page_alloc.c revision 4b4f278c030aa4b6ee0915f396e9a9478d92d610
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/compiler.h> 25#include <linux/kernel.h> 26#include <linux/kmemcheck.h> 27#include <linux/module.h> 28#include <linux/suspend.h> 29#include <linux/pagevec.h> 30#include <linux/blkdev.h> 31#include <linux/slab.h> 32#include <linux/oom.h> 33#include <linux/notifier.h> 34#include <linux/topology.h> 35#include <linux/sysctl.h> 36#include <linux/cpu.h> 37#include <linux/cpuset.h> 38#include <linux/memory_hotplug.h> 39#include <linux/nodemask.h> 40#include <linux/vmalloc.h> 41#include <linux/mempolicy.h> 42#include <linux/stop_machine.h> 43#include <linux/sort.h> 44#include <linux/pfn.h> 45#include <linux/backing-dev.h> 46#include <linux/fault-inject.h> 47#include <linux/page-isolation.h> 48#include <linux/page_cgroup.h> 49#include <linux/debugobjects.h> 50#include <linux/kmemleak.h> 51 52#include <asm/tlbflush.h> 53#include <asm/div64.h> 54#include "internal.h" 55 56/* 57 * Array of node states. 58 */ 59nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 60 [N_POSSIBLE] = NODE_MASK_ALL, 61 [N_ONLINE] = { { [0] = 1UL } }, 62#ifndef CONFIG_NUMA 63 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 64#ifdef CONFIG_HIGHMEM 65 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 66#endif 67 [N_CPU] = { { [0] = 1UL } }, 68#endif /* NUMA */ 69}; 70EXPORT_SYMBOL(node_states); 71 72unsigned long totalram_pages __read_mostly; 73unsigned long totalreserve_pages __read_mostly; 74unsigned long highest_memmap_pfn __read_mostly; 75int percpu_pagelist_fraction; 76gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 77 78#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 79int pageblock_order __read_mostly; 80#endif 81 82static void __free_pages_ok(struct page *page, unsigned int order); 83 84/* 85 * results with 256, 32 in the lowmem_reserve sysctl: 86 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 87 * 1G machine -> (16M dma, 784M normal, 224M high) 88 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 89 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 90 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 91 * 92 * TBD: should special case ZONE_DMA32 machines here - in those we normally 93 * don't need any ZONE_NORMAL reservation 94 */ 95int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 96#ifdef CONFIG_ZONE_DMA 97 256, 98#endif 99#ifdef CONFIG_ZONE_DMA32 100 256, 101#endif 102#ifdef CONFIG_HIGHMEM 103 32, 104#endif 105 32, 106}; 107 108EXPORT_SYMBOL(totalram_pages); 109 110static char * const zone_names[MAX_NR_ZONES] = { 111#ifdef CONFIG_ZONE_DMA 112 "DMA", 113#endif 114#ifdef CONFIG_ZONE_DMA32 115 "DMA32", 116#endif 117 "Normal", 118#ifdef CONFIG_HIGHMEM 119 "HighMem", 120#endif 121 "Movable", 122}; 123 124int min_free_kbytes = 1024; 125 126unsigned long __meminitdata nr_kernel_pages; 127unsigned long __meminitdata nr_all_pages; 128static unsigned long __meminitdata dma_reserve; 129 130#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 131 /* 132 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 133 * ranges of memory (RAM) that may be registered with add_active_range(). 134 * Ranges passed to add_active_range() will be merged if possible 135 * so the number of times add_active_range() can be called is 136 * related to the number of nodes and the number of holes 137 */ 138 #ifdef CONFIG_MAX_ACTIVE_REGIONS 139 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 140 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 141 #else 142 #if MAX_NUMNODES >= 32 143 /* If there can be many nodes, allow up to 50 holes per node */ 144 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 145 #else 146 /* By default, allow up to 256 distinct regions */ 147 #define MAX_ACTIVE_REGIONS 256 148 #endif 149 #endif 150 151 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 152 static int __meminitdata nr_nodemap_entries; 153 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 154 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 155 static unsigned long __initdata required_kernelcore; 156 static unsigned long __initdata required_movablecore; 157 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 158 159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 160 int movable_zone; 161 EXPORT_SYMBOL(movable_zone); 162#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 163 164#if MAX_NUMNODES > 1 165int nr_node_ids __read_mostly = MAX_NUMNODES; 166int nr_online_nodes __read_mostly = 1; 167EXPORT_SYMBOL(nr_node_ids); 168EXPORT_SYMBOL(nr_online_nodes); 169#endif 170 171int page_group_by_mobility_disabled __read_mostly; 172 173static void set_pageblock_migratetype(struct page *page, int migratetype) 174{ 175 176 if (unlikely(page_group_by_mobility_disabled)) 177 migratetype = MIGRATE_UNMOVABLE; 178 179 set_pageblock_flags_group(page, (unsigned long)migratetype, 180 PB_migrate, PB_migrate_end); 181} 182 183bool oom_killer_disabled __read_mostly; 184 185#ifdef CONFIG_DEBUG_VM 186static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 187{ 188 int ret = 0; 189 unsigned seq; 190 unsigned long pfn = page_to_pfn(page); 191 192 do { 193 seq = zone_span_seqbegin(zone); 194 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 195 ret = 1; 196 else if (pfn < zone->zone_start_pfn) 197 ret = 1; 198 } while (zone_span_seqretry(zone, seq)); 199 200 return ret; 201} 202 203static int page_is_consistent(struct zone *zone, struct page *page) 204{ 205 if (!pfn_valid_within(page_to_pfn(page))) 206 return 0; 207 if (zone != page_zone(page)) 208 return 0; 209 210 return 1; 211} 212/* 213 * Temporary debugging check for pages not lying within a given zone. 214 */ 215static int bad_range(struct zone *zone, struct page *page) 216{ 217 if (page_outside_zone_boundaries(zone, page)) 218 return 1; 219 if (!page_is_consistent(zone, page)) 220 return 1; 221 222 return 0; 223} 224#else 225static inline int bad_range(struct zone *zone, struct page *page) 226{ 227 return 0; 228} 229#endif 230 231static void bad_page(struct page *page) 232{ 233 static unsigned long resume; 234 static unsigned long nr_shown; 235 static unsigned long nr_unshown; 236 237 /* 238 * Allow a burst of 60 reports, then keep quiet for that minute; 239 * or allow a steady drip of one report per second. 240 */ 241 if (nr_shown == 60) { 242 if (time_before(jiffies, resume)) { 243 nr_unshown++; 244 goto out; 245 } 246 if (nr_unshown) { 247 printk(KERN_ALERT 248 "BUG: Bad page state: %lu messages suppressed\n", 249 nr_unshown); 250 nr_unshown = 0; 251 } 252 nr_shown = 0; 253 } 254 if (nr_shown++ == 0) 255 resume = jiffies + 60 * HZ; 256 257 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 258 current->comm, page_to_pfn(page)); 259 printk(KERN_ALERT 260 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 261 page, (void *)page->flags, page_count(page), 262 page_mapcount(page), page->mapping, page->index); 263 264 dump_stack(); 265out: 266 /* Leave bad fields for debug, except PageBuddy could make trouble */ 267 __ClearPageBuddy(page); 268 add_taint(TAINT_BAD_PAGE); 269} 270 271/* 272 * Higher-order pages are called "compound pages". They are structured thusly: 273 * 274 * The first PAGE_SIZE page is called the "head page". 275 * 276 * The remaining PAGE_SIZE pages are called "tail pages". 277 * 278 * All pages have PG_compound set. All pages have their ->private pointing at 279 * the head page (even the head page has this). 280 * 281 * The first tail page's ->lru.next holds the address of the compound page's 282 * put_page() function. Its ->lru.prev holds the order of allocation. 283 * This usage means that zero-order pages may not be compound. 284 */ 285 286static void free_compound_page(struct page *page) 287{ 288 __free_pages_ok(page, compound_order(page)); 289} 290 291void prep_compound_page(struct page *page, unsigned long order) 292{ 293 int i; 294 int nr_pages = 1 << order; 295 296 set_compound_page_dtor(page, free_compound_page); 297 set_compound_order(page, order); 298 __SetPageHead(page); 299 for (i = 1; i < nr_pages; i++) { 300 struct page *p = page + i; 301 302 __SetPageTail(p); 303 p->first_page = page; 304 } 305} 306 307static int destroy_compound_page(struct page *page, unsigned long order) 308{ 309 int i; 310 int nr_pages = 1 << order; 311 int bad = 0; 312 313 if (unlikely(compound_order(page) != order) || 314 unlikely(!PageHead(page))) { 315 bad_page(page); 316 bad++; 317 } 318 319 __ClearPageHead(page); 320 321 for (i = 1; i < nr_pages; i++) { 322 struct page *p = page + i; 323 324 if (unlikely(!PageTail(p) || (p->first_page != page))) { 325 bad_page(page); 326 bad++; 327 } 328 __ClearPageTail(p); 329 } 330 331 return bad; 332} 333 334static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 335{ 336 int i; 337 338 /* 339 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 340 * and __GFP_HIGHMEM from hard or soft interrupt context. 341 */ 342 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 343 for (i = 0; i < (1 << order); i++) 344 clear_highpage(page + i); 345} 346 347static inline void set_page_order(struct page *page, int order) 348{ 349 set_page_private(page, order); 350 __SetPageBuddy(page); 351} 352 353static inline void rmv_page_order(struct page *page) 354{ 355 __ClearPageBuddy(page); 356 set_page_private(page, 0); 357} 358 359/* 360 * Locate the struct page for both the matching buddy in our 361 * pair (buddy1) and the combined O(n+1) page they form (page). 362 * 363 * 1) Any buddy B1 will have an order O twin B2 which satisfies 364 * the following equation: 365 * B2 = B1 ^ (1 << O) 366 * For example, if the starting buddy (buddy2) is #8 its order 367 * 1 buddy is #10: 368 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 369 * 370 * 2) Any buddy B will have an order O+1 parent P which 371 * satisfies the following equation: 372 * P = B & ~(1 << O) 373 * 374 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 375 */ 376static inline struct page * 377__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 378{ 379 unsigned long buddy_idx = page_idx ^ (1 << order); 380 381 return page + (buddy_idx - page_idx); 382} 383 384static inline unsigned long 385__find_combined_index(unsigned long page_idx, unsigned int order) 386{ 387 return (page_idx & ~(1 << order)); 388} 389 390/* 391 * This function checks whether a page is free && is the buddy 392 * we can do coalesce a page and its buddy if 393 * (a) the buddy is not in a hole && 394 * (b) the buddy is in the buddy system && 395 * (c) a page and its buddy have the same order && 396 * (d) a page and its buddy are in the same zone. 397 * 398 * For recording whether a page is in the buddy system, we use PG_buddy. 399 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 400 * 401 * For recording page's order, we use page_private(page). 402 */ 403static inline int page_is_buddy(struct page *page, struct page *buddy, 404 int order) 405{ 406 if (!pfn_valid_within(page_to_pfn(buddy))) 407 return 0; 408 409 if (page_zone_id(page) != page_zone_id(buddy)) 410 return 0; 411 412 if (PageBuddy(buddy) && page_order(buddy) == order) { 413 VM_BUG_ON(page_count(buddy) != 0); 414 return 1; 415 } 416 return 0; 417} 418 419/* 420 * Freeing function for a buddy system allocator. 421 * 422 * The concept of a buddy system is to maintain direct-mapped table 423 * (containing bit values) for memory blocks of various "orders". 424 * The bottom level table contains the map for the smallest allocatable 425 * units of memory (here, pages), and each level above it describes 426 * pairs of units from the levels below, hence, "buddies". 427 * At a high level, all that happens here is marking the table entry 428 * at the bottom level available, and propagating the changes upward 429 * as necessary, plus some accounting needed to play nicely with other 430 * parts of the VM system. 431 * At each level, we keep a list of pages, which are heads of continuous 432 * free pages of length of (1 << order) and marked with PG_buddy. Page's 433 * order is recorded in page_private(page) field. 434 * So when we are allocating or freeing one, we can derive the state of the 435 * other. That is, if we allocate a small block, and both were 436 * free, the remainder of the region must be split into blocks. 437 * If a block is freed, and its buddy is also free, then this 438 * triggers coalescing into a block of larger size. 439 * 440 * -- wli 441 */ 442 443static inline void __free_one_page(struct page *page, 444 struct zone *zone, unsigned int order, 445 int migratetype) 446{ 447 unsigned long page_idx; 448 449 if (unlikely(PageCompound(page))) 450 if (unlikely(destroy_compound_page(page, order))) 451 return; 452 453 VM_BUG_ON(migratetype == -1); 454 455 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 456 457 VM_BUG_ON(page_idx & ((1 << order) - 1)); 458 VM_BUG_ON(bad_range(zone, page)); 459 460 while (order < MAX_ORDER-1) { 461 unsigned long combined_idx; 462 struct page *buddy; 463 464 buddy = __page_find_buddy(page, page_idx, order); 465 if (!page_is_buddy(page, buddy, order)) 466 break; 467 468 /* Our buddy is free, merge with it and move up one order. */ 469 list_del(&buddy->lru); 470 zone->free_area[order].nr_free--; 471 rmv_page_order(buddy); 472 combined_idx = __find_combined_index(page_idx, order); 473 page = page + (combined_idx - page_idx); 474 page_idx = combined_idx; 475 order++; 476 } 477 set_page_order(page, order); 478 list_add(&page->lru, 479 &zone->free_area[order].free_list[migratetype]); 480 zone->free_area[order].nr_free++; 481} 482 483#ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT 484/* 485 * free_page_mlock() -- clean up attempts to free and mlocked() page. 486 * Page should not be on lru, so no need to fix that up. 487 * free_pages_check() will verify... 488 */ 489static inline void free_page_mlock(struct page *page) 490{ 491 __dec_zone_page_state(page, NR_MLOCK); 492 __count_vm_event(UNEVICTABLE_MLOCKFREED); 493} 494#else 495static void free_page_mlock(struct page *page) { } 496#endif 497 498static inline int free_pages_check(struct page *page) 499{ 500 if (unlikely(page_mapcount(page) | 501 (page->mapping != NULL) | 502 (atomic_read(&page->_count) != 0) | 503 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 504 bad_page(page); 505 return 1; 506 } 507 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 508 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 509 return 0; 510} 511 512/* 513 * Frees a list of pages. 514 * Assumes all pages on list are in same zone, and of same order. 515 * count is the number of pages to free. 516 * 517 * If the zone was previously in an "all pages pinned" state then look to 518 * see if this freeing clears that state. 519 * 520 * And clear the zone's pages_scanned counter, to hold off the "all pages are 521 * pinned" detection logic. 522 */ 523static void free_pages_bulk(struct zone *zone, int count, 524 struct list_head *list, int order) 525{ 526 spin_lock(&zone->lock); 527 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 528 zone->pages_scanned = 0; 529 530 __mod_zone_page_state(zone, NR_FREE_PAGES, count << order); 531 while (count--) { 532 struct page *page; 533 534 VM_BUG_ON(list_empty(list)); 535 page = list_entry(list->prev, struct page, lru); 536 /* have to delete it as __free_one_page list manipulates */ 537 list_del(&page->lru); 538 __free_one_page(page, zone, order, page_private(page)); 539 } 540 spin_unlock(&zone->lock); 541} 542 543static void free_one_page(struct zone *zone, struct page *page, int order, 544 int migratetype) 545{ 546 spin_lock(&zone->lock); 547 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 548 zone->pages_scanned = 0; 549 550 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 551 __free_one_page(page, zone, order, migratetype); 552 spin_unlock(&zone->lock); 553} 554 555static void __free_pages_ok(struct page *page, unsigned int order) 556{ 557 unsigned long flags; 558 int i; 559 int bad = 0; 560 int wasMlocked = __TestClearPageMlocked(page); 561 562 kmemcheck_free_shadow(page, order); 563 564 for (i = 0 ; i < (1 << order) ; ++i) 565 bad += free_pages_check(page + i); 566 if (bad) 567 return; 568 569 if (!PageHighMem(page)) { 570 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 571 debug_check_no_obj_freed(page_address(page), 572 PAGE_SIZE << order); 573 } 574 arch_free_page(page, order); 575 kernel_map_pages(page, 1 << order, 0); 576 577 local_irq_save(flags); 578 if (unlikely(wasMlocked)) 579 free_page_mlock(page); 580 __count_vm_events(PGFREE, 1 << order); 581 free_one_page(page_zone(page), page, order, 582 get_pageblock_migratetype(page)); 583 local_irq_restore(flags); 584} 585 586/* 587 * permit the bootmem allocator to evade page validation on high-order frees 588 */ 589void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 590{ 591 if (order == 0) { 592 __ClearPageReserved(page); 593 set_page_count(page, 0); 594 set_page_refcounted(page); 595 __free_page(page); 596 } else { 597 int loop; 598 599 prefetchw(page); 600 for (loop = 0; loop < BITS_PER_LONG; loop++) { 601 struct page *p = &page[loop]; 602 603 if (loop + 1 < BITS_PER_LONG) 604 prefetchw(p + 1); 605 __ClearPageReserved(p); 606 set_page_count(p, 0); 607 } 608 609 set_page_refcounted(page); 610 __free_pages(page, order); 611 } 612} 613 614 615/* 616 * The order of subdivision here is critical for the IO subsystem. 617 * Please do not alter this order without good reasons and regression 618 * testing. Specifically, as large blocks of memory are subdivided, 619 * the order in which smaller blocks are delivered depends on the order 620 * they're subdivided in this function. This is the primary factor 621 * influencing the order in which pages are delivered to the IO 622 * subsystem according to empirical testing, and this is also justified 623 * by considering the behavior of a buddy system containing a single 624 * large block of memory acted on by a series of small allocations. 625 * This behavior is a critical factor in sglist merging's success. 