page_alloc.c revision 8d22ba1b74aa9420b6032d856446564fb21f8090
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/compiler.h> 25#include <linux/kernel.h> 26#include <linux/kmemcheck.h> 27#include <linux/module.h> 28#include <linux/suspend.h> 29#include <linux/pagevec.h> 30#include <linux/blkdev.h> 31#include <linux/slab.h> 32#include <linux/oom.h> 33#include <linux/notifier.h> 34#include <linux/topology.h> 35#include <linux/sysctl.h> 36#include <linux/cpu.h> 37#include <linux/cpuset.h> 38#include <linux/memory_hotplug.h> 39#include <linux/nodemask.h> 40#include <linux/vmalloc.h> 41#include <linux/mempolicy.h> 42#include <linux/stop_machine.h> 43#include <linux/sort.h> 44#include <linux/pfn.h> 45#include <linux/backing-dev.h> 46#include <linux/fault-inject.h> 47#include <linux/page-isolation.h> 48#include <linux/page_cgroup.h> 49#include <linux/debugobjects.h> 50#include <linux/kmemleak.h> 51#include <trace/events/kmem.h> 52 53#include <asm/tlbflush.h> 54#include <asm/div64.h> 55#include "internal.h" 56 57/* 58 * Array of node states. 59 */ 60nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 61 [N_POSSIBLE] = NODE_MASK_ALL, 62 [N_ONLINE] = { { [0] = 1UL } }, 63#ifndef CONFIG_NUMA 64 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 65#ifdef CONFIG_HIGHMEM 66 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 67#endif 68 [N_CPU] = { { [0] = 1UL } }, 69#endif /* NUMA */ 70}; 71EXPORT_SYMBOL(node_states); 72 73unsigned long totalram_pages __read_mostly; 74unsigned long totalreserve_pages __read_mostly; 75int percpu_pagelist_fraction; 76gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 77 78#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 79int pageblock_order __read_mostly; 80#endif 81 82static void __free_pages_ok(struct page *page, unsigned int order); 83 84/* 85 * results with 256, 32 in the lowmem_reserve sysctl: 86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 87 * 1G machine -> (16M dma, 784M normal, 224M high) 88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 91 * 92 * TBD: should special case ZONE_DMA32 machines here - in those we normally 93 * don't need any ZONE_NORMAL reservation 94 */ 95int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 96#ifdef CONFIG_ZONE_DMA 97 256, 98#endif 99#ifdef CONFIG_ZONE_DMA32 100 256, 101#endif 102#ifdef CONFIG_HIGHMEM 103 32, 104#endif 105 32, 106}; 107 108EXPORT_SYMBOL(totalram_pages); 109 110static char * const zone_names[MAX_NR_ZONES] = { 111#ifdef CONFIG_ZONE_DMA 112 "DMA", 113#endif 114#ifdef CONFIG_ZONE_DMA32 115 "DMA32", 116#endif 117 "Normal", 118#ifdef CONFIG_HIGHMEM 119 "HighMem", 120#endif 121 "Movable", 122}; 123 124int min_free_kbytes = 1024; 125 126static unsigned long __meminitdata nr_kernel_pages; 127static unsigned long __meminitdata nr_all_pages; 128static unsigned long __meminitdata dma_reserve; 129 130#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 131 /* 132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 133 * ranges of memory (RAM) that may be registered with add_active_range(). 134 * Ranges passed to add_active_range() will be merged if possible 135 * so the number of times add_active_range() can be called is 136 * related to the number of nodes and the number of holes 137 */ 138 #ifdef CONFIG_MAX_ACTIVE_REGIONS 139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 141 #else 142 #if MAX_NUMNODES >= 32 143 /* If there can be many nodes, allow up to 50 holes per node */ 144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 145 #else 146 /* By default, allow up to 256 distinct regions */ 147 #define MAX_ACTIVE_REGIONS 256 148 #endif 149 #endif 150 151 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 152 static int __meminitdata nr_nodemap_entries; 153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 154 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 155 static unsigned long __initdata required_kernelcore; 156 static unsigned long __initdata required_movablecore; 157 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 158 159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 160 int movable_zone; 161 EXPORT_SYMBOL(movable_zone); 162#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 163 164#if MAX_NUMNODES > 1 165int nr_node_ids __read_mostly = MAX_NUMNODES; 166int nr_online_nodes __read_mostly = 1; 167EXPORT_SYMBOL(nr_node_ids); 168EXPORT_SYMBOL(nr_online_nodes); 169#endif 170 171int page_group_by_mobility_disabled __read_mostly; 172 173static void set_pageblock_migratetype(struct page *page, int migratetype) 174{ 175 176 if (unlikely(page_group_by_mobility_disabled)) 177 migratetype = MIGRATE_UNMOVABLE; 178 179 set_pageblock_flags_group(page, (unsigned long)migratetype, 180 PB_migrate, PB_migrate_end); 181} 182 183bool oom_killer_disabled __read_mostly; 184 185#ifdef CONFIG_DEBUG_VM 186static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 187{ 188 int ret = 0; 189 unsigned seq; 190 unsigned long pfn = page_to_pfn(page); 191 192 do { 193 seq = zone_span_seqbegin(zone); 194 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 195 ret = 1; 196 else if (pfn < zone->zone_start_pfn) 197 ret = 1; 198 } while (zone_span_seqretry(zone, seq)); 199 200 return ret; 201} 202 203static int page_is_consistent(struct zone *zone, struct page *page) 204{ 205 if (!pfn_valid_within(page_to_pfn(page))) 206 return 0; 207 if (zone != page_zone(page)) 208 return 0; 209 210 return 1; 211} 212/* 213 * Temporary debugging check for pages not lying within a given zone. 214 */ 215static int bad_range(struct zone *zone, struct page *page) 216{ 217 if (page_outside_zone_boundaries(zone, page)) 218 return 1; 219 if (!page_is_consistent(zone, page)) 220 return 1; 221 222 return 0; 223} 224#else 225static inline int bad_range(struct zone *zone, struct page *page) 226{ 227 return 0; 228} 229#endif 230 231static void bad_page(struct page *page) 232{ 233 static unsigned long resume; 234 static unsigned long nr_shown; 235 static unsigned long nr_unshown; 236 237 /* Don't complain about poisoned pages */ 238 if (PageHWPoison(page)) { 239 __ClearPageBuddy(page); 240 return; 241 } 242 243 /* 244 * Allow a burst of 60 reports, then keep quiet for that minute; 245 * or allow a steady drip of one report per second. 246 */ 247 if (nr_shown == 60) { 248 if (time_before(jiffies, resume)) { 249 nr_unshown++; 250 goto out; 251 } 252 if (nr_unshown) { 253 printk(KERN_ALERT 254 "BUG: Bad page state: %lu messages suppressed\n", 255 nr_unshown); 256 nr_unshown = 0; 257 } 258 nr_shown = 0; 259 } 260 if (nr_shown++ == 0) 261 resume = jiffies + 60 * HZ; 262 263 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 264 current->comm, page_to_pfn(page)); 265 printk(KERN_ALERT 266 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 267 page, (void *)page->flags, page_count(page), 268 page_mapcount(page), page->mapping, page->index); 269 270 dump_stack(); 271out: 272 /* Leave bad fields for debug, except PageBuddy could make trouble */ 273 __ClearPageBuddy(page); 274 add_taint(TAINT_BAD_PAGE); 275} 276 277/* 278 * Higher-order pages are called "compound pages". They are structured thusly: 279 * 280 * The first PAGE_SIZE page is called the "head page". 281 * 282 * The remaining PAGE_SIZE pages are called "tail pages". 283 * 284 * All pages have PG_compound set. All pages have their ->private pointing at 285 * the head page (even the head page has this). 286 * 287 * The first tail page's ->lru.next holds the address of the compound page's 288 * put_page() function. Its ->lru.prev holds the order of allocation. 289 * This usage means that zero-order pages may not be compound. 290 */ 291 292static void free_compound_page(struct page *page) 293{ 294 __free_pages_ok(page, compound_order(page)); 295} 296 297void prep_compound_page(struct page *page, unsigned long order) 298{ 299 int i; 300 int nr_pages = 1 << order; 301 302 set_compound_page_dtor(page, free_compound_page); 303 set_compound_order(page, order); 304 __SetPageHead(page); 305 for (i = 1; i < nr_pages; i++) { 306 struct page *p = page + i; 307 308 __SetPageTail(p); 309 p->first_page = page; 310 } 311} 312 313static int destroy_compound_page(struct page *page, unsigned long order) 314{ 315 int i; 316 int nr_pages = 1 << order; 317 int bad = 0; 318 319 if (unlikely(compound_order(page) != order) || 320 unlikely(!PageHead(page))) { 321 bad_page(page); 322 bad++; 323 } 324 325 __ClearPageHead(page); 326 327 for (i = 1; i < nr_pages; i++) { 328 struct page *p = page + i; 329 330 if (unlikely(!PageTail(p) || (p->first_page != page))) { 331 bad_page(page); 332 bad++; 333 } 334 __ClearPageTail(p); 335 } 336 337 return bad; 338} 339 340static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 341{ 342 int i; 343 344 /* 345 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 346 * and __GFP_HIGHMEM from hard or soft interrupt context. 347 */ 348 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 349 for (i = 0; i < (1 << order); i++) 350 clear_highpage(page + i); 351} 352 353static inline void set_page_order(struct page *page, int order) 354{ 355 set_page_private(page, order); 356 __SetPageBuddy(page); 357} 358 359static inline void rmv_page_order(struct page *page) 360{ 361 __ClearPageBuddy(page); 362 set_page_private(page, 0); 363} 364 365/* 366 * Locate the struct page for both the matching buddy in our 367 * pair (buddy1) and the combined O(n+1) page they form (page). 368 * 369 * 1) Any buddy B1 will have an order O twin B2 which satisfies 370 * the following equation: 371 * B2 = B1 ^ (1 << O) 372 * For example, if the starting buddy (buddy2) is #8 its order 373 * 1 buddy is #10: 374 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 375 * 376 * 2) Any buddy B will have an order O+1 parent P which 377 * satisfies the following equation: 378 * P = B & ~(1 << O) 379 * 380 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 381 */ 382static inline struct page * 383__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 384{ 385 unsigned long buddy_idx = page_idx ^ (1 << order); 386 387 return page + (buddy_idx - page_idx); 388} 389 390static inline unsigned long 391__find_combined_index(unsigned long page_idx, unsigned int order) 392{ 393 return (page_idx & ~(1 << order)); 394} 395 396/* 397 * This function checks whether a page is free && is the buddy 398 * we can do coalesce a page and its buddy if 399 * (a) the buddy is not in a hole && 400 * (b) the buddy is in the buddy system && 401 * (c) a page and its buddy have the same order && 402 * (d) a page and its buddy are in the same zone. 403 * 404 * For recording whether a page is in the buddy system, we use PG_buddy. 405 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 406 * 407 * For recording page's order, we use page_private(page). 408 */ 409static inline int page_is_buddy(struct page *page, struct page *buddy, 410 int order) 411{ 412 if (!pfn_valid_within(page_to_pfn(buddy))) 413 return 0; 414 415 if (page_zone_id(page) != page_zone_id(buddy)) 416 return 0; 417 418 if (PageBuddy(buddy) && page_order(buddy) == order) { 419 VM_BUG_ON(page_count(buddy) != 0); 420 return 1; 421 } 422 return 0; 423} 424 425/* 426 * Freeing function for a buddy system allocator. 427 * 428 * The concept of a buddy system is to maintain direct-mapped table 429 * (containing bit values) for memory blocks of various "orders". 430 * The bottom level table contains the map for the smallest allocatable 431 * units of memory (here, pages), and each level above it describes 432 * pairs of units from the levels below, hence, "buddies". 433 * At a high level, all that happens here is marking the table entry 434 * at the bottom level available, and propagating the changes upward 435 * as necessary, plus some accounting needed to play nicely with other 436 * parts of the VM system. 437 * At each level, we keep a list of pages, which are heads of continuous 438 * free pages of length of (1 << order) and marked with PG_buddy. Page's 439 * order is recorded in page_private(page) field. 440 * So when we are allocating or freeing one, we can derive the state of the 441 * other. That is, if we allocate a small block, and both were 442 * free, the remainder of the region must be split into blocks. 443 * If a block is freed, and its buddy is also free, then this 444 * triggers coalescing into a block of larger size. 445 * 446 * -- wli 447 */ 448 449static inline void __free_one_page(struct page *page, 450 struct zone *zone, unsigned int order, 451 int migratetype) 452{ 453 unsigned long page_idx; 454 455 if (unlikely(PageCompound(page))) 456 if (unlikely(destroy_compound_page(page, order))) 457 return; 458 459 VM_BUG_ON(migratetype == -1); 460 461 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 462 463 VM_BUG_ON(page_idx & ((1 << order) - 1)); 464 VM_BUG_ON(bad_range(zone, page)); 465 466 while (order < MAX_ORDER-1) { 467 unsigned long combined_idx; 468 struct page *buddy; 469 470 buddy = __page_find_buddy(page, page_idx, order); 471 if (!page_is_buddy(page, buddy, order)) 472 break; 473 474 /* Our buddy is free, merge with it and move up one order. */ 475 list_del(&buddy->lru); 476 zone->free_area[order].nr_free--; 477 rmv_page_order(buddy); 478 combined_idx = __find_combined_index(page_idx, order); 479 page = page + (combined_idx - page_idx); 480 page_idx = combined_idx; 481 order++; 482 } 483 set_page_order(page, order); 484 list_add(&page->lru, 485 &zone->free_area[order].free_list[migratetype]); 486 zone->free_area[order].nr_free++; 487} 488 489/* 490 * free_page_mlock() -- clean up attempts to free and mlocked() page. 491 * Page should not be on lru, so no need to fix that up. 492 * free_pages_check() will verify... 493 */ 494static inline void free_page_mlock(struct page *page) 495{ 496 __dec_zone_page_state(page, NR_MLOCK); 497 __count_vm_event(UNEVICTABLE_MLOCKFREED); 498} 499 500static inline int free_pages_check(struct page *page) 501{ 502 if (unlikely(page_mapcount(page) | 503 (page->mapping != NULL) | 504 (atomic_read(&page->_count) != 0) | 505 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 506 bad_page(page); 507 return 1; 508 } 509 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 510 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 511 return 0; 512} 513 514/* 515 * Frees a number of pages from the PCP lists 516 * Assumes all pages on list are in same zone, and of same order. 517 * count is the number of pages to free. 518 * 519 * If the zone was previously in an "all pages pinned" state then look to 520 * see if this freeing clears that state. 521 * 522 * And clear the zone's pages_scanned counter, to hold off the "all pages are 523 * pinned" detection logic. 524 */ 525static void free_pcppages_bulk(struct zone *zone, int count, 526 struct per_cpu_pages *pcp) 527{ 528 int migratetype = 0; 529 int batch_free = 0; 530 531 spin_lock(&zone->lock); 532 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 533 zone->pages_scanned = 0; 534 535 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 536 while (count) { 537 struct page *page; 538 struct list_head *list; 539 540 /* 541 * Remove pages from lists in a round-robin fashion. A 542 * batch_free count is maintained that is incremented when an 543 * empty list is encountered. This is so more pages are freed 544 * off fuller lists instead of spinning excessively around empty 545 * lists 546 */ 547 do { 548 batch_free++; 549 if (++migratetype == MIGRATE_PCPTYPES) 550 migratetype = 0; 551 list = &pcp->lists[migratetype]; 552 } while (list_empty(list)); 553 554 do { 555 page = list_entry(list->prev, struct page, lru); 556 /* must delete as __free_one_page list manipulates */ 557 list_del(&page->lru); 558 __free_one_page(page, zone, 0, migratetype); 559 trace_mm_page_pcpu_drain(page, 0, migratetype); 560 } while (--count && --batch_free && !list_empty(list)); 561 } 562 spin_unlock(&zone->lock); 563} 564 565static void free_one_page(struct zone *zone, struct page *page, int order, 566 int migratetype) 567{ 568 spin_lock(&zone->lock); 569 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 570 zone->pages_scanned = 0; 571 572 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 573 __free_one_page(page, zone, order, migratetype); 574 spin_unlock(&zone->lock); 575} 576 577static void __free_pages_ok(struct page *page, unsigned int order) 578{ 579 unsigned long flags; 580 int i; 581 int bad = 0; 582 int wasMlocked = __TestClearPageMlocked(page); 583 584 kmemcheck_free_shadow(page, order); 585 586 for (i = 0 ; i < (1 << order) ; ++i) 587 bad += free_pages_check(page + i); 588 if (bad) 589 return; 590 591 if (!PageHighMem(page)) { 592 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 593 debug_check_no_obj_freed(page_address(page), 594 PAGE_SIZE << order); 595 } 596 arch_free_page(page, order); 597 kernel_map_pages(page, 1 << order, 0); 598 599 local_irq_save(flags); 600 if (unlikely(wasMlocked)) 601 free_page_mlock(page); 602 __count_vm_events(PGFREE, 1 << order); 603 free_one_page(page_zone(page), page, order, 604 get_pageblock_migratetype(page)); 605 local_irq_restore(flags); 606} 607 608/* 609 * permit the bootmem allocator to evade page validation on high-order frees 610 */ 611void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 612{ 613 if (order == 0) { 614 __ClearPageReserved(page); 615 set_page_count(page, 0); 616 set_page_refcounted(page); 617 __free_page(page); 618 } else { 619 int loop; 620 621 prefetchw(page); 622 for (loop = 0; loop < BITS_PER_LONG; loop++) { 623 struct page *p = &page[loop]; 624 625 if (loop + 1 < BITS_PER_LONG) 626 prefetchw(p + 1); 627 __ClearPageReserved(p); 628 set_page_count(p, 0); 629 } 630 631 set_page_refcounted(page); 632 __free_pages(page, order); 633 } 634} 635 636 637/* 638 * The order of subdivision here is critical for the IO subsystem. 639 * Please do not alter this order without good reasons and regression 640 * testing. Specifically, as large blocks of memory are subdivided, 641 * the order in which smaller blocks are delivered depends on the order 642 * they're subdivided in this function. This is the primary factor 643 * influencing the order in which pages are delivered to the IO 644 * subsystem according to empirical testing, and this is also justified 645 * by considering the behavior of a buddy system containing a single 646 * large block of memory acted on by a series of small allocations. 647 * This behavior is a critical factor in sglist merging's success. 