page_alloc.c revision 945a11136ebdfa7fcce319ee6215958e84cb85f6
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 52#include <asm/tlbflush.h> 53#include <asm/div64.h> 54#include "internal.h" 55 56/* 57 * Array of node states. 58 */ 59nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 60 [N_POSSIBLE] = NODE_MASK_ALL, 61 [N_ONLINE] = { { [0] = 1UL } }, 62#ifndef CONFIG_NUMA 63 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 64#ifdef CONFIG_HIGHMEM 65 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 66#endif 67 [N_CPU] = { { [0] = 1UL } }, 68#endif /* NUMA */ 69}; 70EXPORT_SYMBOL(node_states); 71 72unsigned long totalram_pages __read_mostly; 73unsigned long totalreserve_pages __read_mostly; 74unsigned long highest_memmap_pfn __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 126unsigned long __meminitdata nr_kernel_pages; 127unsigned 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 /* 238 * Allow a burst of 60 reports, then keep quiet for that minute; 239 * or allow a steady drip of one report per second. 240 */ 241 if (nr_shown == 60) { 242 if (time_before(jiffies, resume)) { 243 nr_unshown++; 244 goto out; 245 } 246 if (nr_unshown) { 247 printk(KERN_ALERT 248 "BUG: Bad page state: %lu messages suppressed\n", 249 nr_unshown); 250 nr_unshown = 0; 251 } 252 nr_shown = 0; 253 } 254 if (nr_shown++ == 0) 255 resume = jiffies + 60 * HZ; 256 257 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 258 current->comm, page_to_pfn(page)); 259 printk(KERN_ALERT 260 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 261 page, (void *)page->flags, page_count(page), 262 page_mapcount(page), page->mapping, page->index); 263 264 dump_stack(); 265out: 266 /* Leave bad fields for debug, except PageBuddy could make trouble */ 267 __ClearPageBuddy(page); 268 add_taint(TAINT_BAD_PAGE); 269} 270 271/* 272 * Higher-order pages are called "compound pages". They are structured thusly: 273 * 274 * The first PAGE_SIZE page is called the "head page". 275 * 276 * The remaining PAGE_SIZE pages are called "tail pages". 277 * 278 * All pages have PG_compound set. All pages have their ->private pointing at 279 * the head page (even the head page has this). 280 * 281 * The first tail page's ->lru.next holds the address of the compound page's 282 * put_page() function. Its ->lru.prev holds the order of allocation. 283 * This usage means that zero-order pages may not be compound. 284 */ 285 286static void free_compound_page(struct page *page) 287{ 288 __free_pages_ok(page, compound_order(page)); 289} 290 291void prep_compound_page(struct page *page, unsigned long order) 292{ 293 int i; 294 int nr_pages = 1 << order; 295 296 set_compound_page_dtor(page, free_compound_page); 297 set_compound_order(page, order); 298 __SetPageHead(page); 299 for (i = 1; i < nr_pages; i++) { 300 struct page *p = page + i; 301 302 __SetPageTail(p); 303 p->first_page = page; 304 } 305} 306 307static int destroy_compound_page(struct page *page, unsigned long order) 308{ 309 int i; 310 int nr_pages = 1 << order; 311 int bad = 0; 312 313 if (unlikely(compound_order(page) != order) || 314 unlikely(!PageHead(page))) { 315 bad_page(page); 316 bad++; 317 } 318 319 __ClearPageHead(page); 320 321 for (i = 1; i < nr_pages; i++) { 322 struct page *p = page + i; 323 324 if (unlikely(!PageTail(p) || (p->first_page != page))) { 325 bad_page(page); 326 bad++; 327 } 328 __ClearPageTail(p); 329 } 330 331 return bad; 332} 333 334static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 335{ 336 int i; 337 338 /* 339 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 340 * and __GFP_HIGHMEM from hard or soft interrupt context. 341 */ 342 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 343 for (i = 0; i < (1 << order); i++) 344 clear_highpage(page + i); 345} 346 347static inline void set_page_order(struct page *page, int order) 348{ 349 set_page_private(page, order); 350 __SetPageBuddy(page); 351} 352 353static inline void rmv_page_order(struct page *page) 354{ 355 __ClearPageBuddy(page); 356 set_page_private(page, 0); 357} 358 359/* 360 * Locate the struct page for both the matching buddy in our 361 * pair (buddy1) and the combined O(n+1) page they form (page). 362 * 363 * 1) Any buddy B1 will have an order O twin B2 which satisfies 364 * the following equation: 365 * B2 = B1 ^ (1 << O) 366 * For example, if the starting buddy (buddy2) is #8 its order 367 * 1 buddy is #10: 368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 369 * 370 * 2) Any buddy B will have an order O+1 parent P which 371 * satisfies the following equation: 372 * P = B & ~(1 << O) 373 * 374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 375 */ 376static inline struct page * 377__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 378{ 379 unsigned long buddy_idx = page_idx ^ (1 << order); 380 381 return page + (buddy_idx - page_idx); 382} 383 384static inline unsigned long 385__find_combined_index(unsigned long page_idx, unsigned int order) 386{ 387 return (page_idx & ~(1 << order)); 388} 389 390/* 391 * This function checks whether a page is free && is the buddy 392 * we can do coalesce a page and its buddy if 393 * (a) the buddy is not in a hole && 394 * (b) the buddy is in the buddy system && 395 * (c) a page and its buddy have the same order && 396 * (d) a page and its buddy are in the same zone. 397 * 398 * For recording whether a page is in the buddy system, we use PG_buddy. 399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 400 * 401 * For recording page's order, we use page_private(page). 402 */ 403static inline int page_is_buddy(struct page *page, struct page *buddy, 404 int order) 405{ 406 if (!pfn_valid_within(page_to_pfn(buddy))) 407 return 0; 408 409 if (page_zone_id(page) != page_zone_id(buddy)) 410 return 0; 411 412 if (PageBuddy(buddy) && page_order(buddy) == order) { 413 VM_BUG_ON(page_count(buddy) != 0); 414 return 1; 415 } 416 return 0; 417} 418 419/* 420 * Freeing function for a buddy system allocator. 421 * 422 * The concept of a buddy system is to maintain direct-mapped table 423 * (containing bit values) for memory blocks of various "orders". 424 * The bottom level table contains the map for the smallest allocatable 425 * units of memory (here, pages), and each level above it describes 426 * pairs of units from the levels below, hence, "buddies". 427 * At a high level, all that happens here is marking the table entry 428 * at the bottom level available, and propagating the changes upward 429 * as necessary, plus some accounting needed to play nicely with other 430 * parts of the VM system. 431 * At each level, we keep a list of pages, which are heads of continuous 432 * free pages of length of (1 << order) and marked with PG_buddy. Page's 433 * order is recorded in page_private(page) field. 434 * So when we are allocating or freeing one, we can derive the state of the 435 * other. That is, if we allocate a small block, and both were 436 * free, the remainder of the region must be split into blocks. 437 * If a block is freed, and its buddy is also free, then this 438 * triggers coalescing into a block of larger size. 439 * 440 * -- wli 441 */ 442 443static inline void __free_one_page(struct page *page, 444 struct zone *zone, unsigned int order, 445 int migratetype) 446{ 447 unsigned long page_idx; 448 449 if (unlikely(PageCompound(page))) 450 if (unlikely(destroy_compound_page(page, order))) 451 return; 452 453 VM_BUG_ON(migratetype == -1); 454 455 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 456 457 VM_BUG_ON(page_idx & ((1 << order) - 1)); 458 VM_BUG_ON(bad_range(zone, page)); 459 460 while (order < MAX_ORDER-1) { 461 unsigned long combined_idx; 462 struct page *buddy; 463 464 buddy = __page_find_buddy(page, page_idx, order); 465 if (!page_is_buddy(page, buddy, order)) 466 break; 467 468 /* Our buddy is free, merge with it and move up one order. */ 469 list_del(&buddy->lru); 470 zone->free_area[order].nr_free--; 471 rmv_page_order(buddy); 472 combined_idx = __find_combined_index(page_idx, order); 473 page = page + (combined_idx - page_idx); 474 page_idx = combined_idx; 475 order++; 476 } 477 set_page_order(page, order); 478 list_add(&page->lru, 479 &zone->free_area[order].free_list[migratetype]); 480 zone->free_area[order].nr_free++; 481} 482 483#ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT 484/* 485 * free_page_mlock() -- clean up attempts to free and mlocked() page. 486 * Page should not be on lru, so no need to fix that up. 487 * free_pages_check() will verify... 488 */ 489static inline void free_page_mlock(struct page *page) 490{ 491 __dec_zone_page_state(page, NR_MLOCK); 492 __count_vm_event(UNEVICTABLE_MLOCKFREED); 493} 494#else 495static void free_page_mlock(struct page *page) { } 496#endif 497 498static inline int free_pages_check(struct page *page) 499{ 500 if (unlikely(page_mapcount(page) | 501 (page->mapping != NULL) | 502 (atomic_read(&page->_count) != 0) | 503 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 504 bad_page(page); 505 return 1; 506 } 507 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 508 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 509 return 0; 510} 511 512/* 513 * Frees a list of pages. 514 * Assumes all pages on list are in same zone, and of same order. 515 * count is the number of pages to free. 516 * 517 * If the zone was previously in an "all pages pinned" state then look to 518 * see if this freeing clears that state. 519 * 520 * And clear the zone's pages_scanned counter, to hold off the "all pages are 521 * pinned" detection logic. 522 */ 523static void free_pages_bulk(struct zone *zone, int count, 524 struct list_head *list, int order) 525{ 526 spin_lock(&zone->lock); 527 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 528 zone->pages_scanned = 0; 529 530 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order); 531 while (count--) { 532 struct page *page; 533 534 VM_BUG_ON(list_empty(list)); 535 page = list_entry(list->prev, struct page, lru); 536 /* have to delete it as __free_one_page list manipulates */ 537 list_del(&page->lru); 538 __free_one_page(page, zone, order, page_private(page)); 539 } 540 spin_unlock(&zone->lock); 541} 542 543static void free_one_page(struct zone *zone, struct page *page, int order, 544 int migratetype) 545{ 546 spin_lock(&zone->lock); 547 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 548 zone->pages_scanned = 0; 549 550 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 551 __free_one_page(page, zone, order, migratetype); 552 spin_unlock(&zone->lock); 553} 554 555static void __free_pages_ok(struct page *page, unsigned int order) 556{ 557 unsigned long flags; 558 int i; 559 int bad = 0; 560 int wasMlocked = TestClearPageMlocked(page); 561 562 kmemcheck_free_shadow(page, order); 563 564 for (i = 0 ; i < (1 << order) ; ++i) 565 bad += free_pages_check(page + i); 566 if (bad) 567 return; 568 569 if (!PageHighMem(page)) { 570 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 571 debug_check_no_obj_freed(page_address(page), 572 PAGE_SIZE << order); 573 } 574 arch_free_page(page, order); 575 kernel_map_pages(page, 1 << order, 0); 576 577 local_irq_save(flags); 578 if (unlikely(wasMlocked)) 579 free_page_mlock(page); 580 __count_vm_events(PGFREE, 1 << order); 581 free_one_page(page_zone(page), page, order, 582 get_pageblock_migratetype(page)); 583 local_irq_restore(flags); 584} 585 586/* 587 * permit the bootmem allocator to evade page validation on high-order frees 588 */ 589void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 590{ 591 if (order == 0) { 592 __ClearPageReserved(page); 593 set_page_count(page, 0); 594 set_page_refcounted(page); 595 __free_page(page); 596 } else { 597 int loop; 598 599 prefetchw(page); 600 for (loop = 0; loop < BITS_PER_LONG; loop++) { 601 struct page *p = &page[loop]; 602 603 if (loop + 1 < BITS_PER_LONG) 604 prefetchw(p + 1); 605 __ClearPageReserved(p); 606 set_page_count(p, 0); 607 } 608 609 set_page_refcounted(page); 610 __free_pages(page, order); 611 } 612} 613 614 615/* 616 * The order of subdivision here is critical for the IO subsystem. 617 * Please do not alter this order without good reasons and regression 618 * testing. Specifically, as large blocks of memory are subdivided, 619 * the order in which smaller blocks are delivered depends on the order 620 * they're subdivided in this function. This is the primary factor 621 * influencing the order in which pages are delivered to the IO 622 * subsystem according to empirical testing, and this is also justified 623 * by considering the behavior of a buddy system containing a single 624 * large block of memory acted on by a series of small allocations. 625 * This behavior is a critical factor in sglist merging's success. 626 * 627 * -- wli 628 */ 629static inline void expand(struct zone *zone, struct page *page, 630 int low, int high, struct free_area *area, 631 int migratetype) 632{ 633 unsigned long size = 1 << high; 634 635 while (high > low) { 636 area--; 637 high--; 638 size >>= 1; 639 VM_BUG_ON(bad_range(zone, &page[size])); 640 list_add(&page[size].lru, &area->free_list[migratetype]); 641 area->nr_free++; 642 set_page_order(&page[size], high); 643 } 644} 645 646/* 647 * This page is about to be returned from the page allocator 648 */ 649static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 650{ 651 if (unlikely(page_mapcount(page) | 652 (page->mapping != NULL) | 653 (atomic_read(&page->_count) != 0) | 654 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 655 bad_page(page); 656 return 1; 657 } 658 659 set_page_private(page, 0); 660 set_page_refcounted(page); 661 662 arch_alloc_page(page, order); 663 kernel_map_pages(page, 1 << order, 1); 664 665 if (gfp_flags & __GFP_ZERO) 666 prep_zero_page(page, order, gfp_flags); 667 668 if (order && (gfp_flags & __GFP_COMP)) 669 prep_compound_page(page, order); 670 671 return 0; 672} 673 674/* 675 * Go through the free lists for the given migratetype and remove 676 * the smallest available page from the freelists 677 */ 678static inline 679struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 680 int migratetype) 681{ 682 unsigned int current_order; 683 struct free_area * area; 684 struct page *page; 685 686 /* Find a page of the appropriate size in the preferred list */ 687 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 688 area = &(zone->free_area[current_order]); 689 if (list_empty(&area->free_list[migratetype])) 690 continue; 691 692 page = list_entry(area->free_list[migratetype].next, 693 struct page, lru); 694 list_del(&page->lru); 695 rmv_page_order(page); 696 area->nr_free--; 697 expand(zone, page, order, current_order, area, migratetype); 698 return page; 699 } 700 701 return NULL; 702} 703 704 705/* 706 * This array describes the order lists are fallen back to when 707 * the free lists for the desirable migrate type are depleted 708 */ 709static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 710 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 711 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 712 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 713 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 714}; 715 716/* 717 * Move the free pages in a range to the free lists of the requested type. 718 * Note that start_page and end_pages are not aligned on a pageblock 719 * boundary. If alignment is required, use move_freepages_block() 720 */ 721static int move_freepages(struct zone *zone, 722 struct page *start_page, struct page *end_page, 723 int migratetype) 724{ 725 struct page *page; 726 unsigned long order; 727 int pages_moved = 0; 728 729#ifndef CONFIG_HOLES_IN_ZONE 730 /* 731 * page_zone is not safe to call in this context when 732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 733 * anyway as we check zone boundaries in move_freepages_block(). 734 * Remove at a later date when no bug reports exist related to 735 * grouping pages by mobility 736 */ 737 BUG_ON(page_zone(start_page) != page_zone(end_page)); 738#endif 739 740 for (page = start_page; page <= end_page;) { 741 /* Make sure we are not inadvertently changing nodes */ 742 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 743 744 if (!pfn_valid_within(page_to_pfn(page))) { 745 page++; 746 continue; 747 } 748 749 if (!PageBuddy(page)) { 750 page++; 751 continue; 752 } 753 754 order = page_order(page); 755 list_del(&page->lru); 756 list_add(&page->lru, 757 &zone->free_area[order].free_list[migratetype]); 758 page += 1 << order; 759 pages_moved += 1 << order; 760 } 761 762 return pages_moved; 763} 764 765static int move_freepages_block(struct zone *zone, struct page *page, 766 int migratetype) 767{ 768 unsigned long start_pfn, end_pfn; 769 struct page *start_page, *end_page; 770 771 start_pfn = page_to_pfn(page); 772 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 773 start_page = pfn_to_page(start_pfn); 774 end_page = start_page + pageblock_nr_pages - 1; 775 end_pfn = start_pfn + pageblock_nr_pages - 1; 776 777 /* Do not cross zone boundaries */ 778 if (start_pfn < zone->zone_start_pfn) 779 start_page = page; 780 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 781 return 0; 782 783 return move_freepages(zone, start_page, end_page, migratetype); 784} 785 786/* Remove an element from the buddy allocator from the fallback list */ 787static inline struct page * 788__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 789{ 790 struct free_area * area; 791 int current_order; 792 struct page *page; 793 int migratetype, i; 794 795 /* Find the largest possible block of pages in the other list */ 796 for (current_order = MAX_ORDER-1; current_order >= order; 797 --current_order) { 798 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 799 migratetype = fallbacks[start_migratetype][i]; 800 801 /* MIGRATE_RESERVE handled later if necessary */ 802 if (migratetype == MIGRATE_RESERVE) 803 continue; 804 805 area = &(zone->free_area[current_order]); 806 if (list_empty(&area->free_list[migratetype])) 807 continue; 808 809 page = list_entry(area->free_list[migratetype].next, 810 struct page, lru); 811 area->nr_free--; 812 813 /* 814 * If breaking a large block of pages, move all free 815 * pages to the preferred allocation list. If falling 816 * back for a reclaimable kernel allocation, be more 817 * agressive about taking ownership of free pages 818 */ 819 if (unlikely(current_order >= (pageblock_order >> 1)) || 820 start_migratetype == MIGRATE_RECLAIMABLE || 821 page_group_by_mobility_disabled) { 822 unsigned long pages; 823 pages = move_freepages_block(zone, page, 824 start_migratetype); 825 826 /* Claim the whole block if over half of it is free */ 827 if (pages >= (1 << (pageblock_order-1)) || 828 page_group_by_mobility_disabled) 829 set_pageblock_migratetype(page, 830 start_migratetype); 831 832 migratetype = start_migratetype; 833 } 834 835 /* Remove the page from the freelists */ 836 list_del(&page->lru); 837 rmv_page_order(page); 838 839 if (current_order == pageblock_order) 840 set_pageblock_migratetype(page, 841 start_migratetype); 842 843 expand(zone, page, order, current_order, area, migratetype); 844 return page; 845 } 846 } 847 848 return NULL; 849} 850 851/* 852 * Do the hard work of removing an element from the buddy allocator. 