memcontrol.c revision a433658c30974fc87ba3ff52d7e4e6299762aa3d
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/* 78 * Statistics for memory cgroup. 79 */ 80enum mem_cgroup_stat_index { 81 /* 82 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 83 */ 84 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 85 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ 86 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */ 87 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */ 88 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */ 89 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */ 90 MEM_CGROUP_STAT_NSTATS, 91}; 92 93enum mem_cgroup_events_index { 94 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */ 95 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */ 96 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */ 97 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */ 98 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */ 99 MEM_CGROUP_EVENTS_NSTATS, 100}; 101/* 102 * Per memcg event counter is incremented at every pagein/pageout. With THP, 103 * it will be incremated by the number of pages. This counter is used for 104 * for trigger some periodic events. This is straightforward and better 105 * than using jiffies etc. to handle periodic memcg event. 106 */ 107enum mem_cgroup_events_target { 108 MEM_CGROUP_TARGET_THRESH, 109 MEM_CGROUP_TARGET_SOFTLIMIT, 110 MEM_CGROUP_NTARGETS, 111}; 112#define THRESHOLDS_EVENTS_TARGET (128) 113#define SOFTLIMIT_EVENTS_TARGET (1024) 114 115struct mem_cgroup_stat_cpu { 116 long count[MEM_CGROUP_STAT_NSTATS]; 117 unsigned long events[MEM_CGROUP_EVENTS_NSTATS]; 118 unsigned long targets[MEM_CGROUP_NTARGETS]; 119}; 120 121/* 122 * per-zone information in memory controller. 123 */ 124struct mem_cgroup_per_zone { 125 /* 126 * spin_lock to protect the per cgroup LRU 127 */ 128 struct list_head lists[NR_LRU_LISTS]; 129 unsigned long count[NR_LRU_LISTS]; 130 131 struct zone_reclaim_stat reclaim_stat; 132 struct rb_node tree_node; /* RB tree node */ 133 unsigned long long usage_in_excess;/* Set to the value by which */ 134 /* the soft limit is exceeded*/ 135 bool on_tree; 136 struct mem_cgroup *mem; /* Back pointer, we cannot */ 137 /* use container_of */ 138}; 139/* Macro for accessing counter */ 140#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 141 142struct mem_cgroup_per_node { 143 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 144}; 145 146struct mem_cgroup_lru_info { 147 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 148}; 149 150/* 151 * Cgroups above their limits are maintained in a RB-Tree, independent of 152 * their hierarchy representation 153 */ 154 155struct mem_cgroup_tree_per_zone { 156 struct rb_root rb_root; 157 spinlock_t lock; 158}; 159 160struct mem_cgroup_tree_per_node { 161 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES]; 162}; 163 164struct mem_cgroup_tree { 165 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES]; 166}; 167 168static struct mem_cgroup_tree soft_limit_tree __read_mostly; 169 170struct mem_cgroup_threshold { 171 struct eventfd_ctx *eventfd; 172 u64 threshold; 173}; 174 175/* For threshold */ 176struct mem_cgroup_threshold_ary { 177 /* An array index points to threshold just below usage. */ 178 int current_threshold; 179 /* Size of entries[] */ 180 unsigned int size; 181 /* Array of thresholds */ 182 struct mem_cgroup_threshold entries[0]; 183}; 184 185struct mem_cgroup_thresholds { 186 /* Primary thresholds array */ 187 struct mem_cgroup_threshold_ary *primary; 188 /* 189 * Spare threshold array. 190 * This is needed to make mem_cgroup_unregister_event() "never fail". 191 * It must be able to store at least primary->size - 1 entries. 192 */ 193 struct mem_cgroup_threshold_ary *spare; 194}; 195 196/* for OOM */ 197struct mem_cgroup_eventfd_list { 198 struct list_head list; 199 struct eventfd_ctx *eventfd; 200}; 201 202static void mem_cgroup_threshold(struct mem_cgroup *mem); 203static void mem_cgroup_oom_notify(struct mem_cgroup *mem); 204 205/* 206 * The memory controller data structure. The memory controller controls both 207 * page cache and RSS per cgroup. We would eventually like to provide 208 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 209 * to help the administrator determine what knobs to tune. 210 * 211 * TODO: Add a water mark for the memory controller. Reclaim will begin when 212 * we hit the water mark. May be even add a low water mark, such that 213 * no reclaim occurs from a cgroup at it's low water mark, this is 214 * a feature that will be implemented much later in the future. 215 */ 216struct mem_cgroup { 217 struct cgroup_subsys_state css; 218 /* 219 * the counter to account for memory usage 220 */ 221 struct res_counter res; 222 /* 223 * the counter to account for mem+swap usage. 224 */ 225 struct res_counter memsw; 226 /* 227 * Per cgroup active and inactive list, similar to the 228 * per zone LRU lists. 229 */ 230 struct mem_cgroup_lru_info info; 231 /* 232 * While reclaiming in a hierarchy, we cache the last child we 233 * reclaimed from. 234 */ 235 int last_scanned_child; 236 int last_scanned_node; 237#if MAX_NUMNODES > 1 238 nodemask_t scan_nodes; 239 unsigned long next_scan_node_update; 240#endif 241 /* 242 * Should the accounting and control be hierarchical, per subtree? 243 */ 244 bool use_hierarchy; 245 atomic_t oom_lock; 246 atomic_t refcnt; 247 248 unsigned int swappiness; 249 /* OOM-Killer disable */ 250 int oom_kill_disable; 251 252 /* set when res.limit == memsw.limit */ 253 bool memsw_is_minimum; 254 255 /* protect arrays of thresholds */ 256 struct mutex thresholds_lock; 257 258 /* thresholds for memory usage. RCU-protected */ 259 struct mem_cgroup_thresholds thresholds; 260 261 /* thresholds for mem+swap usage. RCU-protected */ 262 struct mem_cgroup_thresholds memsw_thresholds; 263 264 /* For oom notifier event fd */ 265 struct list_head oom_notify; 266 267 /* 268 * Should we move charges of a task when a task is moved into this 269 * mem_cgroup ? And what type of charges should we move ? 270 */ 271 unsigned long move_charge_at_immigrate; 272 /* 273 * percpu counter. 274 */ 275 struct mem_cgroup_stat_cpu *stat; 276 /* 277 * used when a cpu is offlined or other synchronizations 278 * See mem_cgroup_read_stat(). 279 */ 280 struct mem_cgroup_stat_cpu nocpu_base; 281 spinlock_t pcp_counter_lock; 282}; 283 284/* Stuffs for move charges at task migration. */ 285/* 286 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a 287 * left-shifted bitmap of these types. 288 */ 289enum move_type { 290 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */ 291 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */ 292 NR_MOVE_TYPE, 293}; 294 295/* "mc" and its members are protected by cgroup_mutex */ 296static struct move_charge_struct { 297 spinlock_t lock; /* for from, to */ 298 struct mem_cgroup *from; 299 struct mem_cgroup *to; 300 unsigned long precharge; 301 unsigned long moved_charge; 302 unsigned long moved_swap; 303 struct task_struct *moving_task; /* a task moving charges */ 304 wait_queue_head_t waitq; /* a waitq for other context */ 305} mc = { 306 .lock = __SPIN_LOCK_UNLOCKED(mc.lock), 307 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq), 308}; 309 310static bool move_anon(void) 311{ 312 return test_bit(MOVE_CHARGE_TYPE_ANON, 313 &mc.to->move_charge_at_immigrate); 314} 315 316static bool move_file(void) 317{ 318 return test_bit(MOVE_CHARGE_TYPE_FILE, 319 &mc.to->move_charge_at_immigrate); 320} 321 322/* 323 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft 324 * limit reclaim to prevent infinite loops, if they ever occur. 325 */ 326#define MEM_CGROUP_MAX_RECLAIM_LOOPS (100) 327#define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2) 328 329enum charge_type { 330 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 331 MEM_CGROUP_CHARGE_TYPE_MAPPED, 332 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 333 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 334 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 335 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ 336 NR_CHARGE_TYPE, 337}; 338 339/* for encoding cft->private value on file */ 340#define _MEM (0) 341#define _MEMSWAP (1) 342#define _OOM_TYPE (2) 343#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 344#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 345#define MEMFILE_ATTR(val) ((val) & 0xffff) 346/* Used for OOM nofiier */ 347#define OOM_CONTROL (0) 348 349/* 350 * Reclaim flags for mem_cgroup_hierarchical_reclaim 351 */ 352#define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0 353#define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT) 354#define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1 355#define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT) 356#define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2 357#define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT) 358 359static void mem_cgroup_get(struct mem_cgroup *mem); 360static void mem_cgroup_put(struct mem_cgroup *mem); 361static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); 362static void drain_all_stock_async(void); 363 364static struct mem_cgroup_per_zone * 365mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) 366{ 367 return &mem->info.nodeinfo[nid]->zoneinfo[zid]; 368} 369 370struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem) 371{ 372 return &mem->css; 373} 374 375static struct mem_cgroup_per_zone * 376page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page) 377{ 378 int nid = page_to_nid(page); 379 int zid = page_zonenum(page); 380 381 return mem_cgroup_zoneinfo(mem, nid, zid); 382} 383 384static struct mem_cgroup_tree_per_zone * 385soft_limit_tree_node_zone(int nid, int zid) 386{ 387 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 388} 389 390static struct mem_cgroup_tree_per_zone * 391soft_limit_tree_from_page(struct page *page) 392{ 393 int nid = page_to_nid(page); 394 int zid = page_zonenum(page); 395 396 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid]; 397} 398 399static void 400__mem_cgroup_insert_exceeded(struct mem_cgroup *mem, 401 struct mem_cgroup_per_zone *mz, 402 struct mem_cgroup_tree_per_zone *mctz, 403 unsigned long long new_usage_in_excess) 404{ 405 struct rb_node **p = &mctz->rb_root.rb_node; 406 struct rb_node *parent = NULL; 407 struct mem_cgroup_per_zone *mz_node; 408 409 if (mz->on_tree) 410 return; 411 412 mz->usage_in_excess = new_usage_in_excess; 413 if (!mz->usage_in_excess) 414 return; 415 while (*p) { 416 parent = *p; 417 mz_node = rb_entry(parent, struct mem_cgroup_per_zone, 418 tree_node); 419 if (mz->usage_in_excess < mz_node->usage_in_excess) 420 p = &(*p)->rb_left; 421 /* 422 * We can't avoid mem cgroups that are over their soft 423 * limit by the same amount 424 */ 425 else if (mz->usage_in_excess >= mz_node->usage_in_excess) 426 p = &(*p)->rb_right; 427 } 428 rb_link_node(&mz->tree_node, parent, p); 429 rb_insert_color(&mz->tree_node, &mctz->rb_root); 430 mz->on_tree = true; 431} 432 433static void 434__mem_cgroup_remove_exceeded(struct mem_cgroup *mem, 435 struct mem_cgroup_per_zone *mz, 436 struct mem_cgroup_tree_per_zone *mctz) 437{ 438 if (!mz->on_tree) 439 return; 440 rb_erase(&mz->tree_node, &mctz->rb_root); 441 mz->on_tree = false; 442} 443 444static void 445mem_cgroup_remove_exceeded(struct mem_cgroup *mem, 446 struct mem_cgroup_per_zone *mz, 447 struct mem_cgroup_tree_per_zone *mctz) 448{ 449 spin_lock(&mctz->lock); 450 __mem_cgroup_remove_exceeded(mem, mz, mctz); 451 spin_unlock(&mctz->lock); 452} 453 454 455static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page) 456{ 457 unsigned long long excess; 458 struct mem_cgroup_per_zone *mz; 459 struct mem_cgroup_tree_per_zone *mctz; 460 int nid = page_to_nid(page); 461 int zid = page_zonenum(page); 462 mctz = soft_limit_tree_from_page(page); 463 464 /* 465 * Necessary to update all ancestors when hierarchy is used. 466 * because their event counter is not touched. 467 */ 468 for (; mem; mem = parent_mem_cgroup(mem)) { 469 mz = mem_cgroup_zoneinfo(mem, nid, zid); 470 excess = res_counter_soft_limit_excess(&mem->res); 471 /* 472 * We have to update the tree if mz is on RB-tree or 473 * mem is over its softlimit. 474 */ 475 if (excess || mz->on_tree) { 476 spin_lock(&mctz->lock); 477 /* if on-tree, remove it */ 478 if (mz->on_tree) 479 __mem_cgroup_remove_exceeded(mem, mz, mctz); 480 /* 481 * Insert again. mz->usage_in_excess will be updated. 482 * If excess is 0, no tree ops. 483 */ 484 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess); 485 spin_unlock(&mctz->lock); 486 } 487 } 488} 489 490static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem) 491{ 492 int node, zone; 493 struct mem_cgroup_per_zone *mz; 494 struct mem_cgroup_tree_per_zone *mctz; 495 496 for_each_node_state(node, N_POSSIBLE) { 497 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 498 mz = mem_cgroup_zoneinfo(mem, node, zone); 499 mctz = soft_limit_tree_node_zone(node, zone); 500 mem_cgroup_remove_exceeded(mem, mz, mctz); 501 } 502 } 503} 504 505static struct mem_cgroup_per_zone * 506__mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 507{ 508 struct rb_node *rightmost = NULL; 509 struct mem_cgroup_per_zone *mz; 510 511retry: 512 mz = NULL; 513 rightmost = rb_last(&mctz->rb_root); 514 if (!rightmost) 515 goto done; /* Nothing to reclaim from */ 516 517 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node); 518 /* 519 * Remove the node now but someone else can add it back, 520 * we will to add it back at the end of reclaim to its correct 521 * position in the tree. 522 */ 523 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 524 if (!res_counter_soft_limit_excess(&mz->mem->res) || 525 !css_tryget(&mz->mem->css)) 526 goto retry; 527done: 528 return mz; 529} 530 531static struct mem_cgroup_per_zone * 532mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz) 533{ 534 struct mem_cgroup_per_zone *mz; 535 536 spin_lock(&mctz->lock); 537 mz = __mem_cgroup_largest_soft_limit_node(mctz); 538 spin_unlock(&mctz->lock); 539 return mz; 540} 541 542/* 543 * Implementation Note: reading percpu statistics for memcg. 544 * 545 * Both of vmstat[] and percpu_counter has threshold and do periodic 546 * synchronization to implement "quick" read. There are trade-off between 547 * reading cost and precision of value. Then, we may have a chance to implement 548 * a periodic synchronizion of counter in memcg's counter. 549 * 550 * But this _read() function is used for user interface now. The user accounts 551 * memory usage by memory cgroup and he _always_ requires exact value because 552 * he accounts memory. Even if we provide quick-and-fuzzy read, we always 553 * have to visit all online cpus and make sum. So, for now, unnecessary 554 * synchronization is not implemented. (just implemented for cpu hotplug) 555 * 556 * If there are kernel internal actions which can make use of some not-exact 557 * value, and reading all cpu value can be performance bottleneck in some 558 * common workload, threashold and synchonization as vmstat[] should be 559 * implemented. 560 */ 561static long mem_cgroup_read_stat(struct mem_cgroup *mem, 562 enum mem_cgroup_stat_index idx) 563{ 564 long val = 0; 565 int cpu; 566 567 get_online_cpus(); 568 for_each_online_cpu(cpu) 569 val += per_cpu(mem->stat->count[idx], cpu); 570#ifdef CONFIG_HOTPLUG_CPU 571 spin_lock(&mem->pcp_counter_lock); 572 val += mem->nocpu_base.count[idx]; 573 spin_unlock(&mem->pcp_counter_lock); 574#endif 575 put_online_cpus(); 576 return val; 577} 578 579static long mem_cgroup_local_usage(struct mem_cgroup *mem) 580{ 581 long ret; 582 583 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); 584 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); 585 return ret; 586} 587 588static void mem_cgroup_swap_statistics(struct mem_cgroup *mem, 589 bool charge) 590{ 591 int val = (charge) ? 1 : -1; 592 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val); 593} 594 595void mem_cgroup_pgfault(struct mem_cgroup *mem, int val) 596{ 597 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val); 598} 599 600void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val) 601{ 602 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val); 603} 604 605static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem, 606 enum mem_cgroup_events_index idx) 607{ 608 unsigned long val = 0; 609 int cpu; 610 611 for_each_online_cpu(cpu) 612 val += per_cpu(mem->stat->events[idx], cpu); 613#ifdef CONFIG_HOTPLUG_CPU 614 spin_lock(&mem->pcp_counter_lock); 615 val += mem->nocpu_base.events[idx]; 616 spin_unlock(&mem->pcp_counter_lock); 617#endif 618 return val; 619} 620 621static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, 622 bool file, int nr_pages) 623{ 624 preempt_disable(); 625 626 if (file) 627 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages); 628 else 629 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages); 630 631 /* pagein of a big page is an event. So, ignore page size */ 632 if (nr_pages > 0) 633 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]); 634 else { 635 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]); 636 nr_pages = -nr_pages; /* for event */ 637 } 638 639 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages); 640 641 preempt_enable(); 642} 643 644static unsigned long 645mem_cgroup_get_zonestat_node(struct mem_cgroup *mem, int nid, enum lru_list idx) 646{ 647 struct mem_cgroup_per_zone *mz; 648 u64 total = 0; 649 int zid; 650 651 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 652 mz = mem_cgroup_zoneinfo(mem, nid, zid); 653 total += MEM_CGROUP_ZSTAT(mz, idx); 654 } 655 return total; 656} 657static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, 658 enum lru_list idx) 659{ 660 int nid; 661 u64 total = 0; 662 663 for_each_online_node(nid) 664 total += mem_cgroup_get_zonestat_node(mem, nid, idx); 665 return total; 666} 667 668static bool __memcg_event_check(struct mem_cgroup *mem, int target) 669{ 670 unsigned long val, next; 671 672 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]); 673 next = this_cpu_read(mem->stat->targets[target]); 674 /* from time_after() in jiffies.h */ 675 return ((long)next - (long)val < 0); 676} 677 678static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target) 679{ 680 unsigned long val, next; 681 682 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]); 683 684 switch (target) { 685 case MEM_CGROUP_TARGET_THRESH: 686 next = val + THRESHOLDS_EVENTS_TARGET; 687 break; 688 case MEM_CGROUP_TARGET_SOFTLIMIT: 689 next = val + SOFTLIMIT_EVENTS_TARGET; 690 break; 691 default: 692 return; 693 } 694 695 this_cpu_write(mem->stat->targets[target], next); 696} 697 698/* 699 * Check events in order. 700 * 701 */ 702static void memcg_check_events(struct mem_cgroup *mem, struct page *page) 703{ 704 /* threshold event is triggered in finer grain than soft limit */ 705 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) { 706 mem_cgroup_threshold(mem); 707 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH); 708 if (unlikely(__memcg_event_check(mem, 709 MEM_CGROUP_TARGET_SOFTLIMIT))){ 710 mem_cgroup_update_tree(mem, page); 711 __mem_cgroup_target_update(mem, 712 MEM_CGROUP_TARGET_SOFTLIMIT); 713 } 714 } 715} 716 717static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 718{ 719 return container_of(cgroup_subsys_state(cont, 720 mem_cgroup_subsys_id), struct mem_cgroup, 721 css); 722} 723 724struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 725{ 726 /* 727 * mm_update_next_owner() may clear mm->owner to NULL 728 * if it races with swapoff, page migration, etc. 729 * So this can be called with p == NULL. 730 */ 731 if (unlikely(!p)) 732 return NULL; 733 734 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 735 struct mem_cgroup, css); 736} 737 738struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 739{ 740 struct mem_cgroup *mem = NULL; 741 742 if (!mm) 743 return NULL; 744 /* 745 * Because we have no locks, mm->owner's may be being moved to other 746 * cgroup. We use css_tryget() here even if this looks 747 * pessimistic (rather than adding locks here). 