648 * 649 * -- wli 650 */ 651static inline void expand(struct zone *zone, struct page *page, 652 int low, int high, struct free_area *area, 653 int migratetype) 654{ 655 unsigned long size = 1 << high; 656 657 while (high > low) { 658 area--; 659 high--; 660 size >>= 1; 661 VM_BUG_ON(bad_range(zone, &page[size])); 662 list_add(&page[size].lru, &area->free_list[migratetype]); 663 area->nr_free++; 664 set_page_order(&page[size], high); 665 } 666} 667 668/* 669 * This page is about to be returned from the page allocator 670 */ 671static inline int check_new_page(struct page *page) 672{ 673 if (unlikely(page_mapcount(page) | 674 (page->mapping != NULL) | 675 (atomic_read(&page->_count) != 0) | 676 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 677 bad_page(page); 678 return 1; 679 } 680 return 0; 681} 682 683static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 684{ 685 int i; 686 687 for (i = 0; i < (1 << order); i++) { 688 struct page *p = page + i; 689 if (unlikely(check_new_page(p))) 690 return 1; 691 } 692 693 set_page_private(page, 0); 694 set_page_refcounted(page); 695 696 arch_alloc_page(page, order); 697 kernel_map_pages(page, 1 << order, 1); 698 699 if (gfp_flags & __GFP_ZERO) 700 prep_zero_page(page, order, gfp_flags); 701 702 if (order && (gfp_flags & __GFP_COMP)) 703 prep_compound_page(page, order); 704 705 return 0; 706} 707 708/* 709 * Go through the free lists for the given migratetype and remove 710 * the smallest available page from the freelists 711 */ 712static inline 713struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 714 int migratetype) 715{ 716 unsigned int current_order; 717 struct free_area * area; 718 struct page *page; 719 720 /* Find a page of the appropriate size in the preferred list */ 721 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 722 area = &(zone->free_area[current_order]); 723 if (list_empty(&area->free_list[migratetype])) 724 continue; 725 726 page = list_entry(area->free_list[migratetype].next, 727 struct page, lru); 728 list_del(&page->lru); 729 rmv_page_order(page); 730 area->nr_free--; 731 expand(zone, page, order, current_order, area, migratetype); 732 return page; 733 } 734 735 return NULL; 736} 737 738 739/* 740 * This array describes the order lists are fallen back to when 741 * the free lists for the desirable migrate type are depleted 742 */ 743static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 744 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 745 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 746 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 747 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 748}; 749 750/* 751 * Move the free pages in a range to the free lists of the requested type. 752 * Note that start_page and end_pages are not aligned on a pageblock 753 * boundary. If alignment is required, use move_freepages_block() 754 */ 755static int move_freepages(struct zone *zone, 756 struct page *start_page, struct page *end_page, 757 int migratetype) 758{ 759 struct page *page; 760 unsigned long order; 761 int pages_moved = 0; 762 763#ifndef CONFIG_HOLES_IN_ZONE 764 /* 765 * page_zone is not safe to call in this context when 766 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 767 * anyway as we check zone boundaries in move_freepages_block(). 768 * Remove at a later date when no bug reports exist related to 769 * grouping pages by mobility 770 */ 771 BUG_ON(page_zone(start_page) != page_zone(end_page)); 772#endif 773 774 for (page = start_page; page <= end_page;) { 775 /* Make sure we are not inadvertently changing nodes */ 776 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 777 778 if (!pfn_valid_within(page_to_pfn(page))) { 779 page++; 780 continue; 781 } 782 783 if (!PageBuddy(page)) { 784 page++; 785 continue; 786 } 787 788 order = page_order(page); 789 list_del(&page->lru); 790 list_add(&page->lru, 791 &zone->free_area[order].free_list[migratetype]); 792 page += 1 << order; 793 pages_moved += 1 << order; 794 } 795 796 return pages_moved; 797} 798 799static int move_freepages_block(struct zone *zone, struct page *page, 800 int migratetype) 801{ 802 unsigned long start_pfn, end_pfn; 803 struct page *start_page, *end_page; 804 805 start_pfn = page_to_pfn(page); 806 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 807 start_page = pfn_to_page(start_pfn); 808 end_page = start_page + pageblock_nr_pages - 1; 809 end_pfn = start_pfn + pageblock_nr_pages - 1; 810 811 /* Do not cross zone boundaries */ 812 if (start_pfn < zone->zone_start_pfn) 813 start_page = page; 814 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 815 return 0; 816 817 return move_freepages(zone, start_page, end_page, migratetype); 818} 819 820static void change_pageblock_range(struct page *pageblock_page, 821 int start_order, int migratetype) 822{ 823 int nr_pageblocks = 1 << (start_order - pageblock_order); 824 825 while (nr_pageblocks--) { 826 set_pageblock_migratetype(pageblock_page, migratetype); 827 pageblock_page += pageblock_nr_pages; 828 } 829} 830 831/* Remove an element from the buddy allocator from the fallback list */ 832static inline struct page * 833__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 834{ 835 struct free_area * area; 836 int current_order; 837 struct page *page; 838 int migratetype, i; 839 840 /* Find the largest possible block of pages in the other list */ 841 for (current_order = MAX_ORDER-1; current_order >= order; 842 --current_order) { 843 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 844 migratetype = fallbacks[start_migratetype][i]; 845 846 /* MIGRATE_RESERVE handled later if necessary */ 847 if (migratetype == MIGRATE_RESERVE) 848 continue; 849 850 area = &(zone->free_area[current_order]); 851 if (list_empty(&area->free_list[migratetype])) 852 continue; 853 854 page = list_entry(area->free_list[migratetype].next, 855 struct page, lru); 856 area->nr_free--; 857 858 /* 859 * If breaking a large block of pages, move all free 860 * pages to the preferred allocation list. If falling 861 * back for a reclaimable kernel allocation, be more 862 * agressive about taking ownership of free pages 863 */ 864 if (unlikely(current_order >= (pageblock_order >> 1)) || 865 start_migratetype == MIGRATE_RECLAIMABLE || 866 page_group_by_mobility_disabled) { 867 unsigned long pages; 868 pages = move_freepages_block(zone, page, 869 start_migratetype); 870 871 /* Claim the whole block if over half of it is free */ 872 if (pages >= (1 << (pageblock_order-1)) || 873 page_group_by_mobility_disabled) 874 set_pageblock_migratetype(page, 875 start_migratetype); 876 877 migratetype = start_migratetype; 878 } 879 880 /* Remove the page from the freelists */ 881 list_del(&page->lru); 882 rmv_page_order(page); 883 884 /* Take ownership for orders >= pageblock_order */ 885 if (current_order >= pageblock_order) 886 change_pageblock_range(page, current_order, 887 start_migratetype); 888 889 expand(zone, page, order, current_order, area, migratetype); 890 891 trace_mm_page_alloc_extfrag(page, order, current_order, 892 start_migratetype, migratetype); 893 894 return page; 895 } 896 } 897 898 return NULL; 899} 900 901/* 902 * Do the hard work of removing an element from the buddy allocator. 903 * Call me with the zone->lock already held. 904 */ 905static struct page *__rmqueue(struct zone *zone, unsigned int order, 906 int migratetype) 907{ 908 struct page *page; 909 910retry_reserve: 911 page = __rmqueue_smallest(zone, order, migratetype); 912 913 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 914 page = __rmqueue_fallback(zone, order, migratetype); 915 916 /* 917 * Use MIGRATE_RESERVE rather than fail an allocation. goto 918 * is used because __rmqueue_smallest is an inline function 919 * and we want just one call site 920 */ 921 if (!page) { 922 migratetype = MIGRATE_RESERVE; 923 goto retry_reserve; 924 } 925 } 926 927 trace_mm_page_alloc_zone_locked(page, order, migratetype); 928 return page; 929} 930 931/* 932 * Obtain a specified number of elements from the buddy allocator, all under 933 * a single hold of the lock, for efficiency. Add them to the supplied list. 934 * Returns the number of new pages which were placed at *list. 935 */ 936static int rmqueue_bulk(struct zone *zone, unsigned int order, 937 unsigned long count, struct list_head *list, 938 int migratetype, int cold) 939{ 940 int i; 941 942 spin_lock(&zone->lock); 943 for (i = 0; i < count; ++i) { 944 struct page *page = __rmqueue(zone, order, migratetype); 945 if (unlikely(page == NULL)) 946 break; 947 948 /* 949 * Split buddy pages returned by expand() are received here 950 * in physical page order. The page is added to the callers and 951 * list and the list head then moves forward. From the callers 952 * perspective, the linked list is ordered by page number in 953 * some conditions. This is useful for IO devices that can 954 * merge IO requests if the physical pages are ordered 955 * properly. 956 */ 957 if (likely(cold == 0)) 958 list_add(&page->lru, list); 959 else 960 list_add_tail(&page->lru, list); 961 set_page_private(page, migratetype); 962 list = &page->lru; 963 } 964 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 965 spin_unlock(&zone->lock); 966 return i; 967} 968 969#ifdef CONFIG_NUMA 970/* 971 * Called from the vmstat counter updater to drain pagesets of this 972 * currently executing processor on remote nodes after they have 973 * expired. 974 * 975 * Note that this function must be called with the thread pinned to 976 * a single processor. 977 */ 978void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 979{ 980 unsigned long flags; 981 int to_drain; 982 983 local_irq_save(flags); 984 if (pcp->count >= pcp->batch) 985 to_drain = pcp->batch; 986 else 987 to_drain = pcp->count; 988 free_pcppages_bulk(zone, to_drain, pcp); 989 pcp->count -= to_drain; 990 local_irq_restore(flags); 991} 992#endif 993 994/* 995 * Drain pages of the indicated processor. 996 * 997 * The processor must either be the current processor and the 998 * thread pinned to the current processor or a processor that 999 * is not online. 1000 */ 1001static void drain_pages(unsigned int cpu) 1002{ 1003 unsigned long flags; 1004 struct zone *zone; 1005 1006 for_each_populated_zone(zone) { 1007 struct per_cpu_pageset *pset; 1008 struct per_cpu_pages *pcp; 1009 1010 pset = zone_pcp(zone, cpu); 1011 1012 pcp = &pset->pcp; 1013 local_irq_save(flags); 1014 free_pcppages_bulk(zone, pcp->count, pcp); 1015 pcp->count = 0; 1016 local_irq_restore(flags); 1017 } 1018} 1019 1020/* 1021 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1022 */ 1023void drain_local_pages(void *arg) 1024{ 1025 drain_pages(smp_processor_id()); 1026} 1027 1028/* 1029 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1030 */ 1031void drain_all_pages(void) 1032{ 1033 on_each_cpu(drain_local_pages, NULL, 1); 1034} 1035 1036#ifdef CONFIG_HIBERNATION 1037 1038void mark_free_pages(struct zone *zone) 1039{ 1040 unsigned long pfn, max_zone_pfn; 1041 unsigned long flags; 1042 int order, t; 1043 struct list_head *curr; 1044 1045 if (!zone->spanned_pages) 1046 return; 1047 1048 spin_lock_irqsave(&zone->lock, flags); 1049 1050 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1051 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1052 if (pfn_valid(pfn)) { 1053 struct page *page = pfn_to_page(pfn); 1054 1055 if (!swsusp_page_is_forbidden(page)) 1056 swsusp_unset_page_free(page); 1057 } 1058 1059 for_each_migratetype_order(order, t) { 1060 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1061 unsigned long i; 1062 1063 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1064 for (i = 0; i < (1UL << order); i++) 1065 swsusp_set_page_free(pfn_to_page(pfn + i)); 1066 } 1067 } 1068 spin_unlock_irqrestore(&zone->lock, flags); 1069} 1070#endif /* CONFIG_PM */ 1071 1072/* 1073 * Free a 0-order page 1074 */ 1075static void free_hot_cold_page(struct page *page, int cold) 1076{ 1077 struct zone *zone = page_zone(page); 1078 struct per_cpu_pages *pcp; 1079 unsigned long flags; 1080 int migratetype; 1081 int wasMlocked = __TestClearPageMlocked(page); 1082 1083 kmemcheck_free_shadow(page, 0); 1084 1085 if (PageAnon(page)) 1086 page->mapping = NULL; 1087 if (free_pages_check(page)) 1088 return; 1089 1090 if (!PageHighMem(page)) { 1091 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1092 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1093 } 1094 arch_free_page(page, 0); 1095 kernel_map_pages(page, 1, 0); 1096 1097 pcp = &zone_pcp(zone, get_cpu())->pcp; 1098 migratetype = get_pageblock_migratetype(page); 1099 set_page_private(page, migratetype); 1100 local_irq_save(flags); 1101 if (unlikely(wasMlocked)) 1102 free_page_mlock(page); 1103 __count_vm_event(PGFREE); 1104 1105 /* 1106 * We only track unmovable, reclaimable and movable on pcp lists. 1107 * Free ISOLATE pages back to the allocator because they are being 1108 * offlined but treat RESERVE as movable pages so we can get those 1109 * areas back if necessary. Otherwise, we may have to free 1110 * excessively into the page allocator 1111 */ 1112 if (migratetype >= MIGRATE_PCPTYPES) { 1113 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1114 free_one_page(zone, page, 0, migratetype); 1115 goto out; 1116 } 1117 migratetype = MIGRATE_MOVABLE; 1118 } 1119 1120 if (cold) 1121 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1122 else 1123 list_add(&page->lru, &pcp->lists[migratetype]); 1124 pcp->count++; 1125 if (pcp->count >= pcp->high) { 1126 free_pcppages_bulk(zone, pcp->batch, pcp); 1127 pcp->count -= pcp->batch; 1128 } 1129 1130out: 1131 local_irq_restore(flags); 1132 put_cpu(); 1133} 1134 1135void free_hot_page(struct page *page) 1136{ 1137 trace_mm_page_free_direct(page, 0); 1138 free_hot_cold_page(page, 0); 1139} 1140 1141/* 1142 * split_page takes a non-compound higher-order page, and splits it into 1143 * n (1<<order) sub-pages: page[0..n] 1144 * Each sub-page must be freed individually. 1145 * 1146 * Note: this is probably too low level an operation for use in drivers. 1147 * Please consult with lkml before using this in your driver. 1148 */ 1149void split_page(struct page *page, unsigned int order) 1150{ 1151 int i; 1152 1153 VM_BUG_ON(PageCompound(page)); 1154 VM_BUG_ON(!page_count(page)); 1155 1156#ifdef CONFIG_KMEMCHECK 1157 /* 1158 * Split shadow pages too, because free(page[0]) would 1159 * otherwise free the whole shadow. 1160 */ 1161 if (kmemcheck_page_is_tracked(page)) 1162 split_page(virt_to_page(page[0].shadow), order); 1163#endif 1164 1165 for (i = 1; i < (1 << order); i++) 1166 set_page_refcounted(page + i); 1167} 1168 1169/* 1170 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1171 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1172 * or two. 1173 */ 1174static inline 1175struct page *buffered_rmqueue(struct zone *preferred_zone, 1176 struct zone *zone, int order, gfp_t gfp_flags, 1177 int migratetype) 1178{ 1179 unsigned long flags; 1180 struct page *page; 1181 int cold = !!(gfp_flags & __GFP_COLD); 1182 int cpu; 1183 1184again: 1185 cpu = get_cpu(); 1186 if (likely(order == 0)) { 1187 struct per_cpu_pages *pcp; 1188 struct list_head *list; 1189 1190 pcp = &zone_pcp(zone, cpu)->pcp; 1191 list = &pcp->lists[migratetype]; 1192 local_irq_save(flags); 1193 if (list_empty(list)) { 1194 pcp->count += rmqueue_bulk(zone, 0, 1195 pcp->batch, list, 1196 migratetype, cold); 1197 if (unlikely(list_empty(list))) 1198 goto failed; 1199 } 1200 1201 if (cold) 1202 page = list_entry(list->prev, struct page, lru); 1203 else 1204 page = list_entry(list->next, struct page, lru); 1205 1206 list_del(&page->lru); 1207 pcp->count--; 1208 } else { 1209 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1210 /* 1211 * __GFP_NOFAIL is not to be used in new code. 1212 * 1213 * All __GFP_NOFAIL callers should be fixed so that they 1214 * properly detect and handle allocation failures. 1215 * 1216 * We most definitely don't want callers attempting to 1217 * allocate greater than order-1 page units with 1218 * __GFP_NOFAIL. 1219 */ 1220 WARN_ON_ONCE(order > 1); 1221 } 1222 spin_lock_irqsave(&zone->lock, flags); 1223 page = __rmqueue(zone, order, migratetype); 1224 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1225 spin_unlock(&zone->lock); 1226 if (!page) 1227 goto failed; 1228 } 1229 1230 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1231 zone_statistics(preferred_zone, zone); 1232 local_irq_restore(flags); 1233 put_cpu(); 1234 1235 VM_BUG_ON(bad_range(zone, page)); 1236 if (prep_new_page(page, order, gfp_flags)) 1237 goto again; 1238 return page; 1239 1240failed: 1241 local_irq_restore(flags); 1242 put_cpu(); 1243 return NULL; 1244} 1245 1246/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1247#define ALLOC_WMARK_MIN WMARK_MIN 1248#define ALLOC_WMARK_LOW WMARK_LOW 1249#define ALLOC_WMARK_HIGH WMARK_HIGH 1250#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1251 1252/* Mask to get the watermark bits */ 1253#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1254 1255#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1256#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1257#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1258 1259#ifdef CONFIG_FAIL_PAGE_ALLOC 1260 1261static struct fail_page_alloc_attr { 1262 struct fault_attr attr; 1263 1264 u32 ignore_gfp_highmem; 1265 u32 ignore_gfp_wait; 1266 u32 min_order; 1267 1268#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1269 1270 struct dentry *ignore_gfp_highmem_file; 1271 struct dentry *ignore_gfp_wait_file; 1272 struct dentry *min_order_file; 1273 1274#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1275 1276} fail_page_alloc = { 1277 .attr = FAULT_ATTR_INITIALIZER, 1278 .ignore_gfp_wait = 1, 1279 .ignore_gfp_highmem = 1, 1280 .min_order = 1, 1281}; 1282 1283static int __init setup_fail_page_alloc(char *str) 1284{ 1285 return setup_fault_attr(&fail_page_alloc.attr, str); 1286} 1287__setup("fail_page_alloc=", setup_fail_page_alloc); 1288 1289static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1290{ 1291 if (order < fail_page_alloc.min_order) 1292 return 0; 1293 if (gfp_mask & __GFP_NOFAIL) 1294 return 0; 1295 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1296 return 0; 1297 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1298 return 0; 1299 1300 return should_fail(&fail_page_alloc.attr, 1 << order); 1301} 1302 1303#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1304 1305static int __init fail_page_alloc_debugfs(void) 1306{ 1307 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1308 struct dentry *dir; 1309 int err; 1310 1311 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1312 "fail_page_alloc"); 1313 if (err) 1314 return err; 1315 dir = fail_page_alloc.attr.dentries.dir; 1316 1317 fail_page_alloc.ignore_gfp_wait_file = 1318 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1319 &fail_page_alloc.ignore_gfp_wait); 1320 1321 fail_page_alloc.ignore_gfp_highmem_file = 1322 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1323 &fail_page_alloc.ignore_gfp_highmem); 1324 fail_page_alloc.min_order_file = 1325 debugfs_create_u32("min-order", mode, dir, 1326 &fail_page_alloc.min_order); 1327 1328 if (!fail_page_alloc.ignore_gfp_wait_file || 1329 !fail_page_alloc.ignore_gfp_highmem_file || 1330 !fail_page_alloc.min_order_file) { 1331 err = -ENOMEM; 1332 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1333 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1334 debugfs_remove(fail_page_alloc.min_order_file); 1335 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1336 } 1337 1338 return err; 1339} 1340 1341late_initcall(fail_page_alloc_debugfs); 1342 1343#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1344 1345#else /* CONFIG_FAIL_PAGE_ALLOC */ 1346 1347static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1348{ 1349 return 0; 1350} 1351 1352#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1353 1354/* 1355 * Return 1 if free pages are above 'mark'. This takes into account the order 1356 * of the allocation. 