page_alloc.c revision 08677214e318297f228237be0042aac754f48f1d
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_HUGETLB_PAGE_SIZE_VARIABLE 80int pageblock_order __read_mostly; 81#endif 82 83static void __free_pages_ok(struct page *page, unsigned int order); 84 85/* 86 * results with 256, 32 in the lowmem_reserve sysctl: 87 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 88 * 1G machine -> (16M dma, 784M normal, 224M high) 89 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 90 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 91 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 92 * 93 * TBD: should special case ZONE_DMA32 machines here - in those we normally 94 * don't need any ZONE_NORMAL reservation 95 */ 96int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 97#ifdef CONFIG_ZONE_DMA 98 256, 99#endif 100#ifdef CONFIG_ZONE_DMA32 101 256, 102#endif 103#ifdef CONFIG_HIGHMEM 104 32, 105#endif 106 32, 107}; 108 109EXPORT_SYMBOL(totalram_pages); 110 111static char * const zone_names[MAX_NR_ZONES] = { 112#ifdef CONFIG_ZONE_DMA 113 "DMA", 114#endif 115#ifdef CONFIG_ZONE_DMA32 116 "DMA32", 117#endif 118 "Normal", 119#ifdef CONFIG_HIGHMEM 120 "HighMem", 121#endif 122 "Movable", 123}; 124 125int min_free_kbytes = 1024; 126 127static unsigned long __meminitdata nr_kernel_pages; 128static unsigned long __meminitdata nr_all_pages; 129static unsigned long __meminitdata dma_reserve; 130 131#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 132 /* 133 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 134 * ranges of memory (RAM) that may be registered with add_active_range(). 135 * Ranges passed to add_active_range() will be merged if possible 136 * so the number of times add_active_range() can be called is 137 * related to the number of nodes and the number of holes 138 */ 139 #ifdef CONFIG_MAX_ACTIVE_REGIONS 140 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 141 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 142 #else 143 #if MAX_NUMNODES >= 32 144 /* If there can be many nodes, allow up to 50 holes per node */ 145 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 146 #else 147 /* By default, allow up to 256 distinct regions */ 148 #define MAX_ACTIVE_REGIONS 256 149 #endif 150 #endif 151 152 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 153 static int __meminitdata nr_nodemap_entries; 154 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 155 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 156 static unsigned long __initdata required_kernelcore; 157 static unsigned long __initdata required_movablecore; 158 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 159 160 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 161 int movable_zone; 162 EXPORT_SYMBOL(movable_zone); 163#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 164 165#if MAX_NUMNODES > 1 166int nr_node_ids __read_mostly = MAX_NUMNODES; 167int nr_online_nodes __read_mostly = 1; 168EXPORT_SYMBOL(nr_node_ids); 169EXPORT_SYMBOL(nr_online_nodes); 170#endif 171 172int page_group_by_mobility_disabled __read_mostly; 173 174static void set_pageblock_migratetype(struct page *page, int migratetype) 175{ 176 177 if (unlikely(page_group_by_mobility_disabled)) 178 migratetype = MIGRATE_UNMOVABLE; 179 180 set_pageblock_flags_group(page, (unsigned long)migratetype, 181 PB_migrate, PB_migrate_end); 182} 183 184bool oom_killer_disabled __read_mostly; 185 186#ifdef CONFIG_DEBUG_VM 187static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 188{ 189 int ret = 0; 190 unsigned seq; 191 unsigned long pfn = page_to_pfn(page); 192 193 do { 194 seq = zone_span_seqbegin(zone); 195 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 196 ret = 1; 197 else if (pfn < zone->zone_start_pfn) 198 ret = 1; 199 } while (zone_span_seqretry(zone, seq)); 200 201 return ret; 202} 203 204static int page_is_consistent(struct zone *zone, struct page *page) 205{ 206 if (!pfn_valid_within(page_to_pfn(page))) 207 return 0; 208 if (zone != page_zone(page)) 209 return 0; 210 211 return 1; 212} 213/* 214 * Temporary debugging check for pages not lying within a given zone. 215 */ 216static int bad_range(struct zone *zone, struct page *page) 217{ 218 if (page_outside_zone_boundaries(zone, page)) 219 return 1; 220 if (!page_is_consistent(zone, page)) 221 return 1; 222 223 return 0; 224} 225#else 226static inline int bad_range(struct zone *zone, struct page *page) 227{ 228 return 0; 229} 230#endif 231 232static void bad_page(struct page *page) 233{ 234 static unsigned long resume; 235 static unsigned long nr_shown; 236 static unsigned long nr_unshown; 237 238 /* Don't complain about poisoned pages */ 239 if (PageHWPoison(page)) { 240 __ClearPageBuddy(page); 241 return; 242 } 243 244 /* 245 * Allow a burst of 60 reports, then keep quiet for that minute; 246 * or allow a steady drip of one report per second. 247 */ 248 if (nr_shown == 60) { 249 if (time_before(jiffies, resume)) { 250 nr_unshown++; 251 goto out; 252 } 253 if (nr_unshown) { 254 printk(KERN_ALERT 255 "BUG: Bad page state: %lu messages suppressed\n", 256 nr_unshown); 257 nr_unshown = 0; 258 } 259 nr_shown = 0; 260 } 261 if (nr_shown++ == 0) 262 resume = jiffies + 60 * HZ; 263 264 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 265 current->comm, page_to_pfn(page)); 266 printk(KERN_ALERT 267 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n", 268 page, (void *)page->flags, page_count(page), 269 page_mapcount(page), page->mapping, page->index); 270 271 dump_stack(); 272out: 273 /* Leave bad fields for debug, except PageBuddy could make trouble */ 274 __ClearPageBuddy(page); 275 add_taint(TAINT_BAD_PAGE); 276} 277 278/* 279 * Higher-order pages are called "compound pages". They are structured thusly: 280 * 281 * The first PAGE_SIZE page is called the "head page". 282 * 283 * The remaining PAGE_SIZE pages are called "tail pages". 284 * 285 * All pages have PG_compound set. All pages have their ->private pointing at 286 * the head page (even the head page has this). 287 * 288 * The first tail page's ->lru.next holds the address of the compound page's 289 * put_page() function. Its ->lru.prev holds the order of allocation. 290 * This usage means that zero-order pages may not be compound. 291 */ 292 293static void free_compound_page(struct page *page) 294{ 295 __free_pages_ok(page, compound_order(page)); 296} 297 298void prep_compound_page(struct page *page, unsigned long order) 299{ 300 int i; 301 int nr_pages = 1 << order; 302 303 set_compound_page_dtor(page, free_compound_page); 304 set_compound_order(page, order); 305 __SetPageHead(page); 306 for (i = 1; i < nr_pages; i++) { 307 struct page *p = page + i; 308 309 __SetPageTail(p); 310 p->first_page = page; 311 } 312} 313 314static int destroy_compound_page(struct page *page, unsigned long order) 315{ 316 int i; 317 int nr_pages = 1 << order; 318 int bad = 0; 319 320 if (unlikely(compound_order(page) != order) || 321 unlikely(!PageHead(page))) { 322 bad_page(page); 323 bad++; 324 } 325 326 __ClearPageHead(page); 327 328 for (i = 1; i < nr_pages; i++) { 329 struct page *p = page + i; 330 331 if (unlikely(!PageTail(p) || (p->first_page != page))) { 332 bad_page(page); 333 bad++; 334 } 335 __ClearPageTail(p); 336 } 337 338 return bad; 339} 340 341static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 342{ 343 int i; 344 345 /* 346 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 347 * and __GFP_HIGHMEM from hard or soft interrupt context. 348 */ 349 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 350 for (i = 0; i < (1 << order); i++) 351 clear_highpage(page + i); 352} 353 354static inline void set_page_order(struct page *page, int order) 355{ 356 set_page_private(page, order); 357 __SetPageBuddy(page); 358} 359 360static inline void rmv_page_order(struct page *page) 361{ 362 __ClearPageBuddy(page); 363 set_page_private(page, 0); 364} 365 366/* 367 * Locate the struct page for both the matching buddy in our 368 * pair (buddy1) and the combined O(n+1) page they form (page). 369 * 370 * 1) Any buddy B1 will have an order O twin B2 which satisfies 371 * the following equation: 372 * B2 = B1 ^ (1 << O) 373 * For example, if the starting buddy (buddy2) is #8 its order 374 * 1 buddy is #10: 375 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 376 * 377 * 2) Any buddy B will have an order O+1 parent P which 378 * satisfies the following equation: 379 * P = B & ~(1 << O) 380 * 381 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 382 */ 383static inline struct page * 384__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 385{ 386 unsigned long buddy_idx = page_idx ^ (1 << order); 387 388 return page + (buddy_idx - page_idx); 389} 390 391static inline unsigned long 392__find_combined_index(unsigned long page_idx, unsigned int order) 393{ 394 return (page_idx & ~(1 << order)); 395} 396 397/* 398 * This function checks whether a page is free && is the buddy 399 * we can do coalesce a page and its buddy if 400 * (a) the buddy is not in a hole && 401 * (b) the buddy is in the buddy system && 402 * (c) a page and its buddy have the same order && 403 * (d) a page and its buddy are in the same zone. 404 * 405 * For recording whether a page is in the buddy system, we use PG_buddy. 406 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 407 * 408 * For recording page's order, we use page_private(page). 409 */ 410static inline int page_is_buddy(struct page *page, struct page *buddy, 411 int order) 412{ 413 if (!pfn_valid_within(page_to_pfn(buddy))) 414 return 0; 415 416 if (page_zone_id(page) != page_zone_id(buddy)) 417 return 0; 418 419 if (PageBuddy(buddy) && page_order(buddy) == order) { 420 VM_BUG_ON(page_count(buddy) != 0); 421 return 1; 422 } 423 return 0; 424} 425 426/* 427 * Freeing function for a buddy system allocator. 428 * 429 * The concept of a buddy system is to maintain direct-mapped table 430 * (containing bit values) for memory blocks of various "orders". 431 * The bottom level table contains the map for the smallest allocatable 432 * units of memory (here, pages), and each level above it describes 433 * pairs of units from the levels below, hence, "buddies". 434 * At a high level, all that happens here is marking the table entry 435 * at the bottom level available, and propagating the changes upward 436 * as necessary, plus some accounting needed to play nicely with other 437 * parts of the VM system. 438 * At each level, we keep a list of pages, which are heads of continuous 439 * free pages of length of (1 << order) and marked with PG_buddy. Page's 440 * order is recorded in page_private(page) field. 441 * So when we are allocating or freeing one, we can derive the state of the 442 * other. That is, if we allocate a small block, and both were 443 * free, the remainder of the region must be split into blocks. 444 * If a block is freed, and its buddy is also free, then this 445 * triggers coalescing into a block of larger size. 446 * 447 * -- wli 448 */ 449 450static inline void __free_one_page(struct page *page, 451 struct zone *zone, unsigned int order, 452 int migratetype) 453{ 454 unsigned long page_idx; 455 456 if (unlikely(PageCompound(page))) 457 if (unlikely(destroy_compound_page(page, order))) 458 return; 459 460 VM_BUG_ON(migratetype == -1); 461 462 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 463 464 VM_BUG_ON(page_idx & ((1 << order) - 1)); 465 VM_BUG_ON(bad_range(zone, page)); 466 467 while (order < MAX_ORDER-1) { 468 unsigned long combined_idx; 469 struct page *buddy; 470 471 buddy = __page_find_buddy(page, page_idx, order); 472 if (!page_is_buddy(page, buddy, order)) 473 break; 474 475 /* Our buddy is free, merge with it and move up one order. */ 476 list_del(&buddy->lru); 477 zone->free_area[order].nr_free--; 478 rmv_page_order(buddy); 479 combined_idx = __find_combined_index(page_idx, order); 480 page = page + (combined_idx - page_idx); 481 page_idx = combined_idx; 482 order++; 483 } 484 set_page_order(page, order); 485 list_add(&page->lru, 486 &zone->free_area[order].free_list[migratetype]); 487 zone->free_area[order].nr_free++; 488} 489 490/* 491 * free_page_mlock() -- clean up attempts to free and mlocked() page. 492 * Page should not be on lru, so no need to fix that up. 493 * free_pages_check() will verify... 494 */ 495static inline void free_page_mlock(struct page *page) 496{ 497 __dec_zone_page_state(page, NR_MLOCK); 498 __count_vm_event(UNEVICTABLE_MLOCKFREED); 499} 500 501static inline int free_pages_check(struct page *page) 502{ 503 if (unlikely(page_mapcount(page) | 504 (page->mapping != NULL) | 505 (atomic_read(&page->_count) != 0) | 506 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 507 bad_page(page); 508 return 1; 509 } 510 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 511 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 512 return 0; 513} 514 515/* 516 * Frees a number of pages from the PCP lists 517 * Assumes all pages on list are in same zone, and of same order. 518 * count is the number of pages to free. 519 * 520 * If the zone was previously in an "all pages pinned" state then look to 521 * see if this freeing clears that state. 522 * 523 * And clear the zone's pages_scanned counter, to hold off the "all pages are 524 * pinned" detection logic. 525 */ 526static void free_pcppages_bulk(struct zone *zone, int count, 527 struct per_cpu_pages *pcp) 528{ 529 int migratetype = 0; 530 int batch_free = 0; 531 532 spin_lock(&zone->lock); 533 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 534 zone->pages_scanned = 0; 535 536 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 537 while (count) { 538 struct page *page; 539 struct list_head *list; 540 541 /* 542 * Remove pages from lists in a round-robin fashion. A 543 * batch_free count is maintained that is incremented when an 544 * empty list is encountered. This is so more pages are freed 545 * off fuller lists instead of spinning excessively around empty 546 * lists 547 */ 548 do { 549 batch_free++; 550 if (++migratetype == MIGRATE_PCPTYPES) 551 migratetype = 0; 552 list = &pcp->lists[migratetype]; 553 } while (list_empty(list)); 554 555 do { 556 page = list_entry(list->prev, struct page, lru); 557 /* must delete as __free_one_page list manipulates */ 558 list_del(&page->lru); 559 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 560 __free_one_page(page, zone, 0, page_private(page)); 561 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 562 } while (--count && --batch_free && !list_empty(list)); 563 } 564 spin_unlock(&zone->lock); 565} 566 567static void free_one_page(struct zone *zone, struct page *page, int order, 568 int migratetype) 569{ 570 spin_lock(&zone->lock); 571 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 572 zone->pages_scanned = 0; 573 574 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 575 __free_one_page(page, zone, order, migratetype); 576 spin_unlock(&zone->lock); 577} 578 579static void __free_pages_ok(struct page *page, unsigned int order) 580{ 581 unsigned long flags; 582 int i; 583 int bad = 0; 584 int wasMlocked = __TestClearPageMlocked(page); 585 586 kmemcheck_free_shadow(page, order); 587 588 for (i = 0 ; i < (1 << order) ; ++i) 589 bad += free_pages_check(page + i); 590 if (bad) 591 return; 592 593 if (!PageHighMem(page)) { 594 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 595 debug_check_no_obj_freed(page_address(page), 596 PAGE_SIZE << order); 597 } 598 arch_free_page(page, order); 599 kernel_map_pages(page, 1 << order, 0); 600 601 local_irq_save(flags); 602 if (unlikely(wasMlocked)) 603 free_page_mlock(page); 604 __count_vm_events(PGFREE, 1 << order); 605 free_one_page(page_zone(page), page, order, 606 get_pageblock_migratetype(page)); 607 local_irq_restore(flags); 608} 609 610/* 611 * permit the bootmem allocator to evade page validation on high-order frees 612 */ 613void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 614{ 615 if (order == 0) { 616 __ClearPageReserved(page); 617 set_page_count(page, 0); 618 set_page_refcounted(page); 619 __free_page(page); 620 } else { 621 int loop; 622 623 prefetchw(page); 624 for (loop = 0; loop < BITS_PER_LONG; loop++) { 625 struct page *p = &page[loop]; 626 627 if (loop + 1 < BITS_PER_LONG) 628 prefetchw(p + 1); 629 __ClearPageReserved(p); 630 set_page_count(p, 0); 631 } 632 633 set_page_refcounted(page); 634 __free_pages(page, order); 635 } 636} 637 638 639/* 640 * The order of subdivision here is critical for the IO subsystem. 641 * Please do not alter this order without good reasons and regression 642 * testing. Specifically, as large blocks of memory are subdivided, 643 * the order in which smaller blocks are delivered depends on the order 644 * they're subdivided in this function. This is the primary factor 645 * influencing the order in which pages are delivered to the IO 646 * subsystem according to empirical testing, and this is also justified 647 * by considering the behavior of a buddy system containing a single 648 * large block of memory acted on by a series of small allocations. 649 * This behavior is a critical factor in sglist merging's success. 650 * 651 * -- wli 652 */ 653static inline void expand(struct zone *zone, struct page *page, 654 int low, int high, struct free_area *area, 655 int migratetype) 656{ 657 unsigned long size = 1 << high; 658 659 while (high > low) { 660 area--; 661 high--; 662 size >>= 1; 663 VM_BUG_ON(bad_range(zone, &page[size])); 664 list_add(&page[size].lru, &area->free_list[migratetype]); 665 area->nr_free++; 666 set_page_order(&page[size], high); 667 } 668} 669 670/* 671 * This page is about to be returned from the page allocator 672 */ 673static inline int check_new_page(struct page *page) 674{ 675 if (unlikely(page_mapcount(page) | 676 (page->mapping != NULL) | 677 (atomic_read(&page->_count) != 0) | 678 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 679 bad_page(page); 680 return 1; 681 } 682 return 0; 683} 684 685static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 686{ 687 int i; 688 689 for (i = 0; i < (1 << order); i++) { 690 struct page *p = page + i; 691 if (unlikely(check_new_page(p))) 692 return 1; 693 } 694 695 set_page_private(page, 0); 696 set_page_refcounted(page); 697 698 arch_alloc_page(page, order); 699 kernel_map_pages(page, 1 << order, 1); 700 701 if (gfp_flags & __GFP_ZERO) 702 prep_zero_page(page, order, gfp_flags); 703 704 if (order && (gfp_flags & __GFP_COMP)) 705 prep_compound_page(page, order); 706 707 return 0; 708} 709 710/* 711 * Go through the free lists for the given migratetype and remove 712 * the smallest available page from the freelists 713 */ 714static inline 715struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 716 int migratetype) 717{ 718 unsigned int current_order; 719 struct free_area * area; 720 struct page *page; 721 722 /* Find a page of the appropriate size in the preferred list */ 723 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 724 area = &(zone->free_area[current_order]); 725 if (list_empty(&area->free_list[migratetype])) 726 continue; 727 728 page = list_entry(area->free_list[migratetype].next, 729 struct page, lru); 730 list_del(&page->lru); 731 rmv_page_order(page); 732 area->nr_free--; 733 expand(zone, page, order, current_order, area, migratetype); 734 return page; 735 } 736 737 return NULL; 738} 739 740 741/* 742 * This array describes the order lists are fallen back to when 743 * the free lists for the desirable migrate type are depleted 744 */ 745static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 746 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 747 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 748 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 749 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 750}; 751 752/* 753 * Move the free pages in a range to the free lists of the requested type. 754 * Note that start_page and end_pages are not aligned on a pageblock 755 * boundary. If alignment is required, use move_freepages_block() 756 */ 757static int move_freepages(struct zone *zone, 758 struct page *start_page, struct page *end_page, 759 int migratetype) 760{ 761 struct page *page; 762 unsigned long order; 763 int pages_moved = 0; 764 765#ifndef CONFIG_HOLES_IN_ZONE 766 /* 767 * page_zone is not safe to call in this context when 768 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 769 * anyway as we check zone boundaries in move_freepages_block(). 