page_alloc.c revision 2d30a1f6315b8940537e8e98882c6038fbac9ba5
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/compiler.h> 25#include <linux/kernel.h> 26#include <linux/kmemcheck.h> 27#include <linux/module.h> 28#include <linux/suspend.h> 29#include <linux/pagevec.h> 30#include <linux/blkdev.h> 31#include <linux/slab.h> 32#include <linux/oom.h> 33#include <linux/notifier.h> 34#include <linux/topology.h> 35#include <linux/sysctl.h> 36#include <linux/cpu.h> 37#include <linux/cpuset.h> 38#include <linux/memory_hotplug.h> 39#include <linux/nodemask.h> 40#include <linux/vmalloc.h> 41#include <linux/mempolicy.h> 42#include <linux/stop_machine.h> 43#include <linux/sort.h> 44#include <linux/pfn.h> 45#include <linux/backing-dev.h> 46#include <linux/fault-inject.h> 47#include <linux/page-isolation.h> 48#include <linux/page_cgroup.h> 49#include <linux/debugobjects.h> 50#include <linux/kmemleak.h> 51#include <linux/memory.h> 52#include <trace/events/kmem.h> 53 54#include <asm/tlbflush.h> 55#include <asm/div64.h> 56#include "internal.h" 57 58/* 59 * Array of node states. 60 */ 61nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 62 [N_POSSIBLE] = NODE_MASK_ALL, 63 [N_ONLINE] = { { [0] = 1UL } }, 64#ifndef CONFIG_NUMA 65 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 66#ifdef CONFIG_HIGHMEM 67 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 68#endif 69 [N_CPU] = { { [0] = 1UL } }, 70#endif /* NUMA */ 71}; 72EXPORT_SYMBOL(node_states); 73 74unsigned long totalram_pages __read_mostly; 75unsigned long totalreserve_pages __read_mostly; 76int percpu_pagelist_fraction; 77gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 78 79#ifdef CONFIG_PM_SLEEP 80/* 81 * The following functions are used by the suspend/hibernate code to temporarily 82 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 83 * while devices are suspended. To avoid races with the suspend/hibernate code, 84 * they should always be called with pm_mutex held (gfp_allowed_mask also should 85 * only be modified with pm_mutex held, unless the suspend/hibernate code is 86 * guaranteed not to run in parallel with that modification). 87 */ 88void set_gfp_allowed_mask(gfp_t mask) 89{ 90 WARN_ON(!mutex_is_locked(&pm_mutex)); 91 gfp_allowed_mask = mask; 92} 93 94gfp_t clear_gfp_allowed_mask(gfp_t mask) 95{ 96 gfp_t ret = gfp_allowed_mask; 97 98 WARN_ON(!mutex_is_locked(&pm_mutex)); 99 gfp_allowed_mask &= ~mask; 100 return ret; 101} 102#endif /* CONFIG_PM_SLEEP */ 103 104#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 105int pageblock_order __read_mostly; 106#endif 107 108static void __free_pages_ok(struct page *page, unsigned int order); 109 110/* 111 * results with 256, 32 in the lowmem_reserve sysctl: 112 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 113 * 1G machine -> (16M dma, 784M normal, 224M high) 114 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 115 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 116 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 117 * 118 * TBD: should special case ZONE_DMA32 machines here - in those we normally 119 * don't need any ZONE_NORMAL reservation 120 */ 121int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 122#ifdef CONFIG_ZONE_DMA 123 256, 124#endif 125#ifdef CONFIG_ZONE_DMA32 126 256, 127#endif 128#ifdef CONFIG_HIGHMEM 129 32, 130#endif 131 32, 132}; 133 134EXPORT_SYMBOL(totalram_pages); 135 136static char * const zone_names[MAX_NR_ZONES] = { 137#ifdef CONFIG_ZONE_DMA 138 "DMA", 139#endif 140#ifdef CONFIG_ZONE_DMA32 141 "DMA32", 142#endif 143 "Normal", 144#ifdef CONFIG_HIGHMEM 145 "HighMem", 146#endif 147 "Movable", 148}; 149 150int min_free_kbytes = 1024; 151 152static unsigned long __meminitdata nr_kernel_pages; 153static unsigned long __meminitdata nr_all_pages; 154static unsigned long __meminitdata dma_reserve; 155 156#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 157 /* 158 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 159 * ranges of memory (RAM) that may be registered with add_active_range(). 160 * Ranges passed to add_active_range() will be merged if possible 161 * so the number of times add_active_range() can be called is 162 * related to the number of nodes and the number of holes 163 */ 164 #ifdef CONFIG_MAX_ACTIVE_REGIONS 165 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 166 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 167 #else 168 #if MAX_NUMNODES >= 32 169 /* If there can be many nodes, allow up to 50 holes per node */ 170 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 171 #else 172 /* By default, allow up to 256 distinct regions */ 173 #define MAX_ACTIVE_REGIONS 256 174 #endif 175 #endif 176 177 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 178 static int __meminitdata nr_nodemap_entries; 179 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 180 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 181 static unsigned long __initdata required_kernelcore; 182 static unsigned long __initdata required_movablecore; 183 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 184 185 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 186 int movable_zone; 187 EXPORT_SYMBOL(movable_zone); 188#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 189 190#if MAX_NUMNODES > 1 191int nr_node_ids __read_mostly = MAX_NUMNODES; 192int nr_online_nodes __read_mostly = 1; 193EXPORT_SYMBOL(nr_node_ids); 194EXPORT_SYMBOL(nr_online_nodes); 195#endif 196 197int page_group_by_mobility_disabled __read_mostly; 198 199static void set_pageblock_migratetype(struct page *page, int migratetype) 200{ 201 202 if (unlikely(page_group_by_mobility_disabled)) 203 migratetype = MIGRATE_UNMOVABLE; 204 205 set_pageblock_flags_group(page, (unsigned long)migratetype, 206 PB_migrate, PB_migrate_end); 207} 208 209bool oom_killer_disabled __read_mostly; 210 211#ifdef CONFIG_DEBUG_VM 212static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 213{ 214 int ret = 0; 215 unsigned seq; 216 unsigned long pfn = page_to_pfn(page); 217 218 do { 219 seq = zone_span_seqbegin(zone); 220 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 221 ret = 1; 222 else if (pfn < zone->zone_start_pfn) 223 ret = 1; 224 } while (zone_span_seqretry(zone, seq)); 225 226 return ret; 227} 228 229static int page_is_consistent(struct zone *zone, struct page *page) 230{ 231 if (!pfn_valid_within(page_to_pfn(page))) 232 return 0; 233 if (zone != page_zone(page)) 234 return 0; 235 236 return 1; 237} 238/* 239 * Temporary debugging check for pages not lying within a given zone. 240 */ 241static int bad_range(struct zone *zone, struct page *page) 242{ 243 if (page_outside_zone_boundaries(zone, page)) 244 return 1; 245 if (!page_is_consistent(zone, page)) 246 return 1; 247 248 return 0; 249} 250#else 251static inline int bad_range(struct zone *zone, struct page *page) 252{ 253 return 0; 254} 255#endif 256 257static void bad_page(struct page *page) 258{ 259 static unsigned long resume; 260 static unsigned long nr_shown; 261 static unsigned long nr_unshown; 262 263 /* Don't complain about poisoned pages */ 264 if (PageHWPoison(page)) { 265 __ClearPageBuddy(page); 266 return; 267 } 268 269 /* 270 * Allow a burst of 60 reports, then keep quiet for that minute; 271 * or allow a steady drip of one report per second. 272 */ 273 if (nr_shown == 60) { 274 if (time_before(jiffies, resume)) { 275 nr_unshown++; 276 goto out; 277 } 278 if (nr_unshown) { 279 printk(KERN_ALERT 280 "BUG: Bad page state: %lu messages suppressed\n", 281 nr_unshown); 282 nr_unshown = 0; 283 } 284 nr_shown = 0; 285 } 286 if (nr_shown++ == 0) 287 resume = jiffies + 60 * HZ; 288 289 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 290 current->comm, page_to_pfn(page)); 291 printk(KERN_ALERT 292 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 293 page, (void *)page->flags, page_count(page), 294 page_mapcount(page), page->mapping, page->index); 295 296 dump_stack(); 297out: 298 /* Leave bad fields for debug, except PageBuddy could make trouble */ 299 __ClearPageBuddy(page); 300 add_taint(TAINT_BAD_PAGE); 301} 302 303/* 304 * Higher-order pages are called "compound pages". They are structured thusly: 305 * 306 * The first PAGE_SIZE page is called the "head page". 307 * 308 * The remaining PAGE_SIZE pages are called "tail pages". 309 * 310 * All pages have PG_compound set. All pages have their ->private pointing at 311 * the head page (even the head page has this). 312 * 313 * The first tail page's ->lru.next holds the address of the compound page's 314 * put_page() function. Its ->lru.prev holds the order of allocation. 315 * This usage means that zero-order pages may not be compound. 316 */ 317 318static void free_compound_page(struct page *page) 319{ 320 __free_pages_ok(page, compound_order(page)); 321} 322 323void prep_compound_page(struct page *page, unsigned long order) 324{ 325 int i; 326 int nr_pages = 1 << order; 327 328 set_compound_page_dtor(page, free_compound_page); 329 set_compound_order(page, order); 330 __SetPageHead(page); 331 for (i = 1; i < nr_pages; i++) { 332 struct page *p = page + i; 333 334 __SetPageTail(p); 335 p->first_page = page; 336 } 337} 338 339static int destroy_compound_page(struct page *page, unsigned long order) 340{ 341 int i; 342 int nr_pages = 1 << order; 343 int bad = 0; 344 345 if (unlikely(compound_order(page) != order) || 346 unlikely(!PageHead(page))) { 347 bad_page(page); 348 bad++; 349 } 350 351 __ClearPageHead(page); 352 353 for (i = 1; i < nr_pages; i++) { 354 struct page *p = page + i; 355 356 if (unlikely(!PageTail(p) || (p->first_page != page))) { 357 bad_page(page); 358 bad++; 359 } 360 __ClearPageTail(p); 361 } 362 363 return bad; 364} 365 366static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 367{ 368 int i; 369 370 /* 371 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 372 * and __GFP_HIGHMEM from hard or soft interrupt context. 373 */ 374 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 375 for (i = 0; i < (1 << order); i++) 376 clear_highpage(page + i); 377} 378 379static inline void set_page_order(struct page *page, int order) 380{ 381 set_page_private(page, order); 382 __SetPageBuddy(page); 383} 384 385static inline void rmv_page_order(struct page *page) 386{ 387 __ClearPageBuddy(page); 388 set_page_private(page, 0); 389} 390 391/* 392 * Locate the struct page for both the matching buddy in our 393 * pair (buddy1) and the combined O(n+1) page they form (page). 394 * 395 * 1) Any buddy B1 will have an order O twin B2 which satisfies 396 * the following equation: 397 * B2 = B1 ^ (1 << O) 398 * For example, if the starting buddy (buddy2) is #8 its order 399 * 1 buddy is #10: 400 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 401 * 402 * 2) Any buddy B will have an order O+1 parent P which 403 * satisfies the following equation: 404 * P = B & ~(1 << O) 405 * 406 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 407 */ 408static inline struct page * 409__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 410{ 411 unsigned long buddy_idx = page_idx ^ (1 << order); 412 413 return page + (buddy_idx - page_idx); 414} 415 416static inline unsigned long 417__find_combined_index(unsigned long page_idx, unsigned int order) 418{ 419 return (page_idx & ~(1 << order)); 420} 421 422/* 423 * This function checks whether a page is free && is the buddy 424 * we can do coalesce a page and its buddy if 425 * (a) the buddy is not in a hole && 426 * (b) the buddy is in the buddy system && 427 * (c) a page and its buddy have the same order && 428 * (d) a page and its buddy are in the same zone. 429 * 430 * For recording whether a page is in the buddy system, we use PG_buddy. 431 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 432 * 433 * For recording page's order, we use page_private(page). 434 */ 435static inline int page_is_buddy(struct page *page, struct page *buddy, 436 int order) 437{ 438 if (!pfn_valid_within(page_to_pfn(buddy))) 439 return 0; 440 441 if (page_zone_id(page) != page_zone_id(buddy)) 442 return 0; 443 444 if (PageBuddy(buddy) && page_order(buddy) == order) { 445 VM_BUG_ON(page_count(buddy) != 0); 446 return 1; 447 } 448 return 0; 449} 450 451/* 452 * Freeing function for a buddy system allocator. 453 * 454 * The concept of a buddy system is to maintain direct-mapped table 455 * (containing bit values) for memory blocks of various "orders". 456 * The bottom level table contains the map for the smallest allocatable 457 * units of memory (here, pages), and each level above it describes 458 * pairs of units from the levels below, hence, "buddies". 459 * At a high level, all that happens here is marking the table entry 460 * at the bottom level available, and propagating the changes upward 461 * as necessary, plus some accounting needed to play nicely with other 462 * parts of the VM system. 463 * At each level, we keep a list of pages, which are heads of continuous 464 * free pages of length of (1 << order) and marked with PG_buddy. Page's 465 * order is recorded in page_private(page) field. 466 * So when we are allocating or freeing one, we can derive the state of the 467 * other. That is, if we allocate a small block, and both were 468 * free, the remainder of the region must be split into blocks. 469 * If a block is freed, and its buddy is also free, then this 470 * triggers coalescing into a block of larger size. 471 * 472 * -- wli 473 */ 474 475static inline void __free_one_page(struct page *page, 476 struct zone *zone, unsigned int order, 477 int migratetype) 478{ 479 unsigned long page_idx; 480 481 if (unlikely(PageCompound(page))) 482 if (unlikely(destroy_compound_page(page, order))) 483 return; 484 485 VM_BUG_ON(migratetype == -1); 486 487 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 488 489 VM_BUG_ON(page_idx & ((1 << order) - 1)); 490 VM_BUG_ON(bad_range(zone, page)); 491 492 while (order < MAX_ORDER-1) { 493 unsigned long combined_idx; 494 struct page *buddy; 495 496 buddy = __page_find_buddy(page, page_idx, order); 497 if (!page_is_buddy(page, buddy, order)) 498 break; 499 500 /* Our buddy is free, merge with it and move up one order. */ 501 list_del(&buddy->lru); 502 zone->free_area[order].nr_free--; 503 rmv_page_order(buddy); 504 combined_idx = __find_combined_index(page_idx, order); 505 page = page + (combined_idx - page_idx); 506 page_idx = combined_idx; 507 order++; 508 } 509 set_page_order(page, order); 510 list_add(&page->lru, 511 &zone->free_area[order].free_list[migratetype]); 512 zone->free_area[order].nr_free++; 513} 514 515/* 516 * free_page_mlock() -- clean up attempts to free and mlocked() page. 517 * Page should not be on lru, so no need to fix that up. 518 * free_pages_check() will verify... 519 */ 520static inline void free_page_mlock(struct page *page) 521{ 522 __dec_zone_page_state(page, NR_MLOCK); 523 __count_vm_event(UNEVICTABLE_MLOCKFREED); 524} 525 526static inline int free_pages_check(struct page *page) 527{ 528 if (unlikely(page_mapcount(page) | 529 (page->mapping != NULL) | 530 (atomic_read(&page->_count) != 0) | 531 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 532 bad_page(page); 533 return 1; 534 } 535 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 536 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 537 return 0; 538} 539 540/* 541 * Frees a number of pages from the PCP lists 542 * Assumes all pages on list are in same zone, and of same order. 543 * count is the number of pages to free. 544 * 545 * If the zone was previously in an "all pages pinned" state then look to 546 * see if this freeing clears that state. 547 * 548 * And clear the zone's pages_scanned counter, to hold off the "all pages are 549 * pinned" detection logic. 550 */ 551static void free_pcppages_bulk(struct zone *zone, int count, 552 struct per_cpu_pages *pcp) 553{ 554 int migratetype = 0; 555 int batch_free = 0; 556 557 spin_lock(&zone->lock); 558 zone->all_unreclaimable = 0; 559 zone->pages_scanned = 0; 560 561 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 562 while (count) { 563 struct page *page; 564 struct list_head *list; 565 566 /* 567 * Remove pages from lists in a round-robin fashion. A 568 * batch_free count is maintained that is incremented when an 569 * empty list is encountered. This is so more pages are freed 570 * off fuller lists instead of spinning excessively around empty 571 * lists 572 */ 573 do { 574 batch_free++; 575 if (++migratetype == MIGRATE_PCPTYPES) 576 migratetype = 0; 577 list = &pcp->lists[migratetype]; 578 } while (list_empty(list)); 579 580 do { 581 page = list_entry(list->prev, struct page, lru); 582 /* must delete as __free_one_page list manipulates */ 583 list_del(&page->lru); 584 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 585 __free_one_page(page, zone, 0, page_private(page)); 586 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 587 } while (--count && --batch_free && !list_empty(list)); 588 } 589 spin_unlock(&zone->lock); 590} 591 592static void free_one_page(struct zone *zone, struct page *page, int order, 593 int migratetype) 594{ 595 spin_lock(&zone->lock); 596 zone->all_unreclaimable = 0; 597 zone->pages_scanned = 0; 598 599 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 600 __free_one_page(page, zone, order, migratetype); 601 spin_unlock(&zone->lock); 602} 603 604static void __free_pages_ok(struct page *page, unsigned int order) 605{ 606 unsigned long flags; 607 int i; 608 int bad = 0; 609 int wasMlocked = __TestClearPageMlocked(page); 610 611 trace_mm_page_free_direct(page, order); 612 kmemcheck_free_shadow(page, order); 613 614 for (i = 0 ; i < (1 << order) ; ++i) 615 bad += free_pages_check(page + i); 616 if (bad) 617 return; 618 619 if (!PageHighMem(page)) { 620 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 621 debug_check_no_obj_freed(page_address(page), 622 PAGE_SIZE << order); 623 } 624 arch_free_page(page, order); 625 kernel_map_pages(page, 1 << order, 0); 626 627 local_irq_save(flags); 628 if (unlikely(wasMlocked)) 629 free_page_mlock(page); 630 __count_vm_events(PGFREE, 1 << order); 631 free_one_page(page_zone(page), page, order, 632 get_pageblock_migratetype(page)); 633 local_irq_restore(flags); 634} 635 636/* 637 * permit the bootmem allocator to evade page validation on high-order frees 638 */ 639void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 640{ 641 if (order == 0) { 642 __ClearPageReserved(page); 643 set_page_count(page, 0); 644 set_page_refcounted(page); 645 __free_page(page); 646 } else { 647 int loop; 648 649 prefetchw(page); 650 for (loop = 0; loop < BITS_PER_LONG; loop++) { 651 struct page *p = &page[loop]; 652 653 if (loop + 1 < BITS_PER_LONG) 654 prefetchw(p + 1); 655 __ClearPageReserved(p); 656 set_page_count(p, 0); 657 } 658 659 set_page_refcounted(page); 660 __free_pages(page, order); 661 } 662} 663 664 665/* 666 * The order of subdivision here is critical for the IO subsystem. 667 * Please do not alter this order without good reasons and regression 668 * testing. Specifically, as large blocks of memory are subdivided, 669 * the order in which smaller blocks are delivered depends on the order 670 * they're subdivided in this function. This is the primary factor 671 * influencing the order in which pages are delivered to the IO 672 * subsystem according to empirical testing, and this is also justified 673 * by considering the behavior of a buddy system containing a single 674 * large block of memory acted on by a series of small allocations. 675 * This behavior is a critical factor in sglist merging's success. 