626 * 627 * -- wli 628 */ 629static inline void expand(struct zone *zone, struct page *page, 630 int low, int high, struct free_area *area, 631 int migratetype) 632{ 633 unsigned long size = 1 << high; 634 635 while (high > low) { 636 area--; 637 high--; 638 size >>= 1; 639 VM_BUG_ON(bad_range(zone, &page[size])); 640 list_add(&page[size].lru, &area->free_list[migratetype]); 641 area->nr_free++; 642 set_page_order(&page[size], high); 643 } 644} 645 646/* 647 * This page is about to be returned from the page allocator 648 */ 649static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 650{ 651 if (unlikely(page_mapcount(page) | 652 (page->mapping != NULL) | 653 (atomic_read(&page->_count) != 0) | 654 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 655 bad_page(page); 656 return 1; 657 } 658 659 set_page_private(page, 0); 660 set_page_refcounted(page); 661 662 arch_alloc_page(page, order); 663 kernel_map_pages(page, 1 << order, 1); 664 665 if (gfp_flags & __GFP_ZERO) 666 prep_zero_page(page, order, gfp_flags); 667 668 if (order && (gfp_flags & __GFP_COMP)) 669 prep_compound_page(page, order); 670 671 return 0; 672} 673 674/* 675 * Go through the free lists for the given migratetype and remove 676 * the smallest available page from the freelists 677 */ 678static inline 679struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 680 int migratetype) 681{ 682 unsigned int current_order; 683 struct free_area * area; 684 struct page *page; 685 686 /* Find a page of the appropriate size in the preferred list */ 687 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 688 area = &(zone->free_area[current_order]); 689 if (list_empty(&area->free_list[migratetype])) 690 continue; 691 692 page = list_entry(area->free_list[migratetype].next, 693 struct page, lru); 694 list_del(&page->lru); 695 rmv_page_order(page); 696 area->nr_free--; 697 expand(zone, page, order, current_order, area, migratetype); 698 return page; 699 } 700 701 return NULL; 702} 703 704 705/* 706 * This array describes the order lists are fallen back to when 707 * the free lists for the desirable migrate type are depleted 708 */ 709static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 710 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 711 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 712 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 713 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 714}; 715 716/* 717 * Move the free pages in a range to the free lists of the requested type. 718 * Note that start_page and end_pages are not aligned on a pageblock 719 * boundary. If alignment is required, use move_freepages_block() 720 */ 721static int move_freepages(struct zone *zone, 722 struct page *start_page, struct page *end_page, 723 int migratetype) 724{ 725 struct page *page; 726 unsigned long order; 727 int pages_moved = 0; 728 729#ifndef CONFIG_HOLES_IN_ZONE 730 /* 731 * page_zone is not safe to call in this context when 732 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 733 * anyway as we check zone boundaries in move_freepages_block(). 734 * Remove at a later date when no bug reports exist related to 735 * grouping pages by mobility 736 */ 737 BUG_ON(page_zone(start_page) != page_zone(end_page)); 738#endif 739 740 for (page = start_page; page <= end_page;) { 741 /* Make sure we are not inadvertently changing nodes */ 742 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 743 744 if (!pfn_valid_within(page_to_pfn(page))) { 745 page++; 746 continue; 747 } 748 749 if (!PageBuddy(page)) { 750 page++; 751 continue; 752 } 753 754 order = page_order(page); 755 list_del(&page->lru); 756 list_add(&page->lru, 757 &zone->free_area[order].free_list[migratetype]); 758 page += 1 << order; 759 pages_moved += 1 << order; 760 } 761 762 return pages_moved; 763} 764 765static int move_freepages_block(struct zone *zone, struct page *page, 766 int migratetype) 767{ 768 unsigned long start_pfn, end_pfn; 769 struct page *start_page, *end_page; 770 771 start_pfn = page_to_pfn(page); 772 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 773 start_page = pfn_to_page(start_pfn); 774 end_page = start_page + pageblock_nr_pages - 1; 775 end_pfn = start_pfn + pageblock_nr_pages - 1; 776 777 /* Do not cross zone boundaries */ 778 if (start_pfn < zone->zone_start_pfn) 779 start_page = page; 780 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 781 return 0; 782 783 return move_freepages(zone, start_page, end_page, migratetype); 784} 785 786static void change_pageblock_range(struct page *pageblock_page, 787 int start_order, int migratetype) 788{ 789 int nr_pageblocks = 1 << (start_order - pageblock_order); 790 791 while (nr_pageblocks--) { 792 set_pageblock_migratetype(pageblock_page, migratetype); 793 pageblock_page += pageblock_nr_pages; 794 } 795} 796 797/* Remove an element from the buddy allocator from the fallback list */ 798static inline struct page * 799__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 800{ 801 struct free_area * area; 802 int current_order; 803 struct page *page; 804 int migratetype, i; 805 806 /* Find the largest possible block of pages in the other list */ 807 for (current_order = MAX_ORDER-1; current_order >= order; 808 --current_order) { 809 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 810 migratetype = fallbacks[start_migratetype][i]; 811 812 /* MIGRATE_RESERVE handled later if necessary */ 813 if (migratetype == MIGRATE_RESERVE) 814 continue; 815 816 area = &(zone->free_area[current_order]); 817 if (list_empty(&area->free_list[migratetype])) 818 continue; 819 820 page = list_entry(area->free_list[migratetype].next, 821 struct page, lru); 822 area->nr_free--; 823 824 /* 825 * If breaking a large block of pages, move all free 826 * pages to the preferred allocation list. If falling 827 * back for a reclaimable kernel allocation, be more 828 * agressive about taking ownership of free pages 829 */ 830 if (unlikely(current_order >= (pageblock_order >> 1)) || 831 start_migratetype == MIGRATE_RECLAIMABLE || 832 page_group_by_mobility_disabled) { 833 unsigned long pages; 834 pages = move_freepages_block(zone, page, 835 start_migratetype); 836 837 /* Claim the whole block if over half of it is free */ 838 if (pages >= (1 << (pageblock_order-1)) || 839 page_group_by_mobility_disabled) 840 set_pageblock_migratetype(page, 841 start_migratetype); 842 843 migratetype = start_migratetype; 844 } 845 846 /* Remove the page from the freelists */ 847 list_del(&page->lru); 848 rmv_page_order(page); 849 850 /* Take ownership for orders >= pageblock_order */ 851 if (current_order >= pageblock_order) 852 change_pageblock_range(page, current_order, 853 start_migratetype); 854 855 expand(zone, page, order, current_order, area, migratetype); 856 return page; 857 } 858 } 859 860 return NULL; 861} 862 863/* 864 * Do the hard work of removing an element from the buddy allocator. 865 * Call me with the zone->lock already held. 866 */ 867static struct page *__rmqueue(struct zone *zone, unsigned int order, 868 int migratetype) 869{ 870 struct page *page; 871 872retry_reserve: 873 page = __rmqueue_smallest(zone, order, migratetype); 874 875 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 876 page = __rmqueue_fallback(zone, order, migratetype); 877 878 /* 879 * Use MIGRATE_RESERVE rather than fail an allocation. goto 880 * is used because __rmqueue_smallest is an inline function 881 * and we want just one call site 882 */ 883 if (!page) { 884 migratetype = MIGRATE_RESERVE; 885 goto retry_reserve; 886 } 887 } 888 889 return page; 890} 891 892/* 893 * Obtain a specified number of elements from the buddy allocator, all under 894 * a single hold of the lock, for efficiency. Add them to the supplied list. 895 * Returns the number of new pages which were placed at *list. 896 */ 897static int rmqueue_bulk(struct zone *zone, unsigned int order, 898 unsigned long count, struct list_head *list, 899 int migratetype, int cold) 900{ 901 int i; 902 903 spin_lock(&zone->lock); 904 for (i = 0; i < count; ++i) { 905 struct page *page = __rmqueue(zone, order, migratetype); 906 if (unlikely(page == NULL)) 907 break; 908 909 /* 910 * Split buddy pages returned by expand() are received here 911 * in physical page order. The page is added to the callers and 912 * list and the list head then moves forward. From the callers 913 * perspective, the linked list is ordered by page number in 914 * some conditions. This is useful for IO devices that can 915 * merge IO requests if the physical pages are ordered 916 * properly. 917 */ 918 if (likely(cold == 0)) 919 list_add(&page->lru, list); 920 else 921 list_add_tail(&page->lru, list); 922 set_page_private(page, migratetype); 923 list = &page->lru; 924 } 925 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 926 spin_unlock(&zone->lock); 927 return i; 928} 929 930#ifdef CONFIG_NUMA 931/* 932 * Called from the vmstat counter updater to drain pagesets of this 933 * currently executing processor on remote nodes after they have 934 * expired. 935 * 936 * Note that this function must be called with the thread pinned to 937 * a single processor. 938 */ 939void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 940{ 941 unsigned long flags; 942 int to_drain; 943 944 local_irq_save(flags); 945 if (pcp->count >= pcp->batch) 946 to_drain = pcp->batch; 947 else 948 to_drain = pcp->count; 949 free_pages_bulk(zone, to_drain, &pcp->list, 0); 950 pcp->count -= to_drain; 951 local_irq_restore(flags); 952} 953#endif 954 955/* 956 * Drain pages of the indicated processor. 957 * 958 * The processor must either be the current processor and the 959 * thread pinned to the current processor or a processor that 960 * is not online. 961 */ 962static void drain_pages(unsigned int cpu) 963{ 964 unsigned long flags; 965 struct zone *zone; 966 967 for_each_populated_zone(zone) { 968 struct per_cpu_pageset *pset; 969 struct per_cpu_pages *pcp; 970 971 pset = zone_pcp(zone, cpu); 972 973 pcp = &pset->pcp; 974 local_irq_save(flags); 975 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 976 pcp->count = 0; 977 local_irq_restore(flags); 978 } 979} 980 981/* 982 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 983 */ 984void drain_local_pages(void *arg) 985{ 986 drain_pages(smp_processor_id()); 987} 988 989/* 990 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 991 */ 992void drain_all_pages(void) 993{ 994 on_each_cpu(drain_local_pages, NULL, 1); 995} 996 997#ifdef CONFIG_HIBERNATION 998 999void mark_free_pages(struct zone *zone) 1000{ 1001 unsigned long pfn, max_zone_pfn; 1002 unsigned long flags; 1003 int order, t; 1004 struct list_head *curr; 1005 1006 if (!zone->spanned_pages) 1007 return; 1008 1009 spin_lock_irqsave(&zone->lock, flags); 1010 1011 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1012 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1013 if (pfn_valid(pfn)) { 1014 struct page *page = pfn_to_page(pfn); 1015 1016 if (!swsusp_page_is_forbidden(page)) 1017 swsusp_unset_page_free(page); 1018 } 1019 1020 for_each_migratetype_order(order, t) { 1021 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1022 unsigned long i; 1023 1024 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1025 for (i = 0; i < (1UL << order); i++) 1026 swsusp_set_page_free(pfn_to_page(pfn + i)); 1027 } 1028 } 1029 spin_unlock_irqrestore(&zone->lock, flags); 1030} 1031#endif /* CONFIG_PM */ 1032 1033/* 1034 * Free a 0-order page 1035 */ 1036static void free_hot_cold_page(struct page *page, int cold) 1037{ 1038 struct zone *zone = page_zone(page); 1039 struct per_cpu_pages *pcp; 1040 unsigned long flags; 1041 int wasMlocked = __TestClearPageMlocked(page); 1042 1043 kmemcheck_free_shadow(page, 0); 1044 1045 if (PageAnon(page)) 1046 page->mapping = NULL; 1047 if (free_pages_check(page)) 1048 return; 1049 1050 if (!PageHighMem(page)) { 1051 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1052 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1053 } 1054 arch_free_page(page, 0); 1055 kernel_map_pages(page, 1, 0); 1056 1057 pcp = &zone_pcp(zone, get_cpu())->pcp; 1058 set_page_private(page, get_pageblock_migratetype(page)); 1059 local_irq_save(flags); 1060 if (unlikely(wasMlocked)) 1061 free_page_mlock(page); 1062 __count_vm_event(PGFREE); 1063 1064 if (cold) 1065 list_add_tail(&page->lru, &pcp->list); 1066 else 1067 list_add(&page->lru, &pcp->list); 1068 pcp->count++; 1069 if (pcp->count >= pcp->high) { 1070 free_pages_bulk(zone, pcp->batch, &pcp->list, 0); 1071 pcp->count -= pcp->batch; 1072 } 1073 local_irq_restore(flags); 1074 put_cpu(); 1075} 1076 1077void free_hot_page(struct page *page) 1078{ 1079 trace_mm_page_free_direct(page, 0); 1080 free_hot_cold_page(page, 0); 1081} 1082 1083/* 1084 * split_page takes a non-compound higher-order page, and splits it into 1085 * n (1<<order) sub-pages: page[0..n] 1086 * Each sub-page must be freed individually. 1087 * 1088 * Note: this is probably too low level an operation for use in drivers. 1089 * Please consult with lkml before using this in your driver. 1090 */ 1091void split_page(struct page *page, unsigned int order) 1092{ 1093 int i; 1094 1095 VM_BUG_ON(PageCompound(page)); 1096 VM_BUG_ON(!page_count(page)); 1097 1098#ifdef CONFIG_KMEMCHECK 1099 /* 1100 * Split shadow pages too, because free(page[0]) would 1101 * otherwise free the whole shadow. 1102 */ 1103 if (kmemcheck_page_is_tracked(page)) 1104 split_page(virt_to_page(page[0].shadow), order); 1105#endif 1106 1107 for (i = 1; i < (1 << order); i++) 1108 set_page_refcounted(page + i); 1109} 1110 1111/* 1112 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1113 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1114 * or two. 1115 */ 1116static inline 1117struct page *buffered_rmqueue(struct zone *preferred_zone, 1118 struct zone *zone, int order, gfp_t gfp_flags, 1119 int migratetype) 1120{ 1121 unsigned long flags; 1122 struct page *page; 1123 int cold = !!(gfp_flags & __GFP_COLD); 1124 int cpu; 1125 1126again: 1127 cpu = get_cpu(); 1128 if (likely(order == 0)) { 1129 struct per_cpu_pages *pcp; 1130 1131 pcp = &zone_pcp(zone, cpu)->pcp; 1132 local_irq_save(flags); 1133 if (!pcp->count) { 1134 pcp->count = rmqueue_bulk(zone, 0, 1135 pcp->batch, &pcp->list, 1136 migratetype, cold); 1137 if (unlikely(!pcp->count)) 1138 goto failed; 1139 } 1140 1141 /* Find a page of the appropriate migrate type */ 1142 if (cold) { 1143 list_for_each_entry_reverse(page, &pcp->list, lru) 1144 if (page_private(page) == migratetype) 1145 break; 1146 } else { 1147 list_for_each_entry(page, &pcp->list, lru) 1148 if (page_private(page) == migratetype) 1149 break; 1150 } 1151 1152 /* Allocate more to the pcp list if necessary */ 1153 if (unlikely(&page->lru == &pcp->list)) { 1154 int get_one_page = 0; 1155 1156 pcp->count += rmqueue_bulk(zone, 0, 1157 pcp->batch, &pcp->list, 1158 migratetype, cold); 1159 list_for_each_entry(page, &pcp->list, lru) { 1160 if (get_pageblock_migratetype(page) != 1161 MIGRATE_ISOLATE) { 1162 get_one_page = 1; 1163 break; 1164 } 1165 } 1166 if (!get_one_page) 1167 goto failed; 1168 } 1169 1170 list_del(&page->lru); 1171 pcp->count--; 1172 } else { 1173 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1174 /* 1175 * __GFP_NOFAIL is not to be used in new code. 1176 * 1177 * All __GFP_NOFAIL callers should be fixed so that they 1178 * properly detect and handle allocation failures. 1179 * 1180 * We most definitely don't want callers attempting to 1181 * allocate greater than order-1 page units with 1182 * __GFP_NOFAIL. 1183 */ 1184 WARN_ON_ONCE(order > 1); 1185 } 1186 spin_lock_irqsave(&zone->lock, flags); 1187 page = __rmqueue(zone, order, migratetype); 1188 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1189 spin_unlock(&zone->lock); 1190 if (!page) 1191 goto failed; 1192 } 1193 1194 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1195 zone_statistics(preferred_zone, zone); 1196 local_irq_restore(flags); 1197 put_cpu(); 1198 1199 VM_BUG_ON(bad_range(zone, page)); 1200 if (prep_new_page(page, order, gfp_flags)) 1201 goto again; 1202 return page; 1203 1204failed: 1205 local_irq_restore(flags); 1206 put_cpu(); 1207 return NULL; 1208} 1209 1210/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1211#define ALLOC_WMARK_MIN WMARK_MIN 1212#define ALLOC_WMARK_LOW WMARK_LOW 1213#define ALLOC_WMARK_HIGH WMARK_HIGH 1214#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1215 1216/* Mask to get the watermark bits */ 1217#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1218 1219#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1220#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1221#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1222 1223#ifdef CONFIG_FAIL_PAGE_ALLOC 1224 1225static struct fail_page_alloc_attr { 1226 struct fault_attr attr; 1227 1228 u32 ignore_gfp_highmem; 1229 u32 ignore_gfp_wait; 1230 u32 min_order; 1231 1232#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1233 1234 struct dentry *ignore_gfp_highmem_file; 1235 struct dentry *ignore_gfp_wait_file; 1236 struct dentry *min_order_file; 1237 1238#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1239 1240} fail_page_alloc = { 1241 .attr = FAULT_ATTR_INITIALIZER, 1242 .ignore_gfp_wait = 1, 1243 .ignore_gfp_highmem = 1, 1244 .min_order = 1, 1245}; 1246 1247static int __init setup_fail_page_alloc(char *str) 1248{ 1249 return setup_fault_attr(&fail_page_alloc.attr, str); 1250} 1251__setup("fail_page_alloc=", setup_fail_page_alloc); 1252 1253static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1254{ 1255 if (order < fail_page_alloc.min_order) 1256 return 0; 1257 if (gfp_mask & __GFP_NOFAIL) 1258 return 0; 1259 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1260 return 0; 1261 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1262 return 0; 1263 1264 return should_fail(&fail_page_alloc.attr, 1 << order); 1265} 1266 1267#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1268 1269static int __init fail_page_alloc_debugfs(void) 1270{ 1271 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1272 struct dentry *dir; 1273 int err; 1274 1275 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1276 "fail_page_alloc"); 1277 if (err) 1278 return err; 1279 dir = fail_page_alloc.attr.dentries.dir; 1280 1281 fail_page_alloc.ignore_gfp_wait_file = 1282 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1283 &fail_page_alloc.ignore_gfp_wait); 1284 1285 fail_page_alloc.ignore_gfp_highmem_file = 1286 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1287 &fail_page_alloc.ignore_gfp_highmem); 1288 fail_page_alloc.min_order_file = 1289 debugfs_create_u32("min-order", mode, dir, 1290 &fail_page_alloc.min_order); 1291 1292 if (!fail_page_alloc.ignore_gfp_wait_file || 1293 !fail_page_alloc.ignore_gfp_highmem_file || 1294 !fail_page_alloc.min_order_file) { 1295 err = -ENOMEM; 1296 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1297 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1298 debugfs_remove(fail_page_alloc.min_order_file); 1299 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1300 } 1301 1302 return err; 1303} 1304 1305late_initcall(fail_page_alloc_debugfs); 1306 1307#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1308 1309#else /* CONFIG_FAIL_PAGE_ALLOC */ 1310 1311static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1312{ 1313 return 0; 1314} 1315 1316#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1317 1318/* 1319 * Return 1 if free pages are above 'mark'. This takes into account the order 1320 * of the allocation. 