memcontrol.c revision 261fb61a8bf6d3bd964ae6f1e6af49585d30db51
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 * This program is free software; you can redistribute it and/or modify 10 * it under the terms of the GNU General Public License as published by 11 * the Free Software Foundation; either version 2 of the License, or 12 * (at your option) any later version. 13 * 14 * This program is distributed in the hope that it will be useful, 15 * but WITHOUT ANY WARRANTY; without even the implied warranty of 16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 17 * GNU General Public License for more details. 18 */ 19 20#include <linux/res_counter.h> 21#include <linux/memcontrol.h> 22#include <linux/cgroup.h> 23#include <linux/mm.h> 24#include <linux/pagemap.h> 25#include <linux/smp.h> 26#include <linux/page-flags.h> 27#include <linux/backing-dev.h> 28#include <linux/bit_spinlock.h> 29#include <linux/rcupdate.h> 30#include <linux/limits.h> 31#include <linux/mutex.h> 32#include <linux/slab.h> 33#include <linux/swap.h> 34#include <linux/spinlock.h> 35#include <linux/fs.h> 36#include <linux/seq_file.h> 37#include <linux/vmalloc.h> 38#include <linux/mm_inline.h> 39#include <linux/page_cgroup.h> 40#include "internal.h" 41 42#include <asm/uaccess.h> 43 44struct cgroup_subsys mem_cgroup_subsys __read_mostly; 45#define MEM_CGROUP_RECLAIM_RETRIES 5 46struct mem_cgroup *root_mem_cgroup __read_mostly; 47 48#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 49/* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */ 50int do_swap_account __read_mostly; 51static int really_do_swap_account __initdata = 1; /* for remember boot option*/ 52#else 53#define do_swap_account (0) 54#endif 55 56static DEFINE_MUTEX(memcg_tasklist); /* can be hold under cgroup_mutex */ 57 58/* 59 * Statistics for memory cgroup. 60 */ 61enum mem_cgroup_stat_index { 62 /* 63 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 64 */ 65 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 66 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */ 67 MEM_CGROUP_STAT_MAPPED_FILE, /* # of pages charged as file rss */ 68 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ 69 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ 70 71 MEM_CGROUP_STAT_NSTATS, 72}; 73 74struct mem_cgroup_stat_cpu { 75 s64 count[MEM_CGROUP_STAT_NSTATS]; 76} ____cacheline_aligned_in_smp; 77 78struct mem_cgroup_stat { 79 struct mem_cgroup_stat_cpu cpustat[0]; 80}; 81 82/* 83 * For accounting under irq disable, no need for increment preempt count. 84 */ 85static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat, 86 enum mem_cgroup_stat_index idx, int val) 87{ 88 stat->count[idx] += val; 89} 90 91static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat, 92 enum mem_cgroup_stat_index idx) 93{ 94 int cpu; 95 s64 ret = 0; 96 for_each_possible_cpu(cpu) 97 ret += stat->cpustat[cpu].count[idx]; 98 return ret; 99} 100 101static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat) 102{ 103 s64 ret; 104 105 ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE); 106 ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS); 107 return ret; 108} 109 110/* 111 * per-zone information in memory controller. 112 */ 113struct mem_cgroup_per_zone { 114 /* 115 * spin_lock to protect the per cgroup LRU 116 */ 117 struct list_head lists[NR_LRU_LISTS]; 118 unsigned long count[NR_LRU_LISTS]; 119 120 struct zone_reclaim_stat reclaim_stat; 121}; 122/* Macro for accessing counter */ 123#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 124 125struct mem_cgroup_per_node { 126 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 127}; 128 129struct mem_cgroup_lru_info { 130 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 131}; 132 133/* 134 * The memory controller data structure. The memory controller controls both 135 * page cache and RSS per cgroup. We would eventually like to provide 136 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 137 * to help the administrator determine what knobs to tune. 138 * 139 * TODO: Add a water mark for the memory controller. Reclaim will begin when 140 * we hit the water mark. May be even add a low water mark, such that 141 * no reclaim occurs from a cgroup at it's low water mark, this is 142 * a feature that will be implemented much later in the future. 143 */ 144struct mem_cgroup { 145 struct cgroup_subsys_state css; 146 /* 147 * the counter to account for memory usage 148 */ 149 struct res_counter res; 150 /* 151 * the counter to account for mem+swap usage. 152 */ 153 struct res_counter memsw; 154 /* 155 * Per cgroup active and inactive list, similar to the 156 * per zone LRU lists. 157 */ 158 struct mem_cgroup_lru_info info; 159 160 /* 161 protect against reclaim related member. 162 */ 163 spinlock_t reclaim_param_lock; 164 165 int prev_priority; /* for recording reclaim priority */ 166 167 /* 168 * While reclaiming in a hiearchy, we cache the last child we 169 * reclaimed from. 170 */ 171 int last_scanned_child; 172 /* 173 * Should the accounting and control be hierarchical, per subtree? 174 */ 175 bool use_hierarchy; 176 unsigned long last_oom_jiffies; 177 atomic_t refcnt; 178 179 unsigned int swappiness; 180 181 /* set when res.limit == memsw.limit */ 182 bool memsw_is_minimum; 183 184 /* 185 * statistics. This must be placed at the end of memcg. 186 */ 187 struct mem_cgroup_stat stat; 188}; 189 190enum charge_type { 191 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 192 MEM_CGROUP_CHARGE_TYPE_MAPPED, 193 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 194 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 195 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 196 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */ 197 NR_CHARGE_TYPE, 198}; 199 200/* only for here (for easy reading.) */ 201#define PCGF_CACHE (1UL << PCG_CACHE) 202#define PCGF_USED (1UL << PCG_USED) 203#define PCGF_LOCK (1UL << PCG_LOCK) 204/* Not used, but added here for completeness */ 205#define PCGF_ACCT (1UL << PCG_ACCT) 206 207/* for encoding cft->private value on file */ 208#define _MEM (0) 209#define _MEMSWAP (1) 210#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 211#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 212#define MEMFILE_ATTR(val) ((val) & 0xffff) 213 214static void mem_cgroup_get(struct mem_cgroup *mem); 215static void mem_cgroup_put(struct mem_cgroup *mem); 216static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); 217 218static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, 219 struct page_cgroup *pc, 220 bool charge) 221{ 222 int val = (charge)? 1 : -1; 223 struct mem_cgroup_stat *stat = &mem->stat; 224 struct mem_cgroup_stat_cpu *cpustat; 225 int cpu = get_cpu(); 226 227 cpustat = &stat->cpustat[cpu]; 228 if (PageCgroupCache(pc)) 229 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val); 230 else 231 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val); 232 233 if (charge) 234 __mem_cgroup_stat_add_safe(cpustat, 235 MEM_CGROUP_STAT_PGPGIN_COUNT, 1); 236 else 237 __mem_cgroup_stat_add_safe(cpustat, 238 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1); 239 put_cpu(); 240} 241 242static struct mem_cgroup_per_zone * 243mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) 244{ 245 return &mem->info.nodeinfo[nid]->zoneinfo[zid]; 246} 247 248static struct mem_cgroup_per_zone * 249page_cgroup_zoneinfo(struct page_cgroup *pc) 250{ 251 struct mem_cgroup *mem = pc->mem_cgroup; 252 int nid = page_cgroup_nid(pc); 253 int zid = page_cgroup_zid(pc); 254 255 if (!mem) 256 return NULL; 257 258 return mem_cgroup_zoneinfo(mem, nid, zid); 259} 260 261static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, 262 enum lru_list idx) 263{ 264 int nid, zid; 265 struct mem_cgroup_per_zone *mz; 266 u64 total = 0; 267 268 for_each_online_node(nid) 269 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 270 mz = mem_cgroup_zoneinfo(mem, nid, zid); 271 total += MEM_CGROUP_ZSTAT(mz, idx); 272 } 273 return total; 274} 275 276static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 277{ 278 return container_of(cgroup_subsys_state(cont, 279 mem_cgroup_subsys_id), struct mem_cgroup, 280 css); 281} 282 283struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 284{ 285 /* 286 * mm_update_next_owner() may clear mm->owner to NULL 287 * if it races with swapoff, page migration, etc. 288 * So this can be called with p == NULL. 289 */ 290 if (unlikely(!p)) 291 return NULL; 292 293 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 294 struct mem_cgroup, css); 295} 296 297static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 298{ 299 struct mem_cgroup *mem = NULL; 300 301 if (!mm) 302 return NULL; 303 /* 304 * Because we have no locks, mm->owner's may be being moved to other 305 * cgroup. We use css_tryget() here even if this looks 306 * pessimistic (rather than adding locks here). 