748 */ 749 rcu_read_lock(); 750 do { 751 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 752 if (unlikely(!mem)) 753 break; 754 } while (!css_tryget(&mem->css)); 755 rcu_read_unlock(); 756 return mem; 757} 758 759/* The caller has to guarantee "mem" exists before calling this */ 760static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem) 761{ 762 struct cgroup_subsys_state *css; 763 int found; 764 765 if (!mem) /* ROOT cgroup has the smallest ID */ 766 return root_mem_cgroup; /*css_put/get against root is ignored*/ 767 if (!mem->use_hierarchy) { 768 if (css_tryget(&mem->css)) 769 return mem; 770 return NULL; 771 } 772 rcu_read_lock(); 773 /* 774 * searching a memory cgroup which has the smallest ID under given 775 * ROOT cgroup. (ID >= 1) 776 */ 777 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found); 778 if (css && css_tryget(css)) 779 mem = container_of(css, struct mem_cgroup, css); 780 else 781 mem = NULL; 782 rcu_read_unlock(); 783 return mem; 784} 785 786static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter, 787 struct mem_cgroup *root, 788 bool cond) 789{ 790 int nextid = css_id(&iter->css) + 1; 791 int found; 792 int hierarchy_used; 793 struct cgroup_subsys_state *css; 794 795 hierarchy_used = iter->use_hierarchy; 796 797 css_put(&iter->css); 798 /* If no ROOT, walk all, ignore hierarchy */ 799 if (!cond || (root && !hierarchy_used)) 800 return NULL; 801 802 if (!root) 803 root = root_mem_cgroup; 804 805 do { 806 iter = NULL; 807 rcu_read_lock(); 808 809 css = css_get_next(&mem_cgroup_subsys, nextid, 810 &root->css, &found); 811 if (css && css_tryget(css)) 812 iter = container_of(css, struct mem_cgroup, css); 813 rcu_read_unlock(); 814 /* If css is NULL, no more cgroups will be found */ 815 nextid = found + 1; 816 } while (css && !iter); 817 818 return iter; 819} 820/* 821 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please 822 * be careful that "break" loop is not allowed. We have reference count. 823 * Instead of that modify "cond" to be false and "continue" to exit the loop. 824 */ 825#define for_each_mem_cgroup_tree_cond(iter, root, cond) \ 826 for (iter = mem_cgroup_start_loop(root);\ 827 iter != NULL;\ 828 iter = mem_cgroup_get_next(iter, root, cond)) 829 830#define for_each_mem_cgroup_tree(iter, root) \ 831 for_each_mem_cgroup_tree_cond(iter, root, true) 832 833#define for_each_mem_cgroup_all(iter) \ 834 for_each_mem_cgroup_tree_cond(iter, NULL, true) 835 836 837static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) 838{ 839 return (mem == root_mem_cgroup); 840} 841 842void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx) 843{ 844 struct mem_cgroup *mem; 845 846 if (!mm) 847 return; 848 849 rcu_read_lock(); 850 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 851 if (unlikely(!mem)) 852 goto out; 853 854 switch (idx) { 855 case PGMAJFAULT: 856 mem_cgroup_pgmajfault(mem, 1); 857 break; 858 case PGFAULT: 859 mem_cgroup_pgfault(mem, 1); 860 break; 861 default: 862 BUG(); 863 } 864out: 865 rcu_read_unlock(); 866} 867EXPORT_SYMBOL(mem_cgroup_count_vm_event); 868 869/* 870 * Following LRU functions are allowed to be used without PCG_LOCK. 871 * Operations are called by routine of global LRU independently from memcg. 872 * What we have to take care of here is validness of pc->mem_cgroup. 873 * 874 * Changes to pc->mem_cgroup happens when 875 * 1. charge 876 * 2. moving account 877 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 878 * It is added to LRU before charge. 879 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 880 * When moving account, the page is not on LRU. It's isolated. 881 */ 882 883void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) 884{ 885 struct page_cgroup *pc; 886 struct mem_cgroup_per_zone *mz; 887 888 if (mem_cgroup_disabled()) 889 return; 890 pc = lookup_page_cgroup(page); 891 /* can happen while we handle swapcache. */ 892 if (!TestClearPageCgroupAcctLRU(pc)) 893 return; 894 VM_BUG_ON(!pc->mem_cgroup); 895 /* 896 * We don't check PCG_USED bit. It's cleared when the "page" is finally 897 * removed from global LRU. 898 */ 899 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 900 /* huge page split is done under lru_lock. so, we have no races. */ 901 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page); 902 if (mem_cgroup_is_root(pc->mem_cgroup)) 903 return; 904 VM_BUG_ON(list_empty(&pc->lru)); 905 list_del_init(&pc->lru); 906} 907 908void mem_cgroup_del_lru(struct page *page) 909{ 910 mem_cgroup_del_lru_list(page, page_lru(page)); 911} 912 913/* 914 * Writeback is about to end against a page which has been marked for immediate 915 * reclaim. If it still appears to be reclaimable, move it to the tail of the 916 * inactive list. 917 */ 918void mem_cgroup_rotate_reclaimable_page(struct page *page) 919{ 920 struct mem_cgroup_per_zone *mz; 921 struct page_cgroup *pc; 922 enum lru_list lru = page_lru(page); 923 924 if (mem_cgroup_disabled()) 925 return; 926 927 pc = lookup_page_cgroup(page); 928 /* unused or root page is not rotated. */ 929 if (!PageCgroupUsed(pc)) 930 return; 931 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 932 smp_rmb(); 933 if (mem_cgroup_is_root(pc->mem_cgroup)) 934 return; 935 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 936 list_move_tail(&pc->lru, &mz->lists[lru]); 937} 938 939void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) 940{ 941 struct mem_cgroup_per_zone *mz; 942 struct page_cgroup *pc; 943 944 if (mem_cgroup_disabled()) 945 return; 946 947 pc = lookup_page_cgroup(page); 948 /* unused or root page is not rotated. */ 949 if (!PageCgroupUsed(pc)) 950 return; 951 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 952 smp_rmb(); 953 if (mem_cgroup_is_root(pc->mem_cgroup)) 954 return; 955 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 956 list_move(&pc->lru, &mz->lists[lru]); 957} 958 959void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) 960{ 961 struct page_cgroup *pc; 962 struct mem_cgroup_per_zone *mz; 963 964 if (mem_cgroup_disabled()) 965 return; 966 pc = lookup_page_cgroup(page); 967 VM_BUG_ON(PageCgroupAcctLRU(pc)); 968 if (!PageCgroupUsed(pc)) 969 return; 970 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 971 smp_rmb(); 972 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 973 /* huge page split is done under lru_lock. so, we have no races. */ 974 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page); 975 SetPageCgroupAcctLRU(pc); 976 if (mem_cgroup_is_root(pc->mem_cgroup)) 977 return; 978 list_add(&pc->lru, &mz->lists[lru]); 979} 980 981/* 982 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed 983 * while it's linked to lru because the page may be reused after it's fully 984 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again. 985 * It's done under lock_page and expected that zone->lru_lock isnever held. 986 */ 987static void mem_cgroup_lru_del_before_commit(struct page *page) 988{ 989 unsigned long flags; 990 struct zone *zone = page_zone(page); 991 struct page_cgroup *pc = lookup_page_cgroup(page); 992 993 /* 994 * Doing this check without taking ->lru_lock seems wrong but this 995 * is safe. Because if page_cgroup's USED bit is unset, the page 996 * will not be added to any memcg's LRU. If page_cgroup's USED bit is 997 * set, the commit after this will fail, anyway. 998 * This all charge/uncharge is done under some mutual execustion. 999 * So, we don't need to taking care of changes in USED bit. 1000 */ 1001 if (likely(!PageLRU(page))) 1002 return; 1003 1004 spin_lock_irqsave(&zone->lru_lock, flags); 1005 /* 1006 * Forget old LRU when this page_cgroup is *not* used. This Used bit 1007 * is guarded by lock_page() because the page is SwapCache. 1008 */ 1009 if (!PageCgroupUsed(pc)) 1010 mem_cgroup_del_lru_list(page, page_lru(page)); 1011 spin_unlock_irqrestore(&zone->lru_lock, flags); 1012} 1013 1014static void mem_cgroup_lru_add_after_commit(struct page *page) 1015{ 1016 unsigned long flags; 1017 struct zone *zone = page_zone(page); 1018 struct page_cgroup *pc = lookup_page_cgroup(page); 1019 1020 /* taking care of that the page is added to LRU while we commit it */ 1021 if (likely(!PageLRU(page))) 1022 return; 1023 spin_lock_irqsave(&zone->lru_lock, flags); 1024 /* link when the page is linked to LRU but page_cgroup isn't */ 1025 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) 1026 mem_cgroup_add_lru_list(page, page_lru(page)); 1027 spin_unlock_irqrestore(&zone->lru_lock, flags); 1028} 1029 1030 1031void mem_cgroup_move_lists(struct page *page, 1032 enum lru_list from, enum lru_list to) 1033{ 1034 if (mem_cgroup_disabled()) 1035 return; 1036 mem_cgroup_del_lru_list(page, from); 1037 mem_cgroup_add_lru_list(page, to); 1038} 1039 1040int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) 1041{ 1042 int ret; 1043 struct mem_cgroup *curr = NULL; 1044 struct task_struct *p; 1045 1046 p = find_lock_task_mm(task); 1047 if (!p) 1048 return 0; 1049 curr = try_get_mem_cgroup_from_mm(p->mm); 1050 task_unlock(p); 1051 if (!curr) 1052 return 0; 1053 /* 1054 * We should check use_hierarchy of "mem" not "curr". Because checking 1055 * use_hierarchy of "curr" here make this function true if hierarchy is 1056 * enabled in "curr" and "curr" is a child of "mem" in *cgroup* 1057 * hierarchy(even if use_hierarchy is disabled in "mem"). 1058 */ 1059 if (mem->use_hierarchy) 1060 ret = css_is_ancestor(&curr->css, &mem->css); 1061 else 1062 ret = (curr == mem); 1063 css_put(&curr->css); 1064 return ret; 1065} 1066 1067static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) 1068{ 1069 unsigned long active; 1070 unsigned long inactive; 1071 unsigned long gb; 1072 unsigned long inactive_ratio; 1073 1074 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); 1075 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); 1076 1077 gb = (inactive + active) >> (30 - PAGE_SHIFT); 1078 if (gb) 1079 inactive_ratio = int_sqrt(10 * gb); 1080 else 1081 inactive_ratio = 1; 1082 1083 if (present_pages) { 1084 present_pages[0] = inactive; 1085 present_pages[1] = active; 1086 } 1087 1088 return inactive_ratio; 1089} 1090 1091int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) 1092{ 1093 unsigned long active; 1094 unsigned long inactive; 1095 unsigned long present_pages[2]; 1096 unsigned long inactive_ratio; 1097 1098 inactive_ratio = calc_inactive_ratio(memcg, present_pages); 1099 1100 inactive = present_pages[0]; 1101 active = present_pages[1]; 1102 1103 if (inactive * inactive_ratio < active) 1104 return 1; 1105 1106 return 0; 1107} 1108 1109int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) 1110{ 1111 unsigned long active; 1112 unsigned long inactive; 1113 1114 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); 1115 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); 1116 1117 return (active > inactive); 1118} 1119 1120unsigned long mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, 1121 struct zone *zone, 1122 enum lru_list lru) 1123{ 1124 int nid = zone_to_nid(zone); 1125 int zid = zone_idx(zone); 1126 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1127 1128 return MEM_CGROUP_ZSTAT(mz, lru); 1129} 1130 1131#ifdef CONFIG_NUMA 1132static unsigned long mem_cgroup_node_nr_file_lru_pages(struct mem_cgroup *memcg, 1133 int nid) 1134{ 1135 unsigned long ret; 1136 1137 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_FILE) + 1138 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_FILE); 1139 1140 return ret; 1141} 1142 1143static unsigned long mem_cgroup_nr_file_lru_pages(struct mem_cgroup *memcg) 1144{ 1145 u64 total = 0; 1146 int nid; 1147 1148 for_each_node_state(nid, N_HIGH_MEMORY) 1149 total += mem_cgroup_node_nr_file_lru_pages(memcg, nid); 1150 1151 return total; 1152} 1153 1154static unsigned long mem_cgroup_node_nr_anon_lru_pages(struct mem_cgroup *memcg, 1155 int nid) 1156{ 1157 unsigned long ret; 1158 1159 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_ANON) + 1160 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_ANON); 1161 1162 return ret; 1163} 1164 1165static unsigned long mem_cgroup_nr_anon_lru_pages(struct mem_cgroup *memcg) 1166{ 1167 u64 total = 0; 1168 int nid; 1169 1170 for_each_node_state(nid, N_HIGH_MEMORY) 1171 total += mem_cgroup_node_nr_anon_lru_pages(memcg, nid); 1172 1173 return total; 1174} 1175 1176static unsigned long 1177mem_cgroup_node_nr_unevictable_lru_pages(struct mem_cgroup *memcg, int nid) 1178{ 1179 return mem_cgroup_get_zonestat_node(memcg, nid, LRU_UNEVICTABLE); 1180} 1181 1182static unsigned long 1183mem_cgroup_nr_unevictable_lru_pages(struct mem_cgroup *memcg) 1184{ 1185 u64 total = 0; 1186 int nid; 1187 1188 for_each_node_state(nid, N_HIGH_MEMORY) 1189 total += mem_cgroup_node_nr_unevictable_lru_pages(memcg, nid); 1190 1191 return total; 1192} 1193 1194static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg, 1195 int nid) 1196{ 1197 enum lru_list l; 1198 u64 total = 0; 1199 1200 for_each_lru(l) 1201 total += mem_cgroup_get_zonestat_node(memcg, nid, l); 1202 1203 return total; 1204} 1205 1206static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg) 1207{ 1208 u64 total = 0; 1209 int nid; 1210 1211 for_each_node_state(nid, N_HIGH_MEMORY) 1212 total += mem_cgroup_node_nr_lru_pages(memcg, nid); 1213 1214 return total; 1215} 1216#endif /* CONFIG_NUMA */ 1217 1218struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 1219 struct zone *zone) 1220{ 1221 int nid = zone_to_nid(zone); 1222 int zid = zone_idx(zone); 1223 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 1224 1225 return &mz->reclaim_stat; 1226} 1227 1228struct zone_reclaim_stat * 1229mem_cgroup_get_reclaim_stat_from_page(struct page *page) 1230{ 1231 struct page_cgroup *pc; 1232 struct mem_cgroup_per_zone *mz; 1233 1234 if (mem_cgroup_disabled()) 1235 return NULL; 1236 1237 pc = lookup_page_cgroup(page); 1238 if (!PageCgroupUsed(pc)) 1239 return NULL; 1240 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */ 1241 smp_rmb(); 1242 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page); 1243 return &mz->reclaim_stat; 1244} 1245 1246unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, 1247 struct list_head *dst, 1248 unsigned long *scanned, int order, 1249 int mode, struct zone *z, 1250 struct mem_cgroup *mem_cont, 1251 int active, int file) 1252{ 1253 unsigned long nr_taken = 0; 1254 struct page *page; 1255 unsigned long scan; 1256 LIST_HEAD(pc_list); 1257 struct list_head *src; 1258 struct page_cgroup *pc, *tmp; 1259 int nid = zone_to_nid(z); 1260 int zid = zone_idx(z); 1261 struct mem_cgroup_per_zone *mz; 1262 int lru = LRU_FILE * file + active; 1263 int ret; 1264 1265 BUG_ON(!mem_cont); 1266 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 1267 src = &mz->lists[lru]; 1268 1269 scan = 0; 1270 list_for_each_entry_safe_reverse(pc, tmp, src, lru) { 1271 if (scan >= nr_to_scan) 1272 break; 1273 1274 if (unlikely(!PageCgroupUsed(pc))) 1275 continue; 1276 1277 page = lookup_cgroup_page(pc); 1278 1279 if (unlikely(!PageLRU(page))) 1280 continue; 1281 1282 scan++; 1283 ret = __isolate_lru_page(page, mode, file); 1284 switch (ret) { 1285 case 0: 1286 list_move(&page->lru, dst); 1287 mem_cgroup_del_lru(page); 1288 nr_taken += hpage_nr_pages(page); 1289 break; 1290 case -EBUSY: 1291 /* we don't affect global LRU but rotate in our LRU */ 1292 mem_cgroup_rotate_lru_list(page, page_lru(page)); 1293 break; 1294 default: 1295 break; 1296 } 1297 } 1298 1299 *scanned = scan; 1300 1301 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken, 1302 0, 0, 0, mode); 1303 1304 return nr_taken; 1305} 1306 1307#define mem_cgroup_from_res_counter(counter, member) \ 1308 container_of(counter, struct mem_cgroup, member) 1309 1310/** 1311 * mem_cgroup_margin - calculate chargeable space of a memory cgroup 1312 * @mem: the memory cgroup 1313 * 1314 * Returns the maximum amount of memory @mem can be charged with, in 1315 * pages. 1316 */ 1317static unsigned long mem_cgroup_margin(struct mem_cgroup *mem) 1318{ 1319 unsigned long long margin; 1320 1321 margin = res_counter_margin(&mem->res); 1322 if (do_swap_account) 1323 margin = min(margin, res_counter_margin(&mem->memsw)); 1324 return margin >> PAGE_SHIFT; 1325} 1326 1327static unsigned int get_swappiness(struct mem_cgroup *memcg) 1328{ 1329 struct cgroup *cgrp = memcg->css.cgroup; 1330 1331 /* root ? */ 1332 if (cgrp->parent == NULL) 1333 return vm_swappiness; 1334 1335 return memcg->swappiness; 1336} 1337 1338static void mem_cgroup_start_move(struct mem_cgroup *mem) 1339{ 1340 int cpu; 1341 1342 get_online_cpus(); 1343 spin_lock(&mem->pcp_counter_lock); 1344 for_each_online_cpu(cpu) 1345 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1; 1346 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1; 1347 spin_unlock(&mem->pcp_counter_lock); 1348 put_online_cpus(); 1349 1350 synchronize_rcu(); 1351} 1352 1353static void mem_cgroup_end_move(struct mem_cgroup *mem) 1354{ 1355 int cpu; 1356 1357 if (!mem) 1358 return; 1359 get_online_cpus(); 1360 spin_lock(&mem->pcp_counter_lock); 1361 for_each_online_cpu(cpu) 1362 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1; 1363 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1; 1364 spin_unlock(&mem->pcp_counter_lock); 1365 put_online_cpus(); 1366} 1367/* 1368 * 2 routines for checking "mem" is under move_account() or not. 1369 * 1370 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used 1371 * for avoiding race in accounting. If true, 1372 * pc->mem_cgroup may be overwritten. 1373 * 1374 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or 1375 * under hierarchy of moving cgroups. This is for 1376 * waiting at hith-memory prressure caused by "move". 1377 */ 1378 1379static bool mem_cgroup_stealed(struct mem_cgroup *mem) 1380{ 1381 VM_BUG_ON(!rcu_read_lock_held()); 1382 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0; 1383} 1384 1385static bool mem_cgroup_under_move(struct mem_cgroup *mem) 1386{ 1387 struct mem_cgroup *from; 1388 struct mem_cgroup *to; 1389 bool ret = false; 1390 /* 1391 * Unlike task_move routines, we access mc.to, mc.from not under 1392 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead. 1393 */ 1394 spin_lock(&mc.lock); 1395 from = mc.from; 1396 to = mc.to; 1397 if (!from) 1398 goto unlock; 1399 if (from == mem || to == mem 1400 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css)) 1401 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css))) 1402 ret = true; 1403unlock: 1404 spin_unlock(&mc.lock); 1405 return ret; 1406} 1407 1408static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem) 1409{ 1410 if (mc.moving_task && current != mc.moving_task) { 1411 if (mem_cgroup_under_move(mem)) { 1412 DEFINE_WAIT(wait); 1413 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE); 1414 /* moving charge context might have finished. */ 1415 if (mc.moving_task) 1416 schedule(); 1417 finish_wait(&mc.waitq, &wait); 1418 return true; 1419 } 1420 } 1421 return false; 1422} 1423 1424/** 1425 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode. 1426 * @memcg: The memory cgroup that went over limit 1427 * @p: Task that is going to be killed 1428 * 1429 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 1430 * enabled 1431 */ 1432void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 1433{ 1434 struct cgroup *task_cgrp; 1435 struct cgroup *mem_cgrp; 1436 /* 1437 * Need a buffer in BSS, can't rely on allocations. The code relies 1438 * on the assumption that OOM is serialized for memory controller. 