1321 */ 1322int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1323 int classzone_idx, int alloc_flags) 1324{ 1325 /* free_pages my go negative - that's OK */ 1326 long min = mark; 1327 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1328 int o; 1329 1330 if (alloc_flags & ALLOC_HIGH) 1331 min -= min / 2; 1332 if (alloc_flags & ALLOC_HARDER) 1333 min -= min / 4; 1334 1335 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1336 return 0; 1337 for (o = 0; o < order; o++) { 1338 /* At the next order, this order's pages become unavailable */ 1339 free_pages -= z->free_area[o].nr_free << o; 1340 1341 /* Require fewer higher order pages to be free */ 1342 min >>= 1; 1343 1344 if (free_pages <= min) 1345 return 0; 1346 } 1347 return 1; 1348} 1349 1350#ifdef CONFIG_NUMA 1351/* 1352 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1353 * skip over zones that are not allowed by the cpuset, or that have 1354 * been recently (in last second) found to be nearly full. See further 1355 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1356 * that have to skip over a lot of full or unallowed zones. 1357 * 1358 * If the zonelist cache is present in the passed in zonelist, then 1359 * returns a pointer to the allowed node mask (either the current 1360 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1361 * 1362 * If the zonelist cache is not available for this zonelist, does 1363 * nothing and returns NULL. 1364 * 1365 * If the fullzones BITMAP in the zonelist cache is stale (more than 1366 * a second since last zap'd) then we zap it out (clear its bits.) 1367 * 1368 * We hold off even calling zlc_setup, until after we've checked the 1369 * first zone in the zonelist, on the theory that most allocations will 1370 * be satisfied from that first zone, so best to examine that zone as 1371 * quickly as we can. 1372 */ 1373static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1374{ 1375 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1376 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1377 1378 zlc = zonelist->zlcache_ptr; 1379 if (!zlc) 1380 return NULL; 1381 1382 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1383 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1384 zlc->last_full_zap = jiffies; 1385 } 1386 1387 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1388 &cpuset_current_mems_allowed : 1389 &node_states[N_HIGH_MEMORY]; 1390 return allowednodes; 1391} 1392 1393/* 1394 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1395 * if it is worth looking at further for free memory: 1396 * 1) Check that the zone isn't thought to be full (doesn't have its 1397 * bit set in the zonelist_cache fullzones BITMAP). 1398 * 2) Check that the zones node (obtained from the zonelist_cache 1399 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1400 * Return true (non-zero) if zone is worth looking at further, or 1401 * else return false (zero) if it is not. 1402 * 1403 * This check -ignores- the distinction between various watermarks, 1404 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1405 * found to be full for any variation of these watermarks, it will 1406 * be considered full for up to one second by all requests, unless 1407 * we are so low on memory on all allowed nodes that we are forced 1408 * into the second scan of the zonelist. 1409 * 1410 * In the second scan we ignore this zonelist cache and exactly 1411 * apply the watermarks to all zones, even it is slower to do so. 1412 * We are low on memory in the second scan, and should leave no stone 1413 * unturned looking for a free page. 1414 */ 1415static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1416 nodemask_t *allowednodes) 1417{ 1418 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1419 int i; /* index of *z in zonelist zones */ 1420 int n; /* node that zone *z is on */ 1421 1422 zlc = zonelist->zlcache_ptr; 1423 if (!zlc) 1424 return 1; 1425 1426 i = z - zonelist->_zonerefs; 1427 n = zlc->z_to_n[i]; 1428 1429 /* This zone is worth trying if it is allowed but not full */ 1430 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1431} 1432 1433/* 1434 * Given 'z' scanning a zonelist, set the corresponding bit in 1435 * zlc->fullzones, so that subsequent attempts to allocate a page 1436 * from that zone don't waste time re-examining it. 1437 */ 1438static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1439{ 1440 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1441 int i; /* index of *z in zonelist zones */ 1442 1443 zlc = zonelist->zlcache_ptr; 1444 if (!zlc) 1445 return; 1446 1447 i = z - zonelist->_zonerefs; 1448 1449 set_bit(i, zlc->fullzones); 1450} 1451 1452#else /* CONFIG_NUMA */ 1453 1454static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1455{ 1456 return NULL; 1457} 1458 1459static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1460 nodemask_t *allowednodes) 1461{ 1462 return 1; 1463} 1464 1465static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1466{ 1467} 1468#endif /* CONFIG_NUMA */ 1469 1470/* 1471 * get_page_from_freelist goes through the zonelist trying to allocate 1472 * a page. 1473 */ 1474static struct page * 1475get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1476 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1477 struct zone *preferred_zone, int migratetype) 1478{ 1479 struct zoneref *z; 1480 struct page *page = NULL; 1481 int classzone_idx; 1482 struct zone *zone; 1483 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1484 int zlc_active = 0; /* set if using zonelist_cache */ 1485 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1486 1487 classzone_idx = zone_idx(preferred_zone); 1488zonelist_scan: 1489 /* 1490 * Scan zonelist, looking for a zone with enough free. 1491 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1492 */ 1493 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1494 high_zoneidx, nodemask) { 1495 if (NUMA_BUILD && zlc_active && 1496 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1497 continue; 1498 if ((alloc_flags & ALLOC_CPUSET) && 1499 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1500 goto try_next_zone; 1501 1502 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1503 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1504 unsigned long mark; 1505 int ret; 1506 1507 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1508 if (zone_watermark_ok(zone, order, mark, 1509 classzone_idx, alloc_flags)) 1510 goto try_this_zone; 1511 1512 if (zone_reclaim_mode == 0) 1513 goto this_zone_full; 1514 1515 ret = zone_reclaim(zone, gfp_mask, order); 1516 switch (ret) { 1517 case ZONE_RECLAIM_NOSCAN: 1518 /* did not scan */ 1519 goto try_next_zone; 1520 case ZONE_RECLAIM_FULL: 1521 /* scanned but unreclaimable */ 1522 goto this_zone_full; 1523 default: 1524 /* did we reclaim enough */ 1525 if (!zone_watermark_ok(zone, order, mark, 1526 classzone_idx, alloc_flags)) 1527 goto this_zone_full; 1528 } 1529 } 1530 1531try_this_zone: 1532 page = buffered_rmqueue(preferred_zone, zone, order, 1533 gfp_mask, migratetype); 1534 if (page) 1535 break; 1536this_zone_full: 1537 if (NUMA_BUILD) 1538 zlc_mark_zone_full(zonelist, z); 1539try_next_zone: 1540 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1541 /* 1542 * we do zlc_setup after the first zone is tried but only 1543 * if there are multiple nodes make it worthwhile 1544 */ 1545 allowednodes = zlc_setup(zonelist, alloc_flags); 1546 zlc_active = 1; 1547 did_zlc_setup = 1; 1548 } 1549 } 1550 1551 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1552 /* Disable zlc cache for second zonelist scan */ 1553 zlc_active = 0; 1554 goto zonelist_scan; 1555 } 1556 return page; 1557} 1558 1559static inline int 1560should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1561 unsigned long pages_reclaimed) 1562{ 1563 /* Do not loop if specifically requested */ 1564 if (gfp_mask & __GFP_NORETRY) 1565 return 0; 1566 1567 /* 1568 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1569 * means __GFP_NOFAIL, but that may not be true in other 1570 * implementations. 1571 */ 1572 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1573 return 1; 1574 1575 /* 1576 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1577 * specified, then we retry until we no longer reclaim any pages 1578 * (above), or we've reclaimed an order of pages at least as 1579 * large as the allocation's order. In both cases, if the 1580 * allocation still fails, we stop retrying. 1581 */ 1582 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1583 return 1; 1584 1585 /* 1586 * Don't let big-order allocations loop unless the caller 1587 * explicitly requests that. 1588 */ 1589 if (gfp_mask & __GFP_NOFAIL) 1590 return 1; 1591 1592 return 0; 1593} 1594 1595static inline struct page * 1596__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1597 struct zonelist *zonelist, enum zone_type high_zoneidx, 1598 nodemask_t *nodemask, struct zone *preferred_zone, 1599 int migratetype) 1600{ 1601 struct page *page; 1602 1603 /* Acquire the OOM killer lock for the zones in zonelist */ 1604 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1605 schedule_timeout_uninterruptible(1); 1606 return NULL; 1607 } 1608 1609 /* 1610 * Go through the zonelist yet one more time, keep very high watermark 1611 * here, this is only to catch a parallel oom killing, we must fail if 1612 * we're still under heavy pressure. 1613 */ 1614 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1615 order, zonelist, high_zoneidx, 1616 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1617 preferred_zone, migratetype); 1618 if (page) 1619 goto out; 1620 1621 /* The OOM killer will not help higher order allocs */ 1622 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_NOFAIL)) 1623 goto out; 1624 1625 /* Exhausted what can be done so it's blamo time */ 1626 out_of_memory(zonelist, gfp_mask, order); 1627 1628out: 1629 clear_zonelist_oom(zonelist, gfp_mask); 1630 return page; 1631} 1632 1633/* The really slow allocator path where we enter direct reclaim */ 1634static inline struct page * 1635__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1636 struct zonelist *zonelist, enum zone_type high_zoneidx, 1637 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1638 int migratetype, unsigned long *did_some_progress) 1639{ 1640 struct page *page = NULL; 1641 struct reclaim_state reclaim_state; 1642 struct task_struct *p = current; 1643 1644 cond_resched(); 1645 1646 /* We now go into synchronous reclaim */ 1647 cpuset_memory_pressure_bump(); 1648 p->flags |= PF_MEMALLOC; 1649 lockdep_set_current_reclaim_state(gfp_mask); 1650 reclaim_state.reclaimed_slab = 0; 1651 p->reclaim_state = &reclaim_state; 1652 1653 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1654 1655 p->reclaim_state = NULL; 1656 lockdep_clear_current_reclaim_state(); 1657 p->flags &= ~PF_MEMALLOC; 1658 1659 cond_resched(); 1660 1661 if (order != 0) 1662 drain_all_pages(); 1663 1664 if (likely(*did_some_progress)) 1665 page = get_page_from_freelist(gfp_mask, nodemask, order, 1666 zonelist, high_zoneidx, 1667 alloc_flags, preferred_zone, 1668 migratetype); 1669 return page; 1670} 1671 1672/* 1673 * This is called in the allocator slow-path if the allocation request is of 1674 * sufficient urgency to ignore watermarks and take other desperate measures 1675 */ 1676static inline struct page * 1677__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1678 struct zonelist *zonelist, enum zone_type high_zoneidx, 1679 nodemask_t *nodemask, struct zone *preferred_zone, 1680 int migratetype) 1681{ 1682 struct page *page; 1683 1684 do { 1685 page = get_page_from_freelist(gfp_mask, nodemask, order, 1686 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1687 preferred_zone, migratetype); 1688 1689 if (!page && gfp_mask & __GFP_NOFAIL) 1690 congestion_wait(BLK_RW_ASYNC, HZ/50); 1691 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1692 1693 return page; 1694} 1695 1696static inline 1697void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1698 enum zone_type high_zoneidx) 1699{ 1700 struct zoneref *z; 1701 struct zone *zone; 1702 1703 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1704 wakeup_kswapd(zone, order); 1705} 1706 1707static inline int 1708gfp_to_alloc_flags(gfp_t gfp_mask) 1709{ 1710 struct task_struct *p = current; 1711 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1712 const gfp_t wait = gfp_mask & __GFP_WAIT; 1713 1714 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1715 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1716 1717 /* 1718 * The caller may dip into page reserves a bit more if the caller 1719 * cannot run direct reclaim, or if the caller has realtime scheduling 1720 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1721 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1722 */ 1723 alloc_flags |= (gfp_mask & __GFP_HIGH); 1724 1725 if (!wait) { 1726 alloc_flags |= ALLOC_HARDER; 1727 /* 1728 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1729 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1730 */ 1731 alloc_flags &= ~ALLOC_CPUSET; 1732 } else if (unlikely(rt_task(p))) 1733 alloc_flags |= ALLOC_HARDER; 1734 1735 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1736 if (!in_interrupt() && 1737 ((p->flags & PF_MEMALLOC) || 1738 unlikely(test_thread_flag(TIF_MEMDIE)))) 1739 alloc_flags |= ALLOC_NO_WATERMARKS; 1740 } 1741 1742 return alloc_flags; 1743} 1744 1745static inline struct page * 1746__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1747 struct zonelist *zonelist, enum zone_type high_zoneidx, 1748 nodemask_t *nodemask, struct zone *preferred_zone, 1749 int migratetype) 1750{ 1751 const gfp_t wait = gfp_mask & __GFP_WAIT; 1752 struct page *page = NULL; 1753 int alloc_flags; 1754 unsigned long pages_reclaimed = 0; 1755 unsigned long did_some_progress; 1756 struct task_struct *p = current; 1757 1758 /* 1759 * In the slowpath, we sanity check order to avoid ever trying to 1760 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1761 * be using allocators in order of preference for an area that is 1762 * too large. 1763 */ 1764 if (order >= MAX_ORDER) { 1765 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1766 return NULL; 1767 } 1768 1769 /* 1770 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1771 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1772 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1773 * using a larger set of nodes after it has established that the 1774 * allowed per node queues are empty and that nodes are 1775 * over allocated. 1776 */ 1777 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1778 goto nopage; 1779 1780 wake_all_kswapd(order, zonelist, high_zoneidx); 1781 1782restart: 1783 /* 1784 * OK, we're below the kswapd watermark and have kicked background 1785 * reclaim. Now things get more complex, so set up alloc_flags according 1786 * to how we want to proceed. 1787 */ 1788 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1789 1790 /* This is the last chance, in general, before the goto nopage. */ 1791 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1792 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1793 preferred_zone, migratetype); 1794 if (page) 1795 goto got_pg; 1796 1797rebalance: 1798 /* Allocate without watermarks if the context allows */ 1799 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1800 page = __alloc_pages_high_priority(gfp_mask, order, 1801 zonelist, high_zoneidx, nodemask, 1802 preferred_zone, migratetype); 1803 if (page) 1804 goto got_pg; 1805 } 1806 1807 /* Atomic allocations - we can't balance anything */ 1808 if (!wait) 1809 goto nopage; 1810 1811 /* Avoid recursion of direct reclaim */ 1812 if (p->flags & PF_MEMALLOC) 1813 goto nopage; 1814 1815 /* Avoid allocations with no watermarks from looping endlessly */ 1816 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 1817 goto nopage; 1818 1819 /* Try direct reclaim and then allocating */ 1820 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1821 zonelist, high_zoneidx, 1822 nodemask, 1823 alloc_flags, preferred_zone, 1824 migratetype, &did_some_progress); 1825 if (page) 1826 goto got_pg; 1827 1828 /* 1829 * If we failed to make any progress reclaiming, then we are 1830 * running out of options and have to consider going OOM 1831 */ 1832 if (!did_some_progress) { 1833 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1834 if (oom_killer_disabled) 1835 goto nopage; 1836 page = __alloc_pages_may_oom(gfp_mask, order, 1837 zonelist, high_zoneidx, 1838 nodemask, preferred_zone, 1839 migratetype); 1840 if (page) 1841 goto got_pg; 1842 1843 /* 1844 * The OOM killer does not trigger for high-order 1845 * ~__GFP_NOFAIL allocations so if no progress is being 1846 * made, there are no other options and retrying is 1847 * unlikely to help. 1848 */ 1849 if (order > PAGE_ALLOC_COSTLY_ORDER && 1850 !(gfp_mask & __GFP_NOFAIL)) 1851 goto nopage; 1852 1853 goto restart; 1854 } 1855 } 1856 1857 /* Check if we should retry the allocation */ 1858 pages_reclaimed += did_some_progress; 1859 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1860 /* Wait for some write requests to complete then retry */ 1861 congestion_wait(BLK_RW_ASYNC, HZ/50); 1862 goto rebalance; 1863 } 1864 1865nopage: 1866 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1867 printk(KERN_WARNING "%s: page allocation failure." 1868 " order:%d, mode:0x%x\n", 1869 p->comm, order, gfp_mask); 1870 dump_stack(); 1871 show_mem(); 1872 } 1873 return page; 1874got_pg: 1875 if (kmemcheck_enabled) 1876 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 1877 return page; 1878 1879} 1880 1881/* 1882 * This is the 'heart' of the zoned buddy allocator. 1883 */ 1884struct page * 1885__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1886 struct zonelist *zonelist, nodemask_t *nodemask) 1887{ 1888 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1889 struct zone *preferred_zone; 1890 struct page *page; 1891 int migratetype = allocflags_to_migratetype(gfp_mask); 1892 1893 gfp_mask &= gfp_allowed_mask; 1894 1895 lockdep_trace_alloc(gfp_mask); 1896 1897 might_sleep_if(gfp_mask & __GFP_WAIT); 1898 1899 if (should_fail_alloc_page(gfp_mask, order)) 1900 return NULL; 1901 1902 /* 1903 * Check the zones suitable for the gfp_mask contain at least one 1904 * valid zone. It's possible to have an empty zonelist as a result 1905 * of GFP_THISNODE and a memoryless node 1906 */ 1907 if (unlikely(!zonelist->_zonerefs->zone)) 1908 return NULL; 1909 1910 /* The preferred zone is used for statistics later */ 1911 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1912 if (!preferred_zone) 1913 return NULL; 1914 1915 /* First allocation attempt */ 1916 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1917 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1918 preferred_zone, migratetype); 1919 if (unlikely(!page)) 1920 page = __alloc_pages_slowpath(gfp_mask, order, 1921 zonelist, high_zoneidx, nodemask, 1922 preferred_zone, migratetype); 1923 1924 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 1925 return page; 1926} 1927EXPORT_SYMBOL(__alloc_pages_nodemask); 1928 1929/* 1930 * Common helper functions. 