mm/page_alloc.c

   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/config.h>
  18 #include <linux/stddef.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/interrupt.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/compiler.h>
  25 #include <linux/module.h>
  26 #include <linux/suspend.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/slab.h>
  30 #include <linux/notifier.h>
  31 #include <linux/topology.h>
  32 #include <linux/sysctl.h>
  33 #include <linux/cpu.h>
  34 #include <linux/cpuset.h>
  35 #include <linux/nodemask.h>
  36 #include <linux/vmalloc.h>
  37
  38 #include <asm/tlbflush.h>
  39 #include "internal.h"
  40
  41 /*
  42  * MCD - HACK: Find somewhere to initialize this EARLY, or make this
  43  * initializer cleaner
  44  */
  45 nodemask_t node_online_map = { { [0] = 1UL } };
  46 EXPORT_SYMBOL(node_online_map);
  47 nodemask_t node_possible_map = NODE_MASK_ALL;
  48 EXPORT_SYMBOL(node_possible_map);
  49 struct pglist_data *pgdat_list;
  50 unsigned long totalram_pages;
  51 unsigned long totalhigh_pages;
  52 long nr_swap_pages;
  53
  54 /*
  55  * results with 256, 32 in the lowmem_reserve sysctl:
  56  *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
  57  *      1G machine -> (16M dma, 784M normal, 224M high)
  58  *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
  59  *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
  60  *      HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
  61  */
  62 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 256, 32 };
  63
  64 EXPORT_SYMBOL(totalram_pages);
  65 EXPORT_SYMBOL(nr_swap_pages);
  66
  67 /*
  68  * Used by page_zone() to look up the address of the struct zone whose
  69  * id is encoded in the upper bits of page->flags
  70  */
  71 struct zone *zone_table[1 << (ZONES_SHIFT + NODES_SHIFT)];
  72 EXPORT_SYMBOL(zone_table);
  73
  74 static char *zone_names[MAX_NR_ZONES] = { "DMA", "Normal", "HighMem" };
  75 int min_free_kbytes = 1024;
  76
  77 unsigned long __initdata nr_kernel_pages;
  78 unsigned long __initdata nr_all_pages;
  79
  80 /*
  81  * Temporary debugging check for pages not lying within a given zone.
  82  */
  83 static int bad_range(struct zone *zone, struct page *page)
  84 {
  85         if (page_to_pfn(page) >= zone->zone_start_pfn + zone->spanned_pages)
  86                 return 1;
  87         if (page_to_pfn(page) < zone->zone_start_pfn)
  88                 return 1;
  89 #ifdef CONFIG_HOLES_IN_ZONE
  90         if (!pfn_valid(page_to_pfn(page)))
  91                 return 1;
  92 #endif
  93         if (zone != page_zone(page))
  94                 return 1;
  95         return 0;
  96 }
  97
  98 static void bad_page(const char *function, struct page *page)
  99 {
 100         printk(KERN_EMERG "Bad page state at %s (in process '%s', page %p)\n",
 101                 function, current->comm, page);
 102         printk(KERN_EMERG "flags:0x%0*lx mapping:%p mapcount:%d count:%d\n",
 103                 (int)(2*sizeof(page_flags_t)), (unsigned long)page->flags,
 104                 page->mapping, page_mapcount(page), page_count(page));
 105         printk(KERN_EMERG "Backtrace:\n");
 106         dump_stack();
 107         printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n");
 108         page->flags &= ~(1 << PG_private        |
 109                         1 << PG_locked  |
 110                         1 << PG_lru     |
 111                         1 << PG_active  |
 112                         1 << PG_dirty   |
 113                         1 << PG_swapcache |
 114                         1 << PG_writeback);
 115         set_page_count(page, 0);
 116         reset_page_mapcount(page);
 117         page->mapping = NULL;
 118         tainted |= TAINT_BAD_PAGE;
 119 }
 120
 121 #ifndef CONFIG_HUGETLB_PAGE
 122 #define prep_compound_page(page, order) do { } while (0)
 123 #define destroy_compound_page(page, order) do { } while (0)
 124 #else
 125 /*
 126  * Higher-order pages are called "compound pages".  They are structured thusly:
 127  *
 128  * The first PAGE_SIZE page is called the "head page".