853 * Call me with the zone->lock already held. 854 */ 855static struct page *__rmqueue(struct zone *zone, unsigned int order, 856 int migratetype) 857{ 858 struct page *page; 859 860retry_reserve: 861 page = __rmqueue_smallest(zone, order, migratetype); 862 863 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 864 page = __rmqueue_fallback(zone, order, migratetype); 865 866 /* 867 * Use MIGRATE_RESERVE rather than fail an allocation. goto 868 * is used because __rmqueue_smallest is an inline function 869 * and we want just one call site 870 */ 871 if (!page) { 872 migratetype = MIGRATE_RESERVE; 873 goto retry_reserve; 874 } 875 } 876 877 return page; 878} 879 880/* 881 * Obtain a specified number of elements from the buddy allocator, all under 882 * a single hold of the lock, for efficiency. Add them to the supplied list. 883 * Returns the number of new pages which were placed at *list. 884 */ 885static int rmqueue_bulk(struct zone *zone, unsigned int order, 886 unsigned long count, struct list_head *list, 887 int migratetype, int cold) 888{ 889 int i; 890 891 spin_lock(&zone->lock); 892 for (i = 0; i < count; ++i) { 893 struct page *page = __rmqueue(zone, order, migratetype); 894 if (unlikely(page == NULL)) 895 break; 896 897 /* 898 * Split buddy pages returned by expand() are received here 899 * in physical page order. The page is added to the callers and 900 * list and the list head then moves forward. From the callers 901 * perspective, the linked list is ordered by page number in 902 * some conditions. This is useful for IO devices that can 903 * merge IO requests if the physical pages are ordered 904 * properly. 905 */ 906 if (likely(cold == 0)) 907 list_add(&page->lru, list); 908 else 909 list_add_tail(&page->lru, list); 910 set_page_private(page, migratetype); 911 list = &page->lru; 912 } 913 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 914 spin_unlock(&zone->lock); 915 return i; 916} 917 918#ifdef CONFIG_NUMA 919/* 920 * Called from the vmstat counter updater to drain pagesets of this 921 * currently executing processor on remote nodes after they have 922 * expired. 923 * 924 * Note that this function must be called with the thread pinned to 925 * a single processor. 926 */ 927void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 928{ 929 unsigned long flags; 930 int to_drain; 931 932 local_irq_save(flags); 933 if (pcp->count >= pcp->batch) 934 to_drain = pcp->batch; 935 else 936 to_drain = pcp->count; 937 free_pages_bulk(zone, to_drain, &pcp->list, 0); 938 pcp->count -= to_drain; 939 local_irq_restore(flags); 940} 941#endif 942 943/* 944 * Drain pages of the indicated processor. 945 * 946 * The processor must either be the current processor and the 947 * thread pinned to the current processor or a processor that 948 * is not online. 949 */ 950static void drain_pages(unsigned int cpu) 951{ 952 unsigned long flags; 953 struct zone *zone; 954 955 for_each_populated_zone(zone) { 956 struct per_cpu_pageset *pset; 957 struct per_cpu_pages *pcp; 958 959 pset = zone_pcp(zone, cpu); 960 961 pcp = &pset->pcp; 962 local_irq_save(flags); 963 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 964 pcp->count = 0; 965 local_irq_restore(flags); 966 } 967} 968 969/* 970 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 971 */ 972void drain_local_pages(void *arg) 973{ 974 drain_pages(smp_processor_id()); 975} 976 977/* 978 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 979 */ 980void drain_all_pages(void) 981{ 982 on_each_cpu(drain_local_pages, NULL, 1); 983} 984 985#ifdef CONFIG_HIBERNATION 986 987void mark_free_pages(struct zone *zone) 988{ 989 unsigned long pfn, max_zone_pfn; 990 unsigned long flags; 991 int order, t; 992 struct list_head *curr; 993 994 if (!zone->spanned_pages) 995 return; 996 997 spin_lock_irqsave(&zone->lock, flags); 998 999 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1000 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1001 if (pfn_valid(pfn)) { 1002 struct page *page = pfn_to_page(pfn); 1003 1004 if (!swsusp_page_is_forbidden(page)) 1005 swsusp_unset_page_free(page); 1006 } 1007 1008 for_each_migratetype_order(order, t) { 1009 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1010 unsigned long i; 1011 1012 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1013 for (i = 0; i < (1UL << order); i++) 1014 swsusp_set_page_free(pfn_to_page(pfn + i)); 1015 } 1016 } 1017 spin_unlock_irqrestore(&zone->lock, flags); 1018} 1019#endif /* CONFIG_PM */ 1020 1021/* 1022 * Free a 0-order page 1023 */ 1024static void free_hot_cold_page(struct page *page, int cold) 1025{ 1026 struct zone *zone = page_zone(page); 1027 struct per_cpu_pages *pcp; 1028 unsigned long flags; 1029 int wasMlocked = TestClearPageMlocked(page); 1030 1031 kmemcheck_free_shadow(page, 0); 1032 1033 if (PageAnon(page)) 1034 page->mapping = NULL; 1035 if (free_pages_check(page)) 1036 return; 1037 1038 if (!PageHighMem(page)) { 1039 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1040 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1041 } 1042 arch_free_page(page, 0); 1043 kernel_map_pages(page, 1, 0); 1044 1045 pcp = &zone_pcp(zone, get_cpu())->pcp; 1046 set_page_private(page, get_pageblock_migratetype(page)); 1047 local_irq_save(flags); 1048 if (unlikely(wasMlocked)) 1049 free_page_mlock(page); 1050 __count_vm_event(PGFREE); 1051 1052 if (cold) 1053 list_add_tail(&page->lru, &pcp->list); 1054 else 1055 list_add(&page->lru, &pcp->list); 1056 pcp->count++; 1057 if (pcp->count >= pcp->high) { 1058 free_pages_bulk(zone, pcp->batch, &pcp->list, 0); 1059 pcp->count -= pcp->batch; 1060 } 1061 local_irq_restore(flags); 1062 put_cpu(); 1063} 1064 1065void free_hot_page(struct page *page) 1066{ 1067 free_hot_cold_page(page, 0); 1068} 1069 1070void free_cold_page(struct page *page) 1071{ 1072 free_hot_cold_page(page, 1); 1073} 1074 1075/* 1076 * split_page takes a non-compound higher-order page, and splits it into 1077 * n (1<<order) sub-pages: page[0..n] 1078 * Each sub-page must be freed individually. 1079 * 1080 * Note: this is probably too low level an operation for use in drivers. 1081 * Please consult with lkml before using this in your driver. 1082 */ 1083void split_page(struct page *page, unsigned int order) 1084{ 1085 int i; 1086 1087 VM_BUG_ON(PageCompound(page)); 1088 VM_BUG_ON(!page_count(page)); 1089 1090#ifdef CONFIG_KMEMCHECK 1091 /* 1092 * Split shadow pages too, because free(page[0]) would 1093 * otherwise free the whole shadow. 1094 */ 1095 if (kmemcheck_page_is_tracked(page)) 1096 split_page(virt_to_page(page[0].shadow), order); 1097#endif 1098 1099 for (i = 1; i < (1 << order); i++) 1100 set_page_refcounted(page + i); 1101} 1102 1103/* 1104 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1105 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1106 * or two. 1107 */ 1108static inline 1109struct page *buffered_rmqueue(struct zone *preferred_zone, 1110 struct zone *zone, int order, gfp_t gfp_flags, 1111 int migratetype) 1112{ 1113 unsigned long flags; 1114 struct page *page; 1115 int cold = !!(gfp_flags & __GFP_COLD); 1116 int cpu; 1117 1118again: 1119 cpu = get_cpu(); 1120 if (likely(order == 0)) { 1121 struct per_cpu_pages *pcp; 1122 1123 pcp = &zone_pcp(zone, cpu)->pcp; 1124 local_irq_save(flags); 1125 if (!pcp->count) { 1126 pcp->count = rmqueue_bulk(zone, 0, 1127 pcp->batch, &pcp->list, 1128 migratetype, cold); 1129 if (unlikely(!pcp->count)) 1130 goto failed; 1131 } 1132 1133 /* Find a page of the appropriate migrate type */ 1134 if (cold) { 1135 list_for_each_entry_reverse(page, &pcp->list, lru) 1136 if (page_private(page) == migratetype) 1137 break; 1138 } else { 1139 list_for_each_entry(page, &pcp->list, lru) 1140 if (page_private(page) == migratetype) 1141 break; 1142 } 1143 1144 /* Allocate more to the pcp list if necessary */ 1145 if (unlikely(&page->lru == &pcp->list)) { 1146 int get_one_page = 0; 1147 1148 pcp->count += rmqueue_bulk(zone, 0, 1149 pcp->batch, &pcp->list, 1150 migratetype, cold); 1151 list_for_each_entry(page, &pcp->list, lru) { 1152 if (get_pageblock_migratetype(page) != 1153 MIGRATE_ISOLATE) { 1154 get_one_page = 1; 1155 break; 1156 } 1157 } 1158 if (!get_one_page) 1159 goto failed; 1160 } 1161 1162 list_del(&page->lru); 1163 pcp->count--; 1164 } else { 1165 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1166 /* 1167 * __GFP_NOFAIL is not to be used in new code. 1168 * 1169 * All __GFP_NOFAIL callers should be fixed so that they 1170 * properly detect and handle allocation failures. 1171 * 1172 * We most definitely don't want callers attempting to 1173 * allocate greater than order-1 page units with 1174 * __GFP_NOFAIL. 1175 */ 1176 WARN_ON_ONCE(order > 1); 1177 } 1178 spin_lock_irqsave(&zone->lock, flags); 1179 page = __rmqueue(zone, order, migratetype); 1180 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1181 spin_unlock(&zone->lock); 1182 if (!page) 1183 goto failed; 1184 } 1185 1186 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1187 zone_statistics(preferred_zone, zone); 1188 local_irq_restore(flags); 1189 put_cpu(); 1190 1191 VM_BUG_ON(bad_range(zone, page)); 1192 if (prep_new_page(page, order, gfp_flags)) 1193 goto again; 1194 return page; 1195 1196failed: 1197 local_irq_restore(flags); 1198 put_cpu(); 1199 return NULL; 1200} 1201 1202/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1203#define ALLOC_WMARK_MIN WMARK_MIN 1204#define ALLOC_WMARK_LOW WMARK_LOW 1205#define ALLOC_WMARK_HIGH WMARK_HIGH 1206#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1207 1208/* Mask to get the watermark bits */ 1209#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1210 1211#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1212#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1213#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1214 1215#ifdef CONFIG_FAIL_PAGE_ALLOC 1216 1217static struct fail_page_alloc_attr { 1218 struct fault_attr attr; 1219 1220 u32 ignore_gfp_highmem; 1221 u32 ignore_gfp_wait; 1222 u32 min_order; 1223 1224#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1225 1226 struct dentry *ignore_gfp_highmem_file; 1227 struct dentry *ignore_gfp_wait_file; 1228 struct dentry *min_order_file; 1229 1230#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1231 1232} fail_page_alloc = { 1233 .attr = FAULT_ATTR_INITIALIZER, 1234 .ignore_gfp_wait = 1, 1235 .ignore_gfp_highmem = 1, 1236 .min_order = 1, 1237}; 1238 1239static int __init setup_fail_page_alloc(char *str) 1240{ 1241 return setup_fault_attr(&fail_page_alloc.attr, str); 1242} 1243__setup("fail_page_alloc=", setup_fail_page_alloc); 1244 1245static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1246{ 1247 if (order < fail_page_alloc.min_order) 1248 return 0; 1249 if (gfp_mask & __GFP_NOFAIL) 1250 return 0; 1251 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1252 return 0; 1253 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1254 return 0; 1255 1256 return should_fail(&fail_page_alloc.attr, 1 << order); 1257} 1258 1259#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1260 1261static int __init fail_page_alloc_debugfs(void) 1262{ 1263 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1264 struct dentry *dir; 1265 int err; 1266 1267 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1268 "fail_page_alloc"); 1269 if (err) 1270 return err; 1271 dir = fail_page_alloc.attr.dentries.dir; 1272 1273 fail_page_alloc.ignore_gfp_wait_file = 1274 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1275 &fail_page_alloc.ignore_gfp_wait); 1276 1277 fail_page_alloc.ignore_gfp_highmem_file = 1278 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1279 &fail_page_alloc.ignore_gfp_highmem); 1280 fail_page_alloc.min_order_file = 1281 debugfs_create_u32("min-order", mode, dir, 1282 &fail_page_alloc.min_order); 1283 1284 if (!fail_page_alloc.ignore_gfp_wait_file || 1285 !fail_page_alloc.ignore_gfp_highmem_file || 1286 !fail_page_alloc.min_order_file) { 1287 err = -ENOMEM; 1288 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1289 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1290 debugfs_remove(fail_page_alloc.min_order_file); 1291 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1292 } 1293 1294 return err; 1295} 1296 1297late_initcall(fail_page_alloc_debugfs); 1298 1299#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1300 1301#else /* CONFIG_FAIL_PAGE_ALLOC */ 1302 1303static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1304{ 1305 return 0; 1306} 1307 1308#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1309 1310/* 1311 * Return 1 if free pages are above 'mark'. This takes into account the order 1312 * of the allocation. 1313 */ 1314int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1315 int classzone_idx, int alloc_flags) 1316{ 1317 /* free_pages my go negative - that's OK */ 1318 long min = mark; 1319 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1320 int o; 1321 1322 if (alloc_flags & ALLOC_HIGH) 1323 min -= min / 2; 1324 if (alloc_flags & ALLOC_HARDER) 1325 min -= min / 4; 1326 1327 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1328 return 0; 1329 for (o = 0; o < order; o++) { 1330 /* At the next order, this order's pages become unavailable */ 1331 free_pages -= z->free_area[o].nr_free << o; 1332 1333 /* Require fewer higher order pages to be free */ 1334 min >>= 1; 1335 1336 if (free_pages <= min) 1337 return 0; 1338 } 1339 return 1; 1340} 1341 1342#ifdef CONFIG_NUMA 1343/* 1344 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1345 * skip over zones that are not allowed by the cpuset, or that have 1346 * been recently (in last second) found to be nearly full. See further 1347 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1348 * that have to skip over a lot of full or unallowed zones. 1349 * 1350 * If the zonelist cache is present in the passed in zonelist, then 1351 * returns a pointer to the allowed node mask (either the current 1352 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1353 * 1354 * If the zonelist cache is not available for this zonelist, does 1355 * nothing and returns NULL. 1356 * 1357 * If the fullzones BITMAP in the zonelist cache is stale (more than 1358 * a second since last zap'd) then we zap it out (clear its bits.) 1359 * 1360 * We hold off even calling zlc_setup, until after we've checked the 1361 * first zone in the zonelist, on the theory that most allocations will 1362 * be satisfied from that first zone, so best to examine that zone as 1363 * quickly as we can. 1364 */ 1365static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1366{ 1367 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1368 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1369 1370 zlc = zonelist->zlcache_ptr; 1371 if (!zlc) 1372 return NULL; 1373 1374 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1375 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1376 zlc->last_full_zap = jiffies; 1377 } 1378 1379 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1380 &cpuset_current_mems_allowed : 1381 &node_states[N_HIGH_MEMORY]; 1382 return allowednodes; 1383} 1384 1385/* 1386 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1387 * if it is worth looking at further for free memory: 1388 * 1) Check that the zone isn't thought to be full (doesn't have its 1389 * bit set in the zonelist_cache fullzones BITMAP). 1390 * 2) Check that the zones node (obtained from the zonelist_cache 1391 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1392 * Return true (non-zero) if zone is worth looking at further, or 1393 * else return false (zero) if it is not. 1394 * 1395 * This check -ignores- the distinction between various watermarks, 1396 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1397 * found to be full for any variation of these watermarks, it will 1398 * be considered full for up to one second by all requests, unless 1399 * we are so low on memory on all allowed nodes that we are forced 1400 * into the second scan of the zonelist. 