770 * Remove at a later date when no bug reports exist related to 771 * grouping pages by mobility 772 */ 773 BUG_ON(page_zone(start_page) != page_zone(end_page)); 774#endif 775 776 for (page = start_page; page <= end_page;) { 777 /* Make sure we are not inadvertently changing nodes */ 778 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 779 780 if (!pfn_valid_within(page_to_pfn(page))) { 781 page++; 782 continue; 783 } 784 785 if (!PageBuddy(page)) { 786 page++; 787 continue; 788 } 789 790 order = page_order(page); 791 list_del(&page->lru); 792 list_add(&page->lru, 793 &zone->free_area[order].free_list[migratetype]); 794 page += 1 << order; 795 pages_moved += 1 << order; 796 } 797 798 return pages_moved; 799} 800 801static int move_freepages_block(struct zone *zone, struct page *page, 802 int migratetype) 803{ 804 unsigned long start_pfn, end_pfn; 805 struct page *start_page, *end_page; 806 807 start_pfn = page_to_pfn(page); 808 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 809 start_page = pfn_to_page(start_pfn); 810 end_page = start_page + pageblock_nr_pages - 1; 811 end_pfn = start_pfn + pageblock_nr_pages - 1; 812 813 /* Do not cross zone boundaries */ 814 if (start_pfn < zone->zone_start_pfn) 815 start_page = page; 816 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 817 return 0; 818 819 return move_freepages(zone, start_page, end_page, migratetype); 820} 821 822static void change_pageblock_range(struct page *pageblock_page, 823 int start_order, int migratetype) 824{ 825 int nr_pageblocks = 1 << (start_order - pageblock_order); 826 827 while (nr_pageblocks--) { 828 set_pageblock_migratetype(pageblock_page, migratetype); 829 pageblock_page += pageblock_nr_pages; 830 } 831} 832 833/* Remove an element from the buddy allocator from the fallback list */ 834static inline struct page * 835__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 836{ 837 struct free_area * area; 838 int current_order; 839 struct page *page; 840 int migratetype, i; 841 842 /* Find the largest possible block of pages in the other list */ 843 for (current_order = MAX_ORDER-1; current_order >= order; 844 --current_order) { 845 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 846 migratetype = fallbacks[start_migratetype][i]; 847 848 /* MIGRATE_RESERVE handled later if necessary */ 849 if (migratetype == MIGRATE_RESERVE) 850 continue; 851 852 area = &(zone->free_area[current_order]); 853 if (list_empty(&area->free_list[migratetype])) 854 continue; 855 856 page = list_entry(area->free_list[migratetype].next, 857 struct page, lru); 858 area->nr_free--; 859 860 /* 861 * If breaking a large block of pages, move all free 862 * pages to the preferred allocation list. If falling 863 * back for a reclaimable kernel allocation, be more 864 * agressive about taking ownership of free pages 865 */ 866 if (unlikely(current_order >= (pageblock_order >> 1)) || 867 start_migratetype == MIGRATE_RECLAIMABLE || 868 page_group_by_mobility_disabled) { 869 unsigned long pages; 870 pages = move_freepages_block(zone, page, 871 start_migratetype); 872 873 /* Claim the whole block if over half of it is free */ 874 if (pages >= (1 << (pageblock_order-1)) || 875 page_group_by_mobility_disabled) 876 set_pageblock_migratetype(page, 877 start_migratetype); 878 879 migratetype = start_migratetype; 880 } 881 882 /* Remove the page from the freelists */ 883 list_del(&page->lru); 884 rmv_page_order(page); 885 886 /* Take ownership for orders >= pageblock_order */ 887 if (current_order >= pageblock_order) 888 change_pageblock_range(page, current_order, 889 start_migratetype); 890 891 expand(zone, page, order, current_order, area, migratetype); 892 893 trace_mm_page_alloc_extfrag(page, order, current_order, 894 start_migratetype, migratetype); 895 896 return page; 897 } 898 } 899 900 return NULL; 901} 902 903/* 904 * Do the hard work of removing an element from the buddy allocator. 905 * Call me with the zone->lock already held. 906 */ 907static struct page *__rmqueue(struct zone *zone, unsigned int order, 908 int migratetype) 909{ 910 struct page *page; 911 912retry_reserve: 913 page = __rmqueue_smallest(zone, order, migratetype); 914 915 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 916 page = __rmqueue_fallback(zone, order, migratetype); 917 918 /* 919 * Use MIGRATE_RESERVE rather than fail an allocation. goto 920 * is used because __rmqueue_smallest is an inline function 921 * and we want just one call site 922 */ 923 if (!page) { 924 migratetype = MIGRATE_RESERVE; 925 goto retry_reserve; 926 } 927 } 928 929 trace_mm_page_alloc_zone_locked(page, order, migratetype); 930 return page; 931} 932 933/* 934 * Obtain a specified number of elements from the buddy allocator, all under 935 * a single hold of the lock, for efficiency. Add them to the supplied list. 936 * Returns the number of new pages which were placed at *list. 937 */ 938static int rmqueue_bulk(struct zone *zone, unsigned int order, 939 unsigned long count, struct list_head *list, 940 int migratetype, int cold) 941{ 942 int i; 943 944 spin_lock(&zone->lock); 945 for (i = 0; i < count; ++i) { 946 struct page *page = __rmqueue(zone, order, migratetype); 947 if (unlikely(page == NULL)) 948 break; 949 950 /* 951 * Split buddy pages returned by expand() are received here 952 * in physical page order. The page is added to the callers and 953 * list and the list head then moves forward. From the callers 954 * perspective, the linked list is ordered by page number in 955 * some conditions. This is useful for IO devices that can 956 * merge IO requests if the physical pages are ordered 957 * properly. 958 */ 959 if (likely(cold == 0)) 960 list_add(&page->lru, list); 961 else 962 list_add_tail(&page->lru, list); 963 set_page_private(page, migratetype); 964 list = &page->lru; 965 } 966 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 967 spin_unlock(&zone->lock); 968 return i; 969} 970 971#ifdef CONFIG_NUMA 972/* 973 * Called from the vmstat counter updater to drain pagesets of this 974 * currently executing processor on remote nodes after they have 975 * expired. 976 * 977 * Note that this function must be called with the thread pinned to 978 * a single processor. 979 */ 980void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 981{ 982 unsigned long flags; 983 int to_drain; 984 985 local_irq_save(flags); 986 if (pcp->count >= pcp->batch) 987 to_drain = pcp->batch; 988 else 989 to_drain = pcp->count; 990 free_pcppages_bulk(zone, to_drain, pcp); 991 pcp->count -= to_drain; 992 local_irq_restore(flags); 993} 994#endif 995 996/* 997 * Drain pages of the indicated processor. 998 * 999 * The processor must either be the current processor and the 1000 * thread pinned to the current processor or a processor that 1001 * is not online. 1002 */ 1003static void drain_pages(unsigned int cpu) 1004{ 1005 unsigned long flags; 1006 struct zone *zone; 1007 1008 for_each_populated_zone(zone) { 1009 struct per_cpu_pageset *pset; 1010 struct per_cpu_pages *pcp; 1011 1012 pset = zone_pcp(zone, cpu); 1013 1014 pcp = &pset->pcp; 1015 local_irq_save(flags); 1016 free_pcppages_bulk(zone, pcp->count, pcp); 1017 pcp->count = 0; 1018 local_irq_restore(flags); 1019 } 1020} 1021 1022/* 1023 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1024 */ 1025void drain_local_pages(void *arg) 1026{ 1027 drain_pages(smp_processor_id()); 1028} 1029 1030/* 1031 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1032 */ 1033void drain_all_pages(void) 1034{ 1035 on_each_cpu(drain_local_pages, NULL, 1); 1036} 1037 1038#ifdef CONFIG_HIBERNATION 1039 1040void mark_free_pages(struct zone *zone) 1041{ 1042 unsigned long pfn, max_zone_pfn; 1043 unsigned long flags; 1044 int order, t; 1045 struct list_head *curr; 1046 1047 if (!zone->spanned_pages) 1048 return; 1049 1050 spin_lock_irqsave(&zone->lock, flags); 1051 1052 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1053 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1054 if (pfn_valid(pfn)) { 1055 struct page *page = pfn_to_page(pfn); 1056 1057 if (!swsusp_page_is_forbidden(page)) 1058 swsusp_unset_page_free(page); 1059 } 1060 1061 for_each_migratetype_order(order, t) { 1062 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1063 unsigned long i; 1064 1065 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1066 for (i = 0; i < (1UL << order); i++) 1067 swsusp_set_page_free(pfn_to_page(pfn + i)); 1068 } 1069 } 1070 spin_unlock_irqrestore(&zone->lock, flags); 1071} 1072#endif /* CONFIG_PM */ 1073 1074/* 1075 * Free a 0-order page 1076 */ 1077static void free_hot_cold_page(struct page *page, int cold) 1078{ 1079 struct zone *zone = page_zone(page); 1080 struct per_cpu_pages *pcp; 1081 unsigned long flags; 1082 int migratetype; 1083 int wasMlocked = __TestClearPageMlocked(page); 1084 1085 kmemcheck_free_shadow(page, 0); 1086 1087 if (PageAnon(page)) 1088 page->mapping = NULL; 1089 if (free_pages_check(page)) 1090 return; 1091 1092 if (!PageHighMem(page)) { 1093 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1094 debug_check_no_obj_freed(page_address(page), PAGE_SIZE); 1095 } 1096 arch_free_page(page, 0); 1097 kernel_map_pages(page, 1, 0); 1098 1099 pcp = &zone_pcp(zone, get_cpu())->pcp; 1100 migratetype = get_pageblock_migratetype(page); 1101 set_page_private(page, migratetype); 1102 local_irq_save(flags); 1103 if (unlikely(wasMlocked)) 1104 free_page_mlock(page); 1105 __count_vm_event(PGFREE); 1106 1107 /* 1108 * We only track unmovable, reclaimable and movable on pcp lists. 1109 * Free ISOLATE pages back to the allocator because they are being 1110 * offlined but treat RESERVE as movable pages so we can get those 1111 * areas back if necessary. Otherwise, we may have to free 1112 * excessively into the page allocator 1113 */ 1114 if (migratetype >= MIGRATE_PCPTYPES) { 1115 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1116 free_one_page(zone, page, 0, migratetype); 1117 goto out; 1118 } 1119 migratetype = MIGRATE_MOVABLE; 1120 } 1121 1122 if (cold) 1123 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1124 else 1125 list_add(&page->lru, &pcp->lists[migratetype]); 1126 pcp->count++; 1127 if (pcp->count >= pcp->high) { 1128 free_pcppages_bulk(zone, pcp->batch, pcp); 1129 pcp->count -= pcp->batch; 1130 } 1131 1132out: 1133 local_irq_restore(flags); 1134 put_cpu(); 1135} 1136 1137void free_hot_page(struct page *page) 1138{ 1139 trace_mm_page_free_direct(page, 0); 1140 free_hot_cold_page(page, 0); 1141} 1142 1143/* 1144 * split_page takes a non-compound higher-order page, and splits it into 1145 * n (1<<order) sub-pages: page[0..n] 1146 * Each sub-page must be freed individually. 1147 * 1148 * Note: this is probably too low level an operation for use in drivers. 1149 * Please consult with lkml before using this in your driver. 1150 */ 1151void split_page(struct page *page, unsigned int order) 1152{ 1153 int i; 1154 1155 VM_BUG_ON(PageCompound(page)); 1156 VM_BUG_ON(!page_count(page)); 1157 1158#ifdef CONFIG_KMEMCHECK 1159 /* 1160 * Split shadow pages too, because free(page[0]) would 1161 * otherwise free the whole shadow. 1162 */ 1163 if (kmemcheck_page_is_tracked(page)) 1164 split_page(virt_to_page(page[0].shadow), order); 1165#endif 1166 1167 for (i = 1; i < (1 << order); i++) 1168 set_page_refcounted(page + i); 1169} 1170 1171/* 1172 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1173 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1174 * or two. 1175 */ 1176static inline 1177struct page *buffered_rmqueue(struct zone *preferred_zone, 1178 struct zone *zone, int order, gfp_t gfp_flags, 1179 int migratetype) 1180{ 1181 unsigned long flags; 1182 struct page *page; 1183 int cold = !!(gfp_flags & __GFP_COLD); 1184 int cpu; 1185 1186again: 1187 cpu = get_cpu(); 1188 if (likely(order == 0)) { 1189 struct per_cpu_pages *pcp; 1190 struct list_head *list; 1191 1192 pcp = &zone_pcp(zone, cpu)->pcp; 1193 list = &pcp->lists[migratetype]; 1194 local_irq_save(flags); 1195 if (list_empty(list)) { 1196 pcp->count += rmqueue_bulk(zone, 0, 1197 pcp->batch, list, 1198 migratetype, cold); 1199 if (unlikely(list_empty(list))) 1200 goto failed; 1201 } 1202 1203 if (cold) 1204 page = list_entry(list->prev, struct page, lru); 1205 else 1206 page = list_entry(list->next, struct page, lru); 1207 1208 list_del(&page->lru); 1209 pcp->count--; 1210 } else { 1211 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1212 /* 1213 * __GFP_NOFAIL is not to be used in new code. 1214 * 1215 * All __GFP_NOFAIL callers should be fixed so that they 1216 * properly detect and handle allocation failures. 1217 * 1218 * We most definitely don't want callers attempting to 1219 * allocate greater than order-1 page units with 1220 * __GFP_NOFAIL. 1221 */ 1222 WARN_ON_ONCE(order > 1); 1223 } 1224 spin_lock_irqsave(&zone->lock, flags); 1225 page = __rmqueue(zone, order, migratetype); 1226 spin_unlock(&zone->lock); 1227 if (!page) 1228 goto failed; 1229 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1230 } 1231 1232 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1233 zone_statistics(preferred_zone, zone); 1234 local_irq_restore(flags); 1235 put_cpu(); 1236 1237 VM_BUG_ON(bad_range(zone, page)); 1238 if (prep_new_page(page, order, gfp_flags)) 1239 goto again; 1240 return page; 1241 1242failed: 1243 local_irq_restore(flags); 1244 put_cpu(); 1245 return NULL; 1246} 1247 1248/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1249#define ALLOC_WMARK_MIN WMARK_MIN 1250#define ALLOC_WMARK_LOW WMARK_LOW 1251#define ALLOC_WMARK_HIGH WMARK_HIGH 1252#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1253 1254/* Mask to get the watermark bits */ 1255#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1256 1257#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1258#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1259#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1260 1261#ifdef CONFIG_FAIL_PAGE_ALLOC 1262 1263static struct fail_page_alloc_attr { 1264 struct fault_attr attr; 1265 1266 u32 ignore_gfp_highmem; 1267 u32 ignore_gfp_wait; 1268 u32 min_order; 1269 1270#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1271 1272 struct dentry *ignore_gfp_highmem_file; 1273 struct dentry *ignore_gfp_wait_file; 1274 struct dentry *min_order_file; 1275 1276#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1277 1278} fail_page_alloc = { 1279 .attr = FAULT_ATTR_INITIALIZER, 1280 .ignore_gfp_wait = 1, 1281 .ignore_gfp_highmem = 1, 1282 .min_order = 1, 1283}; 1284 1285static int __init setup_fail_page_alloc(char *str) 1286{ 1287 return setup_fault_attr(&fail_page_alloc.attr, str); 1288} 1289__setup("fail_page_alloc=", setup_fail_page_alloc); 1290 1291static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1292{ 1293 if (order < fail_page_alloc.min_order) 1294 return 0; 1295 if (gfp_mask & __GFP_NOFAIL) 1296 return 0; 1297 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1298 return 0; 1299 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1300 return 0; 1301 1302 return should_fail(&fail_page_alloc.attr, 1 << order); 1303} 1304 1305#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1306 1307static int __init fail_page_alloc_debugfs(void) 1308{ 1309 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1310 struct dentry *dir; 1311 int err; 1312 1313 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1314 "fail_page_alloc"); 1315 if (err) 1316 return err; 1317 dir = fail_page_alloc.attr.dentries.dir; 1318 1319 fail_page_alloc.ignore_gfp_wait_file = 1320 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1321 &fail_page_alloc.ignore_gfp_wait); 1322 1323 fail_page_alloc.ignore_gfp_highmem_file = 1324 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1325 &fail_page_alloc.ignore_gfp_highmem); 1326 fail_page_alloc.min_order_file = 1327 debugfs_create_u32("min-order", mode, dir, 1328 &fail_page_alloc.min_order); 1329 1330 if (!fail_page_alloc.ignore_gfp_wait_file || 1331 !fail_page_alloc.ignore_gfp_highmem_file || 1332 !fail_page_alloc.min_order_file) { 1333 err = -ENOMEM; 1334 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1335 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1336 debugfs_remove(fail_page_alloc.min_order_file); 1337 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1338 } 1339 1340 return err; 1341} 1342 1343late_initcall(fail_page_alloc_debugfs); 1344 1345#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1346 1347#else /* CONFIG_FAIL_PAGE_ALLOC */ 1348 1349static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1350{ 1351 return 0; 1352} 1353 1354#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1355 1356/* 1357 * Return 1 if free pages are above 'mark'. This takes into account the order 1358 * of the allocation. 1359 */ 1360int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1361 int classzone_idx, int alloc_flags) 1362{ 1363 /* free_pages my go negative - that's OK */ 1364 long min = mark; 1365 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1366 int o; 1367 1368 if (alloc_flags & ALLOC_HIGH) 1369 min -= min / 2; 1370 if (alloc_flags & ALLOC_HARDER) 1371 min -= min / 4; 1372 1373 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1374 return 0; 1375 for (o = 0; o < order; o++) { 1376 /* At the next order, this order's pages become unavailable */ 1377 free_pages -= z->free_area[o].nr_free << o; 1378 1379 /* Require fewer higher order pages to be free */ 1380 min >>= 1; 1381 1382 if (free_pages <= min) 1383 return 0; 1384 } 1385 return 1; 1386} 1387 1388#ifdef CONFIG_NUMA 1389/* 1390 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1391 * skip over zones that are not allowed by the cpuset, or that have 1392 * been recently (in last second) found to be nearly full. See further 1393 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1394 * that have to skip over a lot of full or unallowed zones. 