1931 */ 1932unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1933{ 1934 struct page *page; 1935 1936 /* 1937 * __get_free_pages() returns a 32-bit address, which cannot represent 1938 * a highmem page 1939 */ 1940 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1941 1942 page = alloc_pages(gfp_mask, order); 1943 if (!page) 1944 return 0; 1945 return (unsigned long) page_address(page); 1946} 1947EXPORT_SYMBOL(__get_free_pages); 1948 1949unsigned long get_zeroed_page(gfp_t gfp_mask) 1950{ 1951 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 1952} 1953EXPORT_SYMBOL(get_zeroed_page); 1954 1955void __pagevec_free(struct pagevec *pvec) 1956{ 1957 int i = pagevec_count(pvec); 1958 1959 while (--i >= 0) { 1960 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 1961 free_hot_cold_page(pvec->pages[i], pvec->cold); 1962 } 1963} 1964 1965void __free_pages(struct page *page, unsigned int order) 1966{ 1967 if (put_page_testzero(page)) { 1968 trace_mm_page_free_direct(page, order); 1969 if (order == 0) 1970 free_hot_page(page); 1971 else 1972 __free_pages_ok(page, order); 1973 } 1974} 1975 1976EXPORT_SYMBOL(__free_pages); 1977 1978void free_pages(unsigned long addr, unsigned int order) 1979{ 1980 if (addr != 0) { 1981 VM_BUG_ON(!virt_addr_valid((void *)addr)); 1982 __free_pages(virt_to_page((void *)addr), order); 1983 } 1984} 1985 1986EXPORT_SYMBOL(free_pages); 1987 1988/** 1989 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 1990 * @size: the number of bytes to allocate 1991 * @gfp_mask: GFP flags for the allocation 1992 * 1993 * This function is similar to alloc_pages(), except that it allocates the 1994 * minimum number of pages to satisfy the request. alloc_pages() can only 1995 * allocate memory in power-of-two pages. 1996 * 1997 * This function is also limited by MAX_ORDER. 1998 * 1999 * Memory allocated by this function must be released by free_pages_exact(). 2000 */ 2001void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2002{ 2003 unsigned int order = get_order(size); 2004 unsigned long addr; 2005 2006 addr = __get_free_pages(gfp_mask, order); 2007 if (addr) { 2008 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2009 unsigned long used = addr + PAGE_ALIGN(size); 2010 2011 split_page(virt_to_page((void *)addr), order); 2012 while (used < alloc_end) { 2013 free_page(used); 2014 used += PAGE_SIZE; 2015 } 2016 } 2017 2018 return (void *)addr; 2019} 2020EXPORT_SYMBOL(alloc_pages_exact); 2021 2022/** 2023 * free_pages_exact - release memory allocated via alloc_pages_exact() 2024 * @virt: the value returned by alloc_pages_exact. 2025 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2026 * 2027 * Release the memory allocated by a previous call to alloc_pages_exact. 2028 */ 2029void free_pages_exact(void *virt, size_t size) 2030{ 2031 unsigned long addr = (unsigned long)virt; 2032 unsigned long end = addr + PAGE_ALIGN(size); 2033 2034 while (addr < end) { 2035 free_page(addr); 2036 addr += PAGE_SIZE; 2037 } 2038} 2039EXPORT_SYMBOL(free_pages_exact); 2040 2041static unsigned int nr_free_zone_pages(int offset) 2042{ 2043 struct zoneref *z; 2044 struct zone *zone; 2045 2046 /* Just pick one node, since fallback list is circular */ 2047 unsigned int sum = 0; 2048 2049 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2050 2051 for_each_zone_zonelist(zone, z, zonelist, offset) { 2052 unsigned long size = zone->present_pages; 2053 unsigned long high = high_wmark_pages(zone); 2054 if (size > high) 2055 sum += size - high; 2056 } 2057 2058 return sum; 2059} 2060 2061/* 2062 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2063 */ 2064unsigned int nr_free_buffer_pages(void) 2065{ 2066 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2067} 2068EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2069 2070/* 2071 * Amount of free RAM allocatable within all zones 2072 */ 2073unsigned int nr_free_pagecache_pages(void) 2074{ 2075 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2076} 2077 2078static inline void show_node(struct zone *zone) 2079{ 2080 if (NUMA_BUILD) 2081 printk("Node %d ", zone_to_nid(zone)); 2082} 2083 2084void si_meminfo(struct sysinfo *val) 2085{ 2086 val->totalram = totalram_pages; 2087 val->sharedram = 0; 2088 val->freeram = global_page_state(NR_FREE_PAGES); 2089 val->bufferram = nr_blockdev_pages(); 2090 val->totalhigh = totalhigh_pages; 2091 val->freehigh = nr_free_highpages(); 2092 val->mem_unit = PAGE_SIZE; 2093} 2094 2095EXPORT_SYMBOL(si_meminfo); 2096 2097#ifdef CONFIG_NUMA 2098void si_meminfo_node(struct sysinfo *val, int nid) 2099{ 2100 pg_data_t *pgdat = NODE_DATA(nid); 2101 2102 val->totalram = pgdat->node_present_pages; 2103 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2104#ifdef CONFIG_HIGHMEM 2105 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2106 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2107 NR_FREE_PAGES); 2108#else 2109 val->totalhigh = 0; 2110 val->freehigh = 0; 2111#endif 2112 val->mem_unit = PAGE_SIZE; 2113} 2114#endif 2115 2116#define K(x) ((x) << (PAGE_SHIFT-10)) 2117 2118/* 2119 * Show free area list (used inside shift_scroll-lock stuff) 2120 * We also calculate the percentage fragmentation. We do this by counting the 2121 * memory on each free list with the exception of the first item on the list. 2122 */ 2123void show_free_areas(void) 2124{ 2125 int cpu; 2126 struct zone *zone; 2127 2128 for_each_populated_zone(zone) { 2129 show_node(zone); 2130 printk("%s per-cpu:\n", zone->name); 2131 2132 for_each_online_cpu(cpu) { 2133 struct per_cpu_pageset *pageset; 2134 2135 pageset = zone_pcp(zone, cpu); 2136 2137 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2138 cpu, pageset->pcp.high, 2139 pageset->pcp.batch, pageset->pcp.count); 2140 } 2141 } 2142 2143 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2144 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2145 " unevictable:%lu" 2146 " dirty:%lu writeback:%lu unstable:%lu buffer:%lu\n" 2147 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2148 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2149 global_page_state(NR_ACTIVE_ANON), 2150 global_page_state(NR_INACTIVE_ANON), 2151 global_page_state(NR_ISOLATED_ANON), 2152 global_page_state(NR_ACTIVE_FILE), 2153 global_page_state(NR_INACTIVE_FILE), 2154 global_page_state(NR_ISOLATED_FILE), 2155 global_page_state(NR_UNEVICTABLE), 2156 global_page_state(NR_FILE_DIRTY), 2157 global_page_state(NR_WRITEBACK), 2158 global_page_state(NR_UNSTABLE_NFS), 2159 nr_blockdev_pages(), 2160 global_page_state(NR_FREE_PAGES), 2161 global_page_state(NR_SLAB_RECLAIMABLE), 2162 global_page_state(NR_SLAB_UNRECLAIMABLE), 2163 global_page_state(NR_FILE_MAPPED), 2164 global_page_state(NR_SHMEM), 2165 global_page_state(NR_PAGETABLE), 2166 global_page_state(NR_BOUNCE)); 2167 2168 for_each_populated_zone(zone) { 2169 int i; 2170 2171 show_node(zone); 2172 printk("%s" 2173 " free:%lukB" 2174 " min:%lukB" 2175 " low:%lukB" 2176 " high:%lukB" 2177 " active_anon:%lukB" 2178 " inactive_anon:%lukB" 2179 " active_file:%lukB" 2180 " inactive_file:%lukB" 2181 " unevictable:%lukB" 2182 " isolated(anon):%lukB" 2183 " isolated(file):%lukB" 2184 " present:%lukB" 2185 " mlocked:%lukB" 2186 " dirty:%lukB" 2187 " writeback:%lukB" 2188 " mapped:%lukB" 2189 " shmem:%lukB" 2190 " slab_reclaimable:%lukB" 2191 " slab_unreclaimable:%lukB" 2192 " kernel_stack:%lukB" 2193 " pagetables:%lukB" 2194 " unstable:%lukB" 2195 " bounce:%lukB" 2196 " writeback_tmp:%lukB" 2197 " pages_scanned:%lu" 2198 " all_unreclaimable? %s" 2199 "\n", 2200 zone->name, 2201 K(zone_page_state(zone, NR_FREE_PAGES)), 2202 K(min_wmark_pages(zone)), 2203 K(low_wmark_pages(zone)), 2204 K(high_wmark_pages(zone)), 2205 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2206 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2207 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2208 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2209 K(zone_page_state(zone, NR_UNEVICTABLE)), 2210 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2211 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2212 K(zone->present_pages), 2213 K(zone_page_state(zone, NR_MLOCK)), 2214 K(zone_page_state(zone, NR_FILE_DIRTY)), 2215 K(zone_page_state(zone, NR_WRITEBACK)), 2216 K(zone_page_state(zone, NR_FILE_MAPPED)), 2217 K(zone_page_state(zone, NR_SHMEM)), 2218 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2219 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2220 zone_page_state(zone, NR_KERNEL_STACK) * 2221 THREAD_SIZE / 1024, 2222 K(zone_page_state(zone, NR_PAGETABLE)), 2223 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2224 K(zone_page_state(zone, NR_BOUNCE)), 2225 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2226 zone->pages_scanned, 2227 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 2228 ); 2229 printk("lowmem_reserve[]:"); 2230 for (i = 0; i < MAX_NR_ZONES; i++) 2231 printk(" %lu", zone->lowmem_reserve[i]); 2232 printk("\n"); 2233 } 2234 2235 for_each_populated_zone(zone) { 2236 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2237 2238 show_node(zone); 2239 printk("%s: ", zone->name); 2240 2241 spin_lock_irqsave(&zone->lock, flags); 2242 for (order = 0; order < MAX_ORDER; order++) { 2243 nr[order] = zone->free_area[order].nr_free; 2244 total += nr[order] << order; 2245 } 2246 spin_unlock_irqrestore(&zone->lock, flags); 2247 for (order = 0; order < MAX_ORDER; order++) 2248 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2249 printk("= %lukB\n", K(total)); 2250 } 2251 2252 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2253 2254 show_swap_cache_info(); 2255} 2256 2257static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2258{ 2259 zoneref->zone = zone; 2260 zoneref->zone_idx = zone_idx(zone); 2261} 2262 2263/* 2264 * Builds allocation fallback zone lists. 2265 * 2266 * Add all populated zones of a node to the zonelist. 2267 */ 2268static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2269 int nr_zones, enum zone_type zone_type) 2270{ 2271 struct zone *zone; 2272 2273 BUG_ON(zone_type >= MAX_NR_ZONES); 2274 zone_type++; 2275 2276 do { 2277 zone_type--; 2278 zone = pgdat->node_zones + zone_type; 2279 if (populated_zone(zone)) { 2280 zoneref_set_zone(zone, 2281 &zonelist->_zonerefs[nr_zones++]); 2282 check_highest_zone(zone_type); 2283 } 2284 2285 } while (zone_type); 2286 return nr_zones; 2287} 2288 2289 2290/* 2291 * zonelist_order: 2292 * 0 = automatic detection of better ordering. 2293 * 1 = order by ([node] distance, -zonetype) 2294 * 2 = order by (-zonetype, [node] distance) 2295 * 2296 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2297 * the same zonelist. So only NUMA can configure this param. 2298 */ 2299#define ZONELIST_ORDER_DEFAULT 0 2300#define ZONELIST_ORDER_NODE 1 2301#define ZONELIST_ORDER_ZONE 2 2302 2303/* zonelist order in the kernel. 2304 * set_zonelist_order() will set this to NODE or ZONE. 2305 */ 2306static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2307static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2308 2309 2310#ifdef CONFIG_NUMA 2311/* The value user specified ....changed by config */ 2312static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2313/* string for sysctl */ 2314#define NUMA_ZONELIST_ORDER_LEN 16 2315char numa_zonelist_order[16] = "default"; 2316 2317/* 2318 * interface for configure zonelist ordering. 2319 * command line option "numa_zonelist_order" 2320 * = "[dD]efault - default, automatic configuration. 2321 * = "[nN]ode - order by node locality, then by zone within node 2322 * = "[zZ]one - order by zone, then by locality within zone 2323 */ 2324 2325static int __parse_numa_zonelist_order(char *s) 2326{ 2327 if (*s == 'd' || *s == 'D') { 2328 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2329 } else if (*s == 'n' || *s == 'N') { 2330 user_zonelist_order = ZONELIST_ORDER_NODE; 2331 } else if (*s == 'z' || *s == 'Z') { 2332 user_zonelist_order = ZONELIST_ORDER_ZONE; 2333 } else { 2334 printk(KERN_WARNING 2335 "Ignoring invalid numa_zonelist_order value: " 2336 "%s\n", s); 2337 return -EINVAL; 2338 } 2339 return 0; 2340} 2341 2342static __init int setup_numa_zonelist_order(char *s) 2343{ 2344 if (s) 2345 return __parse_numa_zonelist_order(s); 2346 return 0; 2347} 2348early_param("numa_zonelist_order", setup_numa_zonelist_order); 2349 2350/* 2351 * sysctl handler for numa_zonelist_order 2352 */ 2353int numa_zonelist_order_handler(ctl_table *table, int write, 2354 struct file *file, void __user *buffer, size_t *length, 2355 loff_t *ppos) 2356{ 2357 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2358 int ret; 2359 2360 if (write) 2361 strncpy(saved_string, (char*)table->data, 2362 NUMA_ZONELIST_ORDER_LEN); 2363 ret = proc_dostring(table, write, file, buffer, length, ppos); 2364 if (ret) 2365 return ret; 2366 if (write) { 2367 int oldval = user_zonelist_order; 2368 if (__parse_numa_zonelist_order((char*)table->data)) { 2369 /* 2370 * bogus value. restore saved string 2371 */ 2372 strncpy((char*)table->data, saved_string, 2373 NUMA_ZONELIST_ORDER_LEN); 2374 user_zonelist_order = oldval; 2375 } else if (oldval != user_zonelist_order) 2376 build_all_zonelists(); 2377 } 2378 return 0; 2379} 2380 2381 2382#define MAX_NODE_LOAD (nr_online_nodes) 2383static int node_load[MAX_NUMNODES]; 2384 2385/** 2386 * find_next_best_node - find the next node that should appear in a given node's fallback list 2387 * @node: node whose fallback list we're appending 2388 * @used_node_mask: nodemask_t of already used nodes 2389 * 2390 * We use a number of factors to determine which is the next node that should 2391 * appear on a given node's fallback list. The node should not have appeared 2392 * already in @node's fallback list, and it should be the next closest node 2393 * according to the distance array (which contains arbitrary distance values 2394 * from each node to each node in the system), and should also prefer nodes 2395 * with no CPUs, since presumably they'll have very little allocation pressure 2396 * on them otherwise. 2397 * It returns -1 if no node is found. 2398 */ 2399static int find_next_best_node(int node, nodemask_t *used_node_mask) 2400{ 2401 int n, val; 2402 int min_val = INT_MAX; 2403 int best_node = -1; 2404 const struct cpumask *tmp = cpumask_of_node(0); 2405 2406 /* Use the local node if we haven't already */ 2407 if (!node_isset(node, *used_node_mask)) { 2408 node_set(node, *used_node_mask); 2409 return node; 2410 } 2411 2412 for_each_node_state(n, N_HIGH_MEMORY) { 2413 2414 /* Don't want a node to appear more than once */ 2415 if (node_isset(n, *used_node_mask)) 2416 continue; 2417 2418 /* Use the distance array to find the distance */ 2419 val = node_distance(node, n); 2420 2421 /* Penalize nodes under us ("prefer the next node") */ 2422 val += (n < node); 2423 2424 /* Give preference to headless and unused nodes */ 2425 tmp = cpumask_of_node(n); 2426 if (!cpumask_empty(tmp)) 2427 val += PENALTY_FOR_NODE_WITH_CPUS; 2428 2429 /* Slight preference for less loaded node */ 2430 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2431 val += node_load[n]; 2432 2433 if (val < min_val) { 2434 min_val = val; 2435 best_node = n; 2436 } 2437 } 2438 2439 if (best_node >= 0) 2440 node_set(best_node, *used_node_mask); 2441 2442 return best_node; 2443} 2444 2445 2446/* 2447 * Build zonelists ordered by node and zones within node. 2448 * This results in maximum locality--normal zone overflows into local 2449 * DMA zone, if any--but risks exhausting DMA zone. 2450 */ 2451static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2452{ 2453 int j; 2454 struct zonelist *zonelist; 2455 2456 zonelist = &pgdat->node_zonelists[0]; 2457 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2458 ; 2459 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2460 MAX_NR_ZONES - 1); 2461 zonelist->_zonerefs[j].zone = NULL; 2462 zonelist->_zonerefs[j].zone_idx = 0; 2463} 2464 2465/* 2466 * Build gfp_thisnode zonelists 2467 */ 2468static void build_thisnode_zonelists(pg_data_t *pgdat) 2469{ 2470 int j; 2471 struct zonelist *zonelist; 2472 2473 zonelist = &pgdat->node_zonelists[1]; 2474 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2475 zonelist->_zonerefs[j].zone = NULL; 2476 zonelist->_zonerefs[j].zone_idx = 0; 2477} 2478 2479/* 2480 * Build zonelists ordered by zone and nodes within zones. 2481 * This results in conserving DMA zone[s] until all Normal memory is 2482 * exhausted, but results in overflowing to remote node while memory 2483 * may still exist in local DMA zone. 2484 */ 2485static int node_order[MAX_NUMNODES]; 2486 2487static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2488{ 2489 int pos, j, node; 2490 int zone_type; /* needs to be signed */ 2491 struct zone *z; 2492 struct zonelist *zonelist; 2493 2494 zonelist = &pgdat->node_zonelists[0]; 2495 pos = 0; 2496 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2497 for (j = 0; j < nr_nodes; j++) { 2498 node = node_order[j]; 2499 z = &NODE_DATA(node)->node_zones[zone_type]; 2500 if (populated_zone(z)) { 2501 zoneref_set_zone(z, 2502 &zonelist->_zonerefs[pos++]); 2503 check_highest_zone(zone_type); 2504 } 2505 } 2506 } 2507 zonelist->_zonerefs[pos].zone = NULL; 2508 zonelist->_zonerefs[pos].zone_idx = 0; 2509} 2510 2511static int default_zonelist_order(void) 2512{ 2513 int nid, zone_type; 2514 unsigned long low_kmem_size,total_size; 2515 struct zone *z; 2516 int average_size; 2517 /* 2518 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2519 * If they are really small and used heavily, the system can fall 2520 * into OOM very easily. 2521 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2522 */ 2523 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2524 low_kmem_size = 0; 2525 total_size = 0; 2526 for_each_online_node(nid) { 2527 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2528 z = &NODE_DATA(nid)->node_zones[zone_type]; 2529 if (populated_zone(z)) { 2530 if (zone_type < ZONE_NORMAL) 2531 low_kmem_size += z->present_pages; 2532 total_size += z->present_pages; 2533 } 2534 } 2535 } 2536 if (!low_kmem_size || /* there are no DMA area. */ 2537 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2538 return ZONELIST_ORDER_NODE; 2539 /* 2540 * look into each node's config. 2541 * If there is a node whose DMA/DMA32 memory is very big area on 2542 * local memory, NODE_ORDER may be suitable. 