 129  *
 130  * The remaining PAGE_SIZE pages are called "tail pages".
 131  *
 132  * All pages have PG_compound set.  All pages have their ->private pointing at
 133  * the head page (even the head page has this).
 134  *
 135  * The first tail page's ->mapping, if non-zero, holds the address of the
 136  * compound page's put_page() function.
 137  *
 138  * The order of the allocation is stored in the first tail page's ->index
 139  * This is only for debug at present.  This usage means that zero-order pages
 140  * may not be compound.
 141  */
 142 static void prep_compound_page(struct page *page, unsigned long order)
 143 {
 144         int i;
 145         int nr_pages = 1 << order;
 146
 147         page[1].mapping = NULL;
 148         page[1].index = order;
 149         for (i = 0; i < nr_pages; i++) {
 150                 struct page *p = page + i;
 151
 152                 SetPageCompound(p);
 153                 p->private = (unsigned long)page;
 154         }
 155 }
 156
 157 static void destroy_compound_page(struct page *page, unsigned long order)
 158 {
 159         int i;
 160         int nr_pages = 1 << order;
 161
 162         if (!PageCompound(page))
 163                 return;
 164
 165         if (page[1].index != order)
 166                 bad_page(__FUNCTION__, page);
 167
 168         for (i = 0; i < nr_pages; i++) {
 169                 struct page *p = page + i;
 170
 171                 if (!PageCompound(p))
 172                         bad_page(__FUNCTION__, page);
 173                 if (p->private != (unsigned long)page)
 174                         bad_page(__FUNCTION__, page);
 175                 ClearPageCompound(p);
 176         }
 177 }
 178 #endif          /* CONFIG_HUGETLB_PAGE */
 179
 180 /*
 181  * function for dealing with page's order in buddy system.
 182  * zone->lock is already acquired when we use these.
 183  * So, we don't need atomic page->flags operations here.
 184  */
 185 static inline unsigned long page_order(struct page *page) {
 186         return page->private;
 187 }
 188
 189 static inline void set_page_order(struct page *page, int order) {
 190         page->private = order;
 191         __SetPagePrivate(page);
 192 }
 193
 194 static inline void rmv_page_order(struct page *page)
 195 {
 196         __ClearPagePrivate(page);
 197         page->private = 0;
 198 }
 199
 200 /*
 201  * Locate the struct page for both the matching buddy in our
 202  * pair (buddy1) and the combined O(n+1) page they form (page).
 203  *
 204  * 1) Any buddy B1 will have an order O twin B2 which satisfies
 205  * the following equation:
 206  *     B2 = B1 ^ (1 << O)
 207  * For example, if the starting buddy (buddy2) is #8 its order
 208  * 1 buddy is #10:
 209  *     B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
 210  *
 211  * 2) Any buddy B will have an order O+1 parent P which
 212  * satisfies the following equation:
 213  *     P = B & ~(1 << O)
 214  *
 215  * Assumption: *_mem_map is contigious at least up to MAX_ORDER
 216  */
 217 static inline struct page *
 218 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order)
 219 {
 220         unsigned long buddy_idx = page_idx ^ (1 << order);
 221
 222         return page + (buddy_idx - page_idx);
 223 }
 224
 225 static inline unsigned long
 226 __find_combined_index(unsigned long page_idx, unsigned int order)
 227 {
 228         return (page_idx & ~(1 << order));
 229 }
 230
 231 /*
 232  * This function checks whether a page is free && is the buddy
 233  * we can do coalesce a page and its buddy if
 234  * (a) the buddy is free &&
 235  * (b) the buddy is on the buddy system &&
 236  * (c) a page and its buddy have the same order.
 237  * for recording page's order, we use page->private and PG_private.
