long addr, pte_t ptent, swp_entry_t *entry) 4662{ 4663 int usage_count; 4664 struct page *page = NULL; 4665 swp_entry_t ent = pte_to_swp_entry(ptent); 4666 4667 if (!move_anon() || non_swap_entry(ent)) 4668 return NULL; 4669 usage_count = mem_cgroup_count_swap_user(ent, &page); 4670 if (usage_count > 1) { /* we don't move shared anon */ 4671 if (page) 4672 put_page(page); 4673 return NULL; 4674 } 4675 if (do_swap_account) 4676 entry->val = ent.val; 4677 4678 return page; 4679} 4680 4681static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 4682 unsigned long addr, pte_t ptent, swp_entry_t *entry) 4683{ 4684 struct page *page = NULL; 4685 struct inode *inode; 4686 struct address_space *mapping; 4687 pgoff_t pgoff; 4688 4689 if (!vma->vm_file) /* anonymous vma */ 4690 return NULL; 4691 if (!move_file()) 4692 return NULL; 4693 4694 inode = vma->vm_file->f_path.dentry->d_inode; 4695 mapping = vma->vm_file->f_mapping; 4696 if (pte_none(ptent)) 4697 pgoff = linear_page_index(vma, addr); 4698 else /* pte_file(ptent) is true */ 4699 pgoff = pte_to_pgoff(ptent); 4700 4701 /* page is moved even if it's not RSS of this task(page-faulted). */ 4702 if (!mapping_cap_swap_backed(mapping)) { /* normal file */ 4703 page = find_get_page(mapping, pgoff); 4704 } else { /* shmem/tmpfs file. we should take account of swap too. */ 4705 swp_entry_t ent; 4706 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent); 4707 if (do_swap_account) 4708 entry->val = ent.val; 4709 } 4710 4711 return page; 4712} 4713 4714static int is_target_pte_for_mc(struct vm_area_struct *vma, 4715 unsigned long addr, pte_t ptent, union mc_target *target) 4716{ 4717 struct page *page = NULL; 4718 struct page_cgroup *pc; 4719 int ret = 0; 4720 swp_entry_t ent = { .val = 0 }; 4721 4722 if (pte_present(ptent)) 4723 page = mc_handle_present_pte(vma, addr, ptent); 4724 else if (is_swap_pte(ptent)) 4725 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 4726 else if (pte_none(ptent) || pte_file(ptent)) 4727 page = mc_handle_file_pte(vma, addr, ptent, &ent); 4728 4729 if (!page && !ent.val) 4730 return 0; 4731 if (page) { 4732 pc = lookup_page_cgroup(page); 4733 /* 4734 * Do only loose check w/o page_cgroup lock. 4735 * mem_cgroup_move_account() checks the pc is valid or not under 4736 * the lock. 4737 */ 4738 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 4739 ret = MC_TARGET_PAGE; 4740 if (target) 4741 target->page = page; 4742 } 4743 if (!ret || !target) 4744 put_page(page); 4745 } 4746 /* There is a swap entry and a page doesn't exist or isn't charged */ 4747 if (ent.val && !ret && 4748 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { 4749 ret = MC_TARGET_SWAP; 4750 if (target) 4751 target->ent = ent; 4752 } 4753 return ret; 4754} 4755 4756static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 4757 unsigned long addr, unsigned long end, 4758 struct mm_walk *walk) 4759{ 4760 struct vm_area_struct *vma = walk->private; 4761 pte_t *pte; 4762 spinlock_t *ptl; 4763 4764 split_huge_page_pmd(walk->mm, pmd); 4765 4766 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 4767 for (; addr != end; pte++, addr += PAGE_SIZE) 4768 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 4769 mc.precharge++; /* increment precharge temporarily */ 4770 pte_unmap_unlock(pte - 1, ptl); 4771 cond_resched(); 4772 4773 return 0; 4774} 4775 4776static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 4777{ 4778 unsigned long precharge; 4779 struct vm_area_struct *vma; 4780 4781 down_read(&mm->mmap_sem); 4782 for (vma = mm->mmap; vma; vma = vma->vm_next) { 4783 struct mm_walk mem_cgroup_count_precharge_walk = { 4784 .pmd_entry = mem_cgroup_count_precharge_pte_range, 4785 .mm = mm, 4786 .private = vma, 4787 }; 4788 if (is_vm_hugetlb_page(vma)) 4789 continue; 4790 walk_page_range(vma->vm_start, vma->vm_end, 4791 &mem_cgroup_count_precharge_walk); 4792 } 4793 up_read(&mm->mmap_sem); 4794 4795 precharge = mc.precharge; 4796 mc.precharge = 0; 4797 4798 return precharge; 4799} 4800 4801static int mem_cgroup_precharge_mc(struct