1439 * If this assumption is broken, revisit this code. 1440 */ 1441 static char memcg_name[PATH_MAX]; 1442 int ret; 1443 1444 if (!memcg || !p) 1445 return; 1446 1447 1448 rcu_read_lock(); 1449 1450 mem_cgrp = memcg->css.cgroup; 1451 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 1452 1453 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 1454 if (ret < 0) { 1455 /* 1456 * Unfortunately, we are unable to convert to a useful name 1457 * But we'll still print out the usage information 1458 */ 1459 rcu_read_unlock(); 1460 goto done; 1461 } 1462 rcu_read_unlock(); 1463 1464 printk(KERN_INFO "Task in %s killed", memcg_name); 1465 1466 rcu_read_lock(); 1467 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 1468 if (ret < 0) { 1469 rcu_read_unlock(); 1470 goto done; 1471 } 1472 rcu_read_unlock(); 1473 1474 /* 1475 * Continues from above, so we don't need an KERN_ level 1476 */ 1477 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 1478done: 1479 1480 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 1481 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 1482 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 1483 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 1484 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 1485 "failcnt %llu\n", 1486 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 1487 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 1488 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 1489} 1490 1491/* 1492 * This function returns the number of memcg under hierarchy tree. Returns 1493 * 1(self count) if no children. 1494 */ 1495static int mem_cgroup_count_children(struct mem_cgroup *mem) 1496{ 1497 int num = 0; 1498 struct mem_cgroup *iter; 1499 1500 for_each_mem_cgroup_tree(iter, mem) 1501 num++; 1502 return num; 1503} 1504 1505/* 1506 * Return the memory (and swap, if configured) limit for a memcg. 1507 */ 1508u64 mem_cgroup_get_limit(struct mem_cgroup *memcg) 1509{ 1510 u64 limit; 1511 u64 memsw; 1512 1513 limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1514 limit += total_swap_pages << PAGE_SHIFT; 1515 1516 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1517 /* 1518 * If memsw is finite and limits the amount of swap space available 1519 * to this memcg, return that limit. 1520 */ 1521 return min(limit, memsw); 1522} 1523 1524/* 1525 * Visit the first child (need not be the first child as per the ordering 1526 * of the cgroup list, since we track last_scanned_child) of @mem and use 1527 * that to reclaim free pages from. 1528 */ 1529static struct mem_cgroup * 1530mem_cgroup_select_victim(struct mem_cgroup *root_mem) 1531{ 1532 struct mem_cgroup *ret = NULL; 1533 struct cgroup_subsys_state *css; 1534 int nextid, found; 1535 1536 if (!root_mem->use_hierarchy) { 1537 css_get(&root_mem->css); 1538 ret = root_mem; 1539 } 1540 1541 while (!ret) { 1542 rcu_read_lock(); 1543 nextid = root_mem->last_scanned_child + 1; 1544 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, 1545 &found); 1546 if (css && css_tryget(css)) 1547 ret = container_of(css, struct mem_cgroup, css); 1548 1549 rcu_read_unlock(); 1550 /* Updates scanning parameter */ 1551 if (!css) { 1552 /* this means start scan from ID:1 */ 1553 root_mem->last_scanned_child = 0; 1554 } else 1555 root_mem->last_scanned_child = found; 1556 } 1557 1558 return ret; 1559} 1560 1561#if MAX_NUMNODES > 1 1562 1563/* 1564 * Always updating the nodemask is not very good - even if we have an empty 1565 * list or the wrong list here, we can start from some node and traverse all 1566 * nodes based on the zonelist. So update the list loosely once per 10 secs. 1567 * 1568 */ 1569static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem) 1570{ 1571 int nid; 1572 1573 if (time_after(mem->next_scan_node_update, jiffies)) 1574 return; 1575 1576 mem->next_scan_node_update = jiffies + 10*HZ; 1577 /* make a nodemask where this memcg uses memory from */ 1578 mem->scan_nodes = node_states[N_HIGH_MEMORY]; 1579 1580 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) { 1581 1582 if (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_FILE) || 1583 mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_FILE)) 1584 continue; 1585 1586 if (total_swap_pages && 1587 (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_ANON) || 1588 mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_ANON))) 1589 continue; 1590 node_clear(nid, mem->scan_nodes); 1591 } 1592} 1593 1594/* 1595 * Selecting a node where we start reclaim from. Because what we need is just 1596 * reducing usage counter, start from anywhere is O,K. Considering 1597 * memory reclaim from current node, there are pros. and cons. 1598 * 1599 * Freeing memory from current node means freeing memory from a node which 1600 * we'll use or we've used. So, it may make LRU bad. And if several threads 1601 * hit limits, it will see a contention on a node. But freeing from remote 1602 * node means more costs for memory reclaim because of memory latency. 1603 * 1604 * Now, we use round-robin. Better algorithm is welcomed. 1605 */ 1606int mem_cgroup_select_victim_node(struct mem_cgroup *mem) 1607{ 1608 int node; 1609 1610 mem_cgroup_may_update_nodemask(mem); 1611 node = mem->last_scanned_node; 1612 1613 node = next_node(node, mem->scan_nodes); 1614 if (node == MAX_NUMNODES) 1615 node = first_node(mem->scan_nodes); 1616 /* 1617 * We call this when we hit limit, not when pages are added to LRU. 1618 * No LRU may hold pages because all pages are UNEVICTABLE or 1619 * memcg is too small and all pages are not on LRU. In that case, 1620 * we use curret node. 1621 */ 1622 if (unlikely(node == MAX_NUMNODES)) 1623 node = numa_node_id(); 1624 1625 mem->last_scanned_node = node; 1626 return node; 1627} 1628 1629#else 1630int mem_cgroup_select_victim_node(struct mem_cgroup *mem) 1631{ 1632 return 0; 1633} 1634#endif 1635 1636/* 1637 * Scan the hierarchy if needed to reclaim memory. We remember the last child 1638 * we reclaimed from, so that we don't end up penalizing one child extensively 1639 * based on its position in the children list. 1640 * 1641 * root_mem is the original ancestor that we've been reclaim from. 1642 * 1643 * We give up and return to the caller when we visit root_mem twice. 1644 * (other groups can be removed while we're walking....) 1645 * 1646 * If shrink==true, for avoiding to free too much, this returns immedieately. 1647 */ 1648static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, 1649 struct zone *zone, 1650 gfp_t gfp_mask, 1651 unsigned long reclaim_options, 1652 unsigned long *total_scanned) 1653{ 1654 struct mem_cgroup *victim; 1655 int ret, total = 0; 1656 int loop = 0; 1657 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP; 1658 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK; 1659 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT; 1660 unsigned long excess; 1661 unsigned long nr_scanned; 1662 1663 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT; 1664 1665 /* If memsw_is_minimum==1, swap-out is of-no-use. */ 1666 if (root_mem->memsw_is_minimum) 1667 noswap = true; 1668 1669 while (1) { 1670 victim = mem_cgroup_select_victim(root_mem); 1671 if (victim == root_mem) { 1672 loop++; 1673 if (loop >= 1) 1674 drain_all_stock_async(); 1675 if (loop >= 2) { 1676 /* 1677 * If we have not been able to reclaim 1678 * anything, it might because there are 1679 * no reclaimable pages under this hierarchy 1680 */ 1681 if (!check_soft || !total) { 1682 css_put(&victim->css); 1683 break; 1684 } 1685 /* 1686 * We want to do more targeted reclaim. 1687 * excess >> 2 is not to excessive so as to 1688 * reclaim too much, nor too less that we keep 1689 * coming back to reclaim from this cgroup 1690 */ 1691 if (total >= (excess >> 2) || 1692 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) { 1693 css_put(&victim->css); 1694 break; 1695 } 1696 } 1697 } 1698 if (!mem_cgroup_local_usage(victim)) { 1699 /* this cgroup's local usage == 0 */ 1700 css_put(&victim->css); 1701 continue; 1702 } 1703 /* we use swappiness of local cgroup */ 1704 if (check_soft) { 1705 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask, 1706 noswap, get_swappiness(victim), zone, 1707 &nr_scanned); 1708 *total_scanned += nr_scanned; 1709 } else 1710 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, 1711 noswap, get_swappiness(victim)); 1712 css_put(&victim->css); 1713 /* 1714 * At shrinking usage, we can't check we should stop here or 1715 * reclaim more. It's depends on callers. last_scanned_child 1716 * will work enough for keeping fairness under tree. 1717 */ 1718 if (shrink) 1719 return ret; 1720 total += ret; 1721 if (check_soft) { 1722 if (!res_counter_soft_limit_excess(&root_mem->res)) 1723 return total; 1724 } else if (mem_cgroup_margin(root_mem)) 1725 return total; 1726 } 1727 return total; 1728} 1729 1730/* 1731 * Check OOM-Killer is already running under our hierarchy. 1732 * If someone is running, return false. 1733 */ 1734static bool mem_cgroup_oom_lock(struct mem_cgroup *mem) 1735{ 1736 int x, lock_count = 0; 1737 struct mem_cgroup *iter; 1738 1739 for_each_mem_cgroup_tree(iter, mem) { 1740 x = atomic_inc_return(&iter->oom_lock); 1741 lock_count = max(x, lock_count); 1742 } 1743 1744 if (lock_count == 1) 1745 return true; 1746 return false; 1747} 1748 1749static int mem_cgroup_oom_unlock(struct mem_cgroup *mem) 1750{ 1751 struct mem_cgroup *iter; 1752 1753 /* 1754 * When a new child is created while the hierarchy is under oom, 1755 * mem_cgroup_oom_lock() may not be called. We have to use 1756 * atomic_add_unless() here. 1757 */ 1758 for_each_mem_cgroup_tree(iter, mem) 1759 atomic_add_unless(&iter->oom_lock, -1, 0); 1760 return 0; 1761} 1762 1763 1764static DEFINE_MUTEX(memcg_oom_mutex); 1765static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq); 1766 1767struct oom_wait_info { 1768 struct mem_cgroup *mem; 1769 wait_queue_t wait; 1770}; 1771 1772static int memcg_oom_wake_function(wait_queue_t *wait, 1773 unsigned mode, int sync, void *arg) 1774{ 1775 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg; 1776 struct oom_wait_info *oom_wait_info; 1777 1778 oom_wait_info = container_of(wait, struct oom_wait_info, wait); 1779 1780 if (oom_wait_info->mem == wake_mem) 1781 goto wakeup; 1782 /* if no hierarchy, no match */ 1783 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy) 1784 return 0; 1785 /* 1786 * Both of oom_wait_info->mem and wake_mem are stable under us. 1787 * Then we can use css_is_ancestor without taking care of RCU. 1788 */ 1789 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) && 1790 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css)) 1791 return 0; 1792 1793wakeup: 1794 return autoremove_wake_function(wait, mode, sync, arg); 1795} 1796 1797static void memcg_wakeup_oom(struct mem_cgroup *mem) 1798{ 1799 /* for filtering, pass "mem" as argument. */ 1800 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem); 1801} 1802 1803static void memcg_oom_recover(struct mem_cgroup *mem) 1804{ 1805 if (mem && atomic_read(&mem->oom_lock)) 1806 memcg_wakeup_oom(mem); 1807} 1808 1809/* 1810 * try to call OOM killer. returns false if we should exit memory-reclaim loop. 1811 */ 1812bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask) 1813{ 1814 struct oom_wait_info owait; 1815 bool locked, need_to_kill; 1816 1817 owait.mem = mem; 1818 owait.wait.flags = 0; 1819 owait.wait.func = memcg_oom_wake_function; 1820 owait.wait.private = current; 1821 INIT_LIST_HEAD(&owait.wait.task_list); 1822 need_to_kill = true; 1823 /* At first, try to OOM lock hierarchy under mem.*/ 1824 mutex_lock(&memcg_oom_mutex); 1825 locked = mem_cgroup_oom_lock(mem); 1826 /* 1827 * Even if signal_pending(), we can't quit charge() loop without 1828 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL 1829 * under OOM is always welcomed, use TASK_KILLABLE here. 1830 */ 1831 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE); 1832 if (!locked || mem->oom_kill_disable) 1833 need_to_kill = false; 1834 if (locked) 1835 mem_cgroup_oom_notify(mem); 1836 mutex_unlock(&memcg_oom_mutex); 1837 1838 if (need_to_kill) { 1839 finish_wait(&memcg_oom_waitq, &owait.wait); 1840 mem_cgroup_out_of_memory(mem, mask); 1841 } else { 1842 schedule(); 1843 finish_wait(&memcg_oom_waitq, &owait.wait); 1844 } 1845 mutex_lock(&memcg_oom_mutex); 1846 mem_cgroup_oom_unlock(mem); 1847 memcg_wakeup_oom(mem); 1848 mutex_unlock(&memcg_oom_mutex); 1849 1850 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current)) 1851 return false; 1852 /* Give chance to dying process */ 1853 schedule_timeout(1); 1854 return true; 1855} 1856 1857/* 1858 * Currently used to update mapped file statistics, but the routine can be 1859 * generalized to update other statistics as well. 1860 * 1861 * Notes: Race condition 1862 * 1863 * We usually use page_cgroup_lock() for accessing page_cgroup member but 1864 * it tends to be costly. But considering some conditions, we doesn't need 1865 * to do so _always_. 1866 * 1867 * Considering "charge", lock_page_cgroup() is not required because all 1868 * file-stat operations happen after a page is attached to radix-tree. There 1869 * are no race with "charge". 1870 * 1871 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup 1872 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even 1873 * if there are race with "uncharge". Statistics itself is properly handled 1874 * by flags. 1875 * 1876 * Considering "move", this is an only case we see a race. To make the race 1877 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are 1878 * possibility of race condition. If there is, we take a lock. 1879 */ 1880 1881void mem_cgroup_update_page_stat(struct page *page, 1882 enum mem_cgroup_page_stat_item idx, int val) 1883{ 1884 struct mem_cgroup *mem; 1885 struct page_cgroup *pc = lookup_page_cgroup(page); 1886 bool need_unlock = false; 1887 unsigned long uninitialized_var(flags); 1888 1889 if (unlikely(!pc)) 1890 return; 1891 1892 rcu_read_lock(); 1893 mem = pc->mem_cgroup; 1894 if (unlikely(!mem || !PageCgroupUsed(pc))) 1895 goto out; 1896 /* pc->mem_cgroup is unstable ? */ 1897 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) { 1898 /* take a lock against to access pc->mem_cgroup */ 1899 move_lock_page_cgroup(pc, &flags); 1900 need_unlock = true; 1901 mem = pc->mem_cgroup; 1902 if (!mem || !PageCgroupUsed(pc)) 1903 goto out; 1904 } 1905 1906 switch (idx) { 1907 case MEMCG_NR_FILE_MAPPED: 1908 if (val > 0) 1909 SetPageCgroupFileMapped(pc); 1910 else if (!page_mapped(page)) 1911 ClearPageCgroupFileMapped(pc); 1912 idx = MEM_CGROUP_STAT_FILE_MAPPED; 1913 break; 1914 default: 1915 BUG(); 1916 } 1917 1918 this_cpu_add(mem->stat->count[idx], val); 1919 1920out: 1921 if (unlikely(need_unlock)) 1922 move_unlock_page_cgroup(pc, &flags); 1923 rcu_read_unlock(); 1924 return; 1925} 1926EXPORT_SYMBOL(mem_cgroup_update_page_stat); 1927 1928/* 1929 * size of first charge trial. "32" comes from vmscan.c's magic value. 1930 * TODO: maybe necessary to use big numbers in big irons. 1931 */ 1932#define CHARGE_BATCH 32U 1933struct memcg_stock_pcp { 1934 struct mem_cgroup *cached; /* this never be root cgroup */ 1935 unsigned int nr_pages; 1936 struct work_struct work; 1937}; 1938static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock); 1939static atomic_t memcg_drain_count; 1940 1941/* 1942 * Try to consume stocked charge on this cpu. If success, one page is consumed 1943 * from local stock and true is returned. If the stock is 0 or charges from a 1944 * cgroup which is not current target, returns false. This stock will be 1945 * refilled. 1946 */ 1947static bool consume_stock(struct mem_cgroup *mem) 1948{ 1949 struct memcg_stock_pcp *stock; 1950 bool ret = true; 1951 1952 stock = &get_cpu_var(memcg_stock); 1953 if (mem == stock->cached && stock->nr_pages) 1954 stock->nr_pages--; 1955 else /* need to call res_counter_charge */ 1956 ret = false; 1957 put_cpu_var(memcg_stock); 1958 return ret; 1959} 1960 1961/* 1962 * Returns stocks cached in percpu to res_counter and reset cached information. 1963 */ 1964static void drain_stock(struct memcg_stock_pcp *stock) 1965{ 1966 struct mem_cgroup *old = stock->cached; 1967 1968 if (stock->nr_pages) { 1969 unsigned long bytes = stock->nr_pages * PAGE_SIZE; 1970 1971 res_counter_uncharge(&old->res, bytes); 1972 if (do_swap_account) 1973 res_counter_uncharge(&old->memsw, bytes); 1974 stock->nr_pages = 0; 1975 } 1976 stock->cached = NULL; 1977} 1978 1979/* 1980 * This must be called under preempt disabled or must be called by 1981 * a thread which is pinned to local cpu. 1982 */ 1983static void drain_local_stock(struct work_struct *dummy) 1984{ 1985 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock); 1986 drain_stock(stock); 1987} 1988 1989/* 1990 * Cache charges(val) which is from res_counter, to local per_cpu area. 1991 * This will be consumed by consume_stock() function, later. 1992 */ 1993static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages) 1994{ 1995 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock); 1996 1997 if (stock->cached != mem) { /* reset if necessary */ 1998 drain_stock(stock); 1999 stock->cached = mem; 2000 } 2001 stock->nr_pages += nr_pages; 2002 put_cpu_var(memcg_stock); 2003} 2004 2005/* 2006 * Tries to drain stocked charges in other cpus. This function is asynchronous 2007 * and just put a work per cpu for draining localy on each cpu. Caller can 2008 * expects some charges will be back to res_counter later but cannot wait for 2009 * it. 2010 */ 2011static void drain_all_stock_async(void) 2012{ 2013 int cpu; 2014 /* This function is for scheduling "drain" in asynchronous way. 2015 * The result of "drain" is not directly handled by callers. Then, 2016 * if someone is calling drain, we don't have to call drain more. 2017 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if 2018 * there is a race. We just do loose check here. 2019 */ 2020 if (atomic_read(&memcg_drain_count)) 2021 return; 2022 /* Notify other cpus that system-wide "drain" is running */ 2023 atomic_inc(&memcg_drain_count); 2024 get_online_cpus(); 2025 for_each_online_cpu(cpu) { 2026 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu); 2027 schedule_work_on(cpu, &stock->work); 2028 } 2029 put_online_cpus(); 2030 atomic_dec(&memcg_drain_count); 2031 /* We don't wait for flush_work */ 2032} 2033 2034/* This is a synchronous drain interface. */ 2035static void drain_all_stock_sync(void) 2036{ 2037 /* called when force_empty is called */ 2038 atomic_inc(&memcg_drain_count); 2039 schedule_on_each_cpu(drain_local_stock); 2040 atomic_dec(&memcg_drain_count); 2041} 2042 2043/* 2044 * This function drains percpu counter value from DEAD cpu and 2045 * move it to local cpu. Note that this function can be preempted. 