 238  *
 239  */
 240 static inline int page_is_buddy(struct page *page, int order)
 241 {
 242        if (PagePrivate(page)           &&
 243            (page_order(page) == order) &&
 244            !PageReserved(page)         &&
 245             page_count(page) == 0)
 246                return 1;
 247        return 0;
 248 }
 249
 250 /*
 251  * Freeing function for a buddy system allocator.
 252  *
 253  * The concept of a buddy system is to maintain direct-mapped table
 254  * (containing bit values) for memory blocks of various "orders".
 255  * The bottom level table contains the map for the smallest allocatable
 256  * units of memory (here, pages), and each level above it describes
 257  * pairs of units from the levels below, hence, "buddies".
 258  * At a high level, all that happens here is marking the table entry
 259  * at the bottom level available, and propagating the changes upward
 260  * as necessary, plus some accounting needed to play nicely with other
 261  * parts of the VM system.
 262  * At each level, we keep a list of pages, which are heads of continuous
 263  * free pages of length of (1 << order) and marked with PG_Private.Page's
 264  * order is recorded in page->private field.
 265  * So when we are allocating or freeing one, we can derive the state of the
 266  * other.  That is, if we allocate a small block, and both were
 267  * free, the remainder of the region must be split into blocks.
 268  * If a block is freed, and its buddy is also free, then this
 269  * triggers coalescing into a block of larger size.
 270  *
 271  * -- wli
 272  */
 273
 274 static inline void __free_pages_bulk (struct page *page,
 275                 struct zone *zone, unsigned int order)
 276 {
 277         unsigned long page_idx;
 278         int order_size = 1 << order;
 279
 280         if (unlikely(order))
 281                 destroy_compound_page(page, order);
 282
 283         page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
 284
 285         BUG_ON(page_idx & (order_size - 1));
 286         BUG_ON(bad_range(zone, page));
 287
 288         zone->free_pages += order_size;
 289         while (order < MAX_ORDER-1) {
 290                 unsigned long combined_idx;
 291                 struct free_area *area;
 292                 struct page *buddy;
 293
 294                 combined_idx = __find_combined_index(page_idx, order);
 295                 buddy = __page_find_buddy(page, page_idx, order);
 296
 297                 if (bad_range(zone, buddy))
 298                         break;
 299                 if (!page_is_buddy(buddy, order))
 300                         break;          /* Move the buddy up one level. */
 301                 list_del(&buddy->lru);
 302                 area = zone->free_area + order;
 303                 area->nr_free--;
 304                 rmv_page_order(buddy);
 305                 page = page + (combined_idx - page_idx);
 306                 page_idx = combined_idx;
 307                 order++;
 308         }
 309         set_page_order(page, order);
 310         list_add(&page->lru, &zone->free_area[order].free_list);
 311         zone->free_area[order].nr_free++;
 312 }
 313
 314 static inline void free_pages_check(const char *function, struct page *page)
 315 {
 316         if (    page_mapcount(page) ||
 317                 page->mapping != NULL ||
 318                 page_count(page) != 0 ||
 319                 (page->flags & (
 320                         1 << PG_lru     |
 321                         1 << PG_private |
 322                         1 << PG_locked  |
 323                         1 << PG_active  |
 324                         1 << PG_reclaim |
 325                         1 << PG_slab    |
 326                         1 << PG_swapcache |
 327                         1 << PG_writeback )))
 328                 bad_page(function, page);
 329         if (PageDirty(page))
 330                 ClearPageDirty(page);
 331 }
 332
 333 /*
 334  * Frees a list of pages.
 335  * Assumes all pages on list are in same zone, and of same order.
 336  * count is the number of pages to free, or 0 for all on the list.
 337  *
 338  * If the zone was previously in an "all pages pinned" state then look to
 339  * see if this freeing clears that state.
 340  *
 341  * And clear the zone's pages_scanned counter, to hold off the "all pages are
 342  * pinned" detection logic.
