page_alloc.c revision dab48dab37d2770824420d1e01730a107fade1aa
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/module.h> 27#include <linux/suspend.h> 28#include <linux/pagevec.h> 29#include <linux/blkdev.h> 30#include <linux/slab.h> 31#include <linux/oom.h> 32#include <linux/notifier.h> 33#include <linux/topology.h> 34#include <linux/sysctl.h> 35#include <linux/cpu.h> 36#include <linux/cpuset.h> 37#include <linux/memory_hotplug.h> 38#include <linux/nodemask.h> 39#include <linux/vmalloc.h> 40#include <linux/mempolicy.h> 41#include <linux/stop_machine.h> 42#include <linux/sort.h> 43#include <linux/pfn.h> 44#include <linux/backing-dev.h> 45#include <linux/fault-inject.h> 46#include <linux/page-isolation.h> 47#include <linux/page_cgroup.h> 48#include <linux/debugobjects.h> 49#include <linux/kmemleak.h> 50 51#include <asm/tlbflush.h> 52#include <asm/div64.h> 53#include "internal.h" 54 55/* 56 * Array of node states. 57 */ 58nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 59 [N_POSSIBLE] = NODE_MASK_ALL, 60 [N_ONLINE] = { { [0] = 1UL } }, 61#ifndef CONFIG_NUMA 62 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 63#ifdef CONFIG_HIGHMEM 64 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 65#endif 66 [N_CPU] = { { [0] = 1UL } }, 67#endif /* NUMA */ 68}; 69EXPORT_SYMBOL(node_states); 70 71unsigned long totalram_pages __read_mostly; 72unsigned long totalreserve_pages __read_mostly; 73unsigned long highest_memmap_pfn __read_mostly; 74int percpu_pagelist_fraction; 75 76#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 77int pageblock_order __read_mostly; 78#endif 79 80static void __free_pages_ok(struct page *page, unsigned int order); 81 82/* 83 * results with 256, 32 in the lowmem_reserve sysctl: 84 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 85 * 1G machine -> (16M dma, 784M normal, 224M high) 86 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 87 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 88 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 89 * 90 * TBD: should special case ZONE_DMA32 machines here - in those we normally 91 * don't need any ZONE_NORMAL reservation 92 */ 93int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 94#ifdef CONFIG_ZONE_DMA 95 256, 96#endif 97#ifdef CONFIG_ZONE_DMA32 98 256, 99#endif 100#ifdef CONFIG_HIGHMEM 101 32, 102#endif 103 32, 104}; 105 106EXPORT_SYMBOL(totalram_pages); 107 108static char * const zone_names[MAX_NR_ZONES] = { 109#ifdef CONFIG_ZONE_DMA 110 "DMA", 111#endif 112#ifdef CONFIG_ZONE_DMA32 113 "DMA32", 114#endif 115 "Normal", 116#ifdef CONFIG_HIGHMEM 117 "HighMem", 118#endif 119 "Movable", 120}; 121 122int min_free_kbytes = 1024; 123 124unsigned long __meminitdata nr_kernel_pages; 125unsigned long __meminitdata nr_all_pages; 126static unsigned long __meminitdata dma_reserve; 127 128#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 129 /* 130 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 131 * ranges of memory (RAM) that may be registered with add_active_range(). 132 * Ranges passed to add_active_range() will be merged if possible 133 * so the number of times add_active_range() can be called is 134 * related to the number of nodes and the number of holes 135 */ 136 #ifdef CONFIG_MAX_ACTIVE_REGIONS 137 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 138 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 139 #else 140 #if MAX_NUMNODES >= 32 141 /* If there can be many nodes, allow up to 50 holes per node */ 142 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 143 #else 144 /* By default, allow up to 256 distinct regions */ 145 #define MAX_ACTIVE_REGIONS 256 146 #endif 147 #endif 148 149 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 150 static int __meminitdata nr_nodemap_entries; 151 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 152 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 153 static unsigned long __initdata required_kernelcore; 154 static unsigned long __initdata required_movablecore; 155 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 156 157 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 158 int movable_zone; 159 EXPORT_SYMBOL(movable_zone); 160#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 161 162#if MAX_NUMNODES > 1 163int nr_node_ids __read_mostly = MAX_NUMNODES; 164int nr_online_nodes __read_mostly = 1; 165EXPORT_SYMBOL(nr_node_ids); 166EXPORT_SYMBOL(nr_online_nodes); 167#endif 168 169int page_group_by_mobility_disabled __read_mostly; 170 171static void set_pageblock_migratetype(struct page *page, int migratetype) 172{ 173 174 if (unlikely(page_group_by_mobility_disabled)) 175 migratetype = MIGRATE_UNMOVABLE; 176 177 set_pageblock_flags_group(page, (unsigned long)migratetype, 178 PB_migrate, PB_migrate_end); 179} 180 181#ifdef CONFIG_DEBUG_VM 182static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 183{ 184 int ret = 0; 185 unsigned seq; 186 unsigned long pfn = page_to_pfn(page); 187 188 do { 189 seq = zone_span_seqbegin(zone); 190 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 191 ret = 1; 192 else if (pfn < zone->zone_start_pfn) 193 ret = 1; 194 } while (zone_span_seqretry(zone, seq)); 195 196 return ret; 197} 198 199static int page_is_consistent(struct zone *zone, struct page *page) 200{ 201 if (!pfn_valid_within(page_to_pfn(page))) 202 return 0; 203 if (zone != page_zone(page)) 204 return 0; 205 206 return 1; 207} 208/* 209 * Temporary debugging check for pages not lying within a given zone. 210 */ 211static int bad_range(struct zone *zone, struct page *page) 212{ 213 if (page_outside_zone_boundaries(zone, page)) 214 return 1; 215 if (!page_is_consistent(zone, page)) 216 return 1; 217 218 return 0; 219} 220#else 221static inline int bad_range(struct zone *zone, struct page *page) 222{ 223 return 0; 224} 225#endif 226 227static void bad_page(struct page *page) 228{ 229 static unsigned long resume; 230 static unsigned long nr_shown; 231 static unsigned long nr_unshown; 232 233 /* 234 * Allow a burst of 60 reports, then keep quiet for that minute; 235 * or allow a steady drip of one report per second. 236 */ 237 if (nr_shown == 60) { 238 if (time_before(jiffies, resume)) { 239 nr_unshown++; 240 goto out; 241 } 242 if (nr_unshown) { 243 printk(KERN_ALERT 244 "BUG: Bad page state: %lu messages suppressed\n", 245 nr_unshown); 246 nr_unshown = 0; 247 } 248 nr_shown = 0; 249 } 250 if (nr_shown++ == 0) 251 resume = jiffies + 60 * HZ; 252 253 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 254 current->comm, page_to_pfn(page)); 255 printk(KERN_ALERT 256 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 257 page, (void *)page->flags, page_count(page), 258 page_mapcount(page), page->mapping, page->index); 259 260 dump_stack(); 261out: 262 /* Leave bad fields for debug, except PageBuddy could make trouble */ 263 __ClearPageBuddy(page); 264 add_taint(TAINT_BAD_PAGE); 265} 266 267/* 268 * Higher-order pages are called "compound pages". They are structured thusly: 269 * 270 * The first PAGE_SIZE page is called the "head page". 271 * 272 * The remaining PAGE_SIZE pages are called "tail pages". 273 * 274 * All pages have PG_compound set. All pages have their ->private pointing at 275 * the head page (even the head page has this). 276 * 277 * The first tail page's ->lru.next holds the address of the compound page's 278 * put_page() function. Its ->lru.prev holds the order of allocation. 279 * This usage means that zero-order pages may not be compound. 280 */ 281 282static void free_compound_page(struct page *page) 283{ 284 __free_pages_ok(page, compound_order(page)); 285} 286 287void prep_compound_page(struct page *page, unsigned long order) 288{ 289 int i; 290 int nr_pages = 1 << order; 291 292 set_compound_page_dtor(page, free_compound_page); 293 set_compound_order(page, order); 294 __SetPageHead(page); 295 for (i = 1; i < nr_pages; i++) { 296 struct page *p = page + i; 297 298 __SetPageTail(p); 299 p->first_page = page; 300 } 301} 302 303static int destroy_compound_page(struct page *page, unsigned long order) 304{ 305 int i; 306 int nr_pages = 1 << order; 307 int bad = 0; 308 309 if (unlikely(compound_order(page) != order) || 310 unlikely(!PageHead(page))) { 311 bad_page(page); 312 bad++; 313 } 314 315 __ClearPageHead(page); 316 317 for (i = 1; i < nr_pages; i++) { 318 struct page *p = page + i; 319 320 if (unlikely(!PageTail(p) || (p->first_page != page))) { 321 bad_page(page); 322 bad++; 323 } 324 __ClearPageTail(p); 325 } 326 327 return bad; 328} 329 330static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 331{ 332 int i; 333 334 /* 335 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 336 * and __GFP_HIGHMEM from hard or soft interrupt context. 337 */ 338 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 339 for (i = 0; i < (1 << order); i++) 340 clear_highpage(page + i); 341} 342 343static inline void set_page_order(struct page *page, int order) 344{ 345 set_page_private(page, order); 346 __SetPageBuddy(page); 347} 348 349static inline void rmv_page_order(struct page *page) 350{ 351 __ClearPageBuddy(page); 352 set_page_private(page, 0); 353} 354 355/* 356 * Locate the struct page for both the matching buddy in our 357 * pair (buddy1) and the combined O(n+1) page they form (page). 358 * 359 * 1) Any buddy B1 will have an order O twin B2 which satisfies 360 * the following equation: 361 * B2 = B1 ^ (1 << O) 362 * For example, if the starting buddy (buddy2) is #8 its order 363 * 1 buddy is #10: 364 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 365 * 366 * 2) Any buddy B will have an order O+1 parent P which 367 * satisfies the following equation: 368 * P = B & ~(1 << O) 369 * 370 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 371 */ 372static inline struct page * 373__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 374{ 375 unsigned long buddy_idx = page_idx ^ (1 << order); 376 377 return page + (buddy_idx - page_idx); 378} 379 380static inline unsigned long 381__find_combined_index(unsigned long page_idx, unsigned int order) 382{ 383 return (page_idx & ~(1 << order)); 384} 385 386/* 387 * This function checks whether a page is free && is the buddy 388 * we can do coalesce a page and its buddy if 389 * (a) the buddy is not in a hole && 390 * (b) the buddy is in the buddy system && 391 * (c) a page and its buddy have the same order && 392 * (d) a page and its buddy are in the same zone. 393 * 394 * For recording whether a page is in the buddy system, we use PG_buddy. 395 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 396 * 397 * For recording page's order, we use page_private(page). 398 */ 399static inline int page_is_buddy(struct page *page, struct page *buddy, 400 int order) 401{ 402 if (!pfn_valid_within(page_to_pfn(buddy))) 403 return 0; 404 405 if (page_zone_id(page) != page_zone_id(buddy)) 406 return 0; 407 408 if (PageBuddy(buddy) && page_order(buddy) == order) { 409 VM_BUG_ON(page_count(buddy) != 0); 410 return 1; 411 } 412 return 0; 413} 414 415/* 416 * Freeing function for a buddy system allocator. 417 * 418 * The concept of a buddy system is to maintain direct-mapped table 419 * (containing bit values) for memory blocks of various "orders". 420 * The bottom level table contains the map for the smallest allocatable 421 * units of memory (here, pages), and each level above it describes 422 * pairs of units from the levels below, hence, "buddies". 423 * At a high level, all that happens here is marking the table entry 424 * at the bottom level available, and propagating the changes upward 425 * as necessary, plus some accounting needed to play nicely with other 426 * parts of the VM system. 427 * At each level, we keep a list of pages, which are heads of continuous 428 * free pages of length of (1 << order) and marked with PG_buddy. Page's 429 * order is recorded in page_private(page) field. 430 * So when we are allocating or freeing one, we can derive the state of the 431 * other. That is, if we allocate a small block, and both were 432 * free, the remainder of the region must be split into blocks. 433 * If a block is freed, and its buddy is also free, then this 434 * triggers coalescing into a block of larger size. 435 * 436 * -- wli 437 */ 438 439static inline void __free_one_page(struct page *page, 440 struct zone *zone, unsigned int order, 441 int migratetype) 442{ 443 unsigned long page_idx; 444 445 if (unlikely(PageCompound(page))) 446 if (unlikely(destroy_compound_page(page, order))) 447 return; 448 449 VM_BUG_ON(migratetype == -1); 450 451 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 452 453 VM_BUG_ON(page_idx & ((1 << order) - 1)); 454 VM_BUG_ON(bad_range(zone, page)); 455 456 while (order < MAX_ORDER-1) { 457 unsigned long combined_idx; 458 struct page *buddy; 459 460 buddy = __page_find_buddy(page, page_idx, order); 461 if (!page_is_buddy(page, buddy, order)) 462 break; 463 464 /* Our buddy is free, merge with it and move up one order. */ 465 list_del(&buddy->lru); 466 zone->free_area[order].nr_free--; 467 rmv_page_order(buddy); 468 combined_idx = __find_combined_index(page_idx, order); 469 page = page + (combined_idx - page_idx); 470 page_idx = combined_idx; 471 order++; 472 } 473 set_page_order(page, order); 474 list_add(&page->lru, 475 &zone->free_area[order].free_list[migratetype]); 476 zone->free_area[order].nr_free++; 477} 478 479#ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT 480/* 481 * free_page_mlock() -- clean up attempts to free and mlocked() page. 482 * Page should not be on lru, so no need to fix that up. 483 * free_pages_check() will verify... 484 */ 485static inline void free_page_mlock(struct page *page) 486{ 487 __ClearPageMlocked(page); 488 __dec_zone_page_state(page, NR_MLOCK); 489 __count_vm_event(UNEVICTABLE_MLOCKFREED); 490} 491#else 492static void free_page_mlock(struct page *page) { } 493#endif 494 495static inline int free_pages_check(struct page *page) 496{ 497 if (unlikely(page_mapcount(page) | 498 (page->mapping != NULL) | 499 (atomic_read(&page->_count) != 0) | 500 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 501 bad_page(page); 502 return 1; 503 } 504 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 505 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 506 return 0; 507} 508 509/* 510 * Frees a list of pages. 511 * Assumes all pages on list are in same zone, and of same order. 512 * count is the number of pages to free. 513 * 514 * If the zone was previously in an "all pages pinned" state then look to 515 * see if this freeing clears that state. 516 * 517 * And clear the zone's pages_scanned counter, to hold off the "all pages are 518 * pinned" detection logic. 519 */ 520static void free_pages_bulk(struct zone *zone, int count, 521 struct list_head *list, int order) 522{ 523 spin_lock(&zone->lock); 524 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 525 zone->pages_scanned = 0; 526 527 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order); 528 while (count--) { 529 struct page *page; 530 531 VM_BUG_ON(list_empty(list)); 532 page = list_entry(list->prev, struct page, lru); 533 /* have to delete it as __free_one_page list manipulates */ 534 list_del(&page->lru); 535 __free_one_page(page, zone, order, page_private(page)); 536 } 537 spin_unlock(&zone->lock); 538} 539 540static void free_one_page(struct zone *zone, struct page *page, int order, 541 int migratetype) 542{ 543 spin_lock(&zone->lock); 544 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 545 zone->pages_scanned = 0; 546 547 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 548 __free_one_page(page, zone, order, migratetype); 549 spin_unlock(&zone->lock); 550} 551 552static void __free_pages_ok(struct page *page, unsigned int order) 553{ 554 unsigned long flags; 555 int i; 556 int bad = 0; 557 int clearMlocked = PageMlocked(page); 558 559 for (i = 0 ; i < (1 << order) ; ++i) 560 bad += free_pages_check(page + i); 561 if (bad) 562 return; 563 564 if (!PageHighMem(page)) { 565 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 566 debug_check_no_obj_freed(page_address(page), 567 PAGE_SIZE << order); 568 } 569 arch_free_page(page, order); 570 kernel_map_pages(page, 1 << order, 0); 571 572 local_irq_save(flags); 573 if (unlikely(clearMlocked)) 574 free_page_mlock(page); 575 __count_vm_events(PGFREE, 1 << order); 576 free_one_page(page_zone(page), page, order, 577 get_pageblock_migratetype(page)); 578 local_irq_restore(flags); 579} 580 581/* 582 * permit the bootmem allocator to evade page validation on high-order frees 583 */ 584void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 585{ 586 if (order == 0) { 587 __ClearPageReserved(page); 588 set_page_count(page, 0); 589 set_page_refcounted(page); 590 __free_page(page); 591 } else { 592 int loop; 593 594 prefetchw(page); 595 for (loop = 0; loop < BITS_PER_LONG; loop++) { 596 struct page *p = &page[loop]; 597 598 if (loop + 1 < BITS_PER_LONG) 599 prefetchw(p + 1); 600 __ClearPageReserved(p); 601 set_page_count(p, 0); 602 } 603 604 set_page_refcounted(page); 605 __free_pages(page, order); 606 } 607} 608 609 610/* 611 * The order of subdivision here is critical for the IO subsystem. 612 * Please do not alter this order without good reasons and regression 613 * testing. Specifically, as large blocks of memory are subdivided, 614 * the order in which smaller blocks are delivered depends on the order 615 * they're subdivided in this function. This is the primary factor 616 * influencing the order in which pages are delivered to the IO 617 * subsystem according to empirical testing, and this is also justified 618 * by considering the behavior of a buddy system containing a single 619 * large block of memory acted on by a series of small allocations. 620 * This behavior is a critical factor in sglist merging's success. 