1395 * 1396 * If the zonelist cache is present in the passed in zonelist, then 1397 * returns a pointer to the allowed node mask (either the current 1398 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1399 * 1400 * If the zonelist cache is not available for this zonelist, does 1401 * nothing and returns NULL. 1402 * 1403 * If the fullzones BITMAP in the zonelist cache is stale (more than 1404 * a second since last zap'd) then we zap it out (clear its bits.) 1405 * 1406 * We hold off even calling zlc_setup, until after we've checked the 1407 * first zone in the zonelist, on the theory that most allocations will 1408 * be satisfied from that first zone, so best to examine that zone as 1409 * quickly as we can. 1410 */ 1411static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1412{ 1413 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1414 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1415 1416 zlc = zonelist->zlcache_ptr; 1417 if (!zlc) 1418 return NULL; 1419 1420 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1421 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1422 zlc->last_full_zap = jiffies; 1423 } 1424 1425 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1426 &cpuset_current_mems_allowed : 1427 &node_states[N_HIGH_MEMORY]; 1428 return allowednodes; 1429} 1430 1431/* 1432 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1433 * if it is worth looking at further for free memory: 1434 * 1) Check that the zone isn't thought to be full (doesn't have its 1435 * bit set in the zonelist_cache fullzones BITMAP). 1436 * 2) Check that the zones node (obtained from the zonelist_cache 1437 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1438 * Return true (non-zero) if zone is worth looking at further, or 1439 * else return false (zero) if it is not. 1440 * 1441 * This check -ignores- the distinction between various watermarks, 1442 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1443 * found to be full for any variation of these watermarks, it will 1444 * be considered full for up to one second by all requests, unless 1445 * we are so low on memory on all allowed nodes that we are forced 1446 * into the second scan of the zonelist. 1447 * 1448 * In the second scan we ignore this zonelist cache and exactly 1449 * apply the watermarks to all zones, even it is slower to do so. 1450 * We are low on memory in the second scan, and should leave no stone 1451 * unturned looking for a free page. 1452 */ 1453static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1454 nodemask_t *allowednodes) 1455{ 1456 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1457 int i; /* index of *z in zonelist zones */ 1458 int n; /* node that zone *z is on */ 1459 1460 zlc = zonelist->zlcache_ptr; 1461 if (!zlc) 1462 return 1; 1463 1464 i = z - zonelist->_zonerefs; 1465 n = zlc->z_to_n[i]; 1466 1467 /* This zone is worth trying if it is allowed but not full */ 1468 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1469} 1470 1471/* 1472 * Given 'z' scanning a zonelist, set the corresponding bit in 1473 * zlc->fullzones, so that subsequent attempts to allocate a page 1474 * from that zone don't waste time re-examining it. 1475 */ 1476static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1477{ 1478 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1479 int i; /* index of *z in zonelist zones */ 1480 1481 zlc = zonelist->zlcache_ptr; 1482 if (!zlc) 1483 return; 1484 1485 i = z - zonelist->_zonerefs; 1486 1487 set_bit(i, zlc->fullzones); 1488} 1489 1490#else /* CONFIG_NUMA */ 1491 1492static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1493{ 1494 return NULL; 1495} 1496 1497static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1498 nodemask_t *allowednodes) 1499{ 1500 return 1; 1501} 1502 1503static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1504{ 1505} 1506#endif /* CONFIG_NUMA */ 1507 1508/* 1509 * get_page_from_freelist goes through the zonelist trying to allocate 1510 * a page. 1511 */ 1512static struct page * 1513get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1514 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1515 struct zone *preferred_zone, int migratetype) 1516{ 1517 struct zoneref *z; 1518 struct page *page = NULL; 1519 int classzone_idx; 1520 struct zone *zone; 1521 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1522 int zlc_active = 0; /* set if using zonelist_cache */ 1523 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1524 1525 classzone_idx = zone_idx(preferred_zone); 1526zonelist_scan: 1527 /* 1528 * Scan zonelist, looking for a zone with enough free. 1529 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1530 */ 1531 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1532 high_zoneidx, nodemask) { 1533 if (NUMA_BUILD && zlc_active && 1534 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1535 continue; 1536 if ((alloc_flags & ALLOC_CPUSET) && 1537 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1538 goto try_next_zone; 1539 1540 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1541 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1542 unsigned long mark; 1543 int ret; 1544 1545 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1546 if (zone_watermark_ok(zone, order, mark, 1547 classzone_idx, alloc_flags)) 1548 goto try_this_zone; 1549 1550 if (zone_reclaim_mode == 0) 1551 goto this_zone_full; 1552 1553 ret = zone_reclaim(zone, gfp_mask, order); 1554 switch (ret) { 1555 case ZONE_RECLAIM_NOSCAN: 1556 /* did not scan */ 1557 goto try_next_zone; 1558 case ZONE_RECLAIM_FULL: 1559 /* scanned but unreclaimable */ 1560 goto this_zone_full; 1561 default: 1562 /* did we reclaim enough */ 1563 if (!zone_watermark_ok(zone, order, mark, 1564 classzone_idx, alloc_flags)) 1565 goto this_zone_full; 1566 } 1567 } 1568 1569try_this_zone: 1570 page = buffered_rmqueue(preferred_zone, zone, order, 1571 gfp_mask, migratetype); 1572 if (page) 1573 break; 1574this_zone_full: 1575 if (NUMA_BUILD) 1576 zlc_mark_zone_full(zonelist, z); 1577try_next_zone: 1578 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1579 /* 1580 * we do zlc_setup after the first zone is tried but only 1581 * if there are multiple nodes make it worthwhile 1582 */ 1583 allowednodes = zlc_setup(zonelist, alloc_flags); 1584 zlc_active = 1; 1585 did_zlc_setup = 1; 1586 } 1587 } 1588 1589 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1590 /* Disable zlc cache for second zonelist scan */ 1591 zlc_active = 0; 1592 goto zonelist_scan; 1593 } 1594 return page; 1595} 1596 1597static inline int 1598should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1599 unsigned long pages_reclaimed) 1600{ 1601 /* Do not loop if specifically requested */ 1602 if (gfp_mask & __GFP_NORETRY) 1603 return 0; 1604 1605 /* 1606 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1607 * means __GFP_NOFAIL, but that may not be true in other 1608 * implementations. 1609 */ 1610 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1611 return 1; 1612 1613 /* 1614 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1615 * specified, then we retry until we no longer reclaim any pages 1616 * (above), or we've reclaimed an order of pages at least as 1617 * large as the allocation's order. In both cases, if the 1618 * allocation still fails, we stop retrying. 1619 */ 1620 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1621 return 1; 1622 1623 /* 1624 * Don't let big-order allocations loop unless the caller 1625 * explicitly requests that. 1626 */ 1627 if (gfp_mask & __GFP_NOFAIL) 1628 return 1; 1629 1630 return 0; 1631} 1632 1633static inline struct page * 1634__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1635 struct zonelist *zonelist, enum zone_type high_zoneidx, 1636 nodemask_t *nodemask, struct zone *preferred_zone, 1637 int migratetype) 1638{ 1639 struct page *page; 1640 1641 /* Acquire the OOM killer lock for the zones in zonelist */ 1642 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1643 schedule_timeout_uninterruptible(1); 1644 return NULL; 1645 } 1646 1647 /* 1648 * Go through the zonelist yet one more time, keep very high watermark 1649 * here, this is only to catch a parallel oom killing, we must fail if 1650 * we're still under heavy pressure. 1651 */ 1652 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1653 order, zonelist, high_zoneidx, 1654 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1655 preferred_zone, migratetype); 1656 if (page) 1657 goto out; 1658 1659 if (!(gfp_mask & __GFP_NOFAIL)) { 1660 /* The OOM killer will not help higher order allocs */ 1661 if (order > PAGE_ALLOC_COSTLY_ORDER) 1662 goto out; 1663 /* 1664 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1665 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1666 * The caller should handle page allocation failure by itself if 1667 * it specifies __GFP_THISNODE. 1668 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1669 */ 1670 if (gfp_mask & __GFP_THISNODE) 1671 goto out; 1672 } 1673 /* Exhausted what can be done so it's blamo time */ 1674 out_of_memory(zonelist, gfp_mask, order, nodemask); 1675 1676out: 1677 clear_zonelist_oom(zonelist, gfp_mask); 1678 return page; 1679} 1680 1681/* The really slow allocator path where we enter direct reclaim */ 1682static inline struct page * 1683__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1684 struct zonelist *zonelist, enum zone_type high_zoneidx, 1685 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1686 int migratetype, unsigned long *did_some_progress) 1687{ 1688 struct page *page = NULL; 1689 struct reclaim_state reclaim_state; 1690 struct task_struct *p = current; 1691 1692 cond_resched(); 1693 1694 /* We now go into synchronous reclaim */ 1695 cpuset_memory_pressure_bump(); 1696 p->flags |= PF_MEMALLOC; 1697 lockdep_set_current_reclaim_state(gfp_mask); 1698 reclaim_state.reclaimed_slab = 0; 1699 p->reclaim_state = &reclaim_state; 1700 1701 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1702 1703 p->reclaim_state = NULL; 1704 lockdep_clear_current_reclaim_state(); 1705 p->flags &= ~PF_MEMALLOC; 1706 1707 cond_resched(); 1708 1709 if (order != 0) 1710 drain_all_pages(); 1711 1712 if (likely(*did_some_progress)) 1713 page = get_page_from_freelist(gfp_mask, nodemask, order, 1714 zonelist, high_zoneidx, 1715 alloc_flags, preferred_zone, 1716 migratetype); 1717 return page; 1718} 1719 1720/* 1721 * This is called in the allocator slow-path if the allocation request is of 1722 * sufficient urgency to ignore watermarks and take other desperate measures 1723 */ 1724static inline struct page * 1725__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1726 struct zonelist *zonelist, enum zone_type high_zoneidx, 1727 nodemask_t *nodemask, struct zone *preferred_zone, 1728 int migratetype) 1729{ 1730 struct page *page; 1731 1732 do { 1733 page = get_page_from_freelist(gfp_mask, nodemask, order, 1734 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1735 preferred_zone, migratetype); 1736 1737 if (!page && gfp_mask & __GFP_NOFAIL) 1738 congestion_wait(BLK_RW_ASYNC, HZ/50); 1739 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1740 1741 return page; 1742} 1743 1744static inline 1745void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1746 enum zone_type high_zoneidx) 1747{ 1748 struct zoneref *z; 1749 struct zone *zone; 1750 1751 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1752 wakeup_kswapd(zone, order); 1753} 1754 1755static inline int 1756gfp_to_alloc_flags(gfp_t gfp_mask) 1757{ 1758 struct task_struct *p = current; 1759 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1760 const gfp_t wait = gfp_mask & __GFP_WAIT; 1761 1762 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1763 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1764 1765 /* 1766 * The caller may dip into page reserves a bit more if the caller 1767 * cannot run direct reclaim, or if the caller has realtime scheduling 1768 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1769 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1770 */ 1771 alloc_flags |= (gfp_mask & __GFP_HIGH); 1772 1773 if (!wait) { 1774 alloc_flags |= ALLOC_HARDER; 1775 /* 1776 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1777 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1778 */ 1779 alloc_flags &= ~ALLOC_CPUSET; 1780 } else if (unlikely(rt_task(p)) && !in_interrupt()) 1781 alloc_flags |= ALLOC_HARDER; 1782 1783 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1784 if (!in_interrupt() && 1785 ((p->flags & PF_MEMALLOC) || 1786 unlikely(test_thread_flag(TIF_MEMDIE)))) 1787 alloc_flags |= ALLOC_NO_WATERMARKS; 1788 } 1789 1790 return alloc_flags; 1791} 1792 1793static inline struct page * 1794__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1795 struct zonelist *zonelist, enum zone_type high_zoneidx, 1796 nodemask_t *nodemask, struct zone *preferred_zone, 1797 int migratetype) 1798{ 1799 const gfp_t wait = gfp_mask & __GFP_WAIT; 1800 struct page *page = NULL; 1801 int alloc_flags; 1802 unsigned long pages_reclaimed = 0; 1803 unsigned long did_some_progress; 1804 struct task_struct *p = current; 1805 1806 /* 1807 * In the slowpath, we sanity check order to avoid ever trying to 1808 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1809 * be using allocators in order of preference for an area that is 1810 * too large. 1811 */ 1812 if (order >= MAX_ORDER) { 1813 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1814 return NULL; 1815 } 1816 1817 /* 1818 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1819 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1820 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1821 * using a larger set of nodes after it has established that the 1822 * allowed per node queues are empty and that nodes are 1823 * over allocated. 1824 */ 1825 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1826 goto nopage; 1827 1828restart: 1829 wake_all_kswapd(order, zonelist, high_zoneidx); 1830 1831 /* 1832 * OK, we're below the kswapd watermark and have kicked background 1833 * reclaim. Now things get more complex, so set up alloc_flags according 1834 * to how we want to proceed. 1835 */ 1836 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1837 1838 /* This is the last chance, in general, before the goto nopage. */ 1839 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1840 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1841 preferred_zone, migratetype); 1842 if (page) 1843 goto got_pg; 1844 1845rebalance: 1846 /* Allocate without watermarks if the context allows */ 1847 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1848 page = __alloc_pages_high_priority(gfp_mask, order, 1849 zonelist, high_zoneidx, nodemask, 1850 preferred_zone, migratetype); 1851 if (page) 1852 goto got_pg; 1853 } 1854 1855 /* Atomic allocations - we can't balance anything */ 1856 if (!wait) 1857 goto nopage; 1858 1859 /* Avoid recursion of direct reclaim */ 1860 if (p->flags & PF_MEMALLOC) 1861 goto nopage; 1862 1863 /* Avoid allocations with no watermarks from looping endlessly */ 1864 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 1865 goto nopage; 1866 1867 /* Try direct reclaim and then allocating */ 1868 page = __alloc_pages_direct_reclaim(gfp_mask, order, 1869 zonelist, high_zoneidx, 1870 nodemask, 1871 alloc_flags, preferred_zone, 1872 migratetype, &did_some_progress); 1873 if (page) 1874 goto got_pg; 1875 1876 /* 1877 * If we failed to make any progress reclaiming, then we are 1878 * running out of options and have to consider going OOM 1879 */ 1880 if (!did_some_progress) { 1881 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1882 if (oom_killer_disabled) 1883 goto nopage; 1884 page = __alloc_pages_may_oom(gfp_mask, order, 1885 zonelist, high_zoneidx, 1886 nodemask, preferred_zone, 1887 migratetype); 1888 if (page) 1889 goto got_pg; 1890 1891 /* 1892 * The OOM killer does not trigger for high-order 1893 * ~__GFP_NOFAIL allocations so if no progress is being 1894 * made, there are no other options and retrying is 1895 * unlikely to help. 1896 */ 1897 if (order > PAGE_ALLOC_COSTLY_ORDER && 1898 !(gfp_mask & __GFP_NOFAIL)) 1899 goto nopage; 1900 1901 goto restart; 1902 } 1903 } 1904 1905 /* Check if we should retry the allocation */ 1906 pages_reclaimed += did_some_progress; 1907 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 1908 /* Wait for some write requests to complete then retry */ 1909 congestion_wait(BLK_RW_ASYNC, HZ/50); 1910 goto rebalance; 1911 } 1912 1913nopage: 1914 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1915 printk(KERN_WARNING "%s: page allocation failure." 1916 " order:%d, mode:0x%x\n", 1917 p->comm, order, gfp_mask); 1918 dump_stack(); 1919 show_mem(); 1920 } 1921 return page; 1922got_pg: 1923 if (kmemcheck_enabled) 1924 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 1925 return page; 1926 1927} 1928 1929/* 1930 * This is the 'heart' of the zoned buddy allocator. 1931 */ 1932struct page * 1933__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 1934 struct zonelist *zonelist, nodemask_t *nodemask) 1935{ 1936 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 1937 struct zone *preferred_zone; 1938 struct page *page; 1939 int migratetype = allocflags_to_migratetype(gfp_mask); 1940 1941 gfp_mask &= gfp_allowed_mask; 1942 1943 lockdep_trace_alloc(gfp_mask); 1944 1945 might_sleep_if(gfp_mask & __GFP_WAIT); 1946 1947 if (should_fail_alloc_page(gfp_mask, order)) 1948 return NULL; 1949 1950 /* 1951 * Check the zones suitable for the gfp_mask contain at least one 1952 * valid zone. It's possible to have an empty zonelist as a result 1953 * of GFP_THISNODE and a memoryless node 1954 */ 1955 if (unlikely(!zonelist->_zonerefs->zone)) 1956 return NULL; 1957 1958 /* The preferred zone is used for statistics later */ 1959 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 1960 if (!preferred_zone) 1961 return NULL; 1962 1963 /* First allocation attempt */ 1964 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 1965 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 1966 preferred_zone, migratetype); 1967 if (unlikely(!page)) 1968 page = __alloc_pages_slowpath(gfp_mask, order, 1969 zonelist, high_zoneidx, nodemask, 1970 preferred_zone, migratetype); 1971 1972 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 1973 return page; 1974} 1975EXPORT_SYMBOL(__alloc_pages_nodemask); 1976 1977/* 1978 * Common helper functions. 1979 */ 1980unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1981{ 1982 struct page *page; 1983 1984 /* 1985 * __get_free_pages() returns a 32-bit address, which cannot represent 1986 * a highmem page 1987 */ 1988 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1989 1990 page = alloc_pages(gfp_mask, order); 1991 if (!page) 1992 return 0; 1993 return (unsigned long) page_address(page); 1994} 1995EXPORT_SYMBOL(__get_free_pages); 1996 1997unsigned long get_zeroed_page(gfp_t gfp_mask) 1998{ 1999 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2000} 2001EXPORT_SYMBOL(get_zeroed_page); 2002 2003void __pagevec_free(struct pagevec *pvec) 2004{ 2005 int i = pagevec_count(pvec); 2006 2007 while (--i >= 0) { 2008 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2009 free_hot_cold_page(pvec->pages[i], pvec->cold); 2010 } 2011} 2012 2013void __free_pages(struct page *page, unsigned int order) 2014{ 2015 if (put_page_testzero(page)) { 2016 trace_mm_page_free_direct(page, order); 2017 if (order == 0) 2018 free_hot_page(page); 2019 else 2020 __free_pages_ok(page, order); 2021 } 2022} 2023 2024EXPORT_SYMBOL(__free_pages); 2025 2026void free_pages(unsigned long addr, unsigned int order) 2027{ 2028 if (addr != 0) { 2029 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2030 __free_pages(virt_to_page((void *)addr), order); 2031 } 2032} 2033 2034EXPORT_SYMBOL(free_pages); 2035 2036/** 2037 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2038 * @size: the number of bytes to allocate 2039 * @gfp_mask: GFP flags for the allocation 2040 * 2041 * This function is similar to alloc_pages(), except that it allocates the 2042 * minimum number of pages to satisfy the request. alloc_pages() can only 2043 * allocate memory in power-of-two pages. 2044 * 2045 * This function is also limited by MAX_ORDER. 