2543 */ 2544 average_size = total_size / 2545 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2546 for_each_online_node(nid) { 2547 low_kmem_size = 0; 2548 total_size = 0; 2549 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2550 z = &NODE_DATA(nid)->node_zones[zone_type]; 2551 if (populated_zone(z)) { 2552 if (zone_type < ZONE_NORMAL) 2553 low_kmem_size += z->present_pages; 2554 total_size += z->present_pages; 2555 } 2556 } 2557 if (low_kmem_size && 2558 total_size > average_size && /* ignore small node */ 2559 low_kmem_size > total_size * 70/100) 2560 return ZONELIST_ORDER_NODE; 2561 } 2562 return ZONELIST_ORDER_ZONE; 2563} 2564 2565static void set_zonelist_order(void) 2566{ 2567 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2568 current_zonelist_order = default_zonelist_order(); 2569 else 2570 current_zonelist_order = user_zonelist_order; 2571} 2572 2573static void build_zonelists(pg_data_t *pgdat) 2574{ 2575 int j, node, load; 2576 enum zone_type i; 2577 nodemask_t used_mask; 2578 int local_node, prev_node; 2579 struct zonelist *zonelist; 2580 int order = current_zonelist_order; 2581 2582 /* initialize zonelists */ 2583 for (i = 0; i < MAX_ZONELISTS; i++) { 2584 zonelist = pgdat->node_zonelists + i; 2585 zonelist->_zonerefs[0].zone = NULL; 2586 zonelist->_zonerefs[0].zone_idx = 0; 2587 } 2588 2589 /* NUMA-aware ordering of nodes */ 2590 local_node = pgdat->node_id; 2591 load = nr_online_nodes; 2592 prev_node = local_node; 2593 nodes_clear(used_mask); 2594 2595 memset(node_order, 0, sizeof(node_order)); 2596 j = 0; 2597 2598 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2599 int distance = node_distance(local_node, node); 2600 2601 /* 2602 * If another node is sufficiently far away then it is better 2603 * to reclaim pages in a zone before going off node. 2604 */ 2605 if (distance > RECLAIM_DISTANCE) 2606 zone_reclaim_mode = 1; 2607 2608 /* 2609 * We don't want to pressure a particular node. 2610 * So adding penalty to the first node in same 2611 * distance group to make it round-robin. 2612 */ 2613 if (distance != node_distance(local_node, prev_node)) 2614 node_load[node] = load; 2615 2616 prev_node = node; 2617 load--; 2618 if (order == ZONELIST_ORDER_NODE) 2619 build_zonelists_in_node_order(pgdat, node); 2620 else 2621 node_order[j++] = node; /* remember order */ 2622 } 2623 2624 if (order == ZONELIST_ORDER_ZONE) { 2625 /* calculate node order -- i.e., DMA last! */ 2626 build_zonelists_in_zone_order(pgdat, j); 2627 } 2628 2629 build_thisnode_zonelists(pgdat); 2630} 2631 2632/* Construct the zonelist performance cache - see further mmzone.h */ 2633static void build_zonelist_cache(pg_data_t *pgdat) 2634{ 2635 struct zonelist *zonelist; 2636 struct zonelist_cache *zlc; 2637 struct zoneref *z; 2638 2639 zonelist = &pgdat->node_zonelists[0]; 2640 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2641 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2642 for (z = zonelist->_zonerefs; z->zone; z++) 2643 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2644} 2645 2646 2647#else /* CONFIG_NUMA */ 2648 2649static void set_zonelist_order(void) 2650{ 2651 current_zonelist_order = ZONELIST_ORDER_ZONE; 2652} 2653 2654static void build_zonelists(pg_data_t *pgdat) 2655{ 2656 int node, local_node; 2657 enum zone_type j; 2658 struct zonelist *zonelist; 2659 2660 local_node = pgdat->node_id; 2661 2662 zonelist = &pgdat->node_zonelists[0]; 2663 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2664 2665 /* 2666 * Now we build the zonelist so that it contains the zones 2667 * of all the other nodes. 2668 * We don't want to pressure a particular node, so when 2669 * building the zones for node N, we make sure that the 2670 * zones coming right after the local ones are those from 2671 * node N+1 (modulo N) 2672 */ 2673 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2674 if (!node_online(node)) 2675 continue; 2676 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2677 MAX_NR_ZONES - 1); 2678 } 2679 for (node = 0; node < local_node; node++) { 2680 if (!node_online(node)) 2681 continue; 2682 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2683 MAX_NR_ZONES - 1); 2684 } 2685 2686 zonelist->_zonerefs[j].zone = NULL; 2687 zonelist->_zonerefs[j].zone_idx = 0; 2688} 2689 2690/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2691static void build_zonelist_cache(pg_data_t *pgdat) 2692{ 2693 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2694} 2695 2696#endif /* CONFIG_NUMA */ 2697 2698/* return values int ....just for stop_machine() */ 2699static int __build_all_zonelists(void *dummy) 2700{ 2701 int nid; 2702 2703#ifdef CONFIG_NUMA 2704 memset(node_load, 0, sizeof(node_load)); 2705#endif 2706 for_each_online_node(nid) { 2707 pg_data_t *pgdat = NODE_DATA(nid); 2708 2709 build_zonelists(pgdat); 2710 build_zonelist_cache(pgdat); 2711 } 2712 return 0; 2713} 2714 2715void build_all_zonelists(void) 2716{ 2717 set_zonelist_order(); 2718 2719 if (system_state == SYSTEM_BOOTING) { 2720 __build_all_zonelists(NULL); 2721 mminit_verify_zonelist(); 2722 cpuset_init_current_mems_allowed(); 2723 } else { 2724 /* we have to stop all cpus to guarantee there is no user 2725 of zonelist */ 2726 stop_machine(__build_all_zonelists, NULL, NULL); 2727 /* cpuset refresh routine should be here */ 2728 } 2729 vm_total_pages = nr_free_pagecache_pages(); 2730 /* 2731 * Disable grouping by mobility if the number of pages in the 2732 * system is too low to allow the mechanism to work. It would be 2733 * more accurate, but expensive to check per-zone. This check is 2734 * made on memory-hotadd so a system can start with mobility 2735 * disabled and enable it later 2736 */ 2737 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2738 page_group_by_mobility_disabled = 1; 2739 else 2740 page_group_by_mobility_disabled = 0; 2741 2742 printk("Built %i zonelists in %s order, mobility grouping %s. " 2743 "Total pages: %ld\n", 2744 nr_online_nodes, 2745 zonelist_order_name[current_zonelist_order], 2746 page_group_by_mobility_disabled ? "off" : "on", 2747 vm_total_pages); 2748#ifdef CONFIG_NUMA 2749 printk("Policy zone: %s\n", zone_names[policy_zone]); 2750#endif 2751} 2752 2753/* 2754 * Helper functions to size the waitqueue hash table. 2755 * Essentially these want to choose hash table sizes sufficiently 2756 * large so that collisions trying to wait on pages are rare. 2757 * But in fact, the number of active page waitqueues on typical 2758 * systems is ridiculously low, less than 200. So this is even 2759 * conservative, even though it seems large. 2760 * 2761 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2762 * waitqueues, i.e. the size of the waitq table given the number of pages. 2763 */ 2764#define PAGES_PER_WAITQUEUE 256 2765 2766#ifndef CONFIG_MEMORY_HOTPLUG 2767static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2768{ 2769 unsigned long size = 1; 2770 2771 pages /= PAGES_PER_WAITQUEUE; 2772 2773 while (size < pages) 2774 size <<= 1; 2775 2776 /* 2777 * Once we have dozens or even hundreds of threads sleeping 2778 * on IO we've got bigger problems than wait queue collision. 2779 * Limit the size of the wait table to a reasonable size. 2780 */ 2781 size = min(size, 4096UL); 2782 2783 return max(size, 4UL); 2784} 2785#else 2786/* 2787 * A zone's size might be changed by hot-add, so it is not possible to determine 2788 * a suitable size for its wait_table. So we use the maximum size now. 2789 * 2790 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2791 * 2792 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2793 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2794 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2795 * 2796 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2797 * or more by the traditional way. (See above). It equals: 2798 * 2799 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2800 * ia64(16K page size) : = ( 8G + 4M)byte. 2801 * powerpc (64K page size) : = (32G +16M)byte. 2802 */ 2803static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2804{ 2805 return 4096UL; 2806} 2807#endif 2808 2809/* 2810 * This is an integer logarithm so that shifts can be used later 2811 * to extract the more random high bits from the multiplicative 2812 * hash function before the remainder is taken. 2813 */ 2814static inline unsigned long wait_table_bits(unsigned long size) 2815{ 2816 return ffz(~size); 2817} 2818 2819#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2820 2821/* 2822 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2823 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2824 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2825 * higher will lead to a bigger reserve which will get freed as contiguous 2826 * blocks as reclaim kicks in 2827 */ 2828static void setup_zone_migrate_reserve(struct zone *zone) 2829{ 2830 unsigned long start_pfn, pfn, end_pfn; 2831 struct page *page; 2832 unsigned long reserve, block_migratetype; 2833 2834 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2835 start_pfn = zone->zone_start_pfn; 2836 end_pfn = start_pfn + zone->spanned_pages; 2837 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2838 pageblock_order; 2839 2840 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2841 if (!pfn_valid(pfn)) 2842 continue; 2843 page = pfn_to_page(pfn); 2844 2845 /* Watch out for overlapping nodes */ 2846 if (page_to_nid(page) != zone_to_nid(zone)) 2847 continue; 2848 2849 /* Blocks with reserved pages will never free, skip them. */ 2850 if (PageReserved(page)) 2851 continue; 2852 2853 block_migratetype = get_pageblock_migratetype(page); 2854 2855 /* If this block is reserved, account for it */ 2856 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2857 reserve--; 2858 continue; 2859 } 2860 2861 /* Suitable for reserving if this block is movable */ 2862 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2863 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2864 move_freepages_block(zone, page, MIGRATE_RESERVE); 2865 reserve--; 2866 continue; 2867 } 2868 2869 /* 2870 * If the reserve is met and this is a previous reserved block, 2871 * take it back 2872 */ 2873 if (block_migratetype == MIGRATE_RESERVE) { 2874 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2875 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2876 } 2877 } 2878} 2879 2880/* 2881 * Initially all pages are reserved - free ones are freed 2882 * up by free_all_bootmem() once the early boot process is 2883 * done. Non-atomic initialization, single-pass. 2884 */ 2885void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2886 unsigned long start_pfn, enum memmap_context context) 2887{ 2888 struct page *page; 2889 unsigned long end_pfn = start_pfn + size; 2890 unsigned long pfn; 2891 struct zone *z; 2892 2893 if (highest_memmap_pfn < end_pfn - 1) 2894 highest_memmap_pfn = end_pfn - 1; 2895 2896 z = &NODE_DATA(nid)->node_zones[zone]; 2897 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2898 /* 2899 * There can be holes in boot-time mem_map[]s 2900 * handed to this function. They do not 2901 * exist on hotplugged memory. 2902 */ 2903 if (context == MEMMAP_EARLY) { 2904 if (!early_pfn_valid(pfn)) 2905 continue; 2906 if (!early_pfn_in_nid(pfn, nid)) 2907 continue; 2908 } 2909 page = pfn_to_page(pfn); 2910 set_page_links(page, zone, nid, pfn); 2911 mminit_verify_page_links(page, zone, nid, pfn); 2912 init_page_count(page); 2913 reset_page_mapcount(page); 2914 SetPageReserved(page); 2915 /* 2916 * Mark the block movable so that blocks are reserved for 2917 * movable at startup. This will force kernel allocations 2918 * to reserve their blocks rather than leaking throughout 2919 * the address space during boot when many long-lived 2920 * kernel allocations are made. Later some blocks near 2921 * the start are marked MIGRATE_RESERVE by 2922 * setup_zone_migrate_reserve() 2923 * 2924 * bitmap is created for zone's valid pfn range. but memmap 2925 * can be created for invalid pages (for alignment) 2926 * check here not to call set_pageblock_migratetype() against 2927 * pfn out of zone. 2928 */ 2929 if ((z->zone_start_pfn <= pfn) 2930 && (pfn < z->zone_start_pfn + z->spanned_pages) 2931 && !(pfn & (pageblock_nr_pages - 1))) 2932 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2933 2934 INIT_LIST_HEAD(&page->lru); 2935#ifdef WANT_PAGE_VIRTUAL 2936 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2937 if (!is_highmem_idx(zone)) 2938 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2939#endif 2940 } 2941} 2942 2943static void __meminit zone_init_free_lists(struct zone *zone) 2944{ 2945 int order, t; 2946 for_each_migratetype_order(order, t) { 2947 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 2948 zone->free_area[order].nr_free = 0; 2949 } 2950} 2951 2952#ifndef __HAVE_ARCH_MEMMAP_INIT 2953#define memmap_init(size, nid, zone, start_pfn) \ 2954 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 2955#endif 2956 2957static int zone_batchsize(struct zone *zone) 2958{ 2959#ifdef CONFIG_MMU 2960 int batch; 2961 2962 /* 2963 * The per-cpu-pages pools are set to around 1000th of the 2964 * size of the zone. But no more than 1/2 of a meg. 2965 * 2966 * OK, so we don't know how big the cache is. So guess. 2967 */ 2968 batch = zone->present_pages / 1024; 2969 if (batch * PAGE_SIZE > 512 * 1024) 2970 batch = (512 * 1024) / PAGE_SIZE; 2971 batch /= 4; /* We effectively *= 4 below */ 2972 if (batch < 1) 2973 batch = 1; 2974 2975 /* 2976 * Clamp the batch to a 2^n - 1 value. Having a power 2977 * of 2 value was found to be more likely to have 2978 * suboptimal cache aliasing properties in some cases. 2979 * 2980 * For example if 2 tasks are alternately allocating 2981 * batches of pages, one task can end up with a lot 2982 * of pages of one half of the possible page colors 2983 * and the other with pages of the other colors. 2984 */ 2985 batch = rounddown_pow_of_two(batch + batch/2) - 1; 2986 2987 return batch; 2988 2989#else 2990 /* The deferral and batching of frees should be suppressed under NOMMU 2991 * conditions. 2992 * 2993 * The problem is that NOMMU needs to be able to allocate large chunks 2994 * of contiguous memory as there's no hardware page translation to 2995 * assemble apparent contiguous memory from discontiguous pages. 2996 * 2997 * Queueing large contiguous runs of pages for batching, however, 2998 * causes the pages to actually be freed in smaller chunks. As there 2999 * can be a significant delay between the individual batches being 3000 * recycled, this leads to the once large chunks of space being 3001 * fragmented and becoming unavailable for high-order allocations. 3002 */ 3003 return 0; 3004#endif 3005} 3006 3007static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3008{ 3009 struct per_cpu_pages *pcp; 3010 3011 memset(p, 0, sizeof(*p)); 3012 3013 pcp = &p->pcp; 3014 pcp->count = 0; 3015 pcp->high = 6 * batch; 3016 pcp->batch = max(1UL, 1 * batch); 3017 INIT_LIST_HEAD(&pcp->list); 3018} 3019 3020/* 3021 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3022 * to the value high for the pageset p. 3023 */ 3024 3025static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3026 unsigned long high) 3027{ 3028 struct per_cpu_pages *pcp; 3029 3030 pcp = &p->pcp; 3031 pcp->high = high; 3032 pcp->batch = max(1UL, high/4); 3033 if ((high/4) > (PAGE_SHIFT * 8)) 3034 pcp->batch = PAGE_SHIFT * 8; 3035} 3036 3037 3038#ifdef CONFIG_NUMA 3039/* 3040 * Boot pageset table. One per cpu which is going to be used for all 3041 * zones and all nodes. The parameters will be set in such a way 3042 * that an item put on a list will immediately be handed over to 3043 * the buddy list. This is safe since pageset manipulation is done 3044 * with interrupts disabled. 3045 * 3046 * Some NUMA counter updates may also be caught by the boot pagesets. 3047 * 3048 * The boot_pagesets must be kept even after bootup is complete for 3049 * unused processors and/or zones. They do play a role for bootstrapping 3050 * hotplugged processors. 3051 * 3052 * zoneinfo_show() and maybe other functions do 3053 * not check if the processor is online before following the pageset pointer. 3054 * Other parts of the kernel may not check if the zone is available. 3055 */ 3056static struct per_cpu_pageset boot_pageset[NR_CPUS]; 3057 3058/* 3059 * Dynamically allocate memory for the 3060 * per cpu pageset array in struct zone. 3061 */ 3062static int __cpuinit process_zones(int cpu) 3063{ 3064 struct zone *zone, *dzone; 3065 int node = cpu_to_node(cpu); 3066 3067 node_set_state(node, N_CPU); /* this node has a cpu */ 3068 3069 for_each_populated_zone(zone) { 3070 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 3071 GFP_KERNEL, node); 3072 if (!zone_pcp(zone, cpu)) 3073 goto bad; 3074 3075 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 3076 3077 if (percpu_pagelist_fraction) 3078 setup_pagelist_highmark(zone_pcp(zone, cpu), 3079 (zone->present_pages / percpu_pagelist_fraction)); 3080 } 3081 3082 return 0; 3083bad: 3084 for_each_zone(dzone) { 3085 if (!populated_zone(dzone)) 3086 continue; 3087 if (dzone == zone) 3088 break; 3089 kfree(zone_pcp(dzone, cpu)); 3090 zone_pcp(dzone, cpu) = &boot_pageset[cpu]; 3091 } 3092 return -ENOMEM; 3093} 3094 3095static inline void free_zone_pagesets(int cpu) 3096{ 3097 struct zone *zone; 3098 3099 for_each_zone(zone) { 3100 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 3101 3102 /* Free per_cpu_pageset if it is slab allocated */ 3103 if (pset != &boot_pageset[cpu]) 3104 kfree(pset); 3105 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3106 } 3107} 3108 3109static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 3110 unsigned long action, 3111 void *hcpu) 3112{ 3113 int cpu = (long)hcpu; 3114 int ret = NOTIFY_OK; 3115 3116 switch (action) { 3117 case CPU_UP_PREPARE: 3118 case CPU_UP_PREPARE_FROZEN: 3119 if (process_zones(cpu)) 3120 ret = NOTIFY_BAD; 3121 break; 3122 case CPU_UP_CANCELED: 3123 case CPU_UP_CANCELED_FROZEN: 3124 case CPU_DEAD: 3125 case CPU_DEAD_FROZEN: 3126 free_zone_pagesets(cpu); 3127 break; 3128 default: 3129 break; 3130 } 3131 return ret; 3132} 3133 3134static struct notifier_block __cpuinitdata pageset_notifier = 3135 { &pageset_cpuup_callback, NULL, 0 }; 3136 3137void __init setup_per_cpu_pageset(void) 3138{ 3139 int err; 3140 3141 /* Initialize per_cpu_pageset for cpu 0. 3142 * A cpuup callback will do this for every cpu 3143 * as it comes online 3144 */ 3145 err = process_zones(smp_processor_id()); 3146 BUG_ON(err); 3147 register_cpu_notifier(&pageset_notifier); 3148} 3149 3150#endif 3151 3152static noinline __init_refok 3153int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3154{ 3155 int i; 3156 struct pglist_data *pgdat = zone->zone_pgdat; 3157 size_t alloc_size; 3158 3159 /* 3160 * The per-page waitqueue mechanism uses hashed waitqueues 3161 * per zone. 3162 */ 3163 zone->wait_table_hash_nr_entries = 3164 wait_table_hash_nr_entries(zone_size_pages); 3165 zone->wait_table_bits = 3166 wait_table_bits(zone->wait_table_hash_nr_entries); 3167 alloc_size = zone->wait_table_hash_nr_entries 3168 * sizeof(wait_queue_head_t); 3169 3170 if (!slab_is_available()) { 3171 zone->wait_table = (wait_queue_head_t *) 3172 alloc_bootmem_node(pgdat, alloc_size); 3173 } else { 3174 /* 3175 * This case means that a zone whose size was 0 gets new memory 3176 * via memory hot-add. 3177 * But it may be the case that a new node was hot-added. In 3178 * this case vmalloc() will not be able to use this new node's 3179 * memory - this wait_table must be initialized to use this new 3180 * node itself as well. 3181 * To use this new node's memory, further consideration will be 3182 * necessary. 