676 * 677 * -- wli 678 */ 679static inline void expand(struct zone *zone, struct page *page, 680 int low, int high, struct free_area *area, 681 int migratetype) 682{ 683 unsigned long size = 1 << high; 684 685 while (high > low) { 686 area--; 687 high--; 688 size >>= 1; 689 VM_BUG_ON(bad_range(zone, &page[size])); 690 list_add(&page[size].lru, &area->free_list[migratetype]); 691 area->nr_free++; 692 set_page_order(&page[size], high); 693 } 694} 695 696/* 697 * This page is about to be returned from the page allocator 698 */ 699static inline int check_new_page(struct page *page) 700{ 701 if (unlikely(page_mapcount(page) | 702 (page->mapping != NULL) | 703 (atomic_read(&page->_count) != 0) | 704 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 705 bad_page(page); 706 return 1; 707 } 708 return 0; 709} 710 711static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 712{ 713 int i; 714 715 for (i = 0; i < (1 << order); i++) { 716 struct page *p = page + i; 717 if (unlikely(check_new_page(p))) 718 return 1; 719 } 720 721 set_page_private(page, 0); 722 set_page_refcounted(page); 723 724 arch_alloc_page(page, order); 725 kernel_map_pages(page, 1 << order, 1); 726 727 if (gfp_flags & __GFP_ZERO) 728 prep_zero_page(page, order, gfp_flags); 729 730 if (order && (gfp_flags & __GFP_COMP)) 731 prep_compound_page(page, order); 732 733 return 0; 734} 735 736/* 737 * Go through the free lists for the given migratetype and remove 738 * the smallest available page from the freelists 739 */ 740static inline 741struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 742 int migratetype) 743{ 744 unsigned int current_order; 745 struct free_area * area; 746 struct page *page; 747 748 /* Find a page of the appropriate size in the preferred list */ 749 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 750 area = &(zone->free_area[current_order]); 751 if (list_empty(&area->free_list[migratetype])) 752 continue; 753 754 page = list_entry(area->free_list[migratetype].next, 755 struct page, lru); 756 list_del(&page->lru); 757 rmv_page_order(page); 758 area->nr_free--; 759 expand(zone, page, order, current_order, area, migratetype); 760 return page; 761 } 762 763 return NULL; 764} 765 766 767/* 768 * This array describes the order lists are fallen back to when 769 * the free lists for the desirable migrate type are depleted 770 */ 771static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 772 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 773 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 774 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 775 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 776}; 777 778/* 779 * Move the free pages in a range to the free lists of the requested type. 780 * Note that start_page and end_pages are not aligned on a pageblock 781 * boundary. If alignment is required, use move_freepages_block() 782 */ 783static int move_freepages(struct zone *zone, 784 struct page *start_page, struct page *end_page, 785 int migratetype) 786{ 787 struct page *page; 788 unsigned long order; 789 int pages_moved = 0; 790 791#ifndef CONFIG_HOLES_IN_ZONE 792 /* 793 * page_zone is not safe to call in this context when 794 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 795 * anyway as we check zone boundaries in move_freepages_block(). 796 * Remove at a later date when no bug reports exist related to 797 * grouping pages by mobility 798 */ 799 BUG_ON(page_zone(start_page) != page_zone(end_page)); 800#endif 801 802 for (page = start_page; page <= end_page;) { 803 /* Make sure we are not inadvertently changing nodes */ 804 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 805 806 if (!pfn_valid_within(page_to_pfn(page))) { 807 page++; 808 continue; 809 } 810 811 if (!PageBuddy(page)) { 812 page++; 813 continue; 814 } 815 816 order = page_order(page); 817 list_del(&page->lru); 818 list_add(&page->lru, 819 &zone->free_area[order].free_list[migratetype]); 820 page += 1 << order; 821 pages_moved += 1 << order; 822 } 823 824 return pages_moved; 825} 826 827static int move_freepages_block(struct zone *zone, struct page *page, 828 int migratetype) 829{ 830 unsigned long start_pfn, end_pfn; 831 struct page *start_page, *end_page; 832 833 start_pfn = page_to_pfn(page); 834 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 835 start_page = pfn_to_page(start_pfn); 836 end_page = start_page + pageblock_nr_pages - 1; 837 end_pfn = start_pfn + pageblock_nr_pages - 1; 838 839 /* Do not cross zone boundaries */ 840 if (start_pfn < zone->zone_start_pfn) 841 start_page = page; 842 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 843 return 0; 844 845 return move_freepages(zone, start_page, end_page, migratetype); 846} 847 848static void change_pageblock_range(struct page *pageblock_page, 849 int start_order, int migratetype) 850{ 851 int nr_pageblocks = 1 << (start_order - pageblock_order); 852 853 while (nr_pageblocks--) { 854 set_pageblock_migratetype(pageblock_page, migratetype); 855 pageblock_page += pageblock_nr_pages; 856 } 857} 858 859/* Remove an element from the buddy allocator from the fallback list */ 860static inline struct page * 861__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 862{ 863 struct free_area * area; 864 int current_order; 865 struct page *page; 866 int migratetype, i; 867 868 /* Find the largest possible block of pages in the other list */ 869 for (current_order = MAX_ORDER-1; current_order >= order; 870 --current_order) { 871 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 872 migratetype = fallbacks[start_migratetype][i]; 873 874 /* MIGRATE_RESERVE handled later if necessary */ 875 if (migratetype == MIGRATE_RESERVE) 876 continue; 877 878 area = &(zone->free_area[current_order]); 879 if (list_empty(&area->free_list[migratetype])) 880 continue; 881 882 page = list_entry(area->free_list[migratetype].next, 883 struct page, lru); 884 area->nr_free--; 885 886 /* 887 * If breaking a large block of pages, move all free 888 * pages to the preferred allocation list. If falling 889 * back for a reclaimable kernel allocation, be more 890 * agressive about taking ownership of free pages 891 */ 892 if (unlikely(current_order >= (pageblock_order >> 1)) || 893 start_migratetype == MIGRATE_RECLAIMABLE || 894 page_group_by_mobility_disabled) { 895 unsigned long pages; 896 pages = move_freepages_block(zone, page, 897 start_migratetype); 898 899 /* Claim the whole block if over half of it is free */ 900 if (pages >= (1 << (pageblock_order-1)) || 901 page_group_by_mobility_disabled) 902 set_pageblock_migratetype(page, 903 start_migratetype); 904 905 migratetype = start_migratetype; 906 } 907 908 /* Remove the page from the freelists */ 909 list_del(&page->lru); 910 rmv_page_order(page); 911 912 /* Take ownership for orders >= pageblock_order */ 913 if (current_order >= pageblock_order) 914 change_pageblock_range(page, current_order, 915 start_migratetype); 916 917 expand(zone, page, order, current_order, area, migratetype); 918 919 trace_mm_page_alloc_extfrag(page, order, current_order, 920 start_migratetype, migratetype); 921 922 return page; 923 } 924 } 925 926 return NULL; 927} 928 929/* 930 * Do the hard work of removing an element from the buddy allocator. 931 * Call me with the zone->lock already held. 932 */ 933static struct page *__rmqueue(struct zone *zone, unsigned int order, 934 int migratetype) 935{ 936 struct page *page; 937 938retry_reserve: 939 page = __rmqueue_smallest(zone, order, migratetype); 940 941 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 942 page = __rmqueue_fallback(zone, order, migratetype); 943 944 /* 945 * Use MIGRATE_RESERVE rather than fail an allocation. goto 946 * is used because __rmqueue_smallest is an inline function 947 * and we want just one call site 948 */ 949 if (!page) { 950 migratetype = MIGRATE_RESERVE; 951 goto retry_reserve; 952 } 953 } 954 955 trace_mm_page_alloc_zone_locked(page, order, migratetype); 956 return page; 957} 958 959/* 960 * Obtain a specified number of elements from the buddy allocator, all under 961 * a single hold of the lock, for efficiency. Add them to the supplied list. 962 * Returns the number of new pages which were placed at *list. 963 */ 964static int rmqueue_bulk(struct zone *zone, unsigned int order, 965 unsigned long count, struct list_head *list, 966 int migratetype, int cold) 967{ 968 int i; 969 970 spin_lock(&zone->lock); 971 for (i = 0; i < count; ++i) { 972 struct page *page = __rmqueue(zone, order, migratetype); 973 if (unlikely(page == NULL)) 974 break; 975 976 /* 977 * Split buddy pages returned by expand() are received here 978 * in physical page order. The page is added to the callers and 979 * list and the list head then moves forward. From the callers 980 * perspective, the linked list is ordered by page number in 981 * some conditions. This is useful for IO devices that can 982 * merge IO requests if the physical pages are ordered 983 * properly. 984 */ 985 if (likely(cold == 0)) 986 list_add(&page->lru, list); 987 else 988 list_add_tail(&page->lru, list); 989 set_page_private(page, migratetype); 990 list = &page->lru; 991 } 992 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 993 spin_unlock(&zone->lock); 994 return i; 995} 996 997#ifdef CONFIG_NUMA 998/* 999 * Called from the vmstat counter updater to drain pagesets of this 1000 * currently executing processor on remote nodes after they have 1001 * expired. 1002 * 1003 * Note that this function must be called with the thread pinned to 1004 * a single processor. 1005 */ 1006void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1007{ 1008 unsigned long flags; 1009 int to_drain; 1010 1011 local_irq_save(flags); 1012 if (pcp->count >= pcp->batch) 1013 to_drain = pcp->batch; 1014 else 1015 to_drain = pcp->count; 1016 free_pcppages_bulk(zone, to_drain, pcp); 1017 pcp->count -= to_drain; 1018 local_irq_restore(flags); 1019} 1020#endif 1021 1022/* 1023 * Drain pages of the indicated processor. 1024 * 1025 * The processor must either be the current processor and the 1026 * thread pinned to the current processor or a processor that 1027 * is not online. 1028 */ 1029static void drain_pages(unsigned int cpu) 1030{ 1031 unsigned long flags; 1032 struct zone *zone; 1033 1034 for_each_populated_zone(zone) { 1035 struct per_cpu_pageset *pset; 1036 struct per_cpu_pages *pcp; 1037 1038 local_irq_save(flags); 1039 pset = per_cpu_ptr(zone->pageset, cpu); 1040 1041 pcp = &pset->pcp; 1042 free_pcppages_bulk(zone, pcp->count, pcp); 1043 pcp->count = 0; 1044 local_irq_restore(flags); 1045 } 1046} 1047 1048/* 1049 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1050 */ 1051void drain_local_pages(void *arg) 1052{ 1053 drain_pages(smp_processor_id()); 1054} 1055 1056/* 1057 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1058 */ 1059void drain_all_pages(void) 1060{ 1061 on_each_cpu(drain_local_pages, NULL, 1); 1062} 1063 1064#ifdef CONFIG_HIBERNATION 1065 1066void mark_free_pages(struct zone *zone) 1067{ 1068 unsigned long pfn, max_zone_pfn; 1069 unsigned long flags; 1070 int order, t; 1071 struct list_head *curr; 1072 1073 if (!zone->spanned_pages) 1074 return; 1075 1076 spin_lock_irqsave(&zone->lock, flags); 1077 1078 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1079 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1080 if (pfn_valid(pfn)) { 1081 struct page *page = pfn_to_page(pfn); 1082 1083 if (!swsusp_page_is_forbidden(page)) 1084 swsusp_unset_page_free(page); 1085 } 1086 1087 for_each_migratetype_order(order, t) { 1088 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1089 unsigned long i; 1090 1091 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1092 for (i = 0; i < (1UL << order); i++) 1093 swsusp_set_page_free(pfn_to_page(pfn + i)); 1094 } 1095 } 1096 spin_unlock_irqrestore(&zone->lock, flags); 1097} 1098#endif /* CONFIG_PM */ 1099 1100/* 1101 * Free a 0-order page 1102 * cold == 1 ? free a cold page : free a hot page 1103 */ 1104void free_hot_cold_page(struct page *page, int cold) 1105{ 1106 struct zone *zone = page_zone(page); 1107 struct per_cpu_pages *pcp; 1108 unsigned long flags; 1109 int migratetype; 1110 int wasMlocked = __TestClearPageMlocked(page); 1111 1112 trace_mm_page_free_direct(page, 0); 1113 kmemcheck_free_shadow(page, 0); 1114 1115 if (PageAnon(page)) 1116 page->mapping = NULL; 1117 if (free_pages_check(page)) 1118 return; 1119 1120 if (!PageHighMem(page)) { 1121 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1122 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1123 } 1124 arch_free_page(page, 0); 1125 kernel_map_pages(page, 1, 0); 1126 1127 migratetype = get_pageblock_migratetype(page); 1128 set_page_private(page, migratetype); 1129 local_irq_save(flags); 1130 if (unlikely(wasMlocked)) 1131 free_page_mlock(page); 1132 __count_vm_event(PGFREE); 1133 1134 /* 1135 * We only track unmovable, reclaimable and movable on pcp lists. 1136 * Free ISOLATE pages back to the allocator because they are being 1137 * offlined but treat RESERVE as movable pages so we can get those 1138 * areas back if necessary. Otherwise, we may have to free 1139 * excessively into the page allocator 1140 */ 1141 if (migratetype >= MIGRATE_PCPTYPES) { 1142 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1143 free_one_page(zone, page, 0, migratetype); 1144 goto out; 1145 } 1146 migratetype = MIGRATE_MOVABLE; 1147 } 1148 1149 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1150 if (cold) 1151 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1152 else 1153 list_add(&page->lru, &pcp->lists[migratetype]); 1154 pcp->count++; 1155 if (pcp->count >= pcp->high) { 1156 free_pcppages_bulk(zone, pcp->batch, pcp); 1157 pcp->count -= pcp->batch; 1158 } 1159 1160out: 1161 local_irq_restore(flags); 1162} 1163 1164/* 1165 * split_page takes a non-compound higher-order page, and splits it into 1166 * n (1<<order) sub-pages: page[0..n] 1167 * Each sub-page must be freed individually. 1168 * 1169 * Note: this is probably too low level an operation for use in drivers. 1170 * Please consult with lkml before using this in your driver. 1171 */ 1172void split_page(struct page *page, unsigned int order) 1173{ 1174 int i; 1175 1176 VM_BUG_ON(PageCompound(page)); 1177 VM_BUG_ON(!page_count(page)); 1178 1179#ifdef CONFIG_KMEMCHECK 1180 /* 1181 * Split shadow pages too, because free(page[0]) would 1182 * otherwise free the whole shadow. 1183 */ 1184 if (kmemcheck_page_is_tracked(page)) 1185 split_page(virt_to_page(page[0].shadow), order); 1186#endif 1187 1188 for (i = 1; i < (1 << order); i++) 1189 set_page_refcounted(page + i); 1190} 1191 1192/* 1193 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1194 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1195 * or two. 1196 */ 1197static inline 1198struct page *buffered_rmqueue(struct zone *preferred_zone, 1199 struct zone *zone, int order, gfp_t gfp_flags, 1200 int migratetype) 1201{ 1202 unsigned long flags; 1203 struct page *page; 1204 int cold = !!(gfp_flags & __GFP_COLD); 1205 1206again: 1207 if (likely(order == 0)) { 1208 struct per_cpu_pages *pcp; 1209 struct list_head *list; 1210 1211 local_irq_save(flags); 1212 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1213 list = &pcp->lists[migratetype]; 1214 if (list_empty(list)) { 1215 pcp->count += rmqueue_bulk(zone, 0, 1216 pcp->batch, list, 1217 migratetype, cold); 1218 if (unlikely(list_empty(list))) 1219 goto failed; 1220 } 1221 1222 if (cold) 1223 page = list_entry(list->prev, struct page, lru); 1224 else 1225 page = list_entry(list->next, struct page, lru); 1226 1227 list_del(&page->lru); 1228 pcp->count--; 1229 } else { 1230 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1231 /* 1232 * __GFP_NOFAIL is not to be used in new code. 1233 * 1234 * All __GFP_NOFAIL callers should be fixed so that they 1235 * properly detect and handle allocation failures. 1236 * 1237 * We most definitely don't want callers attempting to 1238 * allocate greater than order-1 page units with 1239 * __GFP_NOFAIL. 1240 */ 1241 WARN_ON_ONCE(order > 1); 1242 } 1243 spin_lock_irqsave(&zone->lock, flags); 1244 page = __rmqueue(zone, order, migratetype); 1245 spin_unlock(&zone->lock); 1246 if (!page) 1247 goto failed; 1248 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1249 } 1250 1251 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1252 zone_statistics(preferred_zone, zone); 1253 local_irq_restore(flags); 1254 1255 VM_BUG_ON(bad_range(zone, page)); 1256 if (prep_new_page(page, order, gfp_flags)) 1257 goto again; 1258 return page; 1259 1260failed: 1261 local_irq_restore(flags); 1262 return NULL; 1263} 1264 1265/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1266#define ALLOC_WMARK_MIN WMARK_MIN 1267#define ALLOC_WMARK_LOW WMARK_LOW 1268#define ALLOC_WMARK_HIGH WMARK_HIGH 1269#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1270 1271/* Mask to get the watermark bits */ 1272#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1273 1274#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1275#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1276#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1277 1278#ifdef CONFIG_FAIL_PAGE_ALLOC 1279 1280static struct fail_page_alloc_attr { 1281 struct fault_attr attr; 1282 1283 u32 ignore_gfp_highmem; 1284 u32 ignore_gfp_wait; 1285 u32 min_order; 1286 1287#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1288 1289 struct dentry *ignore_gfp_highmem_file; 1290 struct dentry *ignore_gfp_wait_file; 1291 struct dentry *min_order_file; 1292 1293#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1294 1295} fail_page_alloc = { 1296 .attr = FAULT_ATTR_INITIALIZER, 1297 .ignore_gfp_wait = 1, 1298 .ignore_gfp_highmem = 1, 1299 .min_order = 1, 1300}; 1301 1302static int __init setup_fail_page_alloc(char *str) 1303{ 1304 return setup_fault_attr(&fail_page_alloc.attr, str); 1305} 1306__setup("fail_page_alloc=", setup_fail_page_alloc); 1307 1308static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1309{ 1310 if (order < fail_page_alloc.min_order) 1311 return 0; 1312 if (gfp_mask & __GFP_NOFAIL) 1313 return 0; 1314 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1315 return 0; 1316 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1317 return 0; 1318 1319 return should_fail(&fail_page_alloc.attr, 1 << order); 1320} 1321 1322#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1323 1324static int __init fail_page_alloc_debugfs(void) 1325{ 1326 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1327 struct dentry *dir; 1328 int err; 1329 1330 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1331 "fail_page_alloc"); 1332 if (err) 1333 return err; 1334 dir = fail_page_alloc.attr.dentries.dir; 1335 1336 fail_page_alloc.ignore_gfp_wait_file = 1337 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1338 &fail_page_alloc.ignore_gfp_wait); 1339 1340 fail_page_alloc.ignore_gfp_highmem_file = 1341 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1342 &fail_page_alloc.ignore_gfp_highmem); 1343 fail_page_alloc.min_order_file = 1344 debugfs_create_u32("min-order", mode, dir, 1345 &fail_page_alloc.min_order); 1346 1347 if (!fail_page_alloc.ignore_gfp_wait_file || 1348 !fail_page_alloc.ignore_gfp_highmem_file || 1349 !fail_page_alloc.min_order_file) { 1350 err = -ENOMEM; 1351 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1352 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1353 debugfs_remove(fail_page_alloc.min_order_file); 1354 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1355 } 1356 1357 return err; 1358} 1359 1360late_initcall(fail_page_alloc_debugfs); 1361 1362#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1363 1364#else /* CONFIG_FAIL_PAGE_ALLOC */ 1365 1366static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1367{ 1368 return 0; 1369} 1370 1371#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1372 1373/* 1374 * Return 1 if free pages are above 'mark'. This takes into account the order 1375 * of the allocation. 