621 * 622 * -- wli 623 */ 624static inline void expand(struct zone *zone, struct page *page, 625 int low, int high, struct free_area *area, 626 int migratetype) 627{ 628 unsigned long size = 1 << high; 629 630 while (high > low) { 631 area--; 632 high--; 633 size >>= 1; 634 VM_BUG_ON(bad_range(zone, &page[size])); 635 list_add(&page[size].lru, &area->free_list[migratetype]); 636 area->nr_free++; 637 set_page_order(&page[size], high); 638 } 639} 640 641/* 642 * This page is about to be returned from the page allocator 643 */ 644static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 645{ 646 if (unlikely(page_mapcount(page) | 647 (page->mapping != NULL) | 648 (atomic_read(&page->_count) != 0) | 649 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 650 bad_page(page); 651 return 1; 652 } 653 654 set_page_private(page, 0); 655 set_page_refcounted(page); 656 657 arch_alloc_page(page, order); 658 kernel_map_pages(page, 1 << order, 1); 659 660 if (gfp_flags & __GFP_ZERO) 661 prep_zero_page(page, order, gfp_flags); 662 663 if (order && (gfp_flags & __GFP_COMP)) 664 prep_compound_page(page, order); 665 666 return 0; 667} 668 669/* 670 * Go through the free lists for the given migratetype and remove 671 * the smallest available page from the freelists 672 */ 673static inline 674struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 675 int migratetype) 676{ 677 unsigned int current_order; 678 struct free_area * area; 679 struct page *page; 680 681 /* Find a page of the appropriate size in the preferred list */ 682 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 683 area = &(zone->free_area[current_order]); 684 if (list_empty(&area->free_list[migratetype])) 685 continue; 686 687 page = list_entry(area->free_list[migratetype].next, 688 struct page, lru); 689 list_del(&page->lru); 690 rmv_page_order(page); 691 area->nr_free--; 692 expand(zone, page, order, current_order, area, migratetype); 693 return page; 694 } 695 696 return NULL; 697} 698 699 700/* 701 * This array describes the order lists are fallen back to when 702 * the free lists for the desirable migrate type are depleted 703 */ 704static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 705 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 706 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 707 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 708 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 709}; 710 711/* 712 * Move the free pages in a range to the free lists of the requested type. 713 * Note that start_page and end_pages are not aligned on a pageblock 714 * boundary. If alignment is required, use move_freepages_block() 715 */ 716static int move_freepages(struct zone *zone, 717 struct page *start_page, struct page *end_page, 718 int migratetype) 719{ 720 struct page *page; 721 unsigned long order; 722 int pages_moved = 0; 723 724#ifndef CONFIG_HOLES_IN_ZONE 725 /* 726 * page_zone is not safe to call in this context when 727 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 728 * anyway as we check zone boundaries in move_freepages_block(). 729 * Remove at a later date when no bug reports exist related to 730 * grouping pages by mobility 731 */ 732 BUG_ON(page_zone(start_page) != page_zone(end_page)); 733#endif 734 735 for (page = start_page; page <= end_page;) { 736 /* Make sure we are not inadvertently changing nodes */ 737 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 738 739 if (!pfn_valid_within(page_to_pfn(page))) { 740 page++; 741 continue; 742 } 743 744 if (!PageBuddy(page)) { 745 page++; 746 continue; 747 } 748 749 order = page_order(page); 750 list_del(&page->lru); 751 list_add(&page->lru, 752 &zone->free_area[order].free_list[migratetype]); 753 page += 1 << order; 754 pages_moved += 1 << order; 755 } 756 757 return pages_moved; 758} 759 760static int move_freepages_block(struct zone *zone, struct page *page, 761 int migratetype) 762{ 763 unsigned long start_pfn, end_pfn; 764 struct page *start_page, *end_page; 765 766 start_pfn = page_to_pfn(page); 767 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 768 start_page = pfn_to_page(start_pfn); 769 end_page = start_page + pageblock_nr_pages - 1; 770 end_pfn = start_pfn + pageblock_nr_pages - 1; 771 772 /* Do not cross zone boundaries */ 773 if (start_pfn < zone->zone_start_pfn) 774 start_page = page; 775 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 776 return 0; 777 778 return move_freepages(zone, start_page, end_page, migratetype); 779} 780 781/* Remove an element from the buddy allocator from the fallback list */ 782static inline struct page * 783__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 784{ 785 struct free_area * area; 786 int current_order; 787 struct page *page; 788 int migratetype, i; 789 790 /* Find the largest possible block of pages in the other list */ 791 for (current_order = MAX_ORDER-1; current_order >= order; 792 --current_order) { 793 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 794 migratetype = fallbacks[start_migratetype][i]; 795 796 /* MIGRATE_RESERVE handled later if necessary */ 797 if (migratetype == MIGRATE_RESERVE) 798 continue; 799 800 area = &(zone->free_area[current_order]); 801 if (list_empty(&area->free_list[migratetype])) 802 continue; 803 804 page = list_entry(area->free_list[migratetype].next, 805 struct page, lru); 806 area->nr_free--; 807 808 /* 809 * If breaking a large block of pages, move all free 810 * pages to the preferred allocation list. If falling 811 * back for a reclaimable kernel allocation, be more 812 * agressive about taking ownership of free pages 813 */ 814 if (unlikely(current_order >= (pageblock_order >> 1)) || 815 start_migratetype == MIGRATE_RECLAIMABLE) { 816 unsigned long pages; 817 pages = move_freepages_block(zone, page, 818 start_migratetype); 819 820 /* Claim the whole block if over half of it is free */ 821 if (pages >= (1 << (pageblock_order-1))) 822 set_pageblock_migratetype(page, 823 start_migratetype); 824 825 migratetype = start_migratetype; 826 } 827 828 /* Remove the page from the freelists */ 829 list_del(&page->lru); 830 rmv_page_order(page); 831 832 if (current_order == pageblock_order) 833 set_pageblock_migratetype(page, 834 start_migratetype); 835 836 expand(zone, page, order, current_order, area, migratetype); 837 return page; 838 } 839 } 840 841 return NULL; 842} 843 844/* 845 * Do the hard work of removing an element from the buddy allocator. 846 * Call me with the zone->lock already held. 847 */ 848static struct page *__rmqueue(struct zone *zone, unsigned int order, 849 int migratetype) 850{ 851 struct page *page; 852 853retry_reserve: 854 page = __rmqueue_smallest(zone, order, migratetype); 855 856 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 857 page = __rmqueue_fallback(zone, order, migratetype); 858 859 /* 860 * Use MIGRATE_RESERVE rather than fail an allocation. goto 861 * is used because __rmqueue_smallest is an inline function 862 * and we want just one call site 863 */ 864 if (!page) { 865 migratetype = MIGRATE_RESERVE; 866 goto retry_reserve; 867 } 868 } 869 870 return page; 871} 872 873/* 874 * Obtain a specified number of elements from the buddy allocator, all under 875 * a single hold of the lock, for efficiency. Add them to the supplied list. 876 * Returns the number of new pages which were placed at *list. 877 */ 878static int rmqueue_bulk(struct zone *zone, unsigned int order, 879 unsigned long count, struct list_head *list, 880 int migratetype) 881{ 882 int i; 883 884 spin_lock(&zone->lock); 885 for (i = 0; i < count; ++i) { 886 struct page *page = __rmqueue(zone, order, migratetype); 887 if (unlikely(page == NULL)) 888 break; 889 890 /* 891 * Split buddy pages returned by expand() are received here 892 * in physical page order. The page is added to the callers and 893 * list and the list head then moves forward. From the callers 894 * perspective, the linked list is ordered by page number in 895 * some conditions. This is useful for IO devices that can 896 * merge IO requests if the physical pages are ordered 897 * properly. 898 */ 899 list_add(&page->lru, list); 900 set_page_private(page, migratetype); 901 list = &page->lru; 902 } 903 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 904 spin_unlock(&zone->lock); 905 return i; 906} 907 908#ifdef CONFIG_NUMA 909/* 910 * Called from the vmstat counter updater to drain pagesets of this 911 * currently executing processor on remote nodes after they have 912 * expired. 913 * 914 * Note that this function must be called with the thread pinned to 915 * a single processor. 916 */ 917void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 918{ 919 unsigned long flags; 920 int to_drain; 921 922 local_irq_save(flags); 923 if (pcp->count >= pcp->batch) 924 to_drain = pcp->batch; 925 else 926 to_drain = pcp->count; 927 free_pages_bulk(zone, to_drain, &pcp->list, 0); 928 pcp->count -= to_drain; 929 local_irq_restore(flags); 930} 931#endif 932 933/* 934 * Drain pages of the indicated processor. 935 * 936 * The processor must either be the current processor and the 937 * thread pinned to the current processor or a processor that 938 * is not online. 939 */ 940static void drain_pages(unsigned int cpu) 941{ 942 unsigned long flags; 943 struct zone *zone; 944 945 for_each_populated_zone(zone) { 946 struct per_cpu_pageset *pset; 947 struct per_cpu_pages *pcp; 948 949 pset = zone_pcp(zone, cpu); 950 951 pcp = &pset->pcp; 952 local_irq_save(flags); 953 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 954 pcp->count = 0; 955 local_irq_restore(flags); 956 } 957} 958 959/* 960 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 961 */ 962void drain_local_pages(void *arg) 963{ 964 drain_pages(smp_processor_id()); 965} 966 967/* 968 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 969 */ 970void drain_all_pages(void) 971{ 972 on_each_cpu(drain_local_pages, NULL, 1); 973} 974 975#ifdef CONFIG_HIBERNATION 976 977void mark_free_pages(struct zone *zone) 978{ 979 unsigned long pfn, max_zone_pfn; 980 unsigned long flags; 981 int order, t; 982 struct list_head *curr; 983 984 if (!zone->spanned_pages) 985 return; 986 987 spin_lock_irqsave(&zone->lock, flags); 988 989 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 990 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 991 if (pfn_valid(pfn)) { 992 struct page *page = pfn_to_page(pfn); 993 994 if (!swsusp_page_is_forbidden(page)) 995 swsusp_unset_page_free(page); 996 } 997 998 for_each_migratetype_order(order, t) { 999 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1000 unsigned long i; 1001 1002 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1003 for (i = 0; i < (1UL << order); i++) 1004 swsusp_set_page_free(pfn_to_page(pfn + i)); 1005 } 1006 } 1007 spin_unlock_irqrestore(&zone->lock, flags); 1008} 1009#endif /* CONFIG_PM */ 1010 1011/* 1012 * Free a 0-order page 1013 */ 1014static void free_hot_cold_page(struct page *page, int cold) 1015{ 1016 struct zone *zone = page_zone(page); 1017 struct per_cpu_pages *pcp; 1018 unsigned long flags; 1019 int clearMlocked = PageMlocked(page); 1020 1021 if (PageAnon(page)) 1022 page->mapping = NULL; 1023 if (free_pages_check(page)) 1024 return; 1025 1026 if (!PageHighMem(page)) { 1027 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1028 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1029 } 1030 arch_free_page(page, 0); 1031 kernel_map_pages(page, 1, 0); 1032 1033 pcp = &zone_pcp(zone, get_cpu())->pcp; 1034 set_page_private(page, get_pageblock_migratetype(page)); 1035 local_irq_save(flags); 1036 if (unlikely(clearMlocked)) 1037 free_page_mlock(page); 1038 __count_vm_event(PGFREE); 1039 1040 if (cold) 1041 list_add_tail(&page->lru, &pcp->list); 1042 else 1043 list_add(&page->lru, &pcp->list); 1044 pcp->count++; 1045 if (pcp->count >= pcp->high) { 1046 free_pages_bulk(zone, pcp->batch, &pcp->list, 0); 1047 pcp->count -= pcp->batch; 1048 } 1049 local_irq_restore(flags); 1050 put_cpu(); 1051} 1052 1053void free_hot_page(struct page *page) 1054{ 1055 free_hot_cold_page(page, 0); 1056} 1057 1058void free_cold_page(struct page *page) 1059{ 1060 free_hot_cold_page(page, 1); 1061} 1062 1063/* 1064 * split_page takes a non-compound higher-order page, and splits it into 1065 * n (1<<order) sub-pages: page[0..n] 1066 * Each sub-page must be freed individually. 1067 * 1068 * Note: this is probably too low level an operation for use in drivers. 1069 * Please consult with lkml before using this in your driver. 1070 */ 1071void split_page(struct page *page, unsigned int order) 1072{ 1073 int i; 1074 1075 VM_BUG_ON(PageCompound(page)); 1076 VM_BUG_ON(!page_count(page)); 1077 for (i = 1; i < (1 << order); i++) 1078 set_page_refcounted(page + i); 1079} 1080 1081/* 1082 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1083 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1084 * or two. 1085 */ 1086static inline 1087struct page *buffered_rmqueue(struct zone *preferred_zone, 1088 struct zone *zone, int order, gfp_t gfp_flags, 1089 int migratetype) 1090{ 1091 unsigned long flags; 1092 struct page *page; 1093 int cold = !!(gfp_flags & __GFP_COLD); 1094 int cpu; 1095 1096again: 1097 cpu = get_cpu(); 1098 if (likely(order == 0)) { 1099 struct per_cpu_pages *pcp; 1100 1101 pcp = &zone_pcp(zone, cpu)->pcp; 1102 local_irq_save(flags); 1103 if (!pcp->count) { 1104 pcp->count = rmqueue_bulk(zone, 0, 1105 pcp->batch, &pcp->list, migratetype); 1106 if (unlikely(!pcp->count)) 1107 goto failed; 1108 } 1109 1110 /* Find a page of the appropriate migrate type */ 1111 if (cold) { 1112 list_for_each_entry_reverse(page, &pcp->list, lru) 1113 if (page_private(page) == migratetype) 1114 break; 1115 } else { 1116 list_for_each_entry(page, &pcp->list, lru) 1117 if (page_private(page) == migratetype) 1118 break; 1119 } 1120 1121 /* Allocate more to the pcp list if necessary */ 1122 if (unlikely(&page->lru == &pcp->list)) { 1123 pcp->count += rmqueue_bulk(zone, 0, 1124 pcp->batch, &pcp->list, migratetype); 1125 page = list_entry(pcp->list.next, struct page, lru); 1126 } 1127 1128 list_del(&page->lru); 1129 pcp->count--; 1130 } else { 1131 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1132 /* 1133 * __GFP_NOFAIL is not to be used in new code. 1134 * 1135 * All __GFP_NOFAIL callers should be fixed so that they 1136 * properly detect and handle allocation failures. 1137 * 1138 * We most definitely don't want callers attempting to 1139 * allocate greater than single-page units with 1140 * __GFP_NOFAIL. 1141 */ 1142 WARN_ON_ONCE(order > 0); 1143 } 1144 spin_lock_irqsave(&zone->lock, flags); 1145 page = __rmqueue(zone, order, migratetype); 1146 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1147 spin_unlock(&zone->lock); 1148 if (!page) 1149 goto failed; 1150 } 1151 1152 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1153 zone_statistics(preferred_zone, zone); 1154 local_irq_restore(flags); 1155 put_cpu(); 1156 1157 VM_BUG_ON(bad_range(zone, page)); 1158 if (prep_new_page(page, order, gfp_flags)) 1159 goto again; 1160 return page; 1161 1162failed: 1163 local_irq_restore(flags); 1164 put_cpu(); 1165 return NULL; 1166} 1167 1168/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1169#define ALLOC_WMARK_MIN WMARK_MIN 1170#define ALLOC_WMARK_LOW WMARK_LOW 1171#define ALLOC_WMARK_HIGH WMARK_HIGH 1172#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1173 1174/* Mask to get the watermark bits */ 1175#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1176 1177#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1178#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1179#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1180 1181#ifdef CONFIG_FAIL_PAGE_ALLOC 1182 1183static struct fail_page_alloc_attr { 1184 struct fault_attr attr; 1185 1186 u32 ignore_gfp_highmem; 1187 u32 ignore_gfp_wait; 1188 u32 min_order; 1189 1190#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1191 1192 struct dentry *ignore_gfp_highmem_file; 1193 struct dentry *ignore_gfp_wait_file; 1194 struct dentry *min_order_file; 1195 1196#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1197 1198} fail_page_alloc = { 1199 .attr = FAULT_ATTR_INITIALIZER, 1200 .ignore_gfp_wait = 1, 1201 .ignore_gfp_highmem = 1, 1202 .min_order = 1, 1203}; 1204 1205static int __init setup_fail_page_alloc(char *str) 1206{ 1207 return setup_fault_attr(&fail_page_alloc.attr, str); 1208} 1209__setup("fail_page_alloc=", setup_fail_page_alloc); 1210 1211static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1212{ 1213 if (order < fail_page_alloc.min_order) 1214 return 0; 1215 if (gfp_mask & __GFP_NOFAIL) 1216 return 0; 1217 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1218 return 0; 1219 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1220 return 0; 1221 1222 return should_fail(&fail_page_alloc.attr, 1 << order); 1223} 1224 1225#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1226 1227static int __init fail_page_alloc_debugfs(void) 1228{ 1229 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1230 struct dentry *dir; 1231 int err; 1232 1233 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1234 "fail_page_alloc"); 1235 if (err) 1236 return err; 1237 dir = fail_page_alloc.attr.dentries.dir; 1238 1239 fail_page_alloc.ignore_gfp_wait_file = 1240 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1241 &fail_page_alloc.ignore_gfp_wait); 1242 1243 fail_page_alloc.ignore_gfp_highmem_file = 1244 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1245 &fail_page_alloc.ignore_gfp_highmem); 1246 fail_page_alloc.min_order_file = 1247 debugfs_create_u32("min-order", mode, dir, 1248 &fail_page_alloc.min_order); 1249 1250 if (!fail_page_alloc.ignore_gfp_wait_file || 1251 !fail_page_alloc.ignore_gfp_highmem_file || 1252 !fail_page_alloc.min_order_file) { 1253 err = -ENOMEM; 1254 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1255 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1256 debugfs_remove(fail_page_alloc.min_order_file); 1257 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1258 } 1259 1260 return err; 1261} 1262 1263late_initcall(fail_page_alloc_debugfs); 1264 1265#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1266 1267#else /* CONFIG_FAIL_PAGE_ALLOC */ 1268 1269static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1270{ 1271 return 0; 1272} 1273 1274#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1275 1276/* 1277 * Return 1 if free pages are above 'mark'. This takes into account the order 1278 * of the allocation. 1279 */ 1280int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1281 int classzone_idx, int alloc_flags) 1282{ 1283 /* free_pages my go negative - that's OK */ 1284 long min = mark; 1285 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1286 int o; 1287 1288 if (alloc_flags & ALLOC_HIGH) 1289 min -= min / 2; 1290 if (alloc_flags & ALLOC_HARDER) 1291 min -= min / 4; 1292 1293 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1294 return 0; 1295 for (o = 0; o < order; o++) { 1296 /* At the next order, this order's pages become unavailable */ 1297 free_pages -= z->free_area[o].nr_free << o; 1298 1299 /* Require fewer higher order pages to be free */ 1300 min >>= 1; 1301 1302 if (free_pages <= min) 1303 return 0; 1304 } 1305 return 1; 1306} 1307 1308#ifdef CONFIG_NUMA 1309/* 1310 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1311 * skip over zones that are not allowed by the cpuset, or that have 1312 * been recently (in last second) found to be nearly full. See further 1313 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1314 * that have to skip over a lot of full or unallowed zones. 