1357 */ 1358int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1359 int classzone_idx, int alloc_flags) 1360{ 1361 /* free_pages my go negative - that's OK */ 1362 long min = mark; 1363 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1364 int o; 1365 1366 if (alloc_flags & ALLOC_HIGH) 1367 min -= min / 2; 1368 if (alloc_flags & ALLOC_HARDER) 1369 min -= min / 4; 1370 1371 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1372 return 0; 1373 for (o = 0; o < order; o++) { 1374 /* At the next order, this order's pages become unavailable */ 1375 free_pages -= z->free_area[o].nr_free << o; 1376 1377 /* Require fewer higher order pages to be free */ 1378 min >>= 1; 1379 1380 if (free_pages <= min) 1381 return 0; 1382 } 1383 return 1; 1384} 1385 1386#ifdef CONFIG_NUMA 1387/* 1388 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1389 * skip over zones that are not allowed by the cpuset, or that have 1390 * been recently (in last second) found to be nearly full. See further 1391 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1392 * that have to skip over a lot of full or unallowed zones. 1393 * 1394 * If the zonelist cache is present in the passed in zonelist, then 1395 * returns a pointer to the allowed node mask (either the current 1396 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1397 * 1398 * If the zonelist cache is not available for this zonelist, does 1399 * nothing and returns NULL. 1400 * 1401 * If the fullzones BITMAP in the zonelist cache is stale (more than 1402 * a second since last zap'd) then we zap it out (clear its bits.) 1403 * 1404 * We hold off even calling zlc_setup, until after we've checked the 1405 * first zone in the zonelist, on the theory that most allocations will 1406 * be satisfied from that first zone, so best to examine that zone as 1407 * quickly as we can. 1408 */ 1409static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1410{ 1411 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1412 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1413 1414 zlc = zonelist->zlcache_ptr; 1415 if (!zlc) 1416 return NULL; 1417 1418 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1419 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1420 zlc->last_full_zap = jiffies; 1421 } 1422 1423 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1424 &cpuset_current_mems_allowed : 1425 &node_states[N_HIGH_MEMORY]; 1426 return allowednodes; 1427} 1428 1429/* 1430 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1431 * if it is worth looking at further for free memory: 1432 * 1) Check that the zone isn't thought to be full (doesn't have its 1433 * bit set in the zonelist_cache fullzones BITMAP). 1434 * 2) Check that the zones node (obtained from the zonelist_cache 1435 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1436 * Return true (non-zero) if zone is worth looking at further, or 1437 * else return false (zero) if it is not. 1438 * 1439 * This check -ignores- the distinction between various watermarks, 1440 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1441 * found to be full for any variation of these watermarks, it will 1442 * be considered full for up to one second by all requests, unless 1443 * we are so low on memory on all allowed nodes that we are forced 1444 * into the second scan of the zonelist. 1445 * 1446 * In the second scan we ignore this zonelist cache and exactly 1447 * apply the watermarks to all zones, even it is slower to do so. 1448 * We are low on memory in the second scan, and should leave no stone 1449 * unturned looking for a free page. 1450 */ 1451static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1452 nodemask_t *allowednodes) 1453{ 1454 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1455 int i; /* index of *z in zonelist zones */ 1456 int n; /* node that zone *z is on */ 1457 1458 zlc = zonelist->zlcache_ptr; 1459 if (!zlc) 1460 return 1; 1461 1462 i = z - zonelist->_zonerefs; 1463 n = zlc->z_to_n[i]; 1464 1465 /* This zone is worth trying if it is allowed but not full */ 1466 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1467} 1468 1469/* 1470 * Given 'z' scanning a zonelist, set the corresponding bit in 1471 * zlc->fullzones, so that subsequent attempts to allocate a page 1472 * from that zone don't waste time re-examining it. 1473 */ 1474static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1475{ 1476 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1477 int i; /* index of *z in zonelist zones */ 1478 1479 zlc = zonelist->zlcache_ptr; 1480 if (!zlc) 1481 return; 1482 1483 i = z - zonelist->_zonerefs; 1484 1485 set_bit(i, zlc->fullzones); 1486} 1487 1488#else /* CONFIG_NUMA */ 1489 1490static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1491{ 1492 return NULL; 1493} 1494 1495static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1496 nodemask_t *allowednodes) 1497{ 1498 return 1; 1499} 1500 1501static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1502{ 1503} 1504#endif /* CONFIG_NUMA */ 1505 1506/* 1507 * get_page_from_freelist goes through the zonelist trying to allocate 1508 * a page. 1509 */ 1510static struct page * 1511get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1512 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1513 struct zone *preferred_zone, int migratetype) 1514{ 1515 struct zoneref *z; 1516 struct page *page = NULL; 1517 int classzone_idx; 1518 struct zone *zone; 1519 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1520 int zlc_active = 0; /* set if using zonelist_cache */ 1521 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1522 1523 classzone_idx = zone_idx(preferred_zone); 1524zonelist_scan: 1525 /* 1526 * Scan zonelist, looking for a zone with enough free. 1527 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1528 */ 1529 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1530 high_zoneidx, nodemask) { 1531 if (NUMA_BUILD && zlc_active && 1532 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1533 continue; 1534 if ((alloc_flags & ALLOC_CPUSET) && 1535 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1536 goto try_next_zone; 1537 1538 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1539 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1540 unsigned long mark; 1541 int ret; 1542 1543 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1544 if (zone_watermark_ok(zone, order, mark, 1545 classzone_idx, alloc_flags)) 1546 goto try_this_zone; 1547 1548 if (zone_reclaim_mode == 0) 1549 goto this_zone_full; 1550 1551 ret = zone_reclaim(zone, gfp_mask, order); 1552 switch (ret) { 1553 case ZONE_RECLAIM_NOSCAN: 1554 /* did not scan */ 1555 goto try_next_zone; 1556 case ZONE_RECLAIM_FULL: 1557 /* scanned but unreclaimable */ 1558 goto this_zone_full; 1559 default: 1560 /* did we reclaim enough */ 1561 if (!zone_watermark_ok(zone, order, mark, 1562 classzone_idx, alloc_flags)) 1563 goto this_zone_full; 1564 } 1565 } 1566 1567try_this_zone: 1568 page = buffered_rmqueue(preferred_zone, zone, order, 1569 gfp_mask, migratetype); 1570 if (page) 1571 break; 1572this_zone_full: 1573 if (NUMA_BUILD) 1574 zlc_mark_zone_full(zonelist, z); 1575try_next_zone: 1576 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1577 /* 1578 * we do zlc_setup after the first zone is tried but only 1579 * if there are multiple nodes make it worthwhile 1580 */ 1581 allowednodes = zlc_setup(zonelist, alloc_flags); 1582 zlc_active = 1; 1583 did_zlc_setup = 1; 1584 } 1585 } 1586 1587 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1588 /* Disable zlc cache for second zonelist scan */ 1589 zlc_active = 0; 1590 goto zonelist_scan; 1591 } 1592 return page; 1593} 1594 1595static inline int 1596should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1597 unsigned long pages_reclaimed) 1598{ 1599 /* Do not loop if specifically requested */ 1600 if (gfp_mask & __GFP_NORETRY) 1601 return 0; 1602 1603 /* 1604 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1605 * means __GFP_NOFAIL, but that may not be true in other 1606 * implementations. 1607 */ 1608 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1609 return 1; 1610 1611 /* 1612 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1613 * specified, then we retry until we no longer reclaim any pages 1614 * (above), or we've reclaimed an order of pages at least as 1615 * large as the allocation's order. In both cases, if the 1616 * allocation still fails, we stop retrying. 1617 */ 1618 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1619 return 1; 1620 1621 /* 1622 * Don't let big-order allocations loop unless the caller 1623 * explicitly requests that. 1624 */ 1625 if (gfp_mask & __GFP_NOFAIL) 1626 return 1; 1627 1628 return 0; 1629} 1630 1631static inline struct page * 1632__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1633 struct zonelist *zonelist, enum zone_type high_zoneidx, 1634 nodemask_t *nodemask, struct zone *preferred_zone, 1635 int migratetype) 1636{ 1637 struct page *page; 1638 1639 /* Acquire the OOM killer lock for the zones in zonelist */ 1640 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1641 schedule_timeout_uninterruptible(1); 1642 return NULL; 1643 } 1644 1645 /* 1646 * Go through the zonelist yet one more time, keep very high watermark 1647 * here, this is only to catch a parallel oom killing, we must fail if 1648 * we're still under heavy pressure. 1649 */ 1650 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1651 order, zonelist, high_zoneidx, 1652 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1653 preferred_zone, migratetype); 1654 if (page) 1655 goto out; 1656 1657 /* The OOM killer will not help higher order allocs */ 1658 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL)) 1659 goto out; 1660 1661 /* Exhausted what can be done so it's blamo time */ 1662 out_of_memory(zonelist, gfp_mask, order); 1663 1664out: 1665 clear_zonelist_oom(zonelist, gfp_mask); 1666 return page; 1667} 1668 1669/* The really slow allocator path where we enter direct reclaim */ 1670static inline struct page * 1671__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1672 struct zonelist *zonelist, enum zone_type high_zoneidx, 1673 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1674 int migratetype, unsigned long *did_some_progress) 1675{ 1676 struct page *page = NULL; 1677 struct reclaim_state reclaim_state; 1678 struct task_struct *p = current; 1679 1680 cond_resched(); 1681 1682 /* We now go into synchronous reclaim */ 1683 cpuset_memory_pressure_bump(); 1684 p->flags |= PF_MEMALLOC; 1685 lockdep_set_current_reclaim_state(gfp_mask); 1686 reclaim_state.reclaimed_slab = 0; 1687 p->reclaim_state = &reclaim_state; 1688 1689 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1690 1691 p->reclaim_state = NULL; 1692 lockdep_clear_current_reclaim_state(); 1693 p->flags &= ~PF_MEMALLOC; 1694 1695 cond_resched(); 1696 1697 if (order != 0) 1698 drain_all_pages(); 1699 1700 if (likely(*did_some_progress)) 1701 page = get_page_from_freelist(gfp_mask, nodemask, order, 1702 zonelist, high_zoneidx, 1703 alloc_flags, preferred_zone, 1704 migratetype); 1705 return page; 1706} 1707 1708/* 1709 * This is called in the allocator slow-path if the allocation request is of 1710 * sufficient urgency to ignore watermarks and take other desperate measures 1711 */ 1712static inline struct page * 1713__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1714 struct zonelist *zonelist, enum zone_type high_zoneidx, 1715 nodemask_t *nodemask, struct zone *preferred_zone, 1716 int migratetype) 1717{ 1718 struct page *page; 1719 1720 do { 1721 page = get_page_from_freelist(gfp_mask, nodemask, order, 1722 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1723 preferred_zone, migratetype); 1724 1725 if (!page && gfp_mask & __GFP_NOFAIL) 1726 congestion_wait(BLK_RW_ASYNC, HZ/50); 1727 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1728 1729 return page; 1730} 1731 1732static inline 1733void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1734 enum zone_type high_zoneidx) 1735{ 1736 struct zoneref *z; 1737 struct zone *zone; 1738 1739 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1740 wakeup_kswapd(zone, order); 1741} 1742 1743static inline int 1744gfp_to_alloc_flags(gfp_t gfp_mask) 1745{ 1746 struct task_struct *p = current; 1747 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1748 const gfp_t wait = gfp_mask & __GFP_WAIT; 1749 1750 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1751 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1752 1753 /* 1754 * The caller may dip into page reserves a bit more if the caller 1755 * cannot run direct reclaim, or if the caller has realtime scheduling 1756 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1757 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1758 */ 1759 alloc_flags |= (gfp_mask & __GFP_HIGH); 1760 1761 if (!wait) { 1762 alloc_flags |= ALLOC_HARDER; 1763 /* 1764 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1765 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1766 */ 1767 alloc_flags &= ~ALLOC_CPUSET; 1768 } else if (unlikely(rt_task(p)) && !in_interrupt()) 1769 alloc_flags |= ALLOC_HARDER; 1770 1771 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1772 if (!in_interrupt() && 1773 ((p->flags & PF_MEMALLOC) || 1774 unlikely(test_thread_flag(TIF_MEMDIE)))) 1775 alloc_flags |= ALLOC_NO_WATERMARKS; 1776 } 1777 1778 return alloc_flags; 1779} 1780 1781static inline struct page * 1782__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1783 struct zonelist *zonelist, enum zone_type high_zoneidx, 1784 nodemask_t *nodemask, struct zone *preferred_zone, 1785 int migratetype) 1786{ 1787 const gfp_t wait = gfp_mask & __GFP_WAIT; 1788 struct page *page = NULL; 1789 int alloc_flags; 1790 unsigned long pages_reclaimed = 0; 1791 unsigned long did_some_progress; 1792 struct task_struct *p = current; 1793 1794 /* 1795 * In the slowpath, we sanity check order to avoid ever trying to 1796 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1797 * be using allocators in order of preference for an area that is 1798 * too large. 1799 */ 1800 if (order >= MAX_ORDER) { 1801 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1802 return NULL; 1803 } 1804 1805 /* 1806 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1807 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1808 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1809 * using a larger set of nodes after it has established that the 1810 * allowed per node queues are empty and that nodes are 1811 * over allocated. 1812 */ 1813 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1814 goto nopage; 1815 1816restart: 1817 wake_all_kswapd(order, zonelist, high_zoneidx); 1818 1819 /* 1820 * OK, we're below the kswapd watermark and have kicked background 1821 * reclaim. Now things get more complex, so set up alloc_flags according 1822 * to how we want to proceed. 1823 */ 1824 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1825 1826 /* This is the last chance, in general, before the goto nopage. */ 1827 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1828 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1829 preferred_zone, migratetype); 1830 if (page) 1831 goto got_pg; 1832 1833rebalance: 1834 /* Allocate without watermarks if the context allows */ 1835 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1836 page = __alloc_pages_high_priority(gfp_mask, order, 1837 zonelist, high_zoneidx, nodemask, 1838 preferred_zone, migratetype); 1839 if (page) 1840 goto got_pg; 1841 } 1842 1843 /* Atomic allocations - we can't balance anything */ 1844 if (!wait) 1845 goto nopage; 1846 1847 /* Avoid recursion of direct reclaim */ 1848 if (p->flags & PF_MEMALLOC) 1849 goto nopage; 1850 1851 /* Avoid allocations with no watermarks from looping endlessly */ 1852 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 1853 goto nopage; 1854 1855 /* Try direct reclaim and then allocating */ 1856 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1857 zonelist, high_zoneidx, 1858 nodemask, 1859 alloc_flags, preferred_zone, 1860 migratetype, &did_some_progress); 1861 if (page) 1862 goto got_pg; 1863 1864 /* 1865 * If we failed to make any progress reclaiming, then we are 1866 * running out of options and have to consider going OOM 1867 */ 1868 if (!did_some_progress) { 1869 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1870 if (oom_killer_disabled) 1871 goto nopage; 1872 page = __alloc_pages_may_oom(gfp_mask, order, 1873 zonelist, high_zoneidx, 1874 nodemask, preferred_zone, 1875 migratetype); 1876 if (page) 1877 goto got_pg; 1878 1879 /* 1880 * The OOM killer does not trigger for high-order 1881 * ~__GFP_NOFAIL allocations so if no progress is being 1882 * made, there are no other options and retrying is 1883 * unlikely to help. 1884 */ 1885 if (order > PAGE_ALLOC_COSTLY_ORDER && 1886 !(gfp_mask & __GFP_NOFAIL)) 1887 goto nopage; 1888 1889 goto restart; 1890 } 1891 } 1892 1893 /* Check if we should retry the allocation */ 1894 pages_reclaimed += did_some_progress; 1895 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1896 /* Wait for some write requests to complete then retry */ 1897 congestion_wait(BLK_RW_ASYNC, HZ/50); 1898 goto rebalance; 1899 } 1900 1901nopage: 1902 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1903 printk(KERN_WARNING "%s: page allocation failure." 1904 " order:%d, mode:0x%x\n", 1905 p->comm, order, gfp_mask); 1906 dump_stack(); 1907 show_mem(); 1908 } 1909 return page; 1910got_pg: 1911 if (kmemcheck_enabled) 1912 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 1913 return page; 1914 1915} 1916 1917/* 1918 * This is the 'heart' of the zoned buddy allocator. 1919 */ 1920struct page * 1921__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1922 struct zonelist *zonelist, nodemask_t *nodemask) 1923{ 1924 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1925 struct zone *preferred_zone; 1926 struct page *page; 1927 int migratetype = allocflags_to_migratetype(gfp_mask); 1928 1929 gfp_mask &= gfp_allowed_mask; 1930 1931 lockdep_trace_alloc(gfp_mask); 1932 1933 might_sleep_if(gfp_mask & __GFP_WAIT); 1934 1935 if (should_fail_alloc_page(gfp_mask, order)) 1936 return NULL; 1937 1938 /* 1939 * Check the zones suitable for the gfp_mask contain at least one 1940 * valid zone. It's possible to have an empty zonelist as a result 1941 * of GFP_THISNODE and a memoryless node 1942 */ 1943 if (unlikely(!zonelist->_zonerefs->zone)) 1944 return NULL; 1945 1946 /* The preferred zone is used for statistics later */ 1947 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1948 if (!preferred_zone) 1949 return NULL; 1950 1951 /* First allocation attempt */ 1952 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1953 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1954 preferred_zone, migratetype); 1955 if (unlikely(!page)) 1956 page = __alloc_pages_slowpath(gfp_mask, order, 1957 zonelist, high_zoneidx, nodemask, 1958 preferred_zone, migratetype); 1959 1960 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 1961 return page; 1962} 1963EXPORT_SYMBOL(__alloc_pages_nodemask); 1964 1965/* 1966 * Common helper functions. 