2046 * 2047 * Memory allocated by this function must be released by free_pages_exact(). 2048 */ 2049void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2050{ 2051 unsigned int order = get_order(size); 2052 unsigned long addr; 2053 2054 addr = __get_free_pages(gfp_mask, order); 2055 if (addr) { 2056 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2057 unsigned long used = addr + PAGE_ALIGN(size); 2058 2059 split_page(virt_to_page((void *)addr), order); 2060 while (used < alloc_end) { 2061 free_page(used); 2062 used += PAGE_SIZE; 2063 } 2064 } 2065 2066 return (void *)addr; 2067} 2068EXPORT_SYMBOL(alloc_pages_exact); 2069 2070/** 2071 * free_pages_exact - release memory allocated via alloc_pages_exact() 2072 * @virt: the value returned by alloc_pages_exact. 2073 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2074 * 2075 * Release the memory allocated by a previous call to alloc_pages_exact. 2076 */ 2077void free_pages_exact(void *virt, size_t size) 2078{ 2079 unsigned long addr = (unsigned long)virt; 2080 unsigned long end = addr + PAGE_ALIGN(size); 2081 2082 while (addr < end) { 2083 free_page(addr); 2084 addr += PAGE_SIZE; 2085 } 2086} 2087EXPORT_SYMBOL(free_pages_exact); 2088 2089static unsigned int nr_free_zone_pages(int offset) 2090{ 2091 struct zoneref *z; 2092 struct zone *zone; 2093 2094 /* Just pick one node, since fallback list is circular */ 2095 unsigned int sum = 0; 2096 2097 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2098 2099 for_each_zone_zonelist(zone, z, zonelist, offset) { 2100 unsigned long size = zone->present_pages; 2101 unsigned long high = high_wmark_pages(zone); 2102 if (size > high) 2103 sum += size - high; 2104 } 2105 2106 return sum; 2107} 2108 2109/* 2110 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2111 */ 2112unsigned int nr_free_buffer_pages(void) 2113{ 2114 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2115} 2116EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2117 2118/* 2119 * Amount of free RAM allocatable within all zones 2120 */ 2121unsigned int nr_free_pagecache_pages(void) 2122{ 2123 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2124} 2125 2126static inline void show_node(struct zone *zone) 2127{ 2128 if (NUMA_BUILD) 2129 printk("Node %d ", zone_to_nid(zone)); 2130} 2131 2132void si_meminfo(struct sysinfo *val) 2133{ 2134 val->totalram = totalram_pages; 2135 val->sharedram = 0; 2136 val->freeram = global_page_state(NR_FREE_PAGES); 2137 val->bufferram = nr_blockdev_pages(); 2138 val->totalhigh = totalhigh_pages; 2139 val->freehigh = nr_free_highpages(); 2140 val->mem_unit = PAGE_SIZE; 2141} 2142 2143EXPORT_SYMBOL(si_meminfo); 2144 2145#ifdef CONFIG_NUMA 2146void si_meminfo_node(struct sysinfo *val, int nid) 2147{ 2148 pg_data_t *pgdat = NODE_DATA(nid); 2149 2150 val->totalram = pgdat->node_present_pages; 2151 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2152#ifdef CONFIG_HIGHMEM 2153 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2154 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2155 NR_FREE_PAGES); 2156#else 2157 val->totalhigh = 0; 2158 val->freehigh = 0; 2159#endif 2160 val->mem_unit = PAGE_SIZE; 2161} 2162#endif 2163 2164#define K(x) ((x) << (PAGE_SHIFT-10)) 2165 2166/* 2167 * Show free area list (used inside shift_scroll-lock stuff) 2168 * We also calculate the percentage fragmentation. We do this by counting the 2169 * memory on each free list with the exception of the first item on the list. 2170 */ 2171void show_free_areas(void) 2172{ 2173 int cpu; 2174 struct zone *zone; 2175 2176 for_each_populated_zone(zone) { 2177 show_node(zone); 2178 printk("%s per-cpu:\n", zone->name); 2179 2180 for_each_online_cpu(cpu) { 2181 struct per_cpu_pageset *pageset; 2182 2183 pageset = zone_pcp(zone, cpu); 2184 2185 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2186 cpu, pageset->pcp.high, 2187 pageset->pcp.batch, pageset->pcp.count); 2188 } 2189 } 2190 2191 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2192 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2193 " unevictable:%lu" 2194 " dirty:%lu writeback:%lu unstable:%lu\n" 2195 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2196 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2197 global_page_state(NR_ACTIVE_ANON), 2198 global_page_state(NR_INACTIVE_ANON), 2199 global_page_state(NR_ISOLATED_ANON), 2200 global_page_state(NR_ACTIVE_FILE), 2201 global_page_state(NR_INACTIVE_FILE), 2202 global_page_state(NR_ISOLATED_FILE), 2203 global_page_state(NR_UNEVICTABLE), 2204 global_page_state(NR_FILE_DIRTY), 2205 global_page_state(NR_WRITEBACK), 2206 global_page_state(NR_UNSTABLE_NFS), 2207 global_page_state(NR_FREE_PAGES), 2208 global_page_state(NR_SLAB_RECLAIMABLE), 2209 global_page_state(NR_SLAB_UNRECLAIMABLE), 2210 global_page_state(NR_FILE_MAPPED), 2211 global_page_state(NR_SHMEM), 2212 global_page_state(NR_PAGETABLE), 2213 global_page_state(NR_BOUNCE)); 2214 2215 for_each_populated_zone(zone) { 2216 int i; 2217 2218 show_node(zone); 2219 printk("%s" 2220 " free:%lukB" 2221 " min:%lukB" 2222 " low:%lukB" 2223 " high:%lukB" 2224 " active_anon:%lukB" 2225 " inactive_anon:%lukB" 2226 " active_file:%lukB" 2227 " inactive_file:%lukB" 2228 " unevictable:%lukB" 2229 " isolated(anon):%lukB" 2230 " isolated(file):%lukB" 2231 " present:%lukB" 2232 " mlocked:%lukB" 2233 " dirty:%lukB" 2234 " writeback:%lukB" 2235 " mapped:%lukB" 2236 " shmem:%lukB" 2237 " slab_reclaimable:%lukB" 2238 " slab_unreclaimable:%lukB" 2239 " kernel_stack:%lukB" 2240 " pagetables:%lukB" 2241 " unstable:%lukB" 2242 " bounce:%lukB" 2243 " writeback_tmp:%lukB" 2244 " pages_scanned:%lu" 2245 " all_unreclaimable? %s" 2246 "\n", 2247 zone->name, 2248 K(zone_page_state(zone, NR_FREE_PAGES)), 2249 K(min_wmark_pages(zone)), 2250 K(low_wmark_pages(zone)), 2251 K(high_wmark_pages(zone)), 2252 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2253 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2254 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2255 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2256 K(zone_page_state(zone, NR_UNEVICTABLE)), 2257 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2258 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2259 K(zone->present_pages), 2260 K(zone_page_state(zone, NR_MLOCK)), 2261 K(zone_page_state(zone, NR_FILE_DIRTY)), 2262 K(zone_page_state(zone, NR_WRITEBACK)), 2263 K(zone_page_state(zone, NR_FILE_MAPPED)), 2264 K(zone_page_state(zone, NR_SHMEM)), 2265 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2266 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2267 zone_page_state(zone, NR_KERNEL_STACK) * 2268 THREAD_SIZE / 1024, 2269 K(zone_page_state(zone, NR_PAGETABLE)), 2270 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2271 K(zone_page_state(zone, NR_BOUNCE)), 2272 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2273 zone->pages_scanned, 2274 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 2275 ); 2276 printk("lowmem_reserve[]:"); 2277 for (i = 0; i < MAX_NR_ZONES; i++) 2278 printk(" %lu", zone->lowmem_reserve[i]); 2279 printk("\n"); 2280 } 2281 2282 for_each_populated_zone(zone) { 2283 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2284 2285 show_node(zone); 2286 printk("%s: ", zone->name); 2287 2288 spin_lock_irqsave(&zone->lock, flags); 2289 for (order = 0; order < MAX_ORDER; order++) { 2290 nr[order] = zone->free_area[order].nr_free; 2291 total += nr[order] << order; 2292 } 2293 spin_unlock_irqrestore(&zone->lock, flags); 2294 for (order = 0; order < MAX_ORDER; order++) 2295 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2296 printk("= %lukB\n", K(total)); 2297 } 2298 2299 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2300 2301 show_swap_cache_info(); 2302} 2303 2304static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2305{ 2306 zoneref->zone = zone; 2307 zoneref->zone_idx = zone_idx(zone); 2308} 2309 2310/* 2311 * Builds allocation fallback zone lists. 2312 * 2313 * Add all populated zones of a node to the zonelist. 2314 */ 2315static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2316 int nr_zones, enum zone_type zone_type) 2317{ 2318 struct zone *zone; 2319 2320 BUG_ON(zone_type >= MAX_NR_ZONES); 2321 zone_type++; 2322 2323 do { 2324 zone_type--; 2325 zone = pgdat->node_zones + zone_type; 2326 if (populated_zone(zone)) { 2327 zoneref_set_zone(zone, 2328 &zonelist->_zonerefs[nr_zones++]); 2329 check_highest_zone(zone_type); 2330 } 2331 2332 } while (zone_type); 2333 return nr_zones; 2334} 2335 2336 2337/* 2338 * zonelist_order: 2339 * 0 = automatic detection of better ordering. 2340 * 1 = order by ([node] distance, -zonetype) 2341 * 2 = order by (-zonetype, [node] distance) 2342 * 2343 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2344 * the same zonelist. So only NUMA can configure this param. 2345 */ 2346#define ZONELIST_ORDER_DEFAULT 0 2347#define ZONELIST_ORDER_NODE 1 2348#define ZONELIST_ORDER_ZONE 2 2349 2350/* zonelist order in the kernel. 2351 * set_zonelist_order() will set this to NODE or ZONE. 2352 */ 2353static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2354static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2355 2356 2357#ifdef CONFIG_NUMA 2358/* The value user specified ....changed by config */ 2359static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2360/* string for sysctl */ 2361#define NUMA_ZONELIST_ORDER_LEN 16 2362char numa_zonelist_order[16] = "default"; 2363 2364/* 2365 * interface for configure zonelist ordering. 2366 * command line option "numa_zonelist_order" 2367 * = "[dD]efault - default, automatic configuration. 2368 * = "[nN]ode - order by node locality, then by zone within node 2369 * = "[zZ]one - order by zone, then by locality within zone 2370 */ 2371 2372static int __parse_numa_zonelist_order(char *s) 2373{ 2374 if (*s == 'd' || *s == 'D') { 2375 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2376 } else if (*s == 'n' || *s == 'N') { 2377 user_zonelist_order = ZONELIST_ORDER_NODE; 2378 } else if (*s == 'z' || *s == 'Z') { 2379 user_zonelist_order = ZONELIST_ORDER_ZONE; 2380 } else { 2381 printk(KERN_WARNING 2382 "Ignoring invalid numa_zonelist_order value: " 2383 "%s\n", s); 2384 return -EINVAL; 2385 } 2386 return 0; 2387} 2388 2389static __init int setup_numa_zonelist_order(char *s) 2390{ 2391 if (s) 2392 return __parse_numa_zonelist_order(s); 2393 return 0; 2394} 2395early_param("numa_zonelist_order", setup_numa_zonelist_order); 2396 2397/* 2398 * sysctl handler for numa_zonelist_order 2399 */ 2400int numa_zonelist_order_handler(ctl_table *table, int write, 2401 void __user *buffer, size_t *length, 2402 loff_t *ppos) 2403{ 2404 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2405 int ret; 2406 static DEFINE_MUTEX(zl_order_mutex); 2407 2408 mutex_lock(&zl_order_mutex); 2409 if (write) 2410 strcpy(saved_string, (char*)table->data); 2411 ret = proc_dostring(table, write, buffer, length, ppos); 2412 if (ret) 2413 goto out; 2414 if (write) { 2415 int oldval = user_zonelist_order; 2416 if (__parse_numa_zonelist_order((char*)table->data)) { 2417 /* 2418 * bogus value. restore saved string 2419 */ 2420 strncpy((char*)table->data, saved_string, 2421 NUMA_ZONELIST_ORDER_LEN); 2422 user_zonelist_order = oldval; 2423 } else if (oldval != user_zonelist_order) 2424 build_all_zonelists(); 2425 } 2426out: 2427 mutex_unlock(&zl_order_mutex); 2428 return ret; 2429} 2430 2431 2432#define MAX_NODE_LOAD (nr_online_nodes) 2433static int node_load[MAX_NUMNODES]; 2434 2435/** 2436 * find_next_best_node - find the next node that should appear in a given node's fallback list 2437 * @node: node whose fallback list we're appending 2438 * @used_node_mask: nodemask_t of already used nodes 2439 * 2440 * We use a number of factors to determine which is the next node that should 2441 * appear on a given node's fallback list. The node should not have appeared 2442 * already in @node's fallback list, and it should be the next closest node 2443 * according to the distance array (which contains arbitrary distance values 2444 * from each node to each node in the system), and should also prefer nodes 2445 * with no CPUs, since presumably they'll have very little allocation pressure 2446 * on them otherwise. 2447 * It returns -1 if no node is found. 2448 */ 2449static int find_next_best_node(int node, nodemask_t *used_node_mask) 2450{ 2451 int n, val; 2452 int min_val = INT_MAX; 2453 int best_node = -1; 2454 const struct cpumask *tmp = cpumask_of_node(0); 2455 2456 /* Use the local node if we haven't already */ 2457 if (!node_isset(node, *used_node_mask)) { 2458 node_set(node, *used_node_mask); 2459 return node; 2460 } 2461 2462 for_each_node_state(n, N_HIGH_MEMORY) { 2463 2464 /* Don't want a node to appear more than once */ 2465 if (node_isset(n, *used_node_mask)) 2466 continue; 2467 2468 /* Use the distance array to find the distance */ 2469 val = node_distance(node, n); 2470 2471 /* Penalize nodes under us ("prefer the next node") */ 2472 val += (n < node); 2473 2474 /* Give preference to headless and unused nodes */ 2475 tmp = cpumask_of_node(n); 2476 if (!cpumask_empty(tmp)) 2477 val += PENALTY_FOR_NODE_WITH_CPUS; 2478 2479 /* Slight preference for less loaded node */ 2480 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2481 val += node_load[n]; 2482 2483 if (val < min_val) { 2484 min_val = val; 2485 best_node = n; 2486 } 2487 } 2488 2489 if (best_node >= 0) 2490 node_set(best_node, *used_node_mask); 2491 2492 return best_node; 2493} 2494 2495 2496/* 2497 * Build zonelists ordered by node and zones within node. 2498 * This results in maximum locality--normal zone overflows into local 2499 * DMA zone, if any--but risks exhausting DMA zone. 2500 */ 2501static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2502{ 2503 int j; 2504 struct zonelist *zonelist; 2505 2506 zonelist = &pgdat->node_zonelists[0]; 2507 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2508 ; 2509 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2510 MAX_NR_ZONES - 1); 2511 zonelist->_zonerefs[j].zone = NULL; 2512 zonelist->_zonerefs[j].zone_idx = 0; 2513} 2514 2515/* 2516 * Build gfp_thisnode zonelists 2517 */ 2518static void build_thisnode_zonelists(pg_data_t *pgdat) 2519{ 2520 int j; 2521 struct zonelist *zonelist; 2522 2523 zonelist = &pgdat->node_zonelists[1]; 2524 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2525 zonelist->_zonerefs[j].zone = NULL; 2526 zonelist->_zonerefs[j].zone_idx = 0; 2527} 2528 2529/* 2530 * Build zonelists ordered by zone and nodes within zones. 2531 * This results in conserving DMA zone[s] until all Normal memory is 2532 * exhausted, but results in overflowing to remote node while memory 2533 * may still exist in local DMA zone. 2534 */ 2535static int node_order[MAX_NUMNODES]; 2536 2537static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2538{ 2539 int pos, j, node; 2540 int zone_type; /* needs to be signed */ 2541 struct zone *z; 2542 struct zonelist *zonelist; 2543 2544 zonelist = &pgdat->node_zonelists[0]; 2545 pos = 0; 2546 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2547 for (j = 0; j < nr_nodes; j++) { 2548 node = node_order[j]; 2549 z = &NODE_DATA(node)->node_zones[zone_type]; 2550 if (populated_zone(z)) { 2551 zoneref_set_zone(z, 2552 &zonelist->_zonerefs[pos++]); 2553 check_highest_zone(zone_type); 2554 } 2555 } 2556 } 2557 zonelist->_zonerefs[pos].zone = NULL; 2558 zonelist->_zonerefs[pos].zone_idx = 0; 2559} 2560 2561static int default_zonelist_order(void) 2562{ 2563 int nid, zone_type; 2564 unsigned long low_kmem_size,total_size; 2565 struct zone *z; 2566 int average_size; 2567 /* 2568 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2569 * If they are really small and used heavily, the system can fall 2570 * into OOM very easily. 2571 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2572 */ 2573 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2574 low_kmem_size = 0; 2575 total_size = 0; 2576 for_each_online_node(nid) { 2577 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2578 z = &NODE_DATA(nid)->node_zones[zone_type]; 2579 if (populated_zone(z)) { 2580 if (zone_type < ZONE_NORMAL) 2581 low_kmem_size += z->present_pages; 2582 total_size += z->present_pages; 2583 } 2584 } 2585 } 2586 if (!low_kmem_size || /* there are no DMA area. */ 2587 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2588 return ZONELIST_ORDER_NODE; 2589 /* 2590 * look into each node's config. 2591 * If there is a node whose DMA/DMA32 memory is very big area on 2592 * local memory, NODE_ORDER may be suitable. 2593 */ 2594 average_size = total_size / 2595 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2596 for_each_online_node(nid) { 2597 low_kmem_size = 0; 2598 total_size = 0; 2599 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2600 z = &NODE_DATA(nid)->node_zones[zone_type]; 2601 if (populated_zone(z)) { 2602 if (zone_type < ZONE_NORMAL) 2603 low_kmem_size += z->present_pages; 2604 total_size += z->present_pages; 2605 } 2606 } 2607 if (low_kmem_size && 2608 total_size > average_size && /* ignore small node */ 2609 low_kmem_size > total_size * 70/100) 2610 return ZONELIST_ORDER_NODE; 2611 } 2612 return ZONELIST_ORDER_ZONE; 2613} 2614 2615static void set_zonelist_order(void) 2616{ 2617 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2618 current_zonelist_order = default_zonelist_order(); 2619 else 2620 current_zonelist_order = user_zonelist_order; 2621} 2622 2623static void build_zonelists(pg_data_t *pgdat) 2624{ 2625 int j, node, load; 2626 enum zone_type i; 2627 nodemask_t used_mask; 2628 int local_node, prev_node; 2629 struct zonelist *zonelist; 2630 int order = current_zonelist_order; 2631 2632 /* initialize zonelists */ 2633 for (i = 0; i < MAX_ZONELISTS; i++) { 2634 zonelist = pgdat->node_zonelists + i; 2635 zonelist->_zonerefs[0].zone = NULL; 2636 zonelist->_zonerefs[0].zone_idx = 0; 2637 } 2638 2639 /* NUMA-aware ordering of nodes */ 2640 local_node = pgdat->node_id; 2641 load = nr_online_nodes; 2642 prev_node = local_node; 2643 nodes_clear(used_mask); 2644 2645 memset(node_order, 0, sizeof(node_order)); 2646 j = 0; 2647 2648 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2649 int distance = node_distance(local_node, node); 2650 2651 /* 2652 * If another node is sufficiently far away then it is better 2653 * to reclaim pages in a zone before going off node. 2654 */ 2655 if (distance > RECLAIM_DISTANCE) 2656 zone_reclaim_mode = 1; 2657 2658 /* 2659 * We don't want to pressure a particular node. 2660 * So adding penalty to the first node in same 2661 * distance group to make it round-robin. 2662 */ 2663 if (distance != node_distance(local_node, prev_node)) 2664 node_load[node] = load; 2665 2666 prev_node = node; 2667 load--; 2668 if (order == ZONELIST_ORDER_NODE) 2669 build_zonelists_in_node_order(pgdat, node); 2670 else 2671 node_order[j++] = node; /* remember order */ 2672 } 2673 2674 if (order == ZONELIST_ORDER_ZONE) { 2675 /* calculate node order -- i.e., DMA last! */ 2676 build_zonelists_in_zone_order(pgdat, j); 2677 } 2678 2679 build_thisnode_zonelists(pgdat); 2680} 2681 2682/* Construct the zonelist performance cache - see further mmzone.h */ 2683static void build_zonelist_cache(pg_data_t *pgdat) 2684{ 2685 struct zonelist *zonelist; 2686 struct zonelist_cache *zlc; 2687 struct zoneref *z; 2688 2689 zonelist = &pgdat->node_zonelists[0]; 2690 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2691 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2692 for (z = zonelist->_zonerefs; z->zone; z++) 2693 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2694} 2695 2696 2697#else /* CONFIG_NUMA */ 2698 2699static void set_zonelist_order(void) 2700{ 2701 current_zonelist_order = ZONELIST_ORDER_ZONE; 2702} 2703 2704static void build_zonelists(pg_data_t *pgdat) 2705{ 2706 int node, local_node; 2707 enum zone_type j; 2708 struct zonelist *zonelist; 2709 2710 local_node = pgdat->node_id; 2711 2712 zonelist = &pgdat->node_zonelists[0]; 2713 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2714 2715 /* 2716 * Now we build the zonelist so that it contains the zones 2717 * of all the other nodes. 