3183 */ 3184 zone->wait_table = vmalloc(alloc_size); 3185 } 3186 if (!zone->wait_table) 3187 return -ENOMEM; 3188 3189 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3190 init_waitqueue_head(zone->wait_table + i); 3191 3192 return 0; 3193} 3194 3195static int __zone_pcp_update(void *data) 3196{ 3197 struct zone *zone = data; 3198 int cpu; 3199 unsigned long batch = zone_batchsize(zone), flags; 3200 3201 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3202 struct per_cpu_pageset *pset; 3203 struct per_cpu_pages *pcp; 3204 3205 pset = zone_pcp(zone, cpu); 3206 pcp = &pset->pcp; 3207 3208 local_irq_save(flags); 3209 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 3210 setup_pageset(pset, batch); 3211 local_irq_restore(flags); 3212 } 3213 return 0; 3214} 3215 3216void zone_pcp_update(struct zone *zone) 3217{ 3218 stop_machine(__zone_pcp_update, zone, NULL); 3219} 3220 3221static __meminit void zone_pcp_init(struct zone *zone) 3222{ 3223 int cpu; 3224 unsigned long batch = zone_batchsize(zone); 3225 3226 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3227#ifdef CONFIG_NUMA 3228 /* Early boot. Slab allocator not functional yet */ 3229 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3230 setup_pageset(&boot_pageset[cpu],0); 3231#else 3232 setup_pageset(zone_pcp(zone,cpu), batch); 3233#endif 3234 } 3235 if (zone->present_pages) 3236 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 3237 zone->name, zone->present_pages, batch); 3238} 3239 3240__meminit int init_currently_empty_zone(struct zone *zone, 3241 unsigned long zone_start_pfn, 3242 unsigned long size, 3243 enum memmap_context context) 3244{ 3245 struct pglist_data *pgdat = zone->zone_pgdat; 3246 int ret; 3247 ret = zone_wait_table_init(zone, size); 3248 if (ret) 3249 return ret; 3250 pgdat->nr_zones = zone_idx(zone) + 1; 3251 3252 zone->zone_start_pfn = zone_start_pfn; 3253 3254 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3255 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3256 pgdat->node_id, 3257 (unsigned long)zone_idx(zone), 3258 zone_start_pfn, (zone_start_pfn + size)); 3259 3260 zone_init_free_lists(zone); 3261 3262 return 0; 3263} 3264 3265#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3266/* 3267 * Basic iterator support. Return the first range of PFNs for a node 3268 * Note: nid == MAX_NUMNODES returns first region regardless of node 3269 */ 3270static int __meminit first_active_region_index_in_nid(int nid) 3271{ 3272 int i; 3273 3274 for (i = 0; i < nr_nodemap_entries; i++) 3275 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3276 return i; 3277 3278 return -1; 3279} 3280 3281/* 3282 * Basic iterator support. Return the next active range of PFNs for a node 3283 * Note: nid == MAX_NUMNODES returns next region regardless of node 3284 */ 3285static int __meminit next_active_region_index_in_nid(int index, int nid) 3286{ 3287 for (index = index + 1; index < nr_nodemap_entries; index++) 3288 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3289 return index; 3290 3291 return -1; 3292} 3293 3294#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3295/* 3296 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3297 * Architectures may implement their own version but if add_active_range() 3298 * was used and there are no special requirements, this is a convenient 3299 * alternative 3300 */ 3301int __meminit __early_pfn_to_nid(unsigned long pfn) 3302{ 3303 int i; 3304 3305 for (i = 0; i < nr_nodemap_entries; i++) { 3306 unsigned long start_pfn = early_node_map[i].start_pfn; 3307 unsigned long end_pfn = early_node_map[i].end_pfn; 3308 3309 if (start_pfn <= pfn && pfn < end_pfn) 3310 return early_node_map[i].nid; 3311 } 3312 /* This is a memory hole */ 3313 return -1; 3314} 3315#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3316 3317int __meminit early_pfn_to_nid(unsigned long pfn) 3318{ 3319 int nid; 3320 3321 nid = __early_pfn_to_nid(pfn); 3322 if (nid >= 0) 3323 return nid; 3324 /* just returns 0 */ 3325 return 0; 3326} 3327 3328#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3329bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3330{ 3331 int nid; 3332 3333 nid = __early_pfn_to_nid(pfn); 3334 if (nid >= 0 && nid != node) 3335 return false; 3336 return true; 3337} 3338#endif 3339 3340/* Basic iterator support to walk early_node_map[] */ 3341#define for_each_active_range_index_in_nid(i, nid) \ 3342 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3343 i = next_active_region_index_in_nid(i, nid)) 3344 3345/** 3346 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3347 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3348 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3349 * 3350 * If an architecture guarantees that all ranges registered with 3351 * add_active_ranges() contain no holes and may be freed, this 3352 * this function may be used instead of calling free_bootmem() manually. 3353 */ 3354void __init free_bootmem_with_active_regions(int nid, 3355 unsigned long max_low_pfn) 3356{ 3357 int i; 3358 3359 for_each_active_range_index_in_nid(i, nid) { 3360 unsigned long size_pages = 0; 3361 unsigned long end_pfn = early_node_map[i].end_pfn; 3362 3363 if (early_node_map[i].start_pfn >= max_low_pfn) 3364 continue; 3365 3366 if (end_pfn > max_low_pfn) 3367 end_pfn = max_low_pfn; 3368 3369 size_pages = end_pfn - early_node_map[i].start_pfn; 3370 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3371 PFN_PHYS(early_node_map[i].start_pfn), 3372 size_pages << PAGE_SHIFT); 3373 } 3374} 3375 3376void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3377{ 3378 int i; 3379 int ret; 3380 3381 for_each_active_range_index_in_nid(i, nid) { 3382 ret = work_fn(early_node_map[i].start_pfn, 3383 early_node_map[i].end_pfn, data); 3384 if (ret) 3385 break; 3386 } 3387} 3388/** 3389 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3390 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3391 * 3392 * If an architecture guarantees that all ranges registered with 3393 * add_active_ranges() contain no holes and may be freed, this 3394 * function may be used instead of calling memory_present() manually. 3395 */ 3396void __init sparse_memory_present_with_active_regions(int nid) 3397{ 3398 int i; 3399 3400 for_each_active_range_index_in_nid(i, nid) 3401 memory_present(early_node_map[i].nid, 3402 early_node_map[i].start_pfn, 3403 early_node_map[i].end_pfn); 3404} 3405 3406/** 3407 * get_pfn_range_for_nid - Return the start and end page frames for a node 3408 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3409 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3410 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3411 * 3412 * It returns the start and end page frame of a node based on information 3413 * provided by an arch calling add_active_range(). If called for a node 3414 * with no available memory, a warning is printed and the start and end 3415 * PFNs will be 0. 3416 */ 3417void __meminit get_pfn_range_for_nid(unsigned int nid, 3418 unsigned long *start_pfn, unsigned long *end_pfn) 3419{ 3420 int i; 3421 *start_pfn = -1UL; 3422 *end_pfn = 0; 3423 3424 for_each_active_range_index_in_nid(i, nid) { 3425 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3426 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3427 } 3428 3429 if (*start_pfn == -1UL) 3430 *start_pfn = 0; 3431} 3432 3433/* 3434 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3435 * assumption is made that zones within a node are ordered in monotonic 3436 * increasing memory addresses so that the "highest" populated zone is used 3437 */ 3438static void __init find_usable_zone_for_movable(void) 3439{ 3440 int zone_index; 3441 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3442 if (zone_index == ZONE_MOVABLE) 3443 continue; 3444 3445 if (arch_zone_highest_possible_pfn[zone_index] > 3446 arch_zone_lowest_possible_pfn[zone_index]) 3447 break; 3448 } 3449 3450 VM_BUG_ON(zone_index == -1); 3451 movable_zone = zone_index; 3452} 3453 3454/* 3455 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3456 * because it is sized independant of architecture. Unlike the other zones, 3457 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3458 * in each node depending on the size of each node and how evenly kernelcore 3459 * is distributed. This helper function adjusts the zone ranges 3460 * provided by the architecture for a given node by using the end of the 3461 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3462 * zones within a node are in order of monotonic increases memory addresses 3463 */ 3464static void __meminit adjust_zone_range_for_zone_movable(int nid, 3465 unsigned long zone_type, 3466 unsigned long node_start_pfn, 3467 unsigned long node_end_pfn, 3468 unsigned long *zone_start_pfn, 3469 unsigned long *zone_end_pfn) 3470{ 3471 /* Only adjust if ZONE_MOVABLE is on this node */ 3472 if (zone_movable_pfn[nid]) { 3473 /* Size ZONE_MOVABLE */ 3474 if (zone_type == ZONE_MOVABLE) { 3475 *zone_start_pfn = zone_movable_pfn[nid]; 3476 *zone_end_pfn = min(node_end_pfn, 3477 arch_zone_highest_possible_pfn[movable_zone]); 3478 3479 /* Adjust for ZONE_MOVABLE starting within this range */ 3480 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3481 *zone_end_pfn > zone_movable_pfn[nid]) { 3482 *zone_end_pfn = zone_movable_pfn[nid]; 3483 3484 /* Check if this whole range is within ZONE_MOVABLE */ 3485 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3486 *zone_start_pfn = *zone_end_pfn; 3487 } 3488} 3489 3490/* 3491 * Return the number of pages a zone spans in a node, including holes 3492 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3493 */ 3494static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3495 unsigned long zone_type, 3496 unsigned long *ignored) 3497{ 3498 unsigned long node_start_pfn, node_end_pfn; 3499 unsigned long zone_start_pfn, zone_end_pfn; 3500 3501 /* Get the start and end of the node and zone */ 3502 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3503 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3504 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3505 adjust_zone_range_for_zone_movable(nid, zone_type, 3506 node_start_pfn, node_end_pfn, 3507 &zone_start_pfn, &zone_end_pfn); 3508 3509 /* Check that this node has pages within the zone's required range */ 3510 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3511 return 0; 3512 3513 /* Move the zone boundaries inside the node if necessary */ 3514 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3515 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3516 3517 /* Return the spanned pages */ 3518 return zone_end_pfn - zone_start_pfn; 3519} 3520 3521/* 3522 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3523 * then all holes in the requested range will be accounted for. 3524 */ 3525static unsigned long __meminit __absent_pages_in_range(int nid, 3526 unsigned long range_start_pfn, 3527 unsigned long range_end_pfn) 3528{ 3529 int i = 0; 3530 unsigned long prev_end_pfn = 0, hole_pages = 0; 3531 unsigned long start_pfn; 3532 3533 /* Find the end_pfn of the first active range of pfns in the node */ 3534 i = first_active_region_index_in_nid(nid); 3535 if (i == -1) 3536 return 0; 3537 3538 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3539 3540 /* Account for ranges before physical memory on this node */ 3541 if (early_node_map[i].start_pfn > range_start_pfn) 3542 hole_pages = prev_end_pfn - range_start_pfn; 3543 3544 /* Find all holes for the zone within the node */ 3545 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3546 3547 /* No need to continue if prev_end_pfn is outside the zone */ 3548 if (prev_end_pfn >= range_end_pfn) 3549 break; 3550 3551 /* Make sure the end of the zone is not within the hole */ 3552 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3553 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3554 3555 /* Update the hole size cound and move on */ 3556 if (start_pfn > range_start_pfn) { 3557 BUG_ON(prev_end_pfn > start_pfn); 3558 hole_pages += start_pfn - prev_end_pfn; 3559 } 3560 prev_end_pfn = early_node_map[i].end_pfn; 3561 } 3562 3563 /* Account for ranges past physical memory on this node */ 3564 if (range_end_pfn > prev_end_pfn) 3565 hole_pages += range_end_pfn - 3566 max(range_start_pfn, prev_end_pfn); 3567 3568 return hole_pages; 3569} 3570 3571/** 3572 * absent_pages_in_range - Return number of page frames in holes within a range 3573 * @start_pfn: The start PFN to start searching for holes 3574 * @end_pfn: The end PFN to stop searching for holes 3575 * 3576 * It returns the number of pages frames in memory holes within a range. 3577 */ 3578unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3579 unsigned long end_pfn) 3580{ 3581 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3582} 3583 3584/* Return the number of page frames in holes in a zone on a node */ 3585static unsigned long __meminit zone_absent_pages_in_node(int nid, 3586 unsigned long zone_type, 3587 unsigned long *ignored) 3588{ 3589 unsigned long node_start_pfn, node_end_pfn; 3590 unsigned long zone_start_pfn, zone_end_pfn; 3591 3592 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3593 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3594 node_start_pfn); 3595 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3596 node_end_pfn); 3597 3598 adjust_zone_range_for_zone_movable(nid, zone_type, 3599 node_start_pfn, node_end_pfn, 3600 &zone_start_pfn, &zone_end_pfn); 3601 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3602} 3603 3604#else 3605static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3606 unsigned long zone_type, 3607 unsigned long *zones_size) 3608{ 3609 return zones_size[zone_type]; 3610} 3611 3612static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3613 unsigned long zone_type, 3614 unsigned long *zholes_size) 3615{ 3616 if (!zholes_size) 3617 return 0; 3618 3619 return zholes_size[zone_type]; 3620} 3621 3622#endif 3623 3624static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3625 unsigned long *zones_size, unsigned long *zholes_size) 3626{ 3627 unsigned long realtotalpages, totalpages = 0; 3628 enum zone_type i; 3629 3630 for (i = 0; i < MAX_NR_ZONES; i++) 3631 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3632 zones_size); 3633 pgdat->node_spanned_pages = totalpages; 3634 3635 realtotalpages = totalpages; 3636 for (i = 0; i < MAX_NR_ZONES; i++) 3637 realtotalpages -= 3638 zone_absent_pages_in_node(pgdat->node_id, i, 3639 zholes_size); 3640 pgdat->node_present_pages = realtotalpages; 3641 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3642 realtotalpages); 3643} 3644 3645#ifndef CONFIG_SPARSEMEM 3646/* 3647 * Calculate the size of the zone->blockflags rounded to an unsigned long 3648 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3649 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3650 * round what is now in bits to nearest long in bits, then return it in 3651 * bytes. 3652 */ 3653static unsigned long __init usemap_size(unsigned long zonesize) 3654{ 3655 unsigned long usemapsize; 3656 3657 usemapsize = roundup(zonesize, pageblock_nr_pages); 3658 usemapsize = usemapsize >> pageblock_order; 3659 usemapsize *= NR_PAGEBLOCK_BITS; 3660 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3661 3662 return usemapsize / 8; 3663} 3664 3665static void __init setup_usemap(struct pglist_data *pgdat, 3666 struct zone *zone, unsigned long zonesize) 3667{ 3668 unsigned long usemapsize = usemap_size(zonesize); 3669 zone->pageblock_flags = NULL; 3670 if (usemapsize) 3671 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3672} 3673#else 3674static void inline setup_usemap(struct pglist_data *pgdat, 3675 struct zone *zone, unsigned long zonesize) {} 3676#endif /* CONFIG_SPARSEMEM */ 3677 3678#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3679 3680/* Return a sensible default order for the pageblock size. */ 3681static inline int pageblock_default_order(void) 3682{ 3683 if (HPAGE_SHIFT > PAGE_SHIFT) 3684 return HUGETLB_PAGE_ORDER; 3685 3686 return MAX_ORDER-1; 3687} 3688 3689/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3690static inline void __init set_pageblock_order(unsigned int order) 3691{ 3692 /* Check that pageblock_nr_pages has not already been setup */ 3693 if (pageblock_order) 3694 return; 3695 3696 /* 3697 * Assume the largest contiguous order of interest is a huge page. 3698 * This value may be variable depending on boot parameters on IA64 3699 */ 3700 pageblock_order = order; 3701} 3702#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3703 3704/* 3705 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3706 * and pageblock_default_order() are unused as pageblock_order is set 3707 * at compile-time. See include/linux/pageblock-flags.h for the values of 3708 * pageblock_order based on the kernel config 3709 */ 3710static inline int pageblock_default_order(unsigned int order) 3711{ 3712 return MAX_ORDER-1; 3713} 3714#define set_pageblock_order(x) do {} while (0) 3715 3716#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3717 3718/* 3719 * Set up the zone data structures: 3720 * - mark all pages reserved 3721 * - mark all memory queues empty 3722 * - clear the memory bitmaps 3723 */ 3724static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3725 unsigned long *zones_size, unsigned long *zholes_size) 3726{ 3727 enum zone_type j; 3728 int nid = pgdat->node_id; 3729 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3730 int ret; 3731 3732 pgdat_resize_init(pgdat); 3733 pgdat->nr_zones = 0; 3734 init_waitqueue_head(&pgdat->kswapd_wait); 3735 pgdat->kswapd_max_order = 0; 3736 pgdat_page_cgroup_init(pgdat); 3737 3738 for (j = 0; j < MAX_NR_ZONES; j++) { 3739 struct zone *zone = pgdat->node_zones + j; 3740 unsigned long size, realsize, memmap_pages; 3741 enum lru_list l; 3742 3743 size = zone_spanned_pages_in_node(nid, j, zones_size); 3744 realsize = size - zone_absent_pages_in_node(nid, j, 3745 zholes_size); 3746 3747 /* 3748 * Adjust realsize so that it accounts for how much memory 3749 * is used by this zone for memmap. This affects the watermark 3750 * and per-cpu initialisations 3751 */ 3752 memmap_pages = 3753 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3754 if (realsize >= memmap_pages) { 3755 realsize -= memmap_pages; 3756 if (memmap_pages) 3757 printk(KERN_DEBUG 3758 " %s zone: %lu pages used for memmap\n", 3759 zone_names[j], memmap_pages); 3760 } else 3761 printk(KERN_WARNING 3762 " %s zone: %lu pages exceeds realsize %lu\n", 3763 zone_names[j], memmap_pages, realsize); 3764 3765 /* Account for reserved pages */ 3766 if (j == 0 && realsize > dma_reserve) { 3767 realsize -= dma_reserve; 3768 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3769 zone_names[0], dma_reserve); 3770 } 3771 3772 if (!is_highmem_idx(j)) 3773 nr_kernel_pages += realsize; 3774 nr_all_pages += realsize; 3775 3776 zone->spanned_pages = size; 3777 zone->present_pages = realsize; 3778#ifdef CONFIG_NUMA 3779 zone->node = nid; 3780 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3781 / 100; 3782 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3783#endif 3784 zone->name = zone_names[j]; 3785 spin_lock_init(&zone->lock); 3786 spin_lock_init(&zone->lru_lock); 3787 zone_seqlock_init(zone); 3788 zone->zone_pgdat = pgdat; 3789 3790 zone->prev_priority = DEF_PRIORITY; 3791 3792 zone_pcp_init(zone); 3793 for_each_lru(l) { 3794 INIT_LIST_HEAD(&zone->lru[l].list); 3795 zone->lru[l].nr_saved_scan = 0; 3796 } 3797 zone->reclaim_stat.recent_rotated[0] = 0; 3798 zone->reclaim_stat.recent_rotated[1] = 0; 3799 zone->reclaim_stat.recent_scanned[0] = 0; 3800 zone->reclaim_stat.recent_scanned[1] = 0; 3801 zap_zone_vm_stats(zone); 3802 zone->flags = 0; 3803 if (!size) 3804 continue; 3805 3806 set_pageblock_order(pageblock_default_order()); 3807 setup_usemap(pgdat, zone, size); 3808 ret = init_currently_empty_zone(zone, zone_start_pfn, 3809 size, MEMMAP_EARLY); 3810 BUG_ON(ret); 3811 memmap_init(size, nid, j, zone_start_pfn); 3812 zone_start_pfn += size; 3813 } 3814} 3815 3816static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3817{ 3818 /* Skip empty nodes */ 3819 if (!pgdat->node_spanned_pages) 3820 return; 3821 3822#ifdef CONFIG_FLAT_NODE_MEM_MAP 3823 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3824 if (!pgdat->node_mem_map) { 3825 unsigned long size, start, end; 3826 struct page *map; 3827 3828 /* 3829 * The zone's endpoints aren't required to be MAX_ORDER 3830 * aligned but the node_mem_map endpoints must be in order 3831 * for the buddy allocator to function correctly. 