307 */ 308 rcu_read_lock(); 309 do { 310 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 311 if (unlikely(!mem)) 312 break; 313 } while (!css_tryget(&mem->css)); 314 rcu_read_unlock(); 315 return mem; 316} 317 318/* 319 * Call callback function against all cgroup under hierarchy tree. 320 */ 321static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data, 322 int (*func)(struct mem_cgroup *, void *)) 323{ 324 int found, ret, nextid; 325 struct cgroup_subsys_state *css; 326 struct mem_cgroup *mem; 327 328 if (!root->use_hierarchy) 329 return (*func)(root, data); 330 331 nextid = 1; 332 do { 333 ret = 0; 334 mem = NULL; 335 336 rcu_read_lock(); 337 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css, 338 &found); 339 if (css && css_tryget(css)) 340 mem = container_of(css, struct mem_cgroup, css); 341 rcu_read_unlock(); 342 343 if (mem) { 344 ret = (*func)(mem, data); 345 css_put(&mem->css); 346 } 347 nextid = found + 1; 348 } while (!ret && css); 349 350 return ret; 351} 352 353static inline bool mem_cgroup_is_root(struct mem_cgroup *mem) 354{ 355 return (mem == root_mem_cgroup); 356} 357 358/* 359 * Following LRU functions are allowed to be used without PCG_LOCK. 360 * Operations are called by routine of global LRU independently from memcg. 361 * What we have to take care of here is validness of pc->mem_cgroup. 362 * 363 * Changes to pc->mem_cgroup happens when 364 * 1. charge 365 * 2. moving account 366 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 367 * It is added to LRU before charge. 368 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 369 * When moving account, the page is not on LRU. It's isolated. 370 */ 371 372void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) 373{ 374 struct page_cgroup *pc; 375 struct mem_cgroup_per_zone *mz; 376 377 if (mem_cgroup_disabled()) 378 return; 379 pc = lookup_page_cgroup(page); 380 /* can happen while we handle swapcache. */ 381 if (!TestClearPageCgroupAcctLRU(pc)) 382 return; 383 VM_BUG_ON(!pc->mem_cgroup); 384 /* 385 * We don't check PCG_USED bit. It's cleared when the "page" is finally 386 * removed from global LRU. 387 */ 388 mz = page_cgroup_zoneinfo(pc); 389 MEM_CGROUP_ZSTAT(mz, lru) -= 1; 390 if (mem_cgroup_is_root(pc->mem_cgroup)) 391 return; 392 VM_BUG_ON(list_empty(&pc->lru)); 393 list_del_init(&pc->lru); 394 return; 395} 396 397void mem_cgroup_del_lru(struct page *page) 398{ 399 mem_cgroup_del_lru_list(page, page_lru(page)); 400} 401 402void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru) 403{ 404 struct mem_cgroup_per_zone *mz; 405 struct page_cgroup *pc; 406 407 if (mem_cgroup_disabled()) 408 return; 409 410 pc = lookup_page_cgroup(page); 411 /* 412 * Used bit is set without atomic ops but after smp_wmb(). 413 * For making pc->mem_cgroup visible, insert smp_rmb() here. 414 */ 415 smp_rmb(); 416 /* unused or root page is not rotated. */ 417 if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup)) 418 return; 419 mz = page_cgroup_zoneinfo(pc); 420 list_move(&pc->lru, &mz->lists[lru]); 421} 422 423void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru) 424{ 425 struct page_cgroup *pc; 426 struct mem_cgroup_per_zone *mz; 427 428 if (mem_cgroup_disabled()) 429 return; 430 pc = lookup_page_cgroup(page); 431 VM_BUG_ON(PageCgroupAcctLRU(pc)); 432 /* 433 * Used bit is set without atomic ops but after smp_wmb(). 434 * For making pc->mem_cgroup visible, insert smp_rmb() here. 435 */ 436 smp_rmb(); 437 if (!PageCgroupUsed(pc)) 438 return; 439 440 mz = page_cgroup_zoneinfo(pc); 441 MEM_CGROUP_ZSTAT(mz, lru) += 1; 442 SetPageCgroupAcctLRU(pc); 443 if (mem_cgroup_is_root(pc->mem_cgroup)) 444 return; 445 list_add(&pc->lru, &mz->lists[lru]); 446} 447 448/* 449 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to 450 * lru because the page may.be reused after it's fully uncharged (because of 451 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge 452 * it again. This function is only used to charge SwapCache. It's done under 453 * lock_page and expected that zone->lru_lock is never held. 454 */ 455static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) 456{ 457 unsigned long flags; 458 struct zone *zone = page_zone(page); 459 struct page_cgroup *pc = lookup_page_cgroup(page); 460 461 spin_lock_irqsave(&zone->lru_lock, flags); 462 /* 463 * Forget old LRU when this page_cgroup is *not* used. This Used bit 464 * is guarded by lock_page() because the page is SwapCache. 465 */ 466 if (!PageCgroupUsed(pc)) 467 mem_cgroup_del_lru_list(page, page_lru(page)); 468 spin_unlock_irqrestore(&zone->lru_lock, flags); 469} 470 471static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) 472{ 473 unsigned long flags; 474 struct zone *zone = page_zone(page); 475 struct page_cgroup *pc = lookup_page_cgroup(page); 476 477 spin_lock_irqsave(&zone->lru_lock, flags); 478 /* link when the page is linked to LRU but page_cgroup isn't */ 479 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) 480 mem_cgroup_add_lru_list(page, page_lru(page)); 481 spin_unlock_irqrestore(&zone->lru_lock, flags); 482} 483 484 485void mem_cgroup_move_lists(struct page *page, 486 enum lru_list from, enum lru_list to) 487{ 488 if (mem_cgroup_disabled()) 489 return; 490 mem_cgroup_del_lru_list(page, from); 491 mem_cgroup_add_lru_list(page, to); 492} 493 494int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) 495{ 496 int ret; 497 struct mem_cgroup *curr = NULL; 498 499 task_lock(task); 500 rcu_read_lock(); 501 curr = try_get_mem_cgroup_from_mm(task->mm); 502 rcu_read_unlock(); 503 task_unlock(task); 504 if (!curr) 505 return 0; 506 if (curr->use_hierarchy) 507 ret = css_is_ancestor(&curr->css, &mem->css); 508 else 509 ret = (curr == mem); 510 css_put(&curr->css); 511 return ret; 512} 513 514/* 515 * prev_priority control...this will be used in memory reclaim path. 516 */ 517int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem) 518{ 519 int prev_priority; 520 521 spin_lock(&mem->reclaim_param_lock); 522 prev_priority = mem->prev_priority; 523 spin_unlock(&mem->reclaim_param_lock); 524 525 return prev_priority; 526} 527 528void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority) 529{ 530 spin_lock(&mem->reclaim_param_lock); 531 if (priority < mem->prev_priority) 532 mem->prev_priority = priority; 533 spin_unlock(&mem->reclaim_param_lock); 534} 535 536void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority) 537{ 538 spin_lock(&mem->reclaim_param_lock); 539 mem->prev_priority = priority; 540 spin_unlock(&mem->reclaim_param_lock); 541} 542 543static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) 544{ 545 unsigned long active; 546 unsigned long inactive; 547 unsigned long gb; 548 unsigned long inactive_ratio; 549 550 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); 551 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); 552 553 gb = (inactive + active) >> (30 - PAGE_SHIFT); 554 if (gb) 555 inactive_ratio = int_sqrt(10 * gb); 556 else 557 inactive_ratio = 1; 558 559 if (present_pages) { 560 present_pages[0] = inactive; 561 present_pages[1] = active; 562 } 563 564 return inactive_ratio; 565} 566 567int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) 568{ 569 unsigned long active; 570 unsigned long inactive; 571 unsigned long present_pages[2]; 572 unsigned long inactive_ratio; 573 574 inactive_ratio = calc_inactive_ratio(memcg, present_pages); 575 576 inactive = present_pages[0]; 577 active = present_pages[1]; 578 579 if (inactive * inactive_ratio < active) 580 return 1; 581 582 return 0; 583} 584 585int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg) 586{ 587 unsigned long active; 588 unsigned long inactive; 589 590 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE); 591 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE); 592 593 return (active > inactive); 594} 595 596unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, 597 struct zone *zone, 598 enum lru_list lru) 599{ 600 int nid = zone->zone_pgdat->node_id; 601 int zid = zone_idx(zone); 602 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 603 604 return MEM_CGROUP_ZSTAT(mz, lru); 605} 606 607struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 608 struct zone *zone) 609{ 610 int nid = zone->zone_pgdat->node_id; 611 int zid = zone_idx(zone); 612 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 613 614 return &mz->reclaim_stat; 615} 616 617struct zone_reclaim_stat * 618mem_cgroup_get_reclaim_stat_from_page(struct page *page) 619{ 620 struct page_cgroup *pc; 621 struct mem_cgroup_per_zone *mz; 622 623 if (mem_cgroup_disabled()) 624 return NULL; 625 626 pc = lookup_page_cgroup(page); 627 /* 628 * Used bit is set without atomic ops but after smp_wmb(). 