2718 * We don't want to pressure a particular node, so when 2719 * building the zones for node N, we make sure that the 2720 * zones coming right after the local ones are those from 2721 * node N+1 (modulo N) 2722 */ 2723 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2724 if (!node_online(node)) 2725 continue; 2726 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2727 MAX_NR_ZONES - 1); 2728 } 2729 for (node = 0; node < local_node; node++) { 2730 if (!node_online(node)) 2731 continue; 2732 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2733 MAX_NR_ZONES - 1); 2734 } 2735 2736 zonelist->_zonerefs[j].zone = NULL; 2737 zonelist->_zonerefs[j].zone_idx = 0; 2738} 2739 2740/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2741static void build_zonelist_cache(pg_data_t *pgdat) 2742{ 2743 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2744} 2745 2746#endif /* CONFIG_NUMA */ 2747 2748/* return values int ....just for stop_machine() */ 2749static int __build_all_zonelists(void *dummy) 2750{ 2751 int nid; 2752 2753#ifdef CONFIG_NUMA 2754 memset(node_load, 0, sizeof(node_load)); 2755#endif 2756 for_each_online_node(nid) { 2757 pg_data_t *pgdat = NODE_DATA(nid); 2758 2759 build_zonelists(pgdat); 2760 build_zonelist_cache(pgdat); 2761 } 2762 return 0; 2763} 2764 2765void build_all_zonelists(void) 2766{ 2767 set_zonelist_order(); 2768 2769 if (system_state == SYSTEM_BOOTING) { 2770 __build_all_zonelists(NULL); 2771 mminit_verify_zonelist(); 2772 cpuset_init_current_mems_allowed(); 2773 } else { 2774 /* we have to stop all cpus to guarantee there is no user 2775 of zonelist */ 2776 stop_machine(__build_all_zonelists, NULL, NULL); 2777 /* cpuset refresh routine should be here */ 2778 } 2779 vm_total_pages = nr_free_pagecache_pages(); 2780 /* 2781 * Disable grouping by mobility if the number of pages in the 2782 * system is too low to allow the mechanism to work. It would be 2783 * more accurate, but expensive to check per-zone. This check is 2784 * made on memory-hotadd so a system can start with mobility 2785 * disabled and enable it later 2786 */ 2787 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2788 page_group_by_mobility_disabled = 1; 2789 else 2790 page_group_by_mobility_disabled = 0; 2791 2792 printk("Built %i zonelists in %s order, mobility grouping %s. " 2793 "Total pages: %ld\n", 2794 nr_online_nodes, 2795 zonelist_order_name[current_zonelist_order], 2796 page_group_by_mobility_disabled ? "off" : "on", 2797 vm_total_pages); 2798#ifdef CONFIG_NUMA 2799 printk("Policy zone: %s\n", zone_names[policy_zone]); 2800#endif 2801} 2802 2803/* 2804 * Helper functions to size the waitqueue hash table. 2805 * Essentially these want to choose hash table sizes sufficiently 2806 * large so that collisions trying to wait on pages are rare. 2807 * But in fact, the number of active page waitqueues on typical 2808 * systems is ridiculously low, less than 200. So this is even 2809 * conservative, even though it seems large. 2810 * 2811 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2812 * waitqueues, i.e. the size of the waitq table given the number of pages. 2813 */ 2814#define PAGES_PER_WAITQUEUE 256 2815 2816#ifndef CONFIG_MEMORY_HOTPLUG 2817static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2818{ 2819 unsigned long size = 1; 2820 2821 pages /= PAGES_PER_WAITQUEUE; 2822 2823 while (size < pages) 2824 size <<= 1; 2825 2826 /* 2827 * Once we have dozens or even hundreds of threads sleeping 2828 * on IO we've got bigger problems than wait queue collision. 2829 * Limit the size of the wait table to a reasonable size. 2830 */ 2831 size = min(size, 4096UL); 2832 2833 return max(size, 4UL); 2834} 2835#else 2836/* 2837 * A zone's size might be changed by hot-add, so it is not possible to determine 2838 * a suitable size for its wait_table. So we use the maximum size now. 2839 * 2840 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2841 * 2842 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2843 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2844 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2845 * 2846 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2847 * or more by the traditional way. (See above). It equals: 2848 * 2849 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2850 * ia64(16K page size) : = ( 8G + 4M)byte. 2851 * powerpc (64K page size) : = (32G +16M)byte. 2852 */ 2853static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2854{ 2855 return 4096UL; 2856} 2857#endif 2858 2859/* 2860 * This is an integer logarithm so that shifts can be used later 2861 * to extract the more random high bits from the multiplicative 2862 * hash function before the remainder is taken. 2863 */ 2864static inline unsigned long wait_table_bits(unsigned long size) 2865{ 2866 return ffz(~size); 2867} 2868 2869#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2870 2871/* 2872 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2873 * of blocks reserved is based on min_wmark_pages(zone). The memory within 2874 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 2875 * higher will lead to a bigger reserve which will get freed as contiguous 2876 * blocks as reclaim kicks in 2877 */ 2878static void setup_zone_migrate_reserve(struct zone *zone) 2879{ 2880 unsigned long start_pfn, pfn, end_pfn; 2881 struct page *page; 2882 unsigned long block_migratetype; 2883 int reserve; 2884 2885 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2886 start_pfn = zone->zone_start_pfn; 2887 end_pfn = start_pfn + zone->spanned_pages; 2888 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 2889 pageblock_order; 2890 2891 /* 2892 * Reserve blocks are generally in place to help high-order atomic 2893 * allocations that are short-lived. A min_free_kbytes value that 2894 * would result in more than 2 reserve blocks for atomic allocations 2895 * is assumed to be in place to help anti-fragmentation for the 2896 * future allocation of hugepages at runtime. 2897 */ 2898 reserve = min(2, reserve); 2899 2900 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2901 if (!pfn_valid(pfn)) 2902 continue; 2903 page = pfn_to_page(pfn); 2904 2905 /* Watch out for overlapping nodes */ 2906 if (page_to_nid(page) != zone_to_nid(zone)) 2907 continue; 2908 2909 /* Blocks with reserved pages will never free, skip them. */ 2910 if (PageReserved(page)) 2911 continue; 2912 2913 block_migratetype = get_pageblock_migratetype(page); 2914 2915 /* If this block is reserved, account for it */ 2916 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2917 reserve--; 2918 continue; 2919 } 2920 2921 /* Suitable for reserving if this block is movable */ 2922 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2923 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2924 move_freepages_block(zone, page, MIGRATE_RESERVE); 2925 reserve--; 2926 continue; 2927 } 2928 2929 /* 2930 * If the reserve is met and this is a previous reserved block, 2931 * take it back 2932 */ 2933 if (block_migratetype == MIGRATE_RESERVE) { 2934 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2935 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2936 } 2937 } 2938} 2939 2940/* 2941 * Initially all pages are reserved - free ones are freed 2942 * up by free_all_bootmem() once the early boot process is 2943 * done. Non-atomic initialization, single-pass. 2944 */ 2945void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2946 unsigned long start_pfn, enum memmap_context context) 2947{ 2948 struct page *page; 2949 unsigned long end_pfn = start_pfn + size; 2950 unsigned long pfn; 2951 struct zone *z; 2952 2953 if (highest_memmap_pfn < end_pfn - 1) 2954 highest_memmap_pfn = end_pfn - 1; 2955 2956 z = &NODE_DATA(nid)->node_zones[zone]; 2957 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2958 /* 2959 * There can be holes in boot-time mem_map[]s 2960 * handed to this function. They do not 2961 * exist on hotplugged memory. 2962 */ 2963 if (context == MEMMAP_EARLY) { 2964 if (!early_pfn_valid(pfn)) 2965 continue; 2966 if (!early_pfn_in_nid(pfn, nid)) 2967 continue; 2968 } 2969 page = pfn_to_page(pfn); 2970 set_page_links(page, zone, nid, pfn); 2971 mminit_verify_page_links(page, zone, nid, pfn); 2972 init_page_count(page); 2973 reset_page_mapcount(page); 2974 SetPageReserved(page); 2975 /* 2976 * Mark the block movable so that blocks are reserved for 2977 * movable at startup. This will force kernel allocations 2978 * to reserve their blocks rather than leaking throughout 2979 * the address space during boot when many long-lived 2980 * kernel allocations are made. Later some blocks near 2981 * the start are marked MIGRATE_RESERVE by 2982 * setup_zone_migrate_reserve() 2983 * 2984 * bitmap is created for zone's valid pfn range. but memmap 2985 * can be created for invalid pages (for alignment) 2986 * check here not to call set_pageblock_migratetype() against 2987 * pfn out of zone. 2988 */ 2989 if ((z->zone_start_pfn <= pfn) 2990 && (pfn < z->zone_start_pfn + z->spanned_pages) 2991 && !(pfn & (pageblock_nr_pages - 1))) 2992 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2993 2994 INIT_LIST_HEAD(&page->lru); 2995#ifdef WANT_PAGE_VIRTUAL 2996 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2997 if (!is_highmem_idx(zone)) 2998 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2999#endif 3000 } 3001} 3002 3003static void __meminit zone_init_free_lists(struct zone *zone) 3004{ 3005 int order, t; 3006 for_each_migratetype_order(order, t) { 3007 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3008 zone->free_area[order].nr_free = 0; 3009 } 3010} 3011 3012#ifndef __HAVE_ARCH_MEMMAP_INIT 3013#define memmap_init(size, nid, zone, start_pfn) \ 3014 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3015#endif 3016 3017static int zone_batchsize(struct zone *zone) 3018{ 3019#ifdef CONFIG_MMU 3020 int batch; 3021 3022 /* 3023 * The per-cpu-pages pools are set to around 1000th of the 3024 * size of the zone. But no more than 1/2 of a meg. 3025 * 3026 * OK, so we don't know how big the cache is. So guess. 3027 */ 3028 batch = zone->present_pages / 1024; 3029 if (batch * PAGE_SIZE > 512 * 1024) 3030 batch = (512 * 1024) / PAGE_SIZE; 3031 batch /= 4; /* We effectively *= 4 below */ 3032 if (batch < 1) 3033 batch = 1; 3034 3035 /* 3036 * Clamp the batch to a 2^n - 1 value. Having a power 3037 * of 2 value was found to be more likely to have 3038 * suboptimal cache aliasing properties in some cases. 3039 * 3040 * For example if 2 tasks are alternately allocating 3041 * batches of pages, one task can end up with a lot 3042 * of pages of one half of the possible page colors 3043 * and the other with pages of the other colors. 3044 */ 3045 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3046 3047 return batch; 3048 3049#else 3050 /* The deferral and batching of frees should be suppressed under NOMMU 3051 * conditions. 3052 * 3053 * The problem is that NOMMU needs to be able to allocate large chunks 3054 * of contiguous memory as there's no hardware page translation to 3055 * assemble apparent contiguous memory from discontiguous pages. 3056 * 3057 * Queueing large contiguous runs of pages for batching, however, 3058 * causes the pages to actually be freed in smaller chunks. As there 3059 * can be a significant delay between the individual batches being 3060 * recycled, this leads to the once large chunks of space being 3061 * fragmented and becoming unavailable for high-order allocations. 3062 */ 3063 return 0; 3064#endif 3065} 3066 3067static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3068{ 3069 struct per_cpu_pages *pcp; 3070 int migratetype; 3071 3072 memset(p, 0, sizeof(*p)); 3073 3074 pcp = &p->pcp; 3075 pcp->count = 0; 3076 pcp->high = 6 * batch; 3077 pcp->batch = max(1UL, 1 * batch); 3078 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3079 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3080} 3081 3082/* 3083 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3084 * to the value high for the pageset p. 3085 */ 3086 3087static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3088 unsigned long high) 3089{ 3090 struct per_cpu_pages *pcp; 3091 3092 pcp = &p->pcp; 3093 pcp->high = high; 3094 pcp->batch = max(1UL, high/4); 3095 if ((high/4) > (PAGE_SHIFT * 8)) 3096 pcp->batch = PAGE_SHIFT * 8; 3097} 3098 3099 3100#ifdef CONFIG_NUMA 3101/* 3102 * Boot pageset table. One per cpu which is going to be used for all 3103 * zones and all nodes. The parameters will be set in such a way 3104 * that an item put on a list will immediately be handed over to 3105 * the buddy list. This is safe since pageset manipulation is done 3106 * with interrupts disabled. 3107 * 3108 * Some NUMA counter updates may also be caught by the boot pagesets. 3109 * 3110 * The boot_pagesets must be kept even after bootup is complete for 3111 * unused processors and/or zones. They do play a role for bootstrapping 3112 * hotplugged processors. 3113 * 3114 * zoneinfo_show() and maybe other functions do 3115 * not check if the processor is online before following the pageset pointer. 3116 * Other parts of the kernel may not check if the zone is available. 3117 */ 3118static struct per_cpu_pageset boot_pageset[NR_CPUS]; 3119 3120/* 3121 * Dynamically allocate memory for the 3122 * per cpu pageset array in struct zone. 3123 */ 3124static int __cpuinit process_zones(int cpu) 3125{ 3126 struct zone *zone, *dzone; 3127 int node = cpu_to_node(cpu); 3128 3129 node_set_state(node, N_CPU); /* this node has a cpu */ 3130 3131 for_each_populated_zone(zone) { 3132 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 3133 GFP_KERNEL, node); 3134 if (!zone_pcp(zone, cpu)) 3135 goto bad; 3136 3137 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 3138 3139 if (percpu_pagelist_fraction) 3140 setup_pagelist_highmark(zone_pcp(zone, cpu), 3141 (zone->present_pages / percpu_pagelist_fraction)); 3142 } 3143 3144 return 0; 3145bad: 3146 for_each_zone(dzone) { 3147 if (!populated_zone(dzone)) 3148 continue; 3149 if (dzone == zone) 3150 break; 3151 kfree(zone_pcp(dzone, cpu)); 3152 zone_pcp(dzone, cpu) = &boot_pageset[cpu]; 3153 } 3154 return -ENOMEM; 3155} 3156 3157static inline void free_zone_pagesets(int cpu) 3158{ 3159 struct zone *zone; 3160 3161 for_each_zone(zone) { 3162 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 3163 3164 /* Free per_cpu_pageset if it is slab allocated */ 3165 if (pset != &boot_pageset[cpu]) 3166 kfree(pset); 3167 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3168 } 3169} 3170 3171static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 3172 unsigned long action, 3173 void *hcpu) 3174{ 3175 int cpu = (long)hcpu; 3176 int ret = NOTIFY_OK; 3177 3178 switch (action) { 3179 case CPU_UP_PREPARE: 3180 case CPU_UP_PREPARE_FROZEN: 3181 if (process_zones(cpu)) 3182 ret = NOTIFY_BAD; 3183 break; 3184 case CPU_UP_CANCELED: 3185 case CPU_UP_CANCELED_FROZEN: 3186 case CPU_DEAD: 3187 case CPU_DEAD_FROZEN: 3188 free_zone_pagesets(cpu); 3189 break; 3190 default: 3191 break; 3192 } 3193 return ret; 3194} 3195 3196static struct notifier_block __cpuinitdata pageset_notifier = 3197 { &pageset_cpuup_callback, NULL, 0 }; 3198 3199void __init setup_per_cpu_pageset(void) 3200{ 3201 int err; 3202 3203 /* Initialize per_cpu_pageset for cpu 0. 3204 * A cpuup callback will do this for every cpu 3205 * as it comes online 3206 */ 3207 err = process_zones(smp_processor_id()); 3208 BUG_ON(err); 3209 register_cpu_notifier(&pageset_notifier); 3210} 3211 3212#endif 3213 3214static noinline __init_refok 3215int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3216{ 3217 int i; 3218 struct pglist_data *pgdat = zone->zone_pgdat; 3219 size_t alloc_size; 3220 3221 /* 3222 * The per-page waitqueue mechanism uses hashed waitqueues 3223 * per zone. 3224 */ 3225 zone->wait_table_hash_nr_entries = 3226 wait_table_hash_nr_entries(zone_size_pages); 3227 zone->wait_table_bits = 3228 wait_table_bits(zone->wait_table_hash_nr_entries); 3229 alloc_size = zone->wait_table_hash_nr_entries 3230 * sizeof(wait_queue_head_t); 3231 3232 if (!slab_is_available()) { 3233 zone->wait_table = (wait_queue_head_t *) 3234 alloc_bootmem_node(pgdat, alloc_size); 3235 } else { 3236 /* 3237 * This case means that a zone whose size was 0 gets new memory 3238 * via memory hot-add. 3239 * But it may be the case that a new node was hot-added. In 3240 * this case vmalloc() will not be able to use this new node's 3241 * memory - this wait_table must be initialized to use this new 3242 * node itself as well. 3243 * To use this new node's memory, further consideration will be 3244 * necessary. 