1376 */ 1377int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1378 int classzone_idx, int alloc_flags) 1379{ 1380 /* free_pages my go negative - that's OK */ 1381 long min = mark; 1382 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1383 int o; 1384 1385 if (alloc_flags & ALLOC_HIGH) 1386 min -= min / 2; 1387 if (alloc_flags & ALLOC_HARDER) 1388 min -= min / 4; 1389 1390 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1391 return 0; 1392 for (o = 0; o < order; o++) { 1393 /* At the next order, this order's pages become unavailable */ 1394 free_pages -= z->free_area[o].nr_free << o; 1395 1396 /* Require fewer higher order pages to be free */ 1397 min >>= 1; 1398 1399 if (free_pages <= min) 1400 return 0; 1401 } 1402 return 1; 1403} 1404 1405#ifdef CONFIG_NUMA 1406/* 1407 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1408 * skip over zones that are not allowed by the cpuset, or that have 1409 * been recently (in last second) found to be nearly full. See further 1410 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1411 * that have to skip over a lot of full or unallowed zones. 1412 * 1413 * If the zonelist cache is present in the passed in zonelist, then 1414 * returns a pointer to the allowed node mask (either the current 1415 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1416 * 1417 * If the zonelist cache is not available for this zonelist, does 1418 * nothing and returns NULL. 1419 * 1420 * If the fullzones BITMAP in the zonelist cache is stale (more than 1421 * a second since last zap'd) then we zap it out (clear its bits.) 1422 * 1423 * We hold off even calling zlc_setup, until after we've checked the 1424 * first zone in the zonelist, on the theory that most allocations will 1425 * be satisfied from that first zone, so best to examine that zone as 1426 * quickly as we can. 1427 */ 1428static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1429{ 1430 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1431 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1432 1433 zlc = zonelist->zlcache_ptr; 1434 if (!zlc) 1435 return NULL; 1436 1437 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1438 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1439 zlc->last_full_zap = jiffies; 1440 } 1441 1442 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1443 &cpuset_current_mems_allowed : 1444 &node_states[N_HIGH_MEMORY]; 1445 return allowednodes; 1446} 1447 1448/* 1449 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1450 * if it is worth looking at further for free memory: 1451 * 1) Check that the zone isn't thought to be full (doesn't have its 1452 * bit set in the zonelist_cache fullzones BITMAP). 1453 * 2) Check that the zones node (obtained from the zonelist_cache 1454 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1455 * Return true (non-zero) if zone is worth looking at further, or 1456 * else return false (zero) if it is not. 1457 * 1458 * This check -ignores- the distinction between various watermarks, 1459 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1460 * found to be full for any variation of these watermarks, it will 1461 * be considered full for up to one second by all requests, unless 1462 * we are so low on memory on all allowed nodes that we are forced 1463 * into the second scan of the zonelist. 1464 * 1465 * In the second scan we ignore this zonelist cache and exactly 1466 * apply the watermarks to all zones, even it is slower to do so. 1467 * We are low on memory in the second scan, and should leave no stone 1468 * unturned looking for a free page. 1469 */ 1470static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1471 nodemask_t *allowednodes) 1472{ 1473 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1474 int i; /* index of *z in zonelist zones */ 1475 int n; /* node that zone *z is on */ 1476 1477 zlc = zonelist->zlcache_ptr; 1478 if (!zlc) 1479 return 1; 1480 1481 i = z - zonelist->_zonerefs; 1482 n = zlc->z_to_n[i]; 1483 1484 /* This zone is worth trying if it is allowed but not full */ 1485 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1486} 1487 1488/* 1489 * Given 'z' scanning a zonelist, set the corresponding bit in 1490 * zlc->fullzones, so that subsequent attempts to allocate a page 1491 * from that zone don't waste time re-examining it. 1492 */ 1493static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1494{ 1495 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1496 int i; /* index of *z in zonelist zones */ 1497 1498 zlc = zonelist->zlcache_ptr; 1499 if (!zlc) 1500 return; 1501 1502 i = z - zonelist->_zonerefs; 1503 1504 set_bit(i, zlc->fullzones); 1505} 1506 1507#else /* CONFIG_NUMA */ 1508 1509static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1510{ 1511 return NULL; 1512} 1513 1514static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1515 nodemask_t *allowednodes) 1516{ 1517 return 1; 1518} 1519 1520static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1521{ 1522} 1523#endif /* CONFIG_NUMA */ 1524 1525/* 1526 * get_page_from_freelist goes through the zonelist trying to allocate 1527 * a page. 1528 */ 1529static struct page * 1530get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1531 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1532 struct zone *preferred_zone, int migratetype) 1533{ 1534 struct zoneref *z; 1535 struct page *page = NULL; 1536 int classzone_idx; 1537 struct zone *zone; 1538 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1539 int zlc_active = 0; /* set if using zonelist_cache */ 1540 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1541 1542 classzone_idx = zone_idx(preferred_zone); 1543zonelist_scan: 1544 /* 1545 * Scan zonelist, looking for a zone with enough free. 1546 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1547 */ 1548 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1549 high_zoneidx, nodemask) { 1550 if (NUMA_BUILD && zlc_active && 1551 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1552 continue; 1553 if ((alloc_flags & ALLOC_CPUSET) && 1554 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1555 goto try_next_zone; 1556 1557 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1558 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1559 unsigned long mark; 1560 int ret; 1561 1562 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1563 if (zone_watermark_ok(zone, order, mark, 1564 classzone_idx, alloc_flags)) 1565 goto try_this_zone; 1566 1567 if (zone_reclaim_mode == 0) 1568 goto this_zone_full; 1569 1570 ret = zone_reclaim(zone, gfp_mask, order); 1571 switch (ret) { 1572 case ZONE_RECLAIM_NOSCAN: 1573 /* did not scan */ 1574 goto try_next_zone; 1575 case ZONE_RECLAIM_FULL: 1576 /* scanned but unreclaimable */ 1577 goto this_zone_full; 1578 default: 1579 /* did we reclaim enough */ 1580 if (!zone_watermark_ok(zone, order, mark, 1581 classzone_idx, alloc_flags)) 1582 goto this_zone_full; 1583 } 1584 } 1585 1586try_this_zone: 1587 page = buffered_rmqueue(preferred_zone, zone, order, 1588 gfp_mask, migratetype); 1589 if (page) 1590 break; 1591this_zone_full: 1592 if (NUMA_BUILD) 1593 zlc_mark_zone_full(zonelist, z); 1594try_next_zone: 1595 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1596 /* 1597 * we do zlc_setup after the first zone is tried but only 1598 * if there are multiple nodes make it worthwhile 1599 */ 1600 allowednodes = zlc_setup(zonelist, alloc_flags); 1601 zlc_active = 1; 1602 did_zlc_setup = 1; 1603 } 1604 } 1605 1606 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1607 /* Disable zlc cache for second zonelist scan */ 1608 zlc_active = 0; 1609 goto zonelist_scan; 1610 } 1611 return page; 1612} 1613 1614static inline int 1615should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1616 unsigned long pages_reclaimed) 1617{ 1618 /* Do not loop if specifically requested */ 1619 if (gfp_mask & __GFP_NORETRY) 1620 return 0; 1621 1622 /* 1623 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1624 * means __GFP_NOFAIL, but that may not be true in other 1625 * implementations. 1626 */ 1627 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1628 return 1; 1629 1630 /* 1631 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1632 * specified, then we retry until we no longer reclaim any pages 1633 * (above), or we've reclaimed an order of pages at least as 1634 * large as the allocation's order. In both cases, if the 1635 * allocation still fails, we stop retrying. 1636 */ 1637 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1638 return 1; 1639 1640 /* 1641 * Don't let big-order allocations loop unless the caller 1642 * explicitly requests that. 1643 */ 1644 if (gfp_mask & __GFP_NOFAIL) 1645 return 1; 1646 1647 return 0; 1648} 1649 1650static inline struct page * 1651__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1652 struct zonelist *zonelist, enum zone_type high_zoneidx, 1653 nodemask_t *nodemask, struct zone *preferred_zone, 1654 int migratetype) 1655{ 1656 struct page *page; 1657 1658 /* Acquire the OOM killer lock for the zones in zonelist */ 1659 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1660 schedule_timeout_uninterruptible(1); 1661 return NULL; 1662 } 1663 1664 /* 1665 * Go through the zonelist yet one more time, keep very high watermark 1666 * here, this is only to catch a parallel oom killing, we must fail if 1667 * we're still under heavy pressure. 1668 */ 1669 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1670 order, zonelist, high_zoneidx, 1671 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1672 preferred_zone, migratetype); 1673 if (page) 1674 goto out; 1675 1676 if (!(gfp_mask & __GFP_NOFAIL)) { 1677 /* The OOM killer will not help higher order allocs */ 1678 if (order > PAGE_ALLOC_COSTLY_ORDER) 1679 goto out; 1680 /* 1681 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1682 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1683 * The caller should handle page allocation failure by itself if 1684 * it specifies __GFP_THISNODE. 1685 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1686 */ 1687 if (gfp_mask & __GFP_THISNODE) 1688 goto out; 1689 } 1690 /* Exhausted what can be done so it's blamo time */ 1691 out_of_memory(zonelist, gfp_mask, order, nodemask); 1692 1693out: 1694 clear_zonelist_oom(zonelist, gfp_mask); 1695 return page; 1696} 1697 1698/* The really slow allocator path where we enter direct reclaim */ 1699static inline struct page * 1700__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1701 struct zonelist *zonelist, enum zone_type high_zoneidx, 1702 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1703 int migratetype, unsigned long *did_some_progress) 1704{ 1705 struct page *page = NULL; 1706 struct reclaim_state reclaim_state; 1707 struct task_struct *p = current; 1708 1709 cond_resched(); 1710 1711 /* We now go into synchronous reclaim */ 1712 cpuset_memory_pressure_bump(); 1713 p->flags |= PF_MEMALLOC; 1714 lockdep_set_current_reclaim_state(gfp_mask); 1715 reclaim_state.reclaimed_slab = 0; 1716 p->reclaim_state = &reclaim_state; 1717 1718 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1719 1720 p->reclaim_state = NULL; 1721 lockdep_clear_current_reclaim_state(); 1722 p->flags &= ~PF_MEMALLOC; 1723 1724 cond_resched(); 1725 1726 if (order != 0) 1727 drain_all_pages(); 1728 1729 if (likely(*did_some_progress)) 1730 page = get_page_from_freelist(gfp_mask, nodemask, order, 1731 zonelist, high_zoneidx, 1732 alloc_flags, preferred_zone, 1733 migratetype); 1734 return page; 1735} 1736 1737/* 1738 * This is called in the allocator slow-path if the allocation request is of 1739 * sufficient urgency to ignore watermarks and take other desperate measures 1740 */ 1741static inline struct page * 1742__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1743 struct zonelist *zonelist, enum zone_type high_zoneidx, 1744 nodemask_t *nodemask, struct zone *preferred_zone, 1745 int migratetype) 1746{ 1747 struct page *page; 1748 1749 do { 1750 page = get_page_from_freelist(gfp_mask, nodemask, order, 1751 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1752 preferred_zone, migratetype); 1753 1754 if (!page && gfp_mask & __GFP_NOFAIL) 1755 congestion_wait(BLK_RW_ASYNC, HZ/50); 1756 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1757 1758 return page; 1759} 1760 1761static inline 1762void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1763 enum zone_type high_zoneidx) 1764{ 1765 struct zoneref *z; 1766 struct zone *zone; 1767 1768 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1769 wakeup_kswapd(zone, order); 1770} 1771 1772static inline int 1773gfp_to_alloc_flags(gfp_t gfp_mask) 1774{ 1775 struct task_struct *p = current; 1776 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1777 const gfp_t wait = gfp_mask & __GFP_WAIT; 1778 1779 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1780 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1781 1782 /* 1783 * The caller may dip into page reserves a bit more if the caller 1784 * cannot run direct reclaim, or if the caller has realtime scheduling 1785 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1786 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1787 */ 1788 alloc_flags |= (gfp_mask & __GFP_HIGH); 1789 1790 if (!wait) { 1791 alloc_flags |= ALLOC_HARDER; 1792 /* 1793 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1794 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1795 */ 1796 alloc_flags &= ~ALLOC_CPUSET; 1797 } else if (unlikely(rt_task(p)) && !in_interrupt()) 1798 alloc_flags |= ALLOC_HARDER; 1799 1800 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1801 if (!in_interrupt() && 1802 ((p->flags & PF_MEMALLOC) || 1803 unlikely(test_thread_flag(TIF_MEMDIE)))) 1804 alloc_flags |= ALLOC_NO_WATERMARKS; 1805 } 1806 1807 return alloc_flags; 1808} 1809 1810static inline struct page * 1811__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1812 struct zonelist *zonelist, enum zone_type high_zoneidx, 1813 nodemask_t *nodemask, struct zone *preferred_zone, 1814 int migratetype) 1815{ 1816 const gfp_t wait = gfp_mask & __GFP_WAIT; 1817 struct page *page = NULL; 1818 int alloc_flags; 1819 unsigned long pages_reclaimed = 0; 1820 unsigned long did_some_progress; 1821 struct task_struct *p = current; 1822 1823 /* 1824 * In the slowpath, we sanity check order to avoid ever trying to 1825 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1826 * be using allocators in order of preference for an area that is 1827 * too large. 1828 */ 1829 if (order >= MAX_ORDER) { 1830 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1831 return NULL; 1832 } 1833 1834 /* 1835 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1836 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1837 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1838 * using a larger set of nodes after it has established that the 1839 * allowed per node queues are empty and that nodes are 1840 * over allocated. 1841 */ 1842 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1843 goto nopage; 1844 1845restart: 1846 wake_all_kswapd(order, zonelist, high_zoneidx); 1847 1848 /* 1849 * OK, we're below the kswapd watermark and have kicked background 1850 * reclaim. Now things get more complex, so set up alloc_flags according 1851 * to how we want to proceed. 1852 */ 1853 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1854 1855 /* This is the last chance, in general, before the goto nopage. */ 1856 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1857 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1858 preferred_zone, migratetype); 1859 if (page) 1860 goto got_pg; 1861 1862rebalance: 1863 /* Allocate without watermarks if the context allows */ 1864 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1865 page = __alloc_pages_high_priority(gfp_mask, order, 1866 zonelist, high_zoneidx, nodemask, 1867 preferred_zone, migratetype); 1868 if (page) 1869 goto got_pg; 1870 } 1871 1872 /* Atomic allocations - we can't balance anything */ 1873 if (!wait) 1874 goto nopage; 1875 1876 /* Avoid recursion of direct reclaim */ 1877 if (p->flags & PF_MEMALLOC) 1878 goto nopage; 1879 1880 /* Avoid allocations with no watermarks from looping endlessly */ 1881 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 1882 goto nopage; 1883 1884 /* Try direct reclaim and then allocating */ 1885 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1886 zonelist, high_zoneidx, 1887 nodemask, 1888 alloc_flags, preferred_zone, 1889 migratetype, &did_some_progress); 1890 if (page) 1891 goto got_pg; 1892 1893 /* 1894 * If we failed to make any progress reclaiming, then we are 1895 * running out of options and have to consider going OOM 1896 */ 1897 if (!did_some_progress) { 1898 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1899 if (oom_killer_disabled) 1900 goto nopage; 1901 page = __alloc_pages_may_oom(gfp_mask, order, 1902 zonelist, high_zoneidx, 1903 nodemask, preferred_zone, 1904 migratetype); 1905 if (page) 1906 goto got_pg; 1907 1908 /* 1909 * The OOM killer does not trigger for high-order 1910 * ~__GFP_NOFAIL allocations so if no progress is being 1911 * made, there are no other options and retrying is 1912 * unlikely to help. 1913 */ 1914 if (order > PAGE_ALLOC_COSTLY_ORDER && 1915 !(gfp_mask & __GFP_NOFAIL)) 1916 goto nopage; 1917 1918 goto restart; 1919 } 1920 } 1921 1922 /* Check if we should retry the allocation */ 1923 pages_reclaimed += did_some_progress; 1924 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1925 /* Wait for some write requests to complete then retry */ 1926 congestion_wait(BLK_RW_ASYNC, HZ/50); 1927 goto rebalance; 1928 } 1929 1930nopage: 1931 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1932 printk(KERN_WARNING "%s: page allocation failure." 1933 " order:%d, mode:0x%x\n", 1934 p->comm, order, gfp_mask); 1935 dump_stack(); 1936 show_mem(); 1937 } 1938 return page; 1939got_pg: 1940 if (kmemcheck_enabled) 1941 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 1942 return page; 1943 1944} 1945 1946/* 1947 * This is the 'heart' of the zoned buddy allocator. 1948 */ 1949struct page * 1950__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1951 struct zonelist *zonelist, nodemask_t *nodemask) 1952{ 1953 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1954 struct zone *preferred_zone; 1955 struct page *page; 1956 int migratetype = allocflags_to_migratetype(gfp_mask); 1957 1958 gfp_mask &= gfp_allowed_mask; 1959 1960 lockdep_trace_alloc(gfp_mask); 1961 1962 might_sleep_if(gfp_mask & __GFP_WAIT); 1963 1964 if (should_fail_alloc_page(gfp_mask, order)) 1965 return NULL; 1966 1967 /* 1968 * Check the zones suitable for the gfp_mask contain at least one 1969 * valid zone. It's possible to have an empty zonelist as a result 1970 * of GFP_THISNODE and a memoryless node 1971 */ 1972 if (unlikely(!zonelist->_zonerefs->zone)) 1973 return NULL; 1974 1975 /* The preferred zone is used for statistics later */ 1976 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1977 if (!preferred_zone) 1978 return NULL; 1979 1980 /* First allocation attempt */ 1981 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1982 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1983 preferred_zone, migratetype); 1984 if (unlikely(!page)) 1985 page = __alloc_pages_slowpath(gfp_mask, order, 1986 zonelist, high_zoneidx, nodemask, 1987 preferred_zone, migratetype); 1988 1989 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 1990 return page; 1991} 1992EXPORT_SYMBOL(__alloc_pages_nodemask); 1993 1994/* 1995 * Common helper functions. 