629 * For making pc->mem_cgroup visible, insert smp_rmb() here. 630 */ 631 smp_rmb(); 632 if (!PageCgroupUsed(pc)) 633 return NULL; 634 635 mz = page_cgroup_zoneinfo(pc); 636 if (!mz) 637 return NULL; 638 639 return &mz->reclaim_stat; 640} 641 642unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, 643 struct list_head *dst, 644 unsigned long *scanned, int order, 645 int mode, struct zone *z, 646 struct mem_cgroup *mem_cont, 647 int active, int file) 648{ 649 unsigned long nr_taken = 0; 650 struct page *page; 651 unsigned long scan; 652 LIST_HEAD(pc_list); 653 struct list_head *src; 654 struct page_cgroup *pc, *tmp; 655 int nid = z->zone_pgdat->node_id; 656 int zid = zone_idx(z); 657 struct mem_cgroup_per_zone *mz; 658 int lru = LRU_FILE * file + active; 659 int ret; 660 661 BUG_ON(!mem_cont); 662 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 663 src = &mz->lists[lru]; 664 665 scan = 0; 666 list_for_each_entry_safe_reverse(pc, tmp, src, lru) { 667 if (scan >= nr_to_scan) 668 break; 669 670 page = pc->page; 671 if (unlikely(!PageCgroupUsed(pc))) 672 continue; 673 if (unlikely(!PageLRU(page))) 674 continue; 675 676 scan++; 677 ret = __isolate_lru_page(page, mode, file); 678 switch (ret) { 679 case 0: 680 list_move(&page->lru, dst); 681 mem_cgroup_del_lru(page); 682 nr_taken++; 683 break; 684 case -EBUSY: 685 /* we don't affect global LRU but rotate in our LRU */ 686 mem_cgroup_rotate_lru_list(page, page_lru(page)); 687 break; 688 default: 689 break; 690 } 691 } 692 693 *scanned = scan; 694 return nr_taken; 695} 696 697#define mem_cgroup_from_res_counter(counter, member) \ 698 container_of(counter, struct mem_cgroup, member) 699 700static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem) 701{ 702 if (do_swap_account) { 703 if (res_counter_check_under_limit(&mem->res) && 704 res_counter_check_under_limit(&mem->memsw)) 705 return true; 706 } else 707 if (res_counter_check_under_limit(&mem->res)) 708 return true; 709 return false; 710} 711 712static unsigned int get_swappiness(struct mem_cgroup *memcg) 713{ 714 struct cgroup *cgrp = memcg->css.cgroup; 715 unsigned int swappiness; 716 717 /* root ? */ 718 if (cgrp->parent == NULL) 719 return vm_swappiness; 720 721 spin_lock(&memcg->reclaim_param_lock); 722 swappiness = memcg->swappiness; 723 spin_unlock(&memcg->reclaim_param_lock); 724 725 return swappiness; 726} 727 728static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data) 729{ 730 int *val = data; 731 (*val)++; 732 return 0; 733} 734 735/** 736 * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode. 737 * @memcg: The memory cgroup that went over limit 738 * @p: Task that is going to be killed 739 * 740 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 741 * enabled 742 */ 743void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 744{ 745 struct cgroup *task_cgrp; 746 struct cgroup *mem_cgrp; 747 /* 748 * Need a buffer in BSS, can't rely on allocations. The code relies 749 * on the assumption that OOM is serialized for memory controller. 750 * If this assumption is broken, revisit this code. 751 */ 752 static char memcg_name[PATH_MAX]; 753 int ret; 754 755 if (!memcg) 756 return; 757 758 759 rcu_read_lock(); 760 761 mem_cgrp = memcg->css.cgroup; 762 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 763 764 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 765 if (ret < 0) { 766 /* 767 * Unfortunately, we are unable to convert to a useful name 768 * But we'll still print out the usage information 769 */ 770 rcu_read_unlock(); 771 goto done; 772 } 773 rcu_read_unlock(); 774 775 printk(KERN_INFO "Task in %s killed", memcg_name); 776 777 rcu_read_lock(); 778 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 779 if (ret < 0) { 780 rcu_read_unlock(); 781 goto done; 782 } 783 rcu_read_unlock(); 784 785 /* 786 * Continues from above, so we don't need an KERN_ level 787 */ 788 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 789done: 790 791 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 792 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 793 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 794 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 795 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 796 "failcnt %llu\n", 797 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 798 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 799 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 800} 801 802/* 803 * This function returns the number of memcg under hierarchy tree. Returns 804 * 1(self count) if no children. 805 */ 806static int mem_cgroup_count_children(struct mem_cgroup *mem) 807{ 808 int num = 0; 809 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb); 810 return num; 811} 812 813/* 814 * Visit the first child (need not be the first child as per the ordering 815 * of the cgroup list, since we track last_scanned_child) of @mem and use 816 * that to reclaim free pages from. 817 */ 818static struct mem_cgroup * 819mem_cgroup_select_victim(struct mem_cgroup *root_mem) 820{ 821 struct mem_cgroup *ret = NULL; 822 struct cgroup_subsys_state *css; 823 int nextid, found; 824 825 if (!root_mem->use_hierarchy) { 826 css_get(&root_mem->css); 827 ret = root_mem; 828 } 829 830 while (!ret) { 831 rcu_read_lock(); 832 nextid = root_mem->last_scanned_child + 1; 833 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, 834 &found); 835 if (css && css_tryget(css)) 836 ret = container_of(css, struct mem_cgroup, css); 837 838 rcu_read_unlock(); 839 /* Updates scanning parameter */ 840 spin_lock(&root_mem->reclaim_param_lock); 841 if (!css) { 842 /* this means start scan from ID:1 */ 843 root_mem->last_scanned_child = 0; 844 } else 845 root_mem->last_scanned_child = found; 846 spin_unlock(&root_mem->reclaim_param_lock); 847 } 848 849 return ret; 850} 851 852/* 853 * Scan the hierarchy if needed to reclaim memory. We remember the last child 854 * we reclaimed from, so that we don't end up penalizing one child extensively 855 * based on its position in the children list. 856 * 857 * root_mem is the original ancestor that we've been reclaim from. 858 * 859 * We give up and return to the caller when we visit root_mem twice. 860 * (other groups can be removed while we're walking....) 861 * 862 * If shrink==true, for avoiding to free too much, this returns immedieately. 863 */ 864static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, 865 gfp_t gfp_mask, bool noswap, bool shrink) 866{ 867 struct mem_cgroup *victim; 868 int ret, total = 0; 869 int loop = 0; 870 871 /* If memsw_is_minimum==1, swap-out is of-no-use. */ 872 if (root_mem->memsw_is_minimum) 873 noswap = true; 874 875 while (loop < 2) { 876 victim = mem_cgroup_select_victim(root_mem); 877 if (victim == root_mem) 878 loop++; 879 if (!mem_cgroup_local_usage(&victim->stat)) { 880 /* this cgroup's local usage == 0 */ 881 css_put(&victim->css); 882 continue; 883 } 884 /* we use swappiness of local cgroup */ 885 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap, 886 get_swappiness(victim)); 887 css_put(&victim->css); 888 /* 889 * At shrinking usage, we can't check we should stop here or 890 * reclaim more. It's depends on callers. last_scanned_child 891 * will work enough for keeping fairness under tree. 892 */ 893 if (shrink) 894 return ret; 895 total += ret; 896 if (mem_cgroup_check_under_limit(root_mem)) 897 return 1 + total; 898 } 899 return total; 900} 901 902bool mem_cgroup_oom_called(struct task_struct *task) 903{ 904 bool ret = false; 905 struct mem_cgroup *mem; 906 struct mm_struct *mm; 907 908 rcu_read_lock(); 909 mm = task->mm; 910 if (!mm) 911 mm = &init_mm; 912 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 913 if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10)) 914 ret = true; 915 rcu_read_unlock(); 916 return ret; 917} 918 919static int record_last_oom_cb(struct mem_cgroup *mem, void *data) 920{ 921 mem->last_oom_jiffies = jiffies; 922 return 0; 923} 924 925static void record_last_oom(struct mem_cgroup *mem) 926{ 927 mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb); 928} 929 930/* 931 * Currently used to update mapped file statistics, but the routine can be 932 * generalized to update other statistics as well. 933 */ 934void mem_cgroup_update_mapped_file_stat(struct page *page, int val) 935{ 936 struct mem_cgroup *mem; 937 struct mem_cgroup_stat *stat; 938 struct mem_cgroup_stat_cpu *cpustat; 939 int cpu; 940 struct page_cgroup *pc; 941 942 if (!page_is_file_cache(page)) 943 return; 944 945 pc = lookup_page_cgroup(page); 946 if (unlikely(!pc)) 947 return; 948 949 lock_page_cgroup(pc); 950 mem = pc->mem_cgroup; 951 if (!mem) 952 goto done; 953 954 if (!PageCgroupUsed(pc)) 955 goto done; 956 957 /* 958 * Preemption is already disabled, we don't need get_cpu() 959 */ 960 cpu = smp_processor_id(); 961 stat = &mem->stat; 962 cpustat = &stat->cpustat[cpu]; 963 964 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val); 965done: 966 unlock_page_cgroup(pc); 967} 968 969/* 970 * Unlike exported interface, "oom" parameter is added. if oom==true, 971 * oom-killer can be invoked. 972 */ 973static int __mem_cgroup_try_charge(struct mm_struct *mm, 974 gfp_t gfp_mask, struct mem_cgroup **memcg, 975 bool oom) 976{ 977 struct mem_cgroup *mem, *mem_over_limit; 978 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 979 struct res_counter *fail_res; 980 981 if (unlikely(test_thread_flag(TIF_MEMDIE))) { 982 /* Don't account this! */ 983 *memcg = NULL; 984 return 0; 985 } 986 987 /* 988 * We always charge the cgroup the mm_struct belongs to. 989 * The mm_struct's mem_cgroup changes on task migration if the 990 * thread group leader migrates. It's possible that mm is not 991 * set, if so charge the init_mm (happens for pagecache usage). 992 */ 993 mem = *memcg; 994 if (likely(!mem)) { 995 mem = try_get_mem_cgroup_from_mm(mm); 996 *memcg = mem; 997 } else { 998 css_get(&mem->css); 999 } 1000 if (unlikely(!mem)) 1001 return 0; 1002 1003 VM_BUG_ON(css_is_removed(&mem->css)); 1004 1005 while (1) { 1006 int ret; 1007 bool noswap = false; 1008 1009 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res); 1010 if (likely(!ret)) { 1011 if (!do_swap_account) 1012 break; 1013 ret = res_counter_charge(&mem->memsw, PAGE_SIZE, 1014 &fail_res); 1015 if (likely(!ret)) 1016 break; 1017 /* mem+swap counter fails */ 1018 res_counter_uncharge(&mem->res, PAGE_SIZE); 1019 noswap = true; 1020 mem_over_limit = mem_cgroup_from_res_counter(fail_res, 1021 memsw); 1022 } else 1023 /* mem counter fails */ 1024 mem_over_limit = mem_cgroup_from_res_counter(fail_res, 1025 res); 1026 1027 if (!