3245 */ 3246 zone->wait_table = vmalloc(alloc_size); 3247 } 3248 if (!zone->wait_table) 3249 return -ENOMEM; 3250 3251 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3252 init_waitqueue_head(zone->wait_table + i); 3253 3254 return 0; 3255} 3256 3257static int __zone_pcp_update(void *data) 3258{ 3259 struct zone *zone = data; 3260 int cpu; 3261 unsigned long batch = zone_batchsize(zone), flags; 3262 3263 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3264 struct per_cpu_pageset *pset; 3265 struct per_cpu_pages *pcp; 3266 3267 pset = zone_pcp(zone, cpu); 3268 pcp = &pset->pcp; 3269 3270 local_irq_save(flags); 3271 free_pcppages_bulk(zone, pcp->count, pcp); 3272 setup_pageset(pset, batch); 3273 local_irq_restore(flags); 3274 } 3275 return 0; 3276} 3277 3278void zone_pcp_update(struct zone *zone) 3279{ 3280 stop_machine(__zone_pcp_update, zone, NULL); 3281} 3282 3283static __meminit void zone_pcp_init(struct zone *zone) 3284{ 3285 int cpu; 3286 unsigned long batch = zone_batchsize(zone); 3287 3288 for (cpu = 0; cpu < NR_CPUS; cpu++) { 3289#ifdef CONFIG_NUMA 3290 /* Early boot. Slab allocator not functional yet */ 3291 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 3292 setup_pageset(&boot_pageset[cpu],0); 3293#else 3294 setup_pageset(zone_pcp(zone,cpu), batch); 3295#endif 3296 } 3297 if (zone->present_pages) 3298 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 3299 zone->name, zone->present_pages, batch); 3300} 3301 3302__meminit int init_currently_empty_zone(struct zone *zone, 3303 unsigned long zone_start_pfn, 3304 unsigned long size, 3305 enum memmap_context context) 3306{ 3307 struct pglist_data *pgdat = zone->zone_pgdat; 3308 int ret; 3309 ret = zone_wait_table_init(zone, size); 3310 if (ret) 3311 return ret; 3312 pgdat->nr_zones = zone_idx(zone) + 1; 3313 3314 zone->zone_start_pfn = zone_start_pfn; 3315 3316 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3317 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3318 pgdat->node_id, 3319 (unsigned long)zone_idx(zone), 3320 zone_start_pfn, (zone_start_pfn + size)); 3321 3322 zone_init_free_lists(zone); 3323 3324 return 0; 3325} 3326 3327#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3328/* 3329 * Basic iterator support. Return the first range of PFNs for a node 3330 * Note: nid == MAX_NUMNODES returns first region regardless of node 3331 */ 3332static int __meminit first_active_region_index_in_nid(int nid) 3333{ 3334 int i; 3335 3336 for (i = 0; i < nr_nodemap_entries; i++) 3337 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3338 return i; 3339 3340 return -1; 3341} 3342 3343/* 3344 * Basic iterator support. Return the next active range of PFNs for a node 3345 * Note: nid == MAX_NUMNODES returns next region regardless of node 3346 */ 3347static int __meminit next_active_region_index_in_nid(int index, int nid) 3348{ 3349 for (index = index + 1; index < nr_nodemap_entries; index++) 3350 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3351 return index; 3352 3353 return -1; 3354} 3355 3356#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3357/* 3358 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3359 * Architectures may implement their own version but if add_active_range() 3360 * was used and there are no special requirements, this is a convenient 3361 * alternative 3362 */ 3363int __meminit __early_pfn_to_nid(unsigned long pfn) 3364{ 3365 int i; 3366 3367 for (i = 0; i < nr_nodemap_entries; i++) { 3368 unsigned long start_pfn = early_node_map[i].start_pfn; 3369 unsigned long end_pfn = early_node_map[i].end_pfn; 3370 3371 if (start_pfn <= pfn && pfn < end_pfn) 3372 return early_node_map[i].nid; 3373 } 3374 /* This is a memory hole */ 3375 return -1; 3376} 3377#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3378 3379int __meminit early_pfn_to_nid(unsigned long pfn) 3380{ 3381 int nid; 3382 3383 nid = __early_pfn_to_nid(pfn); 3384 if (nid >= 0) 3385 return nid; 3386 /* just returns 0 */ 3387 return 0; 3388} 3389 3390#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3391bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3392{ 3393 int nid; 3394 3395 nid = __early_pfn_to_nid(pfn); 3396 if (nid >= 0 && nid != node) 3397 return false; 3398 return true; 3399} 3400#endif 3401 3402/* Basic iterator support to walk early_node_map[] */ 3403#define for_each_active_range_index_in_nid(i, nid) \ 3404 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3405 i = next_active_region_index_in_nid(i, nid)) 3406 3407/** 3408 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3409 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3410 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3411 * 3412 * If an architecture guarantees that all ranges registered with 3413 * add_active_ranges() contain no holes and may be freed, this 3414 * this function may be used instead of calling free_bootmem() manually. 3415 */ 3416void __init free_bootmem_with_active_regions(int nid, 3417 unsigned long max_low_pfn) 3418{ 3419 int i; 3420 3421 for_each_active_range_index_in_nid(i, nid) { 3422 unsigned long size_pages = 0; 3423 unsigned long end_pfn = early_node_map[i].end_pfn; 3424 3425 if (early_node_map[i].start_pfn >= max_low_pfn) 3426 continue; 3427 3428 if (end_pfn > max_low_pfn) 3429 end_pfn = max_low_pfn; 3430 3431 size_pages = end_pfn - early_node_map[i].start_pfn; 3432 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3433 PFN_PHYS(early_node_map[i].start_pfn), 3434 size_pages << PAGE_SHIFT); 3435 } 3436} 3437 3438int __init add_from_early_node_map(struct range *range, int az, 3439 int nr_range, int nid) 3440{ 3441 int i; 3442 u64 start, end; 3443 3444 /* need to go over early_node_map to find out good range for node */ 3445 for_each_active_range_index_in_nid(i, nid) { 3446 start = early_node_map[i].start_pfn; 3447 end = early_node_map[i].end_pfn; 3448 nr_range = add_range(range, az, nr_range, start, end); 3449 } 3450 return nr_range; 3451} 3452 3453void * __init __alloc_memory_core_early(int nid, u64 size, u64 align, 3454 u64 goal, u64 limit) 3455{ 3456 int i; 3457 void *ptr; 3458 3459 /* need to go over early_node_map to find out good range for node */ 3460 for_each_active_range_index_in_nid(i, nid) { 3461 u64 addr; 3462 u64 ei_start, ei_last; 3463 3464 ei_last = early_node_map[i].end_pfn; 3465 ei_last <<= PAGE_SHIFT; 3466 ei_start = early_node_map[i].start_pfn; 3467 ei_start <<= PAGE_SHIFT; 3468 addr = find_early_area(ei_start, ei_last, 3469 goal, limit, size, align); 3470 3471 if (addr == -1ULL) 3472 continue; 3473 3474#if 0 3475 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n", 3476 nid, 3477 ei_start, ei_last, goal, limit, size, 3478 align, addr); 3479#endif 3480 3481 ptr = phys_to_virt(addr); 3482 memset(ptr, 0, size); 3483 reserve_early_without_check(addr, addr + size, "BOOTMEM"); 3484 return ptr; 3485 } 3486 3487 return NULL; 3488} 3489 3490 3491void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3492{ 3493 int i; 3494 int ret; 3495 3496 for_each_active_range_index_in_nid(i, nid) { 3497 ret = work_fn(early_node_map[i].start_pfn, 3498 early_node_map[i].end_pfn, data); 3499 if (ret) 3500 break; 3501 } 3502} 3503/** 3504 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3505 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3506 * 3507 * If an architecture guarantees that all ranges registered with 3508 * add_active_ranges() contain no holes and may be freed, this 3509 * function may be used instead of calling memory_present() manually. 3510 */ 3511void __init sparse_memory_present_with_active_regions(int nid) 3512{ 3513 int i; 3514 3515 for_each_active_range_index_in_nid(i, nid) 3516 memory_present(early_node_map[i].nid, 3517 early_node_map[i].start_pfn, 3518 early_node_map[i].end_pfn); 3519} 3520 3521/** 3522 * get_pfn_range_for_nid - Return the start and end page frames for a node 3523 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3524 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3525 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3526 * 3527 * It returns the start and end page frame of a node based on information 3528 * provided by an arch calling add_active_range(). If called for a node 3529 * with no available memory, a warning is printed and the start and end 3530 * PFNs will be 0. 3531 */ 3532void __meminit get_pfn_range_for_nid(unsigned int nid, 3533 unsigned long *start_pfn, unsigned long *end_pfn) 3534{ 3535 int i; 3536 *start_pfn = -1UL; 3537 *end_pfn = 0; 3538 3539 for_each_active_range_index_in_nid(i, nid) { 3540 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3541 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3542 } 3543 3544 if (*start_pfn == -1UL) 3545 *start_pfn = 0; 3546} 3547 3548/* 3549 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3550 * assumption is made that zones within a node are ordered in monotonic 3551 * increasing memory addresses so that the "highest" populated zone is used 3552 */ 3553static void __init find_usable_zone_for_movable(void) 3554{ 3555 int zone_index; 3556 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3557 if (zone_index == ZONE_MOVABLE) 3558 continue; 3559 3560 if (arch_zone_highest_possible_pfn[zone_index] > 3561 arch_zone_lowest_possible_pfn[zone_index]) 3562 break; 3563 } 3564 3565 VM_BUG_ON(zone_index == -1); 3566 movable_zone = zone_index; 3567} 3568 3569/* 3570 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3571 * because it is sized independant of architecture. Unlike the other zones, 3572 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3573 * in each node depending on the size of each node and how evenly kernelcore 3574 * is distributed. This helper function adjusts the zone ranges 3575 * provided by the architecture for a given node by using the end of the 3576 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3577 * zones within a node are in order of monotonic increases memory addresses 3578 */ 3579static void __meminit adjust_zone_range_for_zone_movable(int nid, 3580 unsigned long zone_type, 3581 unsigned long node_start_pfn, 3582 unsigned long node_end_pfn, 3583 unsigned long *zone_start_pfn, 3584 unsigned long *zone_end_pfn) 3585{ 3586 /* Only adjust if ZONE_MOVABLE is on this node */ 3587 if (zone_movable_pfn[nid]) { 3588 /* Size ZONE_MOVABLE */ 3589 if (zone_type == ZONE_MOVABLE) { 3590 *zone_start_pfn = zone_movable_pfn[nid]; 3591 *zone_end_pfn = min(node_end_pfn, 3592 arch_zone_highest_possible_pfn[movable_zone]); 3593 3594 /* Adjust for ZONE_MOVABLE starting within this range */ 3595 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3596 *zone_end_pfn > zone_movable_pfn[nid]) { 3597 *zone_end_pfn = zone_movable_pfn[nid]; 3598 3599 /* Check if this whole range is within ZONE_MOVABLE */ 3600 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3601 *zone_start_pfn = *zone_end_pfn; 3602 } 3603} 3604 3605/* 3606 * Return the number of pages a zone spans in a node, including holes 3607 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3608 */ 3609static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3610 unsigned long zone_type, 3611 unsigned long *ignored) 3612{ 3613 unsigned long node_start_pfn, node_end_pfn; 3614 unsigned long zone_start_pfn, zone_end_pfn; 3615 3616 /* Get the start and end of the node and zone */ 3617 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3618 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3619 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3620 adjust_zone_range_for_zone_movable(nid, zone_type, 3621 node_start_pfn, node_end_pfn, 3622 &zone_start_pfn, &zone_end_pfn); 3623 3624 /* Check that this node has pages within the zone's required range */ 3625 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3626 return 0; 3627 3628 /* Move the zone boundaries inside the node if necessary */ 3629 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3630 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3631 3632 /* Return the spanned pages */ 3633 return zone_end_pfn - zone_start_pfn; 3634} 3635 3636/* 3637 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3638 * then all holes in the requested range will be accounted for. 3639 */ 3640unsigned long __meminit __absent_pages_in_range(int nid, 3641 unsigned long range_start_pfn, 3642 unsigned long range_end_pfn) 3643{ 3644 int i = 0; 3645 unsigned long prev_end_pfn = 0, hole_pages = 0; 3646 unsigned long start_pfn; 3647 3648 /* Find the end_pfn of the first active range of pfns in the node */ 3649 i = first_active_region_index_in_nid(nid); 3650 if (i == -1) 3651 return 0; 3652 3653 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3654 3655 /* Account for ranges before physical memory on this node */ 3656 if (early_node_map[i].start_pfn > range_start_pfn) 3657 hole_pages = prev_end_pfn - range_start_pfn; 3658 3659 /* Find all holes for the zone within the node */ 3660 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3661 3662 /* No need to continue if prev_end_pfn is outside the zone */ 3663 if (prev_end_pfn >= range_end_pfn) 3664 break; 3665 3666 /* Make sure the end of the zone is not within the hole */ 3667 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3668 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3669 3670 /* Update the hole size cound and move on */ 3671 if (start_pfn > range_start_pfn) { 3672 BUG_ON(prev_end_pfn > start_pfn); 3673 hole_pages += start_pfn - prev_end_pfn; 3674 } 3675 prev_end_pfn = early_node_map[i].end_pfn; 3676 } 3677 3678 /* Account for ranges past physical memory on this node */ 3679 if (range_end_pfn > prev_end_pfn) 3680 hole_pages += range_end_pfn - 3681 max(range_start_pfn, prev_end_pfn); 3682 3683 return hole_pages; 3684} 3685 3686/** 3687 * absent_pages_in_range - Return number of page frames in holes within a range 3688 * @start_pfn: The start PFN to start searching for holes 3689 * @end_pfn: The end PFN to stop searching for holes 3690 * 3691 * It returns the number of pages frames in memory holes within a range. 3692 */ 3693unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3694 unsigned long end_pfn) 3695{ 3696 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3697} 3698 3699/* Return the number of page frames in holes in a zone on a node */ 3700static unsigned long __meminit zone_absent_pages_in_node(int nid, 3701 unsigned long zone_type, 3702 unsigned long *ignored) 3703{ 3704 unsigned long node_start_pfn, node_end_pfn; 3705 unsigned long zone_start_pfn, zone_end_pfn; 3706 3707 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3708 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3709 node_start_pfn); 3710 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3711 node_end_pfn); 3712 3713 adjust_zone_range_for_zone_movable(nid, zone_type, 3714 node_start_pfn, node_end_pfn, 3715 &zone_start_pfn, &zone_end_pfn); 3716 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3717} 3718 3719#else 3720static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3721 unsigned long zone_type, 3722 unsigned long *zones_size) 3723{ 3724 return zones_size[zone_type]; 3725} 3726 3727static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3728 unsigned long zone_type, 3729 unsigned long *zholes_size) 3730{ 3731 if (!zholes_size) 3732 return 0; 3733 3734 return zholes_size[zone_type]; 3735} 3736 3737#endif 3738 3739static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3740 unsigned long *zones_size, unsigned long *zholes_size) 3741{ 3742 unsigned long realtotalpages, totalpages = 0; 3743 enum zone_type i; 3744 3745 for (i = 0; i < MAX_NR_ZONES; i++) 3746 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3747 zones_size); 3748 pgdat->node_spanned_pages = totalpages; 3749 3750 realtotalpages = totalpages; 3751 for (i = 0; i < MAX_NR_ZONES; i++) 3752 realtotalpages -= 3753 zone_absent_pages_in_node(pgdat->node_id, i, 3754 zholes_size); 3755 pgdat->node_present_pages = realtotalpages; 3756 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3757 realtotalpages); 3758} 3759 3760#ifndef CONFIG_SPARSEMEM 3761/* 3762 * Calculate the size of the zone->blockflags rounded to an unsigned long 3763 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3764 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3765 * round what is now in bits to nearest long in bits, then return it in 3766 * bytes. 3767 */ 3768static unsigned long __init usemap_size(unsigned long zonesize) 3769{ 3770 unsigned long usemapsize; 3771 3772 usemapsize = roundup(zonesize, pageblock_nr_pages); 3773 usemapsize = usemapsize >> pageblock_order; 3774 usemapsize *= NR_PAGEBLOCK_BITS; 3775 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3776 3777 return usemapsize / 8; 3778} 3779 3780static void __init setup_usemap(struct pglist_data *pgdat, 3781 struct zone *zone, unsigned long zonesize) 3782{ 3783 unsigned long usemapsize = usemap_size(zonesize); 3784 zone->pageblock_flags = NULL; 3785 if (usemapsize) 3786 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3787} 3788#else 3789static void inline setup_usemap(struct pglist_data *pgdat, 3790 struct zone *zone, unsigned long zonesize) {} 3791#endif /* CONFIG_SPARSEMEM */ 3792 3793#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3794 3795/* Return a sensible default order for the pageblock size. */ 3796static inline int pageblock_default_order(void) 3797{ 3798 if (HPAGE_SHIFT > PAGE_SHIFT) 3799 return HUGETLB_PAGE_ORDER; 3800 3801 return MAX_ORDER-1; 3802} 3803 3804/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3805static inline void __init set_pageblock_order(unsigned int order) 3806{ 3807 /* Check that pageblock_nr_pages has not already been setup */ 3808 if (pageblock_order) 3809 return; 3810 3811 /* 3812 * Assume the largest contiguous order of interest is a huge page. 3813 * This value may be variable depending on boot parameters on IA64 3814 */ 3815 pageblock_order = order; 3816} 3817#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3818 3819/* 3820 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3821 * and pageblock_default_order() are unused as pageblock_order is set 3822 * at compile-time. See include/linux/pageblock-flags.h for the values of 3823 * pageblock_order based on the kernel config 3824 */ 3825static inline int pageblock_default_order(unsigned int order) 3826{ 3827 return MAX_ORDER-1; 3828} 3829#define set_pageblock_order(x) do {} while (0) 3830 3831#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3832 3833/* 3834 * Set up the zone data structures: 3835 * - mark all pages reserved 3836 * - mark all memory queues empty 3837 * - clear the memory bitmaps 3838 */ 3839static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3840 unsigned long *zones_size, unsigned long *zholes_size) 3841{ 3842 enum zone_type j; 3843 int nid = pgdat->node_id; 3844 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3845 int ret; 3846 3847 pgdat_resize_init(pgdat); 3848 pgdat->nr_zones = 0; 3849 init_waitqueue_head(&pgdat->kswapd_wait); 3850 pgdat->kswapd_max_order = 0; 3851 pgdat_page_cgroup_init(pgdat); 3852 3853 for (j = 0; j < MAX_NR_ZONES; j++) { 3854 struct zone *zone = pgdat->node_zones + j; 3855 unsigned long size, realsize, memmap_pages; 3856 enum lru_list l; 3857 3858 size = zone_spanned_pages_in_node(nid, j, zones_size); 3859 realsize = size - zone_absent_pages_in_node(nid, j, 3860 zholes_size); 3861 3862 /* 3863 * Adjust realsize so that it accounts for how much memory 3864 * is used by this zone for memmap. This affects the watermark 3865 * and per-cpu initialisations 3866 */ 3867 memmap_pages = 3868 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3869 if (realsize >= memmap_pages) { 3870 realsize -= memmap_pages; 3871 if (memmap_pages) 3872 printk(KERN_DEBUG 3873 " %s zone: %lu pages used for memmap\n", 3874 zone_names[j], memmap_pages); 3875 } else 3876 printk(KERN_WARNING 3877 " %s zone: %lu pages exceeds realsize %lu\n", 3878 zone_names[j], memmap_pages, realsize); 3879 3880 /* Account for reserved pages */ 3881 if (j == 0 && realsize > dma_reserve) { 3882 realsize -= dma_reserve; 3883 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3884 zone_names[0], dma_reserve); 3885 } 3886 3887 if (!is_highmem_idx(j)) 3888 nr_kernel_pages += realsize; 3889 nr_all_pages += realsize; 3890 3891 zone->spanned_pages = size; 3892 zone->present_pages = realsize; 3893#ifdef CONFIG_NUMA 3894 zone->node = nid; 3895 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3896 / 100; 3897 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3898#endif 3899 zone->name = zone_names[j]; 3900 spin_lock_init(&zone->lock); 3901 spin_lock_init(&zone->lru_lock); 3902 zone_seqlock_init(zone); 3903 zone->zone_pgdat = pgdat; 3904 3905 zone->prev_priority = DEF_PRIORITY; 3906 3907 zone_pcp_init(zone); 3908 for_each_lru(l) { 3909 INIT_LIST_HEAD(&zone->lru[l].list); 3910 zone->reclaim_stat.nr_saved_scan[l] = 0; 3911 } 3912 zone->reclaim_stat.recent_rotated[0] = 0; 3913 zone->reclaim_stat.recent_rotated[1] = 0; 3914 zone->reclaim_stat.recent_scanned[0] = 0; 3915 zone->reclaim_stat.recent_scanned[1] = 0; 3916 zap_zone_vm_stats(zone); 3917 zone->flags = 0; 3918 if (!size) 3919 continue; 3920 3921 set_pageblock_order(pageblock_default_order()); 3922 setup_usemap(pgdat, zone, size); 3923 ret = init_currently_empty_zone(zone, zone_start_pfn, 3924 size, MEMMAP_EARLY); 3925 BUG_ON(ret); 3926 memmap_init(size, nid, j, zone_start_pfn); 3927 zone_start_pfn += size; 3928 } 3929} 3930 3931static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3932{ 3933 /* Skip empty nodes */ 3934 if (!pgdat->node_spanned_pages) 3935 return; 3936 3937#ifdef CONFIG_FLAT_NODE_MEM_MAP 3938 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3939 if (!pgdat->node_mem_map) { 3940 unsigned long size, start, end; 3941 struct page *map; 3942 3943 /* 3944 * The zone's endpoints aren't required to be MAX_ORDER 3945 * aligned but the node_mem_map endpoints must be in order 3946 * for the buddy allocator to function correctly. 