2046 */ 2047static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu) 2048{ 2049 int i; 2050 2051 spin_lock(&mem->pcp_counter_lock); 2052 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) { 2053 long x = per_cpu(mem->stat->count[i], cpu); 2054 2055 per_cpu(mem->stat->count[i], cpu) = 0; 2056 mem->nocpu_base.count[i] += x; 2057 } 2058 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) { 2059 unsigned long x = per_cpu(mem->stat->events[i], cpu); 2060 2061 per_cpu(mem->stat->events[i], cpu) = 0; 2062 mem->nocpu_base.events[i] += x; 2063 } 2064 /* need to clear ON_MOVE value, works as a kind of lock. */ 2065 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0; 2066 spin_unlock(&mem->pcp_counter_lock); 2067} 2068 2069static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu) 2070{ 2071 int idx = MEM_CGROUP_ON_MOVE; 2072 2073 spin_lock(&mem->pcp_counter_lock); 2074 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx]; 2075 spin_unlock(&mem->pcp_counter_lock); 2076} 2077 2078static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb, 2079 unsigned long action, 2080 void *hcpu) 2081{ 2082 int cpu = (unsigned long)hcpu; 2083 struct memcg_stock_pcp *stock; 2084 struct mem_cgroup *iter; 2085 2086 if ((action == CPU_ONLINE)) { 2087 for_each_mem_cgroup_all(iter) 2088 synchronize_mem_cgroup_on_move(iter, cpu); 2089 return NOTIFY_OK; 2090 } 2091 2092 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN) 2093 return NOTIFY_OK; 2094 2095 for_each_mem_cgroup_all(iter) 2096 mem_cgroup_drain_pcp_counter(iter, cpu); 2097 2098 stock = &per_cpu(memcg_stock, cpu); 2099 drain_stock(stock); 2100 return NOTIFY_OK; 2101} 2102 2103 2104/* See __mem_cgroup_try_charge() for details */ 2105enum { 2106 CHARGE_OK, /* success */ 2107 CHARGE_RETRY, /* need to retry but retry is not bad */ 2108 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */ 2109 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */ 2110 CHARGE_OOM_DIE, /* the current is killed because of OOM */ 2111}; 2112 2113static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask, 2114 unsigned int nr_pages, bool oom_check) 2115{ 2116 unsigned long csize = nr_pages * PAGE_SIZE; 2117 struct mem_cgroup *mem_over_limit; 2118 struct res_counter *fail_res; 2119 unsigned long flags = 0; 2120 int ret; 2121 2122 ret = res_counter_charge(&mem->res, csize, &fail_res); 2123 2124 if (likely(!ret)) { 2125 if (!do_swap_account) 2126 return CHARGE_OK; 2127 ret = res_counter_charge(&mem->memsw, csize, &fail_res); 2128 if (likely(!ret)) 2129 return CHARGE_OK; 2130 2131 res_counter_uncharge(&mem->res, csize); 2132 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw); 2133 flags |= MEM_CGROUP_RECLAIM_NOSWAP; 2134 } else 2135 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res); 2136 /* 2137 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch 2138 * of regular pages (CHARGE_BATCH), or a single regular page (1). 2139 * 2140 * Never reclaim on behalf of optional batching, retry with a 2141 * single page instead. 2142 */ 2143 if (nr_pages == CHARGE_BATCH) 2144 return CHARGE_RETRY; 2145 2146 if (!(gfp_mask & __GFP_WAIT)) 2147 return CHARGE_WOULDBLOCK; 2148 2149 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL, 2150 gfp_mask, flags, NULL); 2151 if (mem_cgroup_margin(mem_over_limit) >= nr_pages) 2152 return CHARGE_RETRY; 2153 /* 2154 * Even though the limit is exceeded at this point, reclaim 2155 * may have been able to free some pages. Retry the charge 2156 * before killing the task. 2157 * 2158 * Only for regular pages, though: huge pages are rather 2159 * unlikely to succeed so close to the limit, and we fall back 2160 * to regular pages anyway in case of failure. 2161 */ 2162 if (nr_pages == 1 && ret) 2163 return CHARGE_RETRY; 2164 2165 /* 2166 * At task move, charge accounts can be doubly counted. So, it's 2167 * better to wait until the end of task_move if something is going on. 2168 */ 2169 if (mem_cgroup_wait_acct_move(mem_over_limit)) 2170 return CHARGE_RETRY; 2171 2172 /* If we don't need to call oom-killer at el, return immediately */ 2173 if (!oom_check) 2174 return CHARGE_NOMEM; 2175 /* check OOM */ 2176 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) 2177 return CHARGE_OOM_DIE; 2178 2179 return CHARGE_RETRY; 2180} 2181 2182/* 2183 * Unlike exported interface, "oom" parameter is added. if oom==true, 2184 * oom-killer can be invoked. 2185 */ 2186static int __mem_cgroup_try_charge(struct mm_struct *mm, 2187 gfp_t gfp_mask, 2188 unsigned int nr_pages, 2189 struct mem_cgroup **memcg, 2190 bool oom) 2191{ 2192 unsigned int batch = max(CHARGE_BATCH, nr_pages); 2193 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2194 struct mem_cgroup *mem = NULL; 2195 int ret; 2196 2197 /* 2198 * Unlike gloval-vm's OOM-kill, we're not in memory shortage 2199 * in system level. So, allow to go ahead dying process in addition to 2200 * MEMDIE process. 2201 */ 2202 if (unlikely(test_thread_flag(TIF_MEMDIE) 2203 || fatal_signal_pending(current))) 2204 goto bypass; 2205 2206 /* 2207 * We always charge the cgroup the mm_struct belongs to. 2208 * The mm_struct's mem_cgroup changes on task migration if the 2209 * thread group leader migrates. It's possible that mm is not 2210 * set, if so charge the init_mm (happens for pagecache usage). 2211 */ 2212 if (!*memcg && !mm) 2213 goto bypass; 2214again: 2215 if (*memcg) { /* css should be a valid one */ 2216 mem = *memcg; 2217 VM_BUG_ON(css_is_removed(&mem->css)); 2218 if (mem_cgroup_is_root(mem)) 2219 goto done; 2220 if (nr_pages == 1 && consume_stock(mem)) 2221 goto done; 2222 css_get(&mem->css); 2223 } else { 2224 struct task_struct *p; 2225 2226 rcu_read_lock(); 2227 p = rcu_dereference(mm->owner); 2228 /* 2229 * Because we don't have task_lock(), "p" can exit. 2230 * In that case, "mem" can point to root or p can be NULL with 2231 * race with swapoff. Then, we have small risk of mis-accouning. 2232 * But such kind of mis-account by race always happens because 2233 * we don't have cgroup_mutex(). It's overkill and we allo that 2234 * small race, here. 2235 * (*) swapoff at el will charge against mm-struct not against 2236 * task-struct. So, mm->owner can be NULL. 2237 */ 2238 mem = mem_cgroup_from_task(p); 2239 if (!mem || mem_cgroup_is_root(mem)) { 2240 rcu_read_unlock(); 2241 goto done; 2242 } 2243 if (nr_pages == 1 && consume_stock(mem)) { 2244 /* 2245 * It seems dagerous to access memcg without css_get(). 2246 * But considering how consume_stok works, it's not 2247 * necessary. If consume_stock success, some charges 2248 * from this memcg are cached on this cpu. So, we 2249 * don't need to call css_get()/css_tryget() before 2250 * calling consume_stock(). 2251 */ 2252 rcu_read_unlock(); 2253 goto done; 2254 } 2255 /* after here, we may be blocked. we need to get refcnt */ 2256 if (!css_tryget(&mem->css)) { 2257 rcu_read_unlock(); 2258 goto again; 2259 } 2260 rcu_read_unlock(); 2261 } 2262 2263 do { 2264 bool oom_check; 2265 2266 /* If killed, bypass charge */ 2267 if (fatal_signal_pending(current)) { 2268 css_put(&mem->css); 2269 goto bypass; 2270 } 2271 2272 oom_check = false; 2273 if (oom && !nr_oom_retries) { 2274 oom_check = true; 2275 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES; 2276 } 2277 2278 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check); 2279 switch (ret) { 2280 case CHARGE_OK: 2281 break; 2282 case CHARGE_RETRY: /* not in OOM situation but retry */ 2283 batch = nr_pages; 2284 css_put(&mem->css); 2285 mem = NULL; 2286 goto again; 2287 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */ 2288 css_put(&mem->css); 2289 goto nomem; 2290 case CHARGE_NOMEM: /* OOM routine works */ 2291 if (!oom) { 2292 css_put(&mem->css); 2293 goto nomem; 2294 } 2295 /* If oom, we never return -ENOMEM */ 2296 nr_oom_retries--; 2297 break; 2298 case CHARGE_OOM_DIE: /* Killed by OOM Killer */ 2299 css_put(&mem->css); 2300 goto bypass; 2301 } 2302 } while (ret != CHARGE_OK); 2303 2304 if (batch > nr_pages) 2305 refill_stock(mem, batch - nr_pages); 2306 css_put(&mem->css); 2307done: 2308 *memcg = mem; 2309 return 0; 2310nomem: 2311 *memcg = NULL; 2312 return -ENOMEM; 2313bypass: 2314 *memcg = NULL; 2315 return 0; 2316} 2317 2318/* 2319 * Somemtimes we have to undo a charge we got by try_charge(). 2320 * This function is for that and do uncharge, put css's refcnt. 2321 * gotten by try_charge(). 2322 */ 2323static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem, 2324 unsigned int nr_pages) 2325{ 2326 if (!mem_cgroup_is_root(mem)) { 2327 unsigned long bytes = nr_pages * PAGE_SIZE; 2328 2329 res_counter_uncharge(&mem->res, bytes); 2330 if (do_swap_account) 2331 res_counter_uncharge(&mem->memsw, bytes); 2332 } 2333} 2334 2335/* 2336 * A helper function to get mem_cgroup from ID. must be called under 2337 * rcu_read_lock(). The caller must check css_is_removed() or some if 2338 * it's concern. (dropping refcnt from swap can be called against removed 2339 * memcg.) 2340 */ 2341static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 2342{ 2343 struct cgroup_subsys_state *css; 2344 2345 /* ID 0 is unused ID */ 2346 if (!id) 2347 return NULL; 2348 css = css_lookup(&mem_cgroup_subsys, id); 2349 if (!css) 2350 return NULL; 2351 return container_of(css, struct mem_cgroup, css); 2352} 2353 2354struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page) 2355{ 2356 struct mem_cgroup *mem = NULL; 2357 struct page_cgroup *pc; 2358 unsigned short id; 2359 swp_entry_t ent; 2360 2361 VM_BUG_ON(!PageLocked(page)); 2362 2363 pc = lookup_page_cgroup(page); 2364 lock_page_cgroup(pc); 2365 if (PageCgroupUsed(pc)) { 2366 mem = pc->mem_cgroup; 2367 if (mem && !css_tryget(&mem->css)) 2368 mem = NULL; 2369 } else if (PageSwapCache(page)) { 2370 ent.val = page_private(page); 2371 id = lookup_swap_cgroup(ent); 2372 rcu_read_lock(); 2373 mem = mem_cgroup_lookup(id); 2374 if (mem && !css_tryget(&mem->css)) 2375 mem = NULL; 2376 rcu_read_unlock(); 2377 } 2378 unlock_page_cgroup(pc); 2379 return mem; 2380} 2381 2382static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, 2383 struct page *page, 2384 unsigned int nr_pages, 2385 struct page_cgroup *pc, 2386 enum charge_type ctype) 2387{ 2388 lock_page_cgroup(pc); 2389 if (unlikely(PageCgroupUsed(pc))) { 2390 unlock_page_cgroup(pc); 2391 __mem_cgroup_cancel_charge(mem, nr_pages); 2392 return; 2393 } 2394 /* 2395 * we don't need page_cgroup_lock about tail pages, becase they are not 2396 * accessed by any other context at this point. 2397 */ 2398 pc->mem_cgroup = mem; 2399 /* 2400 * We access a page_cgroup asynchronously without lock_page_cgroup(). 2401 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 2402 * is accessed after testing USED bit. To make pc->mem_cgroup visible 2403 * before USED bit, we need memory barrier here. 2404 * See mem_cgroup_add_lru_list(), etc. 2405 */ 2406 smp_wmb(); 2407 switch (ctype) { 2408 case MEM_CGROUP_CHARGE_TYPE_CACHE: 2409 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 2410 SetPageCgroupCache(pc); 2411 SetPageCgroupUsed(pc); 2412 break; 2413 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2414 ClearPageCgroupCache(pc); 2415 SetPageCgroupUsed(pc); 2416 break; 2417 default: 2418 break; 2419 } 2420 2421 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages); 2422 unlock_page_cgroup(pc); 2423 /* 2424 * "charge_statistics" updated event counter. Then, check it. 2425 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree. 2426 * if they exceeds softlimit. 2427 */ 2428 memcg_check_events(mem, page); 2429} 2430 2431#ifdef CONFIG_TRANSPARENT_HUGEPAGE 2432 2433#define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\ 2434 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION)) 2435/* 2436 * Because tail pages are not marked as "used", set it. We're under 2437 * zone->lru_lock, 'splitting on pmd' and compund_lock. 2438 */ 2439void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail) 2440{ 2441 struct page_cgroup *head_pc = lookup_page_cgroup(head); 2442 struct page_cgroup *tail_pc = lookup_page_cgroup(tail); 2443 unsigned long flags; 2444 2445 if (mem_cgroup_disabled()) 2446 return; 2447 /* 2448 * We have no races with charge/uncharge but will have races with 2449 * page state accounting. 2450 */ 2451 move_lock_page_cgroup(head_pc, &flags); 2452 2453 tail_pc->mem_cgroup = head_pc->mem_cgroup; 2454 smp_wmb(); /* see __commit_charge() */ 2455 if (PageCgroupAcctLRU(head_pc)) { 2456 enum lru_list lru; 2457 struct mem_cgroup_per_zone *mz; 2458 2459 /* 2460 * LRU flags cannot be copied because we need to add tail 2461 *.page to LRU by generic call and our hook will be called. 2462 * We hold lru_lock, then, reduce counter directly. 2463 */ 2464 lru = page_lru(head); 2465 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head); 2466 MEM_CGROUP_ZSTAT(mz, lru) -= 1; 2467 } 2468 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT; 2469 move_unlock_page_cgroup(head_pc, &flags); 2470} 2471#endif 2472 2473/** 2474 * mem_cgroup_move_account - move account of the page 2475 * @page: the page 2476 * @nr_pages: number of regular pages (>1 for huge pages) 2477 * @pc: page_cgroup of the page. 2478 * @from: mem_cgroup which the page is moved from. 2479 * @to: mem_cgroup which the page is moved to. @from != @to. 2480 * @uncharge: whether we should call uncharge and css_put against @from. 2481 * 2482 * The caller must confirm following. 2483 * - page is not on LRU (isolate_page() is useful.) 2484 * - compound_lock is held when nr_pages > 1 2485 * 2486 * This function doesn't do "charge" nor css_get to new cgroup. It should be 2487 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is 2488 * true, this function does "uncharge" from old cgroup, but it doesn't if 2489 * @uncharge is false, so a caller should do "uncharge". 2490 */ 2491static int mem_cgroup_move_account(struct page *page, 2492 unsigned int nr_pages, 2493 struct page_cgroup *pc, 2494 struct mem_cgroup *from, 2495 struct mem_cgroup *to, 2496 bool uncharge) 2497{ 2498 unsigned long flags; 2499 int ret; 2500 2501 VM_BUG_ON(from == to); 2502 VM_BUG_ON(PageLRU(page)); 2503 /* 2504 * The page is isolated from LRU. So, collapse function 2505 * will not handle this page. But page splitting can happen. 2506 * Do this check under compound_page_lock(). The caller should 2507 * hold it. 2508 */ 2509 ret = -EBUSY; 2510 if (nr_pages > 1 && !PageTransHuge(page)) 2511 goto out; 2512 2513 lock_page_cgroup(pc); 2514 2515 ret = -EINVAL; 2516 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from) 2517 goto unlock; 2518 2519 move_lock_page_cgroup(pc, &flags); 2520 2521 if (PageCgroupFileMapped(pc)) { 2522 /* Update mapped_file data for mem_cgroup */ 2523 preempt_disable(); 2524 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2525 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]); 2526 preempt_enable(); 2527 } 2528 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages); 2529 if (uncharge) 2530 /* This is not "cancel", but cancel_charge does all we need. */ 2531 __mem_cgroup_cancel_charge(from, nr_pages); 2532 2533 /* caller should have done css_get */ 2534 pc->mem_cgroup = to; 2535 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages); 2536 /* 2537 * We charges against "to" which may not have any tasks. Then, "to" 2538 * can be under rmdir(). But in current implementation, caller of 2539 * this function is just force_empty() and move charge, so it's 2540 * guaranteed that "to" is never removed. So, we don't check rmdir 2541 * status here. 2542 */ 2543 move_unlock_page_cgroup(pc, &flags); 2544 ret = 0; 2545unlock: 2546 unlock_page_cgroup(pc); 2547 /* 2548 * check events 2549 */ 2550 memcg_check_events(to, page); 2551 memcg_check_events(from, page); 2552out: 2553 return ret; 2554} 2555 2556/* 2557 * move charges to its parent. 2558 */ 2559 2560static int mem_cgroup_move_parent(struct page *page, 2561 struct page_cgroup *pc, 2562 struct mem_cgroup *child, 2563 gfp_t gfp_mask) 2564{ 2565 struct cgroup *cg = child->css.cgroup; 2566 struct cgroup *pcg = cg->parent; 2567 struct mem_cgroup *parent; 2568 unsigned int nr_pages; 2569 unsigned long uninitialized_var(flags); 2570 int ret; 2571 2572 /* Is ROOT ? */ 2573 if (!pcg) 2574 return -EINVAL; 2575 2576 ret = -EBUSY; 2577 if (!get_page_unless_zero(page)) 2578 goto out; 2579 if (isolate_lru_page(page)) 2580 goto put; 2581 2582 nr_pages = hpage_nr_pages(page); 2583 2584 parent = mem_cgroup_from_cont(pcg); 2585 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false); 2586 if (ret || !parent) 2587 goto put_back; 2588 2589 if (nr_pages > 1) 2590 flags = compound_lock_irqsave(page); 2591 2592 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true); 2593 if (ret) 2594 __mem_cgroup_cancel_charge(parent, nr_pages); 2595 2596 if (nr_pages > 1) 2597 compound_unlock_irqrestore(page, flags); 2598put_back: 2599 putback_lru_page(page); 2600put: 2601 put_page(page); 2602out: 2603 return ret; 2604} 2605 2606/* 2607 * Charge the memory controller for page usage. 2608 * Return 2609 * 0 if the charge was successful 2610 * < 0 if the cgroup is over its limit 2611 */ 2612static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 2613 gfp_t gfp_mask, enum charge_type ctype) 2614{ 2615 struct mem_cgroup *mem = NULL; 2616 unsigned int nr_pages = 1; 2617 struct page_cgroup *pc; 2618 bool oom = true; 2619 int ret; 2620 2621 if (PageTransHuge(page)) { 2622 nr_pages <<= compound_order(page); 2623 VM_BUG_ON(!PageTransHuge(page)); 2624 /* 2625 * Never OOM-kill a process for a huge page. The 2626 * fault handler will fall back to regular pages. 2627 */ 2628 oom = false; 2629 } 2630 2631 pc = lookup_page_cgroup(page); 2632 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */ 2633 2634 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom); 2635 if (ret || !mem) 2636 return ret; 2637 2638 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype); 2639 return 0; 2640} 2641 2642int mem_cgroup_newpage_charge(struct page *page, 2643 struct mm_struct *mm, gfp_t gfp_mask) 2644{ 2645 if (mem_cgroup_disabled()) 2646 return 0; 2647 /* 2648 * If already mapped, we don't have to account. 2649 * If page cache, page->mapping has address_space. 2650 * But page->mapping may have out-of-use anon_vma pointer, 2651 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 2652 * is NULL. 2653 */ 2654 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 2655 return 0; 2656 if (unlikely(!mm)) 2657 mm = &init_mm; 2658 return mem_cgroup_charge_common(page, mm, gfp_mask, 2659 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2660} 2661 2662static void 2663__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2664 enum charge_type ctype); 2665 2666static void 2667__mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem, 2668 enum charge_type ctype) 2669{ 2670 struct page_cgroup *pc = lookup_page_cgroup(page); 2671 /* 2672 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page 2673 * is already on LRU. It means the page may on some other page_cgroup's 2674 * LRU. Take care of it. 2675 */ 2676 mem_cgroup_lru_del_before_commit(page); 2677 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype); 2678 mem_cgroup_lru_add_after_commit(page); 2679 return; 2680} 2681 2682int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 2683 gfp_t gfp_mask) 2684{ 2685 struct mem_cgroup *mem = NULL; 2686 int ret; 2687 2688 if (mem_cgroup_disabled()) 2689 return 0; 2690 if (PageCompound(page)) 2691 return 0; 2692 /* 2693 * Corner case handling. This is called from add_to_page_cache() 2694 * in usual. But some FS (shmem) precharges this page before calling it 2695 * and call add_to_page_cache() with GFP_NOWAIT. 2696 * 2697 * For GFP_NOWAIT case, the page may be pre-charged before calling 2698 * add_to_page_cache(). (See shmem.c) check it here and avoid to call 2699 * charge twice. (It works but has to pay a bit larger cost.) 2700 * And when the page is SwapCache, it should take swap information 2701 * into account. This is under lock_page() now. 2702 */ 2703 if (!(gfp_mask & __GFP_WAIT)) { 2704 struct page_cgroup *pc; 2705 2706 pc = lookup_page_cgroup(page); 2707 if (!pc) 2708 return 0; 2709 lock_page_cgroup(pc); 2710 if (PageCgroupUsed(pc)) { 2711 unlock_page_cgroup(pc); 2712 return 0; 2713 } 2714 unlock_page_cgroup(pc); 2715 } 2716 2717 if (unlikely(!mm)) 2718 mm = &init_mm; 2719 2720 if (page_is_file_cache(page)) { 2721 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true); 2722 if (ret || !mem) 2723 return ret; 2724 2725 /* 2726 * FUSE reuses pages without going through the final 2727 * put that would remove them from the LRU list, make 2728 * sure that they get relinked properly. 