1315 * 1316 * If the zonelist cache is present in the passed in zonelist, then 1317 * returns a pointer to the allowed node mask (either the current 1318 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1319 * 1320 * If the zonelist cache is not available for this zonelist, does 1321 * nothing and returns NULL. 1322 * 1323 * If the fullzones BITMAP in the zonelist cache is stale (more than 1324 * a second since last zap'd) then we zap it out (clear its bits.) 1325 * 1326 * We hold off even calling zlc_setup, until after we've checked the 1327 * first zone in the zonelist, on the theory that most allocations will 1328 * be satisfied from that first zone, so best to examine that zone as 1329 * quickly as we can. 1330 */ 1331static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1332{ 1333 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1334 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1335 1336 zlc = zonelist->zlcache_ptr; 1337 if (!zlc) 1338 return NULL; 1339 1340 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1341 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1342 zlc->last_full_zap = jiffies; 1343 } 1344 1345 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1346 &cpuset_current_mems_allowed : 1347 &node_states[N_HIGH_MEMORY]; 1348 return allowednodes; 1349} 1350 1351/* 1352 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1353 * if it is worth looking at further for free memory: 1354 * 1) Check that the zone isn't thought to be full (doesn't have its 1355 * bit set in the zonelist_cache fullzones BITMAP). 1356 * 2) Check that the zones node (obtained from the zonelist_cache 1357 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1358 * Return true (non-zero) if zone is worth looking at further, or 1359 * else return false (zero) if it is not. 1360 * 1361 * This check -ignores- the distinction between various watermarks, 1362 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1363 * found to be full for any variation of these watermarks, it will 1364 * be considered full for up to one second by all requests, unless 1365 * we are so low on memory on all allowed nodes that we are forced 1366 * into the second scan of the zonelist. 1367 * 1368 * In the second scan we ignore this zonelist cache and exactly 1369 * apply the watermarks to all zones, even it is slower to do so. 1370 * We are low on memory in the second scan, and should leave no stone 1371 * unturned looking for a free page. 1372 */ 1373static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1374 nodemask_t *allowednodes) 1375{ 1376 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1377 int i; /* index of *z in zonelist zones */ 1378 int n; /* node that zone *z is on */ 1379 1380 zlc = zonelist->zlcache_ptr; 1381 if (!zlc) 1382 return 1; 1383 1384 i = z - zonelist->_zonerefs; 1385 n = zlc->z_to_n[i]; 1386 1387 /* This zone is worth trying if it is allowed but not full */ 1388 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1389} 1390 1391/* 1392 * Given 'z' scanning a zonelist, set the corresponding bit in 1393 * zlc->fullzones, so that subsequent attempts to allocate a page 1394 * from that zone don't waste time re-examining it. 1395 */ 1396static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1397{ 1398 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1399 int i; /* index of *z in zonelist zones */ 1400 1401 zlc = zonelist->zlcache_ptr; 1402 if (!zlc) 1403 return; 1404 1405 i = z - zonelist->_zonerefs; 1406 1407 set_bit(i, zlc->fullzones); 1408} 1409 1410#else /* CONFIG_NUMA */ 1411 1412static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1413{ 1414 return NULL; 1415} 1416 1417static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1418 nodemask_t *allowednodes) 1419{ 1420 return 1; 1421} 1422 1423static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1424{ 1425} 1426#endif /* CONFIG_NUMA */ 1427 1428/* 1429 * get_page_from_freelist goes through the zonelist trying to allocate 1430 * a page. 1431 */ 1432static struct page * 1433get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1434 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1435 struct zone *preferred_zone, int migratetype) 1436{ 1437 struct zoneref *z; 1438 struct page *page = NULL; 1439 int classzone_idx; 1440 struct zone *zone; 1441 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1442 int zlc_active = 0; /* set if using zonelist_cache */ 1443 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1444 1445 classzone_idx = zone_idx(preferred_zone); 1446zonelist_scan: 1447 /* 1448 * Scan zonelist, looking for a zone with enough free. 1449 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1450 */ 1451 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1452 high_zoneidx, nodemask) { 1453 if (NUMA_BUILD && zlc_active && 1454 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1455 continue; 1456 if ((alloc_flags & ALLOC_CPUSET) && 1457 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1458 goto try_next_zone; 1459 1460 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1461 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1462 unsigned long mark; 1463 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1464 if (!zone_watermark_ok(zone, order, mark, 1465 classzone_idx, alloc_flags)) { 1466 if (!zone_reclaim_mode || 1467 !zone_reclaim(zone, gfp_mask, order)) 1468 goto this_zone_full; 1469 } 1470 } 1471 1472 page = buffered_rmqueue(preferred_zone, zone, order, 1473 gfp_mask, migratetype); 1474 if (page) 1475 break; 1476this_zone_full: 1477 if (NUMA_BUILD) 1478 zlc_mark_zone_full(zonelist, z); 1479try_next_zone: 1480 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1481 /* 1482 * we do zlc_setup after the first zone is tried but only 1483 * if there are multiple nodes make it worthwhile 1484 */ 1485 allowednodes = zlc_setup(zonelist, alloc_flags); 1486 zlc_active = 1; 1487 did_zlc_setup = 1; 1488 } 1489 } 1490 1491 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1492 /* Disable zlc cache for second zonelist scan */ 1493 zlc_active = 0; 1494 goto zonelist_scan; 1495 } 1496 return page; 1497} 1498 1499static inline int 1500should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1501 unsigned long pages_reclaimed) 1502{ 1503 /* Do not loop if specifically requested */ 1504 if (gfp_mask & __GFP_NORETRY) 1505 return 0; 1506 1507 /* 1508 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1509 * means __GFP_NOFAIL, but that may not be true in other 1510 * implementations. 1511 */ 1512 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1513 return 1; 1514 1515 /* 1516 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1517 * specified, then we retry until we no longer reclaim any pages 1518 * (above), or we've reclaimed an order of pages at least as 1519 * large as the allocation's order. In both cases, if the 1520 * allocation still fails, we stop retrying. 1521 */ 1522 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1523 return 1; 1524 1525 /* 1526 * Don't let big-order allocations loop unless the caller 1527 * explicitly requests that. 1528 */ 1529 if (gfp_mask & __GFP_NOFAIL) 1530 return 1; 1531 1532 return 0; 1533} 1534 1535static inline struct page * 1536__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1537 struct zonelist *zonelist, enum zone_type high_zoneidx, 1538 nodemask_t *nodemask, struct zone *preferred_zone, 1539 int migratetype) 1540{ 1541 struct page *page; 1542 1543 /* Acquire the OOM killer lock for the zones in zonelist */ 1544 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1545 schedule_timeout_uninterruptible(1); 1546 return NULL; 1547 } 1548 1549 /* 1550 * Go through the zonelist yet one more time, keep very high watermark 1551 * here, this is only to catch a parallel oom killing, we must fail if 1552 * we're still under heavy pressure. 1553 */ 1554 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1555 order, zonelist, high_zoneidx, 1556 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1557 preferred_zone, migratetype); 1558 if (page) 1559 goto out; 1560 1561 /* The OOM killer will not help higher order allocs */ 1562 if (order > PAGE_ALLOC_COSTLY_ORDER) 1563 goto out; 1564 1565 /* Exhausted what can be done so it's blamo time */ 1566 out_of_memory(zonelist, gfp_mask, order); 1567 1568out: 1569 clear_zonelist_oom(zonelist, gfp_mask); 1570 return page; 1571} 1572 1573/* The really slow allocator path where we enter direct reclaim */ 1574static inline struct page * 1575__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1576 struct zonelist *zonelist, enum zone_type high_zoneidx, 1577 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1578 int migratetype, unsigned long *did_some_progress) 1579{ 1580 struct page *page = NULL; 1581 struct reclaim_state reclaim_state; 1582 struct task_struct *p = current; 1583 1584 cond_resched(); 1585 1586 /* We now go into synchronous reclaim */ 1587 cpuset_memory_pressure_bump(); 1588 1589 /* 1590 * The task's cpuset might have expanded its set of allowable nodes 1591 */ 1592 p->flags |= PF_MEMALLOC; 1593 lockdep_set_current_reclaim_state(gfp_mask); 1594 reclaim_state.reclaimed_slab = 0; 1595 p->reclaim_state = &reclaim_state; 1596 1597 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1598 1599 p->reclaim_state = NULL; 1600 lockdep_clear_current_reclaim_state(); 1601 p->flags &= ~PF_MEMALLOC; 1602 1603 cond_resched(); 1604 1605 if (order != 0) 1606 drain_all_pages(); 1607 1608 if (likely(*did_some_progress)) 1609 page = get_page_from_freelist(gfp_mask, nodemask, order, 1610 zonelist, high_zoneidx, 1611 alloc_flags, preferred_zone, 1612 migratetype); 1613 return page; 1614} 1615 1616/* 1617 * This is called in the allocator slow-path if the allocation request is of 1618 * sufficient urgency to ignore watermarks and take other desperate measures 1619 */ 1620static inline struct page * 1621__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1622 struct zonelist *zonelist, enum zone_type high_zoneidx, 1623 nodemask_t *nodemask, struct zone *preferred_zone, 1624 int migratetype) 1625{ 1626 struct page *page; 1627 1628 do { 1629 page = get_page_from_freelist(gfp_mask, nodemask, order, 1630 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1631 preferred_zone, migratetype); 1632 1633 if (!page && gfp_mask & __GFP_NOFAIL) 1634 congestion_wait(WRITE, HZ/50); 1635 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1636 1637 return page; 1638} 1639 1640static inline 1641void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1642 enum zone_type high_zoneidx) 1643{ 1644 struct zoneref *z; 1645 struct zone *zone; 1646 1647 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1648 wakeup_kswapd(zone, order); 1649} 1650 1651static inline int 1652gfp_to_alloc_flags(gfp_t gfp_mask) 1653{ 1654 struct task_struct *p = current; 1655 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1656 const gfp_t wait = gfp_mask & __GFP_WAIT; 1657 1658 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1659 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1660 1661 /* 1662 * The caller may dip into page reserves a bit more if the caller 1663 * cannot run direct reclaim, or if the caller has realtime scheduling 1664 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1665 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1666 */ 1667 alloc_flags |= (gfp_mask & __GFP_HIGH); 1668 1669 if (!wait) { 1670 alloc_flags |= ALLOC_HARDER; 1671 /* 1672 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1673 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1674 */ 1675 alloc_flags &= ~ALLOC_CPUSET; 1676 } else if (unlikely(rt_task(p))) 1677 alloc_flags |= ALLOC_HARDER; 1678 1679 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1680 if (!in_interrupt() && 1681 ((p->flags & PF_MEMALLOC) || 1682 unlikely(test_thread_flag(TIF_MEMDIE)))) 1683 alloc_flags |= ALLOC_NO_WATERMARKS; 1684 } 1685 1686 return alloc_flags; 1687} 1688 1689static inline struct page * 1690__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1691 struct zonelist *zonelist, enum zone_type high_zoneidx, 1692 nodemask_t *nodemask, struct zone *preferred_zone, 1693 int migratetype) 1694{ 1695 const gfp_t wait = gfp_mask & __GFP_WAIT; 1696 struct page *page = NULL; 1697 int alloc_flags; 1698 unsigned long pages_reclaimed = 0; 1699 unsigned long did_some_progress; 1700 struct task_struct *p = current; 1701 1702 /* 1703 * In the slowpath, we sanity check order to avoid ever trying to 1704 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1705 * be using allocators in order of preference for an area that is 1706 * too large. 1707 */ 1708 if (WARN_ON_ONCE(order >= MAX_ORDER)) 1709 return NULL; 1710 1711 /* 1712 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1713 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1714 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1715 * using a larger set of nodes after it has established that the 1716 * allowed per node queues are empty and that nodes are 1717 * over allocated. 1718 */ 1719 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1720 goto nopage; 1721 1722 wake_all_kswapd(order, zonelist, high_zoneidx); 1723 1724 /* 1725 * OK, we're below the kswapd watermark and have kicked background 1726 * reclaim. Now things get more complex, so set up alloc_flags according 1727 * to how we want to proceed. 1728 */ 1729 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1730 1731restart: 1732 /* This is the last chance, in general, before the goto nopage. */ 1733 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1734 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1735 preferred_zone, migratetype); 1736 if (page) 1737 goto got_pg; 1738 1739rebalance: 1740 /* Allocate without watermarks if the context allows */ 1741 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1742 page = __alloc_pages_high_priority(gfp_mask, order, 1743 zonelist, high_zoneidx, nodemask, 1744 preferred_zone, migratetype); 1745 if (page) 1746 goto got_pg; 1747 } 1748 1749 /* Atomic allocations - we can't balance anything */ 1750 if (!wait) 1751 goto nopage; 1752 1753 /* Avoid recursion of direct reclaim */ 1754 if (p->flags & PF_MEMALLOC) 1755 goto nopage; 1756 1757 /* Try direct reclaim and then allocating */ 1758 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1759 zonelist, high_zoneidx, 1760 nodemask, 1761 alloc_flags, preferred_zone, 1762 migratetype, &did_some_progress); 1763 if (page) 1764 goto got_pg; 1765 1766 /* 1767 * If we failed to make any progress reclaiming, then we are 1768 * running out of options and have to consider going OOM 1769 */ 1770 if (!did_some_progress) { 1771 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1772 page = __alloc_pages_may_oom(gfp_mask, order, 1773 zonelist, high_zoneidx, 1774 nodemask, preferred_zone, 1775 migratetype); 1776 if (page) 1777 goto got_pg; 1778 1779 /* 1780 * The OOM killer does not trigger for high-order allocations 1781 * but if no progress is being made, there are no other 1782 * options and retrying is unlikely to help 1783 */ 1784 if (order > PAGE_ALLOC_COSTLY_ORDER) 1785 goto nopage; 1786 1787 goto restart; 1788 } 1789 } 1790 1791 /* Check if we should retry the allocation */ 1792 pages_reclaimed += did_some_progress; 1793 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1794 /* Wait for some write requests to complete then retry */ 1795 congestion_wait(WRITE, HZ/50); 1796 goto rebalance; 1797 } 1798 1799nopage: 1800 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1801 printk(KERN_WARNING "%s: page allocation failure." 1802 " order:%d, mode:0x%x\n", 1803 p->comm, order, gfp_mask); 1804 dump_stack(); 1805 show_mem(); 1806 } 1807got_pg: 1808 return page; 1809 1810} 1811 1812/* 1813 * This is the 'heart' of the zoned buddy allocator. 1814 */ 1815struct page * 1816__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1817 struct zonelist *zonelist, nodemask_t *nodemask) 1818{ 1819 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1820 struct zone *preferred_zone; 1821 struct page *page; 1822 int migratetype = allocflags_to_migratetype(gfp_mask); 1823 1824 lockdep_trace_alloc(gfp_mask); 1825 1826 might_sleep_if(gfp_mask & __GFP_WAIT); 1827 1828 if (should_fail_alloc_page(gfp_mask, order)) 1829 return NULL; 1830 1831 /* 1832 * Check the zones suitable for the gfp_mask contain at least one 1833 * valid zone. It's possible to have an empty zonelist as a result 1834 * of GFP_THISNODE and a memoryless node 1835 */ 1836 if (unlikely(!zonelist->_zonerefs->zone)) 1837 return NULL; 1838 1839 /* The preferred zone is used for statistics later */ 1840 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1841 if (!preferred_zone) 1842 return NULL; 1843 1844 /* First allocation attempt */ 1845 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1846 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1847 preferred_zone, migratetype); 1848 if (unlikely(!page)) 1849 page = __alloc_pages_slowpath(gfp_mask, order, 1850 zonelist, high_zoneidx, nodemask, 1851 preferred_zone, migratetype); 1852 1853 return page; 1854} 1855EXPORT_SYMBOL(__alloc_pages_nodemask); 1856 1857/* 1858 * Common helper functions. 1859 */ 1860unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1861{ 1862 struct page * page; 1863 page = alloc_pages(gfp_mask, order); 1864 if (!page) 1865 return 0; 1866 return (unsigned long) page_address(page); 1867} 1868 1869EXPORT_SYMBOL(__get_free_pages); 1870 1871unsigned long get_zeroed_page(gfp_t gfp_mask) 1872{ 1873 struct page * page; 1874 1875 /* 1876 * get_zeroed_page() returns a 32-bit address, which cannot represent 1877 * a highmem page 1878 */ 1879 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1880 1881 page = alloc_pages(gfp_mask | __GFP_ZERO, 0); 1882 if (page) 1883 return (unsigned long) page_address(page); 1884 return 0; 1885} 1886 1887EXPORT_SYMBOL(get_zeroed_page); 1888 1889void __pagevec_free(struct pagevec *pvec) 1890{ 1891 int i = pagevec_count(pvec); 1892 1893 while (--i >= 0) 1894 free_hot_cold_page(pvec->pages[i], pvec->cold); 1895} 1896 1897void __free_pages(struct page *page, unsigned int order) 1898{ 1899 if (put_page_testzero(page)) { 1900 if (order == 0) 1901 free_hot_page(page); 1902 else 1903 __free_pages_ok(page, order); 1904 } 1905} 1906 1907EXPORT_SYMBOL(__free_pages); 1908 1909void free_pages(unsigned long addr, unsigned int order) 1910{ 1911 if (addr != 0) { 1912 VM_BUG_ON(!virt_addr_valid((void *)addr)); 1913 __free_pages(virt_to_page((void *)addr), order); 1914 } 1915} 1916 1917EXPORT_SYMBOL(free_pages); 1918 1919/** 1920 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 1921 * @size: the number of bytes to allocate 1922 * @gfp_mask: GFP flags for the allocation 1923 * 1924 * This function is similar to alloc_pages(), except that it allocates the 1925 * minimum number of pages to satisfy the request. alloc_pages() can only 1926 * allocate memory in power-of-two pages. 