3832 */ 3833 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3834 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3835 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3836 size = (end - start) * sizeof(struct page); 3837 map = alloc_remap(pgdat->node_id, size); 3838 if (!map) 3839 map = alloc_bootmem_node(pgdat, size); 3840 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3841 } 3842#ifndef CONFIG_NEED_MULTIPLE_NODES 3843 /* 3844 * With no DISCONTIG, the global mem_map is just set as node 0's 3845 */ 3846 if (pgdat == NODE_DATA(0)) { 3847 mem_map = NODE_DATA(0)->node_mem_map; 3848#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3849 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3850 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3851#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3852 } 3853#endif 3854#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3855} 3856 3857void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3858 unsigned long node_start_pfn, unsigned long *zholes_size) 3859{ 3860 pg_data_t *pgdat = NODE_DATA(nid); 3861 3862 pgdat->node_id = nid; 3863 pgdat->node_start_pfn = node_start_pfn; 3864 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3865 3866 alloc_node_mem_map(pgdat); 3867#ifdef CONFIG_FLAT_NODE_MEM_MAP 3868 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3869 nid, (unsigned long)pgdat, 3870 (unsigned long)pgdat->node_mem_map); 3871#endif 3872 3873 free_area_init_core(pgdat, zones_size, zholes_size); 3874} 3875 3876#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3877 3878#if MAX_NUMNODES > 1 3879/* 3880 * Figure out the number of possible node ids. 3881 */ 3882static void __init setup_nr_node_ids(void) 3883{ 3884 unsigned int node; 3885 unsigned int highest = 0; 3886 3887 for_each_node_mask(node, node_possible_map) 3888 highest = node; 3889 nr_node_ids = highest + 1; 3890} 3891#else 3892static inline void setup_nr_node_ids(void) 3893{ 3894} 3895#endif 3896 3897/** 3898 * add_active_range - Register a range of PFNs backed by physical memory 3899 * @nid: The node ID the range resides on 3900 * @start_pfn: The start PFN of the available physical memory 3901 * @end_pfn: The end PFN of the available physical memory 3902 * 3903 * These ranges are stored in an early_node_map[] and later used by 3904 * free_area_init_nodes() to calculate zone sizes and holes. If the 3905 * range spans a memory hole, it is up to the architecture to ensure 3906 * the memory is not freed by the bootmem allocator. If possible 3907 * the range being registered will be merged with existing ranges. 3908 */ 3909void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3910 unsigned long end_pfn) 3911{ 3912 int i; 3913 3914 mminit_dprintk(MMINIT_TRACE, "memory_register", 3915 "Entering add_active_range(%d, %#lx, %#lx) " 3916 "%d entries of %d used\n", 3917 nid, start_pfn, end_pfn, 3918 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3919 3920 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 3921 3922 /* Merge with existing active regions if possible */ 3923 for (i = 0; i < nr_nodemap_entries; i++) { 3924 if (early_node_map[i].nid != nid) 3925 continue; 3926 3927 /* Skip if an existing region covers this new one */ 3928 if (start_pfn >= early_node_map[i].start_pfn && 3929 end_pfn <= early_node_map[i].end_pfn) 3930 return; 3931 3932 /* Merge forward if suitable */ 3933 if (start_pfn <= early_node_map[i].end_pfn && 3934 end_pfn > early_node_map[i].end_pfn) { 3935 early_node_map[i].end_pfn = end_pfn; 3936 return; 3937 } 3938 3939 /* Merge backward if suitable */ 3940 if (start_pfn < early_node_map[i].end_pfn && 3941 end_pfn >= early_node_map[i].start_pfn) { 3942 early_node_map[i].start_pfn = start_pfn; 3943 return; 3944 } 3945 } 3946 3947 /* Check that early_node_map is large enough */ 3948 if (i >= MAX_ACTIVE_REGIONS) { 3949 printk(KERN_CRIT "More than %d memory regions, truncating\n", 3950 MAX_ACTIVE_REGIONS); 3951 return; 3952 } 3953 3954 early_node_map[i].nid = nid; 3955 early_node_map[i].start_pfn = start_pfn; 3956 early_node_map[i].end_pfn = end_pfn; 3957 nr_nodemap_entries = i + 1; 3958} 3959 3960/** 3961 * remove_active_range - Shrink an existing registered range of PFNs 3962 * @nid: The node id the range is on that should be shrunk 3963 * @start_pfn: The new PFN of the range 3964 * @end_pfn: The new PFN of the range 3965 * 3966 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 3967 * The map is kept near the end physical page range that has already been 3968 * registered. This function allows an arch to shrink an existing registered 3969 * range. 3970 */ 3971void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 3972 unsigned long end_pfn) 3973{ 3974 int i, j; 3975 int removed = 0; 3976 3977 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 3978 nid, start_pfn, end_pfn); 3979 3980 /* Find the old active region end and shrink */ 3981 for_each_active_range_index_in_nid(i, nid) { 3982 if (early_node_map[i].start_pfn >= start_pfn && 3983 early_node_map[i].end_pfn <= end_pfn) { 3984 /* clear it */ 3985 early_node_map[i].start_pfn = 0; 3986 early_node_map[i].end_pfn = 0; 3987 removed = 1; 3988 continue; 3989 } 3990 if (early_node_map[i].start_pfn < start_pfn && 3991 early_node_map[i].end_pfn > start_pfn) { 3992 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 3993 early_node_map[i].end_pfn = start_pfn; 3994 if (temp_end_pfn > end_pfn) 3995 add_active_range(nid, end_pfn, temp_end_pfn); 3996 continue; 3997 } 3998 if (early_node_map[i].start_pfn >= start_pfn && 3999 early_node_map[i].end_pfn > end_pfn && 4000 early_node_map[i].start_pfn < end_pfn) { 4001 early_node_map[i].start_pfn = end_pfn; 4002 continue; 4003 } 4004 } 4005 4006 if (!removed) 4007 return; 4008 4009 /* remove the blank ones */ 4010 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4011 if (early_node_map[i].nid != nid) 4012 continue; 4013 if (early_node_map[i].end_pfn) 4014 continue; 4015 /* we found it, get rid of it */ 4016 for (j = i; j < nr_nodemap_entries - 1; j++) 4017 memcpy(&early_node_map[j], &early_node_map[j+1], 4018 sizeof(early_node_map[j])); 4019 j = nr_nodemap_entries - 1; 4020 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4021 nr_nodemap_entries--; 4022 } 4023} 4024 4025/** 4026 * remove_all_active_ranges - Remove all currently registered regions 4027 * 4028 * During discovery, it may be found that a table like SRAT is invalid 4029 * and an alternative discovery method must be used. This function removes 4030 * all currently registered regions. 4031 */ 4032void __init remove_all_active_ranges(void) 4033{ 4034 memset(early_node_map, 0, sizeof(early_node_map)); 4035 nr_nodemap_entries = 0; 4036} 4037 4038/* Compare two active node_active_regions */ 4039static int __init cmp_node_active_region(const void *a, const void *b) 4040{ 4041 struct node_active_region *arange = (struct node_active_region *)a; 4042 struct node_active_region *brange = (struct node_active_region *)b; 4043 4044 /* Done this way to avoid overflows */ 4045 if (arange->start_pfn > brange->start_pfn) 4046 return 1; 4047 if (arange->start_pfn < brange->start_pfn) 4048 return -1; 4049 4050 return 0; 4051} 4052 4053/* sort the node_map by start_pfn */ 4054static void __init sort_node_map(void) 4055{ 4056 sort(early_node_map, (size_t)nr_nodemap_entries, 4057 sizeof(struct node_active_region), 4058 cmp_node_active_region, NULL); 4059} 4060 4061/* Find the lowest pfn for a node */ 4062static unsigned long __init find_min_pfn_for_node(int nid) 4063{ 4064 int i; 4065 unsigned long min_pfn = ULONG_MAX; 4066 4067 /* Assuming a sorted map, the first range found has the starting pfn */ 4068 for_each_active_range_index_in_nid(i, nid) 4069 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4070 4071 if (min_pfn == ULONG_MAX) { 4072 printk(KERN_WARNING 4073 "Could not find start_pfn for node %d\n", nid); 4074 return 0; 4075 } 4076 4077 return min_pfn; 4078} 4079 4080/** 4081 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4082 * 4083 * It returns the minimum PFN based on information provided via 4084 * add_active_range(). 4085 */ 4086unsigned long __init find_min_pfn_with_active_regions(void) 4087{ 4088 return find_min_pfn_for_node(MAX_NUMNODES); 4089} 4090 4091/* 4092 * early_calculate_totalpages() 4093 * Sum pages in active regions for movable zone. 4094 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4095 */ 4096static unsigned long __init early_calculate_totalpages(void) 4097{ 4098 int i; 4099 unsigned long totalpages = 0; 4100 4101 for (i = 0; i < nr_nodemap_entries; i++) { 4102 unsigned long pages = early_node_map[i].end_pfn - 4103 early_node_map[i].start_pfn; 4104 totalpages += pages; 4105 if (pages) 4106 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4107 } 4108 return totalpages; 4109} 4110 4111/* 4112 * Find the PFN the Movable zone begins in each node. Kernel memory 4113 * is spread evenly between nodes as long as the nodes have enough 4114 * memory. When they don't, some nodes will have more kernelcore than 4115 * others 4116 */ 4117static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4118{ 4119 int i, nid; 4120 unsigned long usable_startpfn; 4121 unsigned long kernelcore_node, kernelcore_remaining; 4122 /* save the state before borrow the nodemask */ 4123 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4124 unsigned long totalpages = early_calculate_totalpages(); 4125 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4126 4127 /* 4128 * If movablecore was specified, calculate what size of 4129 * kernelcore that corresponds so that memory usable for 4130 * any allocation type is evenly spread. If both kernelcore 4131 * and movablecore are specified, then the value of kernelcore 4132 * will be used for required_kernelcore if it's greater than 4133 * what movablecore would have allowed. 4134 */ 4135 if (required_movablecore) { 4136 unsigned long corepages; 4137 4138 /* 4139 * Round-up so that ZONE_MOVABLE is at least as large as what 4140 * was requested by the user 4141 */ 4142 required_movablecore = 4143 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4144 corepages = totalpages - required_movablecore; 4145 4146 required_kernelcore = max(required_kernelcore, corepages); 4147 } 4148 4149 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4150 if (!required_kernelcore) 4151 goto out; 4152 4153 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4154 find_usable_zone_for_movable(); 4155 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4156 4157restart: 4158 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4159 kernelcore_node = required_kernelcore / usable_nodes; 4160 for_each_node_state(nid, N_HIGH_MEMORY) { 4161 /* 4162 * Recalculate kernelcore_node if the division per node 4163 * now exceeds what is necessary to satisfy the requested 4164 * amount of memory for the kernel 4165 */ 4166 if (required_kernelcore < kernelcore_node) 4167 kernelcore_node = required_kernelcore / usable_nodes; 4168 4169 /* 4170 * As the map is walked, we track how much memory is usable 4171 * by the kernel using kernelcore_remaining. When it is 4172 * 0, the rest of the node is usable by ZONE_MOVABLE 4173 */ 4174 kernelcore_remaining = kernelcore_node; 4175 4176 /* Go through each range of PFNs within this node */ 4177 for_each_active_range_index_in_nid(i, nid) { 4178 unsigned long start_pfn, end_pfn; 4179 unsigned long size_pages; 4180 4181 start_pfn = max(early_node_map[i].start_pfn, 4182 zone_movable_pfn[nid]); 4183 end_pfn = early_node_map[i].end_pfn; 4184 if (start_pfn >= end_pfn) 4185 continue; 4186 4187 /* Account for what is only usable for kernelcore */ 4188 if (start_pfn < usable_startpfn) { 4189 unsigned long kernel_pages; 4190 kernel_pages = min(end_pfn, usable_startpfn) 4191 - start_pfn; 4192 4193 kernelcore_remaining -= min(kernel_pages, 4194 kernelcore_remaining); 4195 required_kernelcore -= min(kernel_pages, 4196 required_kernelcore); 4197 4198 /* Continue if range is now fully accounted */ 4199 if (end_pfn <= usable_startpfn) { 4200 4201 /* 4202 * Push zone_movable_pfn to the end so 4203 * that if we have to rebalance 4204 * kernelcore across nodes, we will 4205 * not double account here 4206 */ 4207 zone_movable_pfn[nid] = end_pfn; 4208 continue; 4209 } 4210 start_pfn = usable_startpfn; 4211 } 4212 4213 /* 4214 * The usable PFN range for ZONE_MOVABLE is from 4215 * start_pfn->end_pfn. Calculate size_pages as the 4216 * number of pages used as kernelcore 4217 */ 4218 size_pages = end_pfn - start_pfn; 4219 if (size_pages > kernelcore_remaining) 4220 size_pages = kernelcore_remaining; 4221 zone_movable_pfn[nid] = start_pfn + size_pages; 4222 4223 /* 4224 * Some kernelcore has been met, update counts and 4225 * break if the kernelcore for this node has been 4226 * satisified 4227 */ 4228 required_kernelcore -= min(required_kernelcore, 4229 size_pages); 4230 kernelcore_remaining -= size_pages; 4231 if (!kernelcore_remaining) 4232 break; 4233 } 4234 } 4235 4236 /* 4237 * If there is still required_kernelcore, we do another pass with one 4238 * less node in the count. This will push zone_movable_pfn[nid] further 4239 * along on the nodes that still have memory until kernelcore is 4240 * satisified 4241 */ 4242 usable_nodes--; 4243 if (usable_nodes && required_kernelcore > usable_nodes) 4244 goto restart; 4245 4246 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4247 for (nid = 0; nid < MAX_NUMNODES; nid++) 4248 zone_movable_pfn[nid] = 4249 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4250 4251out: 4252 /* restore the node_state */ 4253 node_states[N_HIGH_MEMORY] = saved_node_state; 4254} 4255 4256/* Any regular memory on that node ? */ 4257static void check_for_regular_memory(pg_data_t *pgdat) 4258{ 4259#ifdef CONFIG_HIGHMEM 4260 enum zone_type zone_type; 4261 4262 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4263 struct zone *zone = &pgdat->node_zones[zone_type]; 4264 if (zone->present_pages) 4265 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4266 } 4267#endif 4268} 4269 4270/** 4271 * free_area_init_nodes - Initialise all pg_data_t and zone data 4272 * @max_zone_pfn: an array of max PFNs for each zone 4273 * 4274 * This will call free_area_init_node() for each active node in the system. 4275 * Using the page ranges provided by add_active_range(), the size of each 4276 * zone in each node and their holes is calculated. If the maximum PFN 4277 * between two adjacent zones match, it is assumed that the zone is empty. 4278 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4279 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4280 * starts where the previous one ended. For example, ZONE_DMA32 starts 4281 * at arch_max_dma_pfn. 4282 */ 4283void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4284{ 4285 unsigned long nid; 4286 int i; 4287 4288 /* Sort early_node_map as initialisation assumes it is sorted */ 4289 sort_node_map(); 4290 4291 /* Record where the zone boundaries are */ 4292 memset(arch_zone_lowest_possible_pfn, 0, 4293 sizeof(arch_zone_lowest_possible_pfn)); 4294 memset(arch_zone_highest_possible_pfn, 0, 4295 sizeof(arch_zone_highest_possible_pfn)); 4296 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4297 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4298 for (i = 1; i < MAX_NR_ZONES; i++) { 4299 if (i == ZONE_MOVABLE) 4300 continue; 4301 arch_zone_lowest_possible_pfn[i] = 4302 arch_zone_highest_possible_pfn[i-1]; 4303 arch_zone_highest_possible_pfn[i] = 4304 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4305 } 4306 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4307 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4308 4309 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4310 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4311 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4312 4313 /* Print out the zone ranges */ 4314 printk("Zone PFN ranges:\n"); 4315 for (i = 0; i < MAX_NR_ZONES; i++) { 4316 if (i == ZONE_MOVABLE) 4317 continue; 4318 printk(" %-8s %0#10lx -> %0#10lx\n", 4319 zone_names[i], 4320 arch_zone_lowest_possible_pfn[i], 4321 arch_zone_highest_possible_pfn[i]); 4322 } 4323 4324 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4325 printk("Movable zone start PFN for each node\n"); 4326 for (i = 0; i < MAX_NUMNODES; i++) { 4327 if (zone_movable_pfn[i]) 4328 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4329 } 4330 4331 /* Print out the early_node_map[] */ 4332 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4333 for (i = 0; i < nr_nodemap_entries; i++) 4334 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4335 early_node_map[i].start_pfn, 4336 early_node_map[i].end_pfn); 4337 4338 /* Initialise every node */ 4339 mminit_verify_pageflags_layout(); 4340 setup_nr_node_ids(); 4341 for_each_online_node(nid) { 4342 pg_data_t *pgdat = NODE_DATA(nid); 4343 free_area_init_node(nid, NULL, 4344 find_min_pfn_for_node(nid), NULL); 4345 4346 /* Any memory on that node */ 4347 if (pgdat->node_present_pages) 4348 node_set_state(nid, N_HIGH_MEMORY); 4349 check_for_regular_memory(pgdat); 4350 } 4351} 4352 4353static int __init cmdline_parse_core(char *p, unsigned long *core) 4354{ 4355 unsigned long long coremem; 4356 if (!p) 4357 return -EINVAL; 4358 4359 coremem = memparse(p, &p); 4360 *core = coremem >> PAGE_SHIFT; 4361 4362 /* Paranoid check that UL is enough for the coremem value */ 4363 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4364 4365 return 0; 4366} 4367 4368/* 4369 * kernelcore=size sets the amount of memory for use for allocations that 4370 * cannot be reclaimed or migrated. 4371 */ 4372static int __init cmdline_parse_kernelcore(char *p) 4373{ 4374 return cmdline_parse_core(p, &required_kernelcore); 4375} 4376 4377/* 4378 * movablecore=size sets the amount of memory for use for allocations that 4379 * can be reclaimed or migrated. 