memcontrol.c revision c137b5ece4b111e46981aae7da77315b9909809f
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 46 47#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 48/* Turned on only when memory cgroup is enabled && really_do_swap_account = 0 */ 49int do_swap_account __read_mostly; 50static int really_do_swap_account __initdata = 1; /* for remember boot option*/ 51#else 52#define do_swap_account (0) 53#endif 54 55static DEFINE_MUTEX(memcg_tasklist); /* can be hold under cgroup_mutex */ 56 57/* 58 * Statistics for memory cgroup. 59 */ 60enum mem_cgroup_stat_index { 61 /* 62 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss. 63 */ 64 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */ 65 MEM_CGROUP_STAT_RSS, /* # of pages charged as rss */ 66 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */ 67 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */ 68 69 MEM_CGROUP_STAT_NSTATS, 70}; 71 72struct mem_cgroup_stat_cpu { 73 s64 count[MEM_CGROUP_STAT_NSTATS]; 74} ____cacheline_aligned_in_smp; 75 76struct mem_cgroup_stat { 77 struct mem_cgroup_stat_cpu cpustat[0]; 78}; 79 80/* 81 * For accounting under irq disable, no need for increment preempt count. 82 */ 83static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat, 84 enum mem_cgroup_stat_index idx, int val) 85{ 86 stat->count[idx] += val; 87} 88 89static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat, 90 enum mem_cgroup_stat_index idx) 91{ 92 int cpu; 93 s64 ret = 0; 94 for_each_possible_cpu(cpu) 95 ret += stat->cpustat[cpu].count[idx]; 96 return ret; 97} 98 99static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat) 100{ 101 s64 ret; 102 103 ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE); 104 ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS); 105 return ret; 106} 107 108/* 109 * per-zone information in memory controller. 110 */ 111struct mem_cgroup_per_zone { 112 /* 113 * spin_lock to protect the per cgroup LRU 114 */ 115 struct list_head lists[NR_LRU_LISTS]; 116 unsigned long count[NR_LRU_LISTS]; 117 118 struct zone_reclaim_stat reclaim_stat; 119}; 120/* Macro for accessing counter */ 121#define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)]) 122 123struct mem_cgroup_per_node { 124 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES]; 125}; 126 127struct mem_cgroup_lru_info { 128 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES]; 129}; 130 131/* 132 * The memory controller data structure. The memory controller controls both 133 * page cache and RSS per cgroup. We would eventually like to provide 134 * statistics based on the statistics developed by Rik Van Riel for clock-pro, 135 * to help the administrator determine what knobs to tune. 136 * 137 * TODO: Add a water mark for the memory controller. Reclaim will begin when 138 * we hit the water mark. May be even add a low water mark, such that 139 * no reclaim occurs from a cgroup at it's low water mark, this is 140 * a feature that will be implemented much later in the future. 141 */ 142struct mem_cgroup { 143 struct cgroup_subsys_state css; 144 /* 145 * the counter to account for memory usage 146 */ 147 struct res_counter res; 148 /* 149 * the counter to account for mem+swap usage. 150 */ 151 struct res_counter memsw; 152 /* 153 * Per cgroup active and inactive list, similar to the 154 * per zone LRU lists. 155 */ 156 struct mem_cgroup_lru_info info; 157 158 /* 159 protect against reclaim related member. 160 */ 161 spinlock_t reclaim_param_lock; 162 163 int prev_priority; /* for recording reclaim priority */ 164 165 /* 166 * While reclaiming in a hiearchy, we cache the last child we 167 * reclaimed from. 168 */ 169 int last_scanned_child; 170 /* 171 * Should the accounting and control be hierarchical, per subtree? 172 */ 173 bool use_hierarchy; 174 unsigned long last_oom_jiffies; 175 atomic_t refcnt; 176 177 unsigned int swappiness; 178 179 /* 180 * statistics. This must be placed at the end of memcg. 181 */ 182 struct mem_cgroup_stat stat; 183}; 184 185enum charge_type { 186 MEM_CGROUP_CHARGE_TYPE_CACHE = 0, 187 MEM_CGROUP_CHARGE_TYPE_MAPPED, 188 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */ 189 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */ 190 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */ 191 NR_CHARGE_TYPE, 192}; 193 194/* only for here (for easy reading.) */ 195#define PCGF_CACHE (1UL << PCG_CACHE) 196#define PCGF_USED (1UL << PCG_USED) 197#define PCGF_LOCK (1UL << PCG_LOCK) 198static const unsigned long 199pcg_default_flags[NR_CHARGE_TYPE] = { 200 PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* File Cache */ 201 PCGF_USED | PCGF_LOCK, /* Anon */ 202 PCGF_CACHE | PCGF_USED | PCGF_LOCK, /* Shmem */ 203 0, /* FORCE */ 204}; 205 206/* for encoding cft->private value on file */ 207#define _MEM (0) 208#define _MEMSWAP (1) 209#define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val)) 210#define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff) 211#define MEMFILE_ATTR(val) ((val) & 0xffff) 212 213static void mem_cgroup_get(struct mem_cgroup *mem); 214static void mem_cgroup_put(struct mem_cgroup *mem); 215static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem); 216 217static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, 218 struct page_cgroup *pc, 219 bool charge) 220{ 221 int val = (charge)? 1 : -1; 222 struct mem_cgroup_stat *stat = &mem->stat; 223 struct mem_cgroup_stat_cpu *cpustat; 224 int cpu = get_cpu(); 225 226 cpustat = &stat->cpustat[cpu]; 227 if (PageCgroupCache(pc)) 228 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val); 229 else 230 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val); 231 232 if (charge) 233 __mem_cgroup_stat_add_safe(cpustat, 234 MEM_CGROUP_STAT_PGPGIN_COUNT, 1); 235 else 236 __mem_cgroup_stat_add_safe(cpustat, 237 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1); 238 put_cpu(); 239} 240 241static struct mem_cgroup_per_zone * 242mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid) 243{ 244 return &mem->info.nodeinfo[nid]->zoneinfo[zid]; 245} 246 247static struct mem_cgroup_per_zone * 248page_cgroup_zoneinfo(struct page_cgroup *pc) 249{ 250 struct mem_cgroup *mem = pc->mem_cgroup; 251 int nid = page_cgroup_nid(pc); 252 int zid = page_cgroup_zid(pc); 253 254 if (!mem) 255 return NULL; 256 257 return mem_cgroup_zoneinfo(mem, nid, zid); 258} 259 260static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem, 261 enum lru_list idx) 262{ 263 int nid, zid; 264 struct mem_cgroup_per_zone *mz; 265 u64 total = 0; 266 267 for_each_online_node(nid) 268 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 269 mz = mem_cgroup_zoneinfo(mem, nid, zid); 270 total += MEM_CGROUP_ZSTAT(mz, idx); 271 } 272 return total; 273} 274 275static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont) 276{ 277 return container_of(cgroup_subsys_state(cont, 278 mem_cgroup_subsys_id), struct mem_cgroup, 279 css); 280} 281 282struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p) 283{ 284 /* 285 * mm_update_next_owner() may clear mm->owner to NULL 286 * if it races with swapoff, page migration, etc. 287 * So this can be called with p == NULL. 288 */ 289 if (unlikely(!p)) 290 return NULL; 291 292 return container_of(task_subsys_state(p, mem_cgroup_subsys_id), 293 struct mem_cgroup, css); 294} 295 296static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm) 297{ 298 struct mem_cgroup *mem = NULL; 299 300 if (!mm) 301 return NULL; 302 /* 303 * Because we have no locks, mm->owner's may be being moved to other 304 * cgroup. We use css_tryget() here even if this looks 305 * pessimistic (rather than adding locks here). 306 */ 307 rcu_read_lock(); 308 do { 309 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 310 if (unlikely(!mem)) 311 break; 312 } while (!css_tryget(&mem->css)); 313 rcu_read_unlock(); 314 return mem; 315} 316 317static bool mem_cgroup_is_obsolete(struct mem_cgroup *mem) 318{ 319 if (!mem) 320 return true; 321 return css_is_removed(&mem->css); 322} 323 324 325/* 326 * Call callback function against all cgroup under hierarchy tree. 327 */ 328static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data, 329 int (*func)(struct mem_cgroup *, void *)) 330{ 331 int found, ret, nextid; 332 struct cgroup_subsys_state *css; 333 struct mem_cgroup *mem; 334 335 if (!root->use_hierarchy) 336 return (*func)(root, data); 337 338 nextid = 1; 339 do { 340 ret = 0; 341 mem = NULL; 342 343 rcu_read_lock(); 344 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css, 345 &found); 346 if (css && css_tryget(css)) 347 mem = container_of(css, struct mem_cgroup, css); 348 rcu_read_unlock(); 349 350 if (mem) { 351 ret = (*func)(mem, data); 352 css_put(&mem->css); 353 } 354 nextid = found + 1; 355 } while (!ret && css); 356 357 return ret; 358} 359 360/* 361 * Following LRU functions are allowed to be used without PCG_LOCK. 362 * Operations are called by routine of global LRU independently from memcg. 