 343  */
 344 static int
 345 free_pages_bulk(struct zone *zone, int count,
 346                 struct list_head *list, unsigned int order)
 347 {
 348         unsigned long flags;
 349         struct page *page = NULL;
 350         int ret = 0;
 351
 352         spin_lock_irqsave(&zone->lock, flags);
 353         zone->all_unreclaimable = 0;
 354         zone->pages_scanned = 0;
 355         while (!list_empty(list) && count--) {
 356                 page = list_entry(list->prev, struct page, lru);
 357                 /* have to delete it as __free_pages_bulk list manipulates */
 358                 list_del(&page->lru);
 359                 __free_pages_bulk(page, zone, order);
 360                 ret++;
 361         }
 362         spin_unlock_irqrestore(&zone->lock, flags);
 363         return ret;
 364 }
 365
 366 void __free_pages_ok(struct page *page, unsigned int order)
 367 {
 368         LIST_HEAD(list);
 369         int i;
 370
 371         arch_free_page(page, order);
 372
 373         mod_page_state(pgfree, 1 << order);
 374
 375 #ifndef CONFIG_MMU
 376         if (order > 0)
 377                 for (i = 1 ; i < (1 << order) ; ++i)
 378                         __put_page(page + i);
 379 #endif
 380
 381         for (i = 0 ; i < (1 << order) ; ++i)
 382                 free_pages_check(__FUNCTION__, page + i);
 383         list_add(&page->lru, &list);
 384         kernel_map_pages(page, 1<<order, 0);
 385         free_pages_bulk(page_zone(page), 1, &list, order);
 386 }
 387
 388
 389 /*
 390  * The order of subdivision here is critical for the IO subsystem.
 391  * Please do not alter this order without good reasons and regression
 392  * testing. Specifically, as large blocks of memory are subdivided,
 393  * the order in which smaller blocks are delivered depends on the order
 394  * they're subdivided in this function. This is the primary factor
 395  * influencing the order in which pages are delivered to the IO
 396  * subsystem according to empirical testing, and this is also justified
 397  * by considering the behavior of a buddy system containing a single
 398  * large block of memory acted on by a series of small allocations.
 399  * This behavior is a critical factor in sglist merging's success.
 400  *
 401  * -- wli
 402  */
 403 static inline struct page *
 404 expand(struct zone *zone, struct page *page,
 405         int low, int high, struct free_area *area)
 406 {
 407         unsigned long size = 1 << high;
 408
 409         while (high > low) {
 410                 area--;
 411                 high--;
 412                 size >>= 1;
 413                 BUG_ON(bad_range(zone, &page[size]));
 414                 list_add(&page[size].lru, &area->free_list);
 415                 area->nr_free++;
 416                 set_page_order(&page[size], high);
 417         }
 418         return page;
 419 }
 420
 421 void set_page_refs(struct page *page, int order)
 422 {
 423 #ifdef CONFIG_MMU
 424         set_page_count(page, 1);
 425 #else
 426         int i;
 427
 428         /*
 429          * We need to reference all the pages for this order, otherwise if
 430          * anyone accesses one of the pages with (get/put) it will be freed.
 431          * - eg: access_process_vm()
 432          */
 433         for (i = 0; i < (1 << order); i++)
 434                 set_page_count(page + i, 1);
 435 #endif /* CONFIG_MMU */
 436 }
 437
 438 /*
 439  * This page is about to be returned from the page allocator
 440  */
 441 static void prep_new_page(struct page *page, int order)
 442 {
 443         if (page->mapping || page_mapcount(page) ||
 444             (page->flags & (
 445                         1 << PG_private |
 446                         1 << PG_locked  |
 447                         1 << PG_lru     |
 448                         1 << PG_active  |
 449                         1 << PG_dirty   |
 450                         1 << PG_reclaim |
 451                         1 << PG_swapcache |
 452                         1 << PG_writeback )))
 453                 bad_page(__FUNCTION__, page);
 454
 455         page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
 456                         1 << PG_referenced | 1 << PG_arch_1 |
 457                         1 << PG_checked | 1 << PG_mappedtodisk);
 458         page->private = 0;
 459         set_page_refs(page, order);
 460         kernel_map_pages(page, 1 << order, 1);
 461 }
 462
 463 /*
 464  * Do the hard work of removing an element from the buddy allocator.