1996 */ 1997unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1998{ 1999 struct page *page; 2000 2001 /* 2002 * __get_free_pages() returns a 32-bit address, which cannot represent 2003 * a highmem page 2004 */ 2005 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2006 2007 page = alloc_pages(gfp_mask, order); 2008 if (!page) 2009 return 0; 2010 return (unsigned long) page_address(page); 2011} 2012EXPORT_SYMBOL(__get_free_pages); 2013 2014unsigned long get_zeroed_page(gfp_t gfp_mask) 2015{ 2016 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2017} 2018EXPORT_SYMBOL(get_zeroed_page); 2019 2020void __pagevec_free(struct pagevec *pvec) 2021{ 2022 int i = pagevec_count(pvec); 2023 2024 while (--i >= 0) { 2025 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2026 free_hot_cold_page(pvec->pages[i], pvec->cold); 2027 } 2028} 2029 2030void __free_pages(struct page *page, unsigned int order) 2031{ 2032 if (put_page_testzero(page)) { 2033 if (order == 0) 2034 free_hot_cold_page(page, 0); 2035 else 2036 __free_pages_ok(page, order); 2037 } 2038} 2039 2040EXPORT_SYMBOL(__free_pages); 2041 2042void free_pages(unsigned long addr, unsigned int order) 2043{ 2044 if (addr != 0) { 2045 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2046 __free_pages(virt_to_page((void *)addr), order); 2047 } 2048} 2049 2050EXPORT_SYMBOL(free_pages); 2051 2052/** 2053 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2054 * @size: the number of bytes to allocate 2055 * @gfp_mask: GFP flags for the allocation 2056 * 2057 * This function is similar to alloc_pages(), except that it allocates the 2058 * minimum number of pages to satisfy the request. alloc_pages() can only 2059 * allocate memory in power-of-two pages. 2060 * 2061 * This function is also limited by MAX_ORDER. 2062 * 2063 * Memory allocated by this function must be released by free_pages_exact(). 2064 */ 2065void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2066{ 2067 unsigned int order = get_order(size); 2068 unsigned long addr; 2069 2070 addr = __get_free_pages(gfp_mask, order); 2071 if (addr) { 2072 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2073 unsigned long used = addr + PAGE_ALIGN(size); 2074 2075 split_page(virt_to_page((void *)addr), order); 2076 while (used < alloc_end) { 2077 free_page(used); 2078 used += PAGE_SIZE; 2079 } 2080 } 2081 2082 return (void *)addr; 2083} 2084EXPORT_SYMBOL(alloc_pages_exact); 2085 2086/** 2087 * free_pages_exact - release memory allocated via alloc_pages_exact() 2088 * @virt: the value returned by alloc_pages_exact. 2089 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2090 * 2091 * Release the memory allocated by a previous call to alloc_pages_exact. 2092 */ 2093void free_pages_exact(void *virt, size_t size) 2094{ 2095 unsigned long addr = (unsigned long)virt; 2096 unsigned long end = addr + PAGE_ALIGN(size); 2097 2098 while (addr < end) { 2099 free_page(addr); 2100 addr += PAGE_SIZE; 2101 } 2102} 2103EXPORT_SYMBOL(free_pages_exact); 2104 2105static unsigned int nr_free_zone_pages(int offset) 2106{ 2107 struct zoneref *z; 2108 struct zone *zone; 2109 2110 /* Just pick one node, since fallback list is circular */ 2111 unsigned int sum = 0; 2112 2113 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2114 2115 for_each_zone_zonelist(zone, z, zonelist, offset) { 2116 unsigned long size = zone->present_pages; 2117 unsigned long high = high_wmark_pages(zone); 2118 if (size > high) 2119 sum += size - high; 2120 } 2121 2122 return sum; 2123} 2124 2125/* 2126 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2127 */ 2128unsigned int nr_free_buffer_pages(void) 2129{ 2130 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2131} 2132EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2133 2134/* 2135 * Amount of free RAM allocatable within all zones 2136 */ 2137unsigned int nr_free_pagecache_pages(void) 2138{ 2139 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2140} 2141 2142static inline void show_node(struct zone *zone) 2143{ 2144 if (NUMA_BUILD) 2145 printk("Node %d ", zone_to_nid(zone)); 2146} 2147 2148void si_meminfo(struct sysinfo *val) 2149{ 2150 val->totalram = totalram_pages; 2151 val->sharedram = 0; 2152 val->freeram = global_page_state(NR_FREE_PAGES); 2153 val->bufferram = nr_blockdev_pages(); 2154 val->totalhigh = totalhigh_pages; 2155 val->freehigh = nr_free_highpages(); 2156 val->mem_unit = PAGE_SIZE; 2157} 2158 2159EXPORT_SYMBOL(si_meminfo); 2160 2161#ifdef CONFIG_NUMA 2162void si_meminfo_node(struct sysinfo *val, int nid) 2163{ 2164 pg_data_t *pgdat = NODE_DATA(nid); 2165 2166 val->totalram = pgdat->node_present_pages; 2167 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2168#ifdef CONFIG_HIGHMEM 2169 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2170 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2171 NR_FREE_PAGES); 2172#else 2173 val->totalhigh = 0; 2174 val->freehigh = 0; 2175#endif 2176 val->mem_unit = PAGE_SIZE; 2177} 2178#endif 2179 2180#define K(x) ((x) << (PAGE_SHIFT-10)) 2181 2182/* 2183 * Show free area list (used inside shift_scroll-lock stuff) 2184 * We also calculate the percentage fragmentation. We do this by counting the 2185 * memory on each free list with the exception of the first item on the list. 2186 */ 2187void show_free_areas(void) 2188{ 2189 int cpu; 2190 struct zone *zone; 2191 2192 for_each_populated_zone(zone) { 2193 show_node(zone); 2194 printk("%s per-cpu:\n", zone->name); 2195 2196 for_each_online_cpu(cpu) { 2197 struct per_cpu_pageset *pageset; 2198 2199 pageset = per_cpu_ptr(zone->pageset, cpu); 2200 2201 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2202 cpu, pageset->pcp.high, 2203 pageset->pcp.batch, pageset->pcp.count); 2204 } 2205 } 2206 2207 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2208 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2209 " unevictable:%lu" 2210 " dirty:%lu writeback:%lu unstable:%lu\n" 2211 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2212 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2213 global_page_state(NR_ACTIVE_ANON), 2214 global_page_state(NR_INACTIVE_ANON), 2215 global_page_state(NR_ISOLATED_ANON), 2216 global_page_state(NR_ACTIVE_FILE), 2217 global_page_state(NR_INACTIVE_FILE), 2218 global_page_state(NR_ISOLATED_FILE), 2219 global_page_state(NR_UNEVICTABLE), 2220 global_page_state(NR_FILE_DIRTY), 2221 global_page_state(NR_WRITEBACK), 2222 global_page_state(NR_UNSTABLE_NFS), 2223 global_page_state(NR_FREE_PAGES), 2224 global_page_state(NR_SLAB_RECLAIMABLE), 2225 global_page_state(NR_SLAB_UNRECLAIMABLE), 2226 global_page_state(NR_FILE_MAPPED), 2227 global_page_state(NR_SHMEM), 2228 global_page_state(NR_PAGETABLE), 2229 global_page_state(NR_BOUNCE)); 2230 2231 for_each_populated_zone(zone) { 2232 int i; 2233 2234 show_node(zone); 2235 printk("%s" 2236 " free:%lukB" 2237 " min:%lukB" 2238 " low:%lukB" 2239 " high:%lukB" 2240 " active_anon:%lukB" 2241 " inactive_anon:%lukB" 2242 " active_file:%lukB" 2243 " inactive_file:%lukB" 2244 " unevictable:%lukB" 2245 " isolated(anon):%lukB" 2246 " isolated(file):%lukB" 2247 " present:%lukB" 2248 " mlocked:%lukB" 2249 " dirty:%lukB" 2250 " writeback:%lukB" 2251 " mapped:%lukB" 2252 " shmem:%lukB" 2253 " slab_reclaimable:%lukB" 2254 " slab_unreclaimable:%lukB" 2255 " kernel_stack:%lukB" 2256 " pagetables:%lukB" 2257 " unstable:%lukB" 2258 " bounce:%lukB" 2259 " writeback_tmp:%lukB" 2260 " pages_scanned:%lu" 2261 " all_unreclaimable? %s" 2262 "\n", 2263 zone->name, 2264 K(zone_page_state(zone, NR_FREE_PAGES)), 2265 K(min_wmark_pages(zone)), 2266 K(low_wmark_pages(zone)), 2267 K(high_wmark_pages(zone)), 2268 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2269 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2270 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2271 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2272 K(zone_page_state(zone, NR_UNEVICTABLE)), 2273 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2274 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2275 K(zone->present_pages), 2276 K(zone_page_state(zone, NR_MLOCK)), 2277 K(zone_page_state(zone, NR_FILE_DIRTY)), 2278 K(zone_page_state(zone, NR_WRITEBACK)), 2279 K(zone_page_state(zone, NR_FILE_MAPPED)), 2280 K(zone_page_state(zone, NR_SHMEM)), 2281 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2282 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2283 zone_page_state(zone, NR_KERNEL_STACK) * 2284 THREAD_SIZE / 1024, 2285 K(zone_page_state(zone, NR_PAGETABLE)), 2286 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2287 K(zone_page_state(zone, NR_BOUNCE)), 2288 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2289 zone->pages_scanned, 2290 (zone->all_unreclaimable ? "yes" : "no") 2291 ); 2292 printk("lowmem_reserve[]:"); 2293 for (i = 0; i < MAX_NR_ZONES; i++) 2294 printk(" %lu", zone->lowmem_reserve[i]); 2295 printk("\n"); 2296 } 2297 2298 for_each_populated_zone(zone) { 2299 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2300 2301 show_node(zone); 2302 printk("%s: ", zone->name); 2303 2304 spin_lock_irqsave(&zone->lock, flags); 2305 for (order = 0; order < MAX_ORDER; order++) { 2306 nr[order] = zone->free_area[order].nr_free; 2307 total += nr[order] << order; 2308 } 2309 spin_unlock_irqrestore(&zone->lock, flags); 2310 for (order = 0; order < MAX_ORDER; order++) 2311 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2312 printk("= %lukB\n", K(total)); 2313 } 2314 2315 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2316 2317 show_swap_cache_info(); 2318} 2319 2320static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2321{ 2322 zoneref->zone = zone; 2323 zoneref->zone_idx = zone_idx(zone); 2324} 2325 2326/* 2327 * Builds allocation fallback zone lists. 2328 * 2329 * Add all populated zones of a node to the zonelist. 2330 */ 2331static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2332 int nr_zones, enum zone_type zone_type) 2333{ 2334 struct zone *zone; 2335 2336 BUG_ON(zone_type >= MAX_NR_ZONES); 2337 zone_type++; 2338 2339 do { 2340 zone_type--; 2341 zone = pgdat->node_zones + zone_type; 2342 if (populated_zone(zone)) { 2343 zoneref_set_zone(zone, 2344 &zonelist->_zonerefs[nr_zones++]); 2345 check_highest_zone(zone_type); 2346 } 2347 2348 } while (zone_type); 2349 return nr_zones; 2350} 2351 2352 2353/* 2354 * zonelist_order: 2355 * 0 = automatic detection of better ordering. 2356 * 1 = order by ([node] distance, -zonetype) 2357 * 2 = order by (-zonetype, [node] distance) 2358 * 2359 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2360 * the same zonelist. So only NUMA can configure this param. 2361 */ 2362#define ZONELIST_ORDER_DEFAULT 0 2363#define ZONELIST_ORDER_NODE 1 2364#define ZONELIST_ORDER_ZONE 2 2365 2366/* zonelist order in the kernel. 2367 * set_zonelist_order() will set this to NODE or ZONE. 2368 */ 2369static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2370static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2371 2372 2373#ifdef CONFIG_NUMA 2374/* The value user specified ....changed by config */ 2375static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2376/* string for sysctl */ 2377#define NUMA_ZONELIST_ORDER_LEN 16 2378char numa_zonelist_order[16] = "default"; 2379 2380/* 2381 * interface for configure zonelist ordering. 2382 * command line option "numa_zonelist_order" 2383 * = "[dD]efault - default, automatic configuration. 2384 * = "[nN]ode - order by node locality, then by zone within node 2385 * = "[zZ]one - order by zone, then by locality within zone 2386 */ 2387 2388static int __parse_numa_zonelist_order(char *s) 2389{ 2390 if (*s == 'd' || *s == 'D') { 2391 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2392 } else if (*s == 'n' || *s == 'N') { 2393 user_zonelist_order = ZONELIST_ORDER_NODE; 2394 } else if (*s == 'z' || *s == 'Z') { 2395 user_zonelist_order = ZONELIST_ORDER_ZONE; 2396 } else { 2397 printk(KERN_WARNING 2398 "Ignoring invalid numa_zonelist_order value: " 2399 "%s\n", s); 2400 return -EINVAL; 2401 } 2402 return 0; 2403} 2404 2405static __init int setup_numa_zonelist_order(char *s) 2406{ 2407 if (s) 2408 return __parse_numa_zonelist_order(s); 2409 return 0; 2410} 2411early_param("numa_zonelist_order", setup_numa_zonelist_order); 2412 2413/* 2414 * sysctl handler for numa_zonelist_order 2415 */ 2416int numa_zonelist_order_handler(ctl_table *table, int write, 2417 void __user *buffer, size_t *length, 2418 loff_t *ppos) 2419{ 2420 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2421 int ret; 2422 static DEFINE_MUTEX(zl_order_mutex); 2423 2424 mutex_lock(&zl_order_mutex); 2425 if (write) 2426 strcpy(saved_string, (char*)table->data); 2427 ret = proc_dostring(table, write, buffer, length, ppos); 2428 if (ret) 2429 goto out; 2430 if (write) { 2431 int oldval = user_zonelist_order; 2432 if (__parse_numa_zonelist_order((char*)table->data)) { 2433 /* 2434 * bogus value. restore saved string 2435 */ 2436 strncpy((char*)table->data, saved_string, 2437 NUMA_ZONELIST_ORDER_LEN); 2438 user_zonelist_order = oldval; 2439 } else if (oldval != user_zonelist_order) 2440 build_all_zonelists(); 2441 } 2442out: 2443 mutex_unlock(&zl_order_mutex); 2444 return ret; 2445} 2446 2447 2448#define MAX_NODE_LOAD (nr_online_nodes) 2449static int node_load[MAX_NUMNODES]; 2450 2451/** 2452 * find_next_best_node - find the next node that should appear in a given node's fallback list 2453 * @node: node whose fallback list we're appending 2454 * @used_node_mask: nodemask_t of already used nodes 2455 * 2456 * We use a number of factors to determine which is the next node that should 2457 * appear on a given node's fallback list. The node should not have appeared 2458 * already in @node's fallback list, and it should be the next closest node 2459 * according to the distance array (which contains arbitrary distance values 2460 * from each node to each node in the system), and should also prefer nodes 2461 * with no CPUs, since presumably they'll have very little allocation pressure 2462 * on them otherwise. 2463 * It returns -1 if no node is found. 2464 */ 2465static int find_next_best_node(int node, nodemask_t *used_node_mask) 2466{ 2467 int n, val; 2468 int min_val = INT_MAX; 2469 int best_node = -1; 2470 const struct cpumask *tmp = cpumask_of_node(0); 2471 2472 /* Use the local node if we haven't already */ 2473 if (!node_isset(node, *used_node_mask)) { 2474 node_set(node, *used_node_mask); 2475 return node; 2476 } 2477 2478 for_each_node_state(n, N_HIGH_MEMORY) { 2479 2480 /* Don't want a node to appear more than once */ 2481 if (node_isset(n, *used_node_mask)) 2482 continue; 2483 2484 /* Use the distance array to find the distance */ 2485 val = node_distance(node, n); 2486 2487 /* Penalize nodes under us ("prefer the next node") */ 2488 val += (n < node); 2489 2490 /* Give preference to headless and unused nodes */ 2491 tmp = cpumask_of_node(n); 2492 if (!cpumask_empty(tmp)) 2493 val += PENALTY_FOR_NODE_WITH_CPUS; 2494 2495 /* Slight preference for less loaded node */ 2496 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2497 val += node_load[n]; 2498 2499 if (val < min_val) { 2500 min_val = val; 2501 best_node = n; 2502 } 2503 } 2504 2505 if (best_node >= 0) 2506 node_set(best_node, *used_node_mask); 2507 2508 return best_node; 2509} 2510 2511 2512/* 2513 * Build zonelists ordered by node and zones within node. 2514 * This results in maximum locality--normal zone overflows into local 2515 * DMA zone, if any--but risks exhausting DMA zone. 2516 */ 2517static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2518{ 2519 int j; 2520 struct zonelist *zonelist; 2521 2522 zonelist = &pgdat->node_zonelists[0]; 2523 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2524 ; 2525 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2526 MAX_NR_ZONES - 1); 2527 zonelist->_zonerefs[j].zone = NULL; 2528 zonelist->_zonerefs[j].zone_idx = 0; 2529} 2530 2531/* 2532 * Build gfp_thisnode zonelists 2533 */ 2534static void build_thisnode_zonelists(pg_data_t *pgdat) 2535{ 2536 int j; 2537 struct zonelist *zonelist; 2538 2539 zonelist = &pgdat->node_zonelists[1]; 2540 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2541 zonelist->_zonerefs[j].zone = NULL; 2542 zonelist->_zonerefs[j].zone_idx = 0; 2543} 2544 2545/* 2546 * Build zonelists ordered by zone and nodes within zones. 2547 * This results in conserving DMA zone[s] until all Normal memory is 2548 * exhausted, but results in overflowing to remote node while memory 2549 * may still exist in local DMA zone. 2550 */ 2551static int node_order[MAX_NUMNODES]; 2552 2553static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2554{ 2555 int pos, j, node; 2556 int zone_type; /* needs to be signed */ 2557 struct zone *z; 2558 struct zonelist *zonelist; 2559 2560 zonelist = &pgdat->node_zonelists[0]; 2561 pos = 0; 2562 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2563 for (j = 0; j < nr_nodes; j++) { 2564 node = node_order[j]; 2565 z = &NODE_DATA(node)->node_zones[zone_type]; 2566 if (populated_zone(z)) { 2567 zoneref_set_zone(z, 2568 &zonelist->_zonerefs[pos++]); 2569 check_highest_zone(zone_type); 2570 } 2571 } 2572 } 2573 zonelist->_zonerefs[pos].zone = NULL; 2574 zonelist->_zonerefs[pos].zone_idx = 0; 2575} 2576 2577static int default_zonelist_order(void) 2578{ 2579 int nid, zone_type; 2580 unsigned long low_kmem_size,total_size; 2581 struct zone *z; 2582 int average_size; 2583 /* 2584 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2585 * If they are really small and used heavily, the system can fall 2586 * into OOM very easily. 2587 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2588 */ 2589 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2590 low_kmem_size = 0; 2591 total_size = 0; 2592 for_each_online_node(nid) { 2593 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2594 z = &NODE_DATA(nid)->node_zones[zone_type]; 2595 if (populated_zone(z)) { 2596 if (zone_type < ZONE_NORMAL) 2597 low_kmem_size += z->present_pages; 2598 total_size += z->present_pages; 2599 } 2600 } 2601 } 2602 if (!low_kmem_size || /* there are no DMA area. */ 2603 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2604 return ZONELIST_ORDER_NODE; 2605 /* 2606 * look into each node's config. 2607 * If there is a node whose DMA/DMA32 memory is very big area on 2608 * local memory, NODE_ORDER may be suitable. 2609 */ 2610 average_size = total_size / 2611 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2612 for_each_online_node(nid) { 2613 low_kmem_size = 0; 2614 total_size = 0; 2615 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2616 z = &NODE_DATA(nid)->node_zones[zone_type]; 2617 if (populated_zone(z)) { 2618 if (zone_type < ZONE_NORMAL) 2619 low_kmem_size += z->present_pages; 2620 total_size += z->present_pages; 2621 } 2622 } 2623 if (low_kmem_size && 2624 total_size > average_size && /* ignore small node */ 2625 low_kmem_size > total_size * 70/100) 2626 return ZONELIST_ORDER_NODE; 2627 } 2628 return ZONELIST_ORDER_ZONE; 2629} 2630 2631static void set_zonelist_order(void) 2632{ 2633 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2634 current_zonelist_order = default_zonelist_order(); 2635 else 2636 current_zonelist_order = user_zonelist_order; 2637} 2638 2639static void build_zonelists(pg_data_t *pgdat) 2640{ 2641 int j, node, load; 2642 enum zone_type i; 2643 nodemask_t used_mask; 2644 int local_node, prev_node; 2645 struct zonelist *zonelist; 2646 int order = current_zonelist_order; 2647 2648 /* initialize zonelists */ 2649 for (i = 0; i < MAX_ZONELISTS; i++) { 2650 zonelist = pgdat->node_zonelists + i; 2651 zonelist->_zonerefs[0].zone = NULL; 2652 zonelist->_zonerefs[0].zone_idx = 0; 2653 } 2654 2655 /* NUMA-aware ordering of nodes */ 2656 local_node = pgdat->node_id; 2657 load = nr_online_nodes; 2658 prev_node = local_node; 2659 nodes_clear(used_mask); 2660 2661 memset(node_order, 0, sizeof(node_order)); 2662 j = 0; 2663 2664 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2665 int distance = node_distance(local_node, node); 2666 2667 /* 2668 * If another node is sufficiently far away then it is better 2669 * to reclaim pages in a zone before going off node. 2670 */ 2671 if (distance > RECLAIM_DISTANCE) 2672 zone_reclaim_mode = 1; 2673 2674 /* 2675 * We don't want to pressure a particular node. 2676 * So adding penalty to the first node in same 2677 * distance group to make it round-robin. 