1401 * 1402 * In the second scan we ignore this zonelist cache and exactly 1403 * apply the watermarks to all zones, even it is slower to do so. 1404 * We are low on memory in the second scan, and should leave no stone 1405 * unturned looking for a free page. 1406 */ 1407static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1408 nodemask_t *allowednodes) 1409{ 1410 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1411 int i; /* index of *z in zonelist zones */ 1412 int n; /* node that zone *z is on */ 1413 1414 zlc = zonelist->zlcache_ptr; 1415 if (!zlc) 1416 return 1; 1417 1418 i = z - zonelist->_zonerefs; 1419 n = zlc->z_to_n[i]; 1420 1421 /* This zone is worth trying if it is allowed but not full */ 1422 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1423} 1424 1425/* 1426 * Given 'z' scanning a zonelist, set the corresponding bit in 1427 * zlc->fullzones, so that subsequent attempts to allocate a page 1428 * from that zone don't waste time re-examining it. 1429 */ 1430static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1431{ 1432 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1433 int i; /* index of *z in zonelist zones */ 1434 1435 zlc = zonelist->zlcache_ptr; 1436 if (!zlc) 1437 return; 1438 1439 i = z - zonelist->_zonerefs; 1440 1441 set_bit(i, zlc->fullzones); 1442} 1443 1444#else /* CONFIG_NUMA */ 1445 1446static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1447{ 1448 return NULL; 1449} 1450 1451static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1452 nodemask_t *allowednodes) 1453{ 1454 return 1; 1455} 1456 1457static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1458{ 1459} 1460#endif /* CONFIG_NUMA */ 1461 1462/* 1463 * get_page_from_freelist goes through the zonelist trying to allocate 1464 * a page. 1465 */ 1466static struct page * 1467get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1468 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1469 struct zone *preferred_zone, int migratetype) 1470{ 1471 struct zoneref *z; 1472 struct page *page = NULL; 1473 int classzone_idx; 1474 struct zone *zone; 1475 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1476 int zlc_active = 0; /* set if using zonelist_cache */ 1477 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1478 1479 classzone_idx = zone_idx(preferred_zone); 1480zonelist_scan: 1481 /* 1482 * Scan zonelist, looking for a zone with enough free. 1483 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1484 */ 1485 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1486 high_zoneidx, nodemask) { 1487 if (NUMA_BUILD && zlc_active && 1488 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1489 continue; 1490 if ((alloc_flags & ALLOC_CPUSET) && 1491 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1492 goto try_next_zone; 1493 1494 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1495 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1496 unsigned long mark; 1497 int ret; 1498 1499 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1500 if (zone_watermark_ok(zone, order, mark, 1501 classzone_idx, alloc_flags)) 1502 goto try_this_zone; 1503 1504 if (zone_reclaim_mode == 0) 1505 goto this_zone_full; 1506 1507 ret = zone_reclaim(zone, gfp_mask, order); 1508 switch (ret) { 1509 case ZONE_RECLAIM_NOSCAN: 1510 /* did not scan */ 1511 goto try_next_zone; 1512 case ZONE_RECLAIM_FULL: 1513 /* scanned but unreclaimable */ 1514 goto this_zone_full; 1515 default: 1516 /* did we reclaim enough */ 1517 if (!zone_watermark_ok(zone, order, mark, 1518 classzone_idx, alloc_flags)) 1519 goto this_zone_full; 1520 } 1521 } 1522 1523try_this_zone: 1524 page = buffered_rmqueue(preferred_zone, zone, order, 1525 gfp_mask, migratetype); 1526 if (page) 1527 break; 1528this_zone_full: 1529 if (NUMA_BUILD) 1530 zlc_mark_zone_full(zonelist, z); 1531try_next_zone: 1532 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1533 /* 1534 * we do zlc_setup after the first zone is tried but only 1535 * if there are multiple nodes make it worthwhile 1536 */ 1537 allowednodes = zlc_setup(zonelist, alloc_flags); 1538 zlc_active = 1; 1539 did_zlc_setup = 1; 1540 } 1541 } 1542 1543 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1544 /* Disable zlc cache for second zonelist scan */ 1545 zlc_active = 0; 1546 goto zonelist_scan; 1547 } 1548 return page; 1549} 1550 1551static inline int 1552should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1553 unsigned long pages_reclaimed) 1554{ 1555 /* Do not loop if specifically requested */ 1556 if (gfp_mask & __GFP_NORETRY) 1557 return 0; 1558 1559 /* 1560 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1561 * means __GFP_NOFAIL, but that may not be true in other 1562 * implementations. 1563 */ 1564 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1565 return 1; 1566 1567 /* 1568 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1569 * specified, then we retry until we no longer reclaim any pages 1570 * (above), or we've reclaimed an order of pages at least as 1571 * large as the allocation's order. In both cases, if the 1572 * allocation still fails, we stop retrying. 1573 */ 1574 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1575 return 1; 1576 1577 /* 1578 * Don't let big-order allocations loop unless the caller 1579 * explicitly requests that. 1580 */ 1581 if (gfp_mask & __GFP_NOFAIL) 1582 return 1; 1583 1584 return 0; 1585} 1586 1587static inline struct page * 1588__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1589 struct zonelist *zonelist, enum zone_type high_zoneidx, 1590 nodemask_t *nodemask, struct zone *preferred_zone, 1591 int migratetype) 1592{ 1593 struct page *page; 1594 1595 /* Acquire the OOM killer lock for the zones in zonelist */ 1596 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1597 schedule_timeout_uninterruptible(1); 1598 return NULL; 1599 } 1600 1601 /* 1602 * Go through the zonelist yet one more time, keep very high watermark 1603 * here, this is only to catch a parallel oom killing, we must fail if 1604 * we're still under heavy pressure. 1605 */ 1606 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1607 order, zonelist, high_zoneidx, 1608 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1609 preferred_zone, migratetype); 1610 if (page) 1611 goto out; 1612 1613 /* The OOM killer will not help higher order allocs */ 1614 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL)) 1615 goto out; 1616 1617 /* Exhausted what can be done so it's blamo time */ 1618 out_of_memory(zonelist, gfp_mask, order); 1619 1620out: 1621 clear_zonelist_oom(zonelist, gfp_mask); 1622 return page; 1623} 1624 1625/* The really slow allocator path where we enter direct reclaim */ 1626static inline struct page * 1627__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1628 struct zonelist *zonelist, enum zone_type high_zoneidx, 1629 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1630 int migratetype, unsigned long *did_some_progress) 1631{ 1632 struct page *page = NULL; 1633 struct reclaim_state reclaim_state; 1634 struct task_struct *p = current; 1635 1636 cond_resched(); 1637 1638 /* We now go into synchronous reclaim */ 1639 cpuset_memory_pressure_bump(); 1640 p->flags |= PF_MEMALLOC; 1641 lockdep_set_current_reclaim_state(gfp_mask); 1642 reclaim_state.reclaimed_slab = 0; 1643 p->reclaim_state = &reclaim_state; 1644 1645 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1646 1647 p->reclaim_state = NULL; 1648 lockdep_clear_current_reclaim_state(); 1649 p->flags &= ~PF_MEMALLOC; 1650 1651 cond_resched(); 1652 1653 if (order != 0) 1654 drain_all_pages(); 1655 1656 if (likely(*did_some_progress)) 1657 page = get_page_from_freelist(gfp_mask, nodemask, order, 1658 zonelist, high_zoneidx, 1659 alloc_flags, preferred_zone, 1660 migratetype); 1661 return page; 1662} 1663 1664/* 1665 * This is called in the allocator slow-path if the allocation request is of 1666 * sufficient urgency to ignore watermarks and take other desperate measures 1667 */ 1668static inline struct page * 1669__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1670 struct zonelist *zonelist, enum zone_type high_zoneidx, 1671 nodemask_t *nodemask, struct zone *preferred_zone, 1672 int migratetype) 1673{ 1674 struct page *page; 1675 1676 do { 1677 page = get_page_from_freelist(gfp_mask, nodemask, order, 1678 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1679 preferred_zone, migratetype); 1680 1681 if (!page && gfp_mask & __GFP_NOFAIL) 1682 congestion_wait(BLK_RW_ASYNC, HZ/50); 1683 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1684 1685 return page; 1686} 1687 1688static inline 1689void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1690 enum zone_type high_zoneidx) 1691{ 1692 struct zoneref *z; 1693 struct zone *zone; 1694 1695 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1696 wakeup_kswapd(zone, order); 1697} 1698 1699static inline int 1700gfp_to_alloc_flags(gfp_t gfp_mask) 1701{ 1702 struct task_struct *p = current; 1703 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1704 const gfp_t wait = gfp_mask & __GFP_WAIT; 1705 1706 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1707 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1708 1709 /* 1710 * The caller may dip into page reserves a bit more if the caller 1711 * cannot run direct reclaim, or if the caller has realtime scheduling 1712 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1713 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1714 */ 1715 alloc_flags |= (gfp_mask & __GFP_HIGH); 1716 1717 if (!wait) { 1718 alloc_flags |= ALLOC_HARDER; 1719 /* 1720 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1721 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1722 */ 1723 alloc_flags &= ~ALLOC_CPUSET; 1724 } else if (unlikely(rt_task(p))) 1725 alloc_flags |= ALLOC_HARDER; 1726 1727 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1728 if (!in_interrupt() && 1729 ((p->flags & PF_MEMALLOC) || 1730 unlikely(test_thread_flag(TIF_MEMDIE)))) 1731 alloc_flags |= ALLOC_NO_WATERMARKS; 1732 } 1733 1734 return alloc_flags; 1735} 1736 1737static inline struct page * 1738__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1739 struct zonelist *zonelist, enum zone_type high_zoneidx, 1740 nodemask_t *nodemask, struct zone *preferred_zone, 1741 int migratetype) 1742{ 1743 const gfp_t wait = gfp_mask & __GFP_WAIT; 1744 struct page *page = NULL; 1745 int alloc_flags; 1746 unsigned long pages_reclaimed = 0; 1747 unsigned long did_some_progress; 1748 struct task_struct *p = current; 1749 1750 /* 1751 * In the slowpath, we sanity check order to avoid ever trying to 1752 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1753 * be using allocators in order of preference for an area that is 1754 * too large. 1755 */ 1756 if (order >= MAX_ORDER) { 1757 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1758 return NULL; 1759 } 1760 1761 /* 1762 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1763 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1764 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1765 * using a larger set of nodes after it has established that the 1766 * allowed per node queues are empty and that nodes are 1767 * over allocated. 1768 */ 1769 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1770 goto nopage; 1771 1772 wake_all_kswapd(order, zonelist, high_zoneidx); 1773 1774restart: 1775 /* 1776 * OK, we're below the kswapd watermark and have kicked background 1777 * reclaim. Now things get more complex, so set up alloc_flags according 1778 * to how we want to proceed. 1779 */ 1780 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1781 1782 /* This is the last chance, in general, before the goto nopage. */ 1783 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1784 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1785 preferred_zone, migratetype); 1786 if (page) 1787 goto got_pg; 1788 1789rebalance: 1790 /* Allocate without watermarks if the context allows */ 1791 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1792 page = __alloc_pages_high_priority(gfp_mask, order, 1793 zonelist, high_zoneidx, nodemask, 1794 preferred_zone, migratetype); 1795 if (page) 1796 goto got_pg; 1797 } 1798 1799 /* Atomic allocations - we can't balance anything */ 1800 if (!wait) 1801 goto nopage; 1802 1803 /* Avoid recursion of direct reclaim */ 1804 if (p->flags & PF_MEMALLOC) 1805 goto nopage; 1806 1807 /* Avoid allocations with no watermarks from looping endlessly */ 1808 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 1809 goto nopage; 1810 1811 /* Try direct reclaim and then allocating */ 1812 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1813 zonelist, high_zoneidx, 1814 nodemask, 1815 alloc_flags, preferred_zone, 1816 migratetype, &did_some_progress); 1817 if (page) 1818 goto got_pg; 1819 1820 /* 1821 * If we failed to make any progress reclaiming, then we are 1822 * running out of options and have to consider going OOM 1823 */ 1824 if (!did_some_progress) { 1825 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1826 if (oom_killer_disabled) 1827 goto nopage; 1828 page = __alloc_pages_may_oom(gfp_mask, order, 1829 zonelist, high_zoneidx, 1830 nodemask, preferred_zone, 1831 migratetype); 1832 if (page) 1833 goto got_pg; 1834 1835 /* 1836 * The OOM killer does not trigger for high-order 1837 * ~__GFP_NOFAIL allocations so if no progress is being 1838 * made, there are no other options and retrying is 1839 * unlikely to help. 1840 */ 1841 if (order > PAGE_ALLOC_COSTLY_ORDER && 1842 !(gfp_mask & __GFP_NOFAIL)) 1843 goto nopage; 1844 1845 goto restart; 1846 } 1847 } 1848 1849 /* Check if we should retry the allocation */ 1850 pages_reclaimed += did_some_progress; 1851 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1852 /* Wait for some write requests to complete then retry */ 1853 congestion_wait(BLK_RW_ASYNC, HZ/50); 1854 goto rebalance; 1855 } 1856 1857nopage: 1858 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1859 printk(KERN_WARNING "%s: page allocation failure." 1860 " order:%d, mode:0x%x\n", 1861 p->comm, order, gfp_mask); 1862 dump_stack(); 1863 show_mem(); 1864 } 1865 return page; 1866got_pg: 1867 if (kmemcheck_enabled) 1868 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 1869 return page; 1870 1871} 1872 1873/* 1874 * This is the 'heart' of the zoned buddy allocator. 1875 */ 1876struct page * 1877__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1878 struct zonelist *zonelist, nodemask_t *nodemask) 1879{ 1880 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1881 struct zone *preferred_zone; 1882 struct page *page; 1883 int migratetype = allocflags_to_migratetype(gfp_mask); 1884 1885 gfp_mask &= gfp_allowed_mask; 1886 1887 lockdep_trace_alloc(gfp_mask); 1888 1889 might_sleep_if(gfp_mask & __GFP_WAIT); 1890 1891 if (should_fail_alloc_page(gfp_mask, order)) 1892 return NULL; 1893 1894 /* 1895 * Check the zones suitable for the gfp_mask contain at least one 1896 * valid zone. It's possible to have an empty zonelist as a result 1897 * of GFP_THISNODE and a memoryless node 1898 */ 1899 if (unlikely(!zonelist->_zonerefs->zone)) 1900 return NULL; 1901 1902 /* The preferred zone is used for statistics later */ 1903 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1904 if (!preferred_zone) 1905 return NULL; 1906 1907 /* First allocation attempt */ 1908 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1909 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1910 preferred_zone, migratetype); 1911 if (unlikely(!page)) 1912 page = __alloc_pages_slowpath(gfp_mask, order, 1913 zonelist, high_zoneidx, nodemask, 1914 preferred_zone, migratetype); 1915 1916 return page; 1917} 1918EXPORT_SYMBOL(__alloc_pages_nodemask); 1919 1920/* 1921 * Common helper functions. 1922 */ 1923unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1924{ 1925 struct page *page; 1926 1927 /* 1928 * __get_free_pages() returns a 32-bit address, which cannot represent 1929 * a highmem page 1930 */ 1931 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1932 1933 page = alloc_pages(gfp_mask, order); 1934 if (!page) 1935 return 0; 1936 return (unsigned long) page_address(page); 1937} 1938EXPORT_SYMBOL(__get_free_pages); 1939 1940unsigned long get_zeroed_page(gfp_t gfp_mask) 1941{ 1942 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 1943} 1944EXPORT_SYMBOL(get_zeroed_page); 1945 1946void __pagevec_free(struct pagevec *pvec) 1947{ 1948 int i = pagevec_count(pvec); 1949 1950 while (--i >= 0) 1951 free_hot_cold_page(pvec->pages[i], pvec->cold); 1952} 1953 1954void __free_pages(struct page *page, unsigned int order) 1955{ 1956 if (put_page_testzero(page)) { 1957 if (order == 0) 1958 free_hot_page(page); 1959 else 1960 __free_pages_ok(page, order); 1961 } 1962} 1963 1964EXPORT_SYMBOL(__free_pages); 1965 1966void free_pages(unsigned long addr, unsigned int order) 1967{ 1968 if (addr != 0) { 1969 VM_BUG_ON(!virt_addr_valid((void *)addr)); 1970 __free_pages(virt_to_page((void *)addr), order); 1971 } 1972} 1973 1974EXPORT_SYMBOL(free_pages); 1975 1976/** 1977 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 1978 * @size: the number of bytes to allocate 1979 * @gfp_mask: GFP flags for the allocation 1980 * 1981 * This function is similar to alloc_pages(), except that it allocates the 1982 * minimum number of pages to satisfy the request. alloc_pages() can only 1983 * allocate memory in power-of-two pages. 1984 * 1985 * This function is also limited by MAX_ORDER. 1986 * 1987 * Memory allocated by this function must be released by free_pages_exact(). 1988 */ 1989void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 1990{ 1991 unsigned int order = get_order(size); 1992 unsigned long addr; 1993 1994 addr = __get_free_pages(gfp_mask, order); 1995 if (addr) { 1996 unsigned long alloc_end = addr + (PAGE_SIZE << order); 1997 unsigned long used = addr + PAGE_ALIGN(size); 1998 1999 split_page(virt_to_page((void *)addr), order); 2000 while (used < alloc_end) { 2001 free_page(used); 2002 used += PAGE_SIZE; 2003 } 2004 } 2005 2006 return (void *)addr; 2007} 2008EXPORT_SYMBOL(alloc_pages_exact); 2009 2010/** 2011 * free_pages_exact - release memory allocated via alloc_pages_exact() 2012 * @virt: the value returned by alloc_pages_exact. 2013 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2014 * 2015 * Release the memory allocated by a previous call to alloc_pages_exact. 