1927 * 1928 * This function is also limited by MAX_ORDER. 1929 * 1930 * Memory allocated by this function must be released by free_pages_exact(). 1931 */ 1932void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 1933{ 1934 unsigned int order = get_order(size); 1935 unsigned long addr; 1936 1937 addr = __get_free_pages(gfp_mask, order); 1938 if (addr) { 1939 unsigned long alloc_end = addr + (PAGE_SIZE << order); 1940 unsigned long used = addr + PAGE_ALIGN(size); 1941 1942 split_page(virt_to_page(addr), order); 1943 while (used < alloc_end) { 1944 free_page(used); 1945 used += PAGE_SIZE; 1946 } 1947 } 1948 1949 return (void *)addr; 1950} 1951EXPORT_SYMBOL(alloc_pages_exact); 1952 1953/** 1954 * free_pages_exact - release memory allocated via alloc_pages_exact() 1955 * @virt: the value returned by alloc_pages_exact. 1956 * @size: size of allocation, same value as passed to alloc_pages_exact(). 1957 * 1958 * Release the memory allocated by a previous call to alloc_pages_exact. 1959 */ 1960void free_pages_exact(void *virt, size_t size) 1961{ 1962 unsigned long addr = (unsigned long)virt; 1963 unsigned long end = addr + PAGE_ALIGN(size); 1964 1965 while (addr < end) { 1966 free_page(addr); 1967 addr += PAGE_SIZE; 1968 } 1969} 1970EXPORT_SYMBOL(free_pages_exact); 1971 1972static unsigned int nr_free_zone_pages(int offset) 1973{ 1974 struct zoneref *z; 1975 struct zone *zone; 1976 1977 /* Just pick one node, since fallback list is circular */ 1978 unsigned int sum = 0; 1979 1980 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 1981 1982 for_each_zone_zonelist(zone, z, zonelist, offset) { 1983 unsigned long size = zone->present_pages; 1984 unsigned long high = high_wmark_pages(zone); 1985 if (size > high) 1986 sum += size - high; 1987 } 1988 1989 return sum; 1990} 1991 1992/* 1993 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 1994 */ 1995unsigned int nr_free_buffer_pages(void) 1996{ 1997 return nr_free_zone_pages(gfp_zone(GFP_USER)); 1998} 1999EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2000 2001/* 2002 * Amount of free RAM allocatable within all zones 2003 */ 2004unsigned int nr_free_pagecache_pages(void) 2005{ 2006 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2007} 2008 2009static inline void show_node(struct zone *zone) 2010{ 2011 if (NUMA_BUILD) 2012 printk("Node %d ", zone_to_nid(zone)); 2013} 2014 2015void si_meminfo(struct sysinfo *val) 2016{ 2017 val->totalram = totalram_pages; 2018 val->sharedram = 0; 2019 val->freeram = global_page_state(NR_FREE_PAGES); 2020 val->bufferram = nr_blockdev_pages(); 2021 val->totalhigh = totalhigh_pages; 2022 val->freehigh = nr_free_highpages(); 2023 val->mem_unit = PAGE_SIZE; 2024} 2025 2026EXPORT_SYMBOL(si_meminfo); 2027 2028#ifdef CONFIG_NUMA 2029void si_meminfo_node(struct sysinfo *val, int nid) 2030{ 2031 pg_data_t *pgdat = NODE_DATA(nid); 2032 2033 val->totalram = pgdat->node_present_pages; 2034 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2035#ifdef CONFIG_HIGHMEM 2036 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2037 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2038 NR_FREE_PAGES); 2039#else 2040 val->totalhigh = 0; 2041 val->freehigh = 0; 2042#endif 2043 val->mem_unit = PAGE_SIZE; 2044} 2045#endif 2046 2047#define K(x) ((x) << (PAGE_SHIFT-10)) 2048 2049/* 2050 * Show free area list (used inside shift_scroll-lock stuff) 2051 * We also calculate the percentage fragmentation. We do this by counting the 2052 * memory on each free list with the exception of the first item on the list. 2053 */ 2054void show_free_areas(void) 2055{ 2056 int cpu; 2057 struct zone *zone; 2058 2059 for_each_populated_zone(zone) { 2060 show_node(zone); 2061 printk("%s per-cpu:\n", zone->name); 2062 2063 for_each_online_cpu(cpu) { 2064 struct per_cpu_pageset *pageset; 2065 2066 pageset = zone_pcp(zone, cpu); 2067 2068 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2069 cpu, pageset->pcp.high, 2070 pageset->pcp.batch, pageset->pcp.count); 2071 } 2072 } 2073 2074 printk("Active_anon:%lu active_file:%lu inactive_anon:%lu\n" 2075 " inactive_file:%lu" 2076//TODO: check/adjust line lengths 2077#ifdef CONFIG_UNEVICTABLE_LRU 2078 " unevictable:%lu" 2079#endif 2080 " dirty:%lu writeback:%lu unstable:%lu\n" 2081 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n", 2082 global_page_state(NR_ACTIVE_ANON), 2083 global_page_state(NR_ACTIVE_FILE), 2084 global_page_state(NR_INACTIVE_ANON), 2085 global_page_state(NR_INACTIVE_FILE), 2086#ifdef CONFIG_UNEVICTABLE_LRU 2087 global_page_state(NR_UNEVICTABLE), 2088#endif 2089 global_page_state(NR_FILE_DIRTY), 2090 global_page_state(NR_WRITEBACK), 2091 global_page_state(NR_UNSTABLE_NFS), 2092 global_page_state(NR_FREE_PAGES), 2093 global_page_state(NR_SLAB_RECLAIMABLE) + 2094 global_page_state(NR_SLAB_UNRECLAIMABLE), 2095 global_page_state(NR_FILE_MAPPED), 2096 global_page_state(NR_PAGETABLE), 2097 global_page_state(NR_BOUNCE)); 2098 2099 for_each_populated_zone(zone) { 2100 int i; 2101 2102 show_node(zone); 2103 printk("%s" 2104 " free:%lukB" 2105 " min:%lukB" 2106 " low:%lukB" 2107 " high:%lukB" 2108 " active_anon:%lukB" 2109 " inactive_anon:%lukB" 2110 " active_file:%lukB" 2111 " inactive_file:%lukB" 2112#ifdef CONFIG_UNEVICTABLE_LRU 2113 " unevictable:%lukB" 2114#endif 2115 " present:%lukB" 2116 " pages_scanned:%lu" 2117 " all_unreclaimable? %s" 2118 "\n", 2119 zone->name, 2120 K(zone_page_state(zone, NR_FREE_PAGES)), 2121 K(min_wmark_pages(zone)), 2122 K(low_wmark_pages(zone)), 2123 K(high_wmark_pages(zone)), 2124 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2125 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2126 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2127 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2128#ifdef CONFIG_UNEVICTABLE_LRU 2129 K(zone_page_state(zone, NR_UNEVICTABLE)), 2130#endif 2131 K(zone->present_pages), 2132 zone->pages_scanned, 2133 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 2134 ); 2135 printk("lowmem_reserve[]:"); 2136 for (i = 0; i < MAX_NR_ZONES; i++) 2137 printk(" %lu", zone->lowmem_reserve[i]); 2138 printk("\n"); 2139 } 2140 2141 for_each_populated_zone(zone) { 2142 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2143 2144 show_node(zone); 2145 printk("%s: ", zone->name); 2146 2147 spin_lock_irqsave(&zone->lock, flags); 2148 for (order = 0; order < MAX_ORDER; order++) { 2149 nr[order] = zone->free_area[order].nr_free; 2150 total += nr[order] << order; 2151 } 2152 spin_unlock_irqrestore(&zone->lock, flags); 2153 for (order = 0; order < MAX_ORDER; order++) 2154 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2155 printk("= %lukB\n", K(total)); 2156 } 2157 2158 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2159 2160 show_swap_cache_info(); 2161} 2162 2163static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2164{ 2165 zoneref->zone = zone; 2166 zoneref->zone_idx = zone_idx(zone); 2167} 2168 2169/* 2170 * Builds allocation fallback zone lists. 2171 * 2172 * Add all populated zones of a node to the zonelist. 2173 */ 2174static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2175 int nr_zones, enum zone_type zone_type) 2176{ 2177 struct zone *zone; 2178 2179 BUG_ON(zone_type >= MAX_NR_ZONES); 2180 zone_type++; 2181 2182 do { 2183 zone_type--; 2184 zone = pgdat->node_zones + zone_type; 2185 if (populated_zone(zone)) { 2186 zoneref_set_zone(zone, 2187 &zonelist->_zonerefs[nr_zones++]); 2188 check_highest_zone(zone_type); 2189 } 2190 2191 } while (zone_type); 2192 return nr_zones; 2193} 2194 2195 2196/* 2197 * zonelist_order: 2198 * 0 = automatic detection of better ordering. 2199 * 1 = order by ([node] distance, -zonetype) 2200 * 2 = order by (-zonetype, [node] distance) 2201 * 2202 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2203 * the same zonelist. So only NUMA can configure this param. 2204 */ 2205#define ZONELIST_ORDER_DEFAULT 0 2206#define ZONELIST_ORDER_NODE 1 2207#define ZONELIST_ORDER_ZONE 2 2208 2209/* zonelist order in the kernel. 2210 * set_zonelist_order() will set this to NODE or ZONE. 2211 */ 2212static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2213static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2214 2215 2216#ifdef CONFIG_NUMA 2217/* The value user specified ....changed by config */ 2218static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2219/* string for sysctl */ 2220#define NUMA_ZONELIST_ORDER_LEN 16 2221char numa_zonelist_order[16] = "default"; 2222 2223/* 2224 * interface for configure zonelist ordering. 2225 * command line option "numa_zonelist_order" 2226 * = "[dD]efault - default, automatic configuration. 2227 * = "[nN]ode - order by node locality, then by zone within node 2228 * = "[zZ]one - order by zone, then by locality within zone 2229 */ 2230 2231static int __parse_numa_zonelist_order(char *s) 2232{ 2233 if (*s == 'd' || *s == 'D') { 2234 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2235 } else if (*s == 'n' || *s == 'N') { 2236 user_zonelist_order = ZONELIST_ORDER_NODE; 2237 } else if (*s == 'z' || *s == 'Z') { 2238 user_zonelist_order = ZONELIST_ORDER_ZONE; 2239 } else { 2240 printk(KERN_WARNING 2241 "Ignoring invalid numa_zonelist_order value: " 2242 "%s\n", s); 2243 return -EINVAL; 2244 } 2245 return 0; 2246} 2247 2248static __init int setup_numa_zonelist_order(char *s) 2249{ 2250 if (s) 2251 return __parse_numa_zonelist_order(s); 2252 return 0; 2253} 2254early_param("numa_zonelist_order", setup_numa_zonelist_order); 2255 2256/* 2257 * sysctl handler for numa_zonelist_order 2258 */ 2259int numa_zonelist_order_handler(ctl_table *table, int write, 2260 struct file *file, void __user *buffer, size_t *length, 2261 loff_t *ppos) 2262{ 2263 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2264 int ret; 2265 2266 if (write) 2267 strncpy(saved_string, (char*)table->data, 2268 NUMA_ZONELIST_ORDER_LEN); 2269 ret = proc_dostring(table, write, file, buffer, length, ppos); 2270 if (ret) 2271 return ret; 2272 if (write) { 2273 int oldval = user_zonelist_order; 2274 if (__parse_numa_zonelist_order((char*)table->data)) { 2275 /* 2276 * bogus value. restore saved string 2277 */ 2278 strncpy((char*)table->data, saved_string, 2279 NUMA_ZONELIST_ORDER_LEN); 2280 user_zonelist_order = oldval; 2281 } else if (oldval != user_zonelist_order) 2282 build_all_zonelists(); 2283 } 2284 return 0; 2285} 2286 2287 2288#define MAX_NODE_LOAD (nr_online_nodes) 2289static int node_load[MAX_NUMNODES]; 2290 2291/** 2292 * find_next_best_node - find the next node that should appear in a given node's fallback list 2293 * @node: node whose fallback list we're appending 2294 * @used_node_mask: nodemask_t of already used nodes 2295 * 2296 * We use a number of factors to determine which is the next node that should 2297 * appear on a given node's fallback list. The node should not have appeared 2298 * already in @node's fallback list, and it should be the next closest node 2299 * according to the distance array (which contains arbitrary distance values 2300 * from each node to each node in the system), and should also prefer nodes 2301 * with no CPUs, since presumably they'll have very little allocation pressure 2302 * on them otherwise. 2303 * It returns -1 if no node is found. 2304 */ 2305static int find_next_best_node(int node, nodemask_t *used_node_mask) 2306{ 2307 int n, val; 2308 int min_val = INT_MAX; 2309 int best_node = -1; 2310 const struct cpumask *tmp = cpumask_of_node(0); 2311 2312 /* Use the local node if we haven't already */ 2313 if (!node_isset(node, *used_node_mask)) { 2314 node_set(node, *used_node_mask); 2315 return node; 2316 } 2317 2318 for_each_node_state(n, N_HIGH_MEMORY) { 2319 2320 /* Don't want a node to appear more than once */ 2321 if (node_isset(n, *used_node_mask)) 2322 continue; 2323 2324 /* Use the distance array to find the distance */ 2325 val = node_distance(node, n); 2326 2327 /* Penalize nodes under us ("prefer the next node") */ 2328 val += (n < node); 2329 2330 /* Give preference to headless and unused nodes */ 2331 tmp = cpumask_of_node(n); 2332 if (!cpumask_empty(tmp)) 2333 val += PENALTY_FOR_NODE_WITH_CPUS; 2334 2335 /* Slight preference for less loaded node */ 2336 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2337 val += node_load[n]; 2338 2339 if (val < min_val) { 2340 min_val = val; 2341 best_node = n; 2342 } 2343 } 2344 2345 if (best_node >= 0) 2346 node_set(best_node, *used_node_mask); 2347 2348 return best_node; 2349} 2350 2351 2352/* 2353 * Build zonelists ordered by node and zones within node. 2354 * This results in maximum locality--normal zone overflows into local 2355 * DMA zone, if any--but risks exhausting DMA zone. 2356 */ 2357static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2358{ 2359 int j; 2360 struct zonelist *zonelist; 2361 2362 zonelist = &pgdat->node_zonelists[0]; 2363 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2364 ; 2365 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2366 MAX_NR_ZONES - 1); 2367 zonelist->_zonerefs[j].zone = NULL; 2368 zonelist->_zonerefs[j].zone_idx = 0; 2369} 2370 2371/* 2372 * Build gfp_thisnode zonelists 2373 */ 2374static void build_thisnode_zonelists(pg_data_t *pgdat) 2375{ 2376 int j; 2377 struct zonelist *zonelist; 2378 2379 zonelist = &pgdat->node_zonelists[1]; 2380 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2381 zonelist->_zonerefs[j].zone = NULL; 2382 zonelist->_zonerefs[j].zone_idx = 0; 2383} 2384 2385/* 2386 * Build zonelists ordered by zone and nodes within zones. 2387 * This results in conserving DMA zone[s] until all Normal memory is 2388 * exhausted, but results in overflowing to remote node while memory 2389 * may still exist in local DMA zone. 2390 */ 2391static int node_order[MAX_NUMNODES]; 2392 2393static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2394{ 2395 int pos, j, node; 2396 int zone_type; /* needs to be signed */ 2397 struct zone *z; 2398 struct zonelist *zonelist; 2399 2400 zonelist = &pgdat->node_zonelists[0]; 2401 pos = 0; 2402 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2403 for (j = 0; j < nr_nodes; j++) { 2404 node = node_order[j]; 2405 z = &NODE_DATA(node)->node_zones[zone_type]; 2406 if (populated_zone(z)) { 2407 zoneref_set_zone(z, 2408 &zonelist->_zonerefs[pos++]); 2409 check_highest_zone(zone_type); 2410 } 2411 } 2412 } 2413 zonelist->_zonerefs[pos].zone = NULL; 2414 zonelist->_zonerefs[pos].zone_idx = 0; 2415} 2416 2417static int default_zonelist_order(void) 2418{ 2419 int nid, zone_type; 2420 unsigned long low_kmem_size,total_size; 2421 struct zone *z; 2422 int average_size; 2423 /* 2424 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2425 * If they are really small and used heavily, the system can fall 2426 * into OOM very easily. 2427 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2428 */ 2429 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2430 low_kmem_size = 0; 2431 total_size = 0; 2432 for_each_online_node(nid) { 2433 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2434 z = &NODE_DATA(nid)->node_zones[zone_type]; 2435 if (populated_zone(z)) { 2436 if (zone_type < ZONE_NORMAL) 2437 low_kmem_size += z->present_pages; 2438 total_size += z->present_pages; 2439 } 2440 } 2441 } 2442 if (!low_kmem_size || /* there are no DMA area. */ 2443 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2444 return ZONELIST_ORDER_NODE; 2445 /* 2446 * look into each node's config. 2447 * If there is a node whose DMA/DMA32 memory is very big area on 2448 * local memory, NODE_ORDER may be suitable. 2449 */ 2450 average_size = total_size / 2451 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2452 for_each_online_node(nid) { 2453 low_kmem_size = 0; 2454 total_size = 0; 2455 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2456 z = &NODE_DATA(nid)->node_zones[zone_type]; 2457 if (populated_zone(z)) { 2458 if (zone_type < ZONE_NORMAL) 2459 low_kmem_size += z->present_pages; 2460 total_size += z->present_pages; 2461 } 2462 } 2463 if (low_kmem_size && 2464 total_size > average_size && /* ignore small node */ 2465 low_kmem_size > total_size * 70/100) 2466 return ZONELIST_ORDER_NODE; 2467 } 2468 return ZONELIST_ORDER_ZONE; 2469} 2470 2471static void set_zonelist_order(void) 2472{ 2473 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2474 current_zonelist_order = default_zonelist_order(); 2475 else 2476 current_zonelist_order = user_zonelist_order; 2477} 2478 2479static void build_zonelists(pg_data_t *pgdat) 2480{ 2481 int j, node, load; 2482 enum zone_type i; 2483 nodemask_t used_mask; 2484 int local_node, prev_node; 2485 struct zonelist *zonelist; 2486 int order = current_zonelist_order; 2487 2488 /* initialize zonelists */ 2489 for (i = 0; i < MAX_ZONELISTS; i++) { 2490 zonelist = pgdat->node_zonelists + i; 2491 zonelist->_zonerefs[0].zone = NULL; 2492 zonelist->_zonerefs[0].zone_idx = 0; 2493 } 2494 2495 /* NUMA-aware ordering of nodes */ 2496 local_node = pgdat->node_id; 2497 load = nr_online_nodes; 2498 prev_node = local_node; 2499 nodes_clear(used_mask); 2500 2501 memset(node_load, 0, sizeof(node_load)); 2502 memset(node_order, 0, sizeof(node_order)); 2503 j = 0; 2504 2505 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2506 int distance = node_distance(local_node, node); 2507 2508 /* 2509 * If another node is sufficiently far away then it is better 2510 * to reclaim pages in a zone before going off node. 2511 */ 2512 if (distance > RECLAIM_DISTANCE) 2513 zone_reclaim_mode = 1; 2514 2515 /* 2516 * We don't want to pressure a particular node. 2517 * So adding penalty to the first node in same 2518 * distance group to make it round-robin. 