363 * What we have to take care of here is validness of pc->mem_cgroup. 364 * 365 * Changes to pc->mem_cgroup happens when 366 * 1. charge 367 * 2. moving account 368 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache. 369 * It is added to LRU before charge. 370 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU. 371 * When moving account, the page is not on LRU. It's isolated. 372 */ 373 374void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru) 375{ 376 struct page_cgroup *pc; 377 struct mem_cgroup *mem; 378 struct mem_cgroup_per_zone *mz; 379 380 if (mem_cgroup_disabled()) 381 return; 382 pc = lookup_page_cgroup(page); 383 /* can happen while we handle swapcache. */ 384 if (list_empty(&pc->lru) || !pc->mem_cgroup) 385 return; 386 /* 387 * We don't check PCG_USED bit. It's cleared when the "page" is finally 388 * removed from global LRU. 389 */ 390 mz = page_cgroup_zoneinfo(pc); 391 mem = pc->mem_cgroup; 392 MEM_CGROUP_ZSTAT(mz, lru) -= 1; 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 page is not rotated. */ 417 if (!PageCgroupUsed(pc)) 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 /* 432 * Used bit is set without atomic ops but after smp_wmb(). 433 * For making pc->mem_cgroup visible, insert smp_rmb() here. 434 */ 435 smp_rmb(); 436 if (!PageCgroupUsed(pc)) 437 return; 438 439 mz = page_cgroup_zoneinfo(pc); 440 MEM_CGROUP_ZSTAT(mz, lru) += 1; 441 list_add(&pc->lru, &mz->lists[lru]); 442} 443 444/* 445 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to 446 * lru because the page may.be reused after it's fully uncharged (because of 447 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge 448 * it again. This function is only used to charge SwapCache. It's done under 449 * lock_page and expected that zone->lru_lock is never held. 450 */ 451static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page) 452{ 453 unsigned long flags; 454 struct zone *zone = page_zone(page); 455 struct page_cgroup *pc = lookup_page_cgroup(page); 456 457 spin_lock_irqsave(&zone->lru_lock, flags); 458 /* 459 * Forget old LRU when this page_cgroup is *not* used. This Used bit 460 * is guarded by lock_page() because the page is SwapCache. 461 */ 462 if (!PageCgroupUsed(pc)) 463 mem_cgroup_del_lru_list(page, page_lru(page)); 464 spin_unlock_irqrestore(&zone->lru_lock, flags); 465} 466 467static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page) 468{ 469 unsigned long flags; 470 struct zone *zone = page_zone(page); 471 struct page_cgroup *pc = lookup_page_cgroup(page); 472 473 spin_lock_irqsave(&zone->lru_lock, flags); 474 /* link when the page is linked to LRU but page_cgroup isn't */ 475 if (PageLRU(page) && list_empty(&pc->lru)) 476 mem_cgroup_add_lru_list(page, page_lru(page)); 477 spin_unlock_irqrestore(&zone->lru_lock, flags); 478} 479 480 481void mem_cgroup_move_lists(struct page *page, 482 enum lru_list from, enum lru_list to) 483{ 484 if (mem_cgroup_disabled()) 485 return; 486 mem_cgroup_del_lru_list(page, from); 487 mem_cgroup_add_lru_list(page, to); 488} 489 490int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem) 491{ 492 int ret; 493 struct mem_cgroup *curr = NULL; 494 495 task_lock(task); 496 rcu_read_lock(); 497 curr = try_get_mem_cgroup_from_mm(task->mm); 498 rcu_read_unlock(); 499 task_unlock(task); 500 if (!curr) 501 return 0; 502 if (curr->use_hierarchy) 503 ret = css_is_ancestor(&curr->css, &mem->css); 504 else 505 ret = (curr == mem); 506 css_put(&curr->css); 507 return ret; 508} 509 510/* 511 * prev_priority control...this will be used in memory reclaim path. 512 */ 513int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem) 514{ 515 int prev_priority; 516 517 spin_lock(&mem->reclaim_param_lock); 518 prev_priority = mem->prev_priority; 519 spin_unlock(&mem->reclaim_param_lock); 520 521 return prev_priority; 522} 523 524void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority) 525{ 526 spin_lock(&mem->reclaim_param_lock); 527 if (priority < mem->prev_priority) 528 mem->prev_priority = priority; 529 spin_unlock(&mem->reclaim_param_lock); 530} 531 532void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority) 533{ 534 spin_lock(&mem->reclaim_param_lock); 535 mem->prev_priority = priority; 536 spin_unlock(&mem->reclaim_param_lock); 537} 538 539static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages) 540{ 541 unsigned long active; 542 unsigned long inactive; 543 unsigned long gb; 544 unsigned long inactive_ratio; 545 546 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON); 547 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON); 548 549 gb = (inactive + active) >> (30 - PAGE_SHIFT); 550 if (gb) 551 inactive_ratio = int_sqrt(10 * gb); 552 else 553 inactive_ratio = 1; 554 555 if (present_pages) { 556 present_pages[0] = inactive; 557 present_pages[1] = active; 558 } 559 560 return inactive_ratio; 561} 562 563int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg) 564{ 565 unsigned long active; 566 unsigned long inactive; 567 unsigned long present_pages[2]; 568 unsigned long inactive_ratio; 569 570 inactive_ratio = calc_inactive_ratio(memcg, present_pages); 571 572 inactive = present_pages[0]; 573 active = present_pages[1]; 574 575 if (inactive * inactive_ratio < active) 576 return 1; 577 578 return 0; 579} 580 581unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg, 582 struct zone *zone, 583 enum lru_list lru) 584{ 585 int nid = zone->zone_pgdat->node_id; 586 int zid = zone_idx(zone); 587 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 588 589 return MEM_CGROUP_ZSTAT(mz, lru); 590} 591 592struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg, 593 struct zone *zone) 594{ 595 int nid = zone->zone_pgdat->node_id; 596 int zid = zone_idx(zone); 597 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid); 598 599 return &mz->reclaim_stat; 600} 601 602struct zone_reclaim_stat * 603mem_cgroup_get_reclaim_stat_from_page(struct page *page) 604{ 605 struct page_cgroup *pc; 606 struct mem_cgroup_per_zone *mz; 607 608 if (mem_cgroup_disabled()) 609 return NULL; 610 611 pc = lookup_page_cgroup(page); 612 /* 613 * Used bit is set without atomic ops but after smp_wmb(). 614 * For making pc->mem_cgroup visible, insert smp_rmb() here. 615 */ 616 smp_rmb(); 617 if (!PageCgroupUsed(pc)) 618 return NULL; 619 620 mz = page_cgroup_zoneinfo(pc); 621 if (!mz) 622 return NULL; 623 624 return &mz->reclaim_stat; 625} 626 627unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan, 628 struct list_head *dst, 629 unsigned long *scanned, int order, 630 int mode, struct zone *z, 631 struct mem_cgroup *mem_cont, 632 int active, int file) 633{ 634 unsigned long nr_taken = 0; 635 struct page *page; 636 unsigned long scan; 637 LIST_HEAD(pc_list); 638 struct list_head *src; 639 struct page_cgroup *pc, *tmp; 640 int nid = z->zone_pgdat->node_id; 641 int zid = zone_idx(z); 642 struct mem_cgroup_per_zone *mz; 643 int lru = LRU_FILE * !!file + !!active; 644 645 BUG_ON(!mem_cont); 646 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 647 src = &mz->lists[lru]; 648 649 scan = 0; 650 list_for_each_entry_safe_reverse(pc, tmp, src, lru) { 651 if (scan >= nr_to_scan) 652 break; 653 654 page = pc->page; 655 if (unlikely(!PageCgroupUsed(pc))) 656 continue; 657 if (unlikely(!PageLRU(page))) 658 continue; 659 660 scan++; 661 if (__isolate_lru_page(page, mode, file) == 0) { 662 list_move(&page->lru, dst); 663 nr_taken++; 664 } 665 } 666 667 *scanned = scan; 668 return nr_taken; 669} 670 671#define mem_cgroup_from_res_counter(counter, member) \ 672 container_of(counter, struct mem_cgroup, member) 673 674static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem) 675{ 676 if (do_swap_account) { 677 if (res_counter_check_under_limit(&mem->res) && 678 res_counter_check_under_limit(&mem->memsw)) 679 return true; 680 } else 681 if (res_counter_check_under_limit(&mem->res)) 682 return true; 683 return false; 684} 685 686static unsigned int get_swappiness(struct mem_cgroup *memcg) 687{ 688 struct cgroup *cgrp = memcg->css.cgroup; 689 unsigned int swappiness; 690 691 /* root ? */ 692 if (cgrp->parent == NULL) 693 return vm_swappiness; 694 695 spin_lock(&memcg->reclaim_param_lock); 696 swappiness = memcg->swappiness; 697 spin_unlock(&memcg->reclaim_param_lock); 698 699 return swappiness; 700} 701 702static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data) 703{ 704 int *val = data; 705 (*val)++; 706 return 0; 707} 708 709/** 710 * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode. 711 * @memcg: The memory cgroup that went over limit 712 * @p: Task that is going to be killed 713 * 714 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is 715 * enabled 716 */ 717void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p) 718{ 719 struct cgroup *task_cgrp; 720 struct cgroup *mem_cgrp; 721 /* 722 * Need a buffer in BSS, can't rely on allocations. The code relies 723 * on the assumption that OOM is serialized for memory controller. 