2678 */ 2679 if (distance != node_distance(local_node, prev_node)) 2680 node_load[node] = load; 2681 2682 prev_node = node; 2683 load--; 2684 if (order == ZONELIST_ORDER_NODE) 2685 build_zonelists_in_node_order(pgdat, node); 2686 else 2687 node_order[j++] = node; /* remember order */ 2688 } 2689 2690 if (order == ZONELIST_ORDER_ZONE) { 2691 /* calculate node order -- i.e., DMA last! */ 2692 build_zonelists_in_zone_order(pgdat, j); 2693 } 2694 2695 build_thisnode_zonelists(pgdat); 2696} 2697 2698/* Construct the zonelist performance cache - see further mmzone.h */ 2699static void build_zonelist_cache(pg_data_t *pgdat) 2700{ 2701 struct zonelist *zonelist; 2702 struct zonelist_cache *zlc; 2703 struct zoneref *z; 2704 2705 zonelist = &pgdat->node_zonelists[0]; 2706 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2707 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2708 for (z = zonelist->_zonerefs; z->zone; z++) 2709 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2710} 2711 2712 2713#else /* CONFIG_NUMA */ 2714 2715static void set_zonelist_order(void) 2716{ 2717 current_zonelist_order = ZONELIST_ORDER_ZONE; 2718} 2719 2720static void build_zonelists(pg_data_t *pgdat) 2721{ 2722 int node, local_node; 2723 enum zone_type j; 2724 struct zonelist *zonelist; 2725 2726 local_node = pgdat->node_id; 2727 2728 zonelist = &pgdat->node_zonelists[0]; 2729 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2730 2731 /* 2732 * Now we build the zonelist so that it contains the zones 2733 * of all the other nodes. 2734 * We don't want to pressure a particular node, so when 2735 * building the zones for node N, we make sure that the 2736 * zones coming right after the local ones are those from 2737 * node N+1 (modulo N) 2738 */ 2739 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2740 if (!node_online(node)) 2741 continue; 2742 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2743 MAX_NR_ZONES - 1); 2744 } 2745 for (node = 0; node < local_node; node++) { 2746 if (!node_online(node)) 2747 continue; 2748 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2749 MAX_NR_ZONES - 1); 2750 } 2751 2752 zonelist->_zonerefs[j].zone = NULL; 2753 zonelist->_zonerefs[j].zone_idx = 0; 2754} 2755 2756/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2757static void build_zonelist_cache(pg_data_t *pgdat) 2758{ 2759 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2760} 2761 2762#endif /* CONFIG_NUMA */ 2763 2764/* 2765 * Boot pageset table. One per cpu which is going to be used for all 2766 * zones and all nodes. The parameters will be set in such a way 2767 * that an item put on a list will immediately be handed over to 2768 * the buddy list. This is safe since pageset manipulation is done 2769 * with interrupts disabled. 2770 * 2771 * The boot_pagesets must be kept even after bootup is complete for 2772 * unused processors and/or zones. They do play a role for bootstrapping 2773 * hotplugged processors. 2774 * 2775 * zoneinfo_show() and maybe other functions do 2776 * not check if the processor is online before following the pageset pointer. 2777 * Other parts of the kernel may not check if the zone is available. 2778 */ 2779static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 2780static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 2781 2782/* return values int ....just for stop_machine() */ 2783static int __build_all_zonelists(void *dummy) 2784{ 2785 int nid; 2786 int cpu; 2787 2788#ifdef CONFIG_NUMA 2789 memset(node_load, 0, sizeof(node_load)); 2790#endif 2791 for_each_online_node(nid) { 2792 pg_data_t *pgdat = NODE_DATA(nid); 2793 2794 build_zonelists(pgdat); 2795 build_zonelist_cache(pgdat); 2796 } 2797 2798 /* 2799 * Initialize the boot_pagesets that are going to be used 2800 * for bootstrapping processors. The real pagesets for 2801 * each zone will be allocated later when the per cpu 2802 * allocator is available. 2803 * 2804 * boot_pagesets are used also for bootstrapping offline 2805 * cpus if the system is already booted because the pagesets 2806 * are needed to initialize allocators on a specific cpu too. 2807 * F.e. the percpu allocator needs the page allocator which 2808 * needs the percpu allocator in order to allocate its pagesets 2809 * (a chicken-egg dilemma). 2810 */ 2811 for_each_possible_cpu(cpu) 2812 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 2813 2814 return 0; 2815} 2816 2817void build_all_zonelists(void) 2818{ 2819 set_zonelist_order(); 2820 2821 if (system_state == SYSTEM_BOOTING) { 2822 __build_all_zonelists(NULL); 2823 mminit_verify_zonelist(); 2824 cpuset_init_current_mems_allowed(); 2825 } else { 2826 /* we have to stop all cpus to guarantee there is no user 2827 of zonelist */ 2828 stop_machine(__build_all_zonelists, NULL, NULL); 2829 /* cpuset refresh routine should be here */ 2830 } 2831 vm_total_pages = nr_free_pagecache_pages(); 2832 /* 2833 * Disable grouping by mobility if the number of pages in the 2834 * system is too low to allow the mechanism to work. It would be 2835 * more accurate, but expensive to check per-zone. This check is 2836 * made on memory-hotadd so a system can start with mobility 2837 * disabled and enable it later 2838 */ 2839 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2840 page_group_by_mobility_disabled = 1; 2841 else 2842 page_group_by_mobility_disabled = 0; 2843 2844 printk("Built %i zonelists in %s order, mobility grouping %s. " 2845 "Total pages: %ld\n", 2846 nr_online_nodes, 2847 zonelist_order_name[current_zonelist_order], 2848 page_group_by_mobility_disabled ? "off" : "on", 2849 vm_total_pages); 2850#ifdef CONFIG_NUMA 2851 printk("Policy zone: %s\n", zone_names[policy_zone]); 2852#endif 2853} 2854 2855/* 2856 * Helper functions to size the waitqueue hash table. 2857 * Essentially these want to choose hash table sizes sufficiently 2858 * large so that collisions trying to wait on pages are rare. 2859 * But in fact, the number of active page waitqueues on typical 2860 * systems is ridiculously low, less than 200. So this is even 2861 * conservative, even though it seems large. 2862 * 2863 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2864 * waitqueues, i.e. the size of the waitq table given the number of pages. 2865 */ 2866#define PAGES_PER_WAITQUEUE 256 2867 2868#ifndef CONFIG_MEMORY_HOTPLUG 2869static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2870{ 2871 unsigned long size = 1; 2872 2873 pages /= PAGES_PER_WAITQUEUE; 2874 2875 while (size < pages) 2876 size <<= 1; 2877 2878 /* 2879 * Once we have dozens or even hundreds of threads sleeping 2880 * on IO we've got bigger problems than wait queue collision. 2881 * Limit the size of the wait table to a reasonable size. 2882 */ 2883 size = min(size, 4096UL); 2884 2885 return max(size, 4UL); 2886} 2887#else 2888/* 2889 * A zone's size might be changed by hot-add, so it is not possible to determine 2890 * a suitable size for its wait_table. So we use the maximum size now. 2891 * 2892 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2893 * 2894 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2895 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2896 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2897 * 2898 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2899 * or more by the traditional way. (See above). It equals: 2900 * 2901 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2902 * ia64(16K page size) : = ( 8G + 4M)byte. 2903 * powerpc (64K page size) : = (32G +16M)byte. 2904 */ 2905static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2906{ 2907 return 4096UL; 2908} 2909#endif 2910 2911/* 2912 * This is an integer logarithm so that shifts can be used later 2913 * to extract the more random high bits from the multiplicative 2914 * hash function before the remainder is taken. 2915 */ 2916static inline unsigned long wait_table_bits(unsigned long size) 2917{ 2918 return ffz(~size); 2919} 2920 2921#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2922 2923/* 2924 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2925 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2926 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2927 * higher will lead to a bigger reserve which will get freed as contiguous 2928 * blocks as reclaim kicks in 2929 */ 2930static void setup_zone_migrate_reserve(struct zone *zone) 2931{ 2932 unsigned long start_pfn, pfn, end_pfn; 2933 struct page *page; 2934 unsigned long block_migratetype; 2935 int reserve; 2936 2937 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2938 start_pfn = zone->zone_start_pfn; 2939 end_pfn = start_pfn + zone->spanned_pages; 2940 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2941 pageblock_order; 2942 2943 /* 2944 * Reserve blocks are generally in place to help high-order atomic 2945 * allocations that are short-lived. A min_free_kbytes value that 2946 * would result in more than 2 reserve blocks for atomic allocations 2947 * is assumed to be in place to help anti-fragmentation for the 2948 * future allocation of hugepages at runtime. 2949 */ 2950 reserve = min(2, reserve); 2951 2952 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2953 if (!pfn_valid(pfn)) 2954 continue; 2955 page = pfn_to_page(pfn); 2956 2957 /* Watch out for overlapping nodes */ 2958 if (page_to_nid(page) != zone_to_nid(zone)) 2959 continue; 2960 2961 /* Blocks with reserved pages will never free, skip them. */ 2962 if (PageReserved(page)) 2963 continue; 2964 2965 block_migratetype = get_pageblock_migratetype(page); 2966 2967 /* If this block is reserved, account for it */ 2968 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2969 reserve--; 2970 continue; 2971 } 2972 2973 /* Suitable for reserving if this block is movable */ 2974 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2975 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2976 move_freepages_block(zone, page, MIGRATE_RESERVE); 2977 reserve--; 2978 continue; 2979 } 2980 2981 /* 2982 * If the reserve is met and this is a previous reserved block, 2983 * take it back 2984 */ 2985 if (block_migratetype == MIGRATE_RESERVE) { 2986 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2987 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2988 } 2989 } 2990} 2991 2992/* 2993 * Initially all pages are reserved - free ones are freed 2994 * up by free_all_bootmem() once the early boot process is 2995 * done. Non-atomic initialization, single-pass. 2996 */ 2997void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2998 unsigned long start_pfn, enum memmap_context context) 2999{ 3000 struct page *page; 3001 unsigned long end_pfn = start_pfn + size; 3002 unsigned long pfn; 3003 struct zone *z; 3004 3005 if (highest_memmap_pfn < end_pfn - 1) 3006 highest_memmap_pfn = end_pfn - 1; 3007 3008 z = &NODE_DATA(nid)->node_zones[zone]; 3009 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3010 /* 3011 * There can be holes in boot-time mem_map[]s 3012 * handed to this function. They do not 3013 * exist on hotplugged memory. 3014 */ 3015 if (context == MEMMAP_EARLY) { 3016 if (!early_pfn_valid(pfn)) 3017 continue; 3018 if (!early_pfn_in_nid(pfn, nid)) 3019 continue; 3020 } 3021 page = pfn_to_page(pfn); 3022 set_page_links(page, zone, nid, pfn); 3023 mminit_verify_page_links(page, zone, nid, pfn); 3024 init_page_count(page); 3025 reset_page_mapcount(page); 3026 SetPageReserved(page); 3027 /* 3028 * Mark the block movable so that blocks are reserved for 3029 * movable at startup. This will force kernel allocations 3030 * to reserve their blocks rather than leaking throughout 3031 * the address space during boot when many long-lived 3032 * kernel allocations are made. Later some blocks near 3033 * the start are marked MIGRATE_RESERVE by 3034 * setup_zone_migrate_reserve() 3035 * 3036 * bitmap is created for zone's valid pfn range. but memmap 3037 * can be created for invalid pages (for alignment) 3038 * check here not to call set_pageblock_migratetype() against 3039 * pfn out of zone. 3040 */ 3041 if ((z->zone_start_pfn <= pfn) 3042 && (pfn < z->zone_start_pfn + z->spanned_pages) 3043 && !(pfn & (pageblock_nr_pages - 1))) 3044 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3045 3046 INIT_LIST_HEAD(&page->lru); 3047#ifdef WANT_PAGE_VIRTUAL 3048 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3049 if (!is_highmem_idx(zone)) 3050 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3051#endif 3052 } 3053} 3054 3055static void __meminit zone_init_free_lists(struct zone *zone) 3056{ 3057 int order, t; 3058 for_each_migratetype_order(order, t) { 3059 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3060 zone->free_area[order].nr_free = 0; 3061 } 3062} 3063 3064#ifndef __HAVE_ARCH_MEMMAP_INIT 3065#define memmap_init(size, nid, zone, start_pfn) \ 3066 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3067#endif 3068 3069static int zone_batchsize(struct zone *zone) 3070{ 3071#ifdef CONFIG_MMU 3072 int batch; 3073 3074 /* 3075 * The per-cpu-pages pools are set to around 1000th of the 3076 * size of the zone. But no more than 1/2 of a meg. 3077 * 3078 * OK, so we don't know how big the cache is. So guess. 3079 */ 3080 batch = zone->present_pages / 1024; 3081 if (batch * PAGE_SIZE > 512 * 1024) 3082 batch = (512 * 1024) / PAGE_SIZE; 3083 batch /= 4; /* We effectively *= 4 below */ 3084 if (batch < 1) 3085 batch = 1; 3086 3087 /* 3088 * Clamp the batch to a 2^n - 1 value. Having a power 3089 * of 2 value was found to be more likely to have 3090 * suboptimal cache aliasing properties in some cases. 3091 * 3092 * For example if 2 tasks are alternately allocating 3093 * batches of pages, one task can end up with a lot 3094 * of pages of one half of the possible page colors 3095 * and the other with pages of the other colors. 3096 */ 3097 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3098 3099 return batch; 3100 3101#else 3102 /* The deferral and batching of frees should be suppressed under NOMMU 3103 * conditions. 3104 * 3105 * The problem is that NOMMU needs to be able to allocate large chunks 3106 * of contiguous memory as there's no hardware page translation to 3107 * assemble apparent contiguous memory from discontiguous pages. 3108 * 3109 * Queueing large contiguous runs of pages for batching, however, 3110 * causes the pages to actually be freed in smaller chunks. As there 3111 * can be a significant delay between the individual batches being 3112 * recycled, this leads to the once large chunks of space being 3113 * fragmented and becoming unavailable for high-order allocations. 3114 */ 3115 return 0; 3116#endif 3117} 3118 3119static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3120{ 3121 struct per_cpu_pages *pcp; 3122 int migratetype; 3123 3124 memset(p, 0, sizeof(*p)); 3125 3126 pcp = &p->pcp; 3127 pcp->count = 0; 3128 pcp->high = 6 * batch; 3129 pcp->batch = max(1UL, 1 * batch); 3130 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3131 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3132} 3133 3134/* 3135 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3136 * to the value high for the pageset p. 3137 */ 3138 3139static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3140 unsigned long high) 3141{ 3142 struct per_cpu_pages *pcp; 3143 3144 pcp = &p->pcp; 3145 pcp->high = high; 3146 pcp->batch = max(1UL, high/4); 3147 if ((high/4) > (PAGE_SHIFT * 8)) 3148 pcp->batch = PAGE_SHIFT * 8; 3149} 3150 3151/* 3152 * Allocate per cpu pagesets and initialize them. 3153 * Before this call only boot pagesets were available. 3154 * Boot pagesets will no longer be used by this processorr 3155 * after setup_per_cpu_pageset(). 3156 */ 3157void __init setup_per_cpu_pageset(void) 3158{ 3159 struct zone *zone; 3160 int cpu; 3161 3162 for_each_populated_zone(zone) { 3163 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3164 3165 for_each_possible_cpu(cpu) { 3166 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3167 3168 setup_pageset(pcp, zone_batchsize(zone)); 3169 3170 if (percpu_pagelist_fraction) 3171 setup_pagelist_highmark(pcp, 3172 (zone->present_pages / 3173 percpu_pagelist_fraction)); 3174 } 3175 } 3176} 3177 3178static noinline __init_refok 3179int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3180{ 3181 int i; 3182 struct pglist_data *pgdat = zone->zone_pgdat; 3183 size_t alloc_size; 3184 3185 /* 3186 * The per-page waitqueue mechanism uses hashed waitqueues 3187 * per zone. 3188 */ 3189 zone->wait_table_hash_nr_entries = 3190 wait_table_hash_nr_entries(zone_size_pages); 3191 zone->wait_table_bits = 3192 wait_table_bits(zone->wait_table_hash_nr_entries); 3193 alloc_size = zone->wait_table_hash_nr_entries 3194 * sizeof(wait_queue_head_t); 3195 3196 if (!slab_is_available()) { 3197 zone->wait_table = (wait_queue_head_t *) 3198 alloc_bootmem_node(pgdat, alloc_size); 3199 } else { 3200 /* 3201 * This case means that a zone whose size was 0 gets new memory 3202 * via memory hot-add. 3203 * But it may be the case that a new node was hot-added. In 3204 * this case vmalloc() will not be able to use this new node's 3205 * memory - this wait_table must be initialized to use this new 3206 * node itself as well. 3207 * To use this new node's memory, further consideration will be 3208 * necessary. 