2729 */ 2730 __mem_cgroup_commit_charge_lrucare(page, mem, 2731 MEM_CGROUP_CHARGE_TYPE_CACHE); 2732 return ret; 2733 } 2734 /* shmem */ 2735 if (PageSwapCache(page)) { 2736 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 2737 if (!ret) 2738 __mem_cgroup_commit_charge_swapin(page, mem, 2739 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2740 } else 2741 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 2742 MEM_CGROUP_CHARGE_TYPE_SHMEM); 2743 2744 return ret; 2745} 2746 2747/* 2748 * While swap-in, try_charge -> commit or cancel, the page is locked. 2749 * And when try_charge() successfully returns, one refcnt to memcg without 2750 * struct page_cgroup is acquired. This refcnt will be consumed by 2751 * "commit()" or removed by "cancel()" 2752 */ 2753int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 2754 struct page *page, 2755 gfp_t mask, struct mem_cgroup **ptr) 2756{ 2757 struct mem_cgroup *mem; 2758 int ret; 2759 2760 *ptr = NULL; 2761 2762 if (mem_cgroup_disabled()) 2763 return 0; 2764 2765 if (!do_swap_account) 2766 goto charge_cur_mm; 2767 /* 2768 * A racing thread's fault, or swapoff, may have already updated 2769 * the pte, and even removed page from swap cache: in those cases 2770 * do_swap_page()'s pte_same() test will fail; but there's also a 2771 * KSM case which does need to charge the page. 2772 */ 2773 if (!PageSwapCache(page)) 2774 goto charge_cur_mm; 2775 mem = try_get_mem_cgroup_from_page(page); 2776 if (!mem) 2777 goto charge_cur_mm; 2778 *ptr = mem; 2779 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true); 2780 css_put(&mem->css); 2781 return ret; 2782charge_cur_mm: 2783 if (unlikely(!mm)) 2784 mm = &init_mm; 2785 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true); 2786} 2787 2788static void 2789__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 2790 enum charge_type ctype) 2791{ 2792 if (mem_cgroup_disabled()) 2793 return; 2794 if (!ptr) 2795 return; 2796 cgroup_exclude_rmdir(&ptr->css); 2797 2798 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype); 2799 /* 2800 * Now swap is on-memory. This means this page may be 2801 * counted both as mem and swap....double count. 2802 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 2803 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 2804 * may call delete_from_swap_cache() before reach here. 2805 */ 2806 if (do_swap_account && PageSwapCache(page)) { 2807 swp_entry_t ent = {.val = page_private(page)}; 2808 unsigned short id; 2809 struct mem_cgroup *memcg; 2810 2811 id = swap_cgroup_record(ent, 0); 2812 rcu_read_lock(); 2813 memcg = mem_cgroup_lookup(id); 2814 if (memcg) { 2815 /* 2816 * This recorded memcg can be obsolete one. So, avoid 2817 * calling css_tryget 2818 */ 2819 if (!mem_cgroup_is_root(memcg)) 2820 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 2821 mem_cgroup_swap_statistics(memcg, false); 2822 mem_cgroup_put(memcg); 2823 } 2824 rcu_read_unlock(); 2825 } 2826 /* 2827 * At swapin, we may charge account against cgroup which has no tasks. 2828 * So, rmdir()->pre_destroy() can be called while we do this charge. 2829 * In that case, we need to call pre_destroy() again. check it here. 2830 */ 2831 cgroup_release_and_wakeup_rmdir(&ptr->css); 2832} 2833 2834void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) 2835{ 2836 __mem_cgroup_commit_charge_swapin(page, ptr, 2837 MEM_CGROUP_CHARGE_TYPE_MAPPED); 2838} 2839 2840void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) 2841{ 2842 if (mem_cgroup_disabled()) 2843 return; 2844 if (!mem) 2845 return; 2846 __mem_cgroup_cancel_charge(mem, 1); 2847} 2848 2849static void mem_cgroup_do_uncharge(struct mem_cgroup *mem, 2850 unsigned int nr_pages, 2851 const enum charge_type ctype) 2852{ 2853 struct memcg_batch_info *batch = NULL; 2854 bool uncharge_memsw = true; 2855 2856 /* If swapout, usage of swap doesn't decrease */ 2857 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 2858 uncharge_memsw = false; 2859 2860 batch = ¤t->memcg_batch; 2861 /* 2862 * In usual, we do css_get() when we remember memcg pointer. 2863 * But in this case, we keep res->usage until end of a series of 2864 * uncharges. Then, it's ok to ignore memcg's refcnt. 2865 */ 2866 if (!batch->memcg) 2867 batch->memcg = mem; 2868 /* 2869 * do_batch > 0 when unmapping pages or inode invalidate/truncate. 2870 * In those cases, all pages freed continuously can be expected to be in 2871 * the same cgroup and we have chance to coalesce uncharges. 2872 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE) 2873 * because we want to do uncharge as soon as possible. 2874 */ 2875 2876 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE)) 2877 goto direct_uncharge; 2878 2879 if (nr_pages > 1) 2880 goto direct_uncharge; 2881 2882 /* 2883 * In typical case, batch->memcg == mem. This means we can 2884 * merge a series of uncharges to an uncharge of res_counter. 2885 * If not, we uncharge res_counter ony by one. 2886 */ 2887 if (batch->memcg != mem) 2888 goto direct_uncharge; 2889 /* remember freed charge and uncharge it later */ 2890 batch->nr_pages++; 2891 if (uncharge_memsw) 2892 batch->memsw_nr_pages++; 2893 return; 2894direct_uncharge: 2895 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE); 2896 if (uncharge_memsw) 2897 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE); 2898 if (unlikely(batch->memcg != mem)) 2899 memcg_oom_recover(mem); 2900 return; 2901} 2902 2903/* 2904 * uncharge if !page_mapped(page) 2905 */ 2906static struct mem_cgroup * 2907__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 2908{ 2909 struct mem_cgroup *mem = NULL; 2910 unsigned int nr_pages = 1; 2911 struct page_cgroup *pc; 2912 2913 if (mem_cgroup_disabled()) 2914 return NULL; 2915 2916 if (PageSwapCache(page)) 2917 return NULL; 2918 2919 if (PageTransHuge(page)) { 2920 nr_pages <<= compound_order(page); 2921 VM_BUG_ON(!PageTransHuge(page)); 2922 } 2923 /* 2924 * Check if our page_cgroup is valid 2925 */ 2926 pc = lookup_page_cgroup(page); 2927 if (unlikely(!pc || !PageCgroupUsed(pc))) 2928 return NULL; 2929 2930 lock_page_cgroup(pc); 2931 2932 mem = pc->mem_cgroup; 2933 2934 if (!PageCgroupUsed(pc)) 2935 goto unlock_out; 2936 2937 switch (ctype) { 2938 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 2939 case MEM_CGROUP_CHARGE_TYPE_DROP: 2940 /* See mem_cgroup_prepare_migration() */ 2941 if (page_mapped(page) || PageCgroupMigration(pc)) 2942 goto unlock_out; 2943 break; 2944 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 2945 if (!PageAnon(page)) { /* Shared memory */ 2946 if (page->mapping && !page_is_file_cache(page)) 2947 goto unlock_out; 2948 } else if (page_mapped(page)) /* Anon */ 2949 goto unlock_out; 2950 break; 2951 default: 2952 break; 2953 } 2954 2955 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages); 2956 2957 ClearPageCgroupUsed(pc); 2958 /* 2959 * pc->mem_cgroup is not cleared here. It will be accessed when it's 2960 * freed from LRU. This is safe because uncharged page is expected not 2961 * to be reused (freed soon). Exception is SwapCache, it's handled by 2962 * special functions. 2963 */ 2964 2965 unlock_page_cgroup(pc); 2966 /* 2967 * even after unlock, we have mem->res.usage here and this memcg 2968 * will never be freed. 2969 */ 2970 memcg_check_events(mem, page); 2971 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) { 2972 mem_cgroup_swap_statistics(mem, true); 2973 mem_cgroup_get(mem); 2974 } 2975 if (!mem_cgroup_is_root(mem)) 2976 mem_cgroup_do_uncharge(mem, nr_pages, ctype); 2977 2978 return mem; 2979 2980unlock_out: 2981 unlock_page_cgroup(pc); 2982 return NULL; 2983} 2984 2985void mem_cgroup_uncharge_page(struct page *page) 2986{ 2987 /* early check. */ 2988 if (page_mapped(page)) 2989 return; 2990 if (page->mapping && !PageAnon(page)) 2991 return; 2992 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 2993} 2994 2995void mem_cgroup_uncharge_cache_page(struct page *page) 2996{ 2997 VM_BUG_ON(page_mapped(page)); 2998 VM_BUG_ON(page->mapping); 2999 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 3000} 3001 3002/* 3003 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate. 3004 * In that cases, pages are freed continuously and we can expect pages 3005 * are in the same memcg. All these calls itself limits the number of 3006 * pages freed at once, then uncharge_start/end() is called properly. 3007 * This may be called prural(2) times in a context, 3008 */ 3009 3010void mem_cgroup_uncharge_start(void) 3011{ 3012 current->memcg_batch.do_batch++; 3013 /* We can do nest. */ 3014 if (current->memcg_batch.do_batch == 1) { 3015 current->memcg_batch.memcg = NULL; 3016 current->memcg_batch.nr_pages = 0; 3017 current->memcg_batch.memsw_nr_pages = 0; 3018 } 3019} 3020 3021void mem_cgroup_uncharge_end(void) 3022{ 3023 struct memcg_batch_info *batch = ¤t->memcg_batch; 3024 3025 if (!batch->do_batch) 3026 return; 3027 3028 batch->do_batch--; 3029 if (batch->do_batch) /* If stacked, do nothing. */ 3030 return; 3031 3032 if (!batch->memcg) 3033 return; 3034 /* 3035 * This "batch->memcg" is valid without any css_get/put etc... 3036 * bacause we hide charges behind us. 3037 */ 3038 if (batch->nr_pages) 3039 res_counter_uncharge(&batch->memcg->res, 3040 batch->nr_pages * PAGE_SIZE); 3041 if (batch->memsw_nr_pages) 3042 res_counter_uncharge(&batch->memcg->memsw, 3043 batch->memsw_nr_pages * PAGE_SIZE); 3044 memcg_oom_recover(batch->memcg); 3045 /* forget this pointer (for sanity check) */ 3046 batch->memcg = NULL; 3047} 3048 3049#ifdef CONFIG_SWAP 3050/* 3051 * called after __delete_from_swap_cache() and drop "page" account. 3052 * memcg information is recorded to swap_cgroup of "ent" 3053 */ 3054void 3055mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 3056{ 3057 struct mem_cgroup *memcg; 3058 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 3059 3060 if (!swapout) /* this was a swap cache but the swap is unused ! */ 3061 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 3062 3063 memcg = __mem_cgroup_uncharge_common(page, ctype); 3064 3065 /* 3066 * record memcg information, if swapout && memcg != NULL, 3067 * mem_cgroup_get() was called in uncharge(). 3068 */ 3069 if (do_swap_account && swapout && memcg) 3070 swap_cgroup_record(ent, css_id(&memcg->css)); 3071} 3072#endif 3073 3074#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 3075/* 3076 * called from swap_entry_free(). remove record in swap_cgroup and 3077 * uncharge "memsw" account. 3078 */ 3079void mem_cgroup_uncharge_swap(swp_entry_t ent) 3080{ 3081 struct mem_cgroup *memcg; 3082 unsigned short id; 3083 3084 if (!do_swap_account) 3085 return; 3086 3087 id = swap_cgroup_record(ent, 0); 3088 rcu_read_lock(); 3089 memcg = mem_cgroup_lookup(id); 3090 if (memcg) { 3091 /* 3092 * We uncharge this because swap is freed. 3093 * This memcg can be obsolete one. We avoid calling css_tryget 3094 */ 3095 if (!mem_cgroup_is_root(memcg)) 3096 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 3097 mem_cgroup_swap_statistics(memcg, false); 3098 mem_cgroup_put(memcg); 3099 } 3100 rcu_read_unlock(); 3101} 3102 3103/** 3104 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record. 3105 * @entry: swap entry to be moved 3106 * @from: mem_cgroup which the entry is moved from 3107 * @to: mem_cgroup which the entry is moved to 3108 * @need_fixup: whether we should fixup res_counters and refcounts. 3109 * 3110 * It succeeds only when the swap_cgroup's record for this entry is the same 3111 * as the mem_cgroup's id of @from. 3112 * 3113 * Returns 0 on success, -EINVAL on failure. 3114 * 3115 * The caller must have charged to @to, IOW, called res_counter_charge() about 3116 * both res and memsw, and called css_get(). 3117 */ 3118static int mem_cgroup_move_swap_account(swp_entry_t entry, 3119 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3120{ 3121 unsigned short old_id, new_id; 3122 3123 old_id = css_id(&from->css); 3124 new_id = css_id(&to->css); 3125 3126 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) { 3127 mem_cgroup_swap_statistics(from, false); 3128 mem_cgroup_swap_statistics(to, true); 3129 /* 3130 * This function is only called from task migration context now. 3131 * It postpones res_counter and refcount handling till the end 3132 * of task migration(mem_cgroup_clear_mc()) for performance 3133 * improvement. But we cannot postpone mem_cgroup_get(to) 3134 * because if the process that has been moved to @to does 3135 * swap-in, the refcount of @to might be decreased to 0. 3136 */ 3137 mem_cgroup_get(to); 3138 if (need_fixup) { 3139 if (!mem_cgroup_is_root(from)) 3140 res_counter_uncharge(&from->memsw, PAGE_SIZE); 3141 mem_cgroup_put(from); 3142 /* 3143 * we charged both to->res and to->memsw, so we should 3144 * uncharge to->res. 3145 */ 3146 if (!mem_cgroup_is_root(to)) 3147 res_counter_uncharge(&to->res, PAGE_SIZE); 3148 } 3149 return 0; 3150 } 3151 return -EINVAL; 3152} 3153#else 3154static inline int mem_cgroup_move_swap_account(swp_entry_t entry, 3155 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup) 3156{ 3157 return -EINVAL; 3158} 3159#endif 3160 3161/* 3162 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 3163 * page belongs to. 3164 */ 3165int mem_cgroup_prepare_migration(struct page *page, 3166 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask) 3167{ 3168 struct mem_cgroup *mem = NULL; 3169 struct page_cgroup *pc; 3170 enum charge_type ctype; 3171 int ret = 0; 3172 3173 *ptr = NULL; 3174 3175 VM_BUG_ON(PageTransHuge(page)); 3176 if (mem_cgroup_disabled()) 3177 return 0; 3178 3179 pc = lookup_page_cgroup(page); 3180 lock_page_cgroup(pc); 3181 if (PageCgroupUsed(pc)) { 3182 mem = pc->mem_cgroup; 3183 css_get(&mem->css); 3184 /* 3185 * At migrating an anonymous page, its mapcount goes down 3186 * to 0 and uncharge() will be called. But, even if it's fully 3187 * unmapped, migration may fail and this page has to be 3188 * charged again. We set MIGRATION flag here and delay uncharge 3189 * until end_migration() is called 3190 * 3191 * Corner Case Thinking 3192 * A) 3193 * When the old page was mapped as Anon and it's unmap-and-freed 3194 * while migration was ongoing. 3195 * If unmap finds the old page, uncharge() of it will be delayed 3196 * until end_migration(). If unmap finds a new page, it's 3197 * uncharged when it make mapcount to be 1->0. If unmap code 3198 * finds swap_migration_entry, the new page will not be mapped 3199 * and end_migration() will find it(mapcount==0). 3200 * 3201 * B) 3202 * When the old page was mapped but migraion fails, the kernel 3203 * remaps it. A charge for it is kept by MIGRATION flag even 3204 * if mapcount goes down to 0. We can do remap successfully 3205 * without charging it again. 3206 * 3207 * C) 3208 * The "old" page is under lock_page() until the end of 3209 * migration, so, the old page itself will not be swapped-out. 3210 * If the new page is swapped out before end_migraton, our 3211 * hook to usual swap-out path will catch the event. 3212 */ 3213 if (PageAnon(page)) 3214 SetPageCgroupMigration(pc); 3215 } 3216 unlock_page_cgroup(pc); 3217 /* 3218 * If the page is not charged at this point, 3219 * we return here. 3220 */ 3221 if (!mem) 3222 return 0; 3223 3224 *ptr = mem; 3225 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false); 3226 css_put(&mem->css);/* drop extra refcnt */ 3227 if (ret || *ptr == NULL) { 3228 if (PageAnon(page)) { 3229 lock_page_cgroup(pc); 3230 ClearPageCgroupMigration(pc); 3231 unlock_page_cgroup(pc); 3232 /* 3233 * The old page may be fully unmapped while we kept it. 3234 */ 3235 mem_cgroup_uncharge_page(page); 3236 } 3237 return -ENOMEM; 3238 } 3239 /* 3240 * We charge new page before it's used/mapped. So, even if unlock_page() 3241 * is called before end_migration, we can catch all events on this new 3242 * page. In the case new page is migrated but not remapped, new page's 3243 * mapcount will be finally 0 and we call uncharge in end_migration(). 3244 */ 3245 pc = lookup_page_cgroup(newpage); 3246 if (PageAnon(page)) 3247 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 3248 else if (page_is_file_cache(page)) 3249 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 3250 else 3251 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 3252 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype); 3253 return ret; 3254} 3255 3256/* remove redundant charge if migration failed*/ 3257void mem_cgroup_end_migration(struct mem_cgroup *mem, 3258 struct page *oldpage, struct page *newpage, bool migration_ok) 3259{ 3260 struct page *used, *unused; 3261 struct page_cgroup *pc; 3262 3263 if (!mem) 3264 return; 3265 /* blocks rmdir() */ 3266 cgroup_exclude_rmdir(&mem->css); 3267 if (!migration_ok) { 3268 used = oldpage; 3269 unused = newpage; 3270 } else { 3271 used = newpage; 3272 unused = oldpage; 3273 } 3274 /* 3275 * We disallowed uncharge of pages under migration because mapcount 3276 * of the page goes down to zero, temporarly. 3277 * Clear the flag and check the page should be charged. 3278 */ 3279 pc = lookup_page_cgroup(oldpage); 3280 lock_page_cgroup(pc); 3281 ClearPageCgroupMigration(pc); 3282 unlock_page_cgroup(pc); 3283 3284 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE); 3285 3286 /* 3287 * If a page is a file cache, radix-tree replacement is very atomic 3288 * and we can skip this check. When it was an Anon page, its mapcount 3289 * goes down to 0. But because we added MIGRATION flage, it's not 3290 * uncharged yet. There are several case but page->mapcount check 3291 * and USED bit check in mem_cgroup_uncharge_page() will do enough 3292 * check. (see prepare_charge() also) 3293 */ 3294 if (PageAnon(used)) 3295 mem_cgroup_uncharge_page(used); 3296 /* 3297 * At migration, we may charge account against cgroup which has no 3298 * tasks. 3299 * So, rmdir()->pre_destroy() can be called while we do this charge. 3300 * In that case, we need to call pre_destroy() again. check it here. 3301 */ 3302 cgroup_release_and_wakeup_rmdir(&mem->css); 3303} 3304 3305/* 3306 * A call to try to shrink memory usage on charge failure at shmem's swapin. 3307 * Calling hierarchical_reclaim is not enough because we should update 3308 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. 3309 * Moreover considering hierarchy, we should reclaim from the mem_over_limit, 3310 * not from the memcg which this page would be charged to. 3311 * try_charge_swapin does all of these works properly. 3312 */ 3313int mem_cgroup_shmem_charge_fallback(struct page *page, 3314 struct mm_struct *mm, 3315 gfp_t gfp_mask) 3316{ 3317 struct mem_cgroup *mem; 3318 int ret; 3319 3320 if (mem_cgroup_disabled()) 3321 return 0; 3322 3323 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 3324 if (!ret) 3325 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ 3326 3327 return ret; 3328} 3329 3330#ifdef CONFIG_DEBUG_VM 3331static struct page_cgroup *lookup_page_cgroup_used(struct page *page) 3332{ 3333 struct page_cgroup *pc; 3334 3335 pc = lookup_page_cgroup(page); 3336 if (likely(pc) && PageCgroupUsed(pc)) 3337 return pc; 3338 return NULL; 3339} 3340 3341bool mem_cgroup_bad_page_check(struct page *page) 3342{ 3343 if (mem_cgroup_disabled()) 3344 return false; 3345 3346 return lookup_page_cgroup_used(page) != NULL; 3347} 3348 3349void mem_cgroup_print_bad_page(struct page *page) 3350{ 3351 struct page_cgroup *pc; 3352 3353 pc = lookup_page_cgroup_used(page); 3354 if (pc) { 3355 int ret = -1; 3356 char *path; 3357 3358 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p", 3359 pc, pc->flags, pc->mem_cgroup); 3360 3361 path = kmalloc(PATH_MAX, GFP_KERNEL); 3362 if (path) { 3363 rcu_read_lock(); 3364 ret = cgroup_path(pc->mem_cgroup->css.cgroup, 3365 path, PATH_MAX); 3366 rcu_read_unlock(); 3367 } 3368 3369 printk(KERN_CONT "(%s)\n", 3370 (ret < 0) ? "cannot get the path" : path); 3371 kfree(path); 3372 } 3373} 3374#endif 3375 3376static DEFINE_MUTEX(set_limit_mutex); 3377 3378static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 3379 unsigned long long val) 3380{ 3381 int retry_count; 3382 u64 memswlimit, memlimit; 3383 int ret = 0; 3384 int children = mem_cgroup_count_children(memcg); 3385 u64 curusage, oldusage; 3386 int enlarge; 3387 3388 /* 3389 * For keeping hierarchical_reclaim simple, how long we should retry 3390 * is depends on callers. We set our retry-count to be function 3391 * of # of children which we should visit in this loop. 3392 */ 3393 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 3394 3395 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3396 3397 enlarge = 0; 3398 while (retry_count) { 3399 if (signal_pending(current)) { 3400 ret = -EINTR; 3401 break; 3402 } 3403 /* 3404 * Rather than hide all in some function, I do this in 3405 * open coded manner. You see what this really does. 3406 * We have to guarantee mem->res.limit < mem->memsw.limit. 