1967 */ 1968unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1969{ 1970 struct page *page; 1971 1972 /* 1973 * __get_free_pages() returns a 32-bit address, which cannot represent 1974 * a highmem page 1975 */ 1976 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1977 1978 page = alloc_pages(gfp_mask, order); 1979 if (!page) 1980 return 0; 1981 return (unsigned long) page_address(page); 1982} 1983EXPORT_SYMBOL(__get_free_pages); 1984 1985unsigned long get_zeroed_page(gfp_t gfp_mask) 1986{ 1987 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 1988} 1989EXPORT_SYMBOL(get_zeroed_page); 1990 1991void __pagevec_free(struct pagevec *pvec) 1992{ 1993 int i = pagevec_count(pvec); 1994 1995 while (--i >= 0) { 1996 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 1997 free_hot_cold_page(pvec->pages[i], pvec->cold); 1998 } 1999} 2000 2001void __free_pages(struct page *page, unsigned int order) 2002{ 2003 if (put_page_testzero(page)) { 2004 trace_mm_page_free_direct(page, order); 2005 if (order == 0) 2006 free_hot_page(page); 2007 else 2008 __free_pages_ok(page, order); 2009 } 2010} 2011 2012EXPORT_SYMBOL(__free_pages); 2013 2014void free_pages(unsigned long addr, unsigned int order) 2015{ 2016 if (addr != 0) { 2017 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2018 __free_pages(virt_to_page((void *)addr), order); 2019 } 2020} 2021 2022EXPORT_SYMBOL(free_pages); 2023 2024/** 2025 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2026 * @size: the number of bytes to allocate 2027 * @gfp_mask: GFP flags for the allocation 2028 * 2029 * This function is similar to alloc_pages(), except that it allocates the 2030 * minimum number of pages to satisfy the request. alloc_pages() can only 2031 * allocate memory in power-of-two pages. 2032 * 2033 * This function is also limited by MAX_ORDER. 2034 * 2035 * Memory allocated by this function must be released by free_pages_exact(). 2036 */ 2037void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2038{ 2039 unsigned int order = get_order(size); 2040 unsigned long addr; 2041 2042 addr = __get_free_pages(gfp_mask, order); 2043 if (addr) { 2044 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2045 unsigned long used = addr + PAGE_ALIGN(size); 2046 2047 split_page(virt_to_page((void *)addr), order); 2048 while (used < alloc_end) { 2049 free_page(used); 2050 used += PAGE_SIZE; 2051 } 2052 } 2053 2054 return (void *)addr; 2055} 2056EXPORT_SYMBOL(alloc_pages_exact); 2057 2058/** 2059 * free_pages_exact - release memory allocated via alloc_pages_exact() 2060 * @virt: the value returned by alloc_pages_exact. 2061 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2062 * 2063 * Release the memory allocated by a previous call to alloc_pages_exact. 2064 */ 2065void free_pages_exact(void *virt, size_t size) 2066{ 2067 unsigned long addr = (unsigned long)virt; 2068 unsigned long end = addr + PAGE_ALIGN(size); 2069 2070 while (addr < end) { 2071 free_page(addr); 2072 addr += PAGE_SIZE; 2073 } 2074} 2075EXPORT_SYMBOL(free_pages_exact); 2076 2077static unsigned int nr_free_zone_pages(int offset) 2078{ 2079 struct zoneref *z; 2080 struct zone *zone; 2081 2082 /* Just pick one node, since fallback list is circular */ 2083 unsigned int sum = 0; 2084 2085 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2086 2087 for_each_zone_zonelist(zone, z, zonelist, offset) { 2088 unsigned long size = zone->present_pages; 2089 unsigned long high = high_wmark_pages(zone); 2090 if (size > high) 2091 sum += size - high; 2092 } 2093 2094 return sum; 2095} 2096 2097/* 2098 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2099 */ 2100unsigned int nr_free_buffer_pages(void) 2101{ 2102 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2103} 2104EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2105 2106/* 2107 * Amount of free RAM allocatable within all zones 2108 */ 2109unsigned int nr_free_pagecache_pages(void) 2110{ 2111 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2112} 2113 2114static inline void show_node(struct zone *zone) 2115{ 2116 if (NUMA_BUILD) 2117 printk("Node %d ", zone_to_nid(zone)); 2118} 2119 2120void si_meminfo(struct sysinfo *val) 2121{ 2122 val->totalram = totalram_pages; 2123 val->sharedram = 0; 2124 val->freeram = global_page_state(NR_FREE_PAGES); 2125 val->bufferram = nr_blockdev_pages(); 2126 val->totalhigh = totalhigh_pages; 2127 val->freehigh = nr_free_highpages(); 2128 val->mem_unit = PAGE_SIZE; 2129} 2130 2131EXPORT_SYMBOL(si_meminfo); 2132 2133#ifdef CONFIG_NUMA 2134void si_meminfo_node(struct sysinfo *val, int nid) 2135{ 2136 pg_data_t *pgdat = NODE_DATA(nid); 2137 2138 val->totalram = pgdat->node_present_pages; 2139 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2140#ifdef CONFIG_HIGHMEM 2141 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2142 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2143 NR_FREE_PAGES); 2144#else 2145 val->totalhigh = 0; 2146 val->freehigh = 0; 2147#endif 2148 val->mem_unit = PAGE_SIZE; 2149} 2150#endif 2151 2152#define K(x) ((x) << (PAGE_SHIFT-10)) 2153 2154/* 2155 * Show free area list (used inside shift_scroll-lock stuff) 2156 * We also calculate the percentage fragmentation. We do this by counting the 2157 * memory on each free list with the exception of the first item on the list. 2158 */ 2159void show_free_areas(void) 2160{ 2161 int cpu; 2162 struct zone *zone; 2163 2164 for_each_populated_zone(zone) { 2165 show_node(zone); 2166 printk("%s per-cpu:\n", zone->name); 2167 2168 for_each_online_cpu(cpu) { 2169 struct per_cpu_pageset *pageset; 2170 2171 pageset = zone_pcp(zone, cpu); 2172 2173 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2174 cpu, pageset->pcp.high, 2175 pageset->pcp.batch, pageset->pcp.count); 2176 } 2177 } 2178 2179 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2180 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2181 " unevictable:%lu" 2182 " dirty:%lu writeback:%lu unstable:%lu\n" 2183 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2184 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2185 global_page_state(NR_ACTIVE_ANON), 2186 global_page_state(NR_INACTIVE_ANON), 2187 global_page_state(NR_ISOLATED_ANON), 2188 global_page_state(NR_ACTIVE_FILE), 2189 global_page_state(NR_INACTIVE_FILE), 2190 global_page_state(NR_ISOLATED_FILE), 2191 global_page_state(NR_UNEVICTABLE), 2192 global_page_state(NR_FILE_DIRTY), 2193 global_page_state(NR_WRITEBACK), 2194 global_page_state(NR_UNSTABLE_NFS), 2195 global_page_state(NR_FREE_PAGES), 2196 global_page_state(NR_SLAB_RECLAIMABLE), 2197 global_page_state(NR_SLAB_UNRECLAIMABLE), 2198 global_page_state(NR_FILE_MAPPED), 2199 global_page_state(NR_SHMEM), 2200 global_page_state(NR_PAGETABLE), 2201 global_page_state(NR_BOUNCE)); 2202 2203 for_each_populated_zone(zone) { 2204 int i; 2205 2206 show_node(zone); 2207 printk("%s" 2208 " free:%lukB" 2209 " min:%lukB" 2210 " low:%lukB" 2211 " high:%lukB" 2212 " active_anon:%lukB" 2213 " inactive_anon:%lukB" 2214 " active_file:%lukB" 2215 " inactive_file:%lukB" 2216 " unevictable:%lukB" 2217 " isolated(anon):%lukB" 2218 " isolated(file):%lukB" 2219 " present:%lukB" 2220 " mlocked:%lukB" 2221 " dirty:%lukB" 2222 " writeback:%lukB" 2223 " mapped:%lukB" 2224 " shmem:%lukB" 2225 " slab_reclaimable:%lukB" 2226 " slab_unreclaimable:%lukB" 2227 " kernel_stack:%lukB" 2228 " pagetables:%lukB" 2229 " unstable:%lukB" 2230 " bounce:%lukB" 2231 " writeback_tmp:%lukB" 2232 " pages_scanned:%lu" 2233 " all_unreclaimable? %s" 2234 "\n", 2235 zone->name, 2236 K(zone_page_state(zone, NR_FREE_PAGES)), 2237 K(min_wmark_pages(zone)), 2238 K(low_wmark_pages(zone)), 2239 K(high_wmark_pages(zone)), 2240 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2241 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2242 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2243 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2244 K(zone_page_state(zone, NR_UNEVICTABLE)), 2245 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2246 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2247 K(zone->present_pages), 2248 K(zone_page_state(zone, NR_MLOCK)), 2249 K(zone_page_state(zone, NR_FILE_DIRTY)), 2250 K(zone_page_state(zone, NR_WRITEBACK)), 2251 K(zone_page_state(zone, NR_FILE_MAPPED)), 2252 K(zone_page_state(zone, NR_SHMEM)), 2253 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2254 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2255 zone_page_state(zone, NR_KERNEL_STACK) * 2256 THREAD_SIZE / 1024, 2257 K(zone_page_state(zone, NR_PAGETABLE)), 2258 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2259 K(zone_page_state(zone, NR_BOUNCE)), 2260 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2261 zone->pages_scanned, 2262 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 2263 ); 2264 printk("lowmem_reserve[]:"); 2265 for (i = 0; i < MAX_NR_ZONES; i++) 2266 printk(" %lu", zone->lowmem_reserve[i]); 2267 printk("\n"); 2268 } 2269 2270 for_each_populated_zone(zone) { 2271 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2272 2273 show_node(zone); 2274 printk("%s: ", zone->name); 2275 2276 spin_lock_irqsave(&zone->lock, flags); 2277 for (order = 0; order < MAX_ORDER; order++) { 2278 nr[order] = zone->free_area[order].nr_free; 2279 total += nr[order] << order; 2280 } 2281 spin_unlock_irqrestore(&zone->lock, flags); 2282 for (order = 0; order < MAX_ORDER; order++) 2283 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2284 printk("= %lukB\n", K(total)); 2285 } 2286 2287 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2288 2289 show_swap_cache_info(); 2290} 2291 2292static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2293{ 2294 zoneref->zone = zone; 2295 zoneref->zone_idx = zone_idx(zone); 2296} 2297 2298/* 2299 * Builds allocation fallback zone lists. 2300 * 2301 * Add all populated zones of a node to the zonelist. 2302 */ 2303static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2304 int nr_zones, enum zone_type zone_type) 2305{ 2306 struct zone *zone; 2307 2308 BUG_ON(zone_type >= MAX_NR_ZONES); 2309 zone_type++; 2310 2311 do { 2312 zone_type--; 2313 zone = pgdat->node_zones + zone_type; 2314 if (populated_zone(zone)) { 2315 zoneref_set_zone(zone, 2316 &zonelist->_zonerefs[nr_zones++]); 2317 check_highest_zone(zone_type); 2318 } 2319 2320 } while (zone_type); 2321 return nr_zones; 2322} 2323 2324 2325/* 2326 * zonelist_order: 2327 * 0 = automatic detection of better ordering. 2328 * 1 = order by ([node] distance, -zonetype) 2329 * 2 = order by (-zonetype, [node] distance) 2330 * 2331 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2332 * the same zonelist. So only NUMA can configure this param. 2333 */ 2334#define ZONELIST_ORDER_DEFAULT 0 2335#define ZONELIST_ORDER_NODE 1 2336#define ZONELIST_ORDER_ZONE 2 2337 2338/* zonelist order in the kernel. 2339 * set_zonelist_order() will set this to NODE or ZONE. 2340 */ 2341static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2342static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2343 2344 2345#ifdef CONFIG_NUMA 2346/* The value user specified ....changed by config */ 2347static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2348/* string for sysctl */ 2349#define NUMA_ZONELIST_ORDER_LEN 16 2350char numa_zonelist_order[16] = "default"; 2351 2352/* 2353 * interface for configure zonelist ordering. 2354 * command line option "numa_zonelist_order" 2355 * = "[dD]efault - default, automatic configuration. 2356 * = "[nN]ode - order by node locality, then by zone within node 2357 * = "[zZ]one - order by zone, then by locality within zone 2358 */ 2359 2360static int __parse_numa_zonelist_order(char *s) 2361{ 2362 if (*s == 'd' || *s == 'D') { 2363 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2364 } else if (*s == 'n' || *s == 'N') { 2365 user_zonelist_order = ZONELIST_ORDER_NODE; 2366 } else if (*s == 'z' || *s == 'Z') { 2367 user_zonelist_order = ZONELIST_ORDER_ZONE; 2368 } else { 2369 printk(KERN_WARNING 2370 "Ignoring invalid numa_zonelist_order value: " 2371 "%s\n", s); 2372 return -EINVAL; 2373 } 2374 return 0; 2375} 2376 2377static __init int setup_numa_zonelist_order(char *s) 2378{ 2379 if (s) 2380 return __parse_numa_zonelist_order(s); 2381 return 0; 2382} 2383early_param("numa_zonelist_order", setup_numa_zonelist_order); 2384 2385/* 2386 * sysctl handler for numa_zonelist_order 2387 */ 2388int numa_zonelist_order_handler(ctl_table *table, int write, 2389 void __user *buffer, size_t *length, 2390 loff_t *ppos) 2391{ 2392 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2393 int ret; 2394 2395 if (write) 2396 strncpy(saved_string, (char*)table->data, 2397 NUMA_ZONELIST_ORDER_LEN); 2398 ret = proc_dostring(table, write, buffer, length, ppos); 2399 if (ret) 2400 return ret; 2401 if (write) { 2402 int oldval = user_zonelist_order; 2403 if (__parse_numa_zonelist_order((char*)table->data)) { 2404 /* 2405 * bogus value. restore saved string 2406 */ 2407 strncpy((char*)table->data, saved_string, 2408 NUMA_ZONELIST_ORDER_LEN); 2409 user_zonelist_order = oldval; 2410 } else if (oldval != user_zonelist_order) 2411 build_all_zonelists(); 2412 } 2413 return 0; 2414} 2415 2416 2417#define MAX_NODE_LOAD (nr_online_nodes) 2418static int node_load[MAX_NUMNODES]; 2419 2420/** 2421 * find_next_best_node - find the next node that should appear in a given node's fallback list 2422 * @node: node whose fallback list we're appending 2423 * @used_node_mask: nodemask_t of already used nodes 2424 * 2425 * We use a number of factors to determine which is the next node that should 2426 * appear on a given node's fallback list. The node should not have appeared 2427 * already in @node's fallback list, and it should be the next closest node 2428 * according to the distance array (which contains arbitrary distance values 2429 * from each node to each node in the system), and should also prefer nodes 2430 * with no CPUs, since presumably they'll have very little allocation pressure 2431 * on them otherwise. 2432 * It returns -1 if no node is found. 2433 */ 2434static int find_next_best_node(int node, nodemask_t *used_node_mask) 2435{ 2436 int n, val; 2437 int min_val = INT_MAX; 2438 int best_node = -1; 2439 const struct cpumask *tmp = cpumask_of_node(0); 2440 2441 /* Use the local node if we haven't already */ 2442 if (!node_isset(node, *used_node_mask)) { 2443 node_set(node, *used_node_mask); 2444 return node; 2445 } 2446 2447 for_each_node_state(n, N_HIGH_MEMORY) { 2448 2449 /* Don't want a node to appear more than once */ 2450 if (node_isset(n, *used_node_mask)) 2451 continue; 2452 2453 /* Use the distance array to find the distance */ 2454 val = node_distance(node, n); 2455 2456 /* Penalize nodes under us ("prefer the next node") */ 2457 val += (n < node); 2458 2459 /* Give preference to headless and unused nodes */ 2460 tmp = cpumask_of_node(n); 2461 if (!cpumask_empty(tmp)) 2462 val += PENALTY_FOR_NODE_WITH_CPUS; 2463 2464 /* Slight preference for less loaded node */ 2465 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2466 val += node_load[n]; 2467 2468 if (val < min_val) { 2469 min_val = val; 2470 best_node = n; 2471 } 2472 } 2473 2474 if (best_node >= 0) 2475 node_set(best_node, *used_node_mask); 2476 2477 return best_node; 2478} 2479 2480 2481/* 2482 * Build zonelists ordered by node and zones within node. 2483 * This results in maximum locality--normal zone overflows into local 2484 * DMA zone, if any--but risks exhausting DMA zone. 2485 */ 2486static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2487{ 2488 int j; 2489 struct zonelist *zonelist; 2490 2491 zonelist = &pgdat->node_zonelists[0]; 2492 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2493 ; 2494 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2495 MAX_NR_ZONES - 1); 2496 zonelist->_zonerefs[j].zone = NULL; 2497 zonelist->_zonerefs[j].zone_idx = 0; 2498} 2499 2500/* 2501 * Build gfp_thisnode zonelists 2502 */ 2503static void build_thisnode_zonelists(pg_data_t *pgdat) 2504{ 2505 int j; 2506 struct zonelist *zonelist; 2507 2508 zonelist = &pgdat->node_zonelists[1]; 2509 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2510 zonelist->_zonerefs[j].zone = NULL; 2511 zonelist->_zonerefs[j].zone_idx = 0; 2512} 2513 2514/* 2515 * Build zonelists ordered by zone and nodes within zones. 2516 * This results in conserving DMA zone[s] until all Normal memory is 2517 * exhausted, but results in overflowing to remote node while memory 2518 * may still exist in local DMA zone. 2519 */ 2520static int node_order[MAX_NUMNODES]; 2521 2522static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2523{ 2524 int pos, j, node; 2525 int zone_type; /* needs to be signed */ 2526 struct zone *z; 2527 struct zonelist *zonelist; 2528 2529 zonelist = &pgdat->node_zonelists[0]; 2530 pos = 0; 2531 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2532 for (j = 0; j < nr_nodes; j++) { 2533 node = node_order[j]; 2534 z = &NODE_DATA(node)->node_zones[zone_type]; 2535 if (populated_zone(z)) { 2536 zoneref_set_zone(z, 2537 &zonelist->_zonerefs[pos++]); 2538 check_highest_zone(zone_type); 2539 } 2540 } 2541 } 2542 zonelist->_zonerefs[pos].zone = NULL; 2543 zonelist->_zonerefs[pos].zone_idx = 0; 2544} 2545 2546static int default_zonelist_order(void) 2547{ 2548 int nid, zone_type; 2549 unsigned long low_kmem_size,total_size; 2550 struct zone *z; 2551 int average_size; 2552 /* 2553 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2554 * If they are really small and used heavily, the system can fall 2555 * into OOM very easily. 2556 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2557 */ 2558 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2559 low_kmem_size = 0; 2560 total_size = 0; 2561 for_each_online_node(nid) { 2562 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2563 z = &NODE_DATA(nid)->node_zones[zone_type]; 2564 if (populated_zone(z)) { 2565 if (zone_type < ZONE_NORMAL) 2566 low_kmem_size += z->present_pages; 2567 total_size += z->present_pages; 2568 } 2569 } 2570 } 2571 if (!low_kmem_size || /* there are no DMA area. */ 2572 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2573 return ZONELIST_ORDER_NODE; 2574 /* 2575 * look into each node's config. 2576 * If there is a node whose DMA/DMA32 memory is very big area on 2577 * local memory, NODE_ORDER may be suitable. 2578 */ 2579 average_size = total_size / 2580 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2581 for_each_online_node(nid) { 2582 low_kmem_size = 0; 2583 total_size = 0; 2584 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2585 z = &NODE_DATA(nid)->node_zones[zone_type]; 2586 if (populated_zone(z)) { 2587 if (zone_type < ZONE_NORMAL) 2588 low_kmem_size += z->present_pages; 2589 total_size += z->present_pages; 2590 } 2591 } 2592 if (low_kmem_size && 2593 total_size > average_size && /* ignore small node */ 2594 low_kmem_size > total_size * 70/100) 2595 return ZONELIST_ORDER_NODE; 2596 } 2597 return ZONELIST_ORDER_ZONE; 2598} 2599 2600static void set_zonelist_order(void) 2601{ 2602 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2603 current_zonelist_order = default_zonelist_order(); 2604 else 2605 current_zonelist_order = user_zonelist_order; 2606} 2607 2608static void build_zonelists(pg_data_t *pgdat) 2609{ 2610 int j, node, load; 2611 enum zone_type i; 2612 nodemask_t used_mask; 2613 int local_node, prev_node; 2614 struct zonelist *zonelist; 2615 int order = current_zonelist_order; 2616 2617 /* initialize zonelists */ 2618 for (i = 0; i < MAX_ZONELISTS; i++) { 2619 zonelist = pgdat->node_zonelists + i; 2620 zonelist->_zonerefs[0].zone = NULL; 2621 zonelist->_zonerefs[0].zone_idx = 0; 2622 } 2623 2624 /* NUMA-aware ordering of nodes */ 2625 local_node = pgdat->node_id; 2626 load = nr_online_nodes; 2627 prev_node = local_node; 2628 nodes_clear(used_mask); 2629 2630 memset(node_order, 0, sizeof(node_order)); 2631 j = 0; 2632 2633 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2634 int distance = node_distance(local_node, node); 2635 2636 /* 2637 * If another node is sufficiently far away then it is better 2638 * to reclaim pages in a zone before going off node. 2639 */ 2640 if (distance > RECLAIM_DISTANCE) 2641 zone_reclaim_mode = 1; 2642 2643 /* 2644 * We don't want to pressure a particular node. 2645 * So adding penalty to the first node in same 2646 * distance group to make it round-robin. 