3209 */ 3210 zone->wait_table = vmalloc(alloc_size); 3211 } 3212 if (!zone->wait_table) 3213 return -ENOMEM; 3214 3215 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3216 init_waitqueue_head(zone->wait_table + i); 3217 3218 return 0; 3219} 3220 3221static int __zone_pcp_update(void *data) 3222{ 3223 struct zone *zone = data; 3224 int cpu; 3225 unsigned long batch = zone_batchsize(zone), flags; 3226 3227 for_each_possible_cpu(cpu) { 3228 struct per_cpu_pageset *pset; 3229 struct per_cpu_pages *pcp; 3230 3231 pset = per_cpu_ptr(zone->pageset, cpu); 3232 pcp = &pset->pcp; 3233 3234 local_irq_save(flags); 3235 free_pcppages_bulk(zone, pcp->count, pcp); 3236 setup_pageset(pset, batch); 3237 local_irq_restore(flags); 3238 } 3239 return 0; 3240} 3241 3242void zone_pcp_update(struct zone *zone) 3243{ 3244 stop_machine(__zone_pcp_update, zone, NULL); 3245} 3246 3247static __meminit void zone_pcp_init(struct zone *zone) 3248{ 3249 /* 3250 * per cpu subsystem is not up at this point. The following code 3251 * relies on the ability of the linker to provide the 3252 * offset of a (static) per cpu variable into the per cpu area. 3253 */ 3254 zone->pageset = &boot_pageset; 3255 3256 if (zone->present_pages) 3257 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3258 zone->name, zone->present_pages, 3259 zone_batchsize(zone)); 3260} 3261 3262__meminit int init_currently_empty_zone(struct zone *zone, 3263 unsigned long zone_start_pfn, 3264 unsigned long size, 3265 enum memmap_context context) 3266{ 3267 struct pglist_data *pgdat = zone->zone_pgdat; 3268 int ret; 3269 ret = zone_wait_table_init(zone, size); 3270 if (ret) 3271 return ret; 3272 pgdat->nr_zones = zone_idx(zone) + 1; 3273 3274 zone->zone_start_pfn = zone_start_pfn; 3275 3276 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3277 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3278 pgdat->node_id, 3279 (unsigned long)zone_idx(zone), 3280 zone_start_pfn, (zone_start_pfn + size)); 3281 3282 zone_init_free_lists(zone); 3283 3284 return 0; 3285} 3286 3287#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3288/* 3289 * Basic iterator support. Return the first range of PFNs for a node 3290 * Note: nid == MAX_NUMNODES returns first region regardless of node 3291 */ 3292static int __meminit first_active_region_index_in_nid(int nid) 3293{ 3294 int i; 3295 3296 for (i = 0; i < nr_nodemap_entries; i++) 3297 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3298 return i; 3299 3300 return -1; 3301} 3302 3303/* 3304 * Basic iterator support. Return the next active range of PFNs for a node 3305 * Note: nid == MAX_NUMNODES returns next region regardless of node 3306 */ 3307static int __meminit next_active_region_index_in_nid(int index, int nid) 3308{ 3309 for (index = index + 1; index < nr_nodemap_entries; index++) 3310 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3311 return index; 3312 3313 return -1; 3314} 3315 3316#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3317/* 3318 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3319 * Architectures may implement their own version but if add_active_range() 3320 * was used and there are no special requirements, this is a convenient 3321 * alternative 3322 */ 3323int __meminit __early_pfn_to_nid(unsigned long pfn) 3324{ 3325 int i; 3326 3327 for (i = 0; i < nr_nodemap_entries; i++) { 3328 unsigned long start_pfn = early_node_map[i].start_pfn; 3329 unsigned long end_pfn = early_node_map[i].end_pfn; 3330 3331 if (start_pfn <= pfn && pfn < end_pfn) 3332 return early_node_map[i].nid; 3333 } 3334 /* This is a memory hole */ 3335 return -1; 3336} 3337#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3338 3339int __meminit early_pfn_to_nid(unsigned long pfn) 3340{ 3341 int nid; 3342 3343 nid = __early_pfn_to_nid(pfn); 3344 if (nid >= 0) 3345 return nid; 3346 /* just returns 0 */ 3347 return 0; 3348} 3349 3350#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3351bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3352{ 3353 int nid; 3354 3355 nid = __early_pfn_to_nid(pfn); 3356 if (nid >= 0 && nid != node) 3357 return false; 3358 return true; 3359} 3360#endif 3361 3362/* Basic iterator support to walk early_node_map[] */ 3363#define for_each_active_range_index_in_nid(i, nid) \ 3364 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3365 i = next_active_region_index_in_nid(i, nid)) 3366 3367/** 3368 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3369 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3370 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3371 * 3372 * If an architecture guarantees that all ranges registered with 3373 * add_active_ranges() contain no holes and may be freed, this 3374 * this function may be used instead of calling free_bootmem() manually. 3375 */ 3376void __init free_bootmem_with_active_regions(int nid, 3377 unsigned long max_low_pfn) 3378{ 3379 int i; 3380 3381 for_each_active_range_index_in_nid(i, nid) { 3382 unsigned long size_pages = 0; 3383 unsigned long end_pfn = early_node_map[i].end_pfn; 3384 3385 if (early_node_map[i].start_pfn >= max_low_pfn) 3386 continue; 3387 3388 if (end_pfn > max_low_pfn) 3389 end_pfn = max_low_pfn; 3390 3391 size_pages = end_pfn - early_node_map[i].start_pfn; 3392 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3393 PFN_PHYS(early_node_map[i].start_pfn), 3394 size_pages << PAGE_SHIFT); 3395 } 3396} 3397 3398int __init add_from_early_node_map(struct range *range, int az, 3399 int nr_range, int nid) 3400{ 3401 int i; 3402 u64 start, end; 3403 3404 /* need to go over early_node_map to find out good range for node */ 3405 for_each_active_range_index_in_nid(i, nid) { 3406 start = early_node_map[i].start_pfn; 3407 end = early_node_map[i].end_pfn; 3408 nr_range = add_range(range, az, nr_range, start, end); 3409 } 3410 return nr_range; 3411} 3412 3413#ifdef CONFIG_NO_BOOTMEM 3414void * __init __alloc_memory_core_early(int nid, u64 size, u64 align, 3415 u64 goal, u64 limit) 3416{ 3417 int i; 3418 void *ptr; 3419 3420 /* need to go over early_node_map to find out good range for node */ 3421 for_each_active_range_index_in_nid(i, nid) { 3422 u64 addr; 3423 u64 ei_start, ei_last; 3424 3425 ei_last = early_node_map[i].end_pfn; 3426 ei_last <<= PAGE_SHIFT; 3427 ei_start = early_node_map[i].start_pfn; 3428 ei_start <<= PAGE_SHIFT; 3429 addr = find_early_area(ei_start, ei_last, 3430 goal, limit, size, align); 3431 3432 if (addr == -1ULL) 3433 continue; 3434 3435#if 0 3436 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n", 3437 nid, 3438 ei_start, ei_last, goal, limit, size, 3439 align, addr); 3440#endif 3441 3442 ptr = phys_to_virt(addr); 3443 memset(ptr, 0, size); 3444 reserve_early_without_check(addr, addr + size, "BOOTMEM"); 3445 return ptr; 3446 } 3447 3448 return NULL; 3449} 3450#endif 3451 3452 3453void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3454{ 3455 int i; 3456 int ret; 3457 3458 for_each_active_range_index_in_nid(i, nid) { 3459 ret = work_fn(early_node_map[i].start_pfn, 3460 early_node_map[i].end_pfn, data); 3461 if (ret) 3462 break; 3463 } 3464} 3465/** 3466 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3467 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3468 * 3469 * If an architecture guarantees that all ranges registered with 3470 * add_active_ranges() contain no holes and may be freed, this 3471 * function may be used instead of calling memory_present() manually. 3472 */ 3473void __init sparse_memory_present_with_active_regions(int nid) 3474{ 3475 int i; 3476 3477 for_each_active_range_index_in_nid(i, nid) 3478 memory_present(early_node_map[i].nid, 3479 early_node_map[i].start_pfn, 3480 early_node_map[i].end_pfn); 3481} 3482 3483/** 3484 * get_pfn_range_for_nid - Return the start and end page frames for a node 3485 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3486 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3487 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3488 * 3489 * It returns the start and end page frame of a node based on information 3490 * provided by an arch calling add_active_range(). If called for a node 3491 * with no available memory, a warning is printed and the start and end 3492 * PFNs will be 0. 3493 */ 3494void __meminit get_pfn_range_for_nid(unsigned int nid, 3495 unsigned long *start_pfn, unsigned long *end_pfn) 3496{ 3497 int i; 3498 *start_pfn = -1UL; 3499 *end_pfn = 0; 3500 3501 for_each_active_range_index_in_nid(i, nid) { 3502 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3503 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3504 } 3505 3506 if (*start_pfn == -1UL) 3507 *start_pfn = 0; 3508} 3509 3510/* 3511 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3512 * assumption is made that zones within a node are ordered in monotonic 3513 * increasing memory addresses so that the "highest" populated zone is used 3514 */ 3515static void __init find_usable_zone_for_movable(void) 3516{ 3517 int zone_index; 3518 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3519 if (zone_index == ZONE_MOVABLE) 3520 continue; 3521 3522 if (arch_zone_highest_possible_pfn[zone_index] > 3523 arch_zone_lowest_possible_pfn[zone_index]) 3524 break; 3525 } 3526 3527 VM_BUG_ON(zone_index == -1); 3528 movable_zone = zone_index; 3529} 3530 3531/* 3532 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3533 * because it is sized independant of architecture. Unlike the other zones, 3534 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3535 * in each node depending on the size of each node and how evenly kernelcore 3536 * is distributed. This helper function adjusts the zone ranges 3537 * provided by the architecture for a given node by using the end of the 3538 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3539 * zones within a node are in order of monotonic increases memory addresses 3540 */ 3541static void __meminit adjust_zone_range_for_zone_movable(int nid, 3542 unsigned long zone_type, 3543 unsigned long node_start_pfn, 3544 unsigned long node_end_pfn, 3545 unsigned long *zone_start_pfn, 3546 unsigned long *zone_end_pfn) 3547{ 3548 /* Only adjust if ZONE_MOVABLE is on this node */ 3549 if (zone_movable_pfn[nid]) { 3550 /* Size ZONE_MOVABLE */ 3551 if (zone_type == ZONE_MOVABLE) { 3552 *zone_start_pfn = zone_movable_pfn[nid]; 3553 *zone_end_pfn = min(node_end_pfn, 3554 arch_zone_highest_possible_pfn[movable_zone]); 3555 3556 /* Adjust for ZONE_MOVABLE starting within this range */ 3557 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3558 *zone_end_pfn > zone_movable_pfn[nid]) { 3559 *zone_end_pfn = zone_movable_pfn[nid]; 3560 3561 /* Check if this whole range is within ZONE_MOVABLE */ 3562 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3563 *zone_start_pfn = *zone_end_pfn; 3564 } 3565} 3566 3567/* 3568 * Return the number of pages a zone spans in a node, including holes 3569 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3570 */ 3571static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3572 unsigned long zone_type, 3573 unsigned long *ignored) 3574{ 3575 unsigned long node_start_pfn, node_end_pfn; 3576 unsigned long zone_start_pfn, zone_end_pfn; 3577 3578 /* Get the start and end of the node and zone */ 3579 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3580 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3581 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3582 adjust_zone_range_for_zone_movable(nid, zone_type, 3583 node_start_pfn, node_end_pfn, 3584 &zone_start_pfn, &zone_end_pfn); 3585 3586 /* Check that this node has pages within the zone's required range */ 3587 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3588 return 0; 3589 3590 /* Move the zone boundaries inside the node if necessary */ 3591 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3592 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3593 3594 /* Return the spanned pages */ 3595 return zone_end_pfn - zone_start_pfn; 3596} 3597 3598/* 3599 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3600 * then all holes in the requested range will be accounted for. 3601 */ 3602unsigned long __meminit __absent_pages_in_range(int nid, 3603 unsigned long range_start_pfn, 3604 unsigned long range_end_pfn) 3605{ 3606 int i = 0; 3607 unsigned long prev_end_pfn = 0, hole_pages = 0; 3608 unsigned long start_pfn; 3609 3610 /* Find the end_pfn of the first active range of pfns in the node */ 3611 i = first_active_region_index_in_nid(nid); 3612 if (i == -1) 3613 return 0; 3614 3615 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3616 3617 /* Account for ranges before physical memory on this node */ 3618 if (early_node_map[i].start_pfn > range_start_pfn) 3619 hole_pages = prev_end_pfn - range_start_pfn; 3620 3621 /* Find all holes for the zone within the node */ 3622 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3623 3624 /* No need to continue if prev_end_pfn is outside the zone */ 3625 if (prev_end_pfn >= range_end_pfn) 3626 break; 3627 3628 /* Make sure the end of the zone is not within the hole */ 3629 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3630 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3631 3632 /* Update the hole size cound and move on */ 3633 if (start_pfn > range_start_pfn) { 3634 BUG_ON(prev_end_pfn > start_pfn); 3635 hole_pages += start_pfn - prev_end_pfn; 3636 } 3637 prev_end_pfn = early_node_map[i].end_pfn; 3638 } 3639 3640 /* Account for ranges past physical memory on this node */ 3641 if (range_end_pfn > prev_end_pfn) 3642 hole_pages += range_end_pfn - 3643 max(range_start_pfn, prev_end_pfn); 3644 3645 return hole_pages; 3646} 3647 3648/** 3649 * absent_pages_in_range - Return number of page frames in holes within a range 3650 * @start_pfn: The start PFN to start searching for holes 3651 * @end_pfn: The end PFN to stop searching for holes 3652 * 3653 * It returns the number of pages frames in memory holes within a range. 3654 */ 3655unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3656 unsigned long end_pfn) 3657{ 3658 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3659} 3660 3661/* Return the number of page frames in holes in a zone on a node */ 3662static unsigned long __meminit zone_absent_pages_in_node(int nid, 3663 unsigned long zone_type, 3664 unsigned long *ignored) 3665{ 3666 unsigned long node_start_pfn, node_end_pfn; 3667 unsigned long zone_start_pfn, zone_end_pfn; 3668 3669 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3670 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3671 node_start_pfn); 3672 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3673 node_end_pfn); 3674 3675 adjust_zone_range_for_zone_movable(nid, zone_type, 3676 node_start_pfn, node_end_pfn, 3677 &zone_start_pfn, &zone_end_pfn); 3678 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3679} 3680 3681#else 3682static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3683 unsigned long zone_type, 3684 unsigned long *zones_size) 3685{ 3686 return zones_size[zone_type]; 3687} 3688 3689static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3690 unsigned long zone_type, 3691 unsigned long *zholes_size) 3692{ 3693 if (!zholes_size) 3694 return 0; 3695 3696 return zholes_size[zone_type]; 3697} 3698 3699#endif 3700 3701static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3702 unsigned long *zones_size, unsigned long *zholes_size) 3703{ 3704 unsigned long realtotalpages, totalpages = 0; 3705 enum zone_type i; 3706 3707 for (i = 0; i < MAX_NR_ZONES; i++) 3708 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3709 zones_size); 3710 pgdat->node_spanned_pages = totalpages; 3711 3712 realtotalpages = totalpages; 3713 for (i = 0; i < MAX_NR_ZONES; i++) 3714 realtotalpages -= 3715 zone_absent_pages_in_node(pgdat->node_id, i, 3716 zholes_size); 3717 pgdat->node_present_pages = realtotalpages; 3718 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3719 realtotalpages); 3720} 3721 3722#ifndef CONFIG_SPARSEMEM 3723/* 3724 * Calculate the size of the zone->blockflags rounded to an unsigned long 3725 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3726 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3727 * round what is now in bits to nearest long in bits, then return it in 3728 * bytes. 3729 */ 3730static unsigned long __init usemap_size(unsigned long zonesize) 3731{ 3732 unsigned long usemapsize; 3733 3734 usemapsize = roundup(zonesize, pageblock_nr_pages); 3735 usemapsize = usemapsize >> pageblock_order; 3736 usemapsize *= NR_PAGEBLOCK_BITS; 3737 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3738 3739 return usemapsize / 8; 3740} 3741 3742static void __init setup_usemap(struct pglist_data *pgdat, 3743 struct zone *zone, unsigned long zonesize) 3744{ 3745 unsigned long usemapsize = usemap_size(zonesize); 3746 zone->pageblock_flags = NULL; 3747 if (usemapsize) 3748 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3749} 3750#else 3751static void inline setup_usemap(struct pglist_data *pgdat, 3752 struct zone *zone, unsigned long zonesize) {} 3753#endif /* CONFIG_SPARSEMEM */ 3754 3755#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3756 3757/* Return a sensible default order for the pageblock size. */ 3758static inline int pageblock_default_order(void) 3759{ 3760 if (HPAGE_SHIFT > PAGE_SHIFT) 3761 return HUGETLB_PAGE_ORDER; 3762 3763 return MAX_ORDER-1; 3764} 3765 3766/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3767static inline void __init set_pageblock_order(unsigned int order) 3768{ 3769 /* Check that pageblock_nr_pages has not already been setup */ 3770 if (pageblock_order) 3771 return; 3772 3773 /* 3774 * Assume the largest contiguous order of interest is a huge page. 3775 * This value may be variable depending on boot parameters on IA64 3776 */ 3777 pageblock_order = order; 3778} 3779#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3780 3781/* 3782 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3783 * and pageblock_default_order() are unused as pageblock_order is set 3784 * at compile-time. See include/linux/pageblock-flags.h for the values of 3785 * pageblock_order based on the kernel config 3786 */ 3787static inline int pageblock_default_order(unsigned int order) 3788{ 3789 return MAX_ORDER-1; 3790} 3791#define set_pageblock_order(x) do {} while (0) 3792 3793#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3794 3795/* 3796 * Set up the zone data structures: 3797 * - mark all pages reserved 3798 * - mark all memory queues empty 3799 * - clear the memory bitmaps 3800 */ 3801static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3802 unsigned long *zones_size, unsigned long *zholes_size) 3803{ 3804 enum zone_type j; 3805 int nid = pgdat->node_id; 3806 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3807 int ret; 3808 3809 pgdat_resize_init(pgdat); 3810 pgdat->nr_zones = 0; 3811 init_waitqueue_head(&pgdat->kswapd_wait); 3812 pgdat->kswapd_max_order = 0; 3813 pgdat_page_cgroup_init(pgdat); 3814 3815 for (j = 0; j < MAX_NR_ZONES; j++) { 3816 struct zone *zone = pgdat->node_zones + j; 3817 unsigned long size, realsize, memmap_pages; 3818 enum lru_list l; 3819 3820 size = zone_spanned_pages_in_node(nid, j, zones_size); 3821 realsize = size - zone_absent_pages_in_node(nid, j, 3822 zholes_size); 3823 3824 /* 3825 * Adjust realsize so that it accounts for how much memory 3826 * is used by this zone for memmap. This affects the watermark 3827 * and per-cpu initialisations 3828 */ 3829 memmap_pages = 3830 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3831 if (realsize >= memmap_pages) { 3832 realsize -= memmap_pages; 3833 if (memmap_pages) 3834 printk(KERN_DEBUG 3835 " %s zone: %lu pages used for memmap\n", 3836 zone_names[j], memmap_pages); 3837 } else 3838 printk(KERN_WARNING 3839 " %s zone: %lu pages exceeds realsize %lu\n", 3840 zone_names[j], memmap_pages, realsize); 3841 3842 /* Account for reserved pages */ 3843 if (j == 0 && realsize > dma_reserve) { 3844 realsize -= dma_reserve; 3845 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3846 zone_names[0], dma_reserve); 3847 } 3848 3849 if (!is_highmem_idx(j)) 3850 nr_kernel_pages += realsize; 3851 nr_all_pages += realsize; 3852 3853 zone->spanned_pages = size; 3854 zone->present_pages = realsize; 3855#ifdef CONFIG_NUMA 3856 zone->node = nid; 3857 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3858 / 100; 3859 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3860#endif 3861 zone->name = zone_names[j]; 3862 spin_lock_init(&zone->lock); 3863 spin_lock_init(&zone->lru_lock); 3864 zone_seqlock_init(zone); 3865 zone->zone_pgdat = pgdat; 3866 3867 zone->prev_priority = DEF_PRIORITY; 3868 3869 zone_pcp_init(zone); 3870 for_each_lru(l) { 3871 INIT_LIST_HEAD(&zone->lru[l].list); 3872 zone->reclaim_stat.nr_saved_scan[l] = 0; 3873 } 3874 zone->reclaim_stat.recent_rotated[0] = 0; 3875 zone->reclaim_stat.recent_rotated[1] = 0; 3876 zone->reclaim_stat.recent_scanned[0] = 0; 3877 zone->reclaim_stat.recent_scanned[1] = 0; 3878 zap_zone_vm_stats(zone); 3879 zone->flags = 0; 3880 if (!size) 3881 continue; 3882 3883 set_pageblock_order(pageblock_default_order()); 3884 setup_usemap(pgdat, zone, size); 3885 ret = init_currently_empty_zone(zone, zone_start_pfn, 3886 size, MEMMAP_EARLY); 3887 BUG_ON(ret); 3888 memmap_init(size, nid, j, zone_start_pfn); 3889 zone_start_pfn += size; 3890 } 3891} 3892 3893static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3894{ 3895 /* Skip empty nodes */ 3896 if (!pgdat->node_spanned_pages) 3897 return; 3898 3899#ifdef CONFIG_FLAT_NODE_MEM_MAP 3900 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3901 if (!pgdat->node_mem_map) { 3902 unsigned long size, start, end; 3903 struct page *map; 3904 3905 /* 3906 * The zone's endpoints aren't required to be MAX_ORDER 3907 * aligned but the node_mem_map endpoints must be in order 3908 * for the buddy allocator to function correctly. 