3407 */ 3408 mutex_lock(&set_limit_mutex); 3409 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3410 if (memswlimit < val) { 3411 ret = -EINVAL; 3412 mutex_unlock(&set_limit_mutex); 3413 break; 3414 } 3415 3416 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3417 if (memlimit < val) 3418 enlarge = 1; 3419 3420 ret = res_counter_set_limit(&memcg->res, val); 3421 if (!ret) { 3422 if (memswlimit == val) 3423 memcg->memsw_is_minimum = true; 3424 else 3425 memcg->memsw_is_minimum = false; 3426 } 3427 mutex_unlock(&set_limit_mutex); 3428 3429 if (!ret) 3430 break; 3431 3432 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, 3433 MEM_CGROUP_RECLAIM_SHRINK, 3434 NULL); 3435 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 3436 /* Usage is reduced ? */ 3437 if (curusage >= oldusage) 3438 retry_count--; 3439 else 3440 oldusage = curusage; 3441 } 3442 if (!ret && enlarge) 3443 memcg_oom_recover(memcg); 3444 3445 return ret; 3446} 3447 3448static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 3449 unsigned long long val) 3450{ 3451 int retry_count; 3452 u64 memlimit, memswlimit, oldusage, curusage; 3453 int children = mem_cgroup_count_children(memcg); 3454 int ret = -EBUSY; 3455 int enlarge = 0; 3456 3457 /* see mem_cgroup_resize_res_limit */ 3458 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 3459 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3460 while (retry_count) { 3461 if (signal_pending(current)) { 3462 ret = -EINTR; 3463 break; 3464 } 3465 /* 3466 * Rather than hide all in some function, I do this in 3467 * open coded manner. You see what this really does. 3468 * We have to guarantee mem->res.limit < mem->memsw.limit. 3469 */ 3470 mutex_lock(&set_limit_mutex); 3471 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3472 if (memlimit > val) { 3473 ret = -EINVAL; 3474 mutex_unlock(&set_limit_mutex); 3475 break; 3476 } 3477 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3478 if (memswlimit < val) 3479 enlarge = 1; 3480 ret = res_counter_set_limit(&memcg->memsw, val); 3481 if (!ret) { 3482 if (memlimit == val) 3483 memcg->memsw_is_minimum = true; 3484 else 3485 memcg->memsw_is_minimum = false; 3486 } 3487 mutex_unlock(&set_limit_mutex); 3488 3489 if (!ret) 3490 break; 3491 3492 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL, 3493 MEM_CGROUP_RECLAIM_NOSWAP | 3494 MEM_CGROUP_RECLAIM_SHRINK, 3495 NULL); 3496 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 3497 /* Usage is reduced ? */ 3498 if (curusage >= oldusage) 3499 retry_count--; 3500 else 3501 oldusage = curusage; 3502 } 3503 if (!ret && enlarge) 3504 memcg_oom_recover(memcg); 3505 return ret; 3506} 3507 3508unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order, 3509 gfp_t gfp_mask, 3510 unsigned long *total_scanned) 3511{ 3512 unsigned long nr_reclaimed = 0; 3513 struct mem_cgroup_per_zone *mz, *next_mz = NULL; 3514 unsigned long reclaimed; 3515 int loop = 0; 3516 struct mem_cgroup_tree_per_zone *mctz; 3517 unsigned long long excess; 3518 unsigned long nr_scanned; 3519 3520 if (order > 0) 3521 return 0; 3522 3523 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone)); 3524 /* 3525 * This loop can run a while, specially if mem_cgroup's continuously 3526 * keep exceeding their soft limit and putting the system under 3527 * pressure 3528 */ 3529 do { 3530 if (next_mz) 3531 mz = next_mz; 3532 else 3533 mz = mem_cgroup_largest_soft_limit_node(mctz); 3534 if (!mz) 3535 break; 3536 3537 nr_scanned = 0; 3538 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone, 3539 gfp_mask, 3540 MEM_CGROUP_RECLAIM_SOFT, 3541 &nr_scanned); 3542 nr_reclaimed += reclaimed; 3543 *total_scanned += nr_scanned; 3544 spin_lock(&mctz->lock); 3545 3546 /* 3547 * If we failed to reclaim anything from this memory cgroup 3548 * it is time to move on to the next cgroup 3549 */ 3550 next_mz = NULL; 3551 if (!reclaimed) { 3552 do { 3553 /* 3554 * Loop until we find yet another one. 3555 * 3556 * By the time we get the soft_limit lock 3557 * again, someone might have aded the 3558 * group back on the RB tree. Iterate to 3559 * make sure we get a different mem. 3560 * mem_cgroup_largest_soft_limit_node returns 3561 * NULL if no other cgroup is present on 3562 * the tree 3563 */ 3564 next_mz = 3565 __mem_cgroup_largest_soft_limit_node(mctz); 3566 if (next_mz == mz) 3567 css_put(&next_mz->mem->css); 3568 else /* next_mz == NULL or other memcg */ 3569 break; 3570 } while (1); 3571 } 3572 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz); 3573 excess = res_counter_soft_limit_excess(&mz->mem->res); 3574 /* 3575 * One school of thought says that we should not add 3576 * back the node to the tree if reclaim returns 0. 3577 * But our reclaim could return 0, simply because due 3578 * to priority we are exposing a smaller subset of 3579 * memory to reclaim from. Consider this as a longer 3580 * term TODO. 3581 */ 3582 /* If excess == 0, no tree ops */ 3583 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess); 3584 spin_unlock(&mctz->lock); 3585 css_put(&mz->mem->css); 3586 loop++; 3587 /* 3588 * Could not reclaim anything and there are no more 3589 * mem cgroups to try or we seem to be looping without 3590 * reclaiming anything. 3591 */ 3592 if (!nr_reclaimed && 3593 (next_mz == NULL || 3594 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS)) 3595 break; 3596 } while (!nr_reclaimed); 3597 if (next_mz) 3598 css_put(&next_mz->mem->css); 3599 return nr_reclaimed; 3600} 3601 3602/* 3603 * This routine traverse page_cgroup in given list and drop them all. 3604 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 3605 */ 3606static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, 3607 int node, int zid, enum lru_list lru) 3608{ 3609 struct zone *zone; 3610 struct mem_cgroup_per_zone *mz; 3611 struct page_cgroup *pc, *busy; 3612 unsigned long flags, loop; 3613 struct list_head *list; 3614 int ret = 0; 3615 3616 zone = &NODE_DATA(node)->node_zones[zid]; 3617 mz = mem_cgroup_zoneinfo(mem, node, zid); 3618 list = &mz->lists[lru]; 3619 3620 loop = MEM_CGROUP_ZSTAT(mz, lru); 3621 /* give some margin against EBUSY etc...*/ 3622 loop += 256; 3623 busy = NULL; 3624 while (loop--) { 3625 struct page *page; 3626 3627 ret = 0; 3628 spin_lock_irqsave(&zone->lru_lock, flags); 3629 if (list_empty(list)) { 3630 spin_unlock_irqrestore(&zone->lru_lock, flags); 3631 break; 3632 } 3633 pc = list_entry(list->prev, struct page_cgroup, lru); 3634 if (busy == pc) { 3635 list_move(&pc->lru, list); 3636 busy = NULL; 3637 spin_unlock_irqrestore(&zone->lru_lock, flags); 3638 continue; 3639 } 3640 spin_unlock_irqrestore(&zone->lru_lock, flags); 3641 3642 page = lookup_cgroup_page(pc); 3643 3644 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL); 3645 if (ret == -ENOMEM) 3646 break; 3647 3648 if (ret == -EBUSY || ret == -EINVAL) { 3649 /* found lock contention or "pc" is obsolete. */ 3650 busy = pc; 3651 cond_resched(); 3652 } else 3653 busy = NULL; 3654 } 3655 3656 if (!ret && !list_empty(list)) 3657 return -EBUSY; 3658 return ret; 3659} 3660 3661/* 3662 * make mem_cgroup's charge to be 0 if there is no task. 3663 * This enables deleting this mem_cgroup. 3664 */ 3665static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) 3666{ 3667 int ret; 3668 int node, zid, shrink; 3669 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 3670 struct cgroup *cgrp = mem->css.cgroup; 3671 3672 css_get(&mem->css); 3673 3674 shrink = 0; 3675 /* should free all ? */ 3676 if (free_all) 3677 goto try_to_free; 3678move_account: 3679 do { 3680 ret = -EBUSY; 3681 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 3682 goto out; 3683 ret = -EINTR; 3684 if (signal_pending(current)) 3685 goto out; 3686 /* This is for making all *used* pages to be on LRU. */ 3687 lru_add_drain_all(); 3688 drain_all_stock_sync(); 3689 ret = 0; 3690 mem_cgroup_start_move(mem); 3691 for_each_node_state(node, N_HIGH_MEMORY) { 3692 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 3693 enum lru_list l; 3694 for_each_lru(l) { 3695 ret = mem_cgroup_force_empty_list(mem, 3696 node, zid, l); 3697 if (ret) 3698 break; 3699 } 3700 } 3701 if (ret) 3702 break; 3703 } 3704 mem_cgroup_end_move(mem); 3705 memcg_oom_recover(mem); 3706 /* it seems parent cgroup doesn't have enough mem */ 3707 if (ret == -ENOMEM) 3708 goto try_to_free; 3709 cond_resched(); 3710 /* "ret" should also be checked to ensure all lists are empty. */ 3711 } while (mem->res.usage > 0 || ret); 3712out: 3713 css_put(&mem->css); 3714 return ret; 3715 3716try_to_free: 3717 /* returns EBUSY if there is a task or if we come here twice. */ 3718 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 3719 ret = -EBUSY; 3720 goto out; 3721 } 3722 /* we call try-to-free pages for make this cgroup empty */ 3723 lru_add_drain_all(); 3724 /* try to free all pages in this cgroup */ 3725 shrink = 1; 3726 while (nr_retries && mem->res.usage > 0) { 3727 int progress; 3728 3729 if (signal_pending(current)) { 3730 ret = -EINTR; 3731 goto out; 3732 } 3733 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, 3734 false, get_swappiness(mem)); 3735 if (!progress) { 3736 nr_retries--; 3737 /* maybe some writeback is necessary */ 3738 congestion_wait(BLK_RW_ASYNC, HZ/10); 3739 } 3740 3741 } 3742 lru_add_drain(); 3743 /* try move_account...there may be some *locked* pages. */ 3744 goto move_account; 3745} 3746 3747int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 3748{ 3749 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 3750} 3751 3752 3753static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 3754{ 3755 return mem_cgroup_from_cont(cont)->use_hierarchy; 3756} 3757 3758static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 3759 u64 val) 3760{ 3761 int retval = 0; 3762 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 3763 struct cgroup *parent = cont->parent; 3764 struct mem_cgroup *parent_mem = NULL; 3765 3766 if (parent) 3767 parent_mem = mem_cgroup_from_cont(parent); 3768 3769 cgroup_lock(); 3770 /* 3771 * If parent's use_hierarchy is set, we can't make any modifications 3772 * in the child subtrees. If it is unset, then the change can 3773 * occur, provided the current cgroup has no children. 3774 * 3775 * For the root cgroup, parent_mem is NULL, we allow value to be 3776 * set if there are no children. 3777 */ 3778 if ((!parent_mem || !parent_mem->use_hierarchy) && 3779 (val == 1 || val == 0)) { 3780 if (list_empty(&cont->children)) 3781 mem->use_hierarchy = val; 3782 else 3783 retval = -EBUSY; 3784 } else 3785 retval = -EINVAL; 3786 cgroup_unlock(); 3787 3788 return retval; 3789} 3790 3791 3792static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem, 3793 enum mem_cgroup_stat_index idx) 3794{ 3795 struct mem_cgroup *iter; 3796 long val = 0; 3797 3798 /* Per-cpu values can be negative, use a signed accumulator */ 3799 for_each_mem_cgroup_tree(iter, mem) 3800 val += mem_cgroup_read_stat(iter, idx); 3801 3802 if (val < 0) /* race ? */ 3803 val = 0; 3804 return val; 3805} 3806 3807static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap) 3808{ 3809 u64 val; 3810 3811 if (!mem_cgroup_is_root(mem)) { 3812 if (!swap) 3813 return res_counter_read_u64(&mem->res, RES_USAGE); 3814 else 3815 return res_counter_read_u64(&mem->memsw, RES_USAGE); 3816 } 3817 3818 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE); 3819 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS); 3820 3821 if (swap) 3822 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT); 3823 3824 return val << PAGE_SHIFT; 3825} 3826 3827static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 3828{ 3829 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 3830 u64 val; 3831 int type, name; 3832 3833 type = MEMFILE_TYPE(cft->private); 3834 name = MEMFILE_ATTR(cft->private); 3835 switch (type) { 3836 case _MEM: 3837 if (name == RES_USAGE) 3838 val = mem_cgroup_usage(mem, false); 3839 else 3840 val = res_counter_read_u64(&mem->res, name); 3841 break; 3842 case _MEMSWAP: 3843 if (name == RES_USAGE) 3844 val = mem_cgroup_usage(mem, true); 3845 else 3846 val = res_counter_read_u64(&mem->memsw, name); 3847 break; 3848 default: 3849 BUG(); 3850 break; 3851 } 3852 return val; 3853} 3854/* 3855 * The user of this function is... 3856 * RES_LIMIT. 3857 */ 3858static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 3859 const char *buffer) 3860{ 3861 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 3862 int type, name; 3863 unsigned long long val; 3864 int ret; 3865 3866 type = MEMFILE_TYPE(cft->private); 3867 name = MEMFILE_ATTR(cft->private); 3868 switch (name) { 3869 case RES_LIMIT: 3870 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 3871 ret = -EINVAL; 3872 break; 3873 } 3874 /* This function does all necessary parse...reuse it */ 3875 ret = res_counter_memparse_write_strategy(buffer, &val); 3876 if (ret) 3877 break; 3878 if (type == _MEM) 3879 ret = mem_cgroup_resize_limit(memcg, val); 3880 else 3881 ret = mem_cgroup_resize_memsw_limit(memcg, val); 3882 break; 3883 case RES_SOFT_LIMIT: 3884 ret = res_counter_memparse_write_strategy(buffer, &val); 3885 if (ret) 3886 break; 3887 /* 3888 * For memsw, soft limits are hard to implement in terms 3889 * of semantics, for now, we support soft limits for 3890 * control without swap 3891 */ 3892 if (type == _MEM) 3893 ret = res_counter_set_soft_limit(&memcg->res, val); 3894 else 3895 ret = -EINVAL; 3896 break; 3897 default: 3898 ret = -EINVAL; /* should be BUG() ? */ 3899 break; 3900 } 3901 return ret; 3902} 3903 3904static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 3905 unsigned long long *mem_limit, unsigned long long *memsw_limit) 3906{ 3907 struct cgroup *cgroup; 3908 unsigned long long min_limit, min_memsw_limit, tmp; 3909 3910 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 3911 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3912 cgroup = memcg->css.cgroup; 3913 if (!memcg->use_hierarchy) 3914 goto out; 3915 3916 while (cgroup->parent) { 3917 cgroup = cgroup->parent; 3918 memcg = mem_cgroup_from_cont(cgroup); 3919 if (!memcg->use_hierarchy) 3920 break; 3921 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 3922 min_limit = min(min_limit, tmp); 3923 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 3924 min_memsw_limit = min(min_memsw_limit, tmp); 3925 } 3926out: 3927 *mem_limit = min_limit; 3928 *memsw_limit = min_memsw_limit; 3929 return; 3930} 3931 3932static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 3933{ 3934 struct mem_cgroup *mem; 3935 int type, name; 3936 3937 mem = mem_cgroup_from_cont(cont); 3938 type = MEMFILE_TYPE(event); 3939 name = MEMFILE_ATTR(event); 3940 switch (name) { 3941 case RES_MAX_USAGE: 3942 if (type == _MEM) 3943 res_counter_reset_max(&mem->res); 3944 else 3945 res_counter_reset_max(&mem->memsw); 3946 break; 3947 case RES_FAILCNT: 3948 if (type == _MEM) 3949 res_counter_reset_failcnt(&mem->res); 3950 else 3951 res_counter_reset_failcnt(&mem->memsw); 3952 break; 3953 } 3954 3955 return 0; 3956} 3957 3958static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp, 3959 struct cftype *cft) 3960{ 3961 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate; 3962} 3963 3964#ifdef CONFIG_MMU 3965static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3966 struct cftype *cft, u64 val) 3967{ 3968 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 3969 3970 if (val >= (1 << NR_MOVE_TYPE)) 3971 return -EINVAL; 3972 /* 3973 * We check this value several times in both in can_attach() and 3974 * attach(), so we need cgroup lock to prevent this value from being 3975 * inconsistent. 3976 */ 3977 cgroup_lock(); 3978 mem->move_charge_at_immigrate = val; 3979 cgroup_unlock(); 3980 3981 return 0; 3982} 3983#else 3984static int mem_cgroup_move_charge_write(struct cgroup *cgrp, 3985 struct cftype *cft, u64 val) 3986{ 3987 return -ENOSYS; 3988} 3989#endif 3990 3991 3992/* For read statistics */ 3993enum { 3994 MCS_CACHE, 3995 MCS_RSS, 3996 MCS_FILE_MAPPED, 3997 MCS_PGPGIN, 3998 MCS_PGPGOUT, 3999 MCS_SWAP, 4000 MCS_PGFAULT, 4001 MCS_PGMAJFAULT, 4002 MCS_INACTIVE_ANON, 4003 MCS_ACTIVE_ANON, 4004 MCS_INACTIVE_FILE, 4005 MCS_ACTIVE_FILE, 4006 MCS_UNEVICTABLE, 4007 NR_MCS_STAT, 4008}; 4009 4010struct mcs_total_stat { 4011 s64 stat[NR_MCS_STAT]; 4012}; 4013 4014struct { 4015 char *local_name; 4016 char *total_name; 4017} memcg_stat_strings[NR_MCS_STAT] = { 4018 {"cache", "total_cache"}, 4019 {"rss", "total_rss"}, 4020 {"mapped_file", "total_mapped_file"}, 4021 {"pgpgin", "total_pgpgin"}, 4022 {"pgpgout", "total_pgpgout"}, 4023 {"swap", "total_swap"}, 4024 {"pgfault", "total_pgfault"}, 4025 {"pgmajfault", "total_pgmajfault"}, 4026 {"inactive_anon", "total_inactive_anon"}, 4027 {"active_anon", "total_active_anon"}, 4028 {"inactive_file", "total_inactive_file"}, 4029 {"active_file", "total_active_file"}, 4030 {"unevictable", "total_unevictable"} 4031}; 4032 4033 4034static void 4035mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 4036{ 4037 s64 val; 4038 4039 /* per cpu stat */ 4040 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE); 4041 s->stat[MCS_CACHE] += val * PAGE_SIZE; 4042 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS); 4043 s->stat[MCS_RSS] += val * PAGE_SIZE; 4044 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED); 4045 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE; 4046 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN); 4047 s->stat[MCS_PGPGIN] += val; 4048 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT); 4049 s->stat[MCS_PGPGOUT] += val; 4050 if (do_swap_account) { 4051 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT); 4052 s->stat[MCS_SWAP] += val * PAGE_SIZE; 4053 } 4054 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT); 4055 s->stat[MCS_PGFAULT] += val; 4056 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT); 4057 s->stat[MCS_PGMAJFAULT] += val; 4058 4059 /* per zone stat */ 4060 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); 4061 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 4062 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); 4063 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 4064 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); 4065 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 4066 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); 4067 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 4068 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); 4069 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 4070} 4071 4072static void 4073mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 4074{ 4075 struct mem_cgroup *iter; 