724 * If this assumption is broken, revisit this code. 725 */ 726 static char memcg_name[PATH_MAX]; 727 int ret; 728 729 if (!memcg) 730 return; 731 732 733 rcu_read_lock(); 734 735 mem_cgrp = memcg->css.cgroup; 736 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id); 737 738 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX); 739 if (ret < 0) { 740 /* 741 * Unfortunately, we are unable to convert to a useful name 742 * But we'll still print out the usage information 743 */ 744 rcu_read_unlock(); 745 goto done; 746 } 747 rcu_read_unlock(); 748 749 printk(KERN_INFO "Task in %s killed", memcg_name); 750 751 rcu_read_lock(); 752 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX); 753 if (ret < 0) { 754 rcu_read_unlock(); 755 goto done; 756 } 757 rcu_read_unlock(); 758 759 /* 760 * Continues from above, so we don't need an KERN_ level 761 */ 762 printk(KERN_CONT " as a result of limit of %s\n", memcg_name); 763done: 764 765 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n", 766 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10, 767 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10, 768 res_counter_read_u64(&memcg->res, RES_FAILCNT)); 769 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, " 770 "failcnt %llu\n", 771 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10, 772 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10, 773 res_counter_read_u64(&memcg->memsw, RES_FAILCNT)); 774} 775 776/* 777 * This function returns the number of memcg under hierarchy tree. Returns 778 * 1(self count) if no children. 779 */ 780static int mem_cgroup_count_children(struct mem_cgroup *mem) 781{ 782 int num = 0; 783 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb); 784 return num; 785} 786 787/* 788 * Visit the first child (need not be the first child as per the ordering 789 * of the cgroup list, since we track last_scanned_child) of @mem and use 790 * that to reclaim free pages from. 791 */ 792static struct mem_cgroup * 793mem_cgroup_select_victim(struct mem_cgroup *root_mem) 794{ 795 struct mem_cgroup *ret = NULL; 796 struct cgroup_subsys_state *css; 797 int nextid, found; 798 799 if (!root_mem->use_hierarchy) { 800 css_get(&root_mem->css); 801 ret = root_mem; 802 } 803 804 while (!ret) { 805 rcu_read_lock(); 806 nextid = root_mem->last_scanned_child + 1; 807 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css, 808 &found); 809 if (css && css_tryget(css)) 810 ret = container_of(css, struct mem_cgroup, css); 811 812 rcu_read_unlock(); 813 /* Updates scanning parameter */ 814 spin_lock(&root_mem->reclaim_param_lock); 815 if (!css) { 816 /* this means start scan from ID:1 */ 817 root_mem->last_scanned_child = 0; 818 } else 819 root_mem->last_scanned_child = found; 820 spin_unlock(&root_mem->reclaim_param_lock); 821 } 822 823 return ret; 824} 825 826/* 827 * Scan the hierarchy if needed to reclaim memory. We remember the last child 828 * we reclaimed from, so that we don't end up penalizing one child extensively 829 * based on its position in the children list. 830 * 831 * root_mem is the original ancestor that we've been reclaim from. 832 * 833 * We give up and return to the caller when we visit root_mem twice. 834 * (other groups can be removed while we're walking....) 835 * 836 * If shrink==true, for avoiding to free too much, this returns immedieately. 837 */ 838static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem, 839 gfp_t gfp_mask, bool noswap, bool shrink) 840{ 841 struct mem_cgroup *victim; 842 int ret, total = 0; 843 int loop = 0; 844 845 while (loop < 2) { 846 victim = mem_cgroup_select_victim(root_mem); 847 if (victim == root_mem) 848 loop++; 849 if (!mem_cgroup_local_usage(&victim->stat)) { 850 /* this cgroup's local usage == 0 */ 851 css_put(&victim->css); 852 continue; 853 } 854 /* we use swappiness of local cgroup */ 855 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask, noswap, 856 get_swappiness(victim)); 857 css_put(&victim->css); 858 /* 859 * At shrinking usage, we can't check we should stop here or 860 * reclaim more. It's depends on callers. last_scanned_child 861 * will work enough for keeping fairness under tree. 862 */ 863 if (shrink) 864 return ret; 865 total += ret; 866 if (mem_cgroup_check_under_limit(root_mem)) 867 return 1 + total; 868 } 869 return total; 870} 871 872bool mem_cgroup_oom_called(struct task_struct *task) 873{ 874 bool ret = false; 875 struct mem_cgroup *mem; 876 struct mm_struct *mm; 877 878 rcu_read_lock(); 879 mm = task->mm; 880 if (!mm) 881 mm = &init_mm; 882 mem = mem_cgroup_from_task(rcu_dereference(mm->owner)); 883 if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10)) 884 ret = true; 885 rcu_read_unlock(); 886 return ret; 887} 888 889static int record_last_oom_cb(struct mem_cgroup *mem, void *data) 890{ 891 mem->last_oom_jiffies = jiffies; 892 return 0; 893} 894 895static void record_last_oom(struct mem_cgroup *mem) 896{ 897 mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb); 898} 899 900 901/* 902 * Unlike exported interface, "oom" parameter is added. if oom==true, 903 * oom-killer can be invoked. 904 */ 905static int __mem_cgroup_try_charge(struct mm_struct *mm, 906 gfp_t gfp_mask, struct mem_cgroup **memcg, 907 bool oom) 908{ 909 struct mem_cgroup *mem, *mem_over_limit; 910 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 911 struct res_counter *fail_res; 912 913 if (unlikely(test_thread_flag(TIF_MEMDIE))) { 914 /* Don't account this! */ 915 *memcg = NULL; 916 return 0; 917 } 918 919 /* 920 * We always charge the cgroup the mm_struct belongs to. 921 * The mm_struct's mem_cgroup changes on task migration if the 922 * thread group leader migrates. It's possible that mm is not 923 * set, if so charge the init_mm (happens for pagecache usage). 924 */ 925 mem = *memcg; 926 if (likely(!mem)) { 927 mem = try_get_mem_cgroup_from_mm(mm); 928 *memcg = mem; 929 } else { 930 css_get(&mem->css); 931 } 932 if (unlikely(!mem)) 933 return 0; 934 935 VM_BUG_ON(mem_cgroup_is_obsolete(mem)); 936 937 while (1) { 938 int ret; 939 bool noswap = false; 940 941 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res); 942 if (likely(!ret)) { 943 if (!do_swap_account) 944 break; 945 ret = res_counter_charge(&mem->memsw, PAGE_SIZE, 946 &fail_res); 947 if (likely(!ret)) 948 break; 949 /* mem+swap counter fails */ 950 res_counter_uncharge(&mem->res, PAGE_SIZE); 951 noswap = true; 952 mem_over_limit = mem_cgroup_from_res_counter(fail_res, 953 memsw); 954 } else 955 /* mem counter fails */ 956 mem_over_limit = mem_cgroup_from_res_counter(fail_res, 957 res); 958 959 if (!(gfp_mask & __GFP_WAIT)) 960 goto nomem; 961 962 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, gfp_mask, 963 noswap, false); 964 if (ret) 965 continue; 966 967 /* 968 * try_to_free_mem_cgroup_pages() might not give us a full 969 * picture of reclaim. Some pages are reclaimed and might be 970 * moved to swap cache or just unmapped from the cgroup. 