2519 */ 2520 if (distance != node_distance(local_node, prev_node)) 2521 node_load[node] = load; 2522 2523 prev_node = node; 2524 load--; 2525 if (order == ZONELIST_ORDER_NODE) 2526 build_zonelists_in_node_order(pgdat, node); 2527 else 2528 node_order[j++] = node; /* remember order */ 2529 } 2530 2531 if (order == ZONELIST_ORDER_ZONE) { 2532 /* calculate node order -- i.e., DMA last! */ 2533 build_zonelists_in_zone_order(pgdat, j); 2534 } 2535 2536 build_thisnode_zonelists(pgdat); 2537} 2538 2539/* Construct the zonelist performance cache - see further mmzone.h */ 2540static void build_zonelist_cache(pg_data_t *pgdat) 2541{ 2542 struct zonelist *zonelist; 2543 struct zonelist_cache *zlc; 2544 struct zoneref *z; 2545 2546 zonelist = &pgdat->node_zonelists[0]; 2547 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2548 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2549 for (z = zonelist->_zonerefs; z->zone; z++) 2550 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2551} 2552 2553 2554#else /* CONFIG_NUMA */ 2555 2556static void set_zonelist_order(void) 2557{ 2558 current_zonelist_order = ZONELIST_ORDER_ZONE; 2559} 2560 2561static void build_zonelists(pg_data_t *pgdat) 2562{ 2563 int node, local_node; 2564 enum zone_type j; 2565 struct zonelist *zonelist; 2566 2567 local_node = pgdat->node_id; 2568 2569 zonelist = &pgdat->node_zonelists[0]; 2570 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2571 2572 /* 2573 * Now we build the zonelist so that it contains the zones 2574 * of all the other nodes. 2575 * We don't want to pressure a particular node, so when 2576 * building the zones for node N, we make sure that the 2577 * zones coming right after the local ones are those from 2578 * node N+1 (modulo N) 2579 */ 2580 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2581 if (!node_online(node)) 2582 continue; 2583 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2584 MAX_NR_ZONES - 1); 2585 } 2586 for (node = 0; node < local_node; node++) { 2587 if (!node_online(node)) 2588 continue; 2589 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2590 MAX_NR_ZONES - 1); 2591 } 2592 2593 zonelist->_zonerefs[j].zone = NULL; 2594 zonelist->_zonerefs[j].zone_idx = 0; 2595} 2596 2597/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2598static void build_zonelist_cache(pg_data_t *pgdat) 2599{ 2600 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2601} 2602 2603#endif /* CONFIG_NUMA */ 2604 2605/* return values int ....just for stop_machine() */ 2606static int __build_all_zonelists(void *dummy) 2607{ 2608 int nid; 2609 2610 for_each_online_node(nid) { 2611 pg_data_t *pgdat = NODE_DATA(nid); 2612 2613 build_zonelists(pgdat); 2614 build_zonelist_cache(pgdat); 2615 } 2616 return 0; 2617} 2618 2619void build_all_zonelists(void) 2620{ 2621 set_zonelist_order(); 2622 2623 if (system_state == SYSTEM_BOOTING) { 2624 __build_all_zonelists(NULL); 2625 mminit_verify_zonelist(); 2626 cpuset_init_current_mems_allowed(); 2627 } else { 2628 /* we have to stop all cpus to guarantee there is no user 2629 of zonelist */ 2630 stop_machine(__build_all_zonelists, NULL, NULL); 2631 /* cpuset refresh routine should be here */ 2632 } 2633 vm_total_pages = nr_free_pagecache_pages(); 2634 /* 2635 * Disable grouping by mobility if the number of pages in the 2636 * system is too low to allow the mechanism to work. It would be 2637 * more accurate, but expensive to check per-zone. This check is 2638 * made on memory-hotadd so a system can start with mobility 2639 * disabled and enable it later 2640 */ 2641 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2642 page_group_by_mobility_disabled = 1; 2643 else 2644 page_group_by_mobility_disabled = 0; 2645 2646 printk("Built %i zonelists in %s order, mobility grouping %s. " 2647 "Total pages: %ld\n", 2648 nr_online_nodes, 2649 zonelist_order_name[current_zonelist_order], 2650 page_group_by_mobility_disabled ? "off" : "on", 2651 vm_total_pages); 2652#ifdef CONFIG_NUMA 2653 printk("Policy zone: %s\n", zone_names[policy_zone]); 2654#endif 2655} 2656 2657/* 2658 * Helper functions to size the waitqueue hash table. 2659 * Essentially these want to choose hash table sizes sufficiently 2660 * large so that collisions trying to wait on pages are rare. 2661 * But in fact, the number of active page waitqueues on typical 2662 * systems is ridiculously low, less than 200. So this is even 2663 * conservative, even though it seems large. 2664 * 2665 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2666 * waitqueues, i.e. the size of the waitq table given the number of pages. 2667 */ 2668#define PAGES_PER_WAITQUEUE 256 2669 2670#ifndef CONFIG_MEMORY_HOTPLUG 2671static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2672{ 2673 unsigned long size = 1; 2674 2675 pages /= PAGES_PER_WAITQUEUE; 2676 2677 while (size < pages) 2678 size <<= 1; 2679 2680 /* 2681 * Once we have dozens or even hundreds of threads sleeping 2682 * on IO we've got bigger problems than wait queue collision. 2683 * Limit the size of the wait table to a reasonable size. 2684 */ 2685 size = min(size, 4096UL); 2686 2687 return max(size, 4UL); 2688} 2689#else 2690/* 2691 * A zone's size might be changed by hot-add, so it is not possible to determine 2692 * a suitable size for its wait_table. So we use the maximum size now. 2693 * 2694 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2695 * 2696 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2697 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2698 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2699 * 2700 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2701 * or more by the traditional way. (See above). It equals: 2702 * 2703 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2704 * ia64(16K page size) : = ( 8G + 4M)byte. 2705 * powerpc (64K page size) : = (32G +16M)byte. 2706 */ 2707static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2708{ 2709 return 4096UL; 2710} 2711#endif 2712 2713/* 2714 * This is an integer logarithm so that shifts can be used later 2715 * to extract the more random high bits from the multiplicative 2716 * hash function before the remainder is taken. 2717 */ 2718static inline unsigned long wait_table_bits(unsigned long size) 2719{ 2720 return ffz(~size); 2721} 2722 2723#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2724 2725/* 2726 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2727 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2728 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2729 * higher will lead to a bigger reserve which will get freed as contiguous 2730 * blocks as reclaim kicks in 2731 */ 2732static void setup_zone_migrate_reserve(struct zone *zone) 2733{ 2734 unsigned long start_pfn, pfn, end_pfn; 2735 struct page *page; 2736 unsigned long reserve, block_migratetype; 2737 2738 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2739 start_pfn = zone->zone_start_pfn; 2740 end_pfn = start_pfn + zone->spanned_pages; 2741 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2742 pageblock_order; 2743 2744 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2745 if (!pfn_valid(pfn)) 2746 continue; 2747 page = pfn_to_page(pfn); 2748 2749 /* Watch out for overlapping nodes */ 2750 if (page_to_nid(page) != zone_to_nid(zone)) 2751 continue; 2752 2753 /* Blocks with reserved pages will never free, skip them. */ 2754 if (PageReserved(page)) 2755 continue; 2756 2757 block_migratetype = get_pageblock_migratetype(page); 2758 2759 /* If this block is reserved, account for it */ 2760 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2761 reserve--; 2762 continue; 2763 } 2764 2765 /* Suitable for reserving if this block is movable */ 2766 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2767 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2768 move_freepages_block(zone, page, MIGRATE_RESERVE); 2769 reserve--; 2770 continue; 2771 } 2772 2773 /* 2774 * If the reserve is met and this is a previous reserved block, 2775 * take it back 2776 */ 2777 if (block_migratetype == MIGRATE_RESERVE) { 2778 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2779 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2780 } 2781 } 2782} 2783 2784/* 2785 * Initially all pages are reserved - free ones are freed 2786 * up by free_all_bootmem() once the early boot process is 2787 * done. Non-atomic initialization, single-pass. 2788 */ 2789void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2790 unsigned long start_pfn, enum memmap_context context) 2791{ 2792 struct page *page; 2793 unsigned long end_pfn = start_pfn + size; 2794 unsigned long pfn; 2795 struct zone *z; 2796 2797 if (highest_memmap_pfn < end_pfn - 1) 2798 highest_memmap_pfn = end_pfn - 1; 2799 2800 z = &NODE_DATA(nid)->node_zones[zone]; 2801 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2802 /* 2803 * There can be holes in boot-time mem_map[]s 2804 * handed to this function. They do not 2805 * exist on hotplugged memory. 2806 */ 2807 if (context == MEMMAP_EARLY) { 2808 if (!early_pfn_valid(pfn)) 2809 continue; 2810 if (!early_pfn_in_nid(pfn, nid)) 2811 continue; 2812 } 2813 page = pfn_to_page(pfn); 2814 set_page_links(page, zone, nid, pfn); 2815 mminit_verify_page_links(page, zone, nid, pfn); 2816 init_page_count(page); 2817 reset_page_mapcount(page); 2818 SetPageReserved(page); 2819 /* 2820 * Mark the block movable so that blocks are reserved for 2821 * movable at startup. This will force kernel allocations 2822 * to reserve their blocks rather than leaking throughout 2823 * the address space during boot when many long-lived 2824 * kernel allocations are made. Later some blocks near 2825 * the start are marked MIGRATE_RESERVE by 2826 * setup_zone_migrate_reserve() 2827 * 2828 * bitmap is created for zone's valid pfn range. but memmap 2829 * can be created for invalid pages (for alignment) 2830 * check here not to call set_pageblock_migratetype() against 2831 * pfn out of zone. 2832 */ 2833 if ((z->zone_start_pfn <= pfn) 2834 && (pfn < z->zone_start_pfn + z->spanned_pages) 2835 && !(pfn & (pageblock_nr_pages - 1))) 2836 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2837 2838 INIT_LIST_HEAD(&page->lru); 2839#ifdef WANT_PAGE_VIRTUAL 2840 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2841 if (!is_highmem_idx(zone)) 2842 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2843#endif 2844 } 2845} 2846 2847static void __meminit zone_init_free_lists(struct zone *zone) 2848{ 2849 int order, t; 2850 for_each_migratetype_order(order, t) { 2851 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 2852 zone->free_area[order].nr_free = 0; 2853 } 2854} 2855 2856#ifndef __HAVE_ARCH_MEMMAP_INIT 2857#define memmap_init(size, nid, zone, start_pfn) \ 2858 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 2859#endif 2860 2861static int zone_batchsize(struct zone *zone) 2862{ 2863#ifdef CONFIG_MMU 2864 int batch; 2865 2866 /* 2867 * The per-cpu-pages pools are set to around 1000th of the 2868 * size of the zone. But no more than 1/2 of a meg. 2869 * 2870 * OK, so we don't know how big the cache is. So guess. 2871 */ 2872 batch = zone->present_pages / 1024; 2873 if (batch * PAGE_SIZE > 512 * 1024) 2874 batch = (512 * 1024) / PAGE_SIZE; 2875 batch /= 4; /* We effectively *= 4 below */ 2876 if (batch < 1) 2877 batch = 1; 2878 2879 /* 2880 * Clamp the batch to a 2^n - 1 value. Having a power 2881 * of 2 value was found to be more likely to have 2882 * suboptimal cache aliasing properties in some cases. 2883 * 2884 * For example if 2 tasks are alternately allocating 2885 * batches of pages, one task can end up with a lot 2886 * of pages of one half of the possible page colors 2887 * and the other with pages of the other colors. 2888 */ 2889 batch = rounddown_pow_of_two(batch + batch/2) - 1; 2890 2891 return batch; 2892 2893#else 2894 /* The deferral and batching of frees should be suppressed under NOMMU 2895 * conditions. 2896 * 2897 * The problem is that NOMMU needs to be able to allocate large chunks 2898 * of contiguous memory as there's no hardware page translation to 2899 * assemble apparent contiguous memory from discontiguous pages. 2900 * 2901 * Queueing large contiguous runs of pages for batching, however, 2902 * causes the pages to actually be freed in smaller chunks. As there 2903 * can be a significant delay between the individual batches being 2904 * recycled, this leads to the once large chunks of space being 2905 * fragmented and becoming unavailable for high-order allocations. 2906 */ 2907 return 0; 2908#endif 2909} 2910 2911static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 2912{ 2913 struct per_cpu_pages *pcp; 2914 2915 memset(p, 0, sizeof(*p)); 2916 2917 pcp = &p->pcp; 2918 pcp->count = 0; 2919 pcp->high = 6 * batch; 2920 pcp->batch = max(1UL, 1 * batch); 2921 INIT_LIST_HEAD(&pcp->list); 2922} 2923 2924/* 2925 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 2926 * to the value high for the pageset p. 2927 */ 2928 2929static void setup_pagelist_highmark(struct per_cpu_pageset *p, 2930 unsigned long high) 2931{ 2932 struct per_cpu_pages *pcp; 2933 2934 pcp = &p->pcp; 2935 pcp->high = high; 2936 pcp->batch = max(1UL, high/4); 2937 if ((high/4) > (PAGE_SHIFT * 8)) 2938 pcp->batch = PAGE_SHIFT * 8; 2939} 2940 2941 2942#ifdef CONFIG_NUMA 2943/* 2944 * Boot pageset table. One per cpu which is going to be used for all 2945 * zones and all nodes. The parameters will be set in such a way 2946 * that an item put on a list will immediately be handed over to 2947 * the buddy list. This is safe since pageset manipulation is done 2948 * with interrupts disabled. 2949 * 2950 * Some NUMA counter updates may also be caught by the boot pagesets. 2951 * 2952 * The boot_pagesets must be kept even after bootup is complete for 2953 * unused processors and/or zones. They do play a role for bootstrapping 2954 * hotplugged processors. 2955 * 2956 * zoneinfo_show() and maybe other functions do 2957 * not check if the processor is online before following the pageset pointer. 2958 * Other parts of the kernel may not check if the zone is available. 2959 */ 2960static struct per_cpu_pageset boot_pageset[NR_CPUS]; 2961 2962/* 2963 * Dynamically allocate memory for the 2964 * per cpu pageset array in struct zone. 2965 */ 2966static int __cpuinit process_zones(int cpu) 2967{ 2968 struct zone *zone, *dzone; 2969 int node = cpu_to_node(cpu); 2970 2971 node_set_state(node, N_CPU); /* this node has a cpu */ 2972 2973 for_each_populated_zone(zone) { 2974 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 2975 GFP_KERNEL, node); 2976 if (!zone_pcp(zone, cpu)) 2977 goto bad; 2978 2979 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 2980 2981 if (percpu_pagelist_fraction) 2982 setup_pagelist_highmark(zone_pcp(zone, cpu), 2983 (zone->present_pages / percpu_pagelist_fraction)); 2984 } 2985 2986 return 0; 2987bad: 2988 for_each_zone(dzone) { 2989 if (!populated_zone(dzone)) 2990 continue; 2991 if (dzone == zone) 2992 break; 2993 kfree(zone_pcp(dzone, cpu)); 2994 zone_pcp(dzone, cpu) = NULL; 2995 } 2996 return -ENOMEM; 2997} 2998 2999static inline void free_zone_pagesets(int cpu) 3000{ 3001 struct zone *zone; 3002 3003 for_each_zone(zone) { 3004 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 3005 3006 /* Free per_cpu_pageset if it is slab allocated */ 3007 if (pset != &boot_pageset[cpu]) 3008 kfree(pset); 3009 zone_pcp(zone, cpu) = NULL; 3010 } 3011} 3012 3013static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 3014 unsigned long action, 3015 void *hcpu) 3016{ 3017 int cpu = (long)hcpu; 3018 int ret = NOTIFY_OK; 3019 3020 switch (action) { 3021 case CPU_UP_PREPARE: 3022 case CPU_UP_PREPARE_FROZEN: 3023 if (process_zones(cpu)) 3024 ret = NOTIFY_BAD; 3025 break; 3026 case CPU_UP_CANCELED: 3027 case CPU_UP_CANCELED_FROZEN: 3028 case CPU_DEAD: 3029 case CPU_DEAD_FROZEN: 3030 free_zone_pagesets(cpu); 3031 break; 3032 default: 3033 break; 3034 } 3035 return ret; 3036} 3037 3038static struct notifier_block __cpuinitdata pageset_notifier = 3039 { &pageset_cpuup_callback, NULL, 0 }; 3040 3041void __init setup_per_cpu_pageset(void) 3042{ 3043 int err; 3044 3045 /* Initialize per_cpu_pageset for cpu 0. 3046 * A cpuup callback will do this for every cpu 3047 * as it comes online 3048 */ 3049 err = process_zones(smp_processor_id()); 3050 BUG_ON(err); 3051 register_cpu_notifier(&pageset_notifier); 3052} 3053 3054#endif 3055 3056static noinline __init_refok 3057int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3058{ 3059 int i; 3060 struct pglist_data *pgdat = zone->zone_pgdat; 3061 size_t alloc_size; 3062 3063 /* 3064 * The per-page waitqueue mechanism uses hashed waitqueues 3065 * per zone. 3066 */ 3067 zone->wait_table_hash_nr_entries = 3068 wait_table_hash_nr_entries(zone_size_pages); 3069 zone->wait_table_bits = 3070 wait_table_bits(zone->wait_table_hash_nr_entries); 3071 alloc_size = zone->wait_table_hash_nr_entries 3072 * sizeof(wait_queue_head_t); 3073 3074 if (!slab_is_available()) { 3075 zone->wait_table = (wait_queue_head_t *) 3076 alloc_bootmem_node(pgdat, alloc_size); 3077 } else { 3078 /* 3079 * This case means that a zone whose size was 0 gets new memory 3080 * via memory hot-add. 3081 * But it may be the case that a new node was hot-added. In 3082 * this case vmalloc() will not be able to use this new node's 3083 * memory - this wait_table must be initialized to use this new 3084 * node itself as well. 3085 * To use this new node's memory, further consideration will be 3086 * necessary. 