mm_struct *mm) 4802{ 4803 unsigned long precharge = mem_cgroup_count_precharge(mm); 4804 4805 VM_BUG_ON(mc.moving_task); 4806 mc.moving_task = current; 4807 return mem_cgroup_do_precharge(precharge); 4808} 4809 4810/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 4811static void __mem_cgroup_clear_mc(void) 4812{ 4813 struct mem_cgroup *from = mc.from; 4814 struct mem_cgroup *to = mc.to; 4815 4816 /* we must uncharge all the leftover precharges from mc.to */ 4817 if (mc.precharge) { 4818 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 4819 mc.precharge = 0; 4820 } 4821 /* 4822 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 4823 * we must uncharge here. 4824 */ 4825 if (mc.moved_charge) { 4826 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 4827 mc.moved_charge = 0; 4828 } 4829 /* we must fixup refcnts and charges */ 4830 if (mc.moved_swap) { 4831 /* uncharge swap account from the old cgroup */ 4832 if (!mem_cgroup_is_root(mc.from)) 4833 res_counter_uncharge(&mc.from->memsw, 4834 PAGE_SIZE * mc.moved_swap); 4835 __mem_cgroup_put(mc.from, mc.moved_swap); 4836 4837 if (!mem_cgroup_is_root(mc.to)) { 4838 /* 4839 * we charged both to->res and to->memsw, so we should 4840 * uncharge to->res. 4841 */ 4842 res_counter_uncharge(&mc.to->res, 4843 PAGE_SIZE * mc.moved_swap); 4844 } 4845 /* we've already done mem_cgroup_get(mc.to) */ 4846 mc.moved_swap = 0; 4847 } 4848 memcg_oom_recover(from); 4849 memcg_oom_recover(to); 4850 wake_up_all(&mc.waitq); 4851} 4852 4853static void mem_cgroup_clear_mc(void) 4854{ 4855 struct mem_cgroup *from = mc.from; 4856 4857 /* 4858 * we must clear moving_task before waking up waiters at the end of 4859 * task migration. 4860 */ 4861 mc.moving_task = NULL; 4862 __mem_cgroup_clear_mc(); 4863 spin_lock(&mc.lock); 4864 mc.from = NULL; 4865 mc.to = NULL; 4866 spin_unlock(&mc.lock); 4867 mem_cgroup_end_move(from); 4868} 4869 4870static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 4871 struct cgroup *cgroup, 4872 struct task_struct *p, 4873 bool threadgroup) 4874{ 4875 int ret = 0; 4876 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup); 4877 4878 if (mem->move_charge_at_immigrate) { 4879 struct mm_struct *mm; 4880 struct mem_cgroup *from = mem_cgroup_from_task(p); 4881 4882 VM_BUG_ON(from == mem); 4883 4884 mm = get_task_mm(p); 4885 if (!mm) 4886 return 0; 4887 /* We move charges only when we move a owner of the mm */ 4888 if (mm->owner == p) { 4889 VM_BUG_ON(mc.from); 4890 VM_BUG_ON(mc.to); 4891 VM_BUG_ON(mc.precharge); 4892 VM_BUG_ON(mc.moved_charge); 4893 VM_BUG_ON(mc.moved_swap); 4894 mem_cgroup_start_move(from); 4895 spin_lock(&mc.lock); 4896 mc.from = from; 4897 mc.to = mem; 4898 spin_unlock(&mc.lock); 4899 /* We set mc.moving_task later */ 4900 4901 ret = mem_cgroup_precharge_mc(mm); 4902 if (ret) 4903 mem_cgroup_clear_mc(); 4904 } 4905 mmput(mm); 4906 } 4907 return ret; 4908} 4909 4910static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 4911 struct cgroup *cgroup, 4912 struct task_struct *p, 4913 bool threadgroup) 4914{ 4915 mem_cgroup_clear_mc(); 4916} 4917 4918static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 4919 unsigned long addr, unsigned long end, 4920 struct mm_walk *walk) 4921{ 4922 int ret = 0; 4923 struct vm_area_struct *vma = walk->private; 4924 pte_t *pte; 4925 spinlock_t *ptl; 4926 4927 split_huge_page_pmd(walk->mm, pmd); 4928retry: 4929 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 4930 for (; addr != end; addr += PAGE_SIZE) { 4931 pte_t ptent = *(pte++); 4932 union mc_target target; 4933 int type; 4934 struct page *page; 4935 struct page_cgroup *pc; 4936 swp_entry_t ent; 4937 4938 if (!mc.precharge) 4939 break; 4940 4941 type = is_target_pte_for_mc(vma, addr, ptent, &target); 4942 switch (type) { 4943 case MC_TARGET_PAGE: 4944 page = target.page; 4945 if (isolate_lru_page(page)) 4946 goto put; 4947 pc = lookup_page_cgroup(page); 4948 if (!mem_cgroup_move_account(page, pc, 4949 mc.from, mc.to, false, PAGE_SIZE)) { 4950 mc.precharge--; 4951 /* we uncharge from mc.from later. */ 4952 mc.moved_charge++; 4953 } 4954 putback_lru_page(page); 4955put: /* is_target_pte_for_mc() gets the page */ 4956 put_page(page); 4957 break; 4958 case MC_TARGET_SWAP: 4959 ent = target.ent; 4960 if (!mem_cgroup_move_swap_account(ent, 4961 mc.from, mc.to, false)) { 4962 mc.precharge--; 4963 /* we fixup refcnts and charges later. */ 4964 mc.moved_swap++; 4965 } 4966 break; 4967 default: 4968 break; 4969 } 4970 } 4971 pte_unmap_unlock(pte - 1, ptl); 4972 cond_resched(); 4973 4974 if (addr != end) { 4975 /* 4976 * We have consumed all precharges we got in can_attach(). 