 465  * Call me with the zone->lock already held.
 466  */
 467 static struct page *__rmqueue(struct zone *zone, unsigned int order)
 468 {
 469         struct free_area * area;
 470         unsigned int current_order;
 471         struct page *page;
 472
 473         for (current_order = order; current_order < MAX_ORDER; ++current_order) {
 474                 area = zone->free_area + current_order;
 475                 if (list_empty(&area->free_list))
 476                         continue;
 477
 478                 page = list_entry(area->free_list.next, struct page, lru);
 479                 list_del(&page->lru);
 480                 rmv_page_order(page);
 481                 area->nr_free--;
 482                 zone->free_pages -= 1UL << order;
 483                 return expand(zone, page, order, current_order, area);
 484         }
 485
 486         return NULL;
 487 }
 488
 489 /*
 490  * Obtain a specified number of elements from the buddy allocator, all under
 491  * a single hold of the lock, for efficiency.  Add them to the supplied list.
 492  * Returns the number of new pages which were placed at *list.
 493  */
 494 static int rmqueue_bulk(struct zone *zone, unsigned int order,
 495                         unsigned long count, struct list_head *list)
 496 {
 497         unsigned long flags;
 498         int i;
 499         int allocated = 0;
 500         struct page *page;
 501
 502         spin_lock_irqsave(&zone->lock, flags);
 503         for (i = 0; i < count; ++i) {
 504                 page = __rmqueue(zone, order);
 505                 if (page == NULL)
 506                         break;
 507                 allocated++;
 508                 list_add_tail(&page->lru, list);
 509         }
 510         spin_unlock_irqrestore(&zone->lock, flags);
 511         return allocated;
 512 }
 513
 514 #if defined(CONFIG_PM) || defined(CONFIG_HOTPLUG_CPU)
 515 static void __drain_pages(unsigned int cpu)
 516 {
 517         struct zone *zone;
 518         int i;
 519
 520         for_each_zone(zone) {
 521                 struct per_cpu_pageset *pset;
 522
 523                 pset = &zone->pageset[cpu];
 524                 for (i = 0; i < ARRAY_SIZE(pset->pcp); i++) {
 525                         struct per_cpu_pages *pcp;
 526
 527                         pcp = &pset->pcp[i];
 528                         pcp->count -= free_pages_bulk(zone, pcp->count,
 529                                                 &pcp->list, 0);
 530                 }
 531         }
 532 }
 533 #endif /* CONFIG_PM || CONFIG_HOTPLUG_CPU */
 534
 535 #ifdef CONFIG_PM
 536
 537 void mark_free_pages(struct zone *zone)
 538 {
 539         unsigned long zone_pfn, flags;
 540         int order;
 541         struct list_head *curr;
 542
 543         if (!zone->spanned_pages)
 544                 return;
 545
 546         spin_lock_irqsave(&zone->lock, flags);
 547         for (zone_pfn = 0; zone_pfn < zone->spanned_pages; ++zone_pfn)
 548                 ClearPageNosaveFree(pfn_to_page(zone_pfn + zone->zone_start_pfn));
 549
 550         for (order = MAX_ORDER - 1; order >= 0; --order)
 551                 list_for_each(curr, &zone->free_area[order].free_list) {
 552                         unsigned long start_pfn, i;
 553
 554                         start_pfn = page_to_pfn(list_entry(curr, struct page, lru));
 555
 556                         for (i=0; i < (1<<order); i++)
 557                                 SetPageNosaveFree(pfn_to_page(start_pfn+i));
 558         }
 559         spin_unlock_irqrestore(&zone->lock, flags);
 560 }
 561
 562 /*
 563  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
 564  */
 565 void drain_local_pages(void)
 566 {
 567         unsigned long flags;
 568
 569         local_irq_save(flags);