2647 */ 2648 if (distance != node_distance(local_node, prev_node)) 2649 node_load[node] = load; 2650 2651 prev_node = node; 2652 load--; 2653 if (order == ZONELIST_ORDER_NODE) 2654 build_zonelists_in_node_order(pgdat, node); 2655 else 2656 node_order[j++] = node; /* remember order */ 2657 } 2658 2659 if (order == ZONELIST_ORDER_ZONE) { 2660 /* calculate node order -- i.e., DMA last! */ 2661 build_zonelists_in_zone_order(pgdat, j); 2662 } 2663 2664 build_thisnode_zonelists(pgdat); 2665} 2666 2667/* Construct the zonelist performance cache - see further mmzone.h */ 2668static void build_zonelist_cache(pg_data_t *pgdat) 2669{ 2670 struct zonelist *zonelist; 2671 struct zonelist_cache *zlc; 2672 struct zoneref *z; 2673 2674 zonelist = &pgdat->node_zonelists[0]; 2675 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2676 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2677 for (z = zonelist->_zonerefs; z->zone; z++) 2678 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2679} 2680 2681 2682#else /* CONFIG_NUMA */ 2683 2684static void set_zonelist_order(void) 2685{ 2686 current_zonelist_order = ZONELIST_ORDER_ZONE; 2687} 2688 2689static void build_zonelists(pg_data_t *pgdat) 2690{ 2691 int node, local_node; 2692 enum zone_type j; 2693 struct zonelist *zonelist; 2694 2695 local_node = pgdat->node_id; 2696 2697 zonelist = &pgdat->node_zonelists[0]; 2698 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2699 2700 /* 2701 * Now we build the zonelist so that it contains the zones 2702 * of all the other nodes. 2703 * We don't want to pressure a particular node, so when 2704 * building the zones for node N, we make sure that the 2705 * zones coming right after the local ones are those from 2706 * node N+1 (modulo N) 2707 */ 2708 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2709 if (!node_online(node)) 2710 continue; 2711 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2712 MAX_NR_ZONES - 1); 2713 } 2714 for (node = 0; node < local_node; node++) { 2715 if (!node_online(node)) 2716 continue; 2717 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2718 MAX_NR_ZONES - 1); 2719 } 2720 2721 zonelist->_zonerefs[j].zone = NULL; 2722 zonelist->_zonerefs[j].zone_idx = 0; 2723} 2724 2725/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2726static void build_zonelist_cache(pg_data_t *pgdat) 2727{ 2728 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2729} 2730 2731#endif /* CONFIG_NUMA */ 2732 2733/* return values int ....just for stop_machine() */ 2734static int __build_all_zonelists(void *dummy) 2735{ 2736 int nid; 2737 2738#ifdef CONFIG_NUMA 2739 memset(node_load, 0, sizeof(node_load)); 2740#endif 2741 for_each_online_node(nid) { 2742 pg_data_t *pgdat = NODE_DATA(nid); 2743 2744 build_zonelists(pgdat); 2745 build_zonelist_cache(pgdat); 2746 } 2747 return 0; 2748} 2749 2750void build_all_zonelists(void) 2751{ 2752 set_zonelist_order(); 2753 2754 if (system_state == SYSTEM_BOOTING) { 2755 __build_all_zonelists(NULL); 2756 mminit_verify_zonelist(); 2757 cpuset_init_current_mems_allowed(); 2758 } else { 2759 /* we have to stop all cpus to guarantee there is no user 2760 of zonelist */ 2761 stop_machine(__build_all_zonelists, NULL, NULL); 2762 /* cpuset refresh routine should be here */ 2763 } 2764 vm_total_pages = nr_free_pagecache_pages(); 2765 /* 2766 * Disable grouping by mobility if the number of pages in the 2767 * system is too low to allow the mechanism to work. It would be 2768 * more accurate, but expensive to check per-zone. This check is 2769 * made on memory-hotadd so a system can start with mobility 2770 * disabled and enable it later 2771 */ 2772 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2773 page_group_by_mobility_disabled = 1; 2774 else 2775 page_group_by_mobility_disabled = 0; 2776 2777 printk("Built %i zonelists in %s order, mobility grouping %s. " 2778 "Total pages: %ld\n", 2779 nr_online_nodes, 2780 zonelist_order_name[current_zonelist_order], 2781 page_group_by_mobility_disabled ? "off" : "on", 2782 vm_total_pages); 2783#ifdef CONFIG_NUMA 2784 printk("Policy zone: %s\n", zone_names[policy_zone]); 2785#endif 2786} 2787 2788/* 2789 * Helper functions to size the waitqueue hash table. 2790 * Essentially these want to choose hash table sizes sufficiently 2791 * large so that collisions trying to wait on pages are rare. 2792 * But in fact, the number of active page waitqueues on typical 2793 * systems is ridiculously low, less than 200. So this is even 2794 * conservative, even though it seems large. 2795 * 2796 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2797 * waitqueues, i.e. the size of the waitq table given the number of pages. 2798 */ 2799#define PAGES_PER_WAITQUEUE 256 2800 2801#ifndef CONFIG_MEMORY_HOTPLUG 2802static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2803{ 2804 unsigned long size = 1; 2805 2806 pages /= PAGES_PER_WAITQUEUE; 2807 2808 while (size < pages) 2809 size <<= 1; 2810 2811 /* 2812 * Once we have dozens or even hundreds of threads sleeping 2813 * on IO we've got bigger problems than wait queue collision. 2814 * Limit the size of the wait table to a reasonable size. 2815 */ 2816 size = min(size, 4096UL); 2817 2818 return max(size, 4UL); 2819} 2820#else 2821/* 2822 * A zone's size might be changed by hot-add, so it is not possible to determine 2823 * a suitable size for its wait_table. So we use the maximum size now. 2824 * 2825 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2826 * 2827 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2828 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2829 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2830 * 2831 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2832 * or more by the traditional way. (See above). It equals: 2833 * 2834 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2835 * ia64(16K page size) : = ( 8G + 4M)byte. 2836 * powerpc (64K page size) : = (32G +16M)byte. 2837 */ 2838static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2839{ 2840 return 4096UL; 2841} 2842#endif 2843 2844/* 2845 * This is an integer logarithm so that shifts can be used later 2846 * to extract the more random high bits from the multiplicative 2847 * hash function before the remainder is taken. 2848 */ 2849static inline unsigned long wait_table_bits(unsigned long size) 2850{ 2851 return ffz(~size); 2852} 2853 2854#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2855 2856/* 2857 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2858 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2859 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2860 * higher will lead to a bigger reserve which will get freed as contiguous 2861 * blocks as reclaim kicks in 2862 */ 2863static void setup_zone_migrate_reserve(struct zone *zone) 2864{ 2865 unsigned long start_pfn, pfn, end_pfn; 2866 struct page *page; 2867 unsigned long block_migratetype; 2868 int reserve; 2869 2870 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2871 start_pfn = zone->zone_start_pfn; 2872 end_pfn = start_pfn + zone->spanned_pages; 2873 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2874 pageblock_order; 2875 2876 /* 2877 * Reserve blocks are generally in place to help high-order atomic 2878 * allocations that are short-lived. A min_free_kbytes value that 2879 * would result in more than 2 reserve blocks for atomic allocations 2880 * is assumed to be in place to help anti-fragmentation for the 2881 * future allocation of hugepages at runtime. 2882 */ 2883 reserve = min(2, reserve); 2884 2885 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2886 if (!pfn_valid(pfn)) 2887 continue; 2888 page = pfn_to_page(pfn); 2889 2890 /* Watch out for overlapping nodes */ 2891 if (page_to_nid(page) != zone_to_nid(zone)) 2892 continue; 2893 2894 /* Blocks with reserved pages will never free, skip them. */ 2895 if (PageReserved(page)) 2896 continue; 2897 2898 block_migratetype = get_pageblock_migratetype(page); 2899 2900 /* If this block is reserved, account for it */ 2901 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2902 reserve--; 2903 continue; 2904 } 2905 2906 /* Suitable for reserving if this block is movable */ 2907 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2908 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2909 move_freepages_block(zone, page, MIGRATE_RESERVE); 2910 reserve--; 2911 continue; 2912 } 2913 2914 /* 2915 * If the reserve is met and this is a previous reserved block, 2916 * take it back 2917 */ 2918 if (block_migratetype == MIGRATE_RESERVE) { 2919 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2920 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2921 } 2922 } 2923} 2924 2925/* 2926 * Initially all pages are reserved - free ones are freed 2927 * up by free_all_bootmem() once the early boot process is 2928 * done. Non-atomic initialization, single-pass. 2929 */ 2930void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2931 unsigned long start_pfn, enum memmap_context context) 2932{ 2933 struct page *page; 2934 unsigned long end_pfn = start_pfn + size; 2935 unsigned long pfn; 2936 struct zone *z; 2937 2938 if (highest_memmap_pfn < end_pfn - 1) 2939 highest_memmap_pfn = end_pfn - 1; 2940 2941 z = &NODE_DATA(nid)->node_zones[zone]; 2942 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2943 /* 2944 * There can be holes in boot-time mem_map[]s 2945 * handed to this function. They do not 2946 * exist on hotplugged memory. 2947 */ 2948 if (context == MEMMAP_EARLY) { 2949 if (!early_pfn_valid(pfn)) 2950 continue; 2951 if (!early_pfn_in_nid(pfn, nid)) 2952 continue; 2953 } 2954 page = pfn_to_page(pfn); 2955 set_page_links(page, zone, nid, pfn); 2956 mminit_verify_page_links(page, zone, nid, pfn); 2957 init_page_count(page); 2958 reset_page_mapcount(page); 2959 SetPageReserved(page); 2960 /* 2961 * Mark the block movable so that blocks are reserved for 2962 * movable at startup. This will force kernel allocations 2963 * to reserve their blocks rather than leaking throughout 2964 * the address space during boot when many long-lived 2965 * kernel allocations are made. Later some blocks near 2966 * the start are marked MIGRATE_RESERVE by 2967 * setup_zone_migrate_reserve() 2968 * 2969 * bitmap is created for zone's valid pfn range. but memmap 2970 * can be created for invalid pages (for alignment) 2971 * check here not to call set_pageblock_migratetype() against 2972 * pfn out of zone. 2973 */ 2974 if ((z->zone_start_pfn <= pfn) 2975 && (pfn < z->zone_start_pfn + z->spanned_pages) 2976 && !(pfn & (pageblock_nr_pages - 1))) 2977 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2978 2979 INIT_LIST_HEAD(&page->lru); 2980#ifdef WANT_PAGE_VIRTUAL 2981 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2982 if (!is_highmem_idx(zone)) 2983 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2984#endif 2985 } 2986} 2987 2988static void __meminit zone_init_free_lists(struct zone *zone) 2989{ 2990 int order, t; 2991 for_each_migratetype_order(order, t) { 2992 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 2993 zone->free_area[order].nr_free = 0; 2994 } 2995} 2996 2997#ifndef __HAVE_ARCH_MEMMAP_INIT 2998#define memmap_init(size, nid, zone, start_pfn) \ 2999 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3000#endif 3001 3002static int zone_batchsize(struct zone *zone) 3003{ 3004#ifdef CONFIG_MMU 3005 int batch; 3006 3007 /* 3008 * The per-cpu-pages pools are set to around 1000th of the 3009 * size of the zone. But no more than 1/2 of a meg. 3010 * 3011 * OK, so we don't know how big the cache is. So guess. 3012 */ 3013 batch = zone->present_pages / 1024; 3014 if (batch * PAGE_SIZE > 512 * 1024) 3015 batch = (512 * 1024) / PAGE_SIZE; 3016 batch /= 4; /* We effectively *= 4 below */ 3017 if (batch < 1) 3018 batch = 1; 3019 3020 /* 3021 * Clamp the batch to a 2^n - 1 value. Having a power 3022 * of 2 value was found to be more likely to have 3023 * suboptimal cache aliasing properties in some cases. 3024 * 3025 * For example if 2 tasks are alternately allocating 3026 * batches of pages, one task can end up with a lot 3027 * of pages of one half of the possible page colors 3028 * and the other with pages of the other colors. 3029 */ 3030 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3031 3032 return batch; 3033 3034#else 3035 /* The deferral and batching of frees should be suppressed under NOMMU 3036 * conditions. 3037 * 3038 * The problem is that NOMMU needs to be able to allocate large chunks 3039 * of contiguous memory as there's no hardware page translation to 3040 * assemble apparent contiguous memory from discontiguous pages. 3041 * 3042 * Queueing large contiguous runs of pages for batching, however, 3043 * causes the pages to actually be freed in smaller chunks. As there 3044 * can be a significant delay between the individual batches being 3045 * recycled, this leads to the once large chunks of space being 3046 * fragmented and becoming unavailable for high-order allocations. 3047 */ 3048 return 0; 3049#endif 3050} 3051 3052static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3053{ 3054 struct per_cpu_pages *pcp; 3055 int migratetype; 3056 3057 memset(p, 0, sizeof(*p)); 3058 3059 pcp = &p->pcp; 3060 pcp->count = 0; 3061 pcp->high = 6 * batch; 3062 pcp->batch = max(1UL, 1 * batch); 3063 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3064 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3065} 3066 3067/* 3068 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3069 * to the value high for the pageset p. 3070 */ 3071 3072static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3073 unsigned long high) 3074{ 3075 struct per_cpu_pages *pcp; 3076 3077 pcp = &p->pcp; 3078 pcp->high = high; 3079 pcp->batch = max(1UL, high/4); 3080 if ((high/4) > (PAGE_SHIFT * 8)) 3081 pcp->batch = PAGE_SHIFT * 8; 3082} 3083 3084 3085#ifdef CONFIG_NUMA 3086/* 3087 * Boot pageset table. One per cpu which is going to be used for all 3088 * zones and all nodes. The parameters will be set in such a way 3089 * that an item put on a list will immediately be handed over to 3090 * the buddy list. This is safe since pageset manipulation is done 3091 * with interrupts disabled. 3092 * 3093 * Some NUMA counter updates may also be caught by the boot pagesets. 3094 * 3095 * The boot_pagesets must be kept even after bootup is complete for 3096 * unused processors and/or zones. They do play a role for bootstrapping 3097 * hotplugged processors. 3098 * 3099 * zoneinfo_show() and maybe other functions do 3100 * not check if the processor is online before following the pageset pointer. 3101 * Other parts of the kernel may not check if the zone is available. 3102 */ 3103static struct per_cpu_pageset boot_pageset[NR_CPUS]; 3104 3105/* 3106 * Dynamically allocate memory for the 3107 * per cpu pageset array in struct zone. 3108 */ 3109static int __cpuinit process_zones(int cpu) 3110{ 3111 struct zone *zone, *dzone; 3112 int node = cpu_to_node(cpu); 3113 3114 node_set_state(node, N_CPU); /* this node has a cpu */ 3115 3116 for_each_populated_zone(zone) { 3117 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 3118 GFP_KERNEL, node); 3119 if (!zone_pcp(zone, cpu)) 3120 goto bad; 3121 3122 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 3123 3124 if (percpu_pagelist_fraction) 3125 setup_pagelist_highmark(zone_pcp(zone, cpu), 3126 (zone->present_pages / percpu_pagelist_fraction)); 3127 } 3128 3129 return 0; 3130bad: 3131 for_each_zone(dzone) { 3132 if (!populated_zone(dzone)) 3133 continue; 3134 if (dzone == zone) 3135 break; 3136 kfree(zone_pcp(dzone, cpu)); 3137 zone_pcp(dzone, cpu) = &boot_pageset[cpu]; 3138 } 3139 return -ENOMEM; 3140} 3141 3142static inline void free_zone_pagesets(int cpu) 3143{ 3144 struct zone *zone; 3145 3146 for_each_zone(zone) { 3147 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 3148 3149 /* Free per_cpu_pageset if it is slab allocated */ 3150 if (pset != &boot_pageset[cpu]) 3151 kfree(pset); 3152 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3153 } 3154} 3155 3156static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 3157 unsigned long action, 3158 void *hcpu) 3159{ 3160 int cpu = (long)hcpu; 3161 int ret = NOTIFY_OK; 3162 3163 switch (action) { 3164 case CPU_UP_PREPARE: 3165 case CPU_UP_PREPARE_FROZEN: 3166 if (process_zones(cpu)) 3167 ret = NOTIFY_BAD; 3168 break; 3169 case CPU_UP_CANCELED: 3170 case CPU_UP_CANCELED_FROZEN: 3171 case CPU_DEAD: 3172 case CPU_DEAD_FROZEN: 3173 free_zone_pagesets(cpu); 3174 break; 3175 default: 3176 break; 3177 } 3178 return ret; 3179} 3180 3181static struct notifier_block __cpuinitdata pageset_notifier = 3182 { &pageset_cpuup_callback, NULL, 0 }; 3183 3184void __init setup_per_cpu_pageset(void) 3185{ 3186 int err; 3187 3188 /* Initialize per_cpu_pageset for cpu 0. 3189 * A cpuup callback will do this for every cpu 3190 * as it comes online 3191 */ 3192 err = process_zones(smp_processor_id()); 3193 BUG_ON(err); 3194 register_cpu_notifier(&pageset_notifier); 3195} 3196 3197#endif 3198 3199static noinline __init_refok 3200int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3201{ 3202 int i; 3203 struct pglist_data *pgdat = zone->zone_pgdat; 3204 size_t alloc_size; 3205 3206 /* 3207 * The per-page waitqueue mechanism uses hashed waitqueues 3208 * per zone. 3209 */ 3210 zone->wait_table_hash_nr_entries = 3211 wait_table_hash_nr_entries(zone_size_pages); 3212 zone->wait_table_bits = 3213 wait_table_bits(zone->wait_table_hash_nr_entries); 3214 alloc_size = zone->wait_table_hash_nr_entries 3215 * sizeof(wait_queue_head_t); 3216 3217 if (!slab_is_available()) { 3218 zone->wait_table = (wait_queue_head_t *) 3219 alloc_bootmem_node(pgdat, alloc_size); 3220 } else { 3221 /* 3222 * This case means that a zone whose size was 0 gets new memory 3223 * via memory hot-add. 3224 * But it may be the case that a new node was hot-added. In 3225 * this case vmalloc() will not be able to use this new node's 3226 * memory - this wait_table must be initialized to use this new 3227 * node itself as well. 3228 * To use this new node's memory, further consideration will be 3229 * necessary. 