3947 */ 3948 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3949 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3950 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3951 size = (end - start) * sizeof(struct page); 3952 map = alloc_remap(pgdat->node_id, size); 3953 if (!map) 3954 map = alloc_bootmem_node(pgdat, size); 3955 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3956 } 3957#ifndef CONFIG_NEED_MULTIPLE_NODES 3958 /* 3959 * With no DISCONTIG, the global mem_map is just set as node 0's 3960 */ 3961 if (pgdat == NODE_DATA(0)) { 3962 mem_map = NODE_DATA(0)->node_mem_map; 3963#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3964 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3965 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3966#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3967 } 3968#endif 3969#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3970} 3971 3972void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 3973 unsigned long node_start_pfn, unsigned long *zholes_size) 3974{ 3975 pg_data_t *pgdat = NODE_DATA(nid); 3976 3977 pgdat->node_id = nid; 3978 pgdat->node_start_pfn = node_start_pfn; 3979 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3980 3981 alloc_node_mem_map(pgdat); 3982#ifdef CONFIG_FLAT_NODE_MEM_MAP 3983 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 3984 nid, (unsigned long)pgdat, 3985 (unsigned long)pgdat->node_mem_map); 3986#endif 3987 3988 free_area_init_core(pgdat, zones_size, zholes_size); 3989} 3990 3991#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3992 3993#if MAX_NUMNODES > 1 3994/* 3995 * Figure out the number of possible node ids. 3996 */ 3997static void __init setup_nr_node_ids(void) 3998{ 3999 unsigned int node; 4000 unsigned int highest = 0; 4001 4002 for_each_node_mask(node, node_possible_map) 4003 highest = node; 4004 nr_node_ids = highest + 1; 4005} 4006#else 4007static inline void setup_nr_node_ids(void) 4008{ 4009} 4010#endif 4011 4012/** 4013 * add_active_range - Register a range of PFNs backed by physical memory 4014 * @nid: The node ID the range resides on 4015 * @start_pfn: The start PFN of the available physical memory 4016 * @end_pfn: The end PFN of the available physical memory 4017 * 4018 * These ranges are stored in an early_node_map[] and later used by 4019 * free_area_init_nodes() to calculate zone sizes and holes. If the 4020 * range spans a memory hole, it is up to the architecture to ensure 4021 * the memory is not freed by the bootmem allocator. If possible 4022 * the range being registered will be merged with existing ranges. 4023 */ 4024void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4025 unsigned long end_pfn) 4026{ 4027 int i; 4028 4029 mminit_dprintk(MMINIT_TRACE, "memory_register", 4030 "Entering add_active_range(%d, %#lx, %#lx) " 4031 "%d entries of %d used\n", 4032 nid, start_pfn, end_pfn, 4033 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4034 4035 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4036 4037 /* Merge with existing active regions if possible */ 4038 for (i = 0; i < nr_nodemap_entries; i++) { 4039 if (early_node_map[i].nid != nid) 4040 continue; 4041 4042 /* Skip if an existing region covers this new one */ 4043 if (start_pfn >= early_node_map[i].start_pfn && 4044 end_pfn <= early_node_map[i].end_pfn) 4045 return; 4046 4047 /* Merge forward if suitable */ 4048 if (start_pfn <= early_node_map[i].end_pfn && 4049 end_pfn > early_node_map[i].end_pfn) { 4050 early_node_map[i].end_pfn = end_pfn; 4051 return; 4052 } 4053 4054 /* Merge backward if suitable */ 4055 if (start_pfn < early_node_map[i].start_pfn && 4056 end_pfn >= early_node_map[i].start_pfn) { 4057 early_node_map[i].start_pfn = start_pfn; 4058 return; 4059 } 4060 } 4061 4062 /* Check that early_node_map is large enough */ 4063 if (i >= MAX_ACTIVE_REGIONS) { 4064 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4065 MAX_ACTIVE_REGIONS); 4066 return; 4067 } 4068 4069 early_node_map[i].nid = nid; 4070 early_node_map[i].start_pfn = start_pfn; 4071 early_node_map[i].end_pfn = end_pfn; 4072 nr_nodemap_entries = i + 1; 4073} 4074 4075/** 4076 * remove_active_range - Shrink an existing registered range of PFNs 4077 * @nid: The node id the range is on that should be shrunk 4078 * @start_pfn: The new PFN of the range 4079 * @end_pfn: The new PFN of the range 4080 * 4081 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4082 * The map is kept near the end physical page range that has already been 4083 * registered. This function allows an arch to shrink an existing registered 4084 * range. 4085 */ 4086void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4087 unsigned long end_pfn) 4088{ 4089 int i, j; 4090 int removed = 0; 4091 4092 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4093 nid, start_pfn, end_pfn); 4094 4095 /* Find the old active region end and shrink */ 4096 for_each_active_range_index_in_nid(i, nid) { 4097 if (early_node_map[i].start_pfn >= start_pfn && 4098 early_node_map[i].end_pfn <= end_pfn) { 4099 /* clear it */ 4100 early_node_map[i].start_pfn = 0; 4101 early_node_map[i].end_pfn = 0; 4102 removed = 1; 4103 continue; 4104 } 4105 if (early_node_map[i].start_pfn < start_pfn && 4106 early_node_map[i].end_pfn > start_pfn) { 4107 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4108 early_node_map[i].end_pfn = start_pfn; 4109 if (temp_end_pfn > end_pfn) 4110 add_active_range(nid, end_pfn, temp_end_pfn); 4111 continue; 4112 } 4113 if (early_node_map[i].start_pfn >= start_pfn && 4114 early_node_map[i].end_pfn > end_pfn && 4115 early_node_map[i].start_pfn < end_pfn) { 4116 early_node_map[i].start_pfn = end_pfn; 4117 continue; 4118 } 4119 } 4120 4121 if (!removed) 4122 return; 4123 4124 /* remove the blank ones */ 4125 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4126 if (early_node_map[i].nid != nid) 4127 continue; 4128 if (early_node_map[i].end_pfn) 4129 continue; 4130 /* we found it, get rid of it */ 4131 for (j = i; j < nr_nodemap_entries - 1; j++) 4132 memcpy(&early_node_map[j], &early_node_map[j+1], 4133 sizeof(early_node_map[j])); 4134 j = nr_nodemap_entries - 1; 4135 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4136 nr_nodemap_entries--; 4137 } 4138} 4139 4140/** 4141 * remove_all_active_ranges - Remove all currently registered regions 4142 * 4143 * During discovery, it may be found that a table like SRAT is invalid 4144 * and an alternative discovery method must be used. This function removes 4145 * all currently registered regions. 4146 */ 4147void __init remove_all_active_ranges(void) 4148{ 4149 memset(early_node_map, 0, sizeof(early_node_map)); 4150 nr_nodemap_entries = 0; 4151} 4152 4153/* Compare two active node_active_regions */ 4154static int __init cmp_node_active_region(const void *a, const void *b) 4155{ 4156 struct node_active_region *arange = (struct node_active_region *)a; 4157 struct node_active_region *brange = (struct node_active_region *)b; 4158 4159 /* Done this way to avoid overflows */ 4160 if (arange->start_pfn > brange->start_pfn) 4161 return 1; 4162 if (arange->start_pfn < brange->start_pfn) 4163 return -1; 4164 4165 return 0; 4166} 4167 4168/* sort the node_map by start_pfn */ 4169void __init sort_node_map(void) 4170{ 4171 sort(early_node_map, (size_t)nr_nodemap_entries, 4172 sizeof(struct node_active_region), 4173 cmp_node_active_region, NULL); 4174} 4175 4176/* Find the lowest pfn for a node */ 4177static unsigned long __init find_min_pfn_for_node(int nid) 4178{ 4179 int i; 4180 unsigned long min_pfn = ULONG_MAX; 4181 4182 /* Assuming a sorted map, the first range found has the starting pfn */ 4183 for_each_active_range_index_in_nid(i, nid) 4184 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4185 4186 if (min_pfn == ULONG_MAX) { 4187 printk(KERN_WARNING 4188 "Could not find start_pfn for node %d\n", nid); 4189 return 0; 4190 } 4191 4192 return min_pfn; 4193} 4194 4195/** 4196 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4197 * 4198 * It returns the minimum PFN based on information provided via 4199 * add_active_range(). 4200 */ 4201unsigned long __init find_min_pfn_with_active_regions(void) 4202{ 4203 return find_min_pfn_for_node(MAX_NUMNODES); 4204} 4205 4206/* 4207 * early_calculate_totalpages() 4208 * Sum pages in active regions for movable zone. 4209 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4210 */ 4211static unsigned long __init early_calculate_totalpages(void) 4212{ 4213 int i; 4214 unsigned long totalpages = 0; 4215 4216 for (i = 0; i < nr_nodemap_entries; i++) { 4217 unsigned long pages = early_node_map[i].end_pfn - 4218 early_node_map[i].start_pfn; 4219 totalpages += pages; 4220 if (pages) 4221 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4222 } 4223 return totalpages; 4224} 4225 4226/* 4227 * Find the PFN the Movable zone begins in each node. Kernel memory 4228 * is spread evenly between nodes as long as the nodes have enough 4229 * memory. When they don't, some nodes will have more kernelcore than 4230 * others 4231 */ 4232static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4233{ 4234 int i, nid; 4235 unsigned long usable_startpfn; 4236 unsigned long kernelcore_node, kernelcore_remaining; 4237 /* save the state before borrow the nodemask */ 4238 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4239 unsigned long totalpages = early_calculate_totalpages(); 4240 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4241 4242 /* 4243 * If movablecore was specified, calculate what size of 4244 * kernelcore that corresponds so that memory usable for 4245 * any allocation type is evenly spread. If both kernelcore 4246 * and movablecore are specified, then the value of kernelcore 4247 * will be used for required_kernelcore if it's greater than 4248 * what movablecore would have allowed. 4249 */ 4250 if (required_movablecore) { 4251 unsigned long corepages; 4252 4253 /* 4254 * Round-up so that ZONE_MOVABLE is at least as large as what 4255 * was requested by the user 4256 */ 4257 required_movablecore = 4258 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4259 corepages = totalpages - required_movablecore; 4260 4261 required_kernelcore = max(required_kernelcore, corepages); 4262 } 4263 4264 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4265 if (!required_kernelcore) 4266 goto out; 4267 4268 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4269 find_usable_zone_for_movable(); 4270 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4271 4272restart: 4273 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4274 kernelcore_node = required_kernelcore / usable_nodes; 4275 for_each_node_state(nid, N_HIGH_MEMORY) { 4276 /* 4277 * Recalculate kernelcore_node if the division per node 4278 * now exceeds what is necessary to satisfy the requested 4279 * amount of memory for the kernel 4280 */ 4281 if (required_kernelcore < kernelcore_node) 4282 kernelcore_node = required_kernelcore / usable_nodes; 4283 4284 /* 4285 * As the map is walked, we track how much memory is usable 4286 * by the kernel using kernelcore_remaining. When it is 4287 * 0, the rest of the node is usable by ZONE_MOVABLE 4288 */ 4289 kernelcore_remaining = kernelcore_node; 4290 4291 /* Go through each range of PFNs within this node */ 4292 for_each_active_range_index_in_nid(i, nid) { 4293 unsigned long start_pfn, end_pfn; 4294 unsigned long size_pages; 4295 4296 start_pfn = max(early_node_map[i].start_pfn, 4297 zone_movable_pfn[nid]); 4298 end_pfn = early_node_map[i].end_pfn; 4299 if (start_pfn >= end_pfn) 4300 continue; 4301 4302 /* Account for what is only usable for kernelcore */ 4303 if (start_pfn < usable_startpfn) { 4304 unsigned long kernel_pages; 4305 kernel_pages = min(end_pfn, usable_startpfn) 4306 - start_pfn; 4307 4308 kernelcore_remaining -= min(kernel_pages, 4309 kernelcore_remaining); 4310 required_kernelcore -= min(kernel_pages, 4311 required_kernelcore); 4312 4313 /* Continue if range is now fully accounted */ 4314 if (end_pfn <= usable_startpfn) { 4315 4316 /* 4317 * Push zone_movable_pfn to the end so 4318 * that if we have to rebalance 4319 * kernelcore across nodes, we will 4320 * not double account here 4321 */ 4322 zone_movable_pfn[nid] = end_pfn; 4323 continue; 4324 } 4325 start_pfn = usable_startpfn; 4326 } 4327 4328 /* 4329 * The usable PFN range for ZONE_MOVABLE is from 4330 * start_pfn->end_pfn. Calculate size_pages as the 4331 * number of pages used as kernelcore 4332 */ 4333 size_pages = end_pfn - start_pfn; 4334 if (size_pages > kernelcore_remaining) 4335 size_pages = kernelcore_remaining; 4336 zone_movable_pfn[nid] = start_pfn + size_pages; 4337 4338 /* 4339 * Some kernelcore has been met, update counts and 4340 * break if the kernelcore for this node has been 4341 * satisified 4342 */ 4343 required_kernelcore -= min(required_kernelcore, 4344 size_pages); 4345 kernelcore_remaining -= size_pages; 4346 if (!kernelcore_remaining) 4347 break; 4348 } 4349 } 4350 4351 /* 4352 * If there is still required_kernelcore, we do another pass with one 4353 * less node in the count. This will push zone_movable_pfn[nid] further 4354 * along on the nodes that still have memory until kernelcore is 4355 * satisified 4356 */ 4357 usable_nodes--; 4358 if (usable_nodes && required_kernelcore > usable_nodes) 4359 goto restart; 4360 4361 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4362 for (nid = 0; nid < MAX_NUMNODES; nid++) 4363 zone_movable_pfn[nid] = 4364 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4365 4366out: 4367 /* restore the node_state */ 4368 node_states[N_HIGH_MEMORY] = saved_node_state; 4369} 4370 4371/* Any regular memory on that node ? */ 4372static void check_for_regular_memory(pg_data_t *pgdat) 4373{ 4374#ifdef CONFIG_HIGHMEM 4375 enum zone_type zone_type; 4376 4377 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4378 struct zone *zone = &pgdat->node_zones[zone_type]; 4379 if (zone->present_pages) 4380 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4381 } 4382#endif 4383} 4384 4385/** 4386 * free_area_init_nodes - Initialise all pg_data_t and zone data 4387 * @max_zone_pfn: an array of max PFNs for each zone 4388 * 4389 * This will call free_area_init_node() for each active node in the system. 4390 * Using the page ranges provided by add_active_range(), the size of each 4391 * zone in each node and their holes is calculated. If the maximum PFN 4392 * between two adjacent zones match, it is assumed that the zone is empty. 4393 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4394 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4395 * starts where the previous one ended. For example, ZONE_DMA32 starts 4396 * at arch_max_dma_pfn. 4397 */ 4398void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4399{ 4400 unsigned long nid; 4401 int i; 4402 4403 /* Sort early_node_map as initialisation assumes it is sorted */ 4404 sort_node_map(); 4405 4406 /* Record where the zone boundaries are */ 4407 memset(arch_zone_lowest_possible_pfn, 0, 4408 sizeof(arch_zone_lowest_possible_pfn)); 4409 memset(arch_zone_highest_possible_pfn, 0, 4410 sizeof(arch_zone_highest_possible_pfn)); 4411 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4412 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4413 for (i = 1; i < MAX_NR_ZONES; i++) { 4414 if (i == ZONE_MOVABLE) 4415 continue; 4416 arch_zone_lowest_possible_pfn[i] = 4417 arch_zone_highest_possible_pfn[i-1]; 4418 arch_zone_highest_possible_pfn[i] = 4419 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4420 } 4421 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4422 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4423 4424 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4425 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4426 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4427 4428 /* Print out the zone ranges */ 4429 printk("Zone PFN ranges:\n"); 4430 for (i = 0; i < MAX_NR_ZONES; i++) { 4431 if (i == ZONE_MOVABLE) 4432 continue; 4433 printk(" %-8s %0#10lx -> %0#10lx\n", 4434 zone_names[i], 4435 arch_zone_lowest_possible_pfn[i], 4436 arch_zone_highest_possible_pfn[i]); 4437 } 4438 4439 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4440 printk("Movable zone start PFN for each node\n"); 4441 for (i = 0; i < MAX_NUMNODES; i++) { 4442 if (zone_movable_pfn[i]) 4443 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4444 } 4445 4446 /* Print out the early_node_map[] */ 4447 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4448 for (i = 0; i < nr_nodemap_entries; i++) 4449 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4450 early_node_map[i].start_pfn, 4451 early_node_map[i].end_pfn); 4452 4453 /* Initialise every node */ 4454 mminit_verify_pageflags_layout(); 4455 setup_nr_node_ids(); 4456 for_each_online_node(nid) { 4457 pg_data_t *pgdat = NODE_DATA(nid); 4458 free_area_init_node(nid, NULL, 4459 find_min_pfn_for_node(nid), NULL); 4460 4461 /* Any memory on that node */ 4462 if (pgdat->node_present_pages) 4463 node_set_state(nid, N_HIGH_MEMORY); 4464 check_for_regular_memory(pgdat); 4465 } 4466} 4467 4468static int __init cmdline_parse_core(char *p, unsigned long *core) 4469{ 4470 unsigned long long coremem; 4471 if (!p) 4472 return -EINVAL; 4473 4474 coremem = memparse(p, &p); 4475 *core = coremem >> PAGE_SHIFT; 4476 4477 /* Paranoid check that UL is enough for the coremem value */ 4478 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4479 4480 return 0; 4481} 4482 4483/* 4484 * kernelcore=size sets the amount of memory for use for allocations that 4485 * cannot be reclaimed or migrated. 4486 */ 4487static int __init cmdline_parse_kernelcore(char *p) 4488{ 4489 return cmdline_parse_core(p, &required_kernelcore); 4490} 4491 4492/* 4493 * movablecore=size sets the amount of memory for use for allocations that 4494 * can be reclaimed or migrated. 4495 */ 4496static int __init cmdline_parse_movablecore(char *p) 4497{ 4498 return cmdline_parse_core(p, &required_movablecore); 4499} 4500 4501early_param("kernelcore", cmdline_parse_kernelcore); 4502early_param("movablecore", cmdline_parse_movablecore); 4503 4504#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4505 4506/** 4507 * set_dma_reserve - set the specified number of pages reserved in the first zone 4508 * @new_dma_reserve: The number of pages to mark reserved 4509 * 4510 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4511 * In the DMA zone, a significant percentage may be consumed by kernel image 4512 * and other unfreeable allocations which can skew the watermarks badly. This 4513 * function may optionally be used to account for unfreeable pages in the 4514 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4515 * smaller per-cpu batchsize. 4516 */ 4517void __init set_dma_reserve(unsigned long new_dma_reserve) 4518{ 4519 dma_reserve = new_dma_reserve; 4520} 4521 4522#ifndef CONFIG_NEED_MULTIPLE_NODES 4523struct pglist_data __refdata contig_page_data = { 4524#ifndef CONFIG_NO_BOOTMEM 4525 .bdata = &bootmem_node_data[0] 4526#endif 4527 }; 4528EXPORT_SYMBOL(contig_page_data); 4529#endif 4530 4531void __init free_area_init(unsigned long *zones_size) 4532{ 4533 free_area_init_node(0, zones_size, 4534 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4535} 4536 4537static int page_alloc_cpu_notify(struct notifier_block *self, 4538 unsigned long action, void *hcpu) 4539{ 4540 int cpu = (unsigned long)hcpu; 4541 4542 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4543 drain_pages(cpu); 4544 4545 /* 4546 * Spill the event counters of the dead processor 4547 * into the current processors event counters. 