4076 4077 for_each_mem_cgroup_tree(iter, mem) 4078 mem_cgroup_get_local_stat(iter, s); 4079} 4080 4081#ifdef CONFIG_NUMA 4082static int mem_control_numa_stat_show(struct seq_file *m, void *arg) 4083{ 4084 int nid; 4085 unsigned long total_nr, file_nr, anon_nr, unevictable_nr; 4086 unsigned long node_nr; 4087 struct cgroup *cont = m->private; 4088 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4089 4090 total_nr = mem_cgroup_nr_lru_pages(mem_cont); 4091 seq_printf(m, "total=%lu", total_nr); 4092 for_each_node_state(nid, N_HIGH_MEMORY) { 4093 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid); 4094 seq_printf(m, " N%d=%lu", nid, node_nr); 4095 } 4096 seq_putc(m, '\n'); 4097 4098 file_nr = mem_cgroup_nr_file_lru_pages(mem_cont); 4099 seq_printf(m, "file=%lu", file_nr); 4100 for_each_node_state(nid, N_HIGH_MEMORY) { 4101 node_nr = mem_cgroup_node_nr_file_lru_pages(mem_cont, nid); 4102 seq_printf(m, " N%d=%lu", nid, node_nr); 4103 } 4104 seq_putc(m, '\n'); 4105 4106 anon_nr = mem_cgroup_nr_anon_lru_pages(mem_cont); 4107 seq_printf(m, "anon=%lu", anon_nr); 4108 for_each_node_state(nid, N_HIGH_MEMORY) { 4109 node_nr = mem_cgroup_node_nr_anon_lru_pages(mem_cont, nid); 4110 seq_printf(m, " N%d=%lu", nid, node_nr); 4111 } 4112 seq_putc(m, '\n'); 4113 4114 unevictable_nr = mem_cgroup_nr_unevictable_lru_pages(mem_cont); 4115 seq_printf(m, "unevictable=%lu", unevictable_nr); 4116 for_each_node_state(nid, N_HIGH_MEMORY) { 4117 node_nr = mem_cgroup_node_nr_unevictable_lru_pages(mem_cont, 4118 nid); 4119 seq_printf(m, " N%d=%lu", nid, node_nr); 4120 } 4121 seq_putc(m, '\n'); 4122 return 0; 4123} 4124#endif /* CONFIG_NUMA */ 4125 4126static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 4127 struct cgroup_map_cb *cb) 4128{ 4129 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 4130 struct mcs_total_stat mystat; 4131 int i; 4132 4133 memset(&mystat, 0, sizeof(mystat)); 4134 mem_cgroup_get_local_stat(mem_cont, &mystat); 4135 4136 4137 for (i = 0; i < NR_MCS_STAT; i++) { 4138 if (i == MCS_SWAP && !do_swap_account) 4139 continue; 4140 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 4141 } 4142 4143 /* Hierarchical information */ 4144 { 4145 unsigned long long limit, memsw_limit; 4146 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 4147 cb->fill(cb, "hierarchical_memory_limit", limit); 4148 if (do_swap_account) 4149 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 4150 } 4151 4152 memset(&mystat, 0, sizeof(mystat)); 4153 mem_cgroup_get_total_stat(mem_cont, &mystat); 4154 for (i = 0; i < NR_MCS_STAT; i++) { 4155 if (i == MCS_SWAP && !do_swap_account) 4156 continue; 4157 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 4158 } 4159 4160#ifdef CONFIG_DEBUG_VM 4161 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); 4162 4163 { 4164 int nid, zid; 4165 struct mem_cgroup_per_zone *mz; 4166 unsigned long recent_rotated[2] = {0, 0}; 4167 unsigned long recent_scanned[2] = {0, 0}; 4168 4169 for_each_online_node(nid) 4170 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 4171 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 4172 4173 recent_rotated[0] += 4174 mz->reclaim_stat.recent_rotated[0]; 4175 recent_rotated[1] += 4176 mz->reclaim_stat.recent_rotated[1]; 4177 recent_scanned[0] += 4178 mz->reclaim_stat.recent_scanned[0]; 4179 recent_scanned[1] += 4180 mz->reclaim_stat.recent_scanned[1]; 4181 } 4182 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 4183 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 4184 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 4185 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 4186 } 4187#endif 4188 4189 return 0; 4190} 4191 4192static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 4193{ 4194 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4195 4196 return get_swappiness(memcg); 4197} 4198 4199static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 4200 u64 val) 4201{ 4202 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4203 struct mem_cgroup *parent; 4204 4205 if (val > 100) 4206 return -EINVAL; 4207 4208 if (cgrp->parent == NULL) 4209 return -EINVAL; 4210 4211 parent = mem_cgroup_from_cont(cgrp->parent); 4212 4213 cgroup_lock(); 4214 4215 /* If under hierarchy, only empty-root can set this value */ 4216 if ((parent->use_hierarchy) || 4217 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 4218 cgroup_unlock(); 4219 return -EINVAL; 4220 } 4221 4222 memcg->swappiness = val; 4223 4224 cgroup_unlock(); 4225 4226 return 0; 4227} 4228 4229static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap) 4230{ 4231 struct mem_cgroup_threshold_ary *t; 4232 u64 usage; 4233 int i; 4234 4235 rcu_read_lock(); 4236 if (!swap) 4237 t = rcu_dereference(memcg->thresholds.primary); 4238 else 4239 t = rcu_dereference(memcg->memsw_thresholds.primary); 4240 4241 if (!t) 4242 goto unlock; 4243 4244 usage = mem_cgroup_usage(memcg, swap); 4245 4246 /* 4247 * current_threshold points to threshold just below usage. 4248 * If it's not true, a threshold was crossed after last 4249 * call of __mem_cgroup_threshold(). 4250 */ 4251 i = t->current_threshold; 4252 4253 /* 4254 * Iterate backward over array of thresholds starting from 4255 * current_threshold and check if a threshold is crossed. 4256 * If none of thresholds below usage is crossed, we read 4257 * only one element of the array here. 4258 */ 4259 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--) 4260 eventfd_signal(t->entries[i].eventfd, 1); 4261 4262 /* i = current_threshold + 1 */ 4263 i++; 4264 4265 /* 4266 * Iterate forward over array of thresholds starting from 4267 * current_threshold+1 and check if a threshold is crossed. 4268 * If none of thresholds above usage is crossed, we read 4269 * only one element of the array here. 4270 */ 4271 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++) 4272 eventfd_signal(t->entries[i].eventfd, 1); 4273 4274 /* Update current_threshold */ 4275 t->current_threshold = i - 1; 4276unlock: 4277 rcu_read_unlock(); 4278} 4279 4280static void mem_cgroup_threshold(struct mem_cgroup *memcg) 4281{ 4282 while (memcg) { 4283 __mem_cgroup_threshold(memcg, false); 4284 if (do_swap_account) 4285 __mem_cgroup_threshold(memcg, true); 4286 4287 memcg = parent_mem_cgroup(memcg); 4288 } 4289} 4290 4291static int compare_thresholds(const void *a, const void *b) 4292{ 4293 const struct mem_cgroup_threshold *_a = a; 4294 const struct mem_cgroup_threshold *_b = b; 4295 4296 return _a->threshold - _b->threshold; 4297} 4298 4299static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem) 4300{ 4301 struct mem_cgroup_eventfd_list *ev; 4302 4303 list_for_each_entry(ev, &mem->oom_notify, list) 4304 eventfd_signal(ev->eventfd, 1); 4305 return 0; 4306} 4307 4308static void mem_cgroup_oom_notify(struct mem_cgroup *mem) 4309{ 4310 struct mem_cgroup *iter; 4311 4312 for_each_mem_cgroup_tree(iter, mem) 4313 mem_cgroup_oom_notify_cb(iter); 4314} 4315 4316static int mem_cgroup_usage_register_event(struct cgroup *cgrp, 4317 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4318{ 4319 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4320 struct mem_cgroup_thresholds *thresholds; 4321 struct mem_cgroup_threshold_ary *new; 4322 int type = MEMFILE_TYPE(cft->private); 4323 u64 threshold, usage; 4324 int i, size, ret; 4325 4326 ret = res_counter_memparse_write_strategy(args, &threshold); 4327 if (ret) 4328 return ret; 4329 4330 mutex_lock(&memcg->thresholds_lock); 4331 4332 if (type == _MEM) 4333 thresholds = &memcg->thresholds; 4334 else if (type == _MEMSWAP) 4335 thresholds = &memcg->memsw_thresholds; 4336 else 4337 BUG(); 4338 4339 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4340 4341 /* Check if a threshold crossed before adding a new one */ 4342 if (thresholds->primary) 4343 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4344 4345 size = thresholds->primary ? thresholds->primary->size + 1 : 1; 4346 4347 /* Allocate memory for new array of thresholds */ 4348 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold), 4349 GFP_KERNEL); 4350 if (!new) { 4351 ret = -ENOMEM; 4352 goto unlock; 4353 } 4354 new->size = size; 4355 4356 /* Copy thresholds (if any) to new array */ 4357 if (thresholds->primary) { 4358 memcpy(new->entries, thresholds->primary->entries, (size - 1) * 4359 sizeof(struct mem_cgroup_threshold)); 4360 } 4361 4362 /* Add new threshold */ 4363 new->entries[size - 1].eventfd = eventfd; 4364 new->entries[size - 1].threshold = threshold; 4365 4366 /* Sort thresholds. Registering of new threshold isn't time-critical */ 4367 sort(new->entries, size, sizeof(struct mem_cgroup_threshold), 4368 compare_thresholds, NULL); 4369 4370 /* Find current threshold */ 4371 new->current_threshold = -1; 4372 for (i = 0; i < size; i++) { 4373 if (new->entries[i].threshold < usage) { 4374 /* 4375 * new->current_threshold will not be used until 4376 * rcu_assign_pointer(), so it's safe to increment 4377 * it here. 4378 */ 4379 ++new->current_threshold; 4380 } 4381 } 4382 4383 /* Free old spare buffer and save old primary buffer as spare */ 4384 kfree(thresholds->spare); 4385 thresholds->spare = thresholds->primary; 4386 4387 rcu_assign_pointer(thresholds->primary, new); 4388 4389 /* To be sure that nobody uses thresholds */ 4390 synchronize_rcu(); 4391 4392unlock: 4393 mutex_unlock(&memcg->thresholds_lock); 4394 4395 return ret; 4396} 4397 4398static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp, 4399 struct cftype *cft, struct eventfd_ctx *eventfd) 4400{ 4401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4402 struct mem_cgroup_thresholds *thresholds; 4403 struct mem_cgroup_threshold_ary *new; 4404 int type = MEMFILE_TYPE(cft->private); 4405 u64 usage; 4406 int i, j, size; 4407 4408 mutex_lock(&memcg->thresholds_lock); 4409 if (type == _MEM) 4410 thresholds = &memcg->thresholds; 4411 else if (type == _MEMSWAP) 4412 thresholds = &memcg->memsw_thresholds; 4413 else 4414 BUG(); 4415 4416 /* 4417 * Something went wrong if we trying to unregister a threshold 4418 * if we don't have thresholds 4419 */ 4420 BUG_ON(!thresholds); 4421 4422 usage = mem_cgroup_usage(memcg, type == _MEMSWAP); 4423 4424 /* Check if a threshold crossed before removing */ 4425 __mem_cgroup_threshold(memcg, type == _MEMSWAP); 4426 4427 /* Calculate new number of threshold */ 4428 size = 0; 4429 for (i = 0; i < thresholds->primary->size; i++) { 4430 if (thresholds->primary->entries[i].eventfd != eventfd) 4431 size++; 4432 } 4433 4434 new = thresholds->spare; 4435 4436 /* Set thresholds array to NULL if we don't have thresholds */ 4437 if (!size) { 4438 kfree(new); 4439 new = NULL; 4440 goto swap_buffers; 4441 } 4442 4443 new->size = size; 4444 4445 /* Copy thresholds and find current threshold */ 4446 new->current_threshold = -1; 4447 for (i = 0, j = 0; i < thresholds->primary->size; i++) { 4448 if (thresholds->primary->entries[i].eventfd == eventfd) 4449 continue; 4450 4451 new->entries[j] = thresholds->primary->entries[i]; 4452 if (new->entries[j].threshold < usage) { 4453 /* 4454 * new->current_threshold will not be used 4455 * until rcu_assign_pointer(), so it's safe to increment 4456 * it here. 4457 */ 4458 ++new->current_threshold; 4459 } 4460 j++; 4461 } 4462 4463swap_buffers: 4464 /* Swap primary and spare array */ 4465 thresholds->spare = thresholds->primary; 4466 rcu_assign_pointer(thresholds->primary, new); 4467 4468 /* To be sure that nobody uses thresholds */ 4469 synchronize_rcu(); 4470 4471 mutex_unlock(&memcg->thresholds_lock); 4472} 4473 4474static int mem_cgroup_oom_register_event(struct cgroup *cgrp, 4475 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args) 4476{ 4477 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 4478 struct mem_cgroup_eventfd_list *event; 4479 int type = MEMFILE_TYPE(cft->private); 4480 4481 BUG_ON(type != _OOM_TYPE); 4482 event = kmalloc(sizeof(*event), GFP_KERNEL); 4483 if (!event) 4484 return -ENOMEM; 4485 4486 mutex_lock(&memcg_oom_mutex); 4487 4488 event->eventfd = eventfd; 4489 list_add(&event->list, &memcg->oom_notify); 4490 4491 /* already in OOM ? */ 4492 if (atomic_read(&memcg->oom_lock)) 4493 eventfd_signal(eventfd, 1); 4494 mutex_unlock(&memcg_oom_mutex); 4495 4496 return 0; 4497} 4498 4499static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp, 4500 struct cftype *cft, struct eventfd_ctx *eventfd) 4501{ 4502 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4503 struct mem_cgroup_eventfd_list *ev, *tmp; 4504 int type = MEMFILE_TYPE(cft->private); 4505 4506 BUG_ON(type != _OOM_TYPE); 4507 4508 mutex_lock(&memcg_oom_mutex); 4509 4510 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) { 4511 if (ev->eventfd == eventfd) { 4512 list_del(&ev->list); 4513 kfree(ev); 4514 } 4515 } 4516 4517 mutex_unlock(&memcg_oom_mutex); 4518} 4519 4520static int mem_cgroup_oom_control_read(struct cgroup *cgrp, 4521 struct cftype *cft, struct cgroup_map_cb *cb) 4522{ 4523 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4524 4525 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable); 4526 4527 if (atomic_read(&mem->oom_lock)) 4528 cb->fill(cb, "under_oom", 1); 4529 else 4530 cb->fill(cb, "under_oom", 0); 4531 return 0; 4532} 4533 4534static int mem_cgroup_oom_control_write(struct cgroup *cgrp, 4535 struct cftype *cft, u64 val) 4536{ 4537 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp); 4538 struct mem_cgroup *parent; 4539 4540 /* cannot set to root cgroup and only 0 and 1 are allowed */ 4541 if (!cgrp->parent || !((val == 0) || (val == 1))) 4542 return -EINVAL; 4543 4544 parent = mem_cgroup_from_cont(cgrp->parent); 4545 4546 cgroup_lock(); 4547 /* oom-kill-disable is a flag for subhierarchy. */ 4548 if ((parent->use_hierarchy) || 4549 (mem->use_hierarchy && !list_empty(&cgrp->children))) { 4550 cgroup_unlock(); 4551 return -EINVAL; 4552 } 4553 mem->oom_kill_disable = val; 4554 if (!val) 4555 memcg_oom_recover(mem); 4556 cgroup_unlock(); 4557 return 0; 4558} 4559 4560#ifdef CONFIG_NUMA 4561static const struct file_operations mem_control_numa_stat_file_operations = { 4562 .read = seq_read, 4563 .llseek = seq_lseek, 4564 .release = single_release, 4565}; 4566 4567static int mem_control_numa_stat_open(struct inode *unused, struct file *file) 4568{ 4569 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata; 4570 4571 file->f_op = &mem_control_numa_stat_file_operations; 4572 return single_open(file, mem_control_numa_stat_show, cont); 4573} 4574#endif /* CONFIG_NUMA */ 4575 4576static struct cftype mem_cgroup_files[] = { 4577 { 4578 .name = "usage_in_bytes", 4579 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 4580 .read_u64 = mem_cgroup_read, 4581 .register_event = mem_cgroup_usage_register_event, 4582 .unregister_event = mem_cgroup_usage_unregister_event, 4583 }, 4584 { 4585 .name = "max_usage_in_bytes", 4586 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 4587 .trigger = mem_cgroup_reset, 4588 .read_u64 = mem_cgroup_read, 4589 }, 4590 { 4591 .name = "limit_in_bytes", 4592 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 4593 .write_string = mem_cgroup_write, 4594 .read_u64 = mem_cgroup_read, 4595 }, 4596 { 4597 .name = "soft_limit_in_bytes", 4598 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT), 4599 .write_string = mem_cgroup_write, 4600 .read_u64 = mem_cgroup_read, 4601 }, 4602 { 4603 .name = "failcnt", 4604 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 4605 .trigger = mem_cgroup_reset, 4606 .read_u64 = mem_cgroup_read, 4607 }, 4608 { 4609 .name = "stat", 4610 .read_map = mem_control_stat_show, 4611 }, 4612 { 4613 .name = "force_empty", 4614 .trigger = mem_cgroup_force_empty_write, 4615 }, 4616 { 4617 .name = "use_hierarchy", 4618 .write_u64 = mem_cgroup_hierarchy_write, 4619 .read_u64 = mem_cgroup_hierarchy_read, 4620 }, 4621 { 4622 .name = "swappiness", 4623 .read_u64 = mem_cgroup_swappiness_read, 4624 .write_u64 = mem_cgroup_swappiness_write, 4625 }, 4626 { 4627 .name = "move_charge_at_immigrate", 4628 .read_u64 = mem_cgroup_move_charge_read, 4629 .write_u64 = mem_cgroup_move_charge_write, 4630 }, 4631 { 4632 .name = "oom_control", 4633 .read_map = mem_cgroup_oom_control_read, 4634 .write_u64 = mem_cgroup_oom_control_write, 4635 .register_event = mem_cgroup_oom_register_event, 4636 .unregister_event = mem_cgroup_oom_unregister_event, 4637 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL), 4638 }, 4639#ifdef CONFIG_NUMA 4640 { 4641 .name = "numa_stat", 4642 .open = mem_control_numa_stat_open, 4643 }, 4644#endif 4645}; 4646 4647#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4648static struct cftype memsw_cgroup_files[] = { 4649 { 4650 .name = "memsw.usage_in_bytes", 4651 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 4652 .read_u64 = mem_cgroup_read, 4653 .register_event = mem_cgroup_usage_register_event, 4654 .unregister_event = mem_cgroup_usage_unregister_event, 4655 }, 4656 { 4657 .name = "memsw.max_usage_in_bytes", 4658 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 4659 .trigger = mem_cgroup_reset, 4660 .read_u64 = mem_cgroup_read, 4661 }, 4662 { 4663 .name = "memsw.limit_in_bytes", 4664 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 4665 .write_string = mem_cgroup_write, 4666 .read_u64 = mem_cgroup_read, 4667 }, 4668 { 4669 .name = "memsw.failcnt", 4670 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 4671 .trigger = mem_cgroup_reset, 4672 .read_u64 = mem_cgroup_read, 4673 }, 4674}; 4675 4676static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4677{ 4678 if (!do_swap_account) 4679 return 0; 4680 return cgroup_add_files(cont, ss, memsw_cgroup_files, 4681 ARRAY_SIZE(memsw_cgroup_files)); 4682}; 4683#else 4684static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 4685{ 4686 return 0; 4687} 4688#endif 4689 4690static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 4691{ 4692 struct mem_cgroup_per_node *pn; 4693 struct mem_cgroup_per_zone *mz; 4694 enum lru_list l; 4695 int zone, tmp = node; 4696 /* 4697 * This routine is called against possible nodes. 4698 * But it's BUG to call kmalloc() against offline node. 4699 * 4700 * TODO: this routine can waste much memory for nodes which will 4701 * never be onlined. It's better to use memory hotplug callback 4702 * function. 4703 */ 4704 if (!node_state(node, N_NORMAL_MEMORY)) 4705 tmp = -1; 4706 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 4707 if (!pn) 4708 return 1; 4709 4710 mem->info.nodeinfo[node] = pn; 4711 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4712 mz = &pn->zoneinfo[zone]; 4713 for_each_lru(l) 4714 INIT_LIST_HEAD(&mz->lists[l]); 4715 mz->usage_in_excess = 0; 4716 mz->on_tree = false; 4717 mz->mem = mem; 4718 } 4719 return 0; 4720} 4721 4722static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 4723{ 4724 kfree(mem->info.nodeinfo[node]); 4725} 4726 4727static struct mem_cgroup *mem_cgroup_alloc(void) 4728{ 4729 struct mem_cgroup *mem; 4730 int size = sizeof(struct mem_cgroup); 4731 4732 /* Can be very big if MAX_NUMNODES is very big */ 4733 if (size < PAGE_SIZE) 4734 mem = kzalloc(size, GFP_KERNEL); 4735 else 4736 mem = vzalloc(size); 4737 4738 if (!mem) 4739 return NULL; 4740 4741 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu); 4742 if (!mem->stat) 4743 goto out_free; 4744 spin_lock_init(&mem->pcp_counter_lock); 4745 return mem; 4746 4747out_free: 4748 if (size < PAGE_SIZE) 4749 kfree(mem); 4750 else 4751 vfree(mem); 4752 return NULL; 4753} 4754 4755/* 4756 * At destroying mem_cgroup, references from swap_cgroup can remain. 4757 * (scanning all at force_empty is too costly...) 4758 * 4759 * Instead of clearing all references at force_empty, we remember 4760 * the number of reference from swap_cgroup and free mem_cgroup when 4761 * it goes down to 0. 