2016 */ 2017void free_pages_exact(void *virt, size_t size) 2018{ 2019 unsigned long addr = (unsigned long)virt; 2020 unsigned long end = addr + PAGE_ALIGN(size); 2021 2022 while (addr < end) { 2023 free_page(addr); 2024 addr += PAGE_SIZE; 2025 } 2026} 2027EXPORT_SYMBOL(free_pages_exact); 2028 2029static unsigned int nr_free_zone_pages(int offset) 2030{ 2031 struct zoneref *z; 2032 struct zone *zone; 2033 2034 /* Just pick one node, since fallback list is circular */ 2035 unsigned int sum = 0; 2036 2037 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2038 2039 for_each_zone_zonelist(zone, z, zonelist, offset) { 2040 unsigned long size = zone->present_pages; 2041 unsigned long high = high_wmark_pages(zone); 2042 if (size > high) 2043 sum += size - high; 2044 } 2045 2046 return sum; 2047} 2048 2049/* 2050 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2051 */ 2052unsigned int nr_free_buffer_pages(void) 2053{ 2054 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2055} 2056EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2057 2058/* 2059 * Amount of free RAM allocatable within all zones 2060 */ 2061unsigned int nr_free_pagecache_pages(void) 2062{ 2063 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2064} 2065 2066static inline void show_node(struct zone *zone) 2067{ 2068 if (NUMA_BUILD) 2069 printk("Node %d ", zone_to_nid(zone)); 2070} 2071 2072void si_meminfo(struct sysinfo *val) 2073{ 2074 val->totalram = totalram_pages; 2075 val->sharedram = 0; 2076 val->freeram = global_page_state(NR_FREE_PAGES); 2077 val->bufferram = nr_blockdev_pages(); 2078 val->totalhigh = totalhigh_pages; 2079 val->freehigh = nr_free_highpages(); 2080 val->mem_unit = PAGE_SIZE; 2081} 2082 2083EXPORT_SYMBOL(si_meminfo); 2084 2085#ifdef CONFIG_NUMA 2086void si_meminfo_node(struct sysinfo *val, int nid) 2087{ 2088 pg_data_t *pgdat = NODE_DATA(nid); 2089 2090 val->totalram = pgdat->node_present_pages; 2091 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2092#ifdef CONFIG_HIGHMEM 2093 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2094 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2095 NR_FREE_PAGES); 2096#else 2097 val->totalhigh = 0; 2098 val->freehigh = 0; 2099#endif 2100 val->mem_unit = PAGE_SIZE; 2101} 2102#endif 2103 2104#define K(x) ((x) << (PAGE_SHIFT-10)) 2105 2106/* 2107 * Show free area list (used inside shift_scroll-lock stuff) 2108 * We also calculate the percentage fragmentation. We do this by counting the 2109 * memory on each free list with the exception of the first item on the list. 2110 */ 2111void show_free_areas(void) 2112{ 2113 int cpu; 2114 struct zone *zone; 2115 2116 for_each_populated_zone(zone) { 2117 show_node(zone); 2118 printk("%s per-cpu:\n", zone->name); 2119 2120 for_each_online_cpu(cpu) { 2121 struct per_cpu_pageset *pageset; 2122 2123 pageset = zone_pcp(zone, cpu); 2124 2125 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2126 cpu, pageset->pcp.high, 2127 pageset->pcp.batch, pageset->pcp.count); 2128 } 2129 } 2130 2131 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2132 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2133 " unevictable:%lu" 2134 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n" 2135 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2136 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2137 global_page_state(NR_ACTIVE_ANON), 2138 global_page_state(NR_INACTIVE_ANON), 2139 global_page_state(NR_ISOLATED_ANON), 2140 global_page_state(NR_ACTIVE_FILE), 2141 global_page_state(NR_INACTIVE_FILE), 2142 global_page_state(NR_ISOLATED_FILE), 2143 global_page_state(NR_UNEVICTABLE), 2144 global_page_state(NR_FILE_DIRTY), 2145 global_page_state(NR_WRITEBACK), 2146 global_page_state(NR_UNSTABLE_NFS), 2147 nr_blockdev_pages(), 2148 global_page_state(NR_FREE_PAGES), 2149 global_page_state(NR_SLAB_RECLAIMABLE), 2150 global_page_state(NR_SLAB_UNRECLAIMABLE), 2151 global_page_state(NR_FILE_MAPPED), 2152 global_page_state(NR_SHMEM), 2153 global_page_state(NR_PAGETABLE), 2154 global_page_state(NR_BOUNCE)); 2155 2156 for_each_populated_zone(zone) { 2157 int i; 2158 2159 show_node(zone); 2160 printk("%s" 2161 " free:%lukB" 2162 " min:%lukB" 2163 " low:%lukB" 2164 " high:%lukB" 2165 " active_anon:%lukB" 2166 " inactive_anon:%lukB" 2167 " active_file:%lukB" 2168 " inactive_file:%lukB" 2169 " unevictable:%lukB" 2170 " isolated(anon):%lukB" 2171 " isolated(file):%lukB" 2172 " present:%lukB" 2173 " mlocked:%lukB" 2174 " dirty:%lukB" 2175 " writeback:%lukB" 2176 " mapped:%lukB" 2177 " shmem:%lukB" 2178 " slab_reclaimable:%lukB" 2179 " slab_unreclaimable:%lukB" 2180 " kernel_stack:%lukB" 2181 " pagetables:%lukB" 2182 " unstable:%lukB" 2183 " bounce:%lukB" 2184 " writeback_tmp:%lukB" 2185 " pages_scanned:%lu" 2186 " all_unreclaimable? %s" 2187 "\n", 2188 zone->name, 2189 K(zone_page_state(zone, NR_FREE_PAGES)), 2190 K(min_wmark_pages(zone)), 2191 K(low_wmark_pages(zone)), 2192 K(high_wmark_pages(zone)), 2193 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2194 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2195 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2196 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2197 K(zone_page_state(zone, NR_UNEVICTABLE)), 2198 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2199 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2200 K(zone->present_pages), 2201 K(zone_page_state(zone, NR_MLOCK)), 2202 K(zone_page_state(zone, NR_FILE_DIRTY)), 2203 K(zone_page_state(zone, NR_WRITEBACK)), 2204 K(zone_page_state(zone, NR_FILE_MAPPED)), 2205 K(zone_page_state(zone, NR_SHMEM)), 2206 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2207 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2208 zone_page_state(zone, NR_KERNEL_STACK) * 2209 THREAD_SIZE / 1024, 2210 K(zone_page_state(zone, NR_PAGETABLE)), 2211 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2212 K(zone_page_state(zone, NR_BOUNCE)), 2213 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2214 zone->pages_scanned, 2215 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 2216 ); 2217 printk("lowmem_reserve[]:"); 2218 for (i = 0; i < MAX_NR_ZONES; i++) 2219 printk(" %lu", zone->lowmem_reserve[i]); 2220 printk("\n"); 2221 } 2222 2223 for_each_populated_zone(zone) { 2224 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2225 2226 show_node(zone); 2227 printk("%s: ", zone->name); 2228 2229 spin_lock_irqsave(&zone->lock, flags); 2230 for (order = 0; order < MAX_ORDER; order++) { 2231 nr[order] = zone->free_area[order].nr_free; 2232 total += nr[order] << order; 2233 } 2234 spin_unlock_irqrestore(&zone->lock, flags); 2235 for (order = 0; order < MAX_ORDER; order++) 2236 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2237 printk("= %lukB\n", K(total)); 2238 } 2239 2240 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2241 2242 show_swap_cache_info(); 2243} 2244 2245static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2246{ 2247 zoneref->zone = zone; 2248 zoneref->zone_idx = zone_idx(zone); 2249} 2250 2251/* 2252 * Builds allocation fallback zone lists. 2253 * 2254 * Add all populated zones of a node to the zonelist. 2255 */ 2256static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2257 int nr_zones, enum zone_type zone_type) 2258{ 2259 struct zone *zone; 2260 2261 BUG_ON(zone_type >= MAX_NR_ZONES); 2262 zone_type++; 2263 2264 do { 2265 zone_type--; 2266 zone = pgdat->node_zones + zone_type; 2267 if (populated_zone(zone)) { 2268 zoneref_set_zone(zone, 2269 &zonelist->_zonerefs[nr_zones++]); 2270 check_highest_zone(zone_type); 2271 } 2272 2273 } while (zone_type); 2274 return nr_zones; 2275} 2276 2277 2278/* 2279 * zonelist_order: 2280 * 0 = automatic detection of better ordering. 2281 * 1 = order by ([node] distance, -zonetype) 2282 * 2 = order by (-zonetype, [node] distance) 2283 * 2284 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2285 * the same zonelist. So only NUMA can configure this param. 2286 */ 2287#define ZONELIST_ORDER_DEFAULT 0 2288#define ZONELIST_ORDER_NODE 1 2289#define ZONELIST_ORDER_ZONE 2 2290 2291/* zonelist order in the kernel. 2292 * set_zonelist_order() will set this to NODE or ZONE. 2293 */ 2294static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2295static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2296 2297 2298#ifdef CONFIG_NUMA 2299/* The value user specified ....changed by config */ 2300static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2301/* string for sysctl */ 2302#define NUMA_ZONELIST_ORDER_LEN 16 2303char numa_zonelist_order[16] = "default"; 2304 2305/* 2306 * interface for configure zonelist ordering. 2307 * command line option "numa_zonelist_order" 2308 * = "[dD]efault - default, automatic configuration. 2309 * = "[nN]ode - order by node locality, then by zone within node 2310 * = "[zZ]one - order by zone, then by locality within zone 2311 */ 2312 2313static int __parse_numa_zonelist_order(char *s) 2314{ 2315 if (*s == 'd' || *s == 'D') { 2316 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2317 } else if (*s == 'n' || *s == 'N') { 2318 user_zonelist_order = ZONELIST_ORDER_NODE; 2319 } else if (*s == 'z' || *s == 'Z') { 2320 user_zonelist_order = ZONELIST_ORDER_ZONE; 2321 } else { 2322 printk(KERN_WARNING 2323 "Ignoring invalid numa_zonelist_order value: " 2324 "%s\n", s); 2325 return -EINVAL; 2326 } 2327 return 0; 2328} 2329 2330static __init int setup_numa_zonelist_order(char *s) 2331{ 2332 if (s) 2333 return __parse_numa_zonelist_order(s); 2334 return 0; 2335} 2336early_param("numa_zonelist_order", setup_numa_zonelist_order); 2337 2338/* 2339 * sysctl handler for numa_zonelist_order 2340 */ 2341int numa_zonelist_order_handler(ctl_table *table, int write, 2342 struct file *file, void __user *buffer, size_t *length, 2343 loff_t *ppos) 2344{ 2345 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2346 int ret; 2347 2348 if (write) 2349 strncpy(saved_string, (char*)table->data, 2350 NUMA_ZONELIST_ORDER_LEN); 2351 ret = proc_dostring(table, write, file, buffer, length, ppos); 2352 if (ret) 2353 return ret; 2354 if (write) { 2355 int oldval = user_zonelist_order; 2356 if (__parse_numa_zonelist_order((char*)table->data)) { 2357 /* 2358 * bogus value. restore saved string 2359 */ 2360 strncpy((char*)table->data, saved_string, 2361 NUMA_ZONELIST_ORDER_LEN); 2362 user_zonelist_order = oldval; 2363 } else if (oldval != user_zonelist_order) 2364 build_all_zonelists(); 2365 } 2366 return 0; 2367} 2368 2369 2370#define MAX_NODE_LOAD (nr_online_nodes) 2371static int node_load[MAX_NUMNODES]; 2372 2373/** 2374 * find_next_best_node - find the next node that should appear in a given node's fallback list 2375 * @node: node whose fallback list we're appending 2376 * @used_node_mask: nodemask_t of already used nodes 2377 * 2378 * We use a number of factors to determine which is the next node that should 2379 * appear on a given node's fallback list. The node should not have appeared 2380 * already in @node's fallback list, and it should be the next closest node 2381 * according to the distance array (which contains arbitrary distance values 2382 * from each node to each node in the system), and should also prefer nodes 2383 * with no CPUs, since presumably they'll have very little allocation pressure 2384 * on them otherwise. 2385 * It returns -1 if no node is found. 2386 */ 2387static int find_next_best_node(int node, nodemask_t *used_node_mask) 2388{ 2389 int n, val; 2390 int min_val = INT_MAX; 2391 int best_node = -1; 2392 const struct cpumask *tmp = cpumask_of_node(0); 2393 2394 /* Use the local node if we haven't already */ 2395 if (!node_isset(node, *used_node_mask)) { 2396 node_set(node, *used_node_mask); 2397 return node; 2398 } 2399 2400 for_each_node_state(n, N_HIGH_MEMORY) { 2401 2402 /* Don't want a node to appear more than once */ 2403 if (node_isset(n, *used_node_mask)) 2404 continue; 2405 2406 /* Use the distance array to find the distance */ 2407 val = node_distance(node, n); 2408 2409 /* Penalize nodes under us ("prefer the next node") */ 2410 val += (n < node); 2411 2412 /* Give preference to headless and unused nodes */ 2413 tmp = cpumask_of_node(n); 2414 if (!cpumask_empty(tmp)) 2415 val += PENALTY_FOR_NODE_WITH_CPUS; 2416 2417 /* Slight preference for less loaded node */ 2418 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2419 val += node_load[n]; 2420 2421 if (val < min_val) { 2422 min_val = val; 2423 best_node = n; 2424 } 2425 } 2426 2427 if (best_node >= 0) 2428 node_set(best_node, *used_node_mask); 2429 2430 return best_node; 2431} 2432 2433 2434/* 2435 * Build zonelists ordered by node and zones within node. 2436 * This results in maximum locality--normal zone overflows into local 2437 * DMA zone, if any--but risks exhausting DMA zone. 2438 */ 2439static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2440{ 2441 int j; 2442 struct zonelist *zonelist; 2443 2444 zonelist = &pgdat->node_zonelists[0]; 2445 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2446 ; 2447 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2448 MAX_NR_ZONES - 1); 2449 zonelist->_zonerefs[j].zone = NULL; 2450 zonelist->_zonerefs[j].zone_idx = 0; 2451} 2452 2453/* 2454 * Build gfp_thisnode zonelists 2455 */ 2456static void build_thisnode_zonelists(pg_data_t *pgdat) 2457{ 2458 int j; 2459 struct zonelist *zonelist; 2460 2461 zonelist = &pgdat->node_zonelists[1]; 2462 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2463 zonelist->_zonerefs[j].zone = NULL; 2464 zonelist->_zonerefs[j].zone_idx = 0; 2465} 2466 2467/* 2468 * Build zonelists ordered by zone and nodes within zones. 2469 * This results in conserving DMA zone[s] until all Normal memory is 2470 * exhausted, but results in overflowing to remote node while memory 2471 * may still exist in local DMA zone. 2472 */ 2473static int node_order[MAX_NUMNODES]; 2474 2475static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2476{ 2477 int pos, j, node; 2478 int zone_type; /* needs to be signed */ 2479 struct zone *z; 2480 struct zonelist *zonelist; 2481 2482 zonelist = &pgdat->node_zonelists[0]; 2483 pos = 0; 2484 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2485 for (j = 0; j < nr_nodes; j++) { 2486 node = node_order[j]; 2487 z = &NODE_DATA(node)->node_zones[zone_type]; 2488 if (populated_zone(z)) { 2489 zoneref_set_zone(z, 2490 &zonelist->_zonerefs[pos++]); 2491 check_highest_zone(zone_type); 2492 } 2493 } 2494 } 2495 zonelist->_zonerefs[pos].zone = NULL; 2496 zonelist->_zonerefs[pos].zone_idx = 0; 2497} 2498 2499static int default_zonelist_order(void) 2500{ 2501 int nid, zone_type; 2502 unsigned long low_kmem_size,total_size; 2503 struct zone *z; 2504 int average_size; 2505 /* 2506 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2507 * If they are really small and used heavily, the system can fall 2508 * into OOM very easily. 2509 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2510 */ 2511 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2512 low_kmem_size = 0; 2513 total_size = 0; 2514 for_each_online_node(nid) { 2515 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2516 z = &NODE_DATA(nid)->node_zones[zone_type]; 2517 if (populated_zone(z)) { 2518 if (zone_type < ZONE_NORMAL) 2519 low_kmem_size += z->present_pages; 2520 total_size += z->present_pages; 2521 } 2522 } 2523 } 2524 if (!low_kmem_size || /* there are no DMA area. */ 2525 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2526 return ZONELIST_ORDER_NODE; 2527 /* 2528 * look into each node's config. 2529 * If there is a node whose DMA/DMA32 memory is very big area on 2530 * local memory, NODE_ORDER may be suitable. 2531 */ 2532 average_size = total_size / 2533 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2534 for_each_online_node(nid) { 2535 low_kmem_size = 0; 2536 total_size = 0; 2537 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2538 z = &NODE_DATA(nid)->node_zones[zone_type]; 2539 if (populated_zone(z)) { 2540 if (zone_type < ZONE_NORMAL) 2541 low_kmem_size += z->present_pages; 2542 total_size += z->present_pages; 2543 } 2544 } 2545 if (low_kmem_size && 2546 total_size > average_size && /* ignore small node */ 2547 low_kmem_size > total_size * 70/100) 2548 return ZONELIST_ORDER_NODE; 2549 } 2550 return ZONELIST_ORDER_ZONE; 2551} 2552 2553static void set_zonelist_order(void) 2554{ 2555 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2556 current_zonelist_order = default_zonelist_order(); 2557 else 2558 current_zonelist_order = user_zonelist_order; 2559} 2560 2561static void build_zonelists(pg_data_t *pgdat) 2562{ 2563 int j, node, load; 2564 enum zone_type i; 2565 nodemask_t used_mask; 2566 int local_node, prev_node; 2567 struct zonelist *zonelist; 2568 int order = current_zonelist_order; 2569 2570 /* initialize zonelists */ 2571 for (i = 0; i < MAX_ZONELISTS; i++) { 2572 zonelist = pgdat->node_zonelists + i; 2573 zonelist->_zonerefs[0].zone = NULL; 2574 zonelist->_zonerefs[0].zone_idx = 0; 2575 } 2576 2577 /* NUMA-aware ordering of nodes */ 2578 local_node = pgdat->node_id; 2579 load = nr_online_nodes; 2580 prev_node = local_node; 2581 nodes_clear(used_mask); 2582 2583 memset(node_order, 0, sizeof(node_order)); 2584 j = 0; 2585 2586 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2587 int distance = node_distance(local_node, node); 2588 2589 /* 2590 * If another node is sufficiently far away then it is better 2591 * to reclaim pages in a zone before going off node. 2592 */ 2593 if (distance > RECLAIM_DISTANCE) 2594 zone_reclaim_mode = 1; 2595 2596 /* 2597 * We don't want to pressure a particular node. 2598 * So adding penalty to the first node in same 2599 * distance group to make it round-robin. 2600 */ 2601 if (distance != node_distance(local_node, prev_node)) 2602 node_load[node] = load; 2603 2604 prev_node = node; 2605 load--; 2606 if (order == ZONELIST_ORDER_NODE) 2607 build_zonelists_in_node_order(pgdat, node); 2608 else 2609 node_order[j++] = node; /* remember order */ 2610 } 2611 2612 if (order == ZONELIST_ORDER_ZONE) { 2613 /* calculate node order -- i.e., DMA last! */ 2614 build_zonelists_in_zone_order(pgdat, j); 2615 } 2616 2617 build_thisnode_zonelists(pgdat); 2618} 2619 2620/* Construct the zonelist performance cache - see further mmzone.h */ 2621static void build_zonelist_cache(pg_data_t *pgdat) 2622{ 2623 struct zonelist *zonelist; 2624 struct zonelist_cache *zlc; 2625 struct zoneref *z; 2626 2627 zonelist = &pgdat->node_zonelists[0]; 2628 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2629 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2630 for (z = zonelist->_zonerefs; z->zone; z++) 2631 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2632} 2633 2634 2635#else /* CONFIG_NUMA */ 2636 2637static void set_zonelist_order(void) 2638{ 2639 current_zonelist_order = ZONELIST_ORDER_ZONE; 2640} 2641 2642static void build_zonelists(pg_data_t *pgdat) 2643{ 2644 int node, local_node; 2645 enum zone_type j; 2646 struct zonelist *zonelist; 2647 2648 local_node = pgdat->node_id; 2649 2650 zonelist = &pgdat->node_zonelists[0]; 2651 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2652 2653 /* 2654 * Now we build the zonelist so that it contains the zones 2655 * of all the other nodes. 