971 * Check the limit again to see if the reclaim reduced the 972 * current usage of the cgroup before giving up 973 * 974 */ 975 if (mem_cgroup_check_under_limit(mem_over_limit)) 976 continue; 977 978 if (!nr_retries--) { 979 if (oom) { 980 mutex_lock(&memcg_tasklist); 981 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask); 982 mutex_unlock(&memcg_tasklist); 983 record_last_oom(mem_over_limit); 984 } 985 goto nomem; 986 } 987 } 988 return 0; 989nomem: 990 css_put(&mem->css); 991 return -ENOMEM; 992} 993 994static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page) 995{ 996 struct mem_cgroup *mem; 997 swp_entry_t ent; 998 999 if (!PageSwapCache(page)) 1000 return NULL; 1001 1002 ent.val = page_private(page); 1003 mem = lookup_swap_cgroup(ent); 1004 if (!mem) 1005 return NULL; 1006 if (!css_tryget(&mem->css)) 1007 return NULL; 1008 return mem; 1009} 1010 1011/* 1012 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be 1013 * USED state. If already USED, uncharge and return. 1014 */ 1015 1016static void __mem_cgroup_commit_charge(struct mem_cgroup *mem, 1017 struct page_cgroup *pc, 1018 enum charge_type ctype) 1019{ 1020 /* try_charge() can return NULL to *memcg, taking care of it. */ 1021 if (!mem) 1022 return; 1023 1024 lock_page_cgroup(pc); 1025 if (unlikely(PageCgroupUsed(pc))) { 1026 unlock_page_cgroup(pc); 1027 res_counter_uncharge(&mem->res, PAGE_SIZE); 1028 if (do_swap_account) 1029 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1030 css_put(&mem->css); 1031 return; 1032 } 1033 pc->mem_cgroup = mem; 1034 smp_wmb(); 1035 pc->flags = pcg_default_flags[ctype]; 1036 1037 mem_cgroup_charge_statistics(mem, pc, true); 1038 1039 unlock_page_cgroup(pc); 1040} 1041 1042/** 1043 * mem_cgroup_move_account - move account of the page 1044 * @pc: page_cgroup of the page. 1045 * @from: mem_cgroup which the page is moved from. 1046 * @to: mem_cgroup which the page is moved to. @from != @to. 1047 * 1048 * The caller must confirm following. 1049 * - page is not on LRU (isolate_page() is useful.) 1050 * 1051 * returns 0 at success, 1052 * returns -EBUSY when lock is busy or "pc" is unstable. 1053 * 1054 * This function does "uncharge" from old cgroup but doesn't do "charge" to 1055 * new cgroup. It should be done by a caller. 1056 */ 1057 1058static int mem_cgroup_move_account(struct page_cgroup *pc, 1059 struct mem_cgroup *from, struct mem_cgroup *to) 1060{ 1061 struct mem_cgroup_per_zone *from_mz, *to_mz; 1062 int nid, zid; 1063 int ret = -EBUSY; 1064 1065 VM_BUG_ON(from == to); 1066 VM_BUG_ON(PageLRU(pc->page)); 1067 1068 nid = page_cgroup_nid(pc); 1069 zid = page_cgroup_zid(pc); 1070 from_mz = mem_cgroup_zoneinfo(from, nid, zid); 1071 to_mz = mem_cgroup_zoneinfo(to, nid, zid); 1072 1073 if (!trylock_page_cgroup(pc)) 1074 return ret; 1075 1076 if (!PageCgroupUsed(pc)) 1077 goto out; 1078 1079 if (pc->mem_cgroup != from) 1080 goto out; 1081 1082 res_counter_uncharge(&from->res, PAGE_SIZE); 1083 mem_cgroup_charge_statistics(from, pc, false); 1084 if (do_swap_account) 1085 res_counter_uncharge(&from->memsw, PAGE_SIZE); 1086 css_put(&from->css); 1087 1088 css_get(&to->css); 1089 pc->mem_cgroup = to; 1090 mem_cgroup_charge_statistics(to, pc, true); 1091 ret = 0; 1092out: 1093 unlock_page_cgroup(pc); 1094 return ret; 1095} 1096 1097/* 1098 * move charges to its parent. 1099 */ 1100 1101static int mem_cgroup_move_parent(struct page_cgroup *pc, 1102 struct mem_cgroup *child, 1103 gfp_t gfp_mask) 1104{ 1105 struct page *page = pc->page; 1106 struct cgroup *cg = child->css.cgroup; 1107 struct cgroup *pcg = cg->parent; 1108 struct mem_cgroup *parent; 1109 int ret; 1110 1111 /* Is ROOT ? */ 1112 if (!pcg) 1113 return -EINVAL; 1114 1115 1116 parent = mem_cgroup_from_cont(pcg); 1117 1118 1119 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false); 1120 if (ret || !parent) 1121 return ret; 1122 1123 if (!get_page_unless_zero(page)) { 1124 ret = -EBUSY; 1125 goto uncharge; 1126 } 1127 1128 ret = isolate_lru_page(page); 1129 1130 if (ret) 1131 goto cancel; 1132 1133 ret = mem_cgroup_move_account(pc, child, parent); 1134 1135 putback_lru_page(page); 1136 if (!ret) { 1137 put_page(page); 1138 /* drop extra refcnt by try_charge() */ 1139 css_put(&parent->css); 1140 return 0; 1141 } 1142 1143cancel: 1144 put_page(page); 1145uncharge: 1146 /* drop extra refcnt by try_charge() */ 1147 css_put(&parent->css); 1148 /* uncharge if move fails */ 1149 res_counter_uncharge(&parent->res, PAGE_SIZE); 1150 if (do_swap_account) 1151 res_counter_uncharge(&parent->memsw, PAGE_SIZE); 1152 return ret; 1153} 1154 1155/* 1156 * Charge the memory controller for page usage. 1157 * Return 1158 * 0 if the charge was successful 1159 * < 0 if the cgroup is over its limit 1160 */ 1161static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm, 1162 gfp_t gfp_mask, enum charge_type ctype, 1163 struct mem_cgroup *memcg) 1164{ 1165 struct mem_cgroup *mem; 1166 struct page_cgroup *pc; 1167 int ret; 1168 1169 pc = lookup_page_cgroup(page); 1170 /* can happen at boot */ 1171 if (unlikely(!pc)) 1172 return 0; 1173 prefetchw(pc); 1174 1175 mem = memcg; 1176 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true); 1177 if (ret || !mem) 1178 return ret; 1179 1180 __mem_cgroup_commit_charge(mem, pc, ctype); 1181 return 0; 1182} 1183 1184int mem_cgroup_newpage_charge(struct page *page, 1185 struct mm_struct *mm, gfp_t gfp_mask) 1186{ 1187 if (mem_cgroup_disabled()) 1188 return 0; 1189 if (PageCompound(page)) 1190 return 0; 1191 /* 1192 * If already mapped, we don't have to account. 1193 * If page cache, page->mapping has address_space. 1194 * But page->mapping may have out-of-use anon_vma pointer, 1195 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping 1196 * is NULL. 1197 */ 1198 if (page_mapped(page) || (page->mapping && !PageAnon(page))) 1199 return 0; 1200 if (unlikely(!mm)) 1201 mm = &init_mm; 1202 return mem_cgroup_charge_common(page, mm, gfp_mask, 1203 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL); 1204} 1205 1206int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm, 1207 gfp_t gfp_mask) 1208{ 1209 struct mem_cgroup *mem = NULL; 1210 int ret; 1211 1212 if (mem_cgroup_disabled()) 1213 return 0; 1214 if (PageCompound(page)) 1215 return 0; 1216 /* 1217 * Corner case handling. This is called from add_to_page_cache() 1218 * in usual. But some FS (shmem) precharges this page before calling it 1219 * and call add_to_page_cache() with GFP_NOWAIT. 1220 * 1221 * For GFP_NOWAIT case, the page may be pre-charged before calling 1222 * add_to_page_cache(). (See shmem.c) check it here and avoid to call 1223 * charge twice. (It works but has to pay a bit larger cost.) 1224 * And when the page is SwapCache, it should take swap information 1225 * into account. This is under lock_page() now. 1226 */ 1227 if (!(gfp_mask & __GFP_WAIT)) { 1228 struct page_cgroup *pc; 1229 1230 1231 pc = lookup_page_cgroup(page); 1232 if (!pc) 1233 return 0; 1234 lock_page_cgroup(pc); 1235 if (PageCgroupUsed(pc)) { 1236 unlock_page_cgroup(pc); 1237 return 0; 1238 } 1239 unlock_page_cgroup(pc); 1240 } 1241 1242 if (do_swap_account && PageSwapCache(page)) { 1243 mem = try_get_mem_cgroup_from_swapcache(page); 1244 if (mem) 1245 mm = NULL; 1246 else 1247 mem = NULL; 1248 /* SwapCache may be still linked to LRU now. */ 1249 mem_cgroup_lru_del_before_commit_swapcache(page); 1250 } 1251 1252 if (unlikely(!mm && !mem)) 1253 mm = &init_mm; 1254 1255 if (page_is_file_cache(page)) 1256 return mem_cgroup_charge_common(page, mm, gfp_mask, 1257 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL); 1258 1259 ret = mem_cgroup_charge_common(page, mm, gfp_mask, 1260 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem); 1261 if (mem) 1262 css_put(&mem->css); 1263 if (PageSwapCache(page)) 1264 mem_cgroup_lru_add_after_commit_swapcache(page); 1265 1266 if (do_swap_account && !ret && PageSwapCache(page)) { 1267 swp_entry_t ent = {.val = page_private(page)}; 1268 /* avoid double counting */ 1269 mem = swap_cgroup_record(ent, NULL); 1270 if (mem) { 1271 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1272 mem_cgroup_put(mem); 1273 } 1274 } 1275 return ret; 1276} 1277 1278/* 1279 * While swap-in, try_charge -> commit or cancel, the page is locked. 