4762 * 4763 * Removal of cgroup itself succeeds regardless of refs from swap. 4764 */ 4765 4766static void __mem_cgroup_free(struct mem_cgroup *mem) 4767{ 4768 int node; 4769 4770 mem_cgroup_remove_from_trees(mem); 4771 free_css_id(&mem_cgroup_subsys, &mem->css); 4772 4773 for_each_node_state(node, N_POSSIBLE) 4774 free_mem_cgroup_per_zone_info(mem, node); 4775 4776 free_percpu(mem->stat); 4777 if (sizeof(struct mem_cgroup) < PAGE_SIZE) 4778 kfree(mem); 4779 else 4780 vfree(mem); 4781} 4782 4783static void mem_cgroup_get(struct mem_cgroup *mem) 4784{ 4785 atomic_inc(&mem->refcnt); 4786} 4787 4788static void __mem_cgroup_put(struct mem_cgroup *mem, int count) 4789{ 4790 if (atomic_sub_and_test(count, &mem->refcnt)) { 4791 struct mem_cgroup *parent = parent_mem_cgroup(mem); 4792 __mem_cgroup_free(mem); 4793 if (parent) 4794 mem_cgroup_put(parent); 4795 } 4796} 4797 4798static void mem_cgroup_put(struct mem_cgroup *mem) 4799{ 4800 __mem_cgroup_put(mem, 1); 4801} 4802 4803/* 4804 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 4805 */ 4806static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) 4807{ 4808 if (!mem->res.parent) 4809 return NULL; 4810 return mem_cgroup_from_res_counter(mem->res.parent, res); 4811} 4812 4813#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 4814static void __init enable_swap_cgroup(void) 4815{ 4816 if (!mem_cgroup_disabled() && really_do_swap_account) 4817 do_swap_account = 1; 4818} 4819#else 4820static void __init enable_swap_cgroup(void) 4821{ 4822} 4823#endif 4824 4825static int mem_cgroup_soft_limit_tree_init(void) 4826{ 4827 struct mem_cgroup_tree_per_node *rtpn; 4828 struct mem_cgroup_tree_per_zone *rtpz; 4829 int tmp, node, zone; 4830 4831 for_each_node_state(node, N_POSSIBLE) { 4832 tmp = node; 4833 if (!node_state(node, N_NORMAL_MEMORY)) 4834 tmp = -1; 4835 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp); 4836 if (!rtpn) 4837 return 1; 4838 4839 soft_limit_tree.rb_tree_per_node[node] = rtpn; 4840 4841 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 4842 rtpz = &rtpn->rb_tree_per_zone[zone]; 4843 rtpz->rb_root = RB_ROOT; 4844 spin_lock_init(&rtpz->lock); 4845 } 4846 } 4847 return 0; 4848} 4849 4850static struct cgroup_subsys_state * __ref 4851mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 4852{ 4853 struct mem_cgroup *mem, *parent; 4854 long error = -ENOMEM; 4855 int node; 4856 4857 mem = mem_cgroup_alloc(); 4858 if (!mem) 4859 return ERR_PTR(error); 4860 4861 for_each_node_state(node, N_POSSIBLE) 4862 if (alloc_mem_cgroup_per_zone_info(mem, node)) 4863 goto free_out; 4864 4865 /* root ? */ 4866 if (cont->parent == NULL) { 4867 int cpu; 4868 enable_swap_cgroup(); 4869 parent = NULL; 4870 root_mem_cgroup = mem; 4871 if (mem_cgroup_soft_limit_tree_init()) 4872 goto free_out; 4873 for_each_possible_cpu(cpu) { 4874 struct memcg_stock_pcp *stock = 4875 &per_cpu(memcg_stock, cpu); 4876 INIT_WORK(&stock->work, drain_local_stock); 4877 } 4878 hotcpu_notifier(memcg_cpu_hotplug_callback, 0); 4879 } else { 4880 parent = mem_cgroup_from_cont(cont->parent); 4881 mem->use_hierarchy = parent->use_hierarchy; 4882 mem->oom_kill_disable = parent->oom_kill_disable; 4883 } 4884 4885 if (parent && parent->use_hierarchy) { 4886 res_counter_init(&mem->res, &parent->res); 4887 res_counter_init(&mem->memsw, &parent->memsw); 4888 /* 4889 * We increment refcnt of the parent to ensure that we can 4890 * safely access it on res_counter_charge/uncharge. 4891 * This refcnt will be decremented when freeing this 4892 * mem_cgroup(see mem_cgroup_put). 4893 */ 4894 mem_cgroup_get(parent); 4895 } else { 4896 res_counter_init(&mem->res, NULL); 4897 res_counter_init(&mem->memsw, NULL); 4898 } 4899 mem->last_scanned_child = 0; 4900 mem->last_scanned_node = MAX_NUMNODES; 4901 INIT_LIST_HEAD(&mem->oom_notify); 4902 4903 if (parent) 4904 mem->swappiness = get_swappiness(parent); 4905 atomic_set(&mem->refcnt, 1); 4906 mem->move_charge_at_immigrate = 0; 4907 mutex_init(&mem->thresholds_lock); 4908 return &mem->css; 4909free_out: 4910 __mem_cgroup_free(mem); 4911 root_mem_cgroup = NULL; 4912 return ERR_PTR(error); 4913} 4914 4915static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 4916 struct cgroup *cont) 4917{ 4918 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 4919 4920 return mem_cgroup_force_empty(mem, false); 4921} 4922 4923static void mem_cgroup_destroy(struct cgroup_subsys *ss, 4924 struct cgroup *cont) 4925{ 4926 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 4927 4928 mem_cgroup_put(mem); 4929} 4930 4931static int mem_cgroup_populate(struct cgroup_subsys *ss, 4932 struct cgroup *cont) 4933{ 4934 int ret; 4935 4936 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 4937 ARRAY_SIZE(mem_cgroup_files)); 4938 4939 if (!ret) 4940 ret = register_memsw_files(cont, ss); 4941 return ret; 4942} 4943 4944#ifdef CONFIG_MMU 4945/* Handlers for move charge at task migration. */ 4946#define PRECHARGE_COUNT_AT_ONCE 256 4947static int mem_cgroup_do_precharge(unsigned long count) 4948{ 4949 int ret = 0; 4950 int batch_count = PRECHARGE_COUNT_AT_ONCE; 4951 struct mem_cgroup *mem = mc.to; 4952 4953 if (mem_cgroup_is_root(mem)) { 4954 mc.precharge += count; 4955 /* we don't need css_get for root */ 4956 return ret; 4957 } 4958 /* try to charge at once */ 4959 if (count > 1) { 4960 struct res_counter *dummy; 4961 /* 4962 * "mem" cannot be under rmdir() because we've already checked 4963 * by cgroup_lock_live_cgroup() that it is not removed and we 4964 * are still under the same cgroup_mutex. So we can postpone 4965 * css_get(). 4966 */ 4967 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy)) 4968 goto one_by_one; 4969 if (do_swap_account && res_counter_charge(&mem->memsw, 4970 PAGE_SIZE * count, &dummy)) { 4971 res_counter_uncharge(&mem->res, PAGE_SIZE * count); 4972 goto one_by_one; 4973 } 4974 mc.precharge += count; 4975 return ret; 4976 } 4977one_by_one: 4978 /* fall back to one by one charge */ 4979 while (count--) { 4980 if (signal_pending(current)) { 4981 ret = -EINTR; 4982 break; 4983 } 4984 if (!batch_count--) { 4985 batch_count = PRECHARGE_COUNT_AT_ONCE; 4986 cond_resched(); 4987 } 4988 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false); 4989 if (ret || !mem) 4990 /* mem_cgroup_clear_mc() will do uncharge later */ 4991 return -ENOMEM; 4992 mc.precharge++; 4993 } 4994 return ret; 4995} 4996 4997/** 4998 * is_target_pte_for_mc - check a pte whether it is valid for move charge 4999 * @vma: the vma the pte to be checked belongs 5000 * @addr: the address corresponding to the pte to be checked 5001 * @ptent: the pte to be checked 5002 * @target: the pointer the target page or swap ent will be stored(can be NULL) 5003 * 5004 * Returns 5005 * 0(MC_TARGET_NONE): if the pte is not a target for move charge. 5006 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for 5007 * move charge. if @target is not NULL, the page is stored in target->page 5008 * with extra refcnt got(Callers should handle it). 5009 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a 5010 * target for charge migration. if @target is not NULL, the entry is stored 5011 * in target->ent. 5012 * 5013 * Called with pte lock held. 5014 */ 5015union mc_target { 5016 struct page *page; 5017 swp_entry_t ent; 5018}; 5019 5020enum mc_target_type { 5021 MC_TARGET_NONE, /* not used */ 5022 MC_TARGET_PAGE, 5023 MC_TARGET_SWAP, 5024}; 5025 5026static struct page *mc_handle_present_pte(struct vm_area_struct *vma, 5027 unsigned long addr, pte_t ptent) 5028{ 5029 struct page *page = vm_normal_page(vma, addr, ptent); 5030 5031 if (!page || !page_mapped(page)) 5032 return NULL; 5033 if (PageAnon(page)) { 5034 /* we don't move shared anon */ 5035 if (!move_anon() || page_mapcount(page) > 2) 5036 return NULL; 5037 } else if (!move_file()) 5038 /* we ignore mapcount for file pages */ 5039 return NULL; 5040 if (!get_page_unless_zero(page)) 5041 return NULL; 5042 5043 return page; 5044} 5045 5046static struct page *mc_handle_swap_pte(struct vm_area_struct *vma, 5047 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5048{ 5049 int usage_count; 5050 struct page *page = NULL; 5051 swp_entry_t ent = pte_to_swp_entry(ptent); 5052 5053 if (!move_anon() || non_swap_entry(ent)) 5054 return NULL; 5055 usage_count = mem_cgroup_count_swap_user(ent, &page); 5056 if (usage_count > 1) { /* we don't move shared anon */ 5057 if (page) 5058 put_page(page); 5059 return NULL; 5060 } 5061 if (do_swap_account) 5062 entry->val = ent.val; 5063 5064 return page; 5065} 5066 5067static struct page *mc_handle_file_pte(struct vm_area_struct *vma, 5068 unsigned long addr, pte_t ptent, swp_entry_t *entry) 5069{ 5070 struct page *page = NULL; 5071 struct inode *inode; 5072 struct address_space *mapping; 5073 pgoff_t pgoff; 5074 5075 if (!vma->vm_file) /* anonymous vma */ 5076 return NULL; 5077 if (!move_file()) 5078 return NULL; 5079 5080 inode = vma->vm_file->f_path.dentry->d_inode; 5081 mapping = vma->vm_file->f_mapping; 5082 if (pte_none(ptent)) 5083 pgoff = linear_page_index(vma, addr); 5084 else /* pte_file(ptent) is true */ 5085 pgoff = pte_to_pgoff(ptent); 5086 5087 /* page is moved even if it's not RSS of this task(page-faulted). */ 5088 if (!mapping_cap_swap_backed(mapping)) { /* normal file */ 5089 page = find_get_page(mapping, pgoff); 5090 } else { /* shmem/tmpfs file. we should take account of swap too. */ 5091 swp_entry_t ent; 5092 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent); 5093 if (do_swap_account) 5094 entry->val = ent.val; 5095 } 5096 5097 return page; 5098} 5099 5100static int is_target_pte_for_mc(struct vm_area_struct *vma, 5101 unsigned long addr, pte_t ptent, union mc_target *target) 5102{ 5103 struct page *page = NULL; 5104 struct page_cgroup *pc; 5105 int ret = 0; 5106 swp_entry_t ent = { .val = 0 }; 5107 5108 if (pte_present(ptent)) 5109 page = mc_handle_present_pte(vma, addr, ptent); 5110 else if (is_swap_pte(ptent)) 5111 page = mc_handle_swap_pte(vma, addr, ptent, &ent); 5112 else if (pte_none(ptent) || pte_file(ptent)) 5113 page = mc_handle_file_pte(vma, addr, ptent, &ent); 5114 5115 if (!page && !ent.val) 5116 return 0; 5117 if (page) { 5118 pc = lookup_page_cgroup(page); 5119 /* 5120 * Do only loose check w/o page_cgroup lock. 5121 * mem_cgroup_move_account() checks the pc is valid or not under 5122 * the lock. 5123 */ 5124 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) { 5125 ret = MC_TARGET_PAGE; 5126 if (target) 5127 target->page = page; 5128 } 5129 if (!ret || !target) 5130 put_page(page); 5131 } 5132 /* There is a swap entry and a page doesn't exist or isn't charged */ 5133 if (ent.val && !ret && 5134 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) { 5135 ret = MC_TARGET_SWAP; 5136 if (target) 5137 target->ent = ent; 5138 } 5139 return ret; 5140} 5141 5142static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd, 5143 unsigned long addr, unsigned long end, 5144 struct mm_walk *walk) 5145{ 5146 struct vm_area_struct *vma = walk->private; 5147 pte_t *pte; 5148 spinlock_t *ptl; 5149 5150 split_huge_page_pmd(walk->mm, pmd); 5151 5152 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5153 for (; addr != end; pte++, addr += PAGE_SIZE) 5154 if (is_target_pte_for_mc(vma, addr, *pte, NULL)) 5155 mc.precharge++; /* increment precharge temporarily */ 5156 pte_unmap_unlock(pte - 1, ptl); 5157 cond_resched(); 5158 5159 return 0; 5160} 5161 5162static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm) 5163{ 5164 unsigned long precharge; 5165 struct vm_area_struct *vma; 5166 5167 down_read(&mm->mmap_sem); 5168 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5169 struct mm_walk mem_cgroup_count_precharge_walk = { 5170 .pmd_entry = mem_cgroup_count_precharge_pte_range, 5171 .mm = mm, 5172 .private = vma, 5173 }; 5174 if (is_vm_hugetlb_page(vma)) 5175 continue; 5176 walk_page_range(vma->vm_start, vma->vm_end, 5177 &mem_cgroup_count_precharge_walk); 5178 } 5179 up_read(&mm->mmap_sem); 5180 5181 precharge = mc.precharge; 5182 mc.precharge = 0; 5183 5184 return precharge; 5185} 5186 5187static int mem_cgroup_precharge_mc(struct mm_struct *mm) 5188{ 5189 unsigned long precharge = mem_cgroup_count_precharge(mm); 5190 5191 VM_BUG_ON(mc.moving_task); 5192 mc.moving_task = current; 5193 return mem_cgroup_do_precharge(precharge); 5194} 5195 5196/* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */ 5197static void __mem_cgroup_clear_mc(void) 5198{ 5199 struct mem_cgroup *from = mc.from; 5200 struct mem_cgroup *to = mc.to; 5201 5202 /* we must uncharge all the leftover precharges from mc.to */ 5203 if (mc.precharge) { 5204 __mem_cgroup_cancel_charge(mc.to, mc.precharge); 5205 mc.precharge = 0; 5206 } 5207 /* 5208 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so 5209 * we must uncharge here. 5210 */ 5211 if (mc.moved_charge) { 5212 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge); 5213 mc.moved_charge = 0; 5214 } 5215 /* we must fixup refcnts and charges */ 5216 if (mc.moved_swap) { 5217 /* uncharge swap account from the old cgroup */ 5218 if (!mem_cgroup_is_root(mc.from)) 5219 res_counter_uncharge(&mc.from->memsw, 5220 PAGE_SIZE * mc.moved_swap); 5221 __mem_cgroup_put(mc.from, mc.moved_swap); 5222 5223 if (!mem_cgroup_is_root(mc.to)) { 5224 /* 5225 * we charged both to->res and to->memsw, so we should 5226 * uncharge to->res. 5227 */ 5228 res_counter_uncharge(&mc.to->res, 5229 PAGE_SIZE * mc.moved_swap); 5230 } 5231 /* we've already done mem_cgroup_get(mc.to) */ 5232 mc.moved_swap = 0; 5233 } 5234 memcg_oom_recover(from); 5235 memcg_oom_recover(to); 5236 wake_up_all(&mc.waitq); 5237} 5238 5239static void mem_cgroup_clear_mc(void) 5240{ 5241 struct mem_cgroup *from = mc.from; 5242 5243 /* 5244 * we must clear moving_task before waking up waiters at the end of 5245 * task migration. 5246 */ 5247 mc.moving_task = NULL; 5248 __mem_cgroup_clear_mc(); 5249 spin_lock(&mc.lock); 5250 mc.from = NULL; 5251 mc.to = NULL; 5252 spin_unlock(&mc.lock); 5253 mem_cgroup_end_move(from); 5254} 5255 5256static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5257 struct cgroup *cgroup, 5258 struct task_struct *p) 5259{ 5260 int ret = 0; 5261 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup); 5262 5263 if (mem->move_charge_at_immigrate) { 5264 struct mm_struct *mm; 5265 struct mem_cgroup *from = mem_cgroup_from_task(p); 5266 5267 VM_BUG_ON(from == mem); 5268 5269 mm = get_task_mm(p); 5270 if (!mm) 5271 return 0; 5272 /* We move charges only when we move a owner of the mm */ 5273 if (mm->owner == p) { 5274 VM_BUG_ON(mc.from); 5275 VM_BUG_ON(mc.to); 5276 VM_BUG_ON(mc.precharge); 5277 VM_BUG_ON(mc.moved_charge); 5278 VM_BUG_ON(mc.moved_swap); 5279 mem_cgroup_start_move(from); 5280 spin_lock(&mc.lock); 5281 mc.from = from; 5282 mc.to = mem; 5283 spin_unlock(&mc.lock); 5284 /* We set mc.moving_task later */ 5285 5286 ret = mem_cgroup_precharge_mc(mm); 5287 if (ret) 5288 mem_cgroup_clear_mc(); 5289 } 5290 mmput(mm); 5291 } 5292 return ret; 5293} 5294 5295static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5296 struct cgroup *cgroup, 5297 struct task_struct *p) 5298{ 5299 mem_cgroup_clear_mc(); 5300} 5301 5302static int mem_cgroup_move_charge_pte_range(pmd_t *pmd, 5303 unsigned long addr, unsigned long end, 5304 struct mm_walk *walk) 5305{ 5306 int ret = 0; 5307 struct vm_area_struct *vma = walk->private; 5308 pte_t *pte; 5309 spinlock_t *ptl; 5310 5311 split_huge_page_pmd(walk->mm, pmd); 5312retry: 5313 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl); 5314 for (; addr != end; addr += PAGE_SIZE) { 5315 pte_t ptent = *(pte++); 5316 union mc_target target; 5317 int type; 5318 struct page *page; 5319 struct page_cgroup *pc; 5320 swp_entry_t ent; 5321 5322 if (!mc.precharge) 5323 break; 5324 5325 type = is_target_pte_for_mc(vma, addr, ptent, &target); 5326 switch (type) { 5327 case MC_TARGET_PAGE: 5328 page = target.page; 5329 if (isolate_lru_page(page)) 5330 goto put; 5331 pc = lookup_page_cgroup(page); 5332 if (!mem_cgroup_move_account(page, 1, pc, 5333 mc.from, mc.to, false)) { 5334 mc.precharge--; 5335 /* we uncharge from mc.from later. */ 5336 mc.moved_charge++; 5337 } 5338 putback_lru_page(page); 5339put: /* is_target_pte_for_mc() gets the page */ 5340 put_page(page); 5341 break; 5342 case MC_TARGET_SWAP: 5343 ent = target.ent; 5344 if (!mem_cgroup_move_swap_account(ent, 5345 mc.from, mc.to, false)) { 5346 mc.precharge--; 5347 /* we fixup refcnts and charges later. */ 5348 mc.moved_swap++; 5349 } 5350 break; 5351 default: 5352 break; 5353 } 5354 } 5355 pte_unmap_unlock(pte - 1, ptl); 5356 cond_resched(); 5357 5358 if (addr != end) { 5359 /* 5360 * We have consumed all precharges we got in can_attach(). 5361 * We try charge one by one, but don't do any additional 5362 * charges to mc.to if we have failed in charge once in attach() 5363 * phase. 5364 */ 5365 ret = mem_cgroup_do_precharge(1); 5366 if (!ret) 5367 goto retry; 5368 } 5369 5370 return ret; 5371} 5372 5373static void mem_cgroup_move_charge(struct mm_struct *mm) 5374{ 5375 struct vm_area_struct *vma; 5376 5377 lru_add_drain_all(); 5378retry: 5379 if (unlikely(!down_read_trylock(&mm->mmap_sem))) { 5380 /* 5381 * Someone who are holding the mmap_sem might be waiting in 5382 * waitq. So we cancel all extra charges, wake up all waiters, 5383 * and retry. Because we cancel precharges, we might not be able 5384 * to move enough charges, but moving charge is a best-effort 5385 * feature anyway, so it wouldn't be a big problem. 5386 */ 5387 __mem_cgroup_clear_mc(); 5388 cond_resched(); 5389 goto retry; 5390 } 5391 for (vma = mm->mmap; vma; vma = vma->vm_next) { 5392 int ret; 5393 struct mm_walk mem_cgroup_move_charge_walk = { 5394 .pmd_entry = mem_cgroup_move_charge_pte_range, 5395 .mm = mm, 5396 .private = vma, 5397 }; 5398 if (is_vm_hugetlb_page(vma)) 5399 continue; 5400 ret = walk_page_range(vma->vm_start, vma->vm_end, 5401 &mem_cgroup_move_charge_walk); 5402 if (ret) 5403 /* 5404 * means we have consumed all precharges and failed in 5405 * doing additional charge. Just abandon here. 5406 */ 5407 break; 5408 } 5409 up_read(&mm->mmap_sem); 5410} 5411 5412static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5413 struct cgroup *cont, 5414 struct cgroup *old_cont, 5415 struct task_struct *p) 5416{ 5417 struct mm_struct *mm = get_task_mm(p); 5418 5419 if (mm) { 5420 if (mc.to) 5421 mem_cgroup_move_charge(mm); 5422 put_swap_token(mm); 5423 mmput(mm); 5424 } 5425 if (mc.to) 5426 mem_cgroup_clear_mc(); 5427} 5428#else /* !CONFIG_MMU */ 5429static int mem_cgroup_can_attach(struct cgroup_subsys *ss, 5430 struct cgroup *cgroup, 5431 struct task_struct *p) 5432{ 5433 return 0; 5434} 5435static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss, 5436 struct cgroup *cgroup, 5437 struct task_struct *p) 5438{ 5439} 5440static void mem_cgroup_move_task(struct cgroup_subsys *ss, 5441 struct cgroup *cont, 5442 struct cgroup *old_cont, 5443 struct task_struct *p) 5444{ 5445} 5446#endif 5447 5448struct cgroup_subsys mem_cgroup_subsys = { 5449 .name = "memory", 5450 .subsys_id = mem_cgroup_subsys_id, 5451 .create = mem_cgroup_create, 5452 .pre_destroy = mem_cgroup_pre_destroy, 5453 .destroy = mem_cgroup_destroy, 5454 .populate = mem_cgroup_populate, 5455 .can_attach = mem_cgroup_can_attach, 5456 .cancel_attach = mem_cgroup_cancel_attach, 5457 .attach = mem_cgroup_move_task, 5458 .early_init = 0, 5459 .use_id = 1, 5460}; 5461 5462#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 5463static int __init enable_swap_account(char *s) 5464{ 5465 /* consider enabled if no parameter or 1 is given */ 5466 if (!strcmp(s, "1")) 5467 really_do_swap_account = 1; 5468 else if (!strcmp(s, "0")) 5469 really_do_swap_account = 0; 5470 return 1; 5471} 5472__setup("swapaccount=", enable_swap_account); 5473 5474#endif 5475