2656 * We don't want to pressure a particular node, so when 2657 * building the zones for node N, we make sure that the 2658 * zones coming right after the local ones are those from 2659 * node N+1 (modulo N) 2660 */ 2661 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2662 if (!node_online(node)) 2663 continue; 2664 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2665 MAX_NR_ZONES - 1); 2666 } 2667 for (node = 0; node < local_node; node++) { 2668 if (!node_online(node)) 2669 continue; 2670 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2671 MAX_NR_ZONES - 1); 2672 } 2673 2674 zonelist->_zonerefs[j].zone = NULL; 2675 zonelist->_zonerefs[j].zone_idx = 0; 2676} 2677 2678/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2679static void build_zonelist_cache(pg_data_t *pgdat) 2680{ 2681 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2682} 2683 2684#endif /* CONFIG_NUMA */ 2685 2686/* return values int ....just for stop_machine() */ 2687static int __build_all_zonelists(void *dummy) 2688{ 2689 int nid; 2690 2691#ifdef CONFIG_NUMA 2692 memset(node_load, 0, sizeof(node_load)); 2693#endif 2694 for_each_online_node(nid) { 2695 pg_data_t *pgdat = NODE_DATA(nid); 2696 2697 build_zonelists(pgdat); 2698 build_zonelist_cache(pgdat); 2699 } 2700 return 0; 2701} 2702 2703void build_all_zonelists(void) 2704{ 2705 set_zonelist_order(); 2706 2707 if (system_state == SYSTEM_BOOTING) { 2708 __build_all_zonelists(NULL); 2709 mminit_verify_zonelist(); 2710 cpuset_init_current_mems_allowed(); 2711 } else { 2712 /* we have to stop all cpus to guarantee there is no user 2713 of zonelist */ 2714 stop_machine(__build_all_zonelists, NULL, NULL); 2715 /* cpuset refresh routine should be here */ 2716 } 2717 vm_total_pages = nr_free_pagecache_pages(); 2718 /* 2719 * Disable grouping by mobility if the number of pages in the 2720 * system is too low to allow the mechanism to work. It would be 2721 * more accurate, but expensive to check per-zone. This check is 2722 * made on memory-hotadd so a system can start with mobility 2723 * disabled and enable it later 2724 */ 2725 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2726 page_group_by_mobility_disabled = 1; 2727 else 2728 page_group_by_mobility_disabled = 0; 2729 2730 printk("Built %i zonelists in %s order, mobility grouping %s. " 2731 "Total pages: %ld\n", 2732 nr_online_nodes, 2733 zonelist_order_name[current_zonelist_order], 2734 page_group_by_mobility_disabled ? "off" : "on", 2735 vm_total_pages); 2736#ifdef CONFIG_NUMA 2737 printk("Policy zone: %s\n", zone_names[policy_zone]); 2738#endif 2739} 2740 2741/* 2742 * Helper functions to size the waitqueue hash table. 2743 * Essentially these want to choose hash table sizes sufficiently 2744 * large so that collisions trying to wait on pages are rare. 2745 * But in fact, the number of active page waitqueues on typical 2746 * systems is ridiculously low, less than 200. So this is even 2747 * conservative, even though it seems large. 2748 * 2749 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2750 * waitqueues, i.e. the size of the waitq table given the number of pages. 2751 */ 2752#define PAGES_PER_WAITQUEUE 256 2753 2754#ifndef CONFIG_MEMORY_HOTPLUG 2755static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2756{ 2757 unsigned long size = 1; 2758 2759 pages /= PAGES_PER_WAITQUEUE; 2760 2761 while (size < pages) 2762 size <<= 1; 2763 2764 /* 2765 * Once we have dozens or even hundreds of threads sleeping 2766 * on IO we've got bigger problems than wait queue collision. 2767 * Limit the size of the wait table to a reasonable size. 2768 */ 2769 size = min(size, 4096UL); 2770 2771 return max(size, 4UL); 2772} 2773#else 2774/* 2775 * A zone's size might be changed by hot-add, so it is not possible to determine 2776 * a suitable size for its wait_table. So we use the maximum size now. 2777 * 2778 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2779 * 2780 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2781 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2782 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2783 * 2784 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2785 * or more by the traditional way. (See above). It equals: 2786 * 2787 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2788 * ia64(16K page size) : = ( 8G + 4M)byte. 2789 * powerpc (64K page size) : = (32G +16M)byte. 2790 */ 2791static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2792{ 2793 return 4096UL; 2794} 2795#endif 2796 2797/* 2798 * This is an integer logarithm so that shifts can be used later 2799 * to extract the more random high bits from the multiplicative 2800 * hash function before the remainder is taken. 2801 */ 2802static inline unsigned long wait_table_bits(unsigned long size) 2803{ 2804 return ffz(~size); 2805} 2806 2807#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2808 2809/* 2810 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2811 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2812 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2813 * higher will lead to a bigger reserve which will get freed as contiguous 2814 * blocks as reclaim kicks in 2815 */ 2816static void setup_zone_migrate_reserve(struct zone *zone) 2817{ 2818 unsigned long start_pfn, pfn, end_pfn; 2819 struct page *page; 2820 unsigned long reserve, block_migratetype; 2821 2822 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2823 start_pfn = zone->zone_start_pfn; 2824 end_pfn = start_pfn + zone->spanned_pages; 2825 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2826 pageblock_order; 2827 2828 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2829 if (!pfn_valid(pfn)) 2830 continue; 2831 page = pfn_to_page(pfn); 2832 2833 /* Watch out for overlapping nodes */ 2834 if (page_to_nid(page) != zone_to_nid(zone)) 2835 continue; 2836 2837 /* Blocks with reserved pages will never free, skip them. */ 2838 if (PageReserved(page)) 2839 continue; 2840 2841 block_migratetype = get_pageblock_migratetype(page); 2842 2843 /* If this block is reserved, account for it */ 2844 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2845 reserve--; 2846 continue; 2847 } 2848 2849 /* Suitable for reserving if this block is movable */ 2850 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2851 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2852 move_freepages_block(zone, page, MIGRATE_RESERVE); 2853 reserve--; 2854 continue; 2855 } 2856 2857 /* 2858 * If the reserve is met and this is a previous reserved block, 2859 * take it back 2860 */ 2861 if (block_migratetype == MIGRATE_RESERVE) { 2862 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2863 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2864 } 2865 } 2866} 2867 2868/* 2869 * Initially all pages are reserved - free ones are freed 2870 * up by free_all_bootmem() once the early boot process is 2871 * done. Non-atomic initialization, single-pass. 2872 */ 2873void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2874 unsigned long start_pfn, enum memmap_context context) 2875{ 2876 struct page *page; 2877 unsigned long end_pfn = start_pfn + size; 2878 unsigned long pfn; 2879 struct zone *z; 2880 2881 if (highest_memmap_pfn < end_pfn - 1) 2882 highest_memmap_pfn = end_pfn - 1; 2883 2884 z = &NODE_DATA(nid)->node_zones[zone]; 2885 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2886 /* 2887 * There can be holes in boot-time mem_map[]s 2888 * handed to this function. They do not 2889 * exist on hotplugged memory. 2890 */ 2891 if (context == MEMMAP_EARLY) { 2892 if (!early_pfn_valid(pfn)) 2893 continue; 2894 if (!early_pfn_in_nid(pfn, nid)) 2895 continue; 2896 } 2897 page = pfn_to_page(pfn); 2898 set_page_links(page, zone, nid, pfn); 2899 mminit_verify_page_links(page, zone, nid, pfn); 2900 init_page_count(page); 2901 reset_page_mapcount(page); 2902 SetPageReserved(page); 2903 /* 2904 * Mark the block movable so that blocks are reserved for 2905 * movable at startup. This will force kernel allocations 2906 * to reserve their blocks rather than leaking throughout 2907 * the address space during boot when many long-lived 2908 * kernel allocations are made. Later some blocks near 2909 * the start are marked MIGRATE_RESERVE by 2910 * setup_zone_migrate_reserve() 2911 * 2912 * bitmap is created for zone's valid pfn range. but memmap 2913 * can be created for invalid pages (for alignment) 2914 * check here not to call set_pageblock_migratetype() against 2915 * pfn out of zone. 2916 */ 2917 if ((z->zone_start_pfn <= pfn) 2918 && (pfn < z->zone_start_pfn + z->spanned_pages) 2919 && !(pfn & (pageblock_nr_pages - 1))) 2920 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2921 2922 INIT_LIST_HEAD(&page->lru); 2923#ifdef WANT_PAGE_VIRTUAL 2924 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2925 if (!is_highmem_idx(zone)) 2926 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2927#endif 2928 } 2929} 2930 2931static void __meminit zone_init_free_lists(struct zone *zone) 2932{ 2933 int order, t; 2934 for_each_migratetype_order(order, t) { 2935 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 2936 zone->free_area[order].nr_free = 0; 2937 } 2938} 2939 2940#ifndef __HAVE_ARCH_MEMMAP_INIT 2941#define memmap_init(size, nid, zone, start_pfn) \ 2942 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 2943#endif 2944 2945static int zone_batchsize(struct zone *zone) 2946{ 2947#ifdef CONFIG_MMU 2948 int batch; 2949 2950 /* 2951 * The per-cpu-pages pools are set to around 1000th of the 2952 * size of the zone. But no more than 1/2 of a meg. 2953 * 2954 * OK, so we don't know how big the cache is. So guess. 2955 */ 2956 batch = zone->present_pages / 1024; 2957 if (batch * PAGE_SIZE > 512 * 1024) 2958 batch = (512 * 1024) / PAGE_SIZE; 2959 batch /= 4; /* We effectively *= 4 below */ 2960 if (batch < 1) 2961 batch = 1; 2962 2963 /* 2964 * Clamp the batch to a 2^n - 1 value. Having a power 2965 * of 2 value was found to be more likely to have 2966 * suboptimal cache aliasing properties in some cases. 2967 * 2968 * For example if 2 tasks are alternately allocating 2969 * batches of pages, one task can end up with a lot 2970 * of pages of one half of the possible page colors 2971 * and the other with pages of the other colors. 2972 */ 2973 batch = rounddown_pow_of_two(batch + batch/2) - 1; 2974 2975 return batch; 2976 2977#else 2978 /* The deferral and batching of frees should be suppressed under NOMMU 2979 * conditions. 2980 * 2981 * The problem is that NOMMU needs to be able to allocate large chunks 2982 * of contiguous memory as there's no hardware page translation to 2983 * assemble apparent contiguous memory from discontiguous pages. 2984 * 2985 * Queueing large contiguous runs of pages for batching, however, 2986 * causes the pages to actually be freed in smaller chunks. As there 2987 * can be a significant delay between the individual batches being 2988 * recycled, this leads to the once large chunks of space being 2989 * fragmented and becoming unavailable for high-order allocations. 2990 */ 2991 return 0; 2992#endif 2993} 2994 2995static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 2996{ 2997 struct per_cpu_pages *pcp; 2998 2999 memset(p, 0, sizeof(*p)); 3000 3001 pcp = &p->pcp; 3002 pcp->count = 0; 3003 pcp->high = 6 * batch; 3004 pcp->batch = max(1UL, 1 * batch); 3005 INIT_LIST_HEAD(&pcp->list); 3006} 3007 3008/* 3009 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3010 * to the value high for the pageset p. 3011 */ 3012 3013static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3014 unsigned long high) 3015{ 3016 struct per_cpu_pages *pcp; 3017 3018 pcp = &p->pcp; 3019 pcp->high = high; 3020 pcp->batch = max(1UL, high/4); 3021 if ((high/4) > (PAGE_SHIFT * 8)) 3022 pcp->batch = PAGE_SHIFT * 8; 3023} 3024 3025 3026#ifdef CONFIG_NUMA 3027/* 3028 * Boot pageset table. One per cpu which is going to be used for all 3029 * zones and all nodes. The parameters will be set in such a way 3030 * that an item put on a list will immediately be handed over to 3031 * the buddy list. This is safe since pageset manipulation is done 3032 * with interrupts disabled. 3033 * 3034 * Some NUMA counter updates may also be caught by the boot pagesets. 3035 * 3036 * The boot_pagesets must be kept even after bootup is complete for 3037 * unused processors and/or zones. They do play a role for bootstrapping 3038 * hotplugged processors. 3039 * 3040 * zoneinfo_show() and maybe other functions do 3041 * not check if the processor is online before following the pageset pointer. 3042 * Other parts of the kernel may not check if the zone is available. 3043 */ 3044static struct per_cpu_pageset boot_pageset[NR_CPUS]; 3045 3046/* 3047 * Dynamically allocate memory for the 3048 * per cpu pageset array in struct zone. 3049 */ 3050static int __cpuinit process_zones(int cpu) 3051{ 3052 struct zone *zone, *dzone; 3053 int node = cpu_to_node(cpu); 3054 3055 node_set_state(node, N_CPU); /* this node has a cpu */ 3056 3057 for_each_populated_zone(zone) { 3058 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 3059 GFP_KERNEL, node); 3060 if (!zone_pcp(zone, cpu)) 3061 goto bad; 3062 3063 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 3064 3065 if (percpu_pagelist_fraction) 3066 setup_pagelist_highmark(zone_pcp(zone, cpu), 3067 (zone->present_pages / percpu_pagelist_fraction)); 3068 } 3069 3070 return 0; 3071bad: 3072 for_each_zone(dzone) { 3073 if (!populated_zone(dzone)) 3074 continue; 3075 if (dzone == zone) 3076 break; 3077 kfree(zone_pcp(dzone, cpu)); 3078 zone_pcp(dzone, cpu) = &boot_pageset[cpu]; 3079 } 3080 return -ENOMEM; 3081} 3082 3083static inline void free_zone_pagesets(int cpu) 3084{ 3085 struct zone *zone; 3086 3087 for_each_zone(zone) { 3088 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 3089 3090 /* Free per_cpu_pageset if it is slab allocated */ 3091 if (pset != &boot_pageset[cpu]) 3092 kfree(pset); 3093 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3094 } 3095} 3096 3097static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 3098 unsigned long action, 3099 void *hcpu) 3100{ 3101 int cpu = (long)hcpu; 3102 int ret = NOTIFY_OK; 3103 3104 switch (action) { 3105 case CPU_UP_PREPARE: 3106 case CPU_UP_PREPARE_FROZEN: 3107 if (process_zones(cpu)) 3108 ret = NOTIFY_BAD; 3109 break; 3110 case CPU_UP_CANCELED: 3111 case CPU_UP_CANCELED_FROZEN: 3112 case CPU_DEAD: 3113 case CPU_DEAD_FROZEN: 3114 free_zone_pagesets(cpu); 3115 break; 3116 default: 3117 break; 3118 } 3119 return ret; 3120} 3121 3122static struct notifier_block __cpuinitdata pageset_notifier = 3123 { &pageset_cpuup_callback, NULL, 0 }; 3124 3125void __init setup_per_cpu_pageset(void) 3126{ 3127 int err; 3128 3129 /* Initialize per_cpu_pageset for cpu 0. 3130 * A cpuup callback will do this for every cpu 3131 * as it comes online 3132 */ 3133 err = process_zones(smp_processor_id()); 3134 BUG_ON(err); 3135 register_cpu_notifier(&pageset_notifier); 3136} 3137 3138#endif 3139 3140static noinline __init_refok 3141int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3142{ 3143 int i; 3144 struct pglist_data *pgdat = zone->zone_pgdat; 3145 size_t alloc_size; 3146 3147 /* 3148 * The per-page waitqueue mechanism uses hashed waitqueues 3149 * per zone. 3150 */ 3151 zone->wait_table_hash_nr_entries = 3152 wait_table_hash_nr_entries(zone_size_pages); 3153 zone->wait_table_bits = 3154 wait_table_bits(zone->wait_table_hash_nr_entries); 3155 alloc_size = zone->wait_table_hash_nr_entries 3156 * sizeof(wait_queue_head_t); 3157 3158 if (!slab_is_available()) { 3159 zone->wait_table = (wait_queue_head_t *) 3160 alloc_bootmem_node(pgdat, alloc_size); 3161 } else { 3162 /* 3163 * This case means that a zone whose size was 0 gets new memory 3164 * via memory hot-add. 3165 * But it may be the case that a new node was hot-added. In 3166 * this case vmalloc() will not be able to use this new node's 3167 * memory - this wait_table must be initialized to use this new 3168 * node itself as well. 3169 * To use this new node's memory, further consideration will be 3170 * necessary. 3171 */ 3172 zone->wait_table = vmalloc(alloc_size); 3173 } 3174 if (!zone->wait_table) 3175 return -ENOMEM; 3176 3177 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3178 init_waitqueue_head(zone->wait_table + i); 3179 3180 return 0; 3181} 3182 3183static int __zone_pcp_update(void *data) 3184{ 3185 struct zone *zone = data; 3186 int cpu; 3187 unsigned long batch = zone_batchsize(zone), flags; 3188 3189 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3190 struct per_cpu_pageset *pset; 3191 struct per_cpu_pages *pcp; 3192 3193 pset = zone_pcp(zone, cpu); 3194 pcp = &pset->pcp; 3195 3196 local_irq_save(flags); 3197 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 3198 setup_pageset(pset, batch); 3199 local_irq_restore(flags); 3200 } 3201 return 0; 3202} 3203 3204void zone_pcp_update(struct zone *zone) 3205{ 3206 stop_machine(__zone_pcp_update, zone, NULL); 3207} 3208 3209static __meminit void zone_pcp_init(struct zone *zone) 3210{ 3211 int cpu; 3212 unsigned long batch = zone_batchsize(zone); 3213 3214 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3215#ifdef CONFIG_NUMA 3216 /* Early boot. Slab allocator not functional yet */ 3217 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3218 setup_pageset(&boot_pageset[cpu],0); 3219#else 3220 setup_pageset(zone_pcp(zone,cpu), batch); 3221#endif 3222 } 3223 if (zone->present_pages) 3224 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 3225 zone->name, zone->present_pages, batch); 3226} 3227 3228__meminit int init_currently_empty_zone(struct zone *zone, 3229 unsigned long zone_start_pfn, 3230 unsigned long size, 3231 enum memmap_context context) 3232{ 3233 struct pglist_data *pgdat = zone->zone_pgdat; 3234 int ret; 3235 ret = zone_wait_table_init(zone, size); 3236 if (ret) 3237 return ret; 3238 pgdat->nr_zones = zone_idx(zone) + 1; 3239 3240 zone->zone_start_pfn = zone_start_pfn; 3241 3242 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3243 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3244 pgdat->node_id, 3245 (unsigned long)zone_idx(zone), 3246 zone_start_pfn, (zone_start_pfn + size)); 3247 3248 zone_init_free_lists(zone); 3249 3250 return 0; 3251} 3252 3253#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3254/* 3255 * Basic iterator support. Return the first range of PFNs for a node 3256 * Note: nid == MAX_NUMNODES returns first region regardless of node 3257 */ 3258static int __meminit first_active_region_index_in_nid(int nid) 3259{ 3260 int i; 3261 3262 for (i = 0; i < nr_nodemap_entries; i++) 3263 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3264 return i; 3265 3266 return -1; 3267} 3268 3269/* 3270 * Basic iterator support. Return the next active range of PFNs for a node 3271 * Note: nid == MAX_NUMNODES returns next region regardless of node 3272 */ 3273static int __meminit next_active_region_index_in_nid(int index, int nid) 3274{ 3275 for (index = index + 1; index < nr_nodemap_entries; index++) 3276 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3277 return index; 3278 3279 return -1; 3280} 3281 3282#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3283/* 3284 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3285 * Architectures may implement their own version but if add_active_range() 3286 * was used and there are no special requirements, this is a convenient 3287 * alternative 3288 */ 3289int __meminit __early_pfn_to_nid(unsigned long pfn) 3290{ 3291 int i; 3292 3293 for (i = 0; i < nr_nodemap_entries; i++) { 3294 unsigned long start_pfn = early_node_map[i].start_pfn; 3295 unsigned long end_pfn = early_node_map[i].end_pfn; 3296 3297 if (start_pfn <= pfn && pfn < end_pfn) 3298 return early_node_map[i].nid; 3299 } 3300 /* This is a memory hole */ 3301 return -1; 3302} 3303#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3304 3305int __meminit early_pfn_to_nid(unsigned long pfn) 3306{ 3307 int nid; 3308 3309 nid = __early_pfn_to_nid(pfn); 3310 if (nid >= 0) 3311 return nid; 3312 /* just returns 0 */ 3313 return 0; 3314} 3315 3316#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3317bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3318{ 3319 int nid; 3320 3321 nid = __early_pfn_to_nid(pfn); 3322 if (nid >= 0 && nid != node) 3323 return false; 3324 return true; 3325} 3326#endif 3327 3328/* Basic iterator support to walk early_node_map[] */ 3329#define for_each_active_range_index_in_nid(i, nid) \ 3330 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3331 i = next_active_region_index_in_nid(i, nid)) 3332 3333/** 3334 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3335 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3336 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3337 * 3338 * If an architecture guarantees that all ranges registered with 3339 * add_active_ranges() contain no holes and may be freed, this 3340 * this function may be used instead of calling free_bootmem() manually. 