3909 */ 3910 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3911 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3912 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3913 size = (end - start) * sizeof(struct page); 3914 map = alloc_remap(pgdat->node_id, size); 3915 if (!map) 3916 map = alloc_bootmem_node(pgdat, size); 3917 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3918 } 3919#ifndef CONFIG_NEED_MULTIPLE_NODES 3920 /* 3921 * With no DISCONTIG, the global mem_map is just set as node 0's 3922 */ 3923 if (pgdat == NODE_DATA(0)) { 3924 mem_map = NODE_DATA(0)->node_mem_map; 3925#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3926 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3927 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3928#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3929 } 3930#endif 3931#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3932} 3933 3934void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3935 unsigned long node_start_pfn, unsigned long *zholes_size) 3936{ 3937 pg_data_t *pgdat = NODE_DATA(nid); 3938 3939 pgdat->node_id = nid; 3940 pgdat->node_start_pfn = node_start_pfn; 3941 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3942 3943 alloc_node_mem_map(pgdat); 3944#ifdef CONFIG_FLAT_NODE_MEM_MAP 3945 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3946 nid, (unsigned long)pgdat, 3947 (unsigned long)pgdat->node_mem_map); 3948#endif 3949 3950 free_area_init_core(pgdat, zones_size, zholes_size); 3951} 3952 3953#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3954 3955#if MAX_NUMNODES > 1 3956/* 3957 * Figure out the number of possible node ids. 3958 */ 3959static void __init setup_nr_node_ids(void) 3960{ 3961 unsigned int node; 3962 unsigned int highest = 0; 3963 3964 for_each_node_mask(node, node_possible_map) 3965 highest = node; 3966 nr_node_ids = highest + 1; 3967} 3968#else 3969static inline void setup_nr_node_ids(void) 3970{ 3971} 3972#endif 3973 3974/** 3975 * add_active_range - Register a range of PFNs backed by physical memory 3976 * @nid: The node ID the range resides on 3977 * @start_pfn: The start PFN of the available physical memory 3978 * @end_pfn: The end PFN of the available physical memory 3979 * 3980 * These ranges are stored in an early_node_map[] and later used by 3981 * free_area_init_nodes() to calculate zone sizes and holes. If the 3982 * range spans a memory hole, it is up to the architecture to ensure 3983 * the memory is not freed by the bootmem allocator. If possible 3984 * the range being registered will be merged with existing ranges. 3985 */ 3986void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3987 unsigned long end_pfn) 3988{ 3989 int i; 3990 3991 mminit_dprintk(MMINIT_TRACE, "memory_register", 3992 "Entering add_active_range(%d, %#lx, %#lx) " 3993 "%d entries of %d used\n", 3994 nid, start_pfn, end_pfn, 3995 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3996 3997 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 3998 3999 /* Merge with existing active regions if possible */ 4000 for (i = 0; i < nr_nodemap_entries; i++) { 4001 if (early_node_map[i].nid != nid) 4002 continue; 4003 4004 /* Skip if an existing region covers this new one */ 4005 if (start_pfn >= early_node_map[i].start_pfn && 4006 end_pfn <= early_node_map[i].end_pfn) 4007 return; 4008 4009 /* Merge forward if suitable */ 4010 if (start_pfn <= early_node_map[i].end_pfn && 4011 end_pfn > early_node_map[i].end_pfn) { 4012 early_node_map[i].end_pfn = end_pfn; 4013 return; 4014 } 4015 4016 /* Merge backward if suitable */ 4017 if (start_pfn < early_node_map[i].start_pfn && 4018 end_pfn >= early_node_map[i].start_pfn) { 4019 early_node_map[i].start_pfn = start_pfn; 4020 return; 4021 } 4022 } 4023 4024 /* Check that early_node_map is large enough */ 4025 if (i >= MAX_ACTIVE_REGIONS) { 4026 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4027 MAX_ACTIVE_REGIONS); 4028 return; 4029 } 4030 4031 early_node_map[i].nid = nid; 4032 early_node_map[i].start_pfn = start_pfn; 4033 early_node_map[i].end_pfn = end_pfn; 4034 nr_nodemap_entries = i + 1; 4035} 4036 4037/** 4038 * remove_active_range - Shrink an existing registered range of PFNs 4039 * @nid: The node id the range is on that should be shrunk 4040 * @start_pfn: The new PFN of the range 4041 * @end_pfn: The new PFN of the range 4042 * 4043 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4044 * The map is kept near the end physical page range that has already been 4045 * registered. This function allows an arch to shrink an existing registered 4046 * range. 4047 */ 4048void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4049 unsigned long end_pfn) 4050{ 4051 int i, j; 4052 int removed = 0; 4053 4054 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4055 nid, start_pfn, end_pfn); 4056 4057 /* Find the old active region end and shrink */ 4058 for_each_active_range_index_in_nid(i, nid) { 4059 if (early_node_map[i].start_pfn >= start_pfn && 4060 early_node_map[i].end_pfn <= end_pfn) { 4061 /* clear it */ 4062 early_node_map[i].start_pfn = 0; 4063 early_node_map[i].end_pfn = 0; 4064 removed = 1; 4065 continue; 4066 } 4067 if (early_node_map[i].start_pfn < start_pfn && 4068 early_node_map[i].end_pfn > start_pfn) { 4069 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4070 early_node_map[i].end_pfn = start_pfn; 4071 if (temp_end_pfn > end_pfn) 4072 add_active_range(nid, end_pfn, temp_end_pfn); 4073 continue; 4074 } 4075 if (early_node_map[i].start_pfn >= start_pfn && 4076 early_node_map[i].end_pfn > end_pfn && 4077 early_node_map[i].start_pfn < end_pfn) { 4078 early_node_map[i].start_pfn = end_pfn; 4079 continue; 4080 } 4081 } 4082 4083 if (!removed) 4084 return; 4085 4086 /* remove the blank ones */ 4087 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4088 if (early_node_map[i].nid != nid) 4089 continue; 4090 if (early_node_map[i].end_pfn) 4091 continue; 4092 /* we found it, get rid of it */ 4093 for (j = i; j < nr_nodemap_entries - 1; j++) 4094 memcpy(&early_node_map[j], &early_node_map[j+1], 4095 sizeof(early_node_map[j])); 4096 j = nr_nodemap_entries - 1; 4097 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4098 nr_nodemap_entries--; 4099 } 4100} 4101 4102/** 4103 * remove_all_active_ranges - Remove all currently registered regions 4104 * 4105 * During discovery, it may be found that a table like SRAT is invalid 4106 * and an alternative discovery method must be used. This function removes 4107 * all currently registered regions. 4108 */ 4109void __init remove_all_active_ranges(void) 4110{ 4111 memset(early_node_map, 0, sizeof(early_node_map)); 4112 nr_nodemap_entries = 0; 4113} 4114 4115/* Compare two active node_active_regions */ 4116static int __init cmp_node_active_region(const void *a, const void *b) 4117{ 4118 struct node_active_region *arange = (struct node_active_region *)a; 4119 struct node_active_region *brange = (struct node_active_region *)b; 4120 4121 /* Done this way to avoid overflows */ 4122 if (arange->start_pfn > brange->start_pfn) 4123 return 1; 4124 if (arange->start_pfn < brange->start_pfn) 4125 return -1; 4126 4127 return 0; 4128} 4129 4130/* sort the node_map by start_pfn */ 4131void __init sort_node_map(void) 4132{ 4133 sort(early_node_map, (size_t)nr_nodemap_entries, 4134 sizeof(struct node_active_region), 4135 cmp_node_active_region, NULL); 4136} 4137 4138/* Find the lowest pfn for a node */ 4139static unsigned long __init find_min_pfn_for_node(int nid) 4140{ 4141 int i; 4142 unsigned long min_pfn = ULONG_MAX; 4143 4144 /* Assuming a sorted map, the first range found has the starting pfn */ 4145 for_each_active_range_index_in_nid(i, nid) 4146 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4147 4148 if (min_pfn == ULONG_MAX) { 4149 printk(KERN_WARNING 4150 "Could not find start_pfn for node %d\n", nid); 4151 return 0; 4152 } 4153 4154 return min_pfn; 4155} 4156 4157/** 4158 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4159 * 4160 * It returns the minimum PFN based on information provided via 4161 * add_active_range(). 4162 */ 4163unsigned long __init find_min_pfn_with_active_regions(void) 4164{ 4165 return find_min_pfn_for_node(MAX_NUMNODES); 4166} 4167 4168/* 4169 * early_calculate_totalpages() 4170 * Sum pages in active regions for movable zone. 4171 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4172 */ 4173static unsigned long __init early_calculate_totalpages(void) 4174{ 4175 int i; 4176 unsigned long totalpages = 0; 4177 4178 for (i = 0; i < nr_nodemap_entries; i++) { 4179 unsigned long pages = early_node_map[i].end_pfn - 4180 early_node_map[i].start_pfn; 4181 totalpages += pages; 4182 if (pages) 4183 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4184 } 4185 return totalpages; 4186} 4187 4188/* 4189 * Find the PFN the Movable zone begins in each node. Kernel memory 4190 * is spread evenly between nodes as long as the nodes have enough 4191 * memory. When they don't, some nodes will have more kernelcore than 4192 * others 4193 */ 4194static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4195{ 4196 int i, nid; 4197 unsigned long usable_startpfn; 4198 unsigned long kernelcore_node, kernelcore_remaining; 4199 /* save the state before borrow the nodemask */ 4200 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4201 unsigned long totalpages = early_calculate_totalpages(); 4202 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4203 4204 /* 4205 * If movablecore was specified, calculate what size of 4206 * kernelcore that corresponds so that memory usable for 4207 * any allocation type is evenly spread. If both kernelcore 4208 * and movablecore are specified, then the value of kernelcore 4209 * will be used for required_kernelcore if it's greater than 4210 * what movablecore would have allowed. 4211 */ 4212 if (required_movablecore) { 4213 unsigned long corepages; 4214 4215 /* 4216 * Round-up so that ZONE_MOVABLE is at least as large as what 4217 * was requested by the user 4218 */ 4219 required_movablecore = 4220 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4221 corepages = totalpages - required_movablecore; 4222 4223 required_kernelcore = max(required_kernelcore, corepages); 4224 } 4225 4226 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4227 if (!required_kernelcore) 4228 goto out; 4229 4230 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4231 find_usable_zone_for_movable(); 4232 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4233 4234restart: 4235 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4236 kernelcore_node = required_kernelcore / usable_nodes; 4237 for_each_node_state(nid, N_HIGH_MEMORY) { 4238 /* 4239 * Recalculate kernelcore_node if the division per node 4240 * now exceeds what is necessary to satisfy the requested 4241 * amount of memory for the kernel 4242 */ 4243 if (required_kernelcore < kernelcore_node) 4244 kernelcore_node = required_kernelcore / usable_nodes; 4245 4246 /* 4247 * As the map is walked, we track how much memory is usable 4248 * by the kernel using kernelcore_remaining. When it is 4249 * 0, the rest of the node is usable by ZONE_MOVABLE 4250 */ 4251 kernelcore_remaining = kernelcore_node; 4252 4253 /* Go through each range of PFNs within this node */ 4254 for_each_active_range_index_in_nid(i, nid) { 4255 unsigned long start_pfn, end_pfn; 4256 unsigned long size_pages; 4257 4258 start_pfn = max(early_node_map[i].start_pfn, 4259 zone_movable_pfn[nid]); 4260 end_pfn = early_node_map[i].end_pfn; 4261 if (start_pfn >= end_pfn) 4262 continue; 4263 4264 /* Account for what is only usable for kernelcore */ 4265 if (start_pfn < usable_startpfn) { 4266 unsigned long kernel_pages; 4267 kernel_pages = min(end_pfn, usable_startpfn) 4268 - start_pfn; 4269 4270 kernelcore_remaining -= min(kernel_pages, 4271 kernelcore_remaining); 4272 required_kernelcore -= min(kernel_pages, 4273 required_kernelcore); 4274 4275 /* Continue if range is now fully accounted */ 4276 if (end_pfn <= usable_startpfn) { 4277 4278 /* 4279 * Push zone_movable_pfn to the end so 4280 * that if we have to rebalance 4281 * kernelcore across nodes, we will 4282 * not double account here 4283 */ 4284 zone_movable_pfn[nid] = end_pfn; 4285 continue; 4286 } 4287 start_pfn = usable_startpfn; 4288 } 4289 4290 /* 4291 * The usable PFN range for ZONE_MOVABLE is from 4292 * start_pfn->end_pfn. Calculate size_pages as the 4293 * number of pages used as kernelcore 4294 */ 4295 size_pages = end_pfn - start_pfn; 4296 if (size_pages > kernelcore_remaining) 4297 size_pages = kernelcore_remaining; 4298 zone_movable_pfn[nid] = start_pfn + size_pages; 4299 4300 /* 4301 * Some kernelcore has been met, update counts and 4302 * break if the kernelcore for this node has been 4303 * satisified 4304 */ 4305 required_kernelcore -= min(required_kernelcore, 4306 size_pages); 4307 kernelcore_remaining -= size_pages; 4308 if (!kernelcore_remaining) 4309 break; 4310 } 4311 } 4312 4313 /* 4314 * If there is still required_kernelcore, we do another pass with one 4315 * less node in the count. This will push zone_movable_pfn[nid] further 4316 * along on the nodes that still have memory until kernelcore is 4317 * satisified 4318 */ 4319 usable_nodes--; 4320 if (usable_nodes && required_kernelcore > usable_nodes) 4321 goto restart; 4322 4323 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4324 for (nid = 0; nid < MAX_NUMNODES; nid++) 4325 zone_movable_pfn[nid] = 4326 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4327 4328out: 4329 /* restore the node_state */ 4330 node_states[N_HIGH_MEMORY] = saved_node_state; 4331} 4332 4333/* Any regular memory on that node ? */ 4334static void check_for_regular_memory(pg_data_t *pgdat) 4335{ 4336#ifdef CONFIG_HIGHMEM 4337 enum zone_type zone_type; 4338 4339 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4340 struct zone *zone = &pgdat->node_zones[zone_type]; 4341 if (zone->present_pages) 4342 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4343 } 4344#endif 4345} 4346 4347/** 4348 * free_area_init_nodes - Initialise all pg_data_t and zone data 4349 * @max_zone_pfn: an array of max PFNs for each zone 4350 * 4351 * This will call free_area_init_node() for each active node in the system. 4352 * Using the page ranges provided by add_active_range(), the size of each 4353 * zone in each node and their holes is calculated. If the maximum PFN 4354 * between two adjacent zones match, it is assumed that the zone is empty. 4355 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4356 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4357 * starts where the previous one ended. For example, ZONE_DMA32 starts 4358 * at arch_max_dma_pfn. 4359 */ 4360void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4361{ 4362 unsigned long nid; 4363 int i; 4364 4365 /* Sort early_node_map as initialisation assumes it is sorted */ 4366 sort_node_map(); 4367 4368 /* Record where the zone boundaries are */ 4369 memset(arch_zone_lowest_possible_pfn, 0, 4370 sizeof(arch_zone_lowest_possible_pfn)); 4371 memset(arch_zone_highest_possible_pfn, 0, 4372 sizeof(arch_zone_highest_possible_pfn)); 4373 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4374 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4375 for (i = 1; i < MAX_NR_ZONES; i++) { 4376 if (i == ZONE_MOVABLE) 4377 continue; 4378 arch_zone_lowest_possible_pfn[i] = 4379 arch_zone_highest_possible_pfn[i-1]; 4380 arch_zone_highest_possible_pfn[i] = 4381 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4382 } 4383 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4384 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4385 4386 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4387 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4388 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4389 4390 /* Print out the zone ranges */ 4391 printk("Zone PFN ranges:\n"); 4392 for (i = 0; i < MAX_NR_ZONES; i++) { 4393 if (i == ZONE_MOVABLE) 4394 continue; 4395 printk(" %-8s ", zone_names[i]); 4396 if (arch_zone_lowest_possible_pfn[i] == 4397 arch_zone_highest_possible_pfn[i]) 4398 printk("empty\n"); 4399 else 4400 printk("%0#10lx -> %0#10lx\n", 4401 arch_zone_lowest_possible_pfn[i], 4402 arch_zone_highest_possible_pfn[i]); 4403 } 4404 4405 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4406 printk("Movable zone start PFN for each node\n"); 4407 for (i = 0; i < MAX_NUMNODES; i++) { 4408 if (zone_movable_pfn[i]) 4409 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4410 } 4411 4412 /* Print out the early_node_map[] */ 4413 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4414 for (i = 0; i < nr_nodemap_entries; i++) 4415 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4416 early_node_map[i].start_pfn, 4417 early_node_map[i].end_pfn); 4418 4419 /* Initialise every node */ 4420 mminit_verify_pageflags_layout(); 4421 setup_nr_node_ids(); 4422 for_each_online_node(nid) { 4423 pg_data_t *pgdat = NODE_DATA(nid); 4424 free_area_init_node(nid, NULL, 4425 find_min_pfn_for_node(nid), NULL); 4426 4427 /* Any memory on that node */ 4428 if (pgdat->node_present_pages) 4429 node_set_state(nid, N_HIGH_MEMORY); 4430 check_for_regular_memory(pgdat); 4431 } 4432} 4433 4434static int __init cmdline_parse_core(char *p, unsigned long *core) 4435{ 4436 unsigned long long coremem; 4437 if (!p) 4438 return -EINVAL; 4439 4440 coremem = memparse(p, &p); 4441 *core = coremem >> PAGE_SHIFT; 4442 4443 /* Paranoid check that UL is enough for the coremem value */ 4444 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4445 4446 return 0; 4447} 4448 4449/* 4450 * kernelcore=size sets the amount of memory for use for allocations that 4451 * cannot be reclaimed or migrated. 4452 */ 4453static int __init cmdline_parse_kernelcore(char *p) 4454{ 4455 return cmdline_parse_core(p, &required_kernelcore); 4456} 4457 4458/* 4459 * movablecore=size sets the amount of memory for use for allocations that 4460 * can be reclaimed or migrated. 4461 */ 4462static int __init cmdline_parse_movablecore(char *p) 4463{ 4464 return cmdline_parse_core(p, &required_movablecore); 4465} 4466 4467early_param("kernelcore", cmdline_parse_kernelcore); 4468early_param("movablecore", cmdline_parse_movablecore); 4469 4470#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4471 4472/** 4473 * set_dma_reserve - set the specified number of pages reserved in the first zone 4474 * @new_dma_reserve: The number of pages to mark reserved 4475 * 4476 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4477 * In the DMA zone, a significant percentage may be consumed by kernel image 4478 * and other unfreeable allocations which can skew the watermarks badly. This 4479 * function may optionally be used to account for unfreeable pages in the 4480 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4481 * smaller per-cpu batchsize. 4482 */ 4483void __init set_dma_reserve(unsigned long new_dma_reserve) 4484{ 4485 dma_reserve = new_dma_reserve; 4486} 4487 4488#ifndef CONFIG_NEED_MULTIPLE_NODES 4489struct pglist_data __refdata contig_page_data = { 4490#ifndef CONFIG_NO_BOOTMEM 4491 .bdata = &bootmem_node_data[0] 4492#endif 4493 }; 4494EXPORT_SYMBOL(contig_page_data); 4495#endif 4496 4497void __init free_area_init(unsigned long *zones_size) 4498{ 4499 free_area_init_node(0, zones_size, 4500 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4501} 4502 4503static int page_alloc_cpu_notify(struct notifier_block *self, 4504 unsigned long action, void *hcpu) 4505{ 4506 int cpu = (unsigned long)hcpu; 4507 4508 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4509 drain_pages(cpu); 4510 4511 /* 4512 * Spill the event counters of the dead processor 4513 * into the current processors event counters. 