1280 * And when try_charge() successfully returns, one refcnt to memcg without 1281 * struct page_cgroup is aquired. This refcnt will be cumsumed by 1282 * "commit()" or removed by "cancel()" 1283 */ 1284int mem_cgroup_try_charge_swapin(struct mm_struct *mm, 1285 struct page *page, 1286 gfp_t mask, struct mem_cgroup **ptr) 1287{ 1288 struct mem_cgroup *mem; 1289 int ret; 1290 1291 if (mem_cgroup_disabled()) 1292 return 0; 1293 1294 if (!do_swap_account) 1295 goto charge_cur_mm; 1296 /* 1297 * A racing thread's fault, or swapoff, may have already updated 1298 * the pte, and even removed page from swap cache: return success 1299 * to go on to do_swap_page()'s pte_same() test, which should fail. 1300 */ 1301 if (!PageSwapCache(page)) 1302 return 0; 1303 mem = try_get_mem_cgroup_from_swapcache(page); 1304 if (!mem) 1305 goto charge_cur_mm; 1306 *ptr = mem; 1307 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true); 1308 /* drop extra refcnt from tryget */ 1309 css_put(&mem->css); 1310 return ret; 1311charge_cur_mm: 1312 if (unlikely(!mm)) 1313 mm = &init_mm; 1314 return __mem_cgroup_try_charge(mm, mask, ptr, true); 1315} 1316 1317void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr) 1318{ 1319 struct page_cgroup *pc; 1320 1321 if (mem_cgroup_disabled()) 1322 return; 1323 if (!ptr) 1324 return; 1325 pc = lookup_page_cgroup(page); 1326 mem_cgroup_lru_del_before_commit_swapcache(page); 1327 __mem_cgroup_commit_charge(ptr, pc, MEM_CGROUP_CHARGE_TYPE_MAPPED); 1328 mem_cgroup_lru_add_after_commit_swapcache(page); 1329 /* 1330 * Now swap is on-memory. This means this page may be 1331 * counted both as mem and swap....double count. 1332 * Fix it by uncharging from memsw. Basically, this SwapCache is stable 1333 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page() 1334 * may call delete_from_swap_cache() before reach here. 1335 */ 1336 if (do_swap_account && PageSwapCache(page)) { 1337 swp_entry_t ent = {.val = page_private(page)}; 1338 struct mem_cgroup *memcg; 1339 memcg = swap_cgroup_record(ent, NULL); 1340 if (memcg) { 1341 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 1342 mem_cgroup_put(memcg); 1343 } 1344 1345 } 1346 /* add this page(page_cgroup) to the LRU we want. */ 1347 1348} 1349 1350void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem) 1351{ 1352 if (mem_cgroup_disabled()) 1353 return; 1354 if (!mem) 1355 return; 1356 res_counter_uncharge(&mem->res, PAGE_SIZE); 1357 if (do_swap_account) 1358 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1359 css_put(&mem->css); 1360} 1361 1362 1363/* 1364 * uncharge if !page_mapped(page) 1365 */ 1366static struct mem_cgroup * 1367__mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype) 1368{ 1369 struct page_cgroup *pc; 1370 struct mem_cgroup *mem = NULL; 1371 struct mem_cgroup_per_zone *mz; 1372 1373 if (mem_cgroup_disabled()) 1374 return NULL; 1375 1376 if (PageSwapCache(page)) 1377 return NULL; 1378 1379 /* 1380 * Check if our page_cgroup is valid 1381 */ 1382 pc = lookup_page_cgroup(page); 1383 if (unlikely(!pc || !PageCgroupUsed(pc))) 1384 return NULL; 1385 1386 lock_page_cgroup(pc); 1387 1388 mem = pc->mem_cgroup; 1389 1390 if (!PageCgroupUsed(pc)) 1391 goto unlock_out; 1392 1393 switch (ctype) { 1394 case MEM_CGROUP_CHARGE_TYPE_MAPPED: 1395 if (page_mapped(page)) 1396 goto unlock_out; 1397 break; 1398 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT: 1399 if (!PageAnon(page)) { /* Shared memory */ 1400 if (page->mapping && !page_is_file_cache(page)) 1401 goto unlock_out; 1402 } else if (page_mapped(page)) /* Anon */ 1403 goto unlock_out; 1404 break; 1405 default: 1406 break; 1407 } 1408 1409 res_counter_uncharge(&mem->res, PAGE_SIZE); 1410 if (do_swap_account && (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)) 1411 res_counter_uncharge(&mem->memsw, PAGE_SIZE); 1412 mem_cgroup_charge_statistics(mem, pc, false); 1413 1414 ClearPageCgroupUsed(pc); 1415 /* 1416 * pc->mem_cgroup is not cleared here. It will be accessed when it's 1417 * freed from LRU. This is safe because uncharged page is expected not 1418 * to be reused (freed soon). Exception is SwapCache, it's handled by 1419 * special functions. 1420 */ 1421 1422 mz = page_cgroup_zoneinfo(pc); 1423 unlock_page_cgroup(pc); 1424 1425 /* at swapout, this memcg will be accessed to record to swap */ 1426 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT) 1427 css_put(&mem->css); 1428 1429 return mem; 1430 1431unlock_out: 1432 unlock_page_cgroup(pc); 1433 return NULL; 1434} 1435 1436void mem_cgroup_uncharge_page(struct page *page) 1437{ 1438 /* early check. */ 1439 if (page_mapped(page)) 1440 return; 1441 if (page->mapping && !PageAnon(page)) 1442 return; 1443 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED); 1444} 1445 1446void mem_cgroup_uncharge_cache_page(struct page *page) 1447{ 1448 VM_BUG_ON(page_mapped(page)); 1449 VM_BUG_ON(page->mapping); 1450 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE); 1451} 1452 1453/* 1454 * called from __delete_from_swap_cache() and drop "page" account. 1455 * memcg information is recorded to swap_cgroup of "ent" 1456 */ 1457void mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent) 1458{ 1459 struct mem_cgroup *memcg; 1460 1461 memcg = __mem_cgroup_uncharge_common(page, 1462 MEM_CGROUP_CHARGE_TYPE_SWAPOUT); 1463 /* record memcg information */ 1464 if (do_swap_account && memcg) { 1465 swap_cgroup_record(ent, memcg); 1466 mem_cgroup_get(memcg); 1467 } 1468 if (memcg) 1469 css_put(&memcg->css); 1470} 1471 1472#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 1473/* 1474 * called from swap_entry_free(). remove record in swap_cgroup and 1475 * uncharge "memsw" account. 1476 */ 1477void mem_cgroup_uncharge_swap(swp_entry_t ent) 1478{ 1479 struct mem_cgroup *memcg; 1480 1481 if (!do_swap_account) 1482 return; 1483 1484 memcg = swap_cgroup_record(ent, NULL); 1485 if (memcg) { 1486 res_counter_uncharge(&memcg->memsw, PAGE_SIZE); 1487 mem_cgroup_put(memcg); 1488 } 1489} 1490#endif 1491 1492/* 1493 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old 1494 * page belongs to. 1495 */ 1496int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr) 1497{ 1498 struct page_cgroup *pc; 1499 struct mem_cgroup *mem = NULL; 1500 int ret = 0; 1501 1502 if (mem_cgroup_disabled()) 1503 return 0; 1504 1505 pc = lookup_page_cgroup(page); 1506 lock_page_cgroup(pc); 1507 if (PageCgroupUsed(pc)) { 1508 mem = pc->mem_cgroup; 1509 css_get(&mem->css); 1510 } 1511 unlock_page_cgroup(pc); 1512 1513 if (mem) { 1514 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false); 1515 css_put(&mem->css); 1516 } 1517 *ptr = mem; 1518 return ret; 1519} 1520 1521/* remove redundant charge if migration failed*/ 1522void mem_cgroup_end_migration(struct mem_cgroup *mem, 1523 struct page *oldpage, struct page *newpage) 1524{ 1525 struct page *target, *unused; 1526 struct page_cgroup *pc; 1527 enum charge_type ctype; 1528 1529 if (!mem) 1530 return; 1531 1532 /* at migration success, oldpage->mapping is NULL. */ 1533 if (oldpage->mapping) { 1534 target = oldpage; 1535 unused = NULL; 1536 } else { 1537 target = newpage; 1538 unused = oldpage; 1539 } 1540 1541 if (PageAnon(target)) 1542 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED; 1543 else if (page_is_file_cache(target)) 1544 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE; 1545 else 1546 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM; 1547 1548 /* unused page is not on radix-tree now. */ 1549 if (unused) 1550 __mem_cgroup_uncharge_common(unused, ctype); 1551 1552 pc = lookup_page_cgroup(target); 1553 /* 1554 * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup. 1555 * So, double-counting is effectively avoided. 1556 */ 1557 __mem_cgroup_commit_charge(mem, pc, ctype); 1558 1559 /* 1560 * Both of oldpage and newpage are still under lock_page(). 1561 * Then, we don't have to care about race in radix-tree. 1562 * But we have to be careful that this page is unmapped or not. 1563 * 1564 * There is a case for !page_mapped(). At the start of 1565 * migration, oldpage was mapped. But now, it's zapped. 1566 * But we know *target* page is not freed/reused under us. 1567 * mem_cgroup_uncharge_page() does all necessary checks. 1568 */ 1569 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED) 1570 mem_cgroup_uncharge_page(target); 1571} 1572 1573/* 1574 * A call to try to shrink memory usage under specified resource controller. 1575 * This is typically used for page reclaiming for shmem for reducing side 1576 * effect of page allocation from shmem, which is used by some mem_cgroup. 