4548 * This artificially elevates the count of the current 4549 * processor. 4550 */ 4551 vm_events_fold_cpu(cpu); 4552 4553 /* 4554 * Zero the differential counters of the dead processor 4555 * so that the vm statistics are consistent. 4556 * 4557 * This is only okay since the processor is dead and cannot 4558 * race with what we are doing. 4559 */ 4560 refresh_cpu_vm_stats(cpu); 4561 } 4562 return NOTIFY_OK; 4563} 4564 4565void __init page_alloc_init(void) 4566{ 4567 hotcpu_notifier(page_alloc_cpu_notify, 0); 4568} 4569 4570/* 4571 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4572 * or min_free_kbytes changes. 4573 */ 4574static void calculate_totalreserve_pages(void) 4575{ 4576 struct pglist_data *pgdat; 4577 unsigned long reserve_pages = 0; 4578 enum zone_type i, j; 4579 4580 for_each_online_pgdat(pgdat) { 4581 for (i = 0; i < MAX_NR_ZONES; i++) { 4582 struct zone *zone = pgdat->node_zones + i; 4583 unsigned long max = 0; 4584 4585 /* Find valid and maximum lowmem_reserve in the zone */ 4586 for (j = i; j < MAX_NR_ZONES; j++) { 4587 if (zone->lowmem_reserve[j] > max) 4588 max = zone->lowmem_reserve[j]; 4589 } 4590 4591 /* we treat the high watermark as reserved pages. */ 4592 max += high_wmark_pages(zone); 4593 4594 if (max > zone->present_pages) 4595 max = zone->present_pages; 4596 reserve_pages += max; 4597 } 4598 } 4599 totalreserve_pages = reserve_pages; 4600} 4601 4602/* 4603 * setup_per_zone_lowmem_reserve - called whenever 4604 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4605 * has a correct pages reserved value, so an adequate number of 4606 * pages are left in the zone after a successful __alloc_pages(). 4607 */ 4608static void setup_per_zone_lowmem_reserve(void) 4609{ 4610 struct pglist_data *pgdat; 4611 enum zone_type j, idx; 4612 4613 for_each_online_pgdat(pgdat) { 4614 for (j = 0; j < MAX_NR_ZONES; j++) { 4615 struct zone *zone = pgdat->node_zones + j; 4616 unsigned long present_pages = zone->present_pages; 4617 4618 zone->lowmem_reserve[j] = 0; 4619 4620 idx = j; 4621 while (idx) { 4622 struct zone *lower_zone; 4623 4624 idx--; 4625 4626 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4627 sysctl_lowmem_reserve_ratio[idx] = 1; 4628 4629 lower_zone = pgdat->node_zones + idx; 4630 lower_zone->lowmem_reserve[j] = present_pages / 4631 sysctl_lowmem_reserve_ratio[idx]; 4632 present_pages += lower_zone->present_pages; 4633 } 4634 } 4635 } 4636 4637 /* update totalreserve_pages */ 4638 calculate_totalreserve_pages(); 4639} 4640 4641/** 4642 * setup_per_zone_wmarks - called when min_free_kbytes changes 4643 * or when memory is hot-{added|removed} 4644 * 4645 * Ensures that the watermark[min,low,high] values for each zone are set 4646 * correctly with respect to min_free_kbytes. 4647 */ 4648void setup_per_zone_wmarks(void) 4649{ 4650 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4651 unsigned long lowmem_pages = 0; 4652 struct zone *zone; 4653 unsigned long flags; 4654 4655 /* Calculate total number of !ZONE_HIGHMEM pages */ 4656 for_each_zone(zone) { 4657 if (!is_highmem(zone)) 4658 lowmem_pages += zone->present_pages; 4659 } 4660 4661 for_each_zone(zone) { 4662 u64 tmp; 4663 4664 spin_lock_irqsave(&zone->lock, flags); 4665 tmp = (u64)pages_min * zone->present_pages; 4666 do_div(tmp, lowmem_pages); 4667 if (is_highmem(zone)) { 4668 /* 4669 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4670 * need highmem pages, so cap pages_min to a small 4671 * value here. 4672 * 4673 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4674 * deltas controls asynch page reclaim, and so should 4675 * not be capped for highmem. 4676 */ 4677 int min_pages; 4678 4679 min_pages = zone->present_pages / 1024; 4680 if (min_pages < SWAP_CLUSTER_MAX) 4681 min_pages = SWAP_CLUSTER_MAX; 4682 if (min_pages > 128) 4683 min_pages = 128; 4684 zone->watermark[WMARK_MIN] = min_pages; 4685 } else { 4686 /* 4687 * If it's a lowmem zone, reserve a number of pages 4688 * proportionate to the zone's size. 4689 */ 4690 zone->watermark[WMARK_MIN] = tmp; 4691 } 4692 4693 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4694 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4695 setup_zone_migrate_reserve(zone); 4696 spin_unlock_irqrestore(&zone->lock, flags); 4697 } 4698 4699 /* update totalreserve_pages */ 4700 calculate_totalreserve_pages(); 4701} 4702 4703/* 4704 * The inactive anon list should be small enough that the VM never has to 4705 * do too much work, but large enough that each inactive page has a chance 4706 * to be referenced again before it is swapped out. 4707 * 4708 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4709 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4710 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4711 * the anonymous pages are kept on the inactive list. 4712 * 4713 * total target max 4714 * memory ratio inactive anon 4715 * ------------------------------------- 4716 * 10MB 1 5MB 4717 * 100MB 1 50MB 4718 * 1GB 3 250MB 4719 * 10GB 10 0.9GB 4720 * 100GB 31 3GB 4721 * 1TB 101 10GB 4722 * 10TB 320 32GB 4723 */ 4724void calculate_zone_inactive_ratio(struct zone *zone) 4725{ 4726 unsigned int gb, ratio; 4727 4728 /* Zone size in gigabytes */ 4729 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4730 if (gb) 4731 ratio = int_sqrt(10 * gb); 4732 else 4733 ratio = 1; 4734 4735 zone->inactive_ratio = ratio; 4736} 4737 4738static void __init setup_per_zone_inactive_ratio(void) 4739{ 4740 struct zone *zone; 4741 4742 for_each_zone(zone) 4743 calculate_zone_inactive_ratio(zone); 4744} 4745 4746/* 4747 * Initialise min_free_kbytes. 4748 * 4749 * For small machines we want it small (128k min). For large machines 4750 * we want it large (64MB max). But it is not linear, because network 4751 * bandwidth does not increase linearly with machine size. We use 4752 * 4753 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4754 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4755 * 4756 * which yields 4757 * 4758 * 16MB: 512k 4759 * 32MB: 724k 4760 * 64MB: 1024k 4761 * 128MB: 1448k 4762 * 256MB: 2048k 4763 * 512MB: 2896k 4764 * 1024MB: 4096k 4765 * 2048MB: 5792k 4766 * 4096MB: 8192k 4767 * 8192MB: 11584k 4768 * 16384MB: 16384k 4769 */ 4770static int __init init_per_zone_wmark_min(void) 4771{ 4772 unsigned long lowmem_kbytes; 4773 4774 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4775 4776 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4777 if (min_free_kbytes < 128) 4778 min_free_kbytes = 128; 4779 if (min_free_kbytes > 65536) 4780 min_free_kbytes = 65536; 4781 setup_per_zone_wmarks(); 4782 setup_per_zone_lowmem_reserve(); 4783 setup_per_zone_inactive_ratio(); 4784 return 0; 4785} 4786module_init(init_per_zone_wmark_min) 4787 4788/* 4789 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4790 * that we can call two helper functions whenever min_free_kbytes 4791 * changes. 4792 */ 4793int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4794 void __user *buffer, size_t *length, loff_t *ppos) 4795{ 4796 proc_dointvec(table, write, buffer, length, ppos); 4797 if (write) 4798 setup_per_zone_wmarks(); 4799 return 0; 4800} 4801 4802#ifdef CONFIG_NUMA 4803int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4804 void __user *buffer, size_t *length, loff_t *ppos) 4805{ 4806 struct zone *zone; 4807 int rc; 4808 4809 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4810 if (rc) 4811 return rc; 4812 4813 for_each_zone(zone) 4814 zone->min_unmapped_pages = (zone->present_pages * 4815 sysctl_min_unmapped_ratio) / 100; 4816 return 0; 4817} 4818 4819int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4820 void __user *buffer, size_t *length, loff_t *ppos) 4821{ 4822 struct zone *zone; 4823 int rc; 4824 4825 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4826 if (rc) 4827 return rc; 4828 4829 for_each_zone(zone) 4830 zone->min_slab_pages = (zone->present_pages * 4831 sysctl_min_slab_ratio) / 100; 4832 return 0; 4833} 4834#endif 4835 4836/* 4837 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4838 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4839 * whenever sysctl_lowmem_reserve_ratio changes. 4840 * 4841 * The reserve ratio obviously has absolutely no relation with the 4842 * minimum watermarks. The lowmem reserve ratio can only make sense 4843 * if in function of the boot time zone sizes. 4844 */ 4845int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4846 void __user *buffer, size_t *length, loff_t *ppos) 4847{ 4848 proc_dointvec_minmax(table, write, buffer, length, ppos); 4849 setup_per_zone_lowmem_reserve(); 4850 return 0; 4851} 4852 4853/* 4854 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4855 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4856 * can have before it gets flushed back to buddy allocator. 4857 */ 4858 4859int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4860 void __user *buffer, size_t *length, loff_t *ppos) 4861{ 4862 struct zone *zone; 4863 unsigned int cpu; 4864 int ret; 4865 4866 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 4867 if (!write || (ret == -EINVAL)) 4868 return ret; 4869 for_each_populated_zone(zone) { 4870 for_each_online_cpu(cpu) { 4871 unsigned long high; 4872 high = zone->present_pages / percpu_pagelist_fraction; 4873 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4874 } 4875 } 4876 return 0; 4877} 4878 4879int hashdist = HASHDIST_DEFAULT; 4880 4881#ifdef CONFIG_NUMA 4882static int __init set_hashdist(char *str) 4883{ 4884 if (!str) 4885 return 0; 4886 hashdist = simple_strtoul(str, &str, 0); 4887 return 1; 4888} 4889__setup("hashdist=", set_hashdist); 4890#endif 4891 4892/* 4893 * allocate a large system hash table from bootmem 4894 * - it is assumed that the hash table must contain an exact power-of-2 4895 * quantity of entries 4896 * - limit is the number of hash buckets, not the total allocation size 4897 */ 4898void *__init alloc_large_system_hash(const char *tablename, 4899 unsigned long bucketsize, 4900 unsigned long numentries, 4901 int scale, 4902 int flags, 4903 unsigned int *_hash_shift, 4904 unsigned int *_hash_mask, 4905 unsigned long limit) 4906{ 4907 unsigned long long max = limit; 4908 unsigned long log2qty, size; 4909 void *table = NULL; 4910 4911 /* allow the kernel cmdline to have a say */ 4912 if (!numentries) { 4913 /* round applicable memory size up to nearest megabyte */ 4914 numentries = nr_kernel_pages; 4915 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4916 numentries >>= 20 - PAGE_SHIFT; 4917 numentries <<= 20 - PAGE_SHIFT; 4918 4919 /* limit to 1 bucket per 2^scale bytes of low memory */ 4920 if (scale > PAGE_SHIFT) 4921 numentries >>= (scale - PAGE_SHIFT); 4922 else 4923 numentries <<= (PAGE_SHIFT - scale); 4924 4925 /* Make sure we've got at least a 0-order allocation.. */ 4926 if (unlikely(flags & HASH_SMALL)) { 4927 /* Makes no sense without HASH_EARLY */ 4928 WARN_ON(!(flags & HASH_EARLY)); 4929 if (!(numentries >> *_hash_shift)) { 4930 numentries = 1UL << *_hash_shift; 4931 BUG_ON(!numentries); 4932 } 4933 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4934 numentries = PAGE_SIZE / bucketsize; 4935 } 4936 numentries = roundup_pow_of_two(numentries); 4937 4938 /* limit allocation size to 1/16 total memory by default */ 4939 if (max == 0) { 4940 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4941 do_div(max, bucketsize); 4942 } 4943 4944 if (numentries > max) 4945 numentries = max; 4946 4947 log2qty = ilog2(numentries); 4948 4949 do { 4950 size = bucketsize << log2qty; 4951 if (flags & HASH_EARLY) 4952 table = alloc_bootmem_nopanic(size); 4953 else if (hashdist) 4954 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4955 else { 4956 /* 4957 * If bucketsize is not a power-of-two, we may free 4958 * some pages at the end of hash table which 4959 * alloc_pages_exact() automatically does 4960 */ 4961 if (get_order(size) < MAX_ORDER) { 4962 table = alloc_pages_exact(size, GFP_ATOMIC); 4963 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 4964 } 4965 } 4966 } while (!table && size > PAGE_SIZE && --log2qty); 4967 4968 if (!table) 4969 panic("Failed to allocate %s hash table\n", tablename); 4970 4971 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4972 tablename, 4973 (1U << log2qty), 4974 ilog2(size) - PAGE_SHIFT, 4975 size); 4976 4977 if (_hash_shift) 4978 *_hash_shift = log2qty; 4979 if (_hash_mask) 4980 *_hash_mask = (1 << log2qty) - 1; 4981 4982 return table; 4983} 4984 4985/* Return a pointer to the bitmap storing bits affecting a block of pages */ 4986static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4987 unsigned long pfn) 4988{ 4989#ifdef CONFIG_SPARSEMEM 4990 return __pfn_to_section(pfn)->pageblock_flags; 4991#else 4992 return zone->pageblock_flags; 4993#endif /* CONFIG_SPARSEMEM */ 4994} 4995 4996static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4997{ 4998#ifdef CONFIG_SPARSEMEM 4999 pfn &= (PAGES_PER_SECTION-1); 5000 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5001#else 5002 pfn = pfn - zone->zone_start_pfn; 5003 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5004#endif /* CONFIG_SPARSEMEM */ 5005} 5006 5007/** 5008 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5009 * @page: The page within the block of interest 5010 * @start_bitidx: The first bit of interest to retrieve 5011 * @end_bitidx: The last bit of interest 5012 * returns pageblock_bits flags 5013 */ 5014unsigned long get_pageblock_flags_group(struct page *page, 5015 int start_bitidx, int end_bitidx) 5016{ 5017 struct zone *zone; 5018 unsigned long *bitmap; 5019 unsigned long pfn, bitidx; 5020 unsigned long flags = 0; 5021 unsigned long value = 1; 5022 5023 zone = page_zone(page); 5024 pfn = page_to_pfn(page); 5025 bitmap = get_pageblock_bitmap(zone, pfn); 5026 bitidx = pfn_to_bitidx(zone, pfn); 5027 5028 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5029 if (test_bit(bitidx + start_bitidx, bitmap)) 5030 flags |= value; 5031 5032 return flags; 5033} 5034 5035/** 5036 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5037 * @page: The page within the block of interest 5038 * @start_bitidx: The first bit of interest 5039 * @end_bitidx: The last bit of interest 5040 * @flags: The flags to set 5041 */ 5042void set_pageblock_flags_group(struct page *page, unsigned long flags, 5043 int start_bitidx, int end_bitidx) 5044{ 5045 struct zone *zone; 5046 unsigned long *bitmap; 5047 unsigned long pfn, bitidx; 5048 unsigned long value = 1; 5049 5050 zone = page_zone(page); 5051 pfn = page_to_pfn(page); 5052 bitmap = get_pageblock_bitmap(zone, pfn); 5053 bitidx = pfn_to_bitidx(zone, pfn); 5054 VM_BUG_ON(pfn < zone->zone_start_pfn); 5055 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5056 5057 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5058 if (flags & value) 5059 __set_bit(bitidx + start_bitidx, bitmap); 5060 else 5061 __clear_bit(bitidx + start_bitidx, bitmap); 5062} 5063 5064/* 5065 * This is designed as sub function...plz see page_isolation.c also. 5066 * set/clear page block's type to be ISOLATE. 5067 * page allocater never alloc memory from ISOLATE block. 5068 */ 5069 5070int set_migratetype_isolate(struct page *page) 5071{ 5072 struct zone *zone; 5073 struct page *curr_page; 5074 unsigned long flags, pfn, iter; 5075 unsigned long immobile = 0; 5076 struct memory_isolate_notify arg; 5077 int notifier_ret; 5078 int ret = -EBUSY; 5079 int zone_idx; 5080 5081 zone = page_zone(page); 5082 zone_idx = zone_idx(zone); 5083 5084 spin_lock_irqsave(&zone->lock, flags); 5085 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE || 5086 zone_idx == ZONE_MOVABLE) { 5087 ret = 0; 5088 goto out; 5089 } 5090 5091 pfn = page_to_pfn(page); 5092 arg.start_pfn = pfn; 5093 arg.nr_pages = pageblock_nr_pages; 5094 arg.pages_found = 0; 5095 5096 /* 5097 * It may be possible to isolate a pageblock even if the 5098 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5099 * notifier chain is used by balloon drivers to return the 5100 * number of pages in a range that are held by the balloon 5101 * driver to shrink memory. If all the pages are accounted for 5102 * by balloons, are free, or on the LRU, isolation can continue. 5103 * Later, for example, when memory hotplug notifier runs, these 5104 * pages reported as "can be isolated" should be isolated(freed) 5105 * by the balloon driver through the memory notifier chain. 5106 */ 5107 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5108 notifier_ret = notifier_to_errno(notifier_ret); 5109 if (notifier_ret || !arg.pages_found) 5110 goto out; 5111 5112 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) { 5113 if (!pfn_valid_within(pfn)) 5114 continue; 5115 5116 curr_page = pfn_to_page(iter); 5117 if (!page_count(curr_page) || PageLRU(curr_page)) 5118 continue; 5119 5120 immobile++; 5121 } 5122 5123 if (arg.pages_found == immobile) 5124 ret = 0; 5125 5126out: 5127 if (!ret) { 5128 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5129 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5130 } 5131 5132 spin_unlock_irqrestore(&zone->lock, flags); 5133 if (!ret) 5134 drain_all_pages(); 5135 return ret; 5136} 5137 5138void unset_migratetype_isolate(struct page *page) 5139{ 5140 struct zone *zone; 5141 unsigned long flags; 5142 zone = page_zone(page); 5143 spin_lock_irqsave(&zone->lock, flags); 5144 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5145 goto out; 5146 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5147 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5148out: 5149 spin_unlock_irqrestore(&zone->lock, flags); 5150} 5151 5152#ifdef CONFIG_MEMORY_HOTREMOVE 5153/* 5154 * All pages in the range must be isolated before calling this. 5155 */ 5156void 5157__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5158{ 5159 struct page *page; 5160 struct zone *zone; 5161 int order, i; 5162 unsigned long pfn; 5163 unsigned long flags; 5164 /* find the first valid pfn */ 5165 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5166 if (pfn_valid(pfn)) 5167 break; 5168 if (pfn == end_pfn) 5169 return; 5170 zone = page_zone(pfn_to_page(pfn)); 5171 spin_lock_irqsave(&zone->lock, flags); 5172 pfn = start_pfn; 5173 while (pfn < end_pfn) { 5174 if (!pfn_valid(pfn)) { 5175 pfn++; 5176 continue; 5177 } 5178 page = pfn_to_page(pfn); 5179 BUG_ON(page_count(page)); 5180 BUG_ON(!PageBuddy(page)); 5181 order = page_order(page); 5182#ifdef CONFIG_DEBUG_VM 5183 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5184 pfn, 1 << order, end_pfn); 5185#endif 5186 list_del(&page->lru); 5187 rmv_page_order(page); 5188 zone->free_area[order].nr_free--; 5189 __mod_zone_page_state(zone, NR_FREE_PAGES, 5190 - (1UL << order)); 5191 for (i = 0; i < (1 << order); i++) 5192 SetPageReserved((page+i)); 5193 pfn += (1 << order); 5194 } 5195 spin_unlock_irqrestore(&zone->lock, flags); 5196} 5197#endif 5198 5199#ifdef CONFIG_MEMORY_FAILURE 5200bool is_free_buddy_page(struct page *page) 5201{ 5202 struct zone *zone = page_zone(page); 5203 unsigned long pfn = page_to_pfn(page); 5204 unsigned long flags; 5205 int order; 5206 5207 spin_lock_irqsave(&zone->lock, flags); 5208 for (order = 0; order < MAX_ORDER; order++) { 5209 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5210 5211 if (PageBuddy(page_head) && page_order(page_head) >= order) 5212 break; 5213 } 5214 spin_unlock_irqrestore(&zone->lock, flags); 5215 5216 return order < MAX_ORDER; 5217} 5218#endif 5219