3087 */ 3088 zone->wait_table = vmalloc(alloc_size); 3089 } 3090 if (!zone->wait_table) 3091 return -ENOMEM; 3092 3093 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3094 init_waitqueue_head(zone->wait_table + i); 3095 3096 return 0; 3097} 3098 3099static __meminit void zone_pcp_init(struct zone *zone) 3100{ 3101 int cpu; 3102 unsigned long batch = zone_batchsize(zone); 3103 3104 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3105#ifdef CONFIG_NUMA 3106 /* Early boot. Slab allocator not functional yet */ 3107 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3108 setup_pageset(&boot_pageset[cpu],0); 3109#else 3110 setup_pageset(zone_pcp(zone,cpu), batch); 3111#endif 3112 } 3113 if (zone->present_pages) 3114 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 3115 zone->name, zone->present_pages, batch); 3116} 3117 3118__meminit int init_currently_empty_zone(struct zone *zone, 3119 unsigned long zone_start_pfn, 3120 unsigned long size, 3121 enum memmap_context context) 3122{ 3123 struct pglist_data *pgdat = zone->zone_pgdat; 3124 int ret; 3125 ret = zone_wait_table_init(zone, size); 3126 if (ret) 3127 return ret; 3128 pgdat->nr_zones = zone_idx(zone) + 1; 3129 3130 zone->zone_start_pfn = zone_start_pfn; 3131 3132 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3133 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3134 pgdat->node_id, 3135 (unsigned long)zone_idx(zone), 3136 zone_start_pfn, (zone_start_pfn + size)); 3137 3138 zone_init_free_lists(zone); 3139 3140 return 0; 3141} 3142 3143#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3144/* 3145 * Basic iterator support. Return the first range of PFNs for a node 3146 * Note: nid == MAX_NUMNODES returns first region regardless of node 3147 */ 3148static int __meminit first_active_region_index_in_nid(int nid) 3149{ 3150 int i; 3151 3152 for (i = 0; i < nr_nodemap_entries; i++) 3153 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3154 return i; 3155 3156 return -1; 3157} 3158 3159/* 3160 * Basic iterator support. Return the next active range of PFNs for a node 3161 * Note: nid == MAX_NUMNODES returns next region regardless of node 3162 */ 3163static int __meminit next_active_region_index_in_nid(int index, int nid) 3164{ 3165 for (index = index + 1; index < nr_nodemap_entries; index++) 3166 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3167 return index; 3168 3169 return -1; 3170} 3171 3172#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3173/* 3174 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3175 * Architectures may implement their own version but if add_active_range() 3176 * was used and there are no special requirements, this is a convenient 3177 * alternative 3178 */ 3179int __meminit __early_pfn_to_nid(unsigned long pfn) 3180{ 3181 int i; 3182 3183 for (i = 0; i < nr_nodemap_entries; i++) { 3184 unsigned long start_pfn = early_node_map[i].start_pfn; 3185 unsigned long end_pfn = early_node_map[i].end_pfn; 3186 3187 if (start_pfn <= pfn && pfn < end_pfn) 3188 return early_node_map[i].nid; 3189 } 3190 /* This is a memory hole */ 3191 return -1; 3192} 3193#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3194 3195int __meminit early_pfn_to_nid(unsigned long pfn) 3196{ 3197 int nid; 3198 3199 nid = __early_pfn_to_nid(pfn); 3200 if (nid >= 0) 3201 return nid; 3202 /* just returns 0 */ 3203 return 0; 3204} 3205 3206#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3207bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3208{ 3209 int nid; 3210 3211 nid = __early_pfn_to_nid(pfn); 3212 if (nid >= 0 && nid != node) 3213 return false; 3214 return true; 3215} 3216#endif 3217 3218/* Basic iterator support to walk early_node_map[] */ 3219#define for_each_active_range_index_in_nid(i, nid) \ 3220 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3221 i = next_active_region_index_in_nid(i, nid)) 3222 3223/** 3224 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3225 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3226 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3227 * 3228 * If an architecture guarantees that all ranges registered with 3229 * add_active_ranges() contain no holes and may be freed, this 3230 * this function may be used instead of calling free_bootmem() manually. 3231 */ 3232void __init free_bootmem_with_active_regions(int nid, 3233 unsigned long max_low_pfn) 3234{ 3235 int i; 3236 3237 for_each_active_range_index_in_nid(i, nid) { 3238 unsigned long size_pages = 0; 3239 unsigned long end_pfn = early_node_map[i].end_pfn; 3240 3241 if (early_node_map[i].start_pfn >= max_low_pfn) 3242 continue; 3243 3244 if (end_pfn > max_low_pfn) 3245 end_pfn = max_low_pfn; 3246 3247 size_pages = end_pfn - early_node_map[i].start_pfn; 3248 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3249 PFN_PHYS(early_node_map[i].start_pfn), 3250 size_pages << PAGE_SHIFT); 3251 } 3252} 3253 3254void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3255{ 3256 int i; 3257 int ret; 3258 3259 for_each_active_range_index_in_nid(i, nid) { 3260 ret = work_fn(early_node_map[i].start_pfn, 3261 early_node_map[i].end_pfn, data); 3262 if (ret) 3263 break; 3264 } 3265} 3266/** 3267 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3268 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3269 * 3270 * If an architecture guarantees that all ranges registered with 3271 * add_active_ranges() contain no holes and may be freed, this 3272 * function may be used instead of calling memory_present() manually. 3273 */ 3274void __init sparse_memory_present_with_active_regions(int nid) 3275{ 3276 int i; 3277 3278 for_each_active_range_index_in_nid(i, nid) 3279 memory_present(early_node_map[i].nid, 3280 early_node_map[i].start_pfn, 3281 early_node_map[i].end_pfn); 3282} 3283 3284/** 3285 * get_pfn_range_for_nid - Return the start and end page frames for a node 3286 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3287 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3288 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3289 * 3290 * It returns the start and end page frame of a node based on information 3291 * provided by an arch calling add_active_range(). If called for a node 3292 * with no available memory, a warning is printed and the start and end 3293 * PFNs will be 0. 3294 */ 3295void __meminit get_pfn_range_for_nid(unsigned int nid, 3296 unsigned long *start_pfn, unsigned long *end_pfn) 3297{ 3298 int i; 3299 *start_pfn = -1UL; 3300 *end_pfn = 0; 3301 3302 for_each_active_range_index_in_nid(i, nid) { 3303 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3304 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3305 } 3306 3307 if (*start_pfn == -1UL) 3308 *start_pfn = 0; 3309} 3310 3311/* 3312 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3313 * assumption is made that zones within a node are ordered in monotonic 3314 * increasing memory addresses so that the "highest" populated zone is used 3315 */ 3316static void __init find_usable_zone_for_movable(void) 3317{ 3318 int zone_index; 3319 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3320 if (zone_index == ZONE_MOVABLE) 3321 continue; 3322 3323 if (arch_zone_highest_possible_pfn[zone_index] > 3324 arch_zone_lowest_possible_pfn[zone_index]) 3325 break; 3326 } 3327 3328 VM_BUG_ON(zone_index == -1); 3329 movable_zone = zone_index; 3330} 3331 3332/* 3333 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3334 * because it is sized independant of architecture. Unlike the other zones, 3335 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3336 * in each node depending on the size of each node and how evenly kernelcore 3337 * is distributed. This helper function adjusts the zone ranges 3338 * provided by the architecture for a given node by using the end of the 3339 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3340 * zones within a node are in order of monotonic increases memory addresses 3341 */ 3342static void __meminit adjust_zone_range_for_zone_movable(int nid, 3343 unsigned long zone_type, 3344 unsigned long node_start_pfn, 3345 unsigned long node_end_pfn, 3346 unsigned long *zone_start_pfn, 3347 unsigned long *zone_end_pfn) 3348{ 3349 /* Only adjust if ZONE_MOVABLE is on this node */ 3350 if (zone_movable_pfn[nid]) { 3351 /* Size ZONE_MOVABLE */ 3352 if (zone_type == ZONE_MOVABLE) { 3353 *zone_start_pfn = zone_movable_pfn[nid]; 3354 *zone_end_pfn = min(node_end_pfn, 3355 arch_zone_highest_possible_pfn[movable_zone]); 3356 3357 /* Adjust for ZONE_MOVABLE starting within this range */ 3358 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3359 *zone_end_pfn > zone_movable_pfn[nid]) { 3360 *zone_end_pfn = zone_movable_pfn[nid]; 3361 3362 /* Check if this whole range is within ZONE_MOVABLE */ 3363 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3364 *zone_start_pfn = *zone_end_pfn; 3365 } 3366} 3367 3368/* 3369 * Return the number of pages a zone spans in a node, including holes 3370 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3371 */ 3372static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3373 unsigned long zone_type, 3374 unsigned long *ignored) 3375{ 3376 unsigned long node_start_pfn, node_end_pfn; 3377 unsigned long zone_start_pfn, zone_end_pfn; 3378 3379 /* Get the start and end of the node and zone */ 3380 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3381 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3382 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3383 adjust_zone_range_for_zone_movable(nid, zone_type, 3384 node_start_pfn, node_end_pfn, 3385 &zone_start_pfn, &zone_end_pfn); 3386 3387 /* Check that this node has pages within the zone's required range */ 3388 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3389 return 0; 3390 3391 /* Move the zone boundaries inside the node if necessary */ 3392 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3393 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3394 3395 /* Return the spanned pages */ 3396 return zone_end_pfn - zone_start_pfn; 3397} 3398 3399/* 3400 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3401 * then all holes in the requested range will be accounted for. 3402 */ 3403static unsigned long __meminit __absent_pages_in_range(int nid, 3404 unsigned long range_start_pfn, 3405 unsigned long range_end_pfn) 3406{ 3407 int i = 0; 3408 unsigned long prev_end_pfn = 0, hole_pages = 0; 3409 unsigned long start_pfn; 3410 3411 /* Find the end_pfn of the first active range of pfns in the node */ 3412 i = first_active_region_index_in_nid(nid); 3413 if (i == -1) 3414 return 0; 3415 3416 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3417 3418 /* Account for ranges before physical memory on this node */ 3419 if (early_node_map[i].start_pfn > range_start_pfn) 3420 hole_pages = prev_end_pfn - range_start_pfn; 3421 3422 /* Find all holes for the zone within the node */ 3423 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3424 3425 /* No need to continue if prev_end_pfn is outside the zone */ 3426 if (prev_end_pfn >= range_end_pfn) 3427 break; 3428 3429 /* Make sure the end of the zone is not within the hole */ 3430 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3431 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3432 3433 /* Update the hole size cound and move on */ 3434 if (start_pfn > range_start_pfn) { 3435 BUG_ON(prev_end_pfn > start_pfn); 3436 hole_pages += start_pfn - prev_end_pfn; 3437 } 3438 prev_end_pfn = early_node_map[i].end_pfn; 3439 } 3440 3441 /* Account for ranges past physical memory on this node */ 3442 if (range_end_pfn > prev_end_pfn) 3443 hole_pages += range_end_pfn - 3444 max(range_start_pfn, prev_end_pfn); 3445 3446 return hole_pages; 3447} 3448 3449/** 3450 * absent_pages_in_range - Return number of page frames in holes within a range 3451 * @start_pfn: The start PFN to start searching for holes 3452 * @end_pfn: The end PFN to stop searching for holes 3453 * 3454 * It returns the number of pages frames in memory holes within a range. 3455 */ 3456unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3457 unsigned long end_pfn) 3458{ 3459 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3460} 3461 3462/* Return the number of page frames in holes in a zone on a node */ 3463static unsigned long __meminit zone_absent_pages_in_node(int nid, 3464 unsigned long zone_type, 3465 unsigned long *ignored) 3466{ 3467 unsigned long node_start_pfn, node_end_pfn; 3468 unsigned long zone_start_pfn, zone_end_pfn; 3469 3470 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3471 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3472 node_start_pfn); 3473 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3474 node_end_pfn); 3475 3476 adjust_zone_range_for_zone_movable(nid, zone_type, 3477 node_start_pfn, node_end_pfn, 3478 &zone_start_pfn, &zone_end_pfn); 3479 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3480} 3481 3482#else 3483static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3484 unsigned long zone_type, 3485 unsigned long *zones_size) 3486{ 3487 return zones_size[zone_type]; 3488} 3489 3490static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3491 unsigned long zone_type, 3492 unsigned long *zholes_size) 3493{ 3494 if (!zholes_size) 3495 return 0; 3496 3497 return zholes_size[zone_type]; 3498} 3499 3500#endif 3501 3502static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3503 unsigned long *zones_size, unsigned long *zholes_size) 3504{ 3505 unsigned long realtotalpages, totalpages = 0; 3506 enum zone_type i; 3507 3508 for (i = 0; i < MAX_NR_ZONES; i++) 3509 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3510 zones_size); 3511 pgdat->node_spanned_pages = totalpages; 3512 3513 realtotalpages = totalpages; 3514 for (i = 0; i < MAX_NR_ZONES; i++) 3515 realtotalpages -= 3516 zone_absent_pages_in_node(pgdat->node_id, i, 3517 zholes_size); 3518 pgdat->node_present_pages = realtotalpages; 3519 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3520 realtotalpages); 3521} 3522 3523#ifndef CONFIG_SPARSEMEM 3524/* 3525 * Calculate the size of the zone->blockflags rounded to an unsigned long 3526 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3527 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3528 * round what is now in bits to nearest long in bits, then return it in 3529 * bytes. 3530 */ 3531static unsigned long __init usemap_size(unsigned long zonesize) 3532{ 3533 unsigned long usemapsize; 3534 3535 usemapsize = roundup(zonesize, pageblock_nr_pages); 3536 usemapsize = usemapsize >> pageblock_order; 3537 usemapsize *= NR_PAGEBLOCK_BITS; 3538 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3539 3540 return usemapsize / 8; 3541} 3542 3543static void __init setup_usemap(struct pglist_data *pgdat, 3544 struct zone *zone, unsigned long zonesize) 3545{ 3546 unsigned long usemapsize = usemap_size(zonesize); 3547 zone->pageblock_flags = NULL; 3548 if (usemapsize) 3549 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3550} 3551#else 3552static void inline setup_usemap(struct pglist_data *pgdat, 3553 struct zone *zone, unsigned long zonesize) {} 3554#endif /* CONFIG_SPARSEMEM */ 3555 3556#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3557 3558/* Return a sensible default order for the pageblock size. */ 3559static inline int pageblock_default_order(void) 3560{ 3561 if (HPAGE_SHIFT > PAGE_SHIFT) 3562 return HUGETLB_PAGE_ORDER; 3563 3564 return MAX_ORDER-1; 3565} 3566 3567/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3568static inline void __init set_pageblock_order(unsigned int order) 3569{ 3570 /* Check that pageblock_nr_pages has not already been setup */ 3571 if (pageblock_order) 3572 return; 3573 3574 /* 3575 * Assume the largest contiguous order of interest is a huge page. 3576 * This value may be variable depending on boot parameters on IA64 3577 */ 3578 pageblock_order = order; 3579} 3580#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3581 3582/* 3583 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3584 * and pageblock_default_order() are unused as pageblock_order is set 3585 * at compile-time. See include/linux/pageblock-flags.h for the values of 3586 * pageblock_order based on the kernel config 3587 */ 3588static inline int pageblock_default_order(unsigned int order) 3589{ 3590 return MAX_ORDER-1; 3591} 3592#define set_pageblock_order(x) do {} while (0) 3593 3594#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3595 3596/* 3597 * Set up the zone data structures: 3598 * - mark all pages reserved 3599 * - mark all memory queues empty 3600 * - clear the memory bitmaps 3601 */ 3602static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3603 unsigned long *zones_size, unsigned long *zholes_size) 3604{ 3605 enum zone_type j; 3606 int nid = pgdat->node_id; 3607 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3608 int ret; 3609 3610 pgdat_resize_init(pgdat); 3611 pgdat->nr_zones = 0; 3612 init_waitqueue_head(&pgdat->kswapd_wait); 3613 pgdat->kswapd_max_order = 0; 3614 pgdat_page_cgroup_init(pgdat); 3615 3616 for (j = 0; j < MAX_NR_ZONES; j++) { 3617 struct zone *zone = pgdat->node_zones + j; 3618 unsigned long size, realsize, memmap_pages; 3619 enum lru_list l; 3620 3621 size = zone_spanned_pages_in_node(nid, j, zones_size); 3622 realsize = size - zone_absent_pages_in_node(nid, j, 3623 zholes_size); 3624 3625 /* 3626 * Adjust realsize so that it accounts for how much memory 3627 * is used by this zone for memmap. This affects the watermark 3628 * and per-cpu initialisations 3629 */ 3630 memmap_pages = 3631 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3632 if (realsize >= memmap_pages) { 3633 realsize -= memmap_pages; 3634 if (memmap_pages) 3635 printk(KERN_DEBUG 3636 " %s zone: %lu pages used for memmap\n", 3637 zone_names[j], memmap_pages); 3638 } else 3639 printk(KERN_WARNING 3640 " %s zone: %lu pages exceeds realsize %lu\n", 3641 zone_names[j], memmap_pages, realsize); 3642 3643 /* Account for reserved pages */ 3644 if (j == 0 && realsize > dma_reserve) { 3645 realsize -= dma_reserve; 3646 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3647 zone_names[0], dma_reserve); 3648 } 3649 3650 if (!is_highmem_idx(j)) 3651 nr_kernel_pages += realsize; 3652 nr_all_pages += realsize; 3653 3654 zone->spanned_pages = size; 3655 zone->present_pages = realsize; 3656#ifdef CONFIG_NUMA 3657 zone->node = nid; 3658 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3659 / 100; 3660 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3661#endif 3662 zone->name = zone_names[j]; 3663 spin_lock_init(&zone->lock); 3664 spin_lock_init(&zone->lru_lock); 3665 zone_seqlock_init(zone); 3666 zone->zone_pgdat = pgdat; 3667 3668 zone->prev_priority = DEF_PRIORITY; 3669 3670 zone_pcp_init(zone); 3671 for_each_lru(l) { 3672 INIT_LIST_HEAD(&zone->lru[l].list); 3673 zone->lru[l].nr_saved_scan = 0; 3674 } 3675 zone->reclaim_stat.recent_rotated[0] = 0; 3676 zone->reclaim_stat.recent_rotated[1] = 0; 3677 zone->reclaim_stat.recent_scanned[0] = 0; 3678 zone->reclaim_stat.recent_scanned[1] = 0; 3679 zap_zone_vm_stats(zone); 3680 zone->flags = 0; 3681 if (!size) 3682 continue; 3683 3684 set_pageblock_order(pageblock_default_order()); 3685 setup_usemap(pgdat, zone, size); 3686 ret = init_currently_empty_zone(zone, zone_start_pfn, 3687 size, MEMMAP_EARLY); 3688 BUG_ON(ret); 3689 memmap_init(size, nid, j, zone_start_pfn); 3690 zone_start_pfn += size; 3691 } 3692} 3693 3694static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3695{ 3696 /* Skip empty nodes */ 3697 if (!pgdat->node_spanned_pages) 3698 return; 3699 3700#ifdef CONFIG_FLAT_NODE_MEM_MAP 3701 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3702 if (!pgdat->node_mem_map) { 3703 unsigned long size, start, end; 3704 struct page *map; 3705 3706 /* 3707 * The zone's endpoints aren't required to be MAX_ORDER 3708 * aligned but the node_mem_map endpoints must be in order 3709 * for the buddy allocator to function correctly. 