3341 */ 3342void __init free_bootmem_with_active_regions(int nid, 3343 unsigned long max_low_pfn) 3344{ 3345 int i; 3346 3347 for_each_active_range_index_in_nid(i, nid) { 3348 unsigned long size_pages = 0; 3349 unsigned long end_pfn = early_node_map[i].end_pfn; 3350 3351 if (early_node_map[i].start_pfn >= max_low_pfn) 3352 continue; 3353 3354 if (end_pfn > max_low_pfn) 3355 end_pfn = max_low_pfn; 3356 3357 size_pages = end_pfn - early_node_map[i].start_pfn; 3358 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3359 PFN_PHYS(early_node_map[i].start_pfn), 3360 size_pages << PAGE_SHIFT); 3361 } 3362} 3363 3364void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3365{ 3366 int i; 3367 int ret; 3368 3369 for_each_active_range_index_in_nid(i, nid) { 3370 ret = work_fn(early_node_map[i].start_pfn, 3371 early_node_map[i].end_pfn, data); 3372 if (ret) 3373 break; 3374 } 3375} 3376/** 3377 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3378 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3379 * 3380 * If an architecture guarantees that all ranges registered with 3381 * add_active_ranges() contain no holes and may be freed, this 3382 * function may be used instead of calling memory_present() manually. 3383 */ 3384void __init sparse_memory_present_with_active_regions(int nid) 3385{ 3386 int i; 3387 3388 for_each_active_range_index_in_nid(i, nid) 3389 memory_present(early_node_map[i].nid, 3390 early_node_map[i].start_pfn, 3391 early_node_map[i].end_pfn); 3392} 3393 3394/** 3395 * get_pfn_range_for_nid - Return the start and end page frames for a node 3396 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3397 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3398 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3399 * 3400 * It returns the start and end page frame of a node based on information 3401 * provided by an arch calling add_active_range(). If called for a node 3402 * with no available memory, a warning is printed and the start and end 3403 * PFNs will be 0. 3404 */ 3405void __meminit get_pfn_range_for_nid(unsigned int nid, 3406 unsigned long *start_pfn, unsigned long *end_pfn) 3407{ 3408 int i; 3409 *start_pfn = -1UL; 3410 *end_pfn = 0; 3411 3412 for_each_active_range_index_in_nid(i, nid) { 3413 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3414 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3415 } 3416 3417 if (*start_pfn == -1UL) 3418 *start_pfn = 0; 3419} 3420 3421/* 3422 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3423 * assumption is made that zones within a node are ordered in monotonic 3424 * increasing memory addresses so that the "highest" populated zone is used 3425 */ 3426static void __init find_usable_zone_for_movable(void) 3427{ 3428 int zone_index; 3429 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3430 if (zone_index == ZONE_MOVABLE) 3431 continue; 3432 3433 if (arch_zone_highest_possible_pfn[zone_index] > 3434 arch_zone_lowest_possible_pfn[zone_index]) 3435 break; 3436 } 3437 3438 VM_BUG_ON(zone_index == -1); 3439 movable_zone = zone_index; 3440} 3441 3442/* 3443 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3444 * because it is sized independant of architecture. Unlike the other zones, 3445 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3446 * in each node depending on the size of each node and how evenly kernelcore 3447 * is distributed. This helper function adjusts the zone ranges 3448 * provided by the architecture for a given node by using the end of the 3449 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3450 * zones within a node are in order of monotonic increases memory addresses 3451 */ 3452static void __meminit adjust_zone_range_for_zone_movable(int nid, 3453 unsigned long zone_type, 3454 unsigned long node_start_pfn, 3455 unsigned long node_end_pfn, 3456 unsigned long *zone_start_pfn, 3457 unsigned long *zone_end_pfn) 3458{ 3459 /* Only adjust if ZONE_MOVABLE is on this node */ 3460 if (zone_movable_pfn[nid]) { 3461 /* Size ZONE_MOVABLE */ 3462 if (zone_type == ZONE_MOVABLE) { 3463 *zone_start_pfn = zone_movable_pfn[nid]; 3464 *zone_end_pfn = min(node_end_pfn, 3465 arch_zone_highest_possible_pfn[movable_zone]); 3466 3467 /* Adjust for ZONE_MOVABLE starting within this range */ 3468 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3469 *zone_end_pfn > zone_movable_pfn[nid]) { 3470 *zone_end_pfn = zone_movable_pfn[nid]; 3471 3472 /* Check if this whole range is within ZONE_MOVABLE */ 3473 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3474 *zone_start_pfn = *zone_end_pfn; 3475 } 3476} 3477 3478/* 3479 * Return the number of pages a zone spans in a node, including holes 3480 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3481 */ 3482static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3483 unsigned long zone_type, 3484 unsigned long *ignored) 3485{ 3486 unsigned long node_start_pfn, node_end_pfn; 3487 unsigned long zone_start_pfn, zone_end_pfn; 3488 3489 /* Get the start and end of the node and zone */ 3490 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3491 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3492 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3493 adjust_zone_range_for_zone_movable(nid, zone_type, 3494 node_start_pfn, node_end_pfn, 3495 &zone_start_pfn, &zone_end_pfn); 3496 3497 /* Check that this node has pages within the zone's required range */ 3498 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3499 return 0; 3500 3501 /* Move the zone boundaries inside the node if necessary */ 3502 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3503 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3504 3505 /* Return the spanned pages */ 3506 return zone_end_pfn - zone_start_pfn; 3507} 3508 3509/* 3510 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3511 * then all holes in the requested range will be accounted for. 3512 */ 3513static unsigned long __meminit __absent_pages_in_range(int nid, 3514 unsigned long range_start_pfn, 3515 unsigned long range_end_pfn) 3516{ 3517 int i = 0; 3518 unsigned long prev_end_pfn = 0, hole_pages = 0; 3519 unsigned long start_pfn; 3520 3521 /* Find the end_pfn of the first active range of pfns in the node */ 3522 i = first_active_region_index_in_nid(nid); 3523 if (i == -1) 3524 return 0; 3525 3526 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3527 3528 /* Account for ranges before physical memory on this node */ 3529 if (early_node_map[i].start_pfn > range_start_pfn) 3530 hole_pages = prev_end_pfn - range_start_pfn; 3531 3532 /* Find all holes for the zone within the node */ 3533 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3534 3535 /* No need to continue if prev_end_pfn is outside the zone */ 3536 if (prev_end_pfn >= range_end_pfn) 3537 break; 3538 3539 /* Make sure the end of the zone is not within the hole */ 3540 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3541 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3542 3543 /* Update the hole size cound and move on */ 3544 if (start_pfn > range_start_pfn) { 3545 BUG_ON(prev_end_pfn > start_pfn); 3546 hole_pages += start_pfn - prev_end_pfn; 3547 } 3548 prev_end_pfn = early_node_map[i].end_pfn; 3549 } 3550 3551 /* Account for ranges past physical memory on this node */ 3552 if (range_end_pfn > prev_end_pfn) 3553 hole_pages += range_end_pfn - 3554 max(range_start_pfn, prev_end_pfn); 3555 3556 return hole_pages; 3557} 3558 3559/** 3560 * absent_pages_in_range - Return number of page frames in holes within a range 3561 * @start_pfn: The start PFN to start searching for holes 3562 * @end_pfn: The end PFN to stop searching for holes 3563 * 3564 * It returns the number of pages frames in memory holes within a range. 3565 */ 3566unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3567 unsigned long end_pfn) 3568{ 3569 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3570} 3571 3572/* Return the number of page frames in holes in a zone on a node */ 3573static unsigned long __meminit zone_absent_pages_in_node(int nid, 3574 unsigned long zone_type, 3575 unsigned long *ignored) 3576{ 3577 unsigned long node_start_pfn, node_end_pfn; 3578 unsigned long zone_start_pfn, zone_end_pfn; 3579 3580 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3581 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3582 node_start_pfn); 3583 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3584 node_end_pfn); 3585 3586 adjust_zone_range_for_zone_movable(nid, zone_type, 3587 node_start_pfn, node_end_pfn, 3588 &zone_start_pfn, &zone_end_pfn); 3589 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3590} 3591 3592#else 3593static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3594 unsigned long zone_type, 3595 unsigned long *zones_size) 3596{ 3597 return zones_size[zone_type]; 3598} 3599 3600static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3601 unsigned long zone_type, 3602 unsigned long *zholes_size) 3603{ 3604 if (!zholes_size) 3605 return 0; 3606 3607 return zholes_size[zone_type]; 3608} 3609 3610#endif 3611 3612static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3613 unsigned long *zones_size, unsigned long *zholes_size) 3614{ 3615 unsigned long realtotalpages, totalpages = 0; 3616 enum zone_type i; 3617 3618 for (i = 0; i < MAX_NR_ZONES; i++) 3619 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3620 zones_size); 3621 pgdat->node_spanned_pages = totalpages; 3622 3623 realtotalpages = totalpages; 3624 for (i = 0; i < MAX_NR_ZONES; i++) 3625 realtotalpages -= 3626 zone_absent_pages_in_node(pgdat->node_id, i, 3627 zholes_size); 3628 pgdat->node_present_pages = realtotalpages; 3629 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3630 realtotalpages); 3631} 3632 3633#ifndef CONFIG_SPARSEMEM 3634/* 3635 * Calculate the size of the zone->blockflags rounded to an unsigned long 3636 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3637 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3638 * round what is now in bits to nearest long in bits, then return it in 3639 * bytes. 3640 */ 3641static unsigned long __init usemap_size(unsigned long zonesize) 3642{ 3643 unsigned long usemapsize; 3644 3645 usemapsize = roundup(zonesize, pageblock_nr_pages); 3646 usemapsize = usemapsize >> pageblock_order; 3647 usemapsize *= NR_PAGEBLOCK_BITS; 3648 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3649 3650 return usemapsize / 8; 3651} 3652 3653static void __init setup_usemap(struct pglist_data *pgdat, 3654 struct zone *zone, unsigned long zonesize) 3655{ 3656 unsigned long usemapsize = usemap_size(zonesize); 3657 zone->pageblock_flags = NULL; 3658 if (usemapsize) 3659 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3660} 3661#else 3662static void inline setup_usemap(struct pglist_data *pgdat, 3663 struct zone *zone, unsigned long zonesize) {} 3664#endif /* CONFIG_SPARSEMEM */ 3665 3666#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3667 3668/* Return a sensible default order for the pageblock size. */ 3669static inline int pageblock_default_order(void) 3670{ 3671 if (HPAGE_SHIFT > PAGE_SHIFT) 3672 return HUGETLB_PAGE_ORDER; 3673 3674 return MAX_ORDER-1; 3675} 3676 3677/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3678static inline void __init set_pageblock_order(unsigned int order) 3679{ 3680 /* Check that pageblock_nr_pages has not already been setup */ 3681 if (pageblock_order) 3682 return; 3683 3684 /* 3685 * Assume the largest contiguous order of interest is a huge page. 3686 * This value may be variable depending on boot parameters on IA64 3687 */ 3688 pageblock_order = order; 3689} 3690#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3691 3692/* 3693 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3694 * and pageblock_default_order() are unused as pageblock_order is set 3695 * at compile-time. See include/linux/pageblock-flags.h for the values of 3696 * pageblock_order based on the kernel config 3697 */ 3698static inline int pageblock_default_order(unsigned int order) 3699{ 3700 return MAX_ORDER-1; 3701} 3702#define set_pageblock_order(x) do {} while (0) 3703 3704#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3705 3706/* 3707 * Set up the zone data structures: 3708 * - mark all pages reserved 3709 * - mark all memory queues empty 3710 * - clear the memory bitmaps 3711 */ 3712static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3713 unsigned long *zones_size, unsigned long *zholes_size) 3714{ 3715 enum zone_type j; 3716 int nid = pgdat->node_id; 3717 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3718 int ret; 3719 3720 pgdat_resize_init(pgdat); 3721 pgdat->nr_zones = 0; 3722 init_waitqueue_head(&pgdat->kswapd_wait); 3723 pgdat->kswapd_max_order = 0; 3724 pgdat_page_cgroup_init(pgdat); 3725 3726 for (j = 0; j < MAX_NR_ZONES; j++) { 3727 struct zone *zone = pgdat->node_zones + j; 3728 unsigned long size, realsize, memmap_pages; 3729 enum lru_list l; 3730 3731 size = zone_spanned_pages_in_node(nid, j, zones_size); 3732 realsize = size - zone_absent_pages_in_node(nid, j, 3733 zholes_size); 3734 3735 /* 3736 * Adjust realsize so that it accounts for how much memory 3737 * is used by this zone for memmap. This affects the watermark 3738 * and per-cpu initialisations 3739 */ 3740 memmap_pages = 3741 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3742 if (realsize >= memmap_pages) { 3743 realsize -= memmap_pages; 3744 if (memmap_pages) 3745 printk(KERN_DEBUG 3746 " %s zone: %lu pages used for memmap\n", 3747 zone_names[j], memmap_pages); 3748 } else 3749 printk(KERN_WARNING 3750 " %s zone: %lu pages exceeds realsize %lu\n", 3751 zone_names[j], memmap_pages, realsize); 3752 3753 /* Account for reserved pages */ 3754 if (j == 0 && realsize > dma_reserve) { 3755 realsize -= dma_reserve; 3756 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3757 zone_names[0], dma_reserve); 3758 } 3759 3760 if (!is_highmem_idx(j)) 3761 nr_kernel_pages += realsize; 3762 nr_all_pages += realsize; 3763 3764 zone->spanned_pages = size; 3765 zone->present_pages = realsize; 3766#ifdef CONFIG_NUMA 3767 zone->node = nid; 3768 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3769 / 100; 3770 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3771#endif 3772 zone->name = zone_names[j]; 3773 spin_lock_init(&zone->lock); 3774 spin_lock_init(&zone->lru_lock); 3775 zone_seqlock_init(zone); 3776 zone->zone_pgdat = pgdat; 3777 3778 zone->prev_priority = DEF_PRIORITY; 3779 3780 zone_pcp_init(zone); 3781 for_each_lru(l) { 3782 INIT_LIST_HEAD(&zone->lru[l].list); 3783 zone->lru[l].nr_saved_scan = 0; 3784 } 3785 zone->reclaim_stat.recent_rotated[0] = 0; 3786 zone->reclaim_stat.recent_rotated[1] = 0; 3787 zone->reclaim_stat.recent_scanned[0] = 0; 3788 zone->reclaim_stat.recent_scanned[1] = 0; 3789 zap_zone_vm_stats(zone); 3790 zone->flags = 0; 3791 if (!size) 3792 continue; 3793 3794 set_pageblock_order(pageblock_default_order()); 3795 setup_usemap(pgdat, zone, size); 3796 ret = init_currently_empty_zone(zone, zone_start_pfn, 3797 size, MEMMAP_EARLY); 3798 BUG_ON(ret); 3799 memmap_init(size, nid, j, zone_start_pfn); 3800 zone_start_pfn += size; 3801 } 3802} 3803 3804static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3805{ 3806 /* Skip empty nodes */ 3807 if (!pgdat->node_spanned_pages) 3808 return; 3809 3810#ifdef CONFIG_FLAT_NODE_MEM_MAP 3811 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3812 if (!pgdat->node_mem_map) { 3813 unsigned long size, start, end; 3814 struct page *map; 3815 3816 /* 3817 * The zone's endpoints aren't required to be MAX_ORDER 3818 * aligned but the node_mem_map endpoints must be in order 3819 * for the buddy allocator to function correctly. 