(gfp_mask & __GFP_WAIT)) 1028 goto nomem; 1029 1030 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask, 1031 noswap, false); 1032 if (ret) 1033 continue; 1034 1035 /* 1036 * try_to_free_mem_cgroup_pages() might not give us a full 1037 * picture of reclaim. Some pages are reclaimed and might be 1038 * moved to swap cache or just unmapped from the cgroup. 1039 * Check the limit again to see if the reclaim reduced the 1040 * current usage of the cgroup before giving up 1041 * 1042 */ 1043 if (mem_cgroup_check_under_limit(mem_over_limit)) 1044 continue; 1045 1046 if (!nr_retries--) { 1047 if (oom) { 1048 mutex_lock(&memcg_tasklist); 1049 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask); 1050 mutex_unlock(&memcg_tasklist); 1051 record_last_oom(mem_over_limit); 1052 } 1053 goto nomem; 1054 } 1055 } 1056 return 0; 1057nomem: 1058 css_put(&mem->css); 1059 return -ENOMEM; 1060} 1061 1062 1063/* 1064 * A helper function to get mem_cgroup from ID. must be called under 1065 * rcu_read_lock(). The caller must check css_is_removed() or some if 1066 * it's concern. (dropping refcnt from swap can be called against removed 1067 * memcg.) 1068 */ 1069static struct mem_cgroup *mem_cgroup_lookup(unsigned short id) 1070{ 1071 struct cgroup_subsys_state *css; 1072 1073 /* ID 0 is unused ID */ 1074 if (!id) 1075 return NULL; 1076 css = css_lookup(&mem_cgroup_subsys, id); 1077 if (!css) 1078 return NULL; 1079 return container_of(css, struct mem_cgroup, css); 1080} 1081 1082static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page) 1083{ 1084 struct mem_cgroup *mem; 1085 struct page_cgroup *pc; 1086 unsigned short id; 1087 swp_entry_t ent; 1088 1089 VM_BUG_ON(!PageLocked(page)); 1090 1091 if (!PageSwapCache(page)) 1092 return NULL; 1093 1094 pc = lookup_page_cgroup(page); 1095 lock_page_cgroup(pc); 1096 if (PageCgroupUsed(pc)) { 1097 mem = pc->mem_cgroup; 1098 if (mem && !css_tryget(&mem->css)) 1099 mem = NULL; 1100 } else { 1101 ent.val = page_private(page); 1102 id = lookup_swap_cgroup(ent); 1103 rcu_read_lock(); 1104 mem = mem_cgroup_lookup(id); 1105 if (mem && !css_tryget(&mem->css)) 1106 mem = NULL; 1107 rcu_read_unlock(); 1108 } 1109 unlock_page_cgroup(pc); 1110 return mem; 1111} 1112 1113/* 1114 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be 1115 * USED state. If already USED, uncharge and return. 1116 */ 1117 1118static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, 1119 struct page_cgroup *pc, 1120 enum charge_type ctype) 1121{ 1122 /* try_charge() can return NULL to *memcg, taking care of it. */ 1123 if (!mem) 1124 return; 1125 1126 lock_page_cgroup(pc); 1127 if (unlikely(PageCgroupUsed(pc))) { 1128 unlock_page_cgroup(pc); 1129 res_counter_uncharge(&mem->res, PAGE_SIZE); 1130 if (do_swap_account) 1131 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1132 css_put(&mem->css); 1133 return; 1134 } 1135 1136 pc->mem_cgroup = mem; 1137 /* 1138 * We access a page_cgroup asynchronously without lock_page_cgroup(). 1139 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup 1140 * is accessed after testing USED bit. To make pc->mem_cgroup visible 1141 * before USED bit, we need memory barrier here. 1142 * See mem_cgroup_add_lru_list(), etc. 1143 */ 1144 smp_wmb(); 1145 switch (ctype) { 1146 case MEM_CGROUP_CHARGE_TYPE_CACHE: 1147 case MEM_CGROUP_CHARGE_TYPE_SHMEM: 1148 SetPageCgroupCache(pc); 1149 SetPageCgroupUsed(pc); 1150 break; 1151 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 1152 ClearPageCgroupCache(pc); 1153 SetPageCgroupUsed(pc); 1154 break; 1155 default: 1156 break; 1157 } 1158 1159 mem_cgroup_charge_statistics(mem, pc, true); 1160 1161 unlock_page_cgroup(pc); 1162} 1163 1164/** 1165 * mem_cgroup_move_account - move account of the page 1166 * @pc: page_cgroup of the page. 1167 * @from: mem_cgroup which the page is moved from. 1168 * @to: mem_cgroup which the page is moved to. @from != @to. 1169 * 1170 * The caller must confirm following. 1171 * - page is not on LRU (isolate_page() is useful.) 1172 * 1173 * returns 0 at success, 1174 * returns -EBUSY when lock is busy or "pc" is unstable. 1175 * 1176 * This function does "uncharge" from old cgroup but doesn't do "charge" to 1177 * new cgroup. It should be done by a caller. 1178 */ 1179 1180static int mem_cgroup_move_account(struct page_cgroup *pc, 1181 struct mem_cgroup *from, struct mem_cgroup *to) 1182{ 1183 struct mem_cgroup_per_zone *from_mz, *to_mz; 1184 int nid, zid; 1185 int ret = -EBUSY; 1186 struct page *page; 1187 int cpu; 1188 struct mem_cgroup_stat *stat; 1189 struct mem_cgroup_stat_cpu *cpustat; 1190 1191 VM_BUG_ON(from == to); 1192 VM_BUG_ON(PageLRU(pc->page)); 1193 1194 nid = page_cgroup_nid(pc); 1195 zid = page_cgroup_zid(pc); 1196 from_mz = mem_cgroup_zoneinfo(from, nid, zid); 1197 to_mz = mem_cgroup_zoneinfo(to, nid, zid); 1198 1199 if (!trylock_page_cgroup(pc)) 1200 return ret; 1201 1202 if (!PageCgroupUsed(pc)) 1203 goto out; 1204 1205 if (pc->mem_cgroup != from) 1206 goto out; 1207 1208 res_counter_uncharge(&from->res, PAGE_SIZE); 1209 mem_cgroup_charge_statistics(from, pc, false); 1210 1211 page = pc->page; 1212 if (page_is_file_cache(page) && page_mapped(page)) { 1213 cpu = smp_processor_id(); 1214 /* Update mapped_file data for mem_cgroup "from" */ 1215 stat = &from->stat; 1216 cpustat = &stat->cpustat[cpu]; 1217 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, 1218 -1); 1219 1220 /* Update mapped_file data for mem_cgroup "to" */ 1221 stat = &to->stat; 1222 cpustat = &stat->cpustat[cpu]; 1223 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, 1224 1); 1225 } 1226 1227 if (do_swap_account) 1228 res_counter_uncharge(&from->memsw, PAGE_SIZE); 1229 css_put(&from->css); 1230 1231 css_get(&to->css); 1232 pc->mem_cgroup = to; 1233 mem_cgroup_charge_statistics(to, pc, true); 1234 ret = 0; 1235out: 1236 unlock_page_cgroup(pc); 1237 /* 1238 * We charges against "to" which may not have any tasks. Then, "to" 1239 * can be under rmdir(). But in current implementation, caller of 1240 * this function is just force_empty() and it's garanteed that 1241 * "to" is never removed. So, we don't check rmdir status here. 1242 */ 1243 return ret; 1244} 1245 1246/* 1247 * move charges to its parent. 1248 */ 1249 1250static int mem_cgroup_move_parent(struct page_cgroup *pc, 1251 struct mem_cgroup *child, 1252 gfp_t gfp_mask) 1253{ 1254 struct page *page = pc->page; 1255 struct cgroup *cg = child->css.cgroup; 1256 struct cgroup *pcg = cg->parent; 1257 struct mem_cgroup *parent; 1258 int ret; 1259 1260 /* Is ROOT ? */ 1261 if (!pcg) 1262 return -EINVAL; 1263 1264 1265 parent = mem_cgroup_from_cont(pcg); 1266 1267 1268 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false); 1269 if (ret || !parent) 1270 return ret; 1271 1272 if (!get_page_unless_zero(page)) { 1273 ret = -EBUSY; 1274 goto uncharge; 1275 } 1276 1277 ret = isolate_lru_page(page); 1278 1279 if (ret) 1280 goto cancel; 1281 1282 ret = mem_cgroup_move_account(pc, child, parent); 1283 1284 putback_lru_page(page); 1285 if (!ret) { 1286 put_page(page); 1287 /* drop extra refcnt by try_charge() */ 1288 css_put(&parent->css); 1289 return 0; 1290 } 1291 1292cancel: 1293 put_page(page); 1294uncharge: 1295 /* drop extra refcnt by try_charge() */ 1296 css_put(&parent->css); 1297 /* uncharge if move fails */ 1298 res_counter_uncharge(&parent->res, PAGE_SIZE); 1299 if (do_swap_account) 1300 res_counter_uncharge(&parent->memsw, PAGE_SIZE); 1301 return ret; 1302} 1303 1304/* 1305 * Charge the memory controller for page usage. 1306 * Return 1307 * 0 if the charge was successful 1308 * < 0 if the cgroup is over its limit 1309 */ 1310static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 1311 gfp_t gfp_mask, enum charge_type ctype, 1312 struct mem_cgroup *memcg) 1313{ 1314 struct mem_cgroup *mem; 1315 struct page_cgroup *pc; 1316 int ret; 1317 1318 pc = lookup_page_cgroup(page); 1319 /* can happen at boot */ 1320 if (unlikely(!pc)) 1321 return 0; 1322 prefetchw(pc); 1323 1324 mem = memcg; 1325 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true); 1326 if (ret || !mem) 1327 return ret; 1328 1329 __mem_cgroup_commit_charge(mem, pc, ctype); 1330 return 0; 1331} 1332 1333int mem_cgroup_newpage_charge(struct page *page, 1334 struct mm_struct *mm, gfp_t gfp_mask) 1335{ 1336 if (mem_cgroup_disabled()) 1337 return 0; 1338 if (PageCompound(page)) 1339 return 0; 1340 /* 1341 * If already mapped, we don't have to account. 1342 * If page cache, page->mapping has address_space. 