3710 */ 3711 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3712 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3713 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3714 size = (end - start) * sizeof(struct page); 3715 map = alloc_remap(pgdat->node_id, size); 3716 if (!map) 3717 map = alloc_bootmem_node(pgdat, size); 3718 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3719 } 3720#ifndef CONFIG_NEED_MULTIPLE_NODES 3721 /* 3722 * With no DISCONTIG, the global mem_map is just set as node 0's 3723 */ 3724 if (pgdat == NODE_DATA(0)) { 3725 mem_map = NODE_DATA(0)->node_mem_map; 3726#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3727 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3728 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3729#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3730 } 3731#endif 3732#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3733} 3734 3735void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3736 unsigned long node_start_pfn, unsigned long *zholes_size) 3737{ 3738 pg_data_t *pgdat = NODE_DATA(nid); 3739 3740 pgdat->node_id = nid; 3741 pgdat->node_start_pfn = node_start_pfn; 3742 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3743 3744 alloc_node_mem_map(pgdat); 3745#ifdef CONFIG_FLAT_NODE_MEM_MAP 3746 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3747 nid, (unsigned long)pgdat, 3748 (unsigned long)pgdat->node_mem_map); 3749#endif 3750 3751 free_area_init_core(pgdat, zones_size, zholes_size); 3752} 3753 3754#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3755 3756#if MAX_NUMNODES > 1 3757/* 3758 * Figure out the number of possible node ids. 3759 */ 3760static void __init setup_nr_node_ids(void) 3761{ 3762 unsigned int node; 3763 unsigned int highest = 0; 3764 3765 for_each_node_mask(node, node_possible_map) 3766 highest = node; 3767 nr_node_ids = highest + 1; 3768} 3769#else 3770static inline void setup_nr_node_ids(void) 3771{ 3772} 3773#endif 3774 3775/** 3776 * add_active_range - Register a range of PFNs backed by physical memory 3777 * @nid: The node ID the range resides on 3778 * @start_pfn: The start PFN of the available physical memory 3779 * @end_pfn: The end PFN of the available physical memory 3780 * 3781 * These ranges are stored in an early_node_map[] and later used by 3782 * free_area_init_nodes() to calculate zone sizes and holes. If the 3783 * range spans a memory hole, it is up to the architecture to ensure 3784 * the memory is not freed by the bootmem allocator. If possible 3785 * the range being registered will be merged with existing ranges. 3786 */ 3787void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3788 unsigned long end_pfn) 3789{ 3790 int i; 3791 3792 mminit_dprintk(MMINIT_TRACE, "memory_register", 3793 "Entering add_active_range(%d, %#lx, %#lx) " 3794 "%d entries of %d used\n", 3795 nid, start_pfn, end_pfn, 3796 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3797 3798 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 3799 3800 /* Merge with existing active regions if possible */ 3801 for (i = 0; i < nr_nodemap_entries; i++) { 3802 if (early_node_map[i].nid != nid) 3803 continue; 3804 3805 /* Skip if an existing region covers this new one */ 3806 if (start_pfn >= early_node_map[i].start_pfn && 3807 end_pfn <= early_node_map[i].end_pfn) 3808 return; 3809 3810 /* Merge forward if suitable */ 3811 if (start_pfn <= early_node_map[i].end_pfn && 3812 end_pfn > early_node_map[i].end_pfn) { 3813 early_node_map[i].end_pfn = end_pfn; 3814 return; 3815 } 3816 3817 /* Merge backward if suitable */ 3818 if (start_pfn < early_node_map[i].end_pfn && 3819 end_pfn >= early_node_map[i].start_pfn) { 3820 early_node_map[i].start_pfn = start_pfn; 3821 return; 3822 } 3823 } 3824 3825 /* Check that early_node_map is large enough */ 3826 if (i >= MAX_ACTIVE_REGIONS) { 3827 printk(KERN_CRIT "More than %d memory regions, truncating\n", 3828 MAX_ACTIVE_REGIONS); 3829 return; 3830 } 3831 3832 early_node_map[i].nid = nid; 3833 early_node_map[i].start_pfn = start_pfn; 3834 early_node_map[i].end_pfn = end_pfn; 3835 nr_nodemap_entries = i + 1; 3836} 3837 3838/** 3839 * remove_active_range - Shrink an existing registered range of PFNs 3840 * @nid: The node id the range is on that should be shrunk 3841 * @start_pfn: The new PFN of the range 3842 * @end_pfn: The new PFN of the range 3843 * 3844 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 3845 * The map is kept near the end physical page range that has already been 3846 * registered. This function allows an arch to shrink an existing registered 3847 * range. 3848 */ 3849void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 3850 unsigned long end_pfn) 3851{ 3852 int i, j; 3853 int removed = 0; 3854 3855 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 3856 nid, start_pfn, end_pfn); 3857 3858 /* Find the old active region end and shrink */ 3859 for_each_active_range_index_in_nid(i, nid) { 3860 if (early_node_map[i].start_pfn >= start_pfn && 3861 early_node_map[i].end_pfn <= end_pfn) { 3862 /* clear it */ 3863 early_node_map[i].start_pfn = 0; 3864 early_node_map[i].end_pfn = 0; 3865 removed = 1; 3866 continue; 3867 } 3868 if (early_node_map[i].start_pfn < start_pfn && 3869 early_node_map[i].end_pfn > start_pfn) { 3870 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 3871 early_node_map[i].end_pfn = start_pfn; 3872 if (temp_end_pfn > end_pfn) 3873 add_active_range(nid, end_pfn, temp_end_pfn); 3874 continue; 3875 } 3876 if (early_node_map[i].start_pfn >= start_pfn && 3877 early_node_map[i].end_pfn > end_pfn && 3878 early_node_map[i].start_pfn < end_pfn) { 3879 early_node_map[i].start_pfn = end_pfn; 3880 continue; 3881 } 3882 } 3883 3884 if (!removed) 3885 return; 3886 3887 /* remove the blank ones */ 3888 for (i = nr_nodemap_entries - 1; i > 0; i--) { 3889 if (early_node_map[i].nid != nid) 3890 continue; 3891 if (early_node_map[i].end_pfn) 3892 continue; 3893 /* we found it, get rid of it */ 3894 for (j = i; j < nr_nodemap_entries - 1; j++) 3895 memcpy(&early_node_map[j], &early_node_map[j+1], 3896 sizeof(early_node_map[j])); 3897 j = nr_nodemap_entries - 1; 3898 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 3899 nr_nodemap_entries--; 3900 } 3901} 3902 3903/** 3904 * remove_all_active_ranges - Remove all currently registered regions 3905 * 3906 * During discovery, it may be found that a table like SRAT is invalid 3907 * and an alternative discovery method must be used. This function removes 3908 * all currently registered regions. 3909 */ 3910void __init remove_all_active_ranges(void) 3911{ 3912 memset(early_node_map, 0, sizeof(early_node_map)); 3913 nr_nodemap_entries = 0; 3914} 3915 3916/* Compare two active node_active_regions */ 3917static int __init cmp_node_active_region(const void *a, const void *b) 3918{ 3919 struct node_active_region *arange = (struct node_active_region *)a; 3920 struct node_active_region *brange = (struct node_active_region *)b; 3921 3922 /* Done this way to avoid overflows */ 3923 if (arange->start_pfn > brange->start_pfn) 3924 return 1; 3925 if (arange->start_pfn < brange->start_pfn) 3926 return -1; 3927 3928 return 0; 3929} 3930 3931/* sort the node_map by start_pfn */ 3932static void __init sort_node_map(void) 3933{ 3934 sort(early_node_map, (size_t)nr_nodemap_entries, 3935 sizeof(struct node_active_region), 3936 cmp_node_active_region, NULL); 3937} 3938 3939/* Find the lowest pfn for a node */ 3940static unsigned long __init find_min_pfn_for_node(int nid) 3941{ 3942 int i; 3943 unsigned long min_pfn = ULONG_MAX; 3944 3945 /* Assuming a sorted map, the first range found has the starting pfn */ 3946 for_each_active_range_index_in_nid(i, nid) 3947 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 3948 3949 if (min_pfn == ULONG_MAX) { 3950 printk(KERN_WARNING 3951 "Could not find start_pfn for node %d\n", nid); 3952 return 0; 3953 } 3954 3955 return min_pfn; 3956} 3957 3958/** 3959 * find_min_pfn_with_active_regions - Find the minimum PFN registered 3960 * 3961 * It returns the minimum PFN based on information provided via 3962 * add_active_range(). 3963 */ 3964unsigned long __init find_min_pfn_with_active_regions(void) 3965{ 3966 return find_min_pfn_for_node(MAX_NUMNODES); 3967} 3968 3969/* 3970 * early_calculate_totalpages() 3971 * Sum pages in active regions for movable zone. 3972 * Populate N_HIGH_MEMORY for calculating usable_nodes. 3973 */ 3974static unsigned long __init early_calculate_totalpages(void) 3975{ 3976 int i; 3977 unsigned long totalpages = 0; 3978 3979 for (i = 0; i < nr_nodemap_entries; i++) { 3980 unsigned long pages = early_node_map[i].end_pfn - 3981 early_node_map[i].start_pfn; 3982 totalpages += pages; 3983 if (pages) 3984 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 3985 } 3986 return totalpages; 3987} 3988 3989/* 3990 * Find the PFN the Movable zone begins in each node. Kernel memory 3991 * is spread evenly between nodes as long as the nodes have enough 3992 * memory. When they don't, some nodes will have more kernelcore than 3993 * others 3994 */ 3995static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 3996{ 3997 int i, nid; 3998 unsigned long usable_startpfn; 3999 unsigned long kernelcore_node, kernelcore_remaining; 4000 unsigned long totalpages = early_calculate_totalpages(); 4001 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4002 4003 /* 4004 * If movablecore was specified, calculate what size of 4005 * kernelcore that corresponds so that memory usable for 4006 * any allocation type is evenly spread. If both kernelcore 4007 * and movablecore are specified, then the value of kernelcore 4008 * will be used for required_kernelcore if it's greater than 4009 * what movablecore would have allowed. 4010 */ 4011 if (required_movablecore) { 4012 unsigned long corepages; 4013 4014 /* 4015 * Round-up so that ZONE_MOVABLE is at least as large as what 4016 * was requested by the user 4017 */ 4018 required_movablecore = 4019 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4020 corepages = totalpages - required_movablecore; 4021 4022 required_kernelcore = max(required_kernelcore, corepages); 4023 } 4024 4025 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4026 if (!required_kernelcore) 4027 return; 4028 4029 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4030 find_usable_zone_for_movable(); 4031 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4032 4033restart: 4034 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4035 kernelcore_node = required_kernelcore / usable_nodes; 4036 for_each_node_state(nid, N_HIGH_MEMORY) { 4037 /* 4038 * Recalculate kernelcore_node if the division per node 4039 * now exceeds what is necessary to satisfy the requested 4040 * amount of memory for the kernel 4041 */ 4042 if (required_kernelcore < kernelcore_node) 4043 kernelcore_node = required_kernelcore / usable_nodes; 4044 4045 /* 4046 * As the map is walked, we track how much memory is usable 4047 * by the kernel using kernelcore_remaining. When it is 4048 * 0, the rest of the node is usable by ZONE_MOVABLE 4049 */ 4050 kernelcore_remaining = kernelcore_node; 4051 4052 /* Go through each range of PFNs within this node */ 4053 for_each_active_range_index_in_nid(i, nid) { 4054 unsigned long start_pfn, end_pfn; 4055 unsigned long size_pages; 4056 4057 start_pfn = max(early_node_map[i].start_pfn, 4058 zone_movable_pfn[nid]); 4059 end_pfn = early_node_map[i].end_pfn; 4060 if (start_pfn >= end_pfn) 4061 continue; 4062 4063 /* Account for what is only usable for kernelcore */ 4064 if (start_pfn < usable_startpfn) { 4065 unsigned long kernel_pages; 4066 kernel_pages = min(end_pfn, usable_startpfn) 4067 - start_pfn; 4068 4069 kernelcore_remaining -= min(kernel_pages, 4070 kernelcore_remaining); 4071 required_kernelcore -= min(kernel_pages, 4072 required_kernelcore); 4073 4074 /* Continue if range is now fully accounted */ 4075 if (end_pfn <= usable_startpfn) { 4076 4077 /* 4078 * Push zone_movable_pfn to the end so 4079 * that if we have to rebalance 4080 * kernelcore across nodes, we will 4081 * not double account here 4082 */ 4083 zone_movable_pfn[nid] = end_pfn; 4084 continue; 4085 } 4086 start_pfn = usable_startpfn; 4087 } 4088 4089 /* 4090 * The usable PFN range for ZONE_MOVABLE is from 4091 * start_pfn->end_pfn. Calculate size_pages as the 4092 * number of pages used as kernelcore 4093 */ 4094 size_pages = end_pfn - start_pfn; 4095 if (size_pages > kernelcore_remaining) 4096 size_pages = kernelcore_remaining; 4097 zone_movable_pfn[nid] = start_pfn + size_pages; 4098 4099 /* 4100 * Some kernelcore has been met, update counts and 4101 * break if the kernelcore for this node has been 4102 * satisified 4103 */ 4104 required_kernelcore -= min(required_kernelcore, 4105 size_pages); 4106 kernelcore_remaining -= size_pages; 4107 if (!kernelcore_remaining) 4108 break; 4109 } 4110 } 4111 4112 /* 4113 * If there is still required_kernelcore, we do another pass with one 4114 * less node in the count. This will push zone_movable_pfn[nid] further 4115 * along on the nodes that still have memory until kernelcore is 4116 * satisified 4117 */ 4118 usable_nodes--; 4119 if (usable_nodes && required_kernelcore > usable_nodes) 4120 goto restart; 4121 4122 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4123 for (nid = 0; nid < MAX_NUMNODES; nid++) 4124 zone_movable_pfn[nid] = 4125 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4126} 4127 4128/* Any regular memory on that node ? */ 4129static void check_for_regular_memory(pg_data_t *pgdat) 4130{ 4131#ifdef CONFIG_HIGHMEM 4132 enum zone_type zone_type; 4133 4134 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4135 struct zone *zone = &pgdat->node_zones[zone_type]; 4136 if (zone->present_pages) 4137 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4138 } 4139#endif 4140} 4141 4142/** 4143 * free_area_init_nodes - Initialise all pg_data_t and zone data 4144 * @max_zone_pfn: an array of max PFNs for each zone 4145 * 4146 * This will call free_area_init_node() for each active node in the system. 4147 * Using the page ranges provided by add_active_range(), the size of each 4148 * zone in each node and their holes is calculated. If the maximum PFN 4149 * between two adjacent zones match, it is assumed that the zone is empty. 4150 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4151 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4152 * starts where the previous one ended. For example, ZONE_DMA32 starts 4153 * at arch_max_dma_pfn. 4154 */ 4155void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4156{ 4157 unsigned long nid; 4158 int i; 4159 4160 /* Sort early_node_map as initialisation assumes it is sorted */ 4161 sort_node_map(); 4162 4163 /* Record where the zone boundaries are */ 4164 memset(arch_zone_lowest_possible_pfn, 0, 4165 sizeof(arch_zone_lowest_possible_pfn)); 4166 memset(arch_zone_highest_possible_pfn, 0, 4167 sizeof(arch_zone_highest_possible_pfn)); 4168 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4169 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4170 for (i = 1; i < MAX_NR_ZONES; i++) { 4171 if (i == ZONE_MOVABLE) 4172 continue; 4173 arch_zone_lowest_possible_pfn[i] = 4174 arch_zone_highest_possible_pfn[i-1]; 4175 arch_zone_highest_possible_pfn[i] = 4176 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4177 } 4178 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4179 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4180 4181 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4182 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4183 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4184 4185 /* Print out the zone ranges */ 4186 printk("Zone PFN ranges:\n"); 4187 for (i = 0; i < MAX_NR_ZONES; i++) { 4188 if (i == ZONE_MOVABLE) 4189 continue; 4190 printk(" %-8s %0#10lx -> %0#10lx\n", 4191 zone_names[i], 4192 arch_zone_lowest_possible_pfn[i], 4193 arch_zone_highest_possible_pfn[i]); 4194 } 4195 4196 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4197 printk("Movable zone start PFN for each node\n"); 4198 for (i = 0; i < MAX_NUMNODES; i++) { 4199 if (zone_movable_pfn[i]) 4200 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4201 } 4202 4203 /* Print out the early_node_map[] */ 4204 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4205 for (i = 0; i < nr_nodemap_entries; i++) 4206 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4207 early_node_map[i].start_pfn, 4208 early_node_map[i].end_pfn); 4209 4210 /* Initialise every node */ 4211 mminit_verify_pageflags_layout(); 4212 setup_nr_node_ids(); 4213 for_each_online_node(nid) { 4214 pg_data_t *pgdat = NODE_DATA(nid); 4215 free_area_init_node(nid, NULL, 4216 find_min_pfn_for_node(nid), NULL); 4217 4218 /* Any memory on that node */ 4219 if (pgdat->node_present_pages) 4220 node_set_state(nid, N_HIGH_MEMORY); 4221 check_for_regular_memory(pgdat); 4222 } 4223} 4224 4225static int __init cmdline_parse_core(char *p, unsigned long *core) 4226{ 4227 unsigned long long coremem; 4228 if (!p) 4229 return -EINVAL; 4230 4231 coremem = memparse(p, &p); 4232 *core = coremem >> PAGE_SHIFT; 4233 4234 /* Paranoid check that UL is enough for the coremem value */ 4235 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4236 4237 return 0; 4238} 4239 4240/* 4241 * kernelcore=size sets the amount of memory for use for allocations that 4242 * cannot be reclaimed or migrated. 4243 */ 4244static int __init cmdline_parse_kernelcore(char *p) 4245{ 4246 return cmdline_parse_core(p, &required_kernelcore); 4247} 4248 4249/* 4250 * movablecore=size sets the amount of memory for use for allocations that 4251 * can be reclaimed or migrated. 4252 */ 4253static int __init cmdline_parse_movablecore(char *p) 4254{ 4255 return cmdline_parse_core(p, &required_movablecore); 4256} 4257 4258early_param("kernelcore", cmdline_parse_kernelcore); 4259early_param("movablecore", cmdline_parse_movablecore); 4260 4261#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4262 4263/** 4264 * set_dma_reserve - set the specified number of pages reserved in the first zone 4265 * @new_dma_reserve: The number of pages to mark reserved 4266 * 4267 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4268 * In the DMA zone, a significant percentage may be consumed by kernel image 4269 * and other unfreeable allocations which can skew the watermarks badly. This 4270 * function may optionally be used to account for unfreeable pages in the 4271 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4272 * smaller per-cpu batchsize. 