3230 */ 3231 zone->wait_table = vmalloc(alloc_size); 3232 } 3233 if (!zone->wait_table) 3234 return -ENOMEM; 3235 3236 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3237 init_waitqueue_head(zone->wait_table + i); 3238 3239 return 0; 3240} 3241 3242static int __zone_pcp_update(void *data) 3243{ 3244 struct zone *zone = data; 3245 int cpu; 3246 unsigned long batch = zone_batchsize(zone), flags; 3247 3248 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3249 struct per_cpu_pageset *pset; 3250 struct per_cpu_pages *pcp; 3251 3252 pset = zone_pcp(zone, cpu); 3253 pcp = &pset->pcp; 3254 3255 local_irq_save(flags); 3256 free_pcppages_bulk(zone, pcp->count, pcp); 3257 setup_pageset(pset, batch); 3258 local_irq_restore(flags); 3259 } 3260 return 0; 3261} 3262 3263void zone_pcp_update(struct zone *zone) 3264{ 3265 stop_machine(__zone_pcp_update, zone, NULL); 3266} 3267 3268static __meminit void zone_pcp_init(struct zone *zone) 3269{ 3270 int cpu; 3271 unsigned long batch = zone_batchsize(zone); 3272 3273 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3274#ifdef CONFIG_NUMA 3275 /* Early boot. Slab allocator not functional yet */ 3276 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3277 setup_pageset(&boot_pageset[cpu],0); 3278#else 3279 setup_pageset(zone_pcp(zone,cpu), batch); 3280#endif 3281 } 3282 if (zone->present_pages) 3283 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 3284 zone->name, zone->present_pages, batch); 3285} 3286 3287__meminit int init_currently_empty_zone(struct zone *zone, 3288 unsigned long zone_start_pfn, 3289 unsigned long size, 3290 enum memmap_context context) 3291{ 3292 struct pglist_data *pgdat = zone->zone_pgdat; 3293 int ret; 3294 ret = zone_wait_table_init(zone, size); 3295 if (ret) 3296 return ret; 3297 pgdat->nr_zones = zone_idx(zone) + 1; 3298 3299 zone->zone_start_pfn = zone_start_pfn; 3300 3301 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3302 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3303 pgdat->node_id, 3304 (unsigned long)zone_idx(zone), 3305 zone_start_pfn, (zone_start_pfn + size)); 3306 3307 zone_init_free_lists(zone); 3308 3309 return 0; 3310} 3311 3312#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3313/* 3314 * Basic iterator support. Return the first range of PFNs for a node 3315 * Note: nid == MAX_NUMNODES returns first region regardless of node 3316 */ 3317static int __meminit first_active_region_index_in_nid(int nid) 3318{ 3319 int i; 3320 3321 for (i = 0; i < nr_nodemap_entries; i++) 3322 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3323 return i; 3324 3325 return -1; 3326} 3327 3328/* 3329 * Basic iterator support. Return the next active range of PFNs for a node 3330 * Note: nid == MAX_NUMNODES returns next region regardless of node 3331 */ 3332static int __meminit next_active_region_index_in_nid(int index, int nid) 3333{ 3334 for (index = index + 1; index < nr_nodemap_entries; index++) 3335 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3336 return index; 3337 3338 return -1; 3339} 3340 3341#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3342/* 3343 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3344 * Architectures may implement their own version but if add_active_range() 3345 * was used and there are no special requirements, this is a convenient 3346 * alternative 3347 */ 3348int __meminit __early_pfn_to_nid(unsigned long pfn) 3349{ 3350 int i; 3351 3352 for (i = 0; i < nr_nodemap_entries; i++) { 3353 unsigned long start_pfn = early_node_map[i].start_pfn; 3354 unsigned long end_pfn = early_node_map[i].end_pfn; 3355 3356 if (start_pfn <= pfn && pfn < end_pfn) 3357 return early_node_map[i].nid; 3358 } 3359 /* This is a memory hole */ 3360 return -1; 3361} 3362#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3363 3364int __meminit early_pfn_to_nid(unsigned long pfn) 3365{ 3366 int nid; 3367 3368 nid = __early_pfn_to_nid(pfn); 3369 if (nid >= 0) 3370 return nid; 3371 /* just returns 0 */ 3372 return 0; 3373} 3374 3375#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3376bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3377{ 3378 int nid; 3379 3380 nid = __early_pfn_to_nid(pfn); 3381 if (nid >= 0 && nid != node) 3382 return false; 3383 return true; 3384} 3385#endif 3386 3387/* Basic iterator support to walk early_node_map[] */ 3388#define for_each_active_range_index_in_nid(i, nid) \ 3389 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3390 i = next_active_region_index_in_nid(i, nid)) 3391 3392/** 3393 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3394 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3395 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3396 * 3397 * If an architecture guarantees that all ranges registered with 3398 * add_active_ranges() contain no holes and may be freed, this 3399 * this function may be used instead of calling free_bootmem() manually. 3400 */ 3401void __init free_bootmem_with_active_regions(int nid, 3402 unsigned long max_low_pfn) 3403{ 3404 int i; 3405 3406 for_each_active_range_index_in_nid(i, nid) { 3407 unsigned long size_pages = 0; 3408 unsigned long end_pfn = early_node_map[i].end_pfn; 3409 3410 if (early_node_map[i].start_pfn >= max_low_pfn) 3411 continue; 3412 3413 if (end_pfn > max_low_pfn) 3414 end_pfn = max_low_pfn; 3415 3416 size_pages = end_pfn - early_node_map[i].start_pfn; 3417 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3418 PFN_PHYS(early_node_map[i].start_pfn), 3419 size_pages << PAGE_SHIFT); 3420 } 3421} 3422 3423void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3424{ 3425 int i; 3426 int ret; 3427 3428 for_each_active_range_index_in_nid(i, nid) { 3429 ret = work_fn(early_node_map[i].start_pfn, 3430 early_node_map[i].end_pfn, data); 3431 if (ret) 3432 break; 3433 } 3434} 3435/** 3436 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3437 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3438 * 3439 * If an architecture guarantees that all ranges registered with 3440 * add_active_ranges() contain no holes and may be freed, this 3441 * function may be used instead of calling memory_present() manually. 3442 */ 3443void __init sparse_memory_present_with_active_regions(int nid) 3444{ 3445 int i; 3446 3447 for_each_active_range_index_in_nid(i, nid) 3448 memory_present(early_node_map[i].nid, 3449 early_node_map[i].start_pfn, 3450 early_node_map[i].end_pfn); 3451} 3452 3453/** 3454 * get_pfn_range_for_nid - Return the start and end page frames for a node 3455 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3456 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3457 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3458 * 3459 * It returns the start and end page frame of a node based on information 3460 * provided by an arch calling add_active_range(). If called for a node 3461 * with no available memory, a warning is printed and the start and end 3462 * PFNs will be 0. 3463 */ 3464void __meminit get_pfn_range_for_nid(unsigned int nid, 3465 unsigned long *start_pfn, unsigned long *end_pfn) 3466{ 3467 int i; 3468 *start_pfn = -1UL; 3469 *end_pfn = 0; 3470 3471 for_each_active_range_index_in_nid(i, nid) { 3472 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3473 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3474 } 3475 3476 if (*start_pfn == -1UL) 3477 *start_pfn = 0; 3478} 3479 3480/* 3481 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3482 * assumption is made that zones within a node are ordered in monotonic 3483 * increasing memory addresses so that the "highest" populated zone is used 3484 */ 3485static void __init find_usable_zone_for_movable(void) 3486{ 3487 int zone_index; 3488 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3489 if (zone_index == ZONE_MOVABLE) 3490 continue; 3491 3492 if (arch_zone_highest_possible_pfn[zone_index] > 3493 arch_zone_lowest_possible_pfn[zone_index]) 3494 break; 3495 } 3496 3497 VM_BUG_ON(zone_index == -1); 3498 movable_zone = zone_index; 3499} 3500 3501/* 3502 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3503 * because it is sized independant of architecture. Unlike the other zones, 3504 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3505 * in each node depending on the size of each node and how evenly kernelcore 3506 * is distributed. This helper function adjusts the zone ranges 3507 * provided by the architecture for a given node by using the end of the 3508 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3509 * zones within a node are in order of monotonic increases memory addresses 3510 */ 3511static void __meminit adjust_zone_range_for_zone_movable(int nid, 3512 unsigned long zone_type, 3513 unsigned long node_start_pfn, 3514 unsigned long node_end_pfn, 3515 unsigned long *zone_start_pfn, 3516 unsigned long *zone_end_pfn) 3517{ 3518 /* Only adjust if ZONE_MOVABLE is on this node */ 3519 if (zone_movable_pfn[nid]) { 3520 /* Size ZONE_MOVABLE */ 3521 if (zone_type == ZONE_MOVABLE) { 3522 *zone_start_pfn = zone_movable_pfn[nid]; 3523 *zone_end_pfn = min(node_end_pfn, 3524 arch_zone_highest_possible_pfn[movable_zone]); 3525 3526 /* Adjust for ZONE_MOVABLE starting within this range */ 3527 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3528 *zone_end_pfn > zone_movable_pfn[nid]) { 3529 *zone_end_pfn = zone_movable_pfn[nid]; 3530 3531 /* Check if this whole range is within ZONE_MOVABLE */ 3532 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3533 *zone_start_pfn = *zone_end_pfn; 3534 } 3535} 3536 3537/* 3538 * Return the number of pages a zone spans in a node, including holes 3539 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3540 */ 3541static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3542 unsigned long zone_type, 3543 unsigned long *ignored) 3544{ 3545 unsigned long node_start_pfn, node_end_pfn; 3546 unsigned long zone_start_pfn, zone_end_pfn; 3547 3548 /* Get the start and end of the node and zone */ 3549 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3550 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3551 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3552 adjust_zone_range_for_zone_movable(nid, zone_type, 3553 node_start_pfn, node_end_pfn, 3554 &zone_start_pfn, &zone_end_pfn); 3555 3556 /* Check that this node has pages within the zone's required range */ 3557 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3558 return 0; 3559 3560 /* Move the zone boundaries inside the node if necessary */ 3561 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3562 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3563 3564 /* Return the spanned pages */ 3565 return zone_end_pfn - zone_start_pfn; 3566} 3567 3568/* 3569 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3570 * then all holes in the requested range will be accounted for. 3571 */ 3572static unsigned long __meminit __absent_pages_in_range(int nid, 3573 unsigned long range_start_pfn, 3574 unsigned long range_end_pfn) 3575{ 3576 int i = 0; 3577 unsigned long prev_end_pfn = 0, hole_pages = 0; 3578 unsigned long start_pfn; 3579 3580 /* Find the end_pfn of the first active range of pfns in the node */ 3581 i = first_active_region_index_in_nid(nid); 3582 if (i == -1) 3583 return 0; 3584 3585 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3586 3587 /* Account for ranges before physical memory on this node */ 3588 if (early_node_map[i].start_pfn > range_start_pfn) 3589 hole_pages = prev_end_pfn - range_start_pfn; 3590 3591 /* Find all holes for the zone within the node */ 3592 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3593 3594 /* No need to continue if prev_end_pfn is outside the zone */ 3595 if (prev_end_pfn >= range_end_pfn) 3596 break; 3597 3598 /* Make sure the end of the zone is not within the hole */ 3599 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3600 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3601 3602 /* Update the hole size cound and move on */ 3603 if (start_pfn > range_start_pfn) { 3604 BUG_ON(prev_end_pfn > start_pfn); 3605 hole_pages += start_pfn - prev_end_pfn; 3606 } 3607 prev_end_pfn = early_node_map[i].end_pfn; 3608 } 3609 3610 /* Account for ranges past physical memory on this node */ 3611 if (range_end_pfn > prev_end_pfn) 3612 hole_pages += range_end_pfn - 3613 max(range_start_pfn, prev_end_pfn); 3614 3615 return hole_pages; 3616} 3617 3618/** 3619 * absent_pages_in_range - Return number of page frames in holes within a range 3620 * @start_pfn: The start PFN to start searching for holes 3621 * @end_pfn: The end PFN to stop searching for holes 3622 * 3623 * It returns the number of pages frames in memory holes within a range. 3624 */ 3625unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3626 unsigned long end_pfn) 3627{ 3628 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3629} 3630 3631/* Return the number of page frames in holes in a zone on a node */ 3632static unsigned long __meminit zone_absent_pages_in_node(int nid, 3633 unsigned long zone_type, 3634 unsigned long *ignored) 3635{ 3636 unsigned long node_start_pfn, node_end_pfn; 3637 unsigned long zone_start_pfn, zone_end_pfn; 3638 3639 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3640 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3641 node_start_pfn); 3642 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3643 node_end_pfn); 3644 3645 adjust_zone_range_for_zone_movable(nid, zone_type, 3646 node_start_pfn, node_end_pfn, 3647 &zone_start_pfn, &zone_end_pfn); 3648 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3649} 3650 3651#else 3652static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3653 unsigned long zone_type, 3654 unsigned long *zones_size) 3655{ 3656 return zones_size[zone_type]; 3657} 3658 3659static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3660 unsigned long zone_type, 3661 unsigned long *zholes_size) 3662{ 3663 if (!zholes_size) 3664 return 0; 3665 3666 return zholes_size[zone_type]; 3667} 3668 3669#endif 3670 3671static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3672 unsigned long *zones_size, unsigned long *zholes_size) 3673{ 3674 unsigned long realtotalpages, totalpages = 0; 3675 enum zone_type i; 3676 3677 for (i = 0; i < MAX_NR_ZONES; i++) 3678 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3679 zones_size); 3680 pgdat->node_spanned_pages = totalpages; 3681 3682 realtotalpages = totalpages; 3683 for (i = 0; i < MAX_NR_ZONES; i++) 3684 realtotalpages -= 3685 zone_absent_pages_in_node(pgdat->node_id, i, 3686 zholes_size); 3687 pgdat->node_present_pages = realtotalpages; 3688 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3689 realtotalpages); 3690} 3691 3692#ifndef CONFIG_SPARSEMEM 3693/* 3694 * Calculate the size of the zone->blockflags rounded to an unsigned long 3695 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3696 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3697 * round what is now in bits to nearest long in bits, then return it in 3698 * bytes. 3699 */ 3700static unsigned long __init usemap_size(unsigned long zonesize) 3701{ 3702 unsigned long usemapsize; 3703 3704 usemapsize = roundup(zonesize, pageblock_nr_pages); 3705 usemapsize = usemapsize >> pageblock_order; 3706 usemapsize *= NR_PAGEBLOCK_BITS; 3707 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3708 3709 return usemapsize / 8; 3710} 3711 3712static void __init setup_usemap(struct pglist_data *pgdat, 3713 struct zone *zone, unsigned long zonesize) 3714{ 3715 unsigned long usemapsize = usemap_size(zonesize); 3716 zone->pageblock_flags = NULL; 3717 if (usemapsize) 3718 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3719} 3720#else 3721static void inline setup_usemap(struct pglist_data *pgdat, 3722 struct zone *zone, unsigned long zonesize) {} 3723#endif /* CONFIG_SPARSEMEM */ 3724 3725#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3726 3727/* Return a sensible default order for the pageblock size. */ 3728static inline int pageblock_default_order(void) 3729{ 3730 if (HPAGE_SHIFT > PAGE_SHIFT) 3731 return HUGETLB_PAGE_ORDER; 3732 3733 return MAX_ORDER-1; 3734} 3735 3736/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3737static inline void __init set_pageblock_order(unsigned int order) 3738{ 3739 /* Check that pageblock_nr_pages has not already been setup */ 3740 if (pageblock_order) 3741 return; 3742 3743 /* 3744 * Assume the largest contiguous order of interest is a huge page. 3745 * This value may be variable depending on boot parameters on IA64 3746 */ 3747 pageblock_order = order; 3748} 3749#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3750 3751/* 3752 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3753 * and pageblock_default_order() are unused as pageblock_order is set 3754 * at compile-time. See include/linux/pageblock-flags.h for the values of 3755 * pageblock_order based on the kernel config 3756 */ 3757static inline int pageblock_default_order(unsigned int order) 3758{ 3759 return MAX_ORDER-1; 3760} 3761#define set_pageblock_order(x) do {} while (0) 3762 3763#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3764 3765/* 3766 * Set up the zone data structures: 3767 * - mark all pages reserved 3768 * - mark all memory queues empty 3769 * - clear the memory bitmaps 3770 */ 3771static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3772 unsigned long *zones_size, unsigned long *zholes_size) 3773{ 3774 enum zone_type j; 3775 int nid = pgdat->node_id; 3776 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3777 int ret; 3778 3779 pgdat_resize_init(pgdat); 3780 pgdat->nr_zones = 0; 3781 init_waitqueue_head(&pgdat->kswapd_wait); 3782 pgdat->kswapd_max_order = 0; 3783 pgdat_page_cgroup_init(pgdat); 3784 3785 for (j = 0; j < MAX_NR_ZONES; j++) { 3786 struct zone *zone = pgdat->node_zones + j; 3787 unsigned long size, realsize, memmap_pages; 3788 enum lru_list l; 3789 3790 size = zone_spanned_pages_in_node(nid, j, zones_size); 3791 realsize = size - zone_absent_pages_in_node(nid, j, 3792 zholes_size); 3793 3794 /* 3795 * Adjust realsize so that it accounts for how much memory 3796 * is used by this zone for memmap. This affects the watermark 3797 * and per-cpu initialisations 3798 */ 3799 memmap_pages = 3800 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3801 if (realsize >= memmap_pages) { 3802 realsize -= memmap_pages; 3803 if (memmap_pages) 3804 printk(KERN_DEBUG 3805 " %s zone: %lu pages used for memmap\n", 3806 zone_names[j], memmap_pages); 3807 } else 3808 printk(KERN_WARNING 3809 " %s zone: %lu pages exceeds realsize %lu\n", 3810 zone_names[j], memmap_pages, realsize); 3811 3812 /* Account for reserved pages */ 3813 if (j == 0 && realsize > dma_reserve) { 3814 realsize -= dma_reserve; 3815 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3816 zone_names[0], dma_reserve); 3817 } 3818 3819 if (!is_highmem_idx(j)) 3820 nr_kernel_pages += realsize; 3821 nr_all_pages += realsize; 3822 3823 zone->spanned_pages = size; 3824 zone->present_pages = realsize; 3825#ifdef CONFIG_NUMA 3826 zone->node = nid; 3827 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3828 / 100; 3829 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3830#endif 3831 zone->name = zone_names[j]; 3832 spin_lock_init(&zone->lock); 3833 spin_lock_init(&zone->lru_lock); 3834 zone_seqlock_init(zone); 3835 zone->zone_pgdat = pgdat; 3836 3837 zone->prev_priority = DEF_PRIORITY; 3838 3839 zone_pcp_init(zone); 3840 for_each_lru(l) { 3841 INIT_LIST_HEAD(&zone->lru[l].list); 3842 zone->reclaim_stat.nr_saved_scan[l] = 0; 3843 } 3844 zone->reclaim_stat.recent_rotated[0] = 0; 3845 zone->reclaim_stat.recent_rotated[1] = 0; 3846 zone->reclaim_stat.recent_scanned[0] = 0; 3847 zone->reclaim_stat.recent_scanned[1] = 0; 3848 zap_zone_vm_stats(zone); 3849 zone->flags = 0; 3850 if (!size) 3851 continue; 3852 3853 set_pageblock_order(pageblock_default_order()); 3854 setup_usemap(pgdat, zone, size); 3855 ret = init_currently_empty_zone(zone, zone_start_pfn, 3856 size, MEMMAP_EARLY); 3857 BUG_ON(ret); 3858 memmap_init(size, nid, j, zone_start_pfn); 3859 zone_start_pfn += size; 3860 } 3861} 3862 3863static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3864{ 3865 /* Skip empty nodes */ 3866 if (!pgdat->node_spanned_pages) 3867 return; 3868 3869#ifdef CONFIG_FLAT_NODE_MEM_MAP 3870 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3871 if (!pgdat->node_mem_map) { 3872 unsigned long size, start, end; 3873 struct page *map; 3874 3875 /* 3876 * The zone's endpoints aren't required to be MAX_ORDER 3877 * aligned but the node_mem_map endpoints must be in order 3878 * for the buddy allocator to function correctly. 