1577 */ 1578int mem_cgroup_shrink_usage(struct page *page, 1579 struct mm_struct *mm, 1580 gfp_t gfp_mask) 1581{ 1582 struct mem_cgroup *mem = NULL; 1583 int progress = 0; 1584 int retry = MEM_CGROUP_RECLAIM_RETRIES; 1585 1586 if (mem_cgroup_disabled()) 1587 return 0; 1588 if (page) 1589 mem = try_get_mem_cgroup_from_swapcache(page); 1590 if (!mem && mm) 1591 mem = try_get_mem_cgroup_from_mm(mm); 1592 if (unlikely(!mem)) 1593 return 0; 1594 1595 do { 1596 progress = mem_cgroup_hierarchical_reclaim(mem, 1597 gfp_mask, true, false); 1598 progress += mem_cgroup_check_under_limit(mem); 1599 } while (!progress && --retry); 1600 1601 css_put(&mem->css); 1602 if (!retry) 1603 return -ENOMEM; 1604 return 0; 1605} 1606 1607static DEFINE_MUTEX(set_limit_mutex); 1608 1609static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, 1610 unsigned long long val) 1611{ 1612 int retry_count; 1613 int progress; 1614 u64 memswlimit; 1615 int ret = 0; 1616 int children = mem_cgroup_count_children(memcg); 1617 u64 curusage, oldusage; 1618 1619 /* 1620 * For keeping hierarchical_reclaim simple, how long we should retry 1621 * is depends on callers. We set our retry-count to be function 1622 * of # of children which we should visit in this loop. 1623 */ 1624 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children; 1625 1626 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE); 1627 1628 while (retry_count) { 1629 if (signal_pending(current)) { 1630 ret = -EINTR; 1631 break; 1632 } 1633 /* 1634 * Rather than hide all in some function, I do this in 1635 * open coded manner. You see what this really does. 1636 * We have to guarantee mem->res.limit < mem->memsw.limit. 1637 */ 1638 mutex_lock(&set_limit_mutex); 1639 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1640 if (memswlimit < val) { 1641 ret = -EINVAL; 1642 mutex_unlock(&set_limit_mutex); 1643 break; 1644 } 1645 ret = res_counter_set_limit(&memcg->res, val); 1646 mutex_unlock(&set_limit_mutex); 1647 1648 if (!ret) 1649 break; 1650 1651 progress = mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, 1652 false, true); 1653 curusage = res_counter_read_u64(&memcg->res, RES_USAGE); 1654 /* Usage is reduced ? */ 1655 if (curusage >= oldusage) 1656 retry_count--; 1657 else 1658 oldusage = curusage; 1659 } 1660 1661 return ret; 1662} 1663 1664int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg, 1665 unsigned long long val) 1666{ 1667 int retry_count; 1668 u64 memlimit, oldusage, curusage; 1669 int children = mem_cgroup_count_children(memcg); 1670 int ret = -EBUSY; 1671 1672 if (!do_swap_account) 1673 return -EINVAL; 1674 /* see mem_cgroup_resize_res_limit */ 1675 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES; 1676 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 1677 while (retry_count) { 1678 if (signal_pending(current)) { 1679 ret = -EINTR; 1680 break; 1681 } 1682 /* 1683 * Rather than hide all in some function, I do this in 1684 * open coded manner. You see what this really does. 1685 * We have to guarantee mem->res.limit < mem->memsw.limit. 1686 */ 1687 mutex_lock(&set_limit_mutex); 1688 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1689 if (memlimit > val) { 1690 ret = -EINVAL; 1691 mutex_unlock(&set_limit_mutex); 1692 break; 1693 } 1694 ret = res_counter_set_limit(&memcg->memsw, val); 1695 mutex_unlock(&set_limit_mutex); 1696 1697 if (!ret) 1698 break; 1699 1700 mem_cgroup_hierarchical_reclaim(memcg, GFP_KERNEL, true, true); 1701 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE); 1702 /* Usage is reduced ? */ 1703 if (curusage >= oldusage) 1704 retry_count--; 1705 else 1706 oldusage = curusage; 1707 } 1708 return ret; 1709} 1710 1711/* 1712 * This routine traverse page_cgroup in given list and drop them all. 1713 * *And* this routine doesn't reclaim page itself, just removes page_cgroup. 1714 */ 1715static int mem_cgroup_force_empty_list(struct mem_cgroup *mem, 1716 int node, int zid, enum lru_list lru) 1717{ 1718 struct zone *zone; 1719 struct mem_cgroup_per_zone *mz; 1720 struct page_cgroup *pc, *busy; 1721 unsigned long flags, loop; 1722 struct list_head *list; 1723 int ret = 0; 1724 1725 zone = &NODE_DATA(node)->node_zones[zid]; 1726 mz = mem_cgroup_zoneinfo(mem, node, zid); 1727 list = &mz->lists[lru]; 1728 1729 loop = MEM_CGROUP_ZSTAT(mz, lru); 1730 /* give some margin against EBUSY etc...*/ 1731 loop += 256; 1732 busy = NULL; 1733 while (loop--) { 1734 ret = 0; 1735 spin_lock_irqsave(&zone->lru_lock, flags); 1736 if (list_empty(list)) { 1737 spin_unlock_irqrestore(&zone->lru_lock, flags); 1738 break; 1739 } 1740 pc = list_entry(list->prev, struct page_cgroup, lru); 1741 if (busy == pc) { 1742 list_move(&pc->lru, list); 1743 busy = 0; 1744 spin_unlock_irqrestore(&zone->lru_lock, flags); 1745 continue; 1746 } 1747 spin_unlock_irqrestore(&zone->lru_lock, flags); 1748 1749 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL); 1750 if (ret == -ENOMEM) 1751 break; 1752 1753 if (ret == -EBUSY || ret == -EINVAL) { 1754 /* found lock contention or "pc" is obsolete. */ 1755 busy = pc; 1756 cond_resched(); 1757 } else 1758 busy = NULL; 1759 } 1760 1761 if (!ret && !list_empty(list)) 1762 return -EBUSY; 1763 return ret; 1764} 1765 1766/* 1767 * make mem_cgroup's charge to be 0 if there is no task. 1768 * This enables deleting this mem_cgroup. 1769 */ 1770static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all) 1771{ 1772 int ret; 1773 int node, zid, shrink; 1774 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES; 1775 struct cgroup *cgrp = mem->css.cgroup; 1776 1777 css_get(&mem->css); 1778 1779 shrink = 0; 1780 /* should free all ? */ 1781 if (free_all) 1782 goto try_to_free; 1783move_account: 1784 while (mem->res.usage > 0) { 1785 ret = -EBUSY; 1786 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children)) 1787 goto out; 1788 ret = -EINTR; 1789 if (signal_pending(current)) 1790 goto out; 1791 /* This is for making all *used* pages to be on LRU. */ 1792 lru_add_drain_all(); 1793 ret = 0; 1794 for_each_node_state(node, N_HIGH_MEMORY) { 1795 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) { 1796 enum lru_list l; 1797 for_each_lru(l) { 1798 ret = mem_cgroup_force_empty_list(mem, 1799 node, zid, l); 1800 if (ret) 1801 break; 1802 } 1803 } 1804 if (ret) 1805 break; 1806 } 1807 /* it seems parent cgroup doesn't have enough mem */ 1808 if (ret == -ENOMEM) 1809 goto try_to_free; 1810 cond_resched(); 1811 } 1812 ret = 0; 1813out: 1814 css_put(&mem->css); 1815 return ret; 1816 1817try_to_free: 1818 /* returns EBUSY if there is a task or if we come here twice. */ 1819 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) { 1820 ret = -EBUSY; 1821 goto out; 1822 } 1823 /* we call try-to-free pages for make this cgroup empty */ 1824 lru_add_drain_all(); 1825 /* try to free all pages in this cgroup */ 1826 shrink = 1; 1827 while (nr_retries && mem->res.usage > 0) { 1828 int progress; 1829 1830 if (signal_pending(current)) { 1831 ret = -EINTR; 1832 goto out; 1833 } 1834 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL, 1835 false, get_swappiness(mem)); 1836 if (!progress) { 1837 nr_retries--; 1838 /* maybe some writeback is necessary */ 1839 congestion_wait(WRITE, HZ/10); 1840 } 1841 1842 } 1843 lru_add_drain(); 1844 /* try move_account...there may be some *locked* pages. */ 1845 if (mem->res.usage) 1846 goto move_account; 1847 ret = 0; 1848 goto out; 1849} 1850 1851int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event) 1852{ 1853 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true); 1854} 1855 1856 1857static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft) 1858{ 1859 return mem_cgroup_from_cont(cont)->use_hierarchy; 1860} 1861 1862static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft, 1863 u64 val) 1864{ 1865 int retval = 0; 1866 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 1867 struct cgroup *parent = cont->parent; 1868 struct mem_cgroup *parent_mem = NULL; 1869 1870 if (parent) 1871 parent_mem = mem_cgroup_from_cont(parent); 1872 1873 cgroup_lock(); 1874 /* 1875 * If parent's use_hiearchy is set, we can't make any modifications 1876 * in the child subtrees. If it is unset, then the change can 1877 * occur, provided the current cgroup has no children. 1878 * 1879 * For the root cgroup, parent_mem is NULL, we allow value to be 1880 * set if there are no children. 