1343 * But page->mapping may have out-of-use anon_vma pointer, 1344 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 1345 * is NULL. 1346 */ 1347 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 1348 return 0; 1349 if (unlikely(!mm)) 1350 mm = &init_mm; 1351 return mem_cgroup_charge_common(page, mm, gfp_mask, 1352 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL); 1353} 1354 1355static void 1356__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 1357 enum charge_type ctype); 1358 1359int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 1360 gfp_t gfp_mask) 1361{ 1362 struct mem_cgroup *mem = NULL; 1363 int ret; 1364 1365 if (mem_cgroup_disabled()) 1366 return 0; 1367 if (PageCompound(page)) 1368 return 0; 1369 /* 1370 * Corner case handling. This is called from add_to_page_cache() 1371 * in usual. But some FS (shmem) precharges this page before calling it 1372 * and call add_to_page_cache() with GFP_NOWAIT. 1373 * 1374 * For GFP_NOWAIT case, the page may be pre-charged before calling 1375 * add_to_page_cache(). (See shmem.c) check it here and avoid to call 1376 * charge twice. (It works but has to pay a bit larger cost.) 1377 * And when the page is SwapCache, it should take swap information 1378 * into account. This is under lock_page() now. 1379 */ 1380 if (!(gfp_mask & __GFP_WAIT)) { 1381 struct page_cgroup *pc; 1382 1383 1384 pc = lookup_page_cgroup(page); 1385 if (!pc) 1386 return 0; 1387 lock_page_cgroup(pc); 1388 if (PageCgroupUsed(pc)) { 1389 unlock_page_cgroup(pc); 1390 return 0; 1391 } 1392 unlock_page_cgroup(pc); 1393 } 1394 1395 if (unlikely(!mm && !mem)) 1396 mm = &init_mm; 1397 1398 if (page_is_file_cache(page)) 1399 return mem_cgroup_charge_common(page, mm, gfp_mask, 1400 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL); 1401 1402 /* shmem */ 1403 if (PageSwapCache(page)) { 1404 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 1405 if (!ret) 1406 __mem_cgroup_commit_charge_swapin(page, mem, 1407 MEM_CGROUP_CHARGE_TYPE_SHMEM); 1408 } else 1409 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 1410 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem); 1411 1412 return ret; 1413} 1414 1415/* 1416 * While swap-in, try_charge -> commit or cancel, the page is locked. 1417 * And when try_charge() successfully returns, one refcnt to memcg without 1418 * struct page_cgroup is aquired. This refcnt will be cumsumed by 1419 * "commit()" or removed by "cancel()" 1420 */ 1421int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 1422 struct page *page, 1423 gfp_t mask, struct mem_cgroup **ptr) 1424{ 1425 struct mem_cgroup *mem; 1426 int ret; 1427 1428 if (mem_cgroup_disabled()) 1429 return 0; 1430 1431 if (!do_swap_account) 1432 goto charge_cur_mm; 1433 /* 1434 * A racing thread's fault, or swapoff, may have already updated 1435 * the pte, and even removed page from swap cache: return success 1436 * to go on to do_swap_page()'s pte_same() test, which should fail. 1437 */ 1438 if (!PageSwapCache(page)) 1439 return 0; 1440 mem = try_get_mem_cgroup_from_swapcache(page); 1441 if (!mem) 1442 goto charge_cur_mm; 1443 *ptr = mem; 1444 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true); 1445 /* drop extra refcnt from tryget */ 1446 css_put(&mem->css); 1447 return ret; 1448charge_cur_mm: 1449 if (unlikely(!mm)) 1450 mm = &init_mm; 1451 return __mem_cgroup_try_charge(mm, mask, ptr, true); 1452} 1453 1454static void 1455__mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr, 1456 enum charge_type ctype) 1457{ 1458 struct page_cgroup *pc; 1459 1460 if (mem_cgroup_disabled()) 1461 return; 1462 if (!ptr) 1463 return; 1464 cgroup_exclude_rmdir(&ptr->css); 1465 pc = lookup_page_cgroup(page); 1466 mem_cgroup_lru_del_before_commit_swapcache(page); 1467 __mem_cgroup_commit_charge(ptr, pc, ctype); 1468 mem_cgroup_lru_add_after_commit_swapcache(page); 1469 /* 1470 * Now swap is on-memory. This means this page may be 1471 * counted both as mem and swap....double count. 1472 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 1473 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 1474 * may call delete_from_swap_cache() before reach here. 1475 */ 1476 if (do_swap_account && PageSwapCache(page)) { 1477 swp_entry_t ent = {.val = page_private(page)}; 1478 unsigned short id; 1479 struct mem_cgroup *memcg; 1480 1481 id = swap_cgroup_record(ent, 0); 1482 rcu_read_lock(); 1483 memcg = mem_cgroup_lookup(id); 1484 if (memcg) { 1485 /* 1486 * This recorded memcg can be obsolete one. So, avoid 1487 * calling css_tryget 1488 */ 1489 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 1490 mem_cgroup_put(memcg); 1491 } 1492 rcu_read_unlock(); 1493 } 1494 /* 1495 * At swapin, we may charge account against cgroup which has no tasks. 1496 * So, rmdir()->pre_destroy() can be called while we do this charge. 1497 * In that case, we need to call pre_destroy() again. check it here. 1498 */ 1499 cgroup_release_and_wakeup_rmdir(&ptr->css); 1500} 1501 1502void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) 1503{ 1504 __mem_cgroup_commit_charge_swapin(page, ptr, 1505 MEM_CGROUP_CHARGE_TYPE_MAPPED); 1506} 1507 1508void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) 1509{ 1510 if (mem_cgroup_disabled()) 1511 return; 1512 if (!mem) 1513 return; 1514 res_counter_uncharge(&mem->res, PAGE_SIZE); 1515 if (do_swap_account) 1516 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1517 css_put(&mem->css); 1518} 1519 1520 1521/* 1522 * uncharge if !page_mapped(page) 1523 */ 1524static struct mem_cgroup * 1525__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 1526{ 1527 struct page_cgroup *pc; 1528 struct mem_cgroup *mem = NULL; 1529 struct mem_cgroup_per_zone *mz; 1530 1531 if (mem_cgroup_disabled()) 1532 return NULL; 1533 1534 if (PageSwapCache(page)) 1535 return NULL; 1536 1537 /* 1538 * Check if our page_cgroup is valid 1539 */ 1540 pc = lookup_page_cgroup(page); 1541 if (unlikely(!pc || !PageCgroupUsed(pc))) 1542 return NULL; 1543 1544 lock_page_cgroup(pc); 1545 1546 mem = pc->mem_cgroup; 1547 1548 if (!PageCgroupUsed(pc)) 1549 goto unlock_out; 1550 1551 switch (ctype) { 1552 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 1553 case MEM_CGROUP_CHARGE_TYPE_DROP: 1554 if (page_mapped(page)) 1555 goto unlock_out; 1556 break; 1557 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 1558 if (!PageAnon(page)) { /* Shared memory */ 1559 if (page->mapping && !page_is_file_cache(page)) 1560 goto unlock_out; 1561 } else if (page_mapped(page)) /* Anon */ 1562 goto unlock_out; 1563 break; 1564 default: 1565 break; 1566 } 1567 1568 res_counter_uncharge(&mem->res, PAGE_SIZE); 1569 if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)) 1570 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1571 mem_cgroup_charge_statistics(mem, pc, false); 1572 1573 ClearPageCgroupUsed(pc); 1574 /* 1575 * pc->mem_cgroup is not cleared here. It will be accessed when it's 1576 * freed from LRU. This is safe because uncharged page is expected not 1577 * to be reused (freed soon). Exception is SwapCache, it's handled by 1578 * special functions. 1579 */ 1580 1581 mz = page_cgroup_zoneinfo(pc); 1582 unlock_page_cgroup(pc); 1583 1584 /* at swapout, this memcg will be accessed to record to swap */ 1585 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 1586 css_put(&mem->css); 1587 1588 return mem; 1589 1590unlock_out: 1591 unlock_page_cgroup(pc); 1592 return NULL; 1593} 1594 1595void mem_cgroup_uncharge_page(struct page *page) 1596{ 1597 /* early check. */ 1598 if (page_mapped(page)) 1599 return; 1600 if (page->mapping && !PageAnon(page)) 1601 return; 1602 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 1603} 1604 1605void mem_cgroup_uncharge_cache_page(struct page *page) 1606{ 1607 VM_BUG_ON(page_mapped(page)); 1608 VM_BUG_ON(page->mapping); 1609 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 1610} 1611 1612#ifdef CONFIG_SWAP 1613/* 1614 * called after __delete_from_swap_cache() and drop "page" account. 1615 * memcg information is recorded to swap_cgroup of "ent" 1616 */ 1617void 1618mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout) 1619{ 1620 struct mem_cgroup *memcg; 1621 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT; 1622 1623 if (!swapout) /* this was a swap cache but the swap is unused ! */ 1624 ctype = MEM_CGROUP_CHARGE_TYPE_DROP; 1625 1626 memcg = __mem_cgroup_uncharge_common(page, ctype); 1627 1628 /* record memcg information */ 1629 if (do_swap_account && swapout && memcg) { 1630 swap_cgroup_record(ent, css_id(&memcg->css)); 1631 mem_cgroup_get(memcg); 1632 } 1633 if (swapout && memcg) 1634 css_put(&memcg->css); 1635} 1636#endif 1637 1638#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 1639/* 1640 * called from swap_entry_free(). remove record in swap_cgroup and 1641 * uncharge "memsw" account. 1642 */ 1643void mem_cgroup_uncharge_swap(swp_entry_t ent) 1644{ 1645 struct mem_cgroup *memcg; 1646 unsigned short id; 1647 1648 if (!do_swap_account) 1649 return; 1650 1651 id = swap_cgroup_record(ent, 0); 1652 rcu_read_lock(); 1653 memcg = mem_cgroup_lookup(id); 1654 if (memcg) { 1655 /* 1656 * We uncharge this because swap is freed. 