4514 * This artificially elevates the count of the current 4515 * processor. 4516 */ 4517 vm_events_fold_cpu(cpu); 4518 4519 /* 4520 * Zero the differential counters of the dead processor 4521 * so that the vm statistics are consistent. 4522 * 4523 * This is only okay since the processor is dead and cannot 4524 * race with what we are doing. 4525 */ 4526 refresh_cpu_vm_stats(cpu); 4527 } 4528 return NOTIFY_OK; 4529} 4530 4531void __init page_alloc_init(void) 4532{ 4533 hotcpu_notifier(page_alloc_cpu_notify, 0); 4534} 4535 4536/* 4537 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4538 * or min_free_kbytes changes. 4539 */ 4540static void calculate_totalreserve_pages(void) 4541{ 4542 struct pglist_data *pgdat; 4543 unsigned long reserve_pages = 0; 4544 enum zone_type i, j; 4545 4546 for_each_online_pgdat(pgdat) { 4547 for (i = 0; i < MAX_NR_ZONES; i++) { 4548 struct zone *zone = pgdat->node_zones + i; 4549 unsigned long max = 0; 4550 4551 /* Find valid and maximum lowmem_reserve in the zone */ 4552 for (j = i; j < MAX_NR_ZONES; j++) { 4553 if (zone->lowmem_reserve[j] > max) 4554 max = zone->lowmem_reserve[j]; 4555 } 4556 4557 /* we treat the high watermark as reserved pages. */ 4558 max += high_wmark_pages(zone); 4559 4560 if (max > zone->present_pages) 4561 max = zone->present_pages; 4562 reserve_pages += max; 4563 } 4564 } 4565 totalreserve_pages = reserve_pages; 4566} 4567 4568/* 4569 * setup_per_zone_lowmem_reserve - called whenever 4570 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4571 * has a correct pages reserved value, so an adequate number of 4572 * pages are left in the zone after a successful __alloc_pages(). 4573 */ 4574static void setup_per_zone_lowmem_reserve(void) 4575{ 4576 struct pglist_data *pgdat; 4577 enum zone_type j, idx; 4578 4579 for_each_online_pgdat(pgdat) { 4580 for (j = 0; j < MAX_NR_ZONES; j++) { 4581 struct zone *zone = pgdat->node_zones + j; 4582 unsigned long present_pages = zone->present_pages; 4583 4584 zone->lowmem_reserve[j] = 0; 4585 4586 idx = j; 4587 while (idx) { 4588 struct zone *lower_zone; 4589 4590 idx--; 4591 4592 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4593 sysctl_lowmem_reserve_ratio[idx] = 1; 4594 4595 lower_zone = pgdat->node_zones + idx; 4596 lower_zone->lowmem_reserve[j] = present_pages / 4597 sysctl_lowmem_reserve_ratio[idx]; 4598 present_pages += lower_zone->present_pages; 4599 } 4600 } 4601 } 4602 4603 /* update totalreserve_pages */ 4604 calculate_totalreserve_pages(); 4605} 4606 4607/** 4608 * setup_per_zone_wmarks - called when min_free_kbytes changes 4609 * or when memory is hot-{added|removed} 4610 * 4611 * Ensures that the watermark[min,low,high] values for each zone are set 4612 * correctly with respect to min_free_kbytes. 4613 */ 4614void setup_per_zone_wmarks(void) 4615{ 4616 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4617 unsigned long lowmem_pages = 0; 4618 struct zone *zone; 4619 unsigned long flags; 4620 4621 /* Calculate total number of !ZONE_HIGHMEM pages */ 4622 for_each_zone(zone) { 4623 if (!is_highmem(zone)) 4624 lowmem_pages += zone->present_pages; 4625 } 4626 4627 for_each_zone(zone) { 4628 u64 tmp; 4629 4630 spin_lock_irqsave(&zone->lock, flags); 4631 tmp = (u64)pages_min * zone->present_pages; 4632 do_div(tmp, lowmem_pages); 4633 if (is_highmem(zone)) { 4634 /* 4635 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4636 * need highmem pages, so cap pages_min to a small 4637 * value here. 4638 * 4639 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4640 * deltas controls asynch page reclaim, and so should 4641 * not be capped for highmem. 4642 */ 4643 int min_pages; 4644 4645 min_pages = zone->present_pages / 1024; 4646 if (min_pages < SWAP_CLUSTER_MAX) 4647 min_pages = SWAP_CLUSTER_MAX; 4648 if (min_pages > 128) 4649 min_pages = 128; 4650 zone->watermark[WMARK_MIN] = min_pages; 4651 } else { 4652 /* 4653 * If it's a lowmem zone, reserve a number of pages 4654 * proportionate to the zone's size. 4655 */ 4656 zone->watermark[WMARK_MIN] = tmp; 4657 } 4658 4659 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4660 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4661 setup_zone_migrate_reserve(zone); 4662 spin_unlock_irqrestore(&zone->lock, flags); 4663 } 4664 4665 /* update totalreserve_pages */ 4666 calculate_totalreserve_pages(); 4667} 4668 4669/* 4670 * The inactive anon list should be small enough that the VM never has to 4671 * do too much work, but large enough that each inactive page has a chance 4672 * to be referenced again before it is swapped out. 4673 * 4674 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4675 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4676 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4677 * the anonymous pages are kept on the inactive list. 4678 * 4679 * total target max 4680 * memory ratio inactive anon 4681 * ------------------------------------- 4682 * 10MB 1 5MB 4683 * 100MB 1 50MB 4684 * 1GB 3 250MB 4685 * 10GB 10 0.9GB 4686 * 100GB 31 3GB 4687 * 1TB 101 10GB 4688 * 10TB 320 32GB 4689 */ 4690void calculate_zone_inactive_ratio(struct zone *zone) 4691{ 4692 unsigned int gb, ratio; 4693 4694 /* Zone size in gigabytes */ 4695 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4696 if (gb) 4697 ratio = int_sqrt(10 * gb); 4698 else 4699 ratio = 1; 4700 4701 zone->inactive_ratio = ratio; 4702} 4703 4704static void __init setup_per_zone_inactive_ratio(void) 4705{ 4706 struct zone *zone; 4707 4708 for_each_zone(zone) 4709 calculate_zone_inactive_ratio(zone); 4710} 4711 4712/* 4713 * Initialise min_free_kbytes. 4714 * 4715 * For small machines we want it small (128k min). For large machines 4716 * we want it large (64MB max). But it is not linear, because network 4717 * bandwidth does not increase linearly with machine size. We use 4718 * 4719 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4720 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4721 * 4722 * which yields 4723 * 4724 * 16MB: 512k 4725 * 32MB: 724k 4726 * 64MB: 1024k 4727 * 128MB: 1448k 4728 * 256MB: 2048k 4729 * 512MB: 2896k 4730 * 1024MB: 4096k 4731 * 2048MB: 5792k 4732 * 4096MB: 8192k 4733 * 8192MB: 11584k 4734 * 16384MB: 16384k 4735 */ 4736static int __init init_per_zone_wmark_min(void) 4737{ 4738 unsigned long lowmem_kbytes; 4739 4740 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4741 4742 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4743 if (min_free_kbytes < 128) 4744 min_free_kbytes = 128; 4745 if (min_free_kbytes > 65536) 4746 min_free_kbytes = 65536; 4747 setup_per_zone_wmarks(); 4748 setup_per_zone_lowmem_reserve(); 4749 setup_per_zone_inactive_ratio(); 4750 return 0; 4751} 4752module_init(init_per_zone_wmark_min) 4753 4754/* 4755 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4756 * that we can call two helper functions whenever min_free_kbytes 4757 * changes. 4758 */ 4759int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4760 void __user *buffer, size_t *length, loff_t *ppos) 4761{ 4762 proc_dointvec(table, write, buffer, length, ppos); 4763 if (write) 4764 setup_per_zone_wmarks(); 4765 return 0; 4766} 4767 4768#ifdef CONFIG_NUMA 4769int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4770 void __user *buffer, size_t *length, loff_t *ppos) 4771{ 4772 struct zone *zone; 4773 int rc; 4774 4775 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4776 if (rc) 4777 return rc; 4778 4779 for_each_zone(zone) 4780 zone->min_unmapped_pages = (zone->present_pages * 4781 sysctl_min_unmapped_ratio) / 100; 4782 return 0; 4783} 4784 4785int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4786 void __user *buffer, size_t *length, loff_t *ppos) 4787{ 4788 struct zone *zone; 4789 int rc; 4790 4791 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4792 if (rc) 4793 return rc; 4794 4795 for_each_zone(zone) 4796 zone->min_slab_pages = (zone->present_pages * 4797 sysctl_min_slab_ratio) / 100; 4798 return 0; 4799} 4800#endif 4801 4802/* 4803 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4804 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4805 * whenever sysctl_lowmem_reserve_ratio changes. 4806 * 4807 * The reserve ratio obviously has absolutely no relation with the 4808 * minimum watermarks. The lowmem reserve ratio can only make sense 4809 * if in function of the boot time zone sizes. 4810 */ 4811int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4812 void __user *buffer, size_t *length, loff_t *ppos) 4813{ 4814 proc_dointvec_minmax(table, write, buffer, length, ppos); 4815 setup_per_zone_lowmem_reserve(); 4816 return 0; 4817} 4818 4819/* 4820 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4821 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4822 * can have before it gets flushed back to buddy allocator. 4823 */ 4824 4825int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4826 void __user *buffer, size_t *length, loff_t *ppos) 4827{ 4828 struct zone *zone; 4829 unsigned int cpu; 4830 int ret; 4831 4832 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 4833 if (!write || (ret == -EINVAL)) 4834 return ret; 4835 for_each_populated_zone(zone) { 4836 for_each_possible_cpu(cpu) { 4837 unsigned long high; 4838 high = zone->present_pages / percpu_pagelist_fraction; 4839 setup_pagelist_highmark( 4840 per_cpu_ptr(zone->pageset, cpu), high); 4841 } 4842 } 4843 return 0; 4844} 4845 4846int hashdist = HASHDIST_DEFAULT; 4847 4848#ifdef CONFIG_NUMA 4849static int __init set_hashdist(char *str) 4850{ 4851 if (!str) 4852 return 0; 4853 hashdist = simple_strtoul(str, &str, 0); 4854 return 1; 4855} 4856__setup("hashdist=", set_hashdist); 4857#endif 4858 4859/* 4860 * allocate a large system hash table from bootmem 4861 * - it is assumed that the hash table must contain an exact power-of-2 4862 * quantity of entries 4863 * - limit is the number of hash buckets, not the total allocation size 4864 */ 4865void *__init alloc_large_system_hash(const char *tablename, 4866 unsigned long bucketsize, 4867 unsigned long numentries, 4868 int scale, 4869 int flags, 4870 unsigned int *_hash_shift, 4871 unsigned int *_hash_mask, 4872 unsigned long limit) 4873{ 4874 unsigned long long max = limit; 4875 unsigned long log2qty, size; 4876 void *table = NULL; 4877 4878 /* allow the kernel cmdline to have a say */ 4879 if (!numentries) { 4880 /* round applicable memory size up to nearest megabyte */ 4881 numentries = nr_kernel_pages; 4882 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4883 numentries >>= 20 - PAGE_SHIFT; 4884 numentries <<= 20 - PAGE_SHIFT; 4885 4886 /* limit to 1 bucket per 2^scale bytes of low memory */ 4887 if (scale > PAGE_SHIFT) 4888 numentries >>= (scale - PAGE_SHIFT); 4889 else 4890 numentries <<= (PAGE_SHIFT - scale); 4891 4892 /* Make sure we've got at least a 0-order allocation.. */ 4893 if (unlikely(flags & HASH_SMALL)) { 4894 /* Makes no sense without HASH_EARLY */ 4895 WARN_ON(!(flags & HASH_EARLY)); 4896 if (!(numentries >> *_hash_shift)) { 4897 numentries = 1UL << *_hash_shift; 4898 BUG_ON(!numentries); 4899 } 4900 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4901 numentries = PAGE_SIZE / bucketsize; 4902 } 4903 numentries = roundup_pow_of_two(numentries); 4904 4905 /* limit allocation size to 1/16 total memory by default */ 4906 if (max == 0) { 4907 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4908 do_div(max, bucketsize); 4909 } 4910 4911 if (numentries > max) 4912 numentries = max; 4913 4914 log2qty = ilog2(numentries); 4915 4916 do { 4917 size = bucketsize << log2qty; 4918 if (flags & HASH_EARLY) 4919 table = alloc_bootmem_nopanic(size); 4920 else if (hashdist) 4921 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4922 else { 4923 /* 4924 * If bucketsize is not a power-of-two, we may free 4925 * some pages at the end of hash table which 4926 * alloc_pages_exact() automatically does 4927 */ 4928 if (get_order(size) < MAX_ORDER) { 4929 table = alloc_pages_exact(size, GFP_ATOMIC); 4930 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4931 } 4932 } 4933 } while (!table && size > PAGE_SIZE && --log2qty); 4934 4935 if (!table) 4936 panic("Failed to allocate %s hash table\n", tablename); 4937 4938 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4939 tablename, 4940 (1U << log2qty), 4941 ilog2(size) - PAGE_SHIFT, 4942 size); 4943 4944 if (_hash_shift) 4945 *_hash_shift = log2qty; 4946 if (_hash_mask) 4947 *_hash_mask = (1 << log2qty) - 1; 4948 4949 return table; 4950} 4951 4952/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4953static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4954 unsigned long pfn) 4955{ 4956#ifdef CONFIG_SPARSEMEM 4957 return __pfn_to_section(pfn)->pageblock_flags; 4958#else 4959 return zone->pageblock_flags; 4960#endif /* CONFIG_SPARSEMEM */ 4961} 4962 4963static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4964{ 4965#ifdef CONFIG_SPARSEMEM 4966 pfn &= (PAGES_PER_SECTION-1); 4967 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4968#else 4969 pfn = pfn - zone->zone_start_pfn; 4970 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4971#endif /* CONFIG_SPARSEMEM */ 4972} 4973 4974/** 4975 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4976 * @page: The page within the block of interest 4977 * @start_bitidx: The first bit of interest to retrieve 4978 * @end_bitidx: The last bit of interest 4979 * returns pageblock_bits flags 4980 */ 4981unsigned long get_pageblock_flags_group(struct page *page, 4982 int start_bitidx, int end_bitidx) 4983{ 4984 struct zone *zone; 4985 unsigned long *bitmap; 4986 unsigned long pfn, bitidx; 4987 unsigned long flags = 0; 4988 unsigned long value = 1; 4989 4990 zone = page_zone(page); 4991 pfn = page_to_pfn(page); 4992 bitmap = get_pageblock_bitmap(zone, pfn); 4993 bitidx = pfn_to_bitidx(zone, pfn); 4994 4995 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4996 if (test_bit(bitidx + start_bitidx, bitmap)) 4997 flags |= value; 4998 4999 return flags; 5000} 5001 5002/** 5003 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5004 * @page: The page within the block of interest 5005 * @start_bitidx: The first bit of interest 5006 * @end_bitidx: The last bit of interest 5007 * @flags: The flags to set 5008 */ 5009void set_pageblock_flags_group(struct page *page, unsigned long flags, 5010 int start_bitidx, int end_bitidx) 5011{ 5012 struct zone *zone; 5013 unsigned long *bitmap; 5014 unsigned long pfn, bitidx; 5015 unsigned long value = 1; 5016 5017 zone = page_zone(page); 5018 pfn = page_to_pfn(page); 5019 bitmap = get_pageblock_bitmap(zone, pfn); 5020 bitidx = pfn_to_bitidx(zone, pfn); 5021 VM_BUG_ON(pfn < zone->zone_start_pfn); 5022 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5023 5024 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5025 if (flags & value) 5026 __set_bit(bitidx + start_bitidx, bitmap); 5027 else 5028 __clear_bit(bitidx + start_bitidx, bitmap); 5029} 5030 5031/* 5032 * This is designed as sub function...plz see page_isolation.c also. 5033 * set/clear page block's type to be ISOLATE. 5034 * page allocater never alloc memory from ISOLATE block. 5035 */ 5036 5037int set_migratetype_isolate(struct page *page) 5038{ 5039 struct zone *zone; 5040 struct page *curr_page; 5041 unsigned long flags, pfn, iter; 5042 unsigned long immobile = 0; 5043 struct memory_isolate_notify arg; 5044 int notifier_ret; 5045 int ret = -EBUSY; 5046 int zone_idx; 5047 5048 zone = page_zone(page); 5049 zone_idx = zone_idx(zone); 5050 5051 spin_lock_irqsave(&zone->lock, flags); 5052 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE || 5053 zone_idx == ZONE_MOVABLE) { 5054 ret = 0; 5055 goto out; 5056 } 5057 5058 pfn = page_to_pfn(page); 5059 arg.start_pfn = pfn; 5060 arg.nr_pages = pageblock_nr_pages; 5061 arg.pages_found = 0; 5062 5063 /* 5064 * It may be possible to isolate a pageblock even if the 5065 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5066 * notifier chain is used by balloon drivers to return the 5067 * number of pages in a range that are held by the balloon 5068 * driver to shrink memory. If all the pages are accounted for 5069 * by balloons, are free, or on the LRU, isolation can continue. 5070 * Later, for example, when memory hotplug notifier runs, these 5071 * pages reported as "can be isolated" should be isolated(freed) 5072 * by the balloon driver through the memory notifier chain. 5073 */ 5074 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5075 notifier_ret = notifier_to_errno(notifier_ret); 5076 if (notifier_ret || !arg.pages_found) 5077 goto out; 5078 5079 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) { 5080 if (!pfn_valid_within(pfn)) 5081 continue; 5082 5083 curr_page = pfn_to_page(iter); 5084 if (!page_count(curr_page) || PageLRU(curr_page)) 5085 continue; 5086 5087 immobile++; 5088 } 5089 5090 if (arg.pages_found == immobile) 5091 ret = 0; 5092 5093out: 5094 if (!ret) { 5095 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5096 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5097 } 5098 5099 spin_unlock_irqrestore(&zone->lock, flags); 5100 if (!ret) 5101 drain_all_pages(); 5102 return ret; 5103} 5104 5105void unset_migratetype_isolate(struct page *page) 5106{ 5107 struct zone *zone; 5108 unsigned long flags; 5109 zone = page_zone(page); 5110 spin_lock_irqsave(&zone->lock, flags); 5111 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5112 goto out; 5113 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5114 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5115out: 5116 spin_unlock_irqrestore(&zone->lock, flags); 5117} 5118 5119#ifdef CONFIG_MEMORY_HOTREMOVE 5120/* 5121 * All pages in the range must be isolated before calling this. 5122 */ 5123void 5124__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5125{ 5126 struct page *page; 5127 struct zone *zone; 5128 int order, i; 5129 unsigned long pfn; 5130 unsigned long flags; 5131 /* find the first valid pfn */ 5132 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5133 if (pfn_valid(pfn)) 5134 break; 5135 if (pfn == end_pfn) 5136 return; 5137 zone = page_zone(pfn_to_page(pfn)); 5138 spin_lock_irqsave(&zone->lock, flags); 5139 pfn = start_pfn; 5140 while (pfn < end_pfn) { 5141 if (!pfn_valid(pfn)) { 5142 pfn++; 5143 continue; 5144 } 5145 page = pfn_to_page(pfn); 5146 BUG_ON(page_count(page)); 5147 BUG_ON(!PageBuddy(page)); 5148 order = page_order(page); 5149#ifdef CONFIG_DEBUG_VM 5150 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5151 pfn, 1 << order, end_pfn); 5152#endif 5153 list_del(&page->lru); 5154 rmv_page_order(page); 5155 zone->free_area[order].nr_free--; 5156 __mod_zone_page_state(zone, NR_FREE_PAGES, 5157 - (1UL << order)); 5158 for (i = 0; i < (1 << order); i++) 5159 SetPageReserved((page+i)); 5160 pfn += (1 << order); 5161 } 5162 spin_unlock_irqrestore(&zone->lock, flags); 5163} 5164#endif 5165 5166#ifdef CONFIG_MEMORY_FAILURE 5167bool is_free_buddy_page(struct page *page) 5168{ 5169 struct zone *zone = page_zone(page); 5170 unsigned long pfn = page_to_pfn(page); 5171 unsigned long flags; 5172 int order; 5173 5174 spin_lock_irqsave(&zone->lock, flags); 5175 for (order = 0; order < MAX_ORDER; order++) { 5176 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5177 5178 if (PageBuddy(page_head) && page_order(page_head) >= order) 5179 break; 5180 } 5181 spin_unlock_irqrestore(&zone->lock, flags); 5182 5183 return order < MAX_ORDER; 5184} 5185#endif 5186