3820 */ 3821 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3822 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3823 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3824 size = (end - start) * sizeof(struct page); 3825 map = alloc_remap(pgdat->node_id, size); 3826 if (!map) 3827 map = alloc_bootmem_node(pgdat, size); 3828 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3829 } 3830#ifndef CONFIG_NEED_MULTIPLE_NODES 3831 /* 3832 * With no DISCONTIG, the global mem_map is just set as node 0's 3833 */ 3834 if (pgdat == NODE_DATA(0)) { 3835 mem_map = NODE_DATA(0)->node_mem_map; 3836#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3837 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3838 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3839#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3840 } 3841#endif 3842#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3843} 3844 3845void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3846 unsigned long node_start_pfn, unsigned long *zholes_size) 3847{ 3848 pg_data_t *pgdat = NODE_DATA(nid); 3849 3850 pgdat->node_id = nid; 3851 pgdat->node_start_pfn = node_start_pfn; 3852 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3853 3854 alloc_node_mem_map(pgdat); 3855#ifdef CONFIG_FLAT_NODE_MEM_MAP 3856 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3857 nid, (unsigned long)pgdat, 3858 (unsigned long)pgdat->node_mem_map); 3859#endif 3860 3861 free_area_init_core(pgdat, zones_size, zholes_size); 3862} 3863 3864#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3865 3866#if MAX_NUMNODES > 1 3867/* 3868 * Figure out the number of possible node ids. 3869 */ 3870static void __init setup_nr_node_ids(void) 3871{ 3872 unsigned int node; 3873 unsigned int highest = 0; 3874 3875 for_each_node_mask(node, node_possible_map) 3876 highest = node; 3877 nr_node_ids = highest + 1; 3878} 3879#else 3880static inline void setup_nr_node_ids(void) 3881{ 3882} 3883#endif 3884 3885/** 3886 * add_active_range - Register a range of PFNs backed by physical memory 3887 * @nid: The node ID the range resides on 3888 * @start_pfn: The start PFN of the available physical memory 3889 * @end_pfn: The end PFN of the available physical memory 3890 * 3891 * These ranges are stored in an early_node_map[] and later used by 3892 * free_area_init_nodes() to calculate zone sizes and holes. If the 3893 * range spans a memory hole, it is up to the architecture to ensure 3894 * the memory is not freed by the bootmem allocator. If possible 3895 * the range being registered will be merged with existing ranges. 3896 */ 3897void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3898 unsigned long end_pfn) 3899{ 3900 int i; 3901 3902 mminit_dprintk(MMINIT_TRACE, "memory_register", 3903 "Entering add_active_range(%d, %#lx, %#lx) " 3904 "%d entries of %d used\n", 3905 nid, start_pfn, end_pfn, 3906 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3907 3908 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 3909 3910 /* Merge with existing active regions if possible */ 3911 for (i = 0; i < nr_nodemap_entries; i++) { 3912 if (early_node_map[i].nid != nid) 3913 continue; 3914 3915 /* Skip if an existing region covers this new one */ 3916 if (start_pfn >= early_node_map[i].start_pfn && 3917 end_pfn <= early_node_map[i].end_pfn) 3918 return; 3919 3920 /* Merge forward if suitable */ 3921 if (start_pfn <= early_node_map[i].end_pfn && 3922 end_pfn > early_node_map[i].end_pfn) { 3923 early_node_map[i].end_pfn = end_pfn; 3924 return; 3925 } 3926 3927 /* Merge backward if suitable */ 3928 if (start_pfn < early_node_map[i].end_pfn && 3929 end_pfn >= early_node_map[i].start_pfn) { 3930 early_node_map[i].start_pfn = start_pfn; 3931 return; 3932 } 3933 } 3934 3935 /* Check that early_node_map is large enough */ 3936 if (i >= MAX_ACTIVE_REGIONS) { 3937 printk(KERN_CRIT "More than %d memory regions, truncating\n", 3938 MAX_ACTIVE_REGIONS); 3939 return; 3940 } 3941 3942 early_node_map[i].nid = nid; 3943 early_node_map[i].start_pfn = start_pfn; 3944 early_node_map[i].end_pfn = end_pfn; 3945 nr_nodemap_entries = i + 1; 3946} 3947 3948/** 3949 * remove_active_range - Shrink an existing registered range of PFNs 3950 * @nid: The node id the range is on that should be shrunk 3951 * @start_pfn: The new PFN of the range 3952 * @end_pfn: The new PFN of the range 3953 * 3954 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 3955 * The map is kept near the end physical page range that has already been 3956 * registered. This function allows an arch to shrink an existing registered 3957 * range. 3958 */ 3959void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 3960 unsigned long end_pfn) 3961{ 3962 int i, j; 3963 int removed = 0; 3964 3965 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 3966 nid, start_pfn, end_pfn); 3967 3968 /* Find the old active region end and shrink */ 3969 for_each_active_range_index_in_nid(i, nid) { 3970 if (early_node_map[i].start_pfn >= start_pfn && 3971 early_node_map[i].end_pfn <= end_pfn) { 3972 /* clear it */ 3973 early_node_map[i].start_pfn = 0; 3974 early_node_map[i].end_pfn = 0; 3975 removed = 1; 3976 continue; 3977 } 3978 if (early_node_map[i].start_pfn < start_pfn && 3979 early_node_map[i].end_pfn > start_pfn) { 3980 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 3981 early_node_map[i].end_pfn = start_pfn; 3982 if (temp_end_pfn > end_pfn) 3983 add_active_range(nid, end_pfn, temp_end_pfn); 3984 continue; 3985 } 3986 if (early_node_map[i].start_pfn >= start_pfn && 3987 early_node_map[i].end_pfn > end_pfn && 3988 early_node_map[i].start_pfn < end_pfn) { 3989 early_node_map[i].start_pfn = end_pfn; 3990 continue; 3991 } 3992 } 3993 3994 if (!removed) 3995 return; 3996 3997 /* remove the blank ones */ 3998 for (i = nr_nodemap_entries - 1; i > 0; i--) { 3999 if (early_node_map[i].nid != nid) 4000 continue; 4001 if (early_node_map[i].end_pfn) 4002 continue; 4003 /* we found it, get rid of it */ 4004 for (j = i; j < nr_nodemap_entries - 1; j++) 4005 memcpy(&early_node_map[j], &early_node_map[j+1], 4006 sizeof(early_node_map[j])); 4007 j = nr_nodemap_entries - 1; 4008 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4009 nr_nodemap_entries--; 4010 } 4011} 4012 4013/** 4014 * remove_all_active_ranges - Remove all currently registered regions 4015 * 4016 * During discovery, it may be found that a table like SRAT is invalid 4017 * and an alternative discovery method must be used. This function removes 4018 * all currently registered regions. 4019 */ 4020void __init remove_all_active_ranges(void) 4021{ 4022 memset(early_node_map, 0, sizeof(early_node_map)); 4023 nr_nodemap_entries = 0; 4024} 4025 4026/* Compare two active node_active_regions */ 4027static int __init cmp_node_active_region(const void *a, const void *b) 4028{ 4029 struct node_active_region *arange = (struct node_active_region *)a; 4030 struct node_active_region *brange = (struct node_active_region *)b; 4031 4032 /* Done this way to avoid overflows */ 4033 if (arange->start_pfn > brange->start_pfn) 4034 return 1; 4035 if (arange->start_pfn < brange->start_pfn) 4036 return -1; 4037 4038 return 0; 4039} 4040 4041/* sort the node_map by start_pfn */ 4042static void __init sort_node_map(void) 4043{ 4044 sort(early_node_map, (size_t)nr_nodemap_entries, 4045 sizeof(struct node_active_region), 4046 cmp_node_active_region, NULL); 4047} 4048 4049/* Find the lowest pfn for a node */ 4050static unsigned long __init find_min_pfn_for_node(int nid) 4051{ 4052 int i; 4053 unsigned long min_pfn = ULONG_MAX; 4054 4055 /* Assuming a sorted map, the first range found has the starting pfn */ 4056 for_each_active_range_index_in_nid(i, nid) 4057 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4058 4059 if (min_pfn == ULONG_MAX) { 4060 printk(KERN_WARNING 4061 "Could not find start_pfn for node %d\n", nid); 4062 return 0; 4063 } 4064 4065 return min_pfn; 4066} 4067 4068/** 4069 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4070 * 4071 * It returns the minimum PFN based on information provided via 4072 * add_active_range(). 4073 */ 4074unsigned long __init find_min_pfn_with_active_regions(void) 4075{ 4076 return find_min_pfn_for_node(MAX_NUMNODES); 4077} 4078 4079/* 4080 * early_calculate_totalpages() 4081 * Sum pages in active regions for movable zone. 4082 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4083 */ 4084static unsigned long __init early_calculate_totalpages(void) 4085{ 4086 int i; 4087 unsigned long totalpages = 0; 4088 4089 for (i = 0; i < nr_nodemap_entries; i++) { 4090 unsigned long pages = early_node_map[i].end_pfn - 4091 early_node_map[i].start_pfn; 4092 totalpages += pages; 4093 if (pages) 4094 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4095 } 4096 return totalpages; 4097} 4098 4099/* 4100 * Find the PFN the Movable zone begins in each node. Kernel memory 4101 * is spread evenly between nodes as long as the nodes have enough 4102 * memory. When they don't, some nodes will have more kernelcore than 4103 * others 4104 */ 4105static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4106{ 4107 int i, nid; 4108 unsigned long usable_startpfn; 4109 unsigned long kernelcore_node, kernelcore_remaining; 4110 /* save the state before borrow the nodemask */ 4111 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4112 unsigned long totalpages = early_calculate_totalpages(); 4113 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4114 4115 /* 4116 * If movablecore was specified, calculate what size of 4117 * kernelcore that corresponds so that memory usable for 4118 * any allocation type is evenly spread. If both kernelcore 4119 * and movablecore are specified, then the value of kernelcore 4120 * will be used for required_kernelcore if it's greater than 4121 * what movablecore would have allowed. 4122 */ 4123 if (required_movablecore) { 4124 unsigned long corepages; 4125 4126 /* 4127 * Round-up so that ZONE_MOVABLE is at least as large as what 4128 * was requested by the user 4129 */ 4130 required_movablecore = 4131 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4132 corepages = totalpages - required_movablecore; 4133 4134 required_kernelcore = max(required_kernelcore, corepages); 4135 } 4136 4137 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4138 if (!required_kernelcore) 4139 goto out; 4140 4141 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4142 find_usable_zone_for_movable(); 4143 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4144 4145restart: 4146 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4147 kernelcore_node = required_kernelcore / usable_nodes; 4148 for_each_node_state(nid, N_HIGH_MEMORY) { 4149 /* 4150 * Recalculate kernelcore_node if the division per node 4151 * now exceeds what is necessary to satisfy the requested 4152 * amount of memory for the kernel 4153 */ 4154 if (required_kernelcore < kernelcore_node) 4155 kernelcore_node = required_kernelcore / usable_nodes; 4156 4157 /* 4158 * As the map is walked, we track how much memory is usable 4159 * by the kernel using kernelcore_remaining. When it is 4160 * 0, the rest of the node is usable by ZONE_MOVABLE 4161 */ 4162 kernelcore_remaining = kernelcore_node; 4163 4164 /* Go through each range of PFNs within this node */ 4165 for_each_active_range_index_in_nid(i, nid) { 4166 unsigned long start_pfn, end_pfn; 4167 unsigned long size_pages; 4168 4169 start_pfn = max(early_node_map[i].start_pfn, 4170 zone_movable_pfn[nid]); 4171 end_pfn = early_node_map[i].end_pfn; 4172 if (start_pfn >= end_pfn) 4173 continue; 4174 4175 /* Account for what is only usable for kernelcore */ 4176 if (start_pfn < usable_startpfn) { 4177 unsigned long kernel_pages; 4178 kernel_pages = min(end_pfn, usable_startpfn) 4179 - start_pfn; 4180 4181 kernelcore_remaining -= min(kernel_pages, 4182 kernelcore_remaining); 4183 required_kernelcore -= min(kernel_pages, 4184 required_kernelcore); 4185 4186 /* Continue if range is now fully accounted */ 4187 if (end_pfn <= usable_startpfn) { 4188 4189 /* 4190 * Push zone_movable_pfn to the end so 4191 * that if we have to rebalance 4192 * kernelcore across nodes, we will 4193 * not double account here 4194 */ 4195 zone_movable_pfn[nid] = end_pfn; 4196 continue; 4197 } 4198 start_pfn = usable_startpfn; 4199 } 4200 4201 /* 4202 * The usable PFN range for ZONE_MOVABLE is from 4203 * start_pfn->end_pfn. Calculate size_pages as the 4204 * number of pages used as kernelcore 4205 */ 4206 size_pages = end_pfn - start_pfn; 4207 if (size_pages > kernelcore_remaining) 4208 size_pages = kernelcore_remaining; 4209 zone_movable_pfn[nid] = start_pfn + size_pages; 4210 4211 /* 4212 * Some kernelcore has been met, update counts and 4213 * break if the kernelcore for this node has been 4214 * satisified 4215 */ 4216 required_kernelcore -= min(required_kernelcore, 4217 size_pages); 4218 kernelcore_remaining -= size_pages; 4219 if (!kernelcore_remaining) 4220 break; 4221 } 4222 } 4223 4224 /* 4225 * If there is still required_kernelcore, we do another pass with one 4226 * less node in the count. This will push zone_movable_pfn[nid] further 4227 * along on the nodes that still have memory until kernelcore is 4228 * satisified 4229 */ 4230 usable_nodes--; 4231 if (usable_nodes && required_kernelcore > usable_nodes) 4232 goto restart; 4233 4234 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4235 for (nid = 0; nid < MAX_NUMNODES; nid++) 4236 zone_movable_pfn[nid] = 4237 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4238 4239out: 4240 /* restore the node_state */ 4241 node_states[N_HIGH_MEMORY] = saved_node_state; 4242} 4243 4244/* Any regular memory on that node ? */ 4245static void check_for_regular_memory(pg_data_t *pgdat) 4246{ 4247#ifdef CONFIG_HIGHMEM 4248 enum zone_type zone_type; 4249 4250 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4251 struct zone *zone = &pgdat->node_zones[zone_type]; 4252 if (zone->present_pages) 4253 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4254 } 4255#endif 4256} 4257 4258/** 4259 * free_area_init_nodes - Initialise all pg_data_t and zone data 4260 * @max_zone_pfn: an array of max PFNs for each zone 4261 * 4262 * This will call free_area_init_node() for each active node in the system. 4263 * Using the page ranges provided by add_active_range(), the size of each 4264 * zone in each node and their holes is calculated. If the maximum PFN 4265 * between two adjacent zones match, it is assumed that the zone is empty. 4266 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4267 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4268 * starts where the previous one ended. For example, ZONE_DMA32 starts 4269 * at arch_max_dma_pfn. 4270 */ 4271void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4272{ 4273 unsigned long nid; 4274 int i; 4275 4276 /* Sort early_node_map as initialisation assumes it is sorted */ 4277 sort_node_map(); 4278 4279 /* Record where the zone boundaries are */ 4280 memset(arch_zone_lowest_possible_pfn, 0, 4281 sizeof(arch_zone_lowest_possible_pfn)); 4282 memset(arch_zone_highest_possible_pfn, 0, 4283 sizeof(arch_zone_highest_possible_pfn)); 4284 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4285 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4286 for (i = 1; i < MAX_NR_ZONES; i++) { 4287 if (i == ZONE_MOVABLE) 4288 continue; 4289 arch_zone_lowest_possible_pfn[i] = 4290 arch_zone_highest_possible_pfn[i-1]; 4291 arch_zone_highest_possible_pfn[i] = 4292 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4293 } 4294 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4295 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4296 4297 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4298 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4299 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4300 4301 /* Print out the zone ranges */ 4302 printk("Zone PFN ranges:\n"); 4303 for (i = 0; i < MAX_NR_ZONES; i++) { 4304 if (i == ZONE_MOVABLE) 4305 continue; 4306 printk(" %-8s %0#10lx -> %0#10lx\n", 4307 zone_names[i], 4308 arch_zone_lowest_possible_pfn[i], 4309 arch_zone_highest_possible_pfn[i]); 4310 } 4311 4312 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4313 printk("Movable zone start PFN for each node\n"); 4314 for (i = 0; i < MAX_NUMNODES; i++) { 4315 if (zone_movable_pfn[i]) 4316 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4317 } 4318 4319 /* Print out the early_node_map[] */ 4320 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4321 for (i = 0; i < nr_nodemap_entries; i++) 4322 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4323 early_node_map[i].start_pfn, 4324 early_node_map[i].end_pfn); 4325 4326 /* Initialise every node */ 4327 mminit_verify_pageflags_layout(); 4328 setup_nr_node_ids(); 4329 for_each_online_node(nid) { 4330 pg_data_t *pgdat = NODE_DATA(nid); 4331 free_area_init_node(nid, NULL, 4332 find_min_pfn_for_node(nid), NULL); 4333 4334 /* Any memory on that node */ 4335 if (pgdat->node_present_pages) 4336 node_set_state(nid, N_HIGH_MEMORY); 4337 check_for_regular_memory(pgdat); 4338 } 4339} 4340 4341static int __init cmdline_parse_core(char *p, unsigned long *core) 4342{ 4343 unsigned long long coremem; 4344 if (!p) 4345 return -EINVAL; 4346 4347 coremem = memparse(p, &p); 4348 *core = coremem >> PAGE_SHIFT; 4349 4350 /* Paranoid check that UL is enough for the coremem value */ 4351 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4352 4353 return 0; 4354} 4355 4356/* 4357 * kernelcore=size sets the amount of memory for use for allocations that 4358 * cannot be reclaimed or migrated. 4359 */ 4360static int __init cmdline_parse_kernelcore(char *p) 4361{ 4362 return cmdline_parse_core(p, &required_kernelcore); 4363} 4364 4365/* 4366 * movablecore=size sets the amount of memory for use for allocations that 4367 * can be reclaimed or migrated. 4368 */ 4369static int __init cmdline_parse_movablecore(char *p) 4370{ 4371 return cmdline_parse_core(p, &required_movablecore); 4372} 4373 4374early_param("kernelcore", cmdline_parse_kernelcore); 4375early_param("movablecore", cmdline_parse_movablecore); 4376 4377#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4378 4379/** 4380 * set_dma_reserve - set the specified number of pages reserved in the first zone 4381 * @new_dma_reserve: The number of pages to mark reserved 4382 * 4383 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4384 * In the DMA zone, a significant percentage may be consumed by kernel image 4385 * and other unfreeable allocations which can skew the watermarks badly. This 4386 * function may optionally be used to account for unfreeable pages in the 4387 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4388 * smaller per-cpu batchsize. 