1657 * This memcg can be obsolete one. We avoid calling css_tryget 1658 */ 1659 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 1660 mem_cgroup_put(memcg); 1661 } 1662 rcu_read_unlock(); 1663} 1664#endif 1665 1666/* 1667 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 1668 * page belongs to. 1669 */ 1670int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr) 1671{ 1672 struct page_cgroup *pc; 1673 struct mem_cgroup *mem = NULL; 1674 int ret = 0; 1675 1676 if (mem_cgroup_disabled()) 1677 return 0; 1678 1679 pc = lookup_page_cgroup(page); 1680 lock_page_cgroup(pc); 1681 if (PageCgroupUsed(pc)) { 1682 mem = pc->mem_cgroup; 1683 css_get(&mem->css); 1684 } 1685 unlock_page_cgroup(pc); 1686 1687 if (mem) { 1688 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false); 1689 css_put(&mem->css); 1690 } 1691 *ptr = mem; 1692 return ret; 1693} 1694 1695/* remove redundant charge if migration failed*/ 1696void mem_cgroup_end_migration(struct mem_cgroup *mem, 1697 struct page *oldpage, struct page *newpage) 1698{ 1699 struct page *target, *unused; 1700 struct page_cgroup *pc; 1701 enum charge_type ctype; 1702 1703 if (!mem) 1704 return; 1705 cgroup_exclude_rmdir(&mem->css); 1706 /* at migration success, oldpage->mapping is NULL. */ 1707 if (oldpage->mapping) { 1708 target = oldpage; 1709 unused = NULL; 1710 } else { 1711 target = newpage; 1712 unused = oldpage; 1713 } 1714 1715 if (PageAnon(target)) 1716 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 1717 else if (page_is_file_cache(target)) 1718 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 1719 else 1720 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 1721 1722 /* unused page is not on radix-tree now. */ 1723 if (unused) 1724 __mem_cgroup_uncharge_common(unused, ctype); 1725 1726 pc = lookup_page_cgroup(target); 1727 /* 1728 * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup. 1729 * So, double-counting is effectively avoided. 1730 */ 1731 __mem_cgroup_commit_charge(mem, pc, ctype); 1732 1733 /* 1734 * Both of oldpage and newpage are still under lock_page(). 1735 * Then, we don't have to care about race in radix-tree. 1736 * But we have to be careful that this page is unmapped or not. 1737 * 1738 * There is a case for !page_mapped(). At the start of 1739 * migration, oldpage was mapped. But now, it's zapped. 1740 * But we know *target* page is not freed/reused under us. 1741 * mem_cgroup_uncharge_page() does all necessary checks. 1742 */ 1743 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) 1744 mem_cgroup_uncharge_page(target); 1745 /* 1746 * At migration, we may charge account against cgroup which has no tasks 1747 * So, rmdir()->pre_destroy() can be called while we do this charge. 1748 * In that case, we need to call pre_destroy() again. check it here. 1749 */ 1750 cgroup_release_and_wakeup_rmdir(&mem->css); 1751} 1752 1753/* 1754 * A call to try to shrink memory usage on charge failure at shmem's swapin. 1755 * Calling hierarchical_reclaim is not enough because we should update 1756 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM. 1757 * Moreover considering hierarchy, we should reclaim from the mem_over_limit, 1758 * not from the memcg which this page would be charged to. 1759 * try_charge_swapin does all of these works properly. 1760 */ 1761int mem_cgroup_shmem_charge_fallback(struct page *page, 1762 struct mm_struct *mm, 1763 gfp_t gfp_mask) 1764{ 1765 struct mem_cgroup *mem = NULL; 1766 int ret; 1767 1768 if (mem_cgroup_disabled()) 1769 return 0; 1770 1771 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem); 1772 if (!ret) 1773 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */ 1774 1775 return ret; 1776} 1777 1778static DEFINE_MUTEX(set_limit_mutex); 1779 1780static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 1781 unsigned long long val) 1782{ 1783 int retry_count; 1784 int progress; 1785 u64 memswlimit; 1786 int ret = 0; 1787 int children = mem_cgroup_count_children(memcg); 1788 u64 curusage, oldusage; 1789 1790 /* 1791 * For keeping hierarchical_reclaim simple, how long we should retry 1792 * is depends on callers. We set our retry-count to be function 1793 * of # of children which we should visit in this loop. 1794 */ 1795 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 1796 1797 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 1798 1799 while (retry_count) { 1800 if (signal_pending(current)) { 1801 ret = -EINTR; 1802 break; 1803 } 1804 /* 1805 * Rather than hide all in some function, I do this in 1806 * open coded manner. You see what this really does. 1807 * We have to guarantee mem->res.limit < mem->memsw.limit. 1808 */ 1809 mutex_lock(&set_limit_mutex); 1810 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1811 if (memswlimit < val) { 1812 ret = -EINVAL; 1813 mutex_unlock(&set_limit_mutex); 1814 break; 1815 } 1816 ret = res_counter_set_limit(&memcg->res, val); 1817 if (!ret) { 1818 if (memswlimit == val) 1819 memcg->memsw_is_minimum = true; 1820 else 1821 memcg->memsw_is_minimum = false; 1822 } 1823 mutex_unlock(&set_limit_mutex); 1824 1825 if (!ret) 1826 break; 1827 1828 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, 1829 false, true); 1830 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 1831 /* Usage is reduced ? */ 1832 if (curusage >= oldusage) 1833 retry_count--; 1834 else 1835 oldusage = curusage; 1836 } 1837 1838 return ret; 1839} 1840 1841static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 1842 unsigned long long val) 1843{ 1844 int retry_count; 1845 u64 memlimit, oldusage, curusage; 1846 int children = mem_cgroup_count_children(memcg); 1847 int ret = -EBUSY; 1848 1849 /* see mem_cgroup_resize_res_limit */ 1850 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 1851 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 1852 while (retry_count) { 1853 if (signal_pending(current)) { 1854 ret = -EINTR; 1855 break; 1856 } 1857 /* 1858 * Rather than hide all in some function, I do this in 1859 * open coded manner. You see what this really does. 1860 * We have to guarantee mem->res.limit < mem->memsw.limit. 1861 */ 1862 mutex_lock(&set_limit_mutex); 1863 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1864 if (memlimit > val) { 1865 ret = -EINVAL; 1866 mutex_unlock(&set_limit_mutex); 1867 break; 1868 } 1869 ret = res_counter_set_limit(&memcg->memsw, val); 1870 if (!ret) { 1871 if (memlimit == val) 1872 memcg->memsw_is_minimum = true; 1873 else 1874 memcg->memsw_is_minimum = false; 1875 } 1876 mutex_unlock(&set_limit_mutex); 1877 1878 if (!ret) 1879 break; 1880 1881 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true); 1882 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 1883 /* Usage is reduced ? */ 1884 if (curusage >= oldusage) 1885 retry_count--; 1886 else 1887 oldusage = curusage; 1888 } 1889 return ret; 1890} 1891 1892/* 1893 * This routine traverse page_cgroup in given list and drop them all. 1894 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 1895 */ 1896static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, 1897 int node, int zid, enum lru_list lru) 1898{ 1899 struct zone *zone; 1900 struct mem_cgroup_per_zone *mz; 1901 struct page_cgroup *pc, *busy; 1902 unsigned long flags, loop; 1903 struct list_head *list; 1904 int ret = 0; 1905 1906 zone = &NODE_DATA(node)->node_zones[zid]; 1907 mz = mem_cgroup_zoneinfo(mem, node, zid); 1908 list = &mz->lists[lru]; 1909 1910 loop = MEM_CGROUP_ZSTAT(mz, lru); 1911 /* give some margin against EBUSY etc...*/ 1912 loop += 256; 1913 busy = NULL; 1914 while (loop--) { 1915 ret = 0; 1916 spin_lock_irqsave(&zone->lru_lock, flags); 1917 if (list_empty(list)) { 1918 spin_unlock_irqrestore(&zone->lru_lock, flags); 1919 break; 1920 } 1921 pc = list_entry(list->prev, struct page_cgroup, lru); 1922 if (busy == pc) { 1923 list_move(&pc->lru, list); 1924 busy = 0; 1925 spin_unlock_irqrestore(&zone->lru_lock, flags); 1926 continue; 1927 } 1928 spin_unlock_irqrestore(&zone->lru_lock, flags); 1929 1930 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL); 1931 if (ret == -ENOMEM) 1932 break; 1933 1934 if (ret == -EBUSY || ret == -EINVAL) { 1935 /* found lock contention or "pc" is obsolete. */ 1936 busy = pc; 1937 cond_resched(); 1938 } else 1939 busy = NULL; 1940 } 1941 1942 if (!ret && !list_empty(list)) 1943 return -EBUSY; 1944 return ret; 1945} 1946 1947/* 1948 * make mem_cgroup's charge to be 0 if there is no task. 1949 * This enables deleting this mem_cgroup. 1950 */ 1951static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) 1952{ 1953 int ret; 1954 int node, zid, shrink; 1955 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 1956 struct cgroup *cgrp = mem->css.cgroup; 1957 1958 css_get(&mem->css); 1959 1960 shrink = 0; 1961 /* should free all ? */ 1962 if (free_all) 1963 goto try_to_free; 1964move_account: 1965 while (mem->res.usage > 0) { 1966 ret = -EBUSY; 1967 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 1968 goto out; 1969 ret = -EINTR; 1970 if (signal_pending(current)) 1971 goto out; 1972 /* This is for making all *used* pages to be on LRU. */ 1973 lru_add_drain_all(); 1974 ret = 0; 1975 for_each_node_state(node, N_HIGH_MEMORY) { 1976 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 1977 enum lru_list l; 1978 for_each_lru(l) { 1979 ret = mem_cgroup_force_empty_list(mem, 1980 node, zid, l); 1981 if (ret) 1982 break; 1983 } 1984 } 1985 if (ret) 1986 break; 1987 } 1988 /* it seems parent cgroup doesn't have enough mem */ 1989 if (ret == -ENOMEM) 1990 goto try_to_free; 1991 cond_resched(); 1992 } 1993 ret = 0; 1994out: 1995 css_put(&mem->css); 1996 return ret; 1997 1998try_to_free: 1999 /* returns EBUSY if there is a task or if we come here twice. */ 2000 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 2001 ret = -EBUSY; 2002 goto out; 2003 } 2004 /* we call try-to-free pages for make this cgroup