4273 */ 4274void __init set_dma_reserve(unsigned long new_dma_reserve) 4275{ 4276 dma_reserve = new_dma_reserve; 4277} 4278 4279#ifndef CONFIG_NEED_MULTIPLE_NODES 4280struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; 4281EXPORT_SYMBOL(contig_page_data); 4282#endif 4283 4284void __init free_area_init(unsigned long *zones_size) 4285{ 4286 free_area_init_node(0, zones_size, 4287 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4288} 4289 4290static int page_alloc_cpu_notify(struct notifier_block *self, 4291 unsigned long action, void *hcpu) 4292{ 4293 int cpu = (unsigned long)hcpu; 4294 4295 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4296 drain_pages(cpu); 4297 4298 /* 4299 * Spill the event counters of the dead processor 4300 * into the current processors event counters. 4301 * This artificially elevates the count of the current 4302 * processor. 4303 */ 4304 vm_events_fold_cpu(cpu); 4305 4306 /* 4307 * Zero the differential counters of the dead processor 4308 * so that the vm statistics are consistent. 4309 * 4310 * This is only okay since the processor is dead and cannot 4311 * race with what we are doing. 4312 */ 4313 refresh_cpu_vm_stats(cpu); 4314 } 4315 return NOTIFY_OK; 4316} 4317 4318void __init page_alloc_init(void) 4319{ 4320 hotcpu_notifier(page_alloc_cpu_notify, 0); 4321} 4322 4323/* 4324 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4325 * or min_free_kbytes changes. 4326 */ 4327static void calculate_totalreserve_pages(void) 4328{ 4329 struct pglist_data *pgdat; 4330 unsigned long reserve_pages = 0; 4331 enum zone_type i, j; 4332 4333 for_each_online_pgdat(pgdat) { 4334 for (i = 0; i < MAX_NR_ZONES; i++) { 4335 struct zone *zone = pgdat->node_zones + i; 4336 unsigned long max = 0; 4337 4338 /* Find valid and maximum lowmem_reserve in the zone */ 4339 for (j = i; j < MAX_NR_ZONES; j++) { 4340 if (zone->lowmem_reserve[j] > max) 4341 max = zone->lowmem_reserve[j]; 4342 } 4343 4344 /* we treat the high watermark as reserved pages. */ 4345 max += high_wmark_pages(zone); 4346 4347 if (max > zone->present_pages) 4348 max = zone->present_pages; 4349 reserve_pages += max; 4350 } 4351 } 4352 totalreserve_pages = reserve_pages; 4353} 4354 4355/* 4356 * setup_per_zone_lowmem_reserve - called whenever 4357 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4358 * has a correct pages reserved value, so an adequate number of 4359 * pages are left in the zone after a successful __alloc_pages(). 4360 */ 4361static void setup_per_zone_lowmem_reserve(void) 4362{ 4363 struct pglist_data *pgdat; 4364 enum zone_type j, idx; 4365 4366 for_each_online_pgdat(pgdat) { 4367 for (j = 0; j < MAX_NR_ZONES; j++) { 4368 struct zone *zone = pgdat->node_zones + j; 4369 unsigned long present_pages = zone->present_pages; 4370 4371 zone->lowmem_reserve[j] = 0; 4372 4373 idx = j; 4374 while (idx) { 4375 struct zone *lower_zone; 4376 4377 idx--; 4378 4379 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4380 sysctl_lowmem_reserve_ratio[idx] = 1; 4381 4382 lower_zone = pgdat->node_zones + idx; 4383 lower_zone->lowmem_reserve[j] = present_pages / 4384 sysctl_lowmem_reserve_ratio[idx]; 4385 present_pages += lower_zone->present_pages; 4386 } 4387 } 4388 } 4389 4390 /* update totalreserve_pages */ 4391 calculate_totalreserve_pages(); 4392} 4393 4394/** 4395 * setup_per_zone_pages_min - called when min_free_kbytes changes. 4396 * 4397 * Ensures that the pages_{min,low,high} values for each zone are set correctly 4398 * with respect to min_free_kbytes. 4399 */ 4400void setup_per_zone_pages_min(void) 4401{ 4402 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4403 unsigned long lowmem_pages = 0; 4404 struct zone *zone; 4405 unsigned long flags; 4406 4407 /* Calculate total number of !ZONE_HIGHMEM pages */ 4408 for_each_zone(zone) { 4409 if (!is_highmem(zone)) 4410 lowmem_pages += zone->present_pages; 4411 } 4412 4413 for_each_zone(zone) { 4414 u64 tmp; 4415 4416 spin_lock_irqsave(&zone->lock, flags); 4417 tmp = (u64)pages_min * zone->present_pages; 4418 do_div(tmp, lowmem_pages); 4419 if (is_highmem(zone)) { 4420 /* 4421 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4422 * need highmem pages, so cap pages_min to a small 4423 * value here. 4424 * 4425 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4426 * deltas controls asynch page reclaim, and so should 4427 * not be capped for highmem. 4428 */ 4429 int min_pages; 4430 4431 min_pages = zone->present_pages / 1024; 4432 if (min_pages < SWAP_CLUSTER_MAX) 4433 min_pages = SWAP_CLUSTER_MAX; 4434 if (min_pages > 128) 4435 min_pages = 128; 4436 zone->watermark[WMARK_MIN] = min_pages; 4437 } else { 4438 /* 4439 * If it's a lowmem zone, reserve a number of pages 4440 * proportionate to the zone's size. 4441 */ 4442 zone->watermark[WMARK_MIN] = tmp; 4443 } 4444 4445 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4446 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4447 setup_zone_migrate_reserve(zone); 4448 spin_unlock_irqrestore(&zone->lock, flags); 4449 } 4450 4451 /* update totalreserve_pages */ 4452 calculate_totalreserve_pages(); 4453} 4454 4455/** 4456 * The inactive anon list should be small enough that the VM never has to 4457 * do too much work, but large enough that each inactive page has a chance 4458 * to be referenced again before it is swapped out. 4459 * 4460 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4461 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4462 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4463 * the anonymous pages are kept on the inactive list. 4464 * 4465 * total target max 4466 * memory ratio inactive anon 4467 * ------------------------------------- 4468 * 10MB 1 5MB 4469 * 100MB 1 50MB 4470 * 1GB 3 250MB 4471 * 10GB 10 0.9GB 4472 * 100GB 31 3GB 4473 * 1TB 101 10GB 4474 * 10TB 320 32GB 4475 */ 4476static void __init setup_per_zone_inactive_ratio(void) 4477{ 4478 struct zone *zone; 4479 4480 for_each_zone(zone) { 4481 unsigned int gb, ratio; 4482 4483 /* Zone size in gigabytes */ 4484 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4485 if (gb) 4486 ratio = int_sqrt(10 * gb); 4487 else 4488 ratio = 1; 4489 4490 zone->inactive_ratio = ratio; 4491 } 4492} 4493 4494/* 4495 * Initialise min_free_kbytes. 4496 * 4497 * For small machines we want it small (128k min). For large machines 4498 * we want it large (64MB max). But it is not linear, because network 4499 * bandwidth does not increase linearly with machine size. We use 4500 * 4501 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4502 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4503 * 4504 * which yields 4505 * 4506 * 16MB: 512k 4507 * 32MB: 724k 4508 * 64MB: 1024k 4509 * 128MB: 1448k 4510 * 256MB: 2048k 4511 * 512MB: 2896k 4512 * 1024MB: 4096k 4513 * 2048MB: 5792k 4514 * 4096MB: 8192k 4515 * 8192MB: 11584k 4516 * 16384MB: 16384k 4517 */ 4518static int __init init_per_zone_pages_min(void) 4519{ 4520 unsigned long lowmem_kbytes; 4521 4522 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4523 4524 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4525 if (min_free_kbytes < 128) 4526 min_free_kbytes = 128; 4527 if (min_free_kbytes > 65536) 4528 min_free_kbytes = 65536; 4529 setup_per_zone_pages_min(); 4530 setup_per_zone_lowmem_reserve(); 4531 setup_per_zone_inactive_ratio(); 4532 return 0; 4533} 4534module_init(init_per_zone_pages_min) 4535 4536/* 4537 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4538 * that we can call two helper functions whenever min_free_kbytes 4539 * changes. 4540 */ 4541int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4542 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4543{ 4544 proc_dointvec(table, write, file, buffer, length, ppos); 4545 if (write) 4546 setup_per_zone_pages_min(); 4547 return 0; 4548} 4549 4550#ifdef CONFIG_NUMA 4551int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4552 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4553{ 4554 struct zone *zone; 4555 int rc; 4556 4557 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4558 if (rc) 4559 return rc; 4560 4561 for_each_zone(zone) 4562 zone->min_unmapped_pages = (zone->present_pages * 4563 sysctl_min_unmapped_ratio) / 100; 4564 return 0; 4565} 4566 4567int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4568 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4569{ 4570 struct zone *zone; 4571 int rc; 4572 4573 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4574 if (rc) 4575 return rc; 4576 4577 for_each_zone(zone) 4578 zone->min_slab_pages = (zone->present_pages * 4579 sysctl_min_slab_ratio) / 100; 4580 return 0; 4581} 4582#endif 4583 4584/* 4585 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4586 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4587 * whenever sysctl_lowmem_reserve_ratio changes. 4588 * 4589 * The reserve ratio obviously has absolutely no relation with the 4590 * minimum watermarks. The lowmem reserve ratio can only make sense 4591 * if in function of the boot time zone sizes. 4592 */ 4593int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4594 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4595{ 4596 proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4597 setup_per_zone_lowmem_reserve(); 4598 return 0; 4599} 4600 4601/* 4602 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4603 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4604 * can have before it gets flushed back to buddy allocator. 4605 */ 4606 4607int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4608 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4609{ 4610 struct zone *zone; 4611 unsigned int cpu; 4612 int ret; 4613 4614 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4615 if (!write || (ret == -EINVAL)) 4616 return ret; 4617 for_each_zone(zone) { 4618 for_each_online_cpu(cpu) { 4619 unsigned long high; 4620 high = zone->present_pages / percpu_pagelist_fraction; 4621 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4622 } 4623 } 4624 return 0; 4625} 4626 4627int hashdist = HASHDIST_DEFAULT; 4628 4629#ifdef CONFIG_NUMA 4630static int __init set_hashdist(char *str) 4631{ 4632 if (!str) 4633 return 0; 4634 hashdist = simple_strtoul(str, &str, 0); 4635 return 1; 4636} 4637__setup("hashdist=", set_hashdist); 4638#endif 4639 4640/* 4641 * allocate a large system hash table from bootmem 4642 * - it is assumed that the hash table must contain an exact power-of-2 4643 * quantity of entries 4644 * - limit is the number of hash buckets, not the total allocation size 4645 */ 4646void *__init alloc_large_system_hash(const char *tablename, 4647 unsigned long bucketsize, 4648 unsigned long numentries, 4649 int scale, 4650 int flags, 4651 unsigned int *_hash_shift, 4652 unsigned int *_hash_mask, 4653 unsigned long limit) 4654{ 4655 unsigned long long max = limit; 4656 unsigned long log2qty, size; 4657 void *table = NULL; 4658 4659 /* allow the kernel cmdline to have a say */ 4660 if (!numentries) { 4661 /* round applicable memory size up to nearest megabyte */ 4662 numentries = nr_kernel_pages; 4663 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4664 numentries >>= 20 - PAGE_SHIFT; 4665 numentries <<= 20 - PAGE_SHIFT; 4666 4667 /* limit to 1 bucket per 2^scale bytes of low memory */ 4668 if (scale > PAGE_SHIFT) 4669 numentries >>= (scale - PAGE_SHIFT); 4670 else 4671 numentries <<= (PAGE_SHIFT - scale); 4672 4673 /* Make sure we've got at least a 0-order allocation.. */ 4674 if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4675 numentries = PAGE_SIZE / bucketsize; 4676 } 4677 numentries = roundup_pow_of_two(numentries); 4678 4679 /* limit allocation size to 1/16 total memory by default */ 4680 if (max == 0) { 4681 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4682 do_div(max, bucketsize); 4683 } 4684 4685 if (numentries > max) 4686 numentries = max; 4687 4688 log2qty = ilog2(numentries); 4689 4690 do { 4691 size = bucketsize << log2qty; 4692 if (flags & HASH_EARLY) 4693 table = alloc_bootmem_nopanic(size); 4694 else if (hashdist) 4695 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4696 else { 4697 /* 4698 * If bucketsize is not a power-of-two, we may free 4699 * some pages at the end of hash table which 4700 * alloc_pages_exact() automatically does 4701 */ 4702 if (get_order(size) < MAX_ORDER) 4703 table = alloc_pages_exact(size, GFP_ATOMIC); 4704 } 4705 } while (!table && size > PAGE_SIZE && --log2qty); 4706 4707 if (!table) 4708 panic("Failed to allocate %s hash table\n", tablename); 4709 4710 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4711 tablename, 4712 (1U << log2qty), 4713 ilog2(size) - PAGE_SHIFT, 4714 size); 4715 4716 if (_hash_shift) 4717 *_hash_shift = log2qty; 4718 if (_hash_mask) 4719 *_hash_mask = (1 << log2qty) - 1; 4720 4721 /* 4722 * If hashdist is set, the table allocation is done with __vmalloc() 4723 * which invokes the kmemleak_alloc() callback. This function may also 4724 * be called before the slab and kmemleak are initialised when 4725 * kmemleak simply buffers the request to be executed later 4726 * (GFP_ATOMIC flag ignored in this case). 4727 */ 4728 if (!hashdist) 4729 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4730 4731 return table; 4732} 4733 4734/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4735static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4736 unsigned long pfn) 4737{ 4738#ifdef CONFIG_SPARSEMEM 4739 return __pfn_to_section(pfn)->pageblock_flags; 4740#else 4741 return zone->pageblock_flags; 4742#endif /* CONFIG_SPARSEMEM */ 4743} 4744 4745static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4746{ 4747#ifdef CONFIG_SPARSEMEM 4748 pfn &= (PAGES_PER_SECTION-1); 4749 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4750#else 4751 pfn = pfn - zone->zone_start_pfn; 4752 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4753#endif /* CONFIG_SPARSEMEM */ 4754} 4755 4756/** 4757 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4758 * @page: The page within the block of interest 4759 * @start_bitidx: The first bit of interest to retrieve 4760 * @end_bitidx: The last bit of interest 4761 * returns pageblock_bits flags 4762 */ 4763unsigned long get_pageblock_flags_group(struct page *page, 4764 int start_bitidx, int end_bitidx) 4765{ 4766 struct zone *zone; 4767 unsigned long *bitmap; 4768 unsigned long pfn, bitidx; 4769 unsigned long flags = 0; 4770 unsigned long value = 1; 4771 4772 zone = page_zone(page); 4773 pfn = page_to_pfn(page); 4774 bitmap = get_pageblock_bitmap(zone, pfn); 4775 bitidx = pfn_to_bitidx(zone, pfn); 4776 4777 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4778 if (test_bit(bitidx + start_bitidx, bitmap)) 4779 flags |= value; 4780 4781 return flags; 4782} 4783 4784/** 4785 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 4786 * @page: The page within the block of interest 4787 * @start_bitidx: The first bit of interest 4788 * @end_bitidx: The last bit of interest 4789 * @flags: The flags to set 4790 */ 4791void set_pageblock_flags_group(struct page *page, unsigned long flags, 4792 int start_bitidx, int end_bitidx) 4793{ 4794 struct zone *zone; 4795 unsigned long *bitmap; 4796 unsigned long pfn, bitidx; 4797 unsigned long value = 1; 4798 4799 zone = page_zone(page); 4800 pfn = page_to_pfn(page); 4801 bitmap = get_pageblock_bitmap(zone, pfn); 4802 bitidx = pfn_to_bitidx(zone, pfn); 4803 VM_BUG_ON(pfn < zone->zone_start_pfn); 4804 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 4805 4806 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4807 if (flags & value) 4808 __set_bit(bitidx + start_bitidx, bitmap); 4809 else 4810 __clear_bit(bitidx + start_bitidx, bitmap); 4811} 4812 4813/* 4814 * This is designed as sub function...plz see page_isolation.c also. 4815 * set/clear page block's type to be ISOLATE. 4816 * page allocater never alloc memory from ISOLATE block. 4817 */ 4818 4819int set_migratetype_isolate(struct page *page) 4820{ 4821 struct zone *zone; 4822 unsigned long flags; 4823 int ret = -EBUSY; 4824 4825 zone = page_zone(page); 4826 spin_lock_irqsave(&zone->lock, flags); 4827 /* 4828 * In future, more migrate types will be able to be isolation target. 4829 */ 4830 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE) 4831 goto out; 4832 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 4833 move_freepages_block(zone, page, MIGRATE_ISOLATE); 4834 ret = 0; 4835out: 4836 spin_unlock_irqrestore(&zone->lock, flags); 4837 if (!ret) 4838 drain_all_pages(); 4839 return ret; 4840} 4841 4842void unset_migratetype_isolate(struct page *page) 4843{ 4844 struct zone *zone; 4845 unsigned long flags; 4846 zone = page_zone(page); 4847 spin_lock_irqsave(&zone->lock, flags); 4848 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 4849 goto out; 4850 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4851 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4852out: 4853 spin_unlock_irqrestore(&zone->lock, flags); 4854} 4855 4856#ifdef CONFIG_MEMORY_HOTREMOVE 4857/* 4858 * All pages in the range must be isolated before calling this. 4859 */ 4860void 4861__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 4862{ 4863 struct page *page; 4864 struct zone *zone; 4865 int order, i; 4866 unsigned long pfn; 4867 unsigned long flags; 4868 /* find the first valid pfn */ 4869 for (pfn = start_pfn; pfn < end_pfn; pfn++) 4870 if (pfn_valid(pfn)) 4871 break; 4872 if (pfn == end_pfn) 4873 return; 4874 zone = page_zone(pfn_to_page(pfn)); 4875 spin_lock_irqsave(&zone->lock, flags); 4876 pfn = start_pfn; 4877 while (pfn < end_pfn) { 4878 if (!pfn_valid(pfn)) { 4879 pfn++; 4880 continue; 4881 } 4882 page = pfn_to_page(pfn); 4883 BUG_ON(page_count(page)); 4884 BUG_ON(!PageBuddy(page)); 4885 order = page_order(page); 4886#ifdef CONFIG_DEBUG_VM 4887 printk(KERN_INFO "remove from free list %lx %d %lx\n", 4888 pfn, 1 << order, end_pfn); 4889#endif 4890 list_del(&page->lru); 4891 rmv_page_order(page); 4892 zone->free_area[order].nr_free--; 4893 __mod_zone_page_state(zone, NR_FREE_PAGES, 4894 - (1UL << order)); 4895 for (i = 0; i < (1 << order); i++) 4896 SetPageReserved((page+i)); 4897 pfn += (1 << order); 4898 } 4899 spin_unlock_irqrestore(&zone->lock, flags); 4900} 4901#endif 4902