4389 */ 4390void __init set_dma_reserve(unsigned long new_dma_reserve) 4391{ 4392 dma_reserve = new_dma_reserve; 4393} 4394 4395#ifndef CONFIG_NEED_MULTIPLE_NODES 4396struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; 4397EXPORT_SYMBOL(contig_page_data); 4398#endif 4399 4400void __init free_area_init(unsigned long *zones_size) 4401{ 4402 free_area_init_node(0, zones_size, 4403 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4404} 4405 4406static int page_alloc_cpu_notify(struct notifier_block *self, 4407 unsigned long action, void *hcpu) 4408{ 4409 int cpu = (unsigned long)hcpu; 4410 4411 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4412 drain_pages(cpu); 4413 4414 /* 4415 * Spill the event counters of the dead processor 4416 * into the current processors event counters. 4417 * This artificially elevates the count of the current 4418 * processor. 4419 */ 4420 vm_events_fold_cpu(cpu); 4421 4422 /* 4423 * Zero the differential counters of the dead processor 4424 * so that the vm statistics are consistent. 4425 * 4426 * This is only okay since the processor is dead and cannot 4427 * race with what we are doing. 4428 */ 4429 refresh_cpu_vm_stats(cpu); 4430 } 4431 return NOTIFY_OK; 4432} 4433 4434void __init page_alloc_init(void) 4435{ 4436 hotcpu_notifier(page_alloc_cpu_notify, 0); 4437} 4438 4439/* 4440 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4441 * or min_free_kbytes changes. 4442 */ 4443static void calculate_totalreserve_pages(void) 4444{ 4445 struct pglist_data *pgdat; 4446 unsigned long reserve_pages = 0; 4447 enum zone_type i, j; 4448 4449 for_each_online_pgdat(pgdat) { 4450 for (i = 0; i < MAX_NR_ZONES; i++) { 4451 struct zone *zone = pgdat->node_zones + i; 4452 unsigned long max = 0; 4453 4454 /* Find valid and maximum lowmem_reserve in the zone */ 4455 for (j = i; j < MAX_NR_ZONES; j++) { 4456 if (zone->lowmem_reserve[j] > max) 4457 max = zone->lowmem_reserve[j]; 4458 } 4459 4460 /* we treat the high watermark as reserved pages. */ 4461 max += high_wmark_pages(zone); 4462 4463 if (max > zone->present_pages) 4464 max = zone->present_pages; 4465 reserve_pages += max; 4466 } 4467 } 4468 totalreserve_pages = reserve_pages; 4469} 4470 4471/* 4472 * setup_per_zone_lowmem_reserve - called whenever 4473 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4474 * has a correct pages reserved value, so an adequate number of 4475 * pages are left in the zone after a successful __alloc_pages(). 4476 */ 4477static void setup_per_zone_lowmem_reserve(void) 4478{ 4479 struct pglist_data *pgdat; 4480 enum zone_type j, idx; 4481 4482 for_each_online_pgdat(pgdat) { 4483 for (j = 0; j < MAX_NR_ZONES; j++) { 4484 struct zone *zone = pgdat->node_zones + j; 4485 unsigned long present_pages = zone->present_pages; 4486 4487 zone->lowmem_reserve[j] = 0; 4488 4489 idx = j; 4490 while (idx) { 4491 struct zone *lower_zone; 4492 4493 idx--; 4494 4495 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4496 sysctl_lowmem_reserve_ratio[idx] = 1; 4497 4498 lower_zone = pgdat->node_zones + idx; 4499 lower_zone->lowmem_reserve[j] = present_pages / 4500 sysctl_lowmem_reserve_ratio[idx]; 4501 present_pages += lower_zone->present_pages; 4502 } 4503 } 4504 } 4505 4506 /* update totalreserve_pages */ 4507 calculate_totalreserve_pages(); 4508} 4509 4510/** 4511 * setup_per_zone_wmarks - called when min_free_kbytes changes 4512 * or when memory is hot-{added|removed} 4513 * 4514 * Ensures that the watermark[min,low,high] values for each zone are set 4515 * correctly with respect to min_free_kbytes. 4516 */ 4517void setup_per_zone_wmarks(void) 4518{ 4519 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4520 unsigned long lowmem_pages = 0; 4521 struct zone *zone; 4522 unsigned long flags; 4523 4524 /* Calculate total number of !ZONE_HIGHMEM pages */ 4525 for_each_zone(zone) { 4526 if (!is_highmem(zone)) 4527 lowmem_pages += zone->present_pages; 4528 } 4529 4530 for_each_zone(zone) { 4531 u64 tmp; 4532 4533 spin_lock_irqsave(&zone->lock, flags); 4534 tmp = (u64)pages_min * zone->present_pages; 4535 do_div(tmp, lowmem_pages); 4536 if (is_highmem(zone)) { 4537 /* 4538 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4539 * need highmem pages, so cap pages_min to a small 4540 * value here. 4541 * 4542 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4543 * deltas controls asynch page reclaim, and so should 4544 * not be capped for highmem. 4545 */ 4546 int min_pages; 4547 4548 min_pages = zone->present_pages / 1024; 4549 if (min_pages < SWAP_CLUSTER_MAX) 4550 min_pages = SWAP_CLUSTER_MAX; 4551 if (min_pages > 128) 4552 min_pages = 128; 4553 zone->watermark[WMARK_MIN] = min_pages; 4554 } else { 4555 /* 4556 * If it's a lowmem zone, reserve a number of pages 4557 * proportionate to the zone's size. 4558 */ 4559 zone->watermark[WMARK_MIN] = tmp; 4560 } 4561 4562 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4563 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4564 setup_zone_migrate_reserve(zone); 4565 spin_unlock_irqrestore(&zone->lock, flags); 4566 } 4567 4568 /* update totalreserve_pages */ 4569 calculate_totalreserve_pages(); 4570} 4571 4572/* 4573 * The inactive anon list should be small enough that the VM never has to 4574 * do too much work, but large enough that each inactive page has a chance 4575 * to be referenced again before it is swapped out. 4576 * 4577 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4578 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4579 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4580 * the anonymous pages are kept on the inactive list. 4581 * 4582 * total target max 4583 * memory ratio inactive anon 4584 * ------------------------------------- 4585 * 10MB 1 5MB 4586 * 100MB 1 50MB 4587 * 1GB 3 250MB 4588 * 10GB 10 0.9GB 4589 * 100GB 31 3GB 4590 * 1TB 101 10GB 4591 * 10TB 320 32GB 4592 */ 4593void calculate_zone_inactive_ratio(struct zone *zone) 4594{ 4595 unsigned int gb, ratio; 4596 4597 /* Zone size in gigabytes */ 4598 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4599 if (gb) 4600 ratio = int_sqrt(10 * gb); 4601 else 4602 ratio = 1; 4603 4604 zone->inactive_ratio = ratio; 4605} 4606 4607static void __init setup_per_zone_inactive_ratio(void) 4608{ 4609 struct zone *zone; 4610 4611 for_each_zone(zone) 4612 calculate_zone_inactive_ratio(zone); 4613} 4614 4615/* 4616 * Initialise min_free_kbytes. 4617 * 4618 * For small machines we want it small (128k min). For large machines 4619 * we want it large (64MB max). But it is not linear, because network 4620 * bandwidth does not increase linearly with machine size. We use 4621 * 4622 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4623 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4624 * 4625 * which yields 4626 * 4627 * 16MB: 512k 4628 * 32MB: 724k 4629 * 64MB: 1024k 4630 * 128MB: 1448k 4631 * 256MB: 2048k 4632 * 512MB: 2896k 4633 * 1024MB: 4096k 4634 * 2048MB: 5792k 4635 * 4096MB: 8192k 4636 * 8192MB: 11584k 4637 * 16384MB: 16384k 4638 */ 4639static int __init init_per_zone_wmark_min(void) 4640{ 4641 unsigned long lowmem_kbytes; 4642 4643 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4644 4645 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4646 if (min_free_kbytes < 128) 4647 min_free_kbytes = 128; 4648 if (min_free_kbytes > 65536) 4649 min_free_kbytes = 65536; 4650 setup_per_zone_wmarks(); 4651 setup_per_zone_lowmem_reserve(); 4652 setup_per_zone_inactive_ratio(); 4653 return 0; 4654} 4655module_init(init_per_zone_wmark_min) 4656 4657/* 4658 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4659 * that we can call two helper functions whenever min_free_kbytes 4660 * changes. 4661 */ 4662int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4663 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4664{ 4665 proc_dointvec(table, write, file, buffer, length, ppos); 4666 if (write) 4667 setup_per_zone_wmarks(); 4668 return 0; 4669} 4670 4671#ifdef CONFIG_NUMA 4672int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4673 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4674{ 4675 struct zone *zone; 4676 int rc; 4677 4678 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4679 if (rc) 4680 return rc; 4681 4682 for_each_zone(zone) 4683 zone->min_unmapped_pages = (zone->present_pages * 4684 sysctl_min_unmapped_ratio) / 100; 4685 return 0; 4686} 4687 4688int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4689 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4690{ 4691 struct zone *zone; 4692 int rc; 4693 4694 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4695 if (rc) 4696 return rc; 4697 4698 for_each_zone(zone) 4699 zone->min_slab_pages = (zone->present_pages * 4700 sysctl_min_slab_ratio) / 100; 4701 return 0; 4702} 4703#endif 4704 4705/* 4706 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4707 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4708 * whenever sysctl_lowmem_reserve_ratio changes. 4709 * 4710 * The reserve ratio obviously has absolutely no relation with the 4711 * minimum watermarks. The lowmem reserve ratio can only make sense 4712 * if in function of the boot time zone sizes. 4713 */ 4714int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4715 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4716{ 4717 proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4718 setup_per_zone_lowmem_reserve(); 4719 return 0; 4720} 4721 4722/* 4723 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4724 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4725 * can have before it gets flushed back to buddy allocator. 4726 */ 4727 4728int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4729 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4730{ 4731 struct zone *zone; 4732 unsigned int cpu; 4733 int ret; 4734 4735 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4736 if (!write || (ret == -EINVAL)) 4737 return ret; 4738 for_each_populated_zone(zone) { 4739 for_each_online_cpu(cpu) { 4740 unsigned long high; 4741 high = zone->present_pages / percpu_pagelist_fraction; 4742 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4743 } 4744 } 4745 return 0; 4746} 4747 4748int hashdist = HASHDIST_DEFAULT; 4749 4750#ifdef CONFIG_NUMA 4751static int __init set_hashdist(char *str) 4752{ 4753 if (!str) 4754 return 0; 4755 hashdist = simple_strtoul(str, &str, 0); 4756 return 1; 4757} 4758__setup("hashdist=", set_hashdist); 4759#endif 4760 4761/* 4762 * allocate a large system hash table from bootmem 4763 * - it is assumed that the hash table must contain an exact power-of-2 4764 * quantity of entries 4765 * - limit is the number of hash buckets, not the total allocation size 4766 */ 4767void *__init alloc_large_system_hash(const char *tablename, 4768 unsigned long bucketsize, 4769 unsigned long numentries, 4770 int scale, 4771 int flags, 4772 unsigned int *_hash_shift, 4773 unsigned int *_hash_mask, 4774 unsigned long limit) 4775{ 4776 unsigned long long max = limit; 4777 unsigned long log2qty, size; 4778 void *table = NULL; 4779 4780 /* allow the kernel cmdline to have a say */ 4781 if (!numentries) { 4782 /* round applicable memory size up to nearest megabyte */ 4783 numentries = nr_kernel_pages; 4784 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4785 numentries >>= 20 - PAGE_SHIFT; 4786 numentries <<= 20 - PAGE_SHIFT; 4787 4788 /* limit to 1 bucket per 2^scale bytes of low memory */ 4789 if (scale > PAGE_SHIFT) 4790 numentries >>= (scale - PAGE_SHIFT); 4791 else 4792 numentries <<= (PAGE_SHIFT - scale); 4793 4794 /* Make sure we've got at least a 0-order allocation.. */ 4795 if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4796 numentries = PAGE_SIZE / bucketsize; 4797 } 4798 numentries = roundup_pow_of_two(numentries); 4799 4800 /* limit allocation size to 1/16 total memory by default */ 4801 if (max == 0) { 4802 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4803 do_div(max, bucketsize); 4804 } 4805 4806 if (numentries > max) 4807 numentries = max; 4808 4809 log2qty = ilog2(numentries); 4810 4811 do { 4812 size = bucketsize << log2qty; 4813 if (flags & HASH_EARLY) 4814 table = alloc_bootmem_nopanic(size); 4815 else if (hashdist) 4816 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4817 else { 4818 /* 4819 * If bucketsize is not a power-of-two, we may free 4820 * some pages at the end of hash table which 4821 * alloc_pages_exact() automatically does 4822 */ 4823 if (get_order(size) < MAX_ORDER) { 4824 table = alloc_pages_exact(size, GFP_ATOMIC); 4825 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4826 } 4827 } 4828 } while (!table && size > PAGE_SIZE && --log2qty); 4829 4830 if (!table) 4831 panic("Failed to allocate %s hash table\n", tablename); 4832 4833 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4834 tablename, 4835 (1U << log2qty), 4836 ilog2(size) - PAGE_SHIFT, 4837 size); 4838 4839 if (_hash_shift) 4840 *_hash_shift = log2qty; 4841 if (_hash_mask) 4842 *_hash_mask = (1 << log2qty) - 1; 4843 4844 return table; 4845} 4846 4847/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4848static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4849 unsigned long pfn) 4850{ 4851#ifdef CONFIG_SPARSEMEM 4852 return __pfn_to_section(pfn)->pageblock_flags; 4853#else 4854 return zone->pageblock_flags; 4855#endif /* CONFIG_SPARSEMEM */ 4856} 4857 4858static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4859{ 4860#ifdef CONFIG_SPARSEMEM 4861 pfn &= (PAGES_PER_SECTION-1); 4862 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4863#else 4864 pfn = pfn - zone->zone_start_pfn; 4865 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4866#endif /* CONFIG_SPARSEMEM */ 4867} 4868 4869/** 4870 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4871 * @page: The page within the block of interest 4872 * @start_bitidx: The first bit of interest to retrieve 4873 * @end_bitidx: The last bit of interest 4874 * returns pageblock_bits flags 4875 */ 4876unsigned long get_pageblock_flags_group(struct page *page, 4877 int start_bitidx, int end_bitidx) 4878{ 4879 struct zone *zone; 4880 unsigned long *bitmap; 4881 unsigned long pfn, bitidx; 4882 unsigned long flags = 0; 4883 unsigned long value = 1; 4884 4885 zone = page_zone(page); 4886 pfn = page_to_pfn(page); 4887 bitmap = get_pageblock_bitmap(zone, pfn); 4888 bitidx = pfn_to_bitidx(zone, pfn); 4889 4890 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4891 if (test_bit(bitidx + start_bitidx, bitmap)) 4892 flags |= value; 4893 4894 return flags; 4895} 4896 4897/** 4898 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 4899 * @page: The page within the block of interest 4900 * @start_bitidx: The first bit of interest 4901 * @end_bitidx: The last bit of interest 4902 * @flags: The flags to set 4903 */ 4904void set_pageblock_flags_group(struct page *page, unsigned long flags, 4905 int start_bitidx, int end_bitidx) 4906{ 4907 struct zone *zone; 4908 unsigned long *bitmap; 4909 unsigned long pfn, bitidx; 4910 unsigned long value = 1; 4911 4912 zone = page_zone(page); 4913 pfn = page_to_pfn(page); 4914 bitmap = get_pageblock_bitmap(zone, pfn); 4915 bitidx = pfn_to_bitidx(zone, pfn); 4916 VM_BUG_ON(pfn < zone->zone_start_pfn); 4917 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 4918 4919 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4920 if (flags & value) 4921 __set_bit(bitidx + start_bitidx, bitmap); 4922 else 4923 __clear_bit(bitidx + start_bitidx, bitmap); 4924} 4925 4926/* 4927 * This is designed as sub function...plz see page_isolation.c also. 4928 * set/clear page block's type to be ISOLATE. 4929 * page allocater never alloc memory from ISOLATE block. 4930 */ 4931 4932int set_migratetype_isolate(struct page *page) 4933{ 4934 struct zone *zone; 4935 unsigned long flags; 4936 int ret = -EBUSY; 4937 int zone_idx; 4938 4939 zone = page_zone(page); 4940 zone_idx = zone_idx(zone); 4941 spin_lock_irqsave(&zone->lock, flags); 4942 /* 4943 * In future, more migrate types will be able to be isolation target. 4944 */ 4945 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE && 4946 zone_idx != ZONE_MOVABLE) 4947 goto out; 4948 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 4949 move_freepages_block(zone, page, MIGRATE_ISOLATE); 4950 ret = 0; 4951out: 4952 spin_unlock_irqrestore(&zone->lock, flags); 4953 if (!ret) 4954 drain_all_pages(); 4955 return ret; 4956} 4957 4958void unset_migratetype_isolate(struct page *page) 4959{ 4960 struct zone *zone; 4961 unsigned long flags; 4962 zone = page_zone(page); 4963 spin_lock_irqsave(&zone->lock, flags); 4964 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 4965 goto out; 4966 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4967 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4968out: 4969 spin_unlock_irqrestore(&zone->lock, flags); 4970} 4971 4972#ifdef CONFIG_MEMORY_HOTREMOVE 4973/* 4974 * All pages in the range must be isolated before calling this. 4975 */ 4976void 4977__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 4978{ 4979 struct page *page; 4980 struct zone *zone; 4981 int order, i; 4982 unsigned long pfn; 4983 unsigned long flags; 4984 /* find the first valid pfn */ 4985 for (pfn = start_pfn; pfn < end_pfn; pfn++) 4986 if (pfn_valid(pfn)) 4987 break; 4988 if (pfn == end_pfn) 4989 return; 4990 zone = page_zone(pfn_to_page(pfn)); 4991 spin_lock_irqsave(&zone->lock, flags); 4992 pfn = start_pfn; 4993 while (pfn < end_pfn) { 4994 if (!pfn_valid(pfn)) { 4995 pfn++; 4996 continue; 4997 } 4998 page = pfn_to_page(pfn); 4999 BUG_ON(page_count(page)); 5000 BUG_ON(!PageBuddy(page)); 5001 order = page_order(page); 5002#ifdef CONFIG_DEBUG_VM 5003 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5004 pfn, 1 << order, end_pfn); 5005#endif 5006 list_del(&page->lru); 5007 rmv_page_order(page); 5008 zone->free_area[order].nr_free--; 5009 __mod_zone_page_state(zone, NR_FREE_PAGES, 5010 - (1UL << order)); 5011 for (i = 0; i < (1 << order); i++) 5012 SetPageReserved((page+i)); 5013 pfn += (1 << order); 5014 } 5015 spin_unlock_irqrestore(&zone->lock, flags); 5016} 5017#endif 5018