empty */ 2005 lru_add_drain_all(); 2006 /* try to free all pages in this cgroup */ 2007 shrink = 1; 2008 while (nr_retries && mem->res.usage > 0) { 2009 int progress; 2010 2011 if (signal_pending(current)) { 2012 ret = -EINTR; 2013 goto out; 2014 } 2015 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, 2016 false, get_swappiness(mem)); 2017 if (!progress) { 2018 nr_retries--; 2019 /* maybe some writeback is necessary */ 2020 congestion_wait(BLK_RW_ASYNC, HZ/10); 2021 } 2022 2023 } 2024 lru_add_drain(); 2025 /* try move_account...there may be some *locked* pages. */ 2026 if (mem->res.usage) 2027 goto move_account; 2028 ret = 0; 2029 goto out; 2030} 2031 2032int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 2033{ 2034 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 2035} 2036 2037 2038static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 2039{ 2040 return mem_cgroup_from_cont(cont)->use_hierarchy; 2041} 2042 2043static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 2044 u64 val) 2045{ 2046 int retval = 0; 2047 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2048 struct cgroup *parent = cont->parent; 2049 struct mem_cgroup *parent_mem = NULL; 2050 2051 if (parent) 2052 parent_mem = mem_cgroup_from_cont(parent); 2053 2054 cgroup_lock(); 2055 /* 2056 * If parent's use_hiearchy is set, we can't make any modifications 2057 * in the child subtrees. If it is unset, then the change can 2058 * occur, provided the current cgroup has no children. 2059 * 2060 * For the root cgroup, parent_mem is NULL, we allow value to be 2061 * set if there are no children. 2062 */ 2063 if ((!parent_mem || !parent_mem->use_hierarchy) && 2064 (val == 1 || val == 0)) { 2065 if (list_empty(&cont->children)) 2066 mem->use_hierarchy = val; 2067 else 2068 retval = -EBUSY; 2069 } else 2070 retval = -EINVAL; 2071 cgroup_unlock(); 2072 2073 return retval; 2074} 2075 2076static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 2077{ 2078 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2079 u64 val = 0; 2080 int type, name; 2081 2082 type = MEMFILE_TYPE(cft->private); 2083 name = MEMFILE_ATTR(cft->private); 2084 switch (type) { 2085 case _MEM: 2086 val = res_counter_read_u64(&mem->res, name); 2087 break; 2088 case _MEMSWAP: 2089 val = res_counter_read_u64(&mem->memsw, name); 2090 break; 2091 default: 2092 BUG(); 2093 break; 2094 } 2095 return val; 2096} 2097/* 2098 * The user of this function is... 2099 * RES_LIMIT. 2100 */ 2101static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 2102 const char *buffer) 2103{ 2104 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 2105 int type, name; 2106 unsigned long long val; 2107 int ret; 2108 2109 type = MEMFILE_TYPE(cft->private); 2110 name = MEMFILE_ATTR(cft->private); 2111 switch (name) { 2112 case RES_LIMIT: 2113 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */ 2114 ret = -EINVAL; 2115 break; 2116 } 2117 /* This function does all necessary parse...reuse it */ 2118 ret = res_counter_memparse_write_strategy(buffer, &val); 2119 if (ret) 2120 break; 2121 if (type == _MEM) 2122 ret = mem_cgroup_resize_limit(memcg, val); 2123 else 2124 ret = mem_cgroup_resize_memsw_limit(memcg, val); 2125 break; 2126 default: 2127 ret = -EINVAL; /* should be BUG() ? */ 2128 break; 2129 } 2130 return ret; 2131} 2132 2133static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 2134 unsigned long long *mem_limit, unsigned long long *memsw_limit) 2135{ 2136 struct cgroup *cgroup; 2137 unsigned long long min_limit, min_memsw_limit, tmp; 2138 2139 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 2140 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 2141 cgroup = memcg->css.cgroup; 2142 if (!memcg->use_hierarchy) 2143 goto out; 2144 2145 while (cgroup->parent) { 2146 cgroup = cgroup->parent; 2147 memcg = mem_cgroup_from_cont(cgroup); 2148 if (!memcg->use_hierarchy) 2149 break; 2150 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 2151 min_limit = min(min_limit, tmp); 2152 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 2153 min_memsw_limit = min(min_memsw_limit, tmp); 2154 } 2155out: 2156 *mem_limit = min_limit; 2157 *memsw_limit = min_memsw_limit; 2158 return; 2159} 2160 2161static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 2162{ 2163 struct mem_cgroup *mem; 2164 int type, name; 2165 2166 mem = mem_cgroup_from_cont(cont); 2167 type = MEMFILE_TYPE(event); 2168 name = MEMFILE_ATTR(event); 2169 switch (name) { 2170 case RES_MAX_USAGE: 2171 if (type == _MEM) 2172 res_counter_reset_max(&mem->res); 2173 else 2174 res_counter_reset_max(&mem->memsw); 2175 break; 2176 case RES_FAILCNT: 2177 if (type == _MEM) 2178 res_counter_reset_failcnt(&mem->res); 2179 else 2180 res_counter_reset_failcnt(&mem->memsw); 2181 break; 2182 } 2183 return 0; 2184} 2185 2186 2187/* For read statistics */ 2188enum { 2189 MCS_CACHE, 2190 MCS_RSS, 2191 MCS_MAPPED_FILE, 2192 MCS_PGPGIN, 2193 MCS_PGPGOUT, 2194 MCS_INACTIVE_ANON, 2195 MCS_ACTIVE_ANON, 2196 MCS_INACTIVE_FILE, 2197 MCS_ACTIVE_FILE, 2198 MCS_UNEVICTABLE, 2199 NR_MCS_STAT, 2200}; 2201 2202struct mcs_total_stat { 2203 s64 stat[NR_MCS_STAT]; 2204}; 2205 2206struct { 2207 char *local_name; 2208 char *total_name; 2209} memcg_stat_strings[NR_MCS_STAT] = { 2210 {"cache", "total_cache"}, 2211 {"rss", "total_rss"}, 2212 {"mapped_file", "total_mapped_file"}, 2213 {"pgpgin", "total_pgpgin"}, 2214 {"pgpgout", "total_pgpgout"}, 2215 {"inactive_anon", "total_inactive_anon"}, 2216 {"active_anon", "total_active_anon"}, 2217 {"inactive_file", "total_inactive_file"}, 2218 {"active_file", "total_active_file"}, 2219 {"unevictable", "total_unevictable"} 2220}; 2221 2222 2223static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data) 2224{ 2225 struct mcs_total_stat *s = data; 2226 s64 val; 2227 2228 /* per cpu stat */ 2229 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE); 2230 s->stat[MCS_CACHE] += val * PAGE_SIZE; 2231 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS); 2232 s->stat[MCS_RSS] += val * PAGE_SIZE; 2233 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE); 2234 s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE; 2235 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT); 2236 s->stat[MCS_PGPGIN] += val; 2237 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT); 2238 s->stat[MCS_PGPGOUT] += val; 2239 2240 /* per zone stat */ 2241 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); 2242 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 2243 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); 2244 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 2245 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); 2246 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 2247 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); 2248 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 2249 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); 2250 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 2251 return 0; 2252} 2253 2254static void 2255mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 2256{ 2257 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat); 2258} 2259 2260static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 2261 struct cgroup_map_cb *cb) 2262{ 2263 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 2264 struct mcs_total_stat mystat; 2265 int i; 2266 2267 memset(&mystat, 0, sizeof(mystat)); 2268 mem_cgroup_get_local_stat(mem_cont, &mystat); 2269 2270 for (i = 0; i < NR_MCS_STAT; i++) 2271 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 2272 2273 /* Hierarchical information */ 2274 { 2275 unsigned long long limit, memsw_limit; 2276 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 2277 cb->fill(cb, "hierarchical_memory_limit", limit); 2278 if (do_swap_account) 2279 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 2280 } 2281 2282 memset(&mystat, 0, sizeof(mystat)); 2283 mem_cgroup_get_total_stat(mem_cont, &mystat); 2284 for (i = 0; i < NR_MCS_STAT; i++) 2285 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 2286 2287 2288#ifdef CONFIG_DEBUG_VM 2289 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); 2290 2291 { 2292 int nid, zid; 2293 struct mem_cgroup_per_zone *mz; 2294 unsigned long recent_rotated[2] = {0, 0}; 2295 unsigned long recent_scanned[2] = {0, 0}; 2296 2297 for_each_online_node(nid) 2298 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 2299 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 2300 2301 recent_rotated[0] += 2302 mz->reclaim_stat.recent_rotated[0]; 2303 recent_rotated[1] += 2304 mz->reclaim_stat.recent_rotated[1]; 2305 recent_scanned[0] += 2306 mz->reclaim_stat.recent_scanned[0]; 2307 recent_scanned[1] += 2308 mz->reclaim_stat.recent_scanned[1]; 2309 } 2310 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 2311 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 2312 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 2313 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 2314 } 2315#endif 2316 2317 return 0; 2318} 2319 2320static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 2321{ 2322 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 2323 2324 return get_swappiness(memcg); 2325} 2326 2327static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 2328 u64 