3879 */ 3880 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3881 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3882 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3883 size = (end - start) * sizeof(struct page); 3884 map = alloc_remap(pgdat->node_id, size); 3885 if (!map) 3886 map = alloc_bootmem_node(pgdat, size); 3887 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3888 } 3889#ifndef CONFIG_NEED_MULTIPLE_NODES 3890 /* 3891 * With no DISCONTIG, the global mem_map is just set as node 0's 3892 */ 3893 if (pgdat == NODE_DATA(0)) { 3894 mem_map = NODE_DATA(0)->node_mem_map; 3895#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3896 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3897 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3898#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3899 } 3900#endif 3901#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3902} 3903 3904void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3905 unsigned long node_start_pfn, unsigned long *zholes_size) 3906{ 3907 pg_data_t *pgdat = NODE_DATA(nid); 3908 3909 pgdat->node_id = nid; 3910 pgdat->node_start_pfn = node_start_pfn; 3911 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3912 3913 alloc_node_mem_map(pgdat); 3914#ifdef CONFIG_FLAT_NODE_MEM_MAP 3915 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3916 nid, (unsigned long)pgdat, 3917 (unsigned long)pgdat->node_mem_map); 3918#endif 3919 3920 free_area_init_core(pgdat, zones_size, zholes_size); 3921} 3922 3923#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3924 3925#if MAX_NUMNODES > 1 3926/* 3927 * Figure out the number of possible node ids. 3928 */ 3929static void __init setup_nr_node_ids(void) 3930{ 3931 unsigned int node; 3932 unsigned int highest = 0; 3933 3934 for_each_node_mask(node, node_possible_map) 3935 highest = node; 3936 nr_node_ids = highest + 1; 3937} 3938#else 3939static inline void setup_nr_node_ids(void) 3940{ 3941} 3942#endif 3943 3944/** 3945 * add_active_range - Register a range of PFNs backed by physical memory 3946 * @nid: The node ID the range resides on 3947 * @start_pfn: The start PFN of the available physical memory 3948 * @end_pfn: The end PFN of the available physical memory 3949 * 3950 * These ranges are stored in an early_node_map[] and later used by 3951 * free_area_init_nodes() to calculate zone sizes and holes. If the 3952 * range spans a memory hole, it is up to the architecture to ensure 3953 * the memory is not freed by the bootmem allocator. If possible 3954 * the range being registered will be merged with existing ranges. 3955 */ 3956void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3957 unsigned long end_pfn) 3958{ 3959 int i; 3960 3961 mminit_dprintk(MMINIT_TRACE, "memory_register", 3962 "Entering add_active_range(%d, %#lx, %#lx) " 3963 "%d entries of %d used\n", 3964 nid, start_pfn, end_pfn, 3965 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3966 3967 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 3968 3969 /* Merge with existing active regions if possible */ 3970 for (i = 0; i < nr_nodemap_entries; i++) { 3971 if (early_node_map[i].nid != nid) 3972 continue; 3973 3974 /* Skip if an existing region covers this new one */ 3975 if (start_pfn >= early_node_map[i].start_pfn && 3976 end_pfn <= early_node_map[i].end_pfn) 3977 return; 3978 3979 /* Merge forward if suitable */ 3980 if (start_pfn <= early_node_map[i].end_pfn && 3981 end_pfn > early_node_map[i].end_pfn) { 3982 early_node_map[i].end_pfn = end_pfn; 3983 return; 3984 } 3985 3986 /* Merge backward if suitable */ 3987 if (start_pfn < early_node_map[i].end_pfn && 3988 end_pfn >= early_node_map[i].start_pfn) { 3989 early_node_map[i].start_pfn = start_pfn; 3990 return; 3991 } 3992 } 3993 3994 /* Check that early_node_map is large enough */ 3995 if (i >= MAX_ACTIVE_REGIONS) { 3996 printk(KERN_CRIT "More than %d memory regions, truncating\n", 3997 MAX_ACTIVE_REGIONS); 3998 return; 3999 } 4000 4001 early_node_map[i].nid = nid; 4002 early_node_map[i].start_pfn = start_pfn; 4003 early_node_map[i].end_pfn = end_pfn; 4004 nr_nodemap_entries = i + 1; 4005} 4006 4007/** 4008 * remove_active_range - Shrink an existing registered range of PFNs 4009 * @nid: The node id the range is on that should be shrunk 4010 * @start_pfn: The new PFN of the range 4011 * @end_pfn: The new PFN of the range 4012 * 4013 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4014 * The map is kept near the end physical page range that has already been 4015 * registered. This function allows an arch to shrink an existing registered 4016 * range. 4017 */ 4018void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4019 unsigned long end_pfn) 4020{ 4021 int i, j; 4022 int removed = 0; 4023 4024 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4025 nid, start_pfn, end_pfn); 4026 4027 /* Find the old active region end and shrink */ 4028 for_each_active_range_index_in_nid(i, nid) { 4029 if (early_node_map[i].start_pfn >= start_pfn && 4030 early_node_map[i].end_pfn <= end_pfn) { 4031 /* clear it */ 4032 early_node_map[i].start_pfn = 0; 4033 early_node_map[i].end_pfn = 0; 4034 removed = 1; 4035 continue; 4036 } 4037 if (early_node_map[i].start_pfn < start_pfn && 4038 early_node_map[i].end_pfn > start_pfn) { 4039 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4040 early_node_map[i].end_pfn = start_pfn; 4041 if (temp_end_pfn > end_pfn) 4042 add_active_range(nid, end_pfn, temp_end_pfn); 4043 continue; 4044 } 4045 if (early_node_map[i].start_pfn >= start_pfn && 4046 early_node_map[i].end_pfn > end_pfn && 4047 early_node_map[i].start_pfn < end_pfn) { 4048 early_node_map[i].start_pfn = end_pfn; 4049 continue; 4050 } 4051 } 4052 4053 if (!removed) 4054 return; 4055 4056 /* remove the blank ones */ 4057 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4058 if (early_node_map[i].nid != nid) 4059 continue; 4060 if (early_node_map[i].end_pfn) 4061 continue; 4062 /* we found it, get rid of it */ 4063 for (j = i; j < nr_nodemap_entries - 1; j++) 4064 memcpy(&early_node_map[j], &early_node_map[j+1], 4065 sizeof(early_node_map[j])); 4066 j = nr_nodemap_entries - 1; 4067 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4068 nr_nodemap_entries--; 4069 } 4070} 4071 4072/** 4073 * remove_all_active_ranges - Remove all currently registered regions 4074 * 4075 * During discovery, it may be found that a table like SRAT is invalid 4076 * and an alternative discovery method must be used. This function removes 4077 * all currently registered regions. 4078 */ 4079void __init remove_all_active_ranges(void) 4080{ 4081 memset(early_node_map, 0, sizeof(early_node_map)); 4082 nr_nodemap_entries = 0; 4083} 4084 4085/* Compare two active node_active_regions */ 4086static int __init cmp_node_active_region(const void *a, const void *b) 4087{ 4088 struct node_active_region *arange = (struct node_active_region *)a; 4089 struct node_active_region *brange = (struct node_active_region *)b; 4090 4091 /* Done this way to avoid overflows */ 4092 if (arange->start_pfn > brange->start_pfn) 4093 return 1; 4094 if (arange->start_pfn < brange->start_pfn) 4095 return -1; 4096 4097 return 0; 4098} 4099 4100/* sort the node_map by start_pfn */ 4101static void __init sort_node_map(void) 4102{ 4103 sort(early_node_map, (size_t)nr_nodemap_entries, 4104 sizeof(struct node_active_region), 4105 cmp_node_active_region, NULL); 4106} 4107 4108/* Find the lowest pfn for a node */ 4109static unsigned long __init find_min_pfn_for_node(int nid) 4110{ 4111 int i; 4112 unsigned long min_pfn = ULONG_MAX; 4113 4114 /* Assuming a sorted map, the first range found has the starting pfn */ 4115 for_each_active_range_index_in_nid(i, nid) 4116 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4117 4118 if (min_pfn == ULONG_MAX) { 4119 printk(KERN_WARNING 4120 "Could not find start_pfn for node %d\n", nid); 4121 return 0; 4122 } 4123 4124 return min_pfn; 4125} 4126 4127/** 4128 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4129 * 4130 * It returns the minimum PFN based on information provided via 4131 * add_active_range(). 4132 */ 4133unsigned long __init find_min_pfn_with_active_regions(void) 4134{ 4135 return find_min_pfn_for_node(MAX_NUMNODES); 4136} 4137 4138/* 4139 * early_calculate_totalpages() 4140 * Sum pages in active regions for movable zone. 4141 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4142 */ 4143static unsigned long __init early_calculate_totalpages(void) 4144{ 4145 int i; 4146 unsigned long totalpages = 0; 4147 4148 for (i = 0; i < nr_nodemap_entries; i++) { 4149 unsigned long pages = early_node_map[i].end_pfn - 4150 early_node_map[i].start_pfn; 4151 totalpages += pages; 4152 if (pages) 4153 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4154 } 4155 return totalpages; 4156} 4157 4158/* 4159 * Find the PFN the Movable zone begins in each node. Kernel memory 4160 * is spread evenly between nodes as long as the nodes have enough 4161 * memory. When they don't, some nodes will have more kernelcore than 4162 * others 4163 */ 4164static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4165{ 4166 int i, nid; 4167 unsigned long usable_startpfn; 4168 unsigned long kernelcore_node, kernelcore_remaining; 4169 /* save the state before borrow the nodemask */ 4170 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4171 unsigned long totalpages = early_calculate_totalpages(); 4172 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4173 4174 /* 4175 * If movablecore was specified, calculate what size of 4176 * kernelcore that corresponds so that memory usable for 4177 * any allocation type is evenly spread. If both kernelcore 4178 * and movablecore are specified, then the value of kernelcore 4179 * will be used for required_kernelcore if it's greater than 4180 * what movablecore would have allowed. 4181 */ 4182 if (required_movablecore) { 4183 unsigned long corepages; 4184 4185 /* 4186 * Round-up so that ZONE_MOVABLE is at least as large as what 4187 * was requested by the user 4188 */ 4189 required_movablecore = 4190 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4191 corepages = totalpages - required_movablecore; 4192 4193 required_kernelcore = max(required_kernelcore, corepages); 4194 } 4195 4196 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4197 if (!required_kernelcore) 4198 goto out; 4199 4200 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4201 find_usable_zone_for_movable(); 4202 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4203 4204restart: 4205 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4206 kernelcore_node = required_kernelcore / usable_nodes; 4207 for_each_node_state(nid, N_HIGH_MEMORY) { 4208 /* 4209 * Recalculate kernelcore_node if the division per node 4210 * now exceeds what is necessary to satisfy the requested 4211 * amount of memory for the kernel 4212 */ 4213 if (required_kernelcore < kernelcore_node) 4214 kernelcore_node = required_kernelcore / usable_nodes; 4215 4216 /* 4217 * As the map is walked, we track how much memory is usable 4218 * by the kernel using kernelcore_remaining. When it is 4219 * 0, the rest of the node is usable by ZONE_MOVABLE 4220 */ 4221 kernelcore_remaining = kernelcore_node; 4222 4223 /* Go through each range of PFNs within this node */ 4224 for_each_active_range_index_in_nid(i, nid) { 4225 unsigned long start_pfn, end_pfn; 4226 unsigned long size_pages; 4227 4228 start_pfn = max(early_node_map[i].start_pfn, 4229 zone_movable_pfn[nid]); 4230 end_pfn = early_node_map[i].end_pfn; 4231 if (start_pfn >= end_pfn) 4232 continue; 4233 4234 /* Account for what is only usable for kernelcore */ 4235 if (start_pfn < usable_startpfn) { 4236 unsigned long kernel_pages; 4237 kernel_pages = min(end_pfn, usable_startpfn) 4238 - start_pfn; 4239 4240 kernelcore_remaining -= min(kernel_pages, 4241 kernelcore_remaining); 4242 required_kernelcore -= min(kernel_pages, 4243 required_kernelcore); 4244 4245 /* Continue if range is now fully accounted */ 4246 if (end_pfn <= usable_startpfn) { 4247 4248 /* 4249 * Push zone_movable_pfn to the end so 4250 * that if we have to rebalance 4251 * kernelcore across nodes, we will 4252 * not double account here 4253 */ 4254 zone_movable_pfn[nid] = end_pfn; 4255 continue; 4256 } 4257 start_pfn = usable_startpfn; 4258 } 4259 4260 /* 4261 * The usable PFN range for ZONE_MOVABLE is from 4262 * start_pfn->end_pfn. Calculate size_pages as the 4263 * number of pages used as kernelcore 4264 */ 4265 size_pages = end_pfn - start_pfn; 4266 if (size_pages > kernelcore_remaining) 4267 size_pages = kernelcore_remaining; 4268 zone_movable_pfn[nid] = start_pfn + size_pages; 4269 4270 /* 4271 * Some kernelcore has been met, update counts and 4272 * break if the kernelcore for this node has been 4273 * satisified 4274 */ 4275 required_kernelcore -= min(required_kernelcore, 4276 size_pages); 4277 kernelcore_remaining -= size_pages; 4278 if (!kernelcore_remaining) 4279 break; 4280 } 4281 } 4282 4283 /* 4284 * If there is still required_kernelcore, we do another pass with one 4285 * less node in the count. This will push zone_movable_pfn[nid] further 4286 * along on the nodes that still have memory until kernelcore is 4287 * satisified 4288 */ 4289 usable_nodes--; 4290 if (usable_nodes && required_kernelcore > usable_nodes) 4291 goto restart; 4292 4293 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4294 for (nid = 0; nid < MAX_NUMNODES; nid++) 4295 zone_movable_pfn[nid] = 4296 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4297 4298out: 4299 /* restore the node_state */ 4300 node_states[N_HIGH_MEMORY] = saved_node_state; 4301} 4302 4303/* Any regular memory on that node ? */ 4304static void check_for_regular_memory(pg_data_t *pgdat) 4305{ 4306#ifdef CONFIG_HIGHMEM 4307 enum zone_type zone_type; 4308 4309 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4310 struct zone *zone = &pgdat->node_zones[zone_type]; 4311 if (zone->present_pages) 4312 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4313 } 4314#endif 4315} 4316 4317/** 4318 * free_area_init_nodes - Initialise all pg_data_t and zone data 4319 * @max_zone_pfn: an array of max PFNs for each zone 4320 * 4321 * This will call free_area_init_node() for each active node in the system. 4322 * Using the page ranges provided by add_active_range(), the size of each 4323 * zone in each node and their holes is calculated. If the maximum PFN 4324 * between two adjacent zones match, it is assumed that the zone is empty. 4325 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4326 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4327 * starts where the previous one ended. For example, ZONE_DMA32 starts 4328 * at arch_max_dma_pfn. 4329 */ 4330void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4331{ 4332 unsigned long nid; 4333 int i; 4334 4335 /* Sort early_node_map as initialisation assumes it is sorted */ 4336 sort_node_map(); 4337 4338 /* Record where the zone boundaries are */ 4339 memset(arch_zone_lowest_possible_pfn, 0, 4340 sizeof(arch_zone_lowest_possible_pfn)); 4341 memset(arch_zone_highest_possible_pfn, 0, 4342 sizeof(arch_zone_highest_possible_pfn)); 4343 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4344 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4345 for (i = 1; i < MAX_NR_ZONES; i++) { 4346 if (i == ZONE_MOVABLE) 4347 continue; 4348 arch_zone_lowest_possible_pfn[i] = 4349 arch_zone_highest_possible_pfn[i-1]; 4350 arch_zone_highest_possible_pfn[i] = 4351 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4352 } 4353 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4354 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4355 4356 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4357 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4358 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4359 4360 /* Print out the zone ranges */ 4361 printk("Zone PFN ranges:\n"); 4362 for (i = 0; i < MAX_NR_ZONES; i++) { 4363 if (i == ZONE_MOVABLE) 4364 continue; 4365 printk(" %-8s %0#10lx -> %0#10lx\n", 4366 zone_names[i], 4367 arch_zone_lowest_possible_pfn[i], 4368 arch_zone_highest_possible_pfn[i]); 4369 } 4370 4371 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4372 printk("Movable zone start PFN for each node\n"); 4373 for (i = 0; i < MAX_NUMNODES; i++) { 4374 if (zone_movable_pfn[i]) 4375 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4376 } 4377 4378 /* Print out the early_node_map[] */ 4379 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4380 for (i = 0; i < nr_nodemap_entries; i++) 4381 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4382 early_node_map[i].start_pfn, 4383 early_node_map[i].end_pfn); 4384 4385 /* Initialise every node */ 4386 mminit_verify_pageflags_layout(); 4387 setup_nr_node_ids(); 4388 for_each_online_node(nid) { 4389 pg_data_t *pgdat = NODE_DATA(nid); 4390 free_area_init_node(nid, NULL, 4391 find_min_pfn_for_node(nid), NULL); 4392 4393 /* Any memory on that node */ 4394 if (pgdat->node_present_pages) 4395 node_set_state(nid, N_HIGH_MEMORY); 4396 check_for_regular_memory(pgdat); 4397 } 4398} 4399 4400static int __init cmdline_parse_core(char *p, unsigned long *core) 4401{ 4402 unsigned long long coremem; 4403 if (!p) 4404 return -EINVAL; 4405 4406 coremem = memparse(p, &p); 4407 *core = coremem >> PAGE_SHIFT; 4408 4409 /* Paranoid check that UL is enough for the coremem value */ 4410 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4411 4412 return 0; 4413} 4414 4415/* 4416 * kernelcore=size sets the amount of memory for use for allocations that 4417 * cannot be reclaimed or migrated. 4418 */ 4419static int __init cmdline_parse_kernelcore(char *p) 4420{ 4421 return cmdline_parse_core(p, &required_kernelcore); 4422} 4423 4424/* 4425 * movablecore=size sets the amount of memory for use for allocations that 4426 * can be reclaimed or migrated. 4427 */ 4428static int __init cmdline_parse_movablecore(char *p) 4429{ 4430 return cmdline_parse_core(p, &required_movablecore); 4431} 4432 4433early_param("kernelcore", cmdline_parse_kernelcore); 4434early_param("movablecore", cmdline_parse_movablecore); 4435 4436#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4437 4438/** 4439 * set_dma_reserve - set the specified number of pages reserved in the first zone 4440 * @new_dma_reserve: The number of pages to mark reserved 4441 * 4442 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4443 * In the DMA zone, a significant percentage may be consumed by kernel image 4444 * and other unfreeable allocations which can skew the watermarks badly. This 4445 * function may optionally be used to account for unfreeable pages in the 4446 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4447 * smaller per-cpu batchsize. 