1881 */ 1882 if ((!parent_mem || !parent_mem->use_hierarchy) && 1883 (val == 1 || val == 0)) { 1884 if (list_empty(&cont->children)) 1885 mem->use_hierarchy = val; 1886 else 1887 retval = -EBUSY; 1888 } else 1889 retval = -EINVAL; 1890 cgroup_unlock(); 1891 1892 return retval; 1893} 1894 1895static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft) 1896{ 1897 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 1898 u64 val = 0; 1899 int type, name; 1900 1901 type = MEMFILE_TYPE(cft->private); 1902 name = MEMFILE_ATTR(cft->private); 1903 switch (type) { 1904 case _MEM: 1905 val = res_counter_read_u64(&mem->res, name); 1906 break; 1907 case _MEMSWAP: 1908 if (do_swap_account) 1909 val = res_counter_read_u64(&mem->memsw, name); 1910 break; 1911 default: 1912 BUG(); 1913 break; 1914 } 1915 return val; 1916} 1917/* 1918 * The user of this function is... 1919 * RES_LIMIT. 1920 */ 1921static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft, 1922 const char *buffer) 1923{ 1924 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont); 1925 int type, name; 1926 unsigned long long val; 1927 int ret; 1928 1929 type = MEMFILE_TYPE(cft->private); 1930 name = MEMFILE_ATTR(cft->private); 1931 switch (name) { 1932 case RES_LIMIT: 1933 /* This function does all necessary parse...reuse it */ 1934 ret = res_counter_memparse_write_strategy(buffer, &val); 1935 if (ret) 1936 break; 1937 if (type == _MEM) 1938 ret = mem_cgroup_resize_limit(memcg, val); 1939 else 1940 ret = mem_cgroup_resize_memsw_limit(memcg, val); 1941 break; 1942 default: 1943 ret = -EINVAL; /* should be BUG() ? */ 1944 break; 1945 } 1946 return ret; 1947} 1948 1949static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg, 1950 unsigned long long *mem_limit, unsigned long long *memsw_limit) 1951{ 1952 struct cgroup *cgroup; 1953 unsigned long long min_limit, min_memsw_limit, tmp; 1954 1955 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT); 1956 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1957 cgroup = memcg->css.cgroup; 1958 if (!memcg->use_hierarchy) 1959 goto out; 1960 1961 while (cgroup->parent) { 1962 cgroup = cgroup->parent; 1963 memcg = mem_cgroup_from_cont(cgroup); 1964 if (!memcg->use_hierarchy) 1965 break; 1966 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT); 1967 min_limit = min(min_limit, tmp); 1968 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT); 1969 min_memsw_limit = min(min_memsw_limit, tmp); 1970 } 1971out: 1972 *mem_limit = min_limit; 1973 *memsw_limit = min_memsw_limit; 1974 return; 1975} 1976 1977static int mem_cgroup_reset(struct cgroup *cont, unsigned int event) 1978{ 1979 struct mem_cgroup *mem; 1980 int type, name; 1981 1982 mem = mem_cgroup_from_cont(cont); 1983 type = MEMFILE_TYPE(event); 1984 name = MEMFILE_ATTR(event); 1985 switch (name) { 1986 case RES_MAX_USAGE: 1987 if (type == _MEM) 1988 res_counter_reset_max(&mem->res); 1989 else 1990 res_counter_reset_max(&mem->memsw); 1991 break; 1992 case RES_FAILCNT: 1993 if (type == _MEM) 1994 res_counter_reset_failcnt(&mem->res); 1995 else 1996 res_counter_reset_failcnt(&mem->memsw); 1997 break; 1998 } 1999 return 0; 2000} 2001 2002 2003/* For read statistics */ 2004enum { 2005 MCS_CACHE, 2006 MCS_RSS, 2007 MCS_PGPGIN, 2008 MCS_PGPGOUT, 2009 MCS_INACTIVE_ANON, 2010 MCS_ACTIVE_ANON, 2011 MCS_INACTIVE_FILE, 2012 MCS_ACTIVE_FILE, 2013 MCS_UNEVICTABLE, 2014 NR_MCS_STAT, 2015}; 2016 2017struct mcs_total_stat { 2018 s64 stat[NR_MCS_STAT]; 2019}; 2020 2021struct { 2022 char *local_name; 2023 char *total_name; 2024} memcg_stat_strings[NR_MCS_STAT] = { 2025 {"cache", "total_cache"}, 2026 {"rss", "total_rss"}, 2027 {"pgpgin", "total_pgpgin"}, 2028 {"pgpgout", "total_pgpgout"}, 2029 {"inactive_anon", "total_inactive_anon"}, 2030 {"active_anon", "total_active_anon"}, 2031 {"inactive_file", "total_inactive_file"}, 2032 {"active_file", "total_active_file"}, 2033 {"unevictable", "total_unevictable"} 2034}; 2035 2036 2037static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data) 2038{ 2039 struct mcs_total_stat *s = data; 2040 s64 val; 2041 2042 /* per cpu stat */ 2043 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE); 2044 s->stat[MCS_CACHE] += val * PAGE_SIZE; 2045 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS); 2046 s->stat[MCS_RSS] += val * PAGE_SIZE; 2047 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT); 2048 s->stat[MCS_PGPGIN] += val; 2049 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT); 2050 s->stat[MCS_PGPGOUT] += val; 2051 2052 /* per zone stat */ 2053 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON); 2054 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE; 2055 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON); 2056 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE; 2057 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE); 2058 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE; 2059 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE); 2060 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE; 2061 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE); 2062 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE; 2063 return 0; 2064} 2065 2066static void 2067mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s) 2068{ 2069 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat); 2070} 2071 2072static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft, 2073 struct cgroup_map_cb *cb) 2074{ 2075 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont); 2076 struct mcs_total_stat mystat; 2077 int i; 2078 2079 memset(&mystat, 0, sizeof(mystat)); 2080 mem_cgroup_get_local_stat(mem_cont, &mystat); 2081 2082 for (i = 0; i < NR_MCS_STAT; i++) 2083 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]); 2084 2085 /* Hierarchical information */ 2086 { 2087 unsigned long long limit, memsw_limit; 2088 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit); 2089 cb->fill(cb, "hierarchical_memory_limit", limit); 2090 if (do_swap_account) 2091 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit); 2092 } 2093 2094 memset(&mystat, 0, sizeof(mystat)); 2095 mem_cgroup_get_total_stat(mem_cont, &mystat); 2096 for (i = 0; i < NR_MCS_STAT; i++) 2097 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]); 2098 2099 2100#ifdef CONFIG_DEBUG_VM 2101 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL)); 2102 2103 { 2104 int nid, zid; 2105 struct mem_cgroup_per_zone *mz; 2106 unsigned long recent_rotated[2] = {0, 0}; 2107 unsigned long recent_scanned[2] = {0, 0}; 2108 2109 for_each_online_node(nid) 2110 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 2111 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid); 2112 2113 recent_rotated[0] += 2114 mz->reclaim_stat.recent_rotated[0]; 2115 recent_rotated[1] += 2116 mz->reclaim_stat.recent_rotated[1]; 2117 recent_scanned[0] += 2118 mz->reclaim_stat.recent_scanned[0]; 2119 recent_scanned[1] += 2120 mz->reclaim_stat.recent_scanned[1]; 2121 } 2122 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]); 2123 cb->fill(cb, "recent_rotated_file", recent_rotated[1]); 2124 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]); 2125 cb->fill(cb, "recent_scanned_file", recent_scanned[1]); 2126 } 2127#endif 2128 2129 return 0; 2130} 2131 2132static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft) 2133{ 2134 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 2135 2136 return get_swappiness(memcg); 2137} 2138 2139static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft, 2140 u64 val) 2141{ 2142 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp); 2143 struct mem_cgroup *parent; 2144 2145 if (val > 100) 2146 return -EINVAL; 2147 2148 if (cgrp->parent == NULL) 2149 return -EINVAL; 2150 2151 parent = mem_cgroup_from_cont(cgrp->parent); 2152 2153 cgroup_lock(); 2154 2155 /* If under hierarchy, only empty-root can set this value */ 2156 if ((parent->use_hierarchy) || 2157 (memcg->use_hierarchy && !list_empty(&cgrp->children))) { 2158 cgroup_unlock(); 2159 return -EINVAL; 2160 } 2161 2162 spin_lock(&memcg->reclaim_param_lock); 2163 memcg->swappiness = val; 2164 spin_unlock(&memcg->reclaim_param_lock); 2165 2166 cgroup_unlock(); 2167 2168 return 0; 2169} 2170 2171 2172static struct cftype mem_cgroup_files[] = { 2173 { 2174 .name = "usage_in_bytes", 2175 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE), 2176 .read_u64 = mem_cgroup_read, 2177 }, 2178 { 2179 .name = "max_usage_in_bytes", 2180 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE), 2181 .trigger = mem_cgroup_reset, 2182 .read_u64 = mem_cgroup_read, 2183 }, 2184 { 2185 .name = "limit_in_bytes", 2186 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT), 2187 .write_string = mem_cgroup_write, 2188 .read_u64 = mem_cgroup_read, 2189 }, 2190 { 2191 .name = "failcnt", 2192 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT), 2193 .trigger = mem_cgroup_reset, 2194 .read_u64 = mem_cgroup_read, 2195 }, 2196 { 2197 .name = "stat", 2198 .read_map = mem_control_stat_show, 2199 }, 2200 { 2201 .name = "force_empty", 2202 .trigger = mem_cgroup_force_empty_write, 2203 }, 2204 { 2205 .name = "use_hierarchy", 2206 .write_u64 = mem_cgroup_hierarchy_write, 2207 .read_u64 = mem_cgroup_hierarchy_read, 2208 }, 2209 { 2210 .name = "swappiness", 2211 .read_u64 = mem_cgroup_swappiness_read, 2212 .write_u64 = mem_cgroup_swappiness_write, 2213 }, 2214}; 2215 2216#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2217static struct cftype memsw_cgroup_files[] = { 2218 { 2219 .name = "memsw.usage_in_bytes", 2220 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE), 2221 .read_u64 = mem_cgroup_read, 2222 }, 2223 { 2224 .name = "memsw.max_usage_in_bytes", 2225 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE), 2226 .trigger = mem_cgroup_reset, 2227 .read_u64 = mem_cgroup_read, 2228 }, 2229 { 2230 .name = "memsw.limit_in_bytes", 2231 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT), 2232 .write_string = mem_cgroup_write, 2233 .read_u64 = mem_cgroup_read, 2234 }, 2235 { 2236 .name = "memsw.failcnt", 2237 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT), 2238 .trigger = mem_cgroup_reset, 2239 .read_u64 = mem_cgroup_read, 2240 }, 2241}; 2242 2243static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 2244{ 2245 if (!do_swap_account) 2246 return 0; 2247 return cgroup_add_files(cont, ss, memsw_cgroup_files, 2248 ARRAY_SIZE(memsw_cgroup_files)); 2249}; 2250#else 2251static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss) 2252{ 2253 return 0; 2254} 2255#endif 2256 2257static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 2258{ 2259 struct mem_cgroup_per_node *pn; 2260 struct mem_cgroup_per_zone *mz; 2261 enum lru_list l; 2262 int zone, tmp = node; 2263 /* 2264 * This routine is called against possible nodes. 2265 * But it's BUG to call kmalloc() against offline node. 2266 * 2267 * TODO: this routine can waste much memory for nodes which will 2268 * never be onlined. It's better to use memory hotplug callback 2269 * function. 2270 */ 2271 if (!node_state(node, N_NORMAL_MEMORY)) 2272 tmp = -1; 2273 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp); 2274 if (!pn) 2275 return 1; 2276 2277 mem->info.nodeinfo[node] = pn; 2278 memset(pn, 0, sizeof(*pn)); 2279 2280 for (zone = 0; zone < MAX_NR_ZONES; zone++) { 2281 mz = &pn->zoneinfo[zone]; 2282 for_each_lru(l) 2283 INIT_LIST_HEAD(&mz->lists[l]); 2284 } 2285 return 0; 2286} 2287 2288static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node) 2289{ 2290 kfree(mem->info.nodeinfo[node]); 2291} 2292 2293static int mem_cgroup_size(void) 2294{ 2295 int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu); 2296 return sizeof(struct mem_cgroup) + cpustat_size; 2297} 2298 2299static struct mem_cgroup *mem_cgroup_alloc(void) 2300{ 2301 struct mem_cgroup *mem; 2302 int size = mem_cgroup_size(); 2303 2304 if (size < PAGE_SIZE) 2305 mem = kmalloc(size, GFP_KERNEL); 2306 else 2307 mem = vmalloc(size); 2308 2309 if (mem) 2310 memset(mem, 0, size); 2311 return mem; 2312} 2313 2314/* 2315 * At destroying mem_cgroup, references from swap_cgroup can remain. 2316 * (scanning all at force_empty is too costly...) 2317 * 2318 * Instead of clearing all references at force_empty, we remember 2319 * the number of reference from swap_cgroup and free mem_cgroup when 2320 * it goes down to 0. 2321 * 2322 * Removal of cgroup itself succeeds regardless of refs from swap. 2323 */ 2324 2325static void __mem_cgroup_free(struct mem_cgroup *mem) 2326{ 2327 int node; 2328 2329 free_css_id(&mem_cgroup_subsys, &mem->css); 2330 2331 for_each_node_state(node, N_POSSIBLE) 2332 free_mem_cgroup_per_zone_info(mem, node); 2333 2334 if (mem_cgroup_size() < PAGE_SIZE) 2335 kfree(mem); 2336 else 2337 vfree(mem); 2338} 2339 2340static void mem_cgroup_get(struct mem_cgroup *mem) 2341{ 2342 atomic_inc(&mem->refcnt); 2343} 2344 2345static void mem_cgroup_put(struct mem_cgroup *mem) 2346{ 2347 if (atomic_dec_and_test(&mem->refcnt)) { 2348 struct mem_cgroup *parent = parent_mem_cgroup(mem); 2349 __mem_cgroup_free(mem); 2350 if (parent) 2351 mem_cgroup_put(parent); 2352 } 2353} 2354 2355/* 2356 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled. 2357 */ 2358static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem) 2359{ 2360 if (!mem->res.parent) 2361 return NULL; 2362 return mem_cgroup_from_res_counter(mem->res.parent, res); 2363} 2364 2365#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2366static void __init enable_swap_cgroup(void) 2367{ 2368 if (!mem_cgroup_disabled() && really_do_swap_account) 2369 do_swap_account = 1; 2370} 2371#else 2372static void __init enable_swap_cgroup(void) 2373{ 2374} 2375#endif 2376 2377static struct cgroup_subsys_state * __ref 2378mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont) 2379{ 2380 struct mem_cgroup *mem, *parent; 2381 long error = -ENOMEM; 2382 int node; 2383 2384 mem = mem_cgroup_alloc(); 2385 if (!mem) 2386 return ERR_PTR(error); 2387 2388 for_each_node_state(node, N_POSSIBLE) 2389 if (alloc_mem_cgroup_per_zone_info(mem, node)) 2390 goto free_out; 2391 /* root ? */ 2392 if (cont->parent == NULL) { 2393 enable_swap_cgroup(); 2394 parent = NULL; 2395 } else { 2396 parent = mem_cgroup_from_cont(cont->parent); 2397 mem->use_hierarchy = parent->use_hierarchy; 2398 } 2399 2400 if (parent && parent->use_hierarchy) { 2401 res_counter_init(&mem->res, &parent->res); 2402 res_counter_init(&mem->memsw, &parent->memsw); 2403 /* 2404 * We increment refcnt of the parent to ensure that we can 2405 * safely access it on res_counter_charge/uncharge. 2406 * This refcnt will be decremented when freeing this 2407 * mem_cgroup(see mem_cgroup_put). 2408 */ 2409 mem_cgroup_get(parent); 2410 } else { 2411 res_counter_init(&mem->res, NULL); 2412 res_counter_init(&mem->memsw, NULL); 2413 } 2414 mem->last_scanned_child = 0; 2415 spin_lock_init(&mem->reclaim_param_lock); 2416 2417 if (parent) 2418 mem->swappiness = get_swappiness(parent); 2419 atomic_set(&mem->refcnt, 1); 2420 return &mem->css; 2421free_out: 2422 __mem_cgroup_free(mem); 2423 return ERR_PTR(error); 2424} 2425 2426static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss, 2427 struct cgroup *cont) 2428{ 2429 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2430 2431 return mem_cgroup_force_empty(mem, false); 2432} 2433 2434static void mem_cgroup_destroy(struct cgroup_subsys *ss, 2435 struct cgroup *cont) 2436{ 2437 struct mem_cgroup *mem = mem_cgroup_from_cont(cont); 2438 2439 mem_cgroup_put(mem); 2440} 2441 2442static int mem_cgroup_populate(struct cgroup_subsys *ss, 2443 struct cgroup *cont) 2444{ 2445 int ret; 2446 2447 ret = cgroup_add_files(cont, ss, mem_cgroup_files, 2448 ARRAY_SIZE(mem_cgroup_files)); 2449 2450 if (!ret) 2451 ret = register_memsw_files(cont, ss); 2452 return ret; 2453} 2454 2455static void mem_cgroup_move_task(struct cgroup_subsys *ss, 2456 struct cgroup *cont, 2457 struct cgroup *old_cont, 2458 struct task_struct *p) 2459{ 2460 mutex_lock(&memcg_tasklist); 2461 /* 2462 * FIXME: It's better to move charges of this process from old 2463 * memcg to new memcg. But it's just on TODO-List now. 2464 */ 2465 mutex_unlock(&memcg_tasklist); 2466} 2467 2468struct cgroup_subsys mem_cgroup_subsys = { 2469 .name = "memory", 2470 .subsys_id = mem_cgroup_subsys_id, 2471 .create = mem_cgroup_create, 2472 .pre_destroy = mem_cgroup_pre_destroy, 2473 .destroy = mem_cgroup_destroy, 2474 .populate = mem_cgroup_populate, 2475 .attach = mem_cgroup_move_task, 2476 .early_init = 0, 2477 .use_id = 1, 2478}; 2479 2480#ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP 2481 2482static int __init disable_swap_account(char *s) 2483{ 2484 really_do_swap_account = 0; 2485 return 1; 2486} 2487__setup("noswapaccount", disable_swap_account); 2488#endif 2489