val) 2329{ 2330 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 2331 struct mem_cgroup *parent; 2332 2333 if (val > 100) 2334 return -EINVAL; 2335 2336 if (cgrp->parent == NULL) 2337 return -EINVAL; 2338 2339 parent = mem_cgroup_from_cont(cgrp->parent); 2340 2341 cgroup_lock(); 2342 2343 /* If under hierarchy, only empty-root can set this value */ 2344 if ((parent->use_hierarchy) || 2345 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 2346 cgroup_unlock(); 2347 return -EINVAL; 2348 } 2349 2350 spin_lock(&memcg->reclaim_param_lock); 2351 memcg->swappiness = val; 2352 spin_unlock(&memcg->reclaim_param_lock); 2353 2354 cgroup_unlock(); 2355 2356 return 0; 2357} 2358 2359 2360static struct cftype mem_cgroup_files[] = { 2361 { 2362 .name = "usage_in_bytes", 2363 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 2364 .read_u64 = mem_cgroup_read, 2365 }, 2366 { 2367 .name = "max_usage_in_bytes", 2368 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 2369 .trigger = mem_cgroup_reset, 2370 .read_u64 = mem_cgroup_read, 2371 }, 2372 { 2373 .name = "limit_in_bytes", 2374 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 2375 .write_string = mem_cgroup_write, 2376 .read_u64 = mem_cgroup_read, 2377 }, 2378 { 2379 .name = "failcnt", 2380 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 2381 .trigger = mem_cgroup_reset, 2382 .read_u64 = mem_cgroup_read, 2383 }, 2384 { 2385 .name = "stat", 2386 .read_map = mem_control_stat_show, 2387 }, 2388 { 2389 .name = "force_empty", 2390 .trigger = mem_cgroup_force_empty_write, 2391 }, 2392 { 2393 .name = "use_hierarchy", 2394 .write_u64 = mem_cgroup_hierarchy_write, 2395 .read_u64 = mem_cgroup_hierarchy_read, 2396 }, 2397 { 2398 .name = "swappiness", 2399 .read_u64 = mem_cgroup_swappiness_read, 2400 .write_u64 = mem_cgroup_swappiness_write, 2401 }, 2402}; 2403 2404#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2405static struct cftype memsw_cgroup_files[] = { 2406 { 2407 .name = "memsw.usage_in_bytes", 2408 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 2409 .read_u64 = mem_cgroup_read, 2410 }, 2411 { 2412 .name = "memsw.max_usage_in_bytes", 2413 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 2414 .trigger = mem_cgroup_reset, 2415 .read_u64 = mem_cgroup_read, 2416 }, 2417 { 2418 .name = "memsw.limit_in_bytes", 2419 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 2420 .write_string = mem_cgroup_write, 2421 .read_u64 = mem_cgroup_read, 2422 }, 2423 { 2424 .name = "memsw.failcnt", 2425 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 2426 .trigger = mem_cgroup_reset, 2427 .read_u64 = mem_cgroup_read, 2428 }, 2429}; 2430 2431static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 2432{ 2433 if (!do_swap_account) 2434 return 0; 2435 return cgroup_add_files(cont, ss, memsw_cgroup_files, 2436 ARRAY_SIZE(memsw_cgroup_files)); 2437}; 2438#else 2439static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 2440{ 2441 return 0; 2442} 2443#endif 2444 2445static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 2446{ 2447 struct mem_cgroup_per_node *pn; 2448 struct mem_cgroup_per_zone *mz; 2449 enum lru_list l; 2450 int zone, tmp = node; 2451 /* 2452 * This routine is called against possible nodes. 2453 * But it's BUG to call kmalloc() against offline node. 2454 * 2455 * TODO: this routine can waste much memory for nodes which will 2456 * never be onlined. It's better to use memory hotplug callback 2457 * function. 2458 */ 2459 if (!node_state(node, N_NORMAL_MEMORY)) 2460 tmp = -1; 2461 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 2462 if (!pn) 2463 return 1; 2464 2465 mem->info.nodeinfo[node] = pn; 2466 memset(pn, 0, sizeof(*pn)); 2467 2468 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 2469 mz = &pn->zoneinfo[zone]; 2470 for_each_lru(l) 2471 INIT_LIST_HEAD(&mz->lists[l]); 2472 } 2473 return 0; 2474} 2475 2476static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 2477{ 2478 kfree(mem->info.nodeinfo[node]); 2479} 2480 2481static int mem_cgroup_size(void) 2482{ 2483 int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu); 2484 return sizeof(struct mem_cgroup) + cpustat_size; 2485} 2486 2487static struct mem_cgroup *mem_cgroup_alloc(void) 2488{ 2489 struct mem_cgroup *mem; 2490 int size = mem_cgroup_size(); 2491 2492 if (size < PAGE_SIZE) 2493 mem = kmalloc(size, GFP_KERNEL); 2494 else 2495 mem = vmalloc(size); 2496 2497 if (mem) 2498 memset(mem, 0, size); 2499 return mem; 2500} 2501 2502/* 2503 * At destroying mem_cgroup, references from swap_cgroup can remain. 2504 * (scanning all at force_empty is too costly...) 2505 * 2506 * Instead of clearing all references at force_empty, we remember 2507 * the number of reference from swap_cgroup and free mem_cgroup when 2508 * it goes down to 0. 2509 * 2510 * Removal of cgroup itself succeeds regardless of refs from swap. 2511 */ 2512 2513static void __mem_cgroup_free(struct mem_cgroup *mem) 2514{ 2515 int node; 2516 2517 free_css_id(&mem_cgroup_subsys, &mem->css); 2518 2519 for_each_node_state(node, N_POSSIBLE) 2520 free_mem_cgroup_per_zone_info(mem, node); 2521 2522 if (mem_cgroup_size() < PAGE_SIZE) 2523 kfree(mem); 2524 else 2525 vfree(mem); 2526} 2527 2528static void mem_cgroup_get(struct mem_cgroup *mem) 2529{ 2530 atomic_inc(&mem->refcnt); 2531} 2532 2533static void mem_cgroup_put(struct mem_cgroup *mem) 2534{ 2535 if (atomic_dec_and_test(&mem->refcnt)) { 2536 struct mem_cgroup *parent = parent_mem_cgroup(mem); 2537 __mem_cgroup_free(mem); 2538 if (parent) 2539 mem_cgroup_put(parent); 2540 } 2541} 2542 2543/* 2544 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 2545 */ 2546static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) 2547{ 2548 if (!mem->res.parent) 2549 return NULL; 2550 return mem_cgroup_from_res_counter(mem->res.parent, res); 2551} 2552 2553#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2554static void __init enable_swap_cgroup(void) 2555{ 2556 if (!mem_cgroup_disabled() && really_do_swap_account) 2557 do_swap_account = 1; 2558} 2559#else 2560static void __init enable_swap_cgroup(void) 2561{ 2562} 2563#endif 2564 2565static struct cgroup_subsys_state * __ref 2566mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 2567{ 2568 struct mem_cgroup *mem, *parent; 2569 long error = -ENOMEM; 2570 int node; 2571 2572 mem = mem_cgroup_alloc(); 2573 if (!mem) 2574 return ERR_PTR(error); 2575 2576 for_each_node_state(node, N_POSSIBLE) 2577 if (alloc_mem_cgroup_per_zone_info(mem, node)) 2578 goto free_out; 2579 /* root ? */ 2580 if (cont->parent == NULL) { 2581 enable_swap_cgroup(); 2582 parent = NULL; 2583 root_mem_cgroup = mem; 2584 } else { 2585 parent = mem_cgroup_from_cont(cont->parent); 2586 mem->use_hierarchy = parent->use_hierarchy; 2587 } 2588 2589 if (parent && parent->use_hierarchy) { 2590 res_counter_init(&mem->res, &parent->res); 2591 res_counter_init(&mem->memsw, &parent->memsw); 2592 /* 2593 * We increment refcnt of the parent to ensure that we can 2594 * safely access it on res_counter_charge/uncharge. 2595 * This refcnt will be decremented when freeing this 2596 * mem_cgroup(see mem_cgroup_put). 2597 */ 2598 mem_cgroup_get(parent); 2599 } else { 2600 res_counter_init(&mem->res, NULL); 2601 res_counter_init(&mem->memsw, NULL); 2602 } 2603 mem->last_scanned_child = 0; 2604 spin_lock_init(&mem->reclaim_param_lock); 2605 2606 if (parent) 2607 mem->swappiness = get_swappiness(parent); 2608 atomic_set(&mem->refcnt, 1); 2609 return &mem->css; 2610free_out: 2611 __mem_cgroup_free(mem); 2612 root_mem_cgroup = NULL; 2613 return ERR_PTR(error); 2614} 2615 2616static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 2617 struct cgroup *cont) 2618{ 2619 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2620 2621 return mem_cgroup_force_empty(mem, false); 2622} 2623 2624static void mem_cgroup_destroy(struct cgroup_subsys *ss, 2625 struct cgroup *cont) 2626{ 2627 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2628 2629 mem_cgroup_put(mem); 2630} 2631 2632static int mem_cgroup_populate(struct cgroup_subsys *ss, 2633 struct cgroup *cont) 2634{ 2635 int ret; 2636 2637 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 2638 ARRAY_SIZE(mem_cgroup_files)); 2639 2640 if (!ret) 2641 ret = register_memsw_files(cont, ss); 2642 return ret; 2643} 2644 2645static void mem_cgroup_move_task(struct cgroup_subsys *ss, 2646 struct cgroup *cont, 2647 struct cgroup *old_cont, 2648 struct task_struct *p, 2649 bool threadgroup) 2650{ 2651 mutex_lock(&memcg_tasklist); 2652 /* 2653 * FIXME: It's better to move charges of this process from old 2654 * memcg to new memcg. But it's just on TODO-List now. 2655 */ 2656 mutex_unlock(&memcg_tasklist); 2657} 2658 2659struct cgroup_subsys mem_cgroup_subsys = { 2660 .name = "memory", 2661 .subsys_id = mem_cgroup_subsys_id, 2662 .create = mem_cgroup_create, 2663 .pre_destroy = mem_cgroup_pre_destroy, 2664 .destroy = mem_cgroup_destroy, 2665 .populate = mem_cgroup_populate, 2666 .attach = mem_cgroup_move_task, 2667 .early_init = 0, 2668 .use_id = 1, 2669}; 2670 2671#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2672 2673static int __init disable_swap_account(char *s) 2674{ 2675 really_do_swap_account = 0; 2676 return 1; 2677} 2678__setup("noswapaccount", disable_swap_account); 2679#endif 2680