4448 */ 4449void __init set_dma_reserve(unsigned long new_dma_reserve) 4450{ 4451 dma_reserve = new_dma_reserve; 4452} 4453 4454#ifndef CONFIG_NEED_MULTIPLE_NODES 4455struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; 4456EXPORT_SYMBOL(contig_page_data); 4457#endif 4458 4459void __init free_area_init(unsigned long *zones_size) 4460{ 4461 free_area_init_node(0, zones_size, 4462 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4463} 4464 4465static int page_alloc_cpu_notify(struct notifier_block *self, 4466 unsigned long action, void *hcpu) 4467{ 4468 int cpu = (unsigned long)hcpu; 4469 4470 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4471 drain_pages(cpu); 4472 4473 /* 4474 * Spill the event counters of the dead processor 4475 * into the current processors event counters. 4476 * This artificially elevates the count of the current 4477 * processor. 4478 */ 4479 vm_events_fold_cpu(cpu); 4480 4481 /* 4482 * Zero the differential counters of the dead processor 4483 * so that the vm statistics are consistent. 4484 * 4485 * This is only okay since the processor is dead and cannot 4486 * race with what we are doing. 4487 */ 4488 refresh_cpu_vm_stats(cpu); 4489 } 4490 return NOTIFY_OK; 4491} 4492 4493void __init page_alloc_init(void) 4494{ 4495 hotcpu_notifier(page_alloc_cpu_notify, 0); 4496} 4497 4498/* 4499 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4500 * or min_free_kbytes changes. 4501 */ 4502static void calculate_totalreserve_pages(void) 4503{ 4504 struct pglist_data *pgdat; 4505 unsigned long reserve_pages = 0; 4506 enum zone_type i, j; 4507 4508 for_each_online_pgdat(pgdat) { 4509 for (i = 0; i < MAX_NR_ZONES; i++) { 4510 struct zone *zone = pgdat->node_zones + i; 4511 unsigned long max = 0; 4512 4513 /* Find valid and maximum lowmem_reserve in the zone */ 4514 for (j = i; j < MAX_NR_ZONES; j++) { 4515 if (zone->lowmem_reserve[j] > max) 4516 max = zone->lowmem_reserve[j]; 4517 } 4518 4519 /* we treat the high watermark as reserved pages. */ 4520 max += high_wmark_pages(zone); 4521 4522 if (max > zone->present_pages) 4523 max = zone->present_pages; 4524 reserve_pages += max; 4525 } 4526 } 4527 totalreserve_pages = reserve_pages; 4528} 4529 4530/* 4531 * setup_per_zone_lowmem_reserve - called whenever 4532 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4533 * has a correct pages reserved value, so an adequate number of 4534 * pages are left in the zone after a successful __alloc_pages(). 4535 */ 4536static void setup_per_zone_lowmem_reserve(void) 4537{ 4538 struct pglist_data *pgdat; 4539 enum zone_type j, idx; 4540 4541 for_each_online_pgdat(pgdat) { 4542 for (j = 0; j < MAX_NR_ZONES; j++) { 4543 struct zone *zone = pgdat->node_zones + j; 4544 unsigned long present_pages = zone->present_pages; 4545 4546 zone->lowmem_reserve[j] = 0; 4547 4548 idx = j; 4549 while (idx) { 4550 struct zone *lower_zone; 4551 4552 idx--; 4553 4554 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4555 sysctl_lowmem_reserve_ratio[idx] = 1; 4556 4557 lower_zone = pgdat->node_zones + idx; 4558 lower_zone->lowmem_reserve[j] = present_pages / 4559 sysctl_lowmem_reserve_ratio[idx]; 4560 present_pages += lower_zone->present_pages; 4561 } 4562 } 4563 } 4564 4565 /* update totalreserve_pages */ 4566 calculate_totalreserve_pages(); 4567} 4568 4569/** 4570 * setup_per_zone_wmarks - called when min_free_kbytes changes 4571 * or when memory is hot-{added|removed} 4572 * 4573 * Ensures that the watermark[min,low,high] values for each zone are set 4574 * correctly with respect to min_free_kbytes. 4575 */ 4576void setup_per_zone_wmarks(void) 4577{ 4578 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4579 unsigned long lowmem_pages = 0; 4580 struct zone *zone; 4581 unsigned long flags; 4582 4583 /* Calculate total number of !ZONE_HIGHMEM pages */ 4584 for_each_zone(zone) { 4585 if (!is_highmem(zone)) 4586 lowmem_pages += zone->present_pages; 4587 } 4588 4589 for_each_zone(zone) { 4590 u64 tmp; 4591 4592 spin_lock_irqsave(&zone->lock, flags); 4593 tmp = (u64)pages_min * zone->present_pages; 4594 do_div(tmp, lowmem_pages); 4595 if (is_highmem(zone)) { 4596 /* 4597 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4598 * need highmem pages, so cap pages_min to a small 4599 * value here. 4600 * 4601 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4602 * deltas controls asynch page reclaim, and so should 4603 * not be capped for highmem. 4604 */ 4605 int min_pages; 4606 4607 min_pages = zone->present_pages / 1024; 4608 if (min_pages < SWAP_CLUSTER_MAX) 4609 min_pages = SWAP_CLUSTER_MAX; 4610 if (min_pages > 128) 4611 min_pages = 128; 4612 zone->watermark[WMARK_MIN] = min_pages; 4613 } else { 4614 /* 4615 * If it's a lowmem zone, reserve a number of pages 4616 * proportionate to the zone's size. 4617 */ 4618 zone->watermark[WMARK_MIN] = tmp; 4619 } 4620 4621 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4622 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4623 setup_zone_migrate_reserve(zone); 4624 spin_unlock_irqrestore(&zone->lock, flags); 4625 } 4626 4627 /* update totalreserve_pages */ 4628 calculate_totalreserve_pages(); 4629} 4630 4631/* 4632 * The inactive anon list should be small enough that the VM never has to 4633 * do too much work, but large enough that each inactive page has a chance 4634 * to be referenced again before it is swapped out. 4635 * 4636 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4637 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4638 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4639 * the anonymous pages are kept on the inactive list. 4640 * 4641 * total target max 4642 * memory ratio inactive anon 4643 * ------------------------------------- 4644 * 10MB 1 5MB 4645 * 100MB 1 50MB 4646 * 1GB 3 250MB 4647 * 10GB 10 0.9GB 4648 * 100GB 31 3GB 4649 * 1TB 101 10GB 4650 * 10TB 320 32GB 4651 */ 4652void calculate_zone_inactive_ratio(struct zone *zone) 4653{ 4654 unsigned int gb, ratio; 4655 4656 /* Zone size in gigabytes */ 4657 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4658 if (gb) 4659 ratio = int_sqrt(10 * gb); 4660 else 4661 ratio = 1; 4662 4663 zone->inactive_ratio = ratio; 4664} 4665 4666static void __init setup_per_zone_inactive_ratio(void) 4667{ 4668 struct zone *zone; 4669 4670 for_each_zone(zone) 4671 calculate_zone_inactive_ratio(zone); 4672} 4673 4674/* 4675 * Initialise min_free_kbytes. 4676 * 4677 * For small machines we want it small (128k min). For large machines 4678 * we want it large (64MB max). But it is not linear, because network 4679 * bandwidth does not increase linearly with machine size. We use 4680 * 4681 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4682 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4683 * 4684 * which yields 4685 * 4686 * 16MB: 512k 4687 * 32MB: 724k 4688 * 64MB: 1024k 4689 * 128MB: 1448k 4690 * 256MB: 2048k 4691 * 512MB: 2896k 4692 * 1024MB: 4096k 4693 * 2048MB: 5792k 4694 * 4096MB: 8192k 4695 * 8192MB: 11584k 4696 * 16384MB: 16384k 4697 */ 4698static int __init init_per_zone_wmark_min(void) 4699{ 4700 unsigned long lowmem_kbytes; 4701 4702 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4703 4704 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4705 if (min_free_kbytes < 128) 4706 min_free_kbytes = 128; 4707 if (min_free_kbytes > 65536) 4708 min_free_kbytes = 65536; 4709 setup_per_zone_wmarks(); 4710 setup_per_zone_lowmem_reserve(); 4711 setup_per_zone_inactive_ratio(); 4712 return 0; 4713} 4714module_init(init_per_zone_wmark_min) 4715 4716/* 4717 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4718 * that we can call two helper functions whenever min_free_kbytes 4719 * changes. 4720 */ 4721int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4722 void __user *buffer, size_t *length, loff_t *ppos) 4723{ 4724 proc_dointvec(table, write, buffer, length, ppos); 4725 if (write) 4726 setup_per_zone_wmarks(); 4727 return 0; 4728} 4729 4730#ifdef CONFIG_NUMA 4731int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4732 void __user *buffer, size_t *length, loff_t *ppos) 4733{ 4734 struct zone *zone; 4735 int rc; 4736 4737 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4738 if (rc) 4739 return rc; 4740 4741 for_each_zone(zone) 4742 zone->min_unmapped_pages = (zone->present_pages * 4743 sysctl_min_unmapped_ratio) / 100; 4744 return 0; 4745} 4746 4747int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4748 void __user *buffer, size_t *length, loff_t *ppos) 4749{ 4750 struct zone *zone; 4751 int rc; 4752 4753 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4754 if (rc) 4755 return rc; 4756 4757 for_each_zone(zone) 4758 zone->min_slab_pages = (zone->present_pages * 4759 sysctl_min_slab_ratio) / 100; 4760 return 0; 4761} 4762#endif 4763 4764/* 4765 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4766 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4767 * whenever sysctl_lowmem_reserve_ratio changes. 4768 * 4769 * The reserve ratio obviously has absolutely no relation with the 4770 * minimum watermarks. The lowmem reserve ratio can only make sense 4771 * if in function of the boot time zone sizes. 4772 */ 4773int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4774 void __user *buffer, size_t *length, loff_t *ppos) 4775{ 4776 proc_dointvec_minmax(table, write, buffer, length, ppos); 4777 setup_per_zone_lowmem_reserve(); 4778 return 0; 4779} 4780 4781/* 4782 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4783 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4784 * can have before it gets flushed back to buddy allocator. 4785 */ 4786 4787int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4788 void __user *buffer, size_t *length, loff_t *ppos) 4789{ 4790 struct zone *zone; 4791 unsigned int cpu; 4792 int ret; 4793 4794 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 4795 if (!write || (ret == -EINVAL)) 4796 return ret; 4797 for_each_populated_zone(zone) { 4798 for_each_online_cpu(cpu) { 4799 unsigned long high; 4800 high = zone->present_pages / percpu_pagelist_fraction; 4801 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4802 } 4803 } 4804 return 0; 4805} 4806 4807int hashdist = HASHDIST_DEFAULT; 4808 4809#ifdef CONFIG_NUMA 4810static int __init set_hashdist(char *str) 4811{ 4812 if (!str) 4813 return 0; 4814 hashdist = simple_strtoul(str, &str, 0); 4815 return 1; 4816} 4817__setup("hashdist=", set_hashdist); 4818#endif 4819 4820/* 4821 * allocate a large system hash table from bootmem 4822 * - it is assumed that the hash table must contain an exact power-of-2 4823 * quantity of entries 4824 * - limit is the number of hash buckets, not the total allocation size 4825 */ 4826void *__init alloc_large_system_hash(const char *tablename, 4827 unsigned long bucketsize, 4828 unsigned long numentries, 4829 int scale, 4830 int flags, 4831 unsigned int *_hash_shift, 4832 unsigned int *_hash_mask, 4833 unsigned long limit) 4834{ 4835 unsigned long long max = limit; 4836 unsigned long log2qty, size; 4837 void *table = NULL; 4838 4839 /* allow the kernel cmdline to have a say */ 4840 if (!numentries) { 4841 /* round applicable memory size up to nearest megabyte */ 4842 numentries = nr_kernel_pages; 4843 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4844 numentries >>= 20 - PAGE_SHIFT; 4845 numentries <<= 20 - PAGE_SHIFT; 4846 4847 /* limit to 1 bucket per 2^scale bytes of low memory */ 4848 if (scale > PAGE_SHIFT) 4849 numentries >>= (scale - PAGE_SHIFT); 4850 else 4851 numentries <<= (PAGE_SHIFT - scale); 4852 4853 /* Make sure we've got at least a 0-order allocation.. */ 4854 if (unlikely(flags & HASH_SMALL)) { 4855 /* Makes no sense without HASH_EARLY */ 4856 WARN_ON(!(flags & HASH_EARLY)); 4857 if (!(numentries >> *_hash_shift)) { 4858 numentries = 1UL << *_hash_shift; 4859 BUG_ON(!numentries); 4860 } 4861 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4862 numentries = PAGE_SIZE / bucketsize; 4863 } 4864 numentries = roundup_pow_of_two(numentries); 4865 4866 /* limit allocation size to 1/16 total memory by default */ 4867 if (max == 0) { 4868 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4869 do_div(max, bucketsize); 4870 } 4871 4872 if (numentries > max) 4873 numentries = max; 4874 4875 log2qty = ilog2(numentries); 4876 4877 do { 4878 size = bucketsize << log2qty; 4879 if (flags & HASH_EARLY) 4880 table = alloc_bootmem_nopanic(size); 4881 else if (hashdist) 4882 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4883 else { 4884 /* 4885 * If bucketsize is not a power-of-two, we may free 4886 * some pages at the end of hash table which 4887 * alloc_pages_exact() automatically does 4888 */ 4889 if (get_order(size) < MAX_ORDER) { 4890 table = alloc_pages_exact(size, GFP_ATOMIC); 4891 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4892 } 4893 } 4894 } while (!table && size > PAGE_SIZE && --log2qty); 4895 4896 if (!table) 4897 panic("Failed to allocate %s hash table\n", tablename); 4898 4899 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4900 tablename, 4901 (1U << log2qty), 4902 ilog2(size) - PAGE_SHIFT, 4903 size); 4904 4905 if (_hash_shift) 4906 *_hash_shift = log2qty; 4907 if (_hash_mask) 4908 *_hash_mask = (1 << log2qty) - 1; 4909 4910 return table; 4911} 4912 4913/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4914static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4915 unsigned long pfn) 4916{ 4917#ifdef CONFIG_SPARSEMEM 4918 return __pfn_to_section(pfn)->pageblock_flags; 4919#else 4920 return zone->pageblock_flags; 4921#endif /* CONFIG_SPARSEMEM */ 4922} 4923 4924static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4925{ 4926#ifdef CONFIG_SPARSEMEM 4927 pfn &= (PAGES_PER_SECTION-1); 4928 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4929#else 4930 pfn = pfn - zone->zone_start_pfn; 4931 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4932#endif /* CONFIG_SPARSEMEM */ 4933} 4934 4935/** 4936 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4937 * @page: The page within the block of interest 4938 * @start_bitidx: The first bit of interest to retrieve 4939 * @end_bitidx: The last bit of interest 4940 * returns pageblock_bits flags 4941 */ 4942unsigned long get_pageblock_flags_group(struct page *page, 4943 int start_bitidx, int end_bitidx) 4944{ 4945 struct zone *zone; 4946 unsigned long *bitmap; 4947 unsigned long pfn, bitidx; 4948 unsigned long flags = 0; 4949 unsigned long value = 1; 4950 4951 zone = page_zone(page); 4952 pfn = page_to_pfn(page); 4953 bitmap = get_pageblock_bitmap(zone, pfn); 4954 bitidx = pfn_to_bitidx(zone, pfn); 4955 4956 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4957 if (test_bit(bitidx + start_bitidx, bitmap)) 4958 flags |= value; 4959 4960 return flags; 4961} 4962 4963/** 4964 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 4965 * @page: The page within the block of interest 4966 * @start_bitidx: The first bit of interest 4967 * @end_bitidx: The last bit of interest 4968 * @flags: The flags to set 4969 */ 4970void set_pageblock_flags_group(struct page *page, unsigned long flags, 4971 int start_bitidx, int end_bitidx) 4972{ 4973 struct zone *zone; 4974 unsigned long *bitmap; 4975 unsigned long pfn, bitidx; 4976 unsigned long value = 1; 4977 4978 zone = page_zone(page); 4979 pfn = page_to_pfn(page); 4980 bitmap = get_pageblock_bitmap(zone, pfn); 4981 bitidx = pfn_to_bitidx(zone, pfn); 4982 VM_BUG_ON(pfn < zone->zone_start_pfn); 4983 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 4984 4985 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4986 if (flags & value) 4987 __set_bit(bitidx + start_bitidx, bitmap); 4988 else 4989 __clear_bit(bitidx + start_bitidx, bitmap); 4990} 4991 4992/* 4993 * This is designed as sub function...plz see page_isolation.c also. 4994 * set/clear page block's type to be ISOLATE. 4995 * page allocater never alloc memory from ISOLATE block. 4996 */ 4997 4998int set_migratetype_isolate(struct page *page) 4999{ 5000 struct zone *zone; 5001 unsigned long flags; 5002 int ret = -EBUSY; 5003 int zone_idx; 5004 5005 zone = page_zone(page); 5006 zone_idx = zone_idx(zone); 5007 spin_lock_irqsave(&zone->lock, flags); 5008 /* 5009 * In future, more migrate types will be able to be isolation target. 5010 */ 5011 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE && 5012 zone_idx != ZONE_MOVABLE) 5013 goto out; 5014 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5015 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5016 ret = 0; 5017out: 5018 spin_unlock_irqrestore(&zone->lock, flags); 5019 if (!ret) 5020 drain_all_pages(); 5021 return ret; 5022} 5023 5024void unset_migratetype_isolate(struct page *page) 5025{ 5026 struct zone *zone; 5027 unsigned long flags; 5028 zone = page_zone(page); 5029 spin_lock_irqsave(&zone->lock, flags); 5030 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5031 goto out; 5032 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5033 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5034out: 5035 spin_unlock_irqrestore(&zone->lock, flags); 5036} 5037 5038#ifdef CONFIG_MEMORY_HOTREMOVE 5039/* 5040 * All pages in the range must be isolated before calling this. 5041 */ 5042void 5043__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5044{ 5045 struct page *page; 5046 struct zone *zone; 5047 int order, i; 5048 unsigned long pfn; 5049 unsigned long flags; 5050 /* find the first valid pfn */ 5051 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5052 if (pfn_valid(pfn)) 5053 break; 5054 if (pfn == end_pfn) 5055 return; 5056 zone = page_zone(pfn_to_page(pfn)); 5057 spin_lock_irqsave(&zone->lock, flags); 5058 pfn = start_pfn; 5059 while (pfn < end_pfn) { 5060 if (!pfn_valid(pfn)) { 5061 pfn++; 5062 continue; 5063 } 5064 page = pfn_to_page(pfn); 5065 BUG_ON(page_count(page)); 5066 BUG_ON(!PageBuddy(page)); 5067 order = page_order(page); 5068#ifdef CONFIG_DEBUG_VM 5069 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5070 pfn, 1 << order, end_pfn); 5071#endif 5072 list_del(&page->lru); 5073 rmv_page_order(page); 5074 zone->free_area[order].nr_free--; 5075 __mod_zone_page_state(zone, NR_FREE_PAGES, 5076 - (1UL << order)); 5077 for (i = 0; i < (1 << order); i++) 5078 SetPageReserved((page+i)); 5079 pfn += (1 << order); 5080 } 5081 spin_unlock_irqrestore(&zone->lock, flags); 5082} 5083#endif 5084 5085#ifdef CONFIG_MEMORY_FAILURE 5086bool is_free_buddy_page(struct page *page) 5087{ 5088 struct zone *zone = page_zone(page); 5089 unsigned long pfn = page_to_pfn(page); 5090 unsigned long flags; 5091 int order; 5092 5093 spin_lock_irqsave(&zone->lock, flags); 5094 for (order = 0; order < MAX_ORDER; order++) { 5095 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5096 5097 if (PageBuddy(page_head) && page_order(page_head) >= order) 5098 break; 5099 } 5100 spin_unlock_irqrestore(&zone->lock, flags); 5101 5102 return order < MAX_ORDER; 5103} 5104#endif 5105