4977 * We try charge one by one, but don't do any additional 4978 * charges to mc.to if we have failed in charge once in attach() 4979 * phase. 4980 */ 4981 ret = mem_cgroup_do_precharge(1); 4982 if (!ret) 4983 goto retry; 4984 } 4985 4986 return ret; 4987} 4988 4989static void mem_cgroup_move_charge(struct mm_struct *mm) 4990{ 4991 struct vm_area_struct *vma; 4992 4993 lru_add_drain_all(); 4994retry: 4995 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 4996 /* 4997 * Someone who are holding the mmap_sem might be waiting in 4998 * waitq. So we cancel all extra charges, wake up all waiters, 4999 * and retry. Because we cancel precharges, we might not be able 5000 * to move enough charges, but moving charge is a best-effort 5001 * feature anyway, so it wouldn't be a big problem. 5002 */ 5003 __mem_cgroup_clear_mc(); 5004 cond_resched(); 5005 goto retry; 5006 } 5007 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5008 int ret; 5009 struct mm_walk mem_cgroup_move_charge_walk = { 5010 .pmd_entry = mem_cgroup_move_charge_pte_range, 5011 .mm = mm, 5012 .private = vma, 5013 }; 5014 if (is_vm_hugetlb_page(vma)) 5015 continue; 5016 ret = walk_page_range(vma->vm_start, vma->vm_end, 5017 &mem_cgroup_move_charge_walk); 5018 if (ret) 5019 /* 5020 * means we have consumed all precharges and failed in 5021 * doing additional charge. Just abandon here. 5022 */ 5023 break; 5024 } 5025 up_read(&mm->mmap_sem); 5026} 5027 5028static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5029 struct cgroup *cont, 5030 struct cgroup *old_cont, 5031 struct task_struct *p, 5032 bool threadgroup) 5033{ 5034 struct mm_struct *mm; 5035 5036 if (!mc.to) 5037 /* no need to move charge */ 5038 return; 5039 5040 mm = get_task_mm(p); 5041 if (mm) { 5042 mem_cgroup_move_charge(mm); 5043 mmput(mm); 5044 } 5045 mem_cgroup_clear_mc(); 5046} 5047#else /* !CONFIG_MMU */ 5048static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5049 struct cgroup *cgroup, 5050 struct task_struct *p, 5051 bool threadgroup) 5052{ 5053 return 0; 5054} 5055static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5056 struct cgroup *cgroup, 5057 struct task_struct *p, 5058 bool threadgroup) 5059{ 5060} 5061static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5062 struct cgroup *cont, 5063 struct cgroup *old_cont, 5064 struct task_struct *p, 5065 bool threadgroup) 5066{ 5067} 5068#endif 5069 5070struct cgroup_subsys mem_cgroup_subsys = { 5071 .name = "memory", 5072 .subsys_id = mem_cgroup_subsys_id, 5073 .create = mem_cgroup_create, 5074 .pre_destroy = mem_cgroup_pre_destroy, 5075 .destroy = mem_cgroup_destroy, 5076 .populate = mem_cgroup_populate, 5077 .can_attach = mem_cgroup_can_attach, 5078 .cancel_attach = mem_cgroup_cancel_attach, 5079 .attach = mem_cgroup_move_task, 5080 .early_init = 0, 5081 .use_id = 1, 5082}; 5083 5084#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5085static int __init enable_swap_account(char *s) 5086{ 5087 /* consider enabled if no parameter or 1 is given */ 5088 if (!(*s) || !strcmp(s, "=1")) 5089 really_do_swap_account = 1; 5090 else if (!strcmp(s, "=0")) 5091 really_do_swap_account = 0; 5092 return 1; 5093} 5094__setup("swapaccount", enable_swap_account); 5095 5096static int __init disable_swap_account(char *s) 5097{ 5098 printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n"); 5099 enable_swap_account("=0"); 5100 return 1; 5101} 5102__setup("noswapaccount", disable_swap_account); 5103#endif 5104