4380 */ 4381static int __init cmdline_parse_movablecore(char *p) 4382{ 4383 return cmdline_parse_core(p, &required_movablecore); 4384} 4385 4386early_param("kernelcore", cmdline_parse_kernelcore); 4387early_param("movablecore", cmdline_parse_movablecore); 4388 4389#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4390 4391/** 4392 * set_dma_reserve - set the specified number of pages reserved in the first zone 4393 * @new_dma_reserve: The number of pages to mark reserved 4394 * 4395 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4396 * In the DMA zone, a significant percentage may be consumed by kernel image 4397 * and other unfreeable allocations which can skew the watermarks badly. This 4398 * function may optionally be used to account for unfreeable pages in the 4399 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4400 * smaller per-cpu batchsize. 4401 */ 4402void __init set_dma_reserve(unsigned long new_dma_reserve) 4403{ 4404 dma_reserve = new_dma_reserve; 4405} 4406 4407#ifndef CONFIG_NEED_MULTIPLE_NODES 4408struct pglist_data __refdata contig_page_data = { .bdata = &bootmem_node_data[0] }; 4409EXPORT_SYMBOL(contig_page_data); 4410#endif 4411 4412void __init free_area_init(unsigned long *zones_size) 4413{ 4414 free_area_init_node(0, zones_size, 4415 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4416} 4417 4418static int page_alloc_cpu_notify(struct notifier_block *self, 4419 unsigned long action, void *hcpu) 4420{ 4421 int cpu = (unsigned long)hcpu; 4422 4423 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4424 drain_pages(cpu); 4425 4426 /* 4427 * Spill the event counters of the dead processor 4428 * into the current processors event counters. 4429 * This artificially elevates the count of the current 4430 * processor. 4431 */ 4432 vm_events_fold_cpu(cpu); 4433 4434 /* 4435 * Zero the differential counters of the dead processor 4436 * so that the vm statistics are consistent. 4437 * 4438 * This is only okay since the processor is dead and cannot 4439 * race with what we are doing. 4440 */ 4441 refresh_cpu_vm_stats(cpu); 4442 } 4443 return NOTIFY_OK; 4444} 4445 4446void __init page_alloc_init(void) 4447{ 4448 hotcpu_notifier(page_alloc_cpu_notify, 0); 4449} 4450 4451/* 4452 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4453 * or min_free_kbytes changes. 4454 */ 4455static void calculate_totalreserve_pages(void) 4456{ 4457 struct pglist_data *pgdat; 4458 unsigned long reserve_pages = 0; 4459 enum zone_type i, j; 4460 4461 for_each_online_pgdat(pgdat) { 4462 for (i = 0; i < MAX_NR_ZONES; i++) { 4463 struct zone *zone = pgdat->node_zones + i; 4464 unsigned long max = 0; 4465 4466 /* Find valid and maximum lowmem_reserve in the zone */ 4467 for (j = i; j < MAX_NR_ZONES; j++) { 4468 if (zone->lowmem_reserve[j] > max) 4469 max = zone->lowmem_reserve[j]; 4470 } 4471 4472 /* we treat the high watermark as reserved pages. */ 4473 max += high_wmark_pages(zone); 4474 4475 if (max > zone->present_pages) 4476 max = zone->present_pages; 4477 reserve_pages += max; 4478 } 4479 } 4480 totalreserve_pages = reserve_pages; 4481} 4482 4483/* 4484 * setup_per_zone_lowmem_reserve - called whenever 4485 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4486 * has a correct pages reserved value, so an adequate number of 4487 * pages are left in the zone after a successful __alloc_pages(). 4488 */ 4489static void setup_per_zone_lowmem_reserve(void) 4490{ 4491 struct pglist_data *pgdat; 4492 enum zone_type j, idx; 4493 4494 for_each_online_pgdat(pgdat) { 4495 for (j = 0; j < MAX_NR_ZONES; j++) { 4496 struct zone *zone = pgdat->node_zones + j; 4497 unsigned long present_pages = zone->present_pages; 4498 4499 zone->lowmem_reserve[j] = 0; 4500 4501 idx = j; 4502 while (idx) { 4503 struct zone *lower_zone; 4504 4505 idx--; 4506 4507 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4508 sysctl_lowmem_reserve_ratio[idx] = 1; 4509 4510 lower_zone = pgdat->node_zones + idx; 4511 lower_zone->lowmem_reserve[j] = present_pages / 4512 sysctl_lowmem_reserve_ratio[idx]; 4513 present_pages += lower_zone->present_pages; 4514 } 4515 } 4516 } 4517 4518 /* update totalreserve_pages */ 4519 calculate_totalreserve_pages(); 4520} 4521 4522/** 4523 * setup_per_zone_wmarks - called when min_free_kbytes changes 4524 * or when memory is hot-{added|removed} 4525 * 4526 * Ensures that the watermark[min,low,high] values for each zone are set 4527 * correctly with respect to min_free_kbytes. 4528 */ 4529void setup_per_zone_wmarks(void) 4530{ 4531 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4532 unsigned long lowmem_pages = 0; 4533 struct zone *zone; 4534 unsigned long flags; 4535 4536 /* Calculate total number of !ZONE_HIGHMEM pages */ 4537 for_each_zone(zone) { 4538 if (!is_highmem(zone)) 4539 lowmem_pages += zone->present_pages; 4540 } 4541 4542 for_each_zone(zone) { 4543 u64 tmp; 4544 4545 spin_lock_irqsave(&zone->lock, flags); 4546 tmp = (u64)pages_min * zone->present_pages; 4547 do_div(tmp, lowmem_pages); 4548 if (is_highmem(zone)) { 4549 /* 4550 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4551 * need highmem pages, so cap pages_min to a small 4552 * value here. 4553 * 4554 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4555 * deltas controls asynch page reclaim, and so should 4556 * not be capped for highmem. 4557 */ 4558 int min_pages; 4559 4560 min_pages = zone->present_pages / 1024; 4561 if (min_pages < SWAP_CLUSTER_MAX) 4562 min_pages = SWAP_CLUSTER_MAX; 4563 if (min_pages > 128) 4564 min_pages = 128; 4565 zone->watermark[WMARK_MIN] = min_pages; 4566 } else { 4567 /* 4568 * If it's a lowmem zone, reserve a number of pages 4569 * proportionate to the zone's size. 4570 */ 4571 zone->watermark[WMARK_MIN] = tmp; 4572 } 4573 4574 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4575 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4576 setup_zone_migrate_reserve(zone); 4577 spin_unlock_irqrestore(&zone->lock, flags); 4578 } 4579 4580 /* update totalreserve_pages */ 4581 calculate_totalreserve_pages(); 4582} 4583 4584/* 4585 * The inactive anon list should be small enough that the VM never has to 4586 * do too much work, but large enough that each inactive page has a chance 4587 * to be referenced again before it is swapped out. 4588 * 4589 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4590 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4591 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4592 * the anonymous pages are kept on the inactive list. 4593 * 4594 * total target max 4595 * memory ratio inactive anon 4596 * ------------------------------------- 4597 * 10MB 1 5MB 4598 * 100MB 1 50MB 4599 * 1GB 3 250MB 4600 * 10GB 10 0.9GB 4601 * 100GB 31 3GB 4602 * 1TB 101 10GB 4603 * 10TB 320 32GB 4604 */ 4605void calculate_zone_inactive_ratio(struct zone *zone) 4606{ 4607 unsigned int gb, ratio; 4608 4609 /* Zone size in gigabytes */ 4610 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4611 if (gb) 4612 ratio = int_sqrt(10 * gb); 4613 else 4614 ratio = 1; 4615 4616 zone->inactive_ratio = ratio; 4617} 4618 4619static void __init setup_per_zone_inactive_ratio(void) 4620{ 4621 struct zone *zone; 4622 4623 for_each_zone(zone) 4624 calculate_zone_inactive_ratio(zone); 4625} 4626 4627/* 4628 * Initialise min_free_kbytes. 4629 * 4630 * For small machines we want it small (128k min). For large machines 4631 * we want it large (64MB max). But it is not linear, because network 4632 * bandwidth does not increase linearly with machine size. We use 4633 * 4634 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4635 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4636 * 4637 * which yields 4638 * 4639 * 16MB: 512k 4640 * 32MB: 724k 4641 * 64MB: 1024k 4642 * 128MB: 1448k 4643 * 256MB: 2048k 4644 * 512MB: 2896k 4645 * 1024MB: 4096k 4646 * 2048MB: 5792k 4647 * 4096MB: 8192k 4648 * 8192MB: 11584k 4649 * 16384MB: 16384k 4650 */ 4651static int __init init_per_zone_wmark_min(void) 4652{ 4653 unsigned long lowmem_kbytes; 4654 4655 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4656 4657 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4658 if (min_free_kbytes < 128) 4659 min_free_kbytes = 128; 4660 if (min_free_kbytes > 65536) 4661 min_free_kbytes = 65536; 4662 setup_per_zone_wmarks(); 4663 setup_per_zone_lowmem_reserve(); 4664 setup_per_zone_inactive_ratio(); 4665 return 0; 4666} 4667module_init(init_per_zone_wmark_min) 4668 4669/* 4670 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4671 * that we can call two helper functions whenever min_free_kbytes 4672 * changes. 4673 */ 4674int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4675 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4676{ 4677 proc_dointvec(table, write, file, buffer, length, ppos); 4678 if (write) 4679 setup_per_zone_wmarks(); 4680 return 0; 4681} 4682 4683#ifdef CONFIG_NUMA 4684int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4685 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4686{ 4687 struct zone *zone; 4688 int rc; 4689 4690 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4691 if (rc) 4692 return rc; 4693 4694 for_each_zone(zone) 4695 zone->min_unmapped_pages = (zone->present_pages * 4696 sysctl_min_unmapped_ratio) / 100; 4697 return 0; 4698} 4699 4700int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4701 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4702{ 4703 struct zone *zone; 4704 int rc; 4705 4706 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4707 if (rc) 4708 return rc; 4709 4710 for_each_zone(zone) 4711 zone->min_slab_pages = (zone->present_pages * 4712 sysctl_min_slab_ratio) / 100; 4713 return 0; 4714} 4715#endif 4716 4717/* 4718 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4719 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4720 * whenever sysctl_lowmem_reserve_ratio changes. 4721 * 4722 * The reserve ratio obviously has absolutely no relation with the 4723 * minimum watermarks. The lowmem reserve ratio can only make sense 4724 * if in function of the boot time zone sizes. 4725 */ 4726int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4727 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4728{ 4729 proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4730 setup_per_zone_lowmem_reserve(); 4731 return 0; 4732} 4733 4734/* 4735 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4736 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4737 * can have before it gets flushed back to buddy allocator. 4738 */ 4739 4740int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4741 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4742{ 4743 struct zone *zone; 4744 unsigned int cpu; 4745 int ret; 4746 4747 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4748 if (!write || (ret == -EINVAL)) 4749 return ret; 4750 for_each_populated_zone(zone) { 4751 for_each_online_cpu(cpu) { 4752 unsigned long high; 4753 high = zone->present_pages / percpu_pagelist_fraction; 4754 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4755 } 4756 } 4757 return 0; 4758} 4759 4760int hashdist = HASHDIST_DEFAULT; 4761 4762#ifdef CONFIG_NUMA 4763static int __init set_hashdist(char *str) 4764{ 4765 if (!str) 4766 return 0; 4767 hashdist = simple_strtoul(str, &str, 0); 4768 return 1; 4769} 4770__setup("hashdist=", set_hashdist); 4771#endif 4772 4773/* 4774 * allocate a large system hash table from bootmem 4775 * - it is assumed that the hash table must contain an exact power-of-2 4776 * quantity of entries 4777 * - limit is the number of hash buckets, not the total allocation size 4778 */ 4779void *__init alloc_large_system_hash(const char *tablename, 4780 unsigned long bucketsize, 4781 unsigned long numentries, 4782 int scale, 4783 int flags, 4784 unsigned int *_hash_shift, 4785 unsigned int *_hash_mask, 4786 unsigned long limit) 4787{ 4788 unsigned long long max = limit; 4789 unsigned long log2qty, size; 4790 void *table = NULL; 4791 4792 /* allow the kernel cmdline to have a say */ 4793 if (!numentries) { 4794 /* round applicable memory size up to nearest megabyte */ 4795 numentries = nr_kernel_pages; 4796 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4797 numentries >>= 20 - PAGE_SHIFT; 4798 numentries <<= 20 - PAGE_SHIFT; 4799 4800 /* limit to 1 bucket per 2^scale bytes of low memory */ 4801 if (scale > PAGE_SHIFT) 4802 numentries >>= (scale - PAGE_SHIFT); 4803 else 4804 numentries <<= (PAGE_SHIFT - scale); 4805 4806 /* Make sure we've got at least a 0-order allocation.. */ 4807 if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4808 numentries = PAGE_SIZE / bucketsize; 4809 } 4810 numentries = roundup_pow_of_two(numentries); 4811 4812 /* limit allocation size to 1/16 total memory by default */ 4813 if (max == 0) { 4814 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4815 do_div(max, bucketsize); 4816 } 4817 4818 if (numentries > max) 4819 numentries = max; 4820 4821 log2qty = ilog2(numentries); 4822 4823 do { 4824 size = bucketsize << log2qty; 4825 if (flags & HASH_EARLY) 4826 table = alloc_bootmem_nopanic(size); 4827 else if (hashdist) 4828 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4829 else { 4830 /* 4831 * If bucketsize is not a power-of-two, we may free 4832 * some pages at the end of hash table which 4833 * alloc_pages_exact() automatically does 4834 */ 4835 if (get_order(size) < MAX_ORDER) { 4836 table = alloc_pages_exact(size, GFP_ATOMIC); 4837 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4838 } 4839 } 4840 } while (!table && size > PAGE_SIZE && --log2qty); 4841 4842 if (!table) 4843 panic("Failed to allocate %s hash table\n", tablename); 4844 4845 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4846 tablename, 4847 (1U << log2qty), 4848 ilog2(size) - PAGE_SHIFT, 4849 size); 4850 4851 if (_hash_shift) 4852 *_hash_shift = log2qty; 4853 if (_hash_mask) 4854 *_hash_mask = (1 << log2qty) - 1; 4855 4856 return table; 4857} 4858 4859/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4860static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4861 unsigned long pfn) 4862{ 4863#ifdef CONFIG_SPARSEMEM 4864 return __pfn_to_section(pfn)->pageblock_flags; 4865#else 4866 return zone->pageblock_flags; 4867#endif /* CONFIG_SPARSEMEM */ 4868} 4869 4870static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4871{ 4872#ifdef CONFIG_SPARSEMEM 4873 pfn &= (PAGES_PER_SECTION-1); 4874 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4875#else 4876 pfn = pfn - zone->zone_start_pfn; 4877 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4878#endif /* CONFIG_SPARSEMEM */ 4879} 4880 4881/** 4882 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4883 * @page: The page within the block of interest 4884 * @start_bitidx: The first bit of interest to retrieve 4885 * @end_bitidx: The last bit of interest 4886 * returns pageblock_bits flags 4887 */ 4888unsigned long get_pageblock_flags_group(struct page *page, 4889 int start_bitidx, int end_bitidx) 4890{ 4891 struct zone *zone; 4892 unsigned long *bitmap; 4893 unsigned long pfn, bitidx; 4894 unsigned long flags = 0; 4895 unsigned long value = 1; 4896 4897 zone = page_zone(page); 4898 pfn = page_to_pfn(page); 4899 bitmap = get_pageblock_bitmap(zone, pfn); 4900 bitidx = pfn_to_bitidx(zone, pfn); 4901 4902 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4903 if (test_bit(bitidx + start_bitidx, bitmap)) 4904 flags |= value; 4905 4906 return flags; 4907} 4908 4909/** 4910 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 4911 * @page: The page within the block of interest 4912 * @start_bitidx: The first bit of interest 4913 * @end_bitidx: The last bit of interest 4914 * @flags: The flags to set 4915 */ 4916void set_pageblock_flags_group(struct page *page, unsigned long flags, 4917 int start_bitidx, int end_bitidx) 4918{ 4919 struct zone *zone; 4920 unsigned long *bitmap; 4921 unsigned long pfn, bitidx; 4922 unsigned long value = 1; 4923 4924 zone = page_zone(page); 4925 pfn = page_to_pfn(page); 4926 bitmap = get_pageblock_bitmap(zone, pfn); 4927 bitidx = pfn_to_bitidx(zone, pfn); 4928 VM_BUG_ON(pfn < zone->zone_start_pfn); 4929 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 4930 4931 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4932 if (flags & value) 4933 __set_bit(bitidx + start_bitidx, bitmap); 4934 else 4935 __clear_bit(bitidx + start_bitidx, bitmap); 4936} 4937 4938/* 4939 * This is designed as sub function...plz see page_isolation.c also. 4940 * set/clear page block's type to be ISOLATE. 4941 * page allocater never alloc memory from ISOLATE block. 4942 */ 4943 4944int set_migratetype_isolate(struct page *page) 4945{ 4946 struct zone *zone; 4947 unsigned long flags; 4948 int ret = -EBUSY; 4949 int zone_idx; 4950 4951 zone = page_zone(page); 4952 zone_idx = zone_idx(zone); 4953 spin_lock_irqsave(&zone->lock, flags); 4954 /* 4955 * In future, more migrate types will be able to be isolation target. 4956 */ 4957 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE && 4958 zone_idx != ZONE_MOVABLE) 4959 goto out; 4960 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 4961 move_freepages_block(zone, page, MIGRATE_ISOLATE); 4962 ret = 0; 4963out: 4964 spin_unlock_irqrestore(&zone->lock, flags); 4965 if (!ret) 4966 drain_all_pages(); 4967 return ret; 4968} 4969 4970void unset_migratetype_isolate(struct page *page) 4971{ 4972 struct zone *zone; 4973 unsigned long flags; 4974 zone = page_zone(page); 4975 spin_lock_irqsave(&zone->lock, flags); 4976 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 4977 goto out; 4978 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4979 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4980out: 4981 spin_unlock_irqrestore(&zone->lock, flags); 4982} 4983 4984#ifdef CONFIG_MEMORY_HOTREMOVE 4985/* 4986 * All pages in the range must be isolated before calling this. 4987 */ 4988void 4989__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 4990{ 4991 struct page *page; 4992 struct zone *zone; 4993 int order, i; 4994 unsigned long pfn; 4995 unsigned long flags; 4996 /* find the first valid pfn */ 4997 for (pfn = start_pfn; pfn < end_pfn; pfn++) 4998 if (pfn_valid(pfn)) 4999 break; 5000 if (pfn == end_pfn) 5001 return; 5002 zone = page_zone(pfn_to_page(pfn)); 5003 spin_lock_irqsave(&zone->lock, flags); 5004 pfn = start_pfn; 5005 while (pfn < end_pfn) { 5006 if (!pfn_valid(pfn)) { 5007 pfn++; 5008 continue; 5009 } 5010 page = pfn_to_page(pfn); 5011 BUG_ON(page_count(page)); 5012 BUG_ON(!PageBuddy(page)); 5013 order = page_order(page); 5014#ifdef CONFIG_DEBUG_VM 5015 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5016 pfn, 1 << order, end_pfn); 5017#endif 5018 list_del(&page->lru); 5019 rmv_page_order(page); 5020 zone->free_area[order].nr_free--; 5021 __mod_zone_page_state(zone, NR_FREE_PAGES, 5022 - (1UL << order)); 5023 for (i = 0; i < (1 << order); i++) 5024 SetPageReserved((page+i)); 5025 pfn += (1 << order); 5026 } 5027 spin_unlock_irqrestore(&zone->lock, flags); 5028} 5029#endif 5030