 570         __drain_pages(smp_processor_id());
 571         local_irq_restore(flags);
 572 }
 573 #endif /* CONFIG_PM */
 574
 575 static void zone_statistics(struct zonelist *zonelist, struct zone *z)
 576 {
 577 #ifdef CONFIG_NUMA
 578         unsigned long flags;
 579         int cpu;
 580         pg_data_t *pg = z->zone_pgdat;
 581         pg_data_t *orig = zonelist->zones[0]->zone_pgdat;
 582         struct per_cpu_pageset *p;
 583
 584         local_irq_save(flags);
 585         cpu = smp_processor_id();
 586         p = &z->pageset[cpu];
 587         if (pg == orig) {
 588                 z->pageset[cpu].numa_hit++;
 589         } else {
 590                 p->numa_miss++;
 591                 zonelist->zones[0]->pageset[cpu].numa_foreign++;
 592         }
 593         if (pg == NODE_DATA(numa_node_id()))
 594                 p->local_node++;
 595         else
 596                 p->other_node++;
 597         local_irq_restore(flags);
 598 #endif
 599 }
 600
 601 /*
 602  * Free a 0-order page
 603  */
 604 static void FASTCALL(free_hot_cold_page(struct page *page, int cold));
 605 static void fastcall free_hot_cold_page(struct page *page, int cold)
 606 {
 607         struct zone *zone = page_zone(page);
 608         struct per_cpu_pages *pcp;
 609         unsigned long flags;
 610
 611         arch_free_page(page, 0);
 612
 613         kernel_map_pages(page, 1, 0);
 614         inc_page_state(pgfree);
 615         if (PageAnon(page))
 616                 page->mapping = NULL;
 617         free_pages_check(__FUNCTION__, page);
 618         pcp = &zone->pageset[get_cpu()].pcp[cold];
 619         local_irq_save(flags);
 620         if (pcp->count >= pcp->high)
 621                 pcp->count -= free_pages_bulk(zone, pcp->batch, &pcp->list, 0);
 622         list_add(&page->lru, &pcp->list);
 623         pcp->count++;
 624         local_irq_restore(flags);
 625         put_cpu();
 626 }
 627
 628 void fastcall free_hot_page(struct page *page)
 629 {
 630         free_hot_cold_page(page, 0);
 631 }
 632
 633 void fastcall free_cold_page(struct page *page)
 634 {
 635         free_hot_cold_page(page, 1);
 636 }
 637
 638 static inline void prep_zero_page(struct page *page, int order, unsigned int __nocast gfp_flags)
 639 {
 640         int i;
 641
 642         BUG_ON((gfp_flags & (__GFP_WAIT | __GFP_HIGHMEM)) == __GFP_HIGHMEM);
 643         for(i = 0; i < (1 << order); i++)
 644                 clear_highpage(page + i);
 645 }
 646
 647 /*
 648  * Really, prep_compound_page() should be called from __rmqueue_bulk().  But
 649  * we cheat by calling it from here, in the order > 0 path.  Saves a branch
 650  * or two.
 651  */
 652 static struct page *
 653 buffered_rmqueue(struct zone *zone, int order, unsigned int __nocast gfp_flags)
 654 {
 655         unsigned long flags;
 656         struct page *page = NULL;
 657         int cold = !!(gfp_flags & __GFP_COLD);
 658
 659         if (order == 0) {
 660                 struct per_cpu_pages *pcp;
 661
 662                 pcp = &zone->pageset[get_cpu()].pcp[cold];
 663                 local_irq_save(flags);
 664                 if (pcp->count <= pcp->low)
 665                         pcp->count += rmqueue_bulk(zone, 0,
 666                                                 pcp->batch, &pcp->list);
 667                 if (pcp->count) {
 668                         page = list_entry(pcp->list.next, struct page, lru);
 669                         list_del(&page->lru);
 670                         pcp->count--;
 671                 }
 672                 local_irq_restore(flags);
 673                 put_cpu();
 674         }
 675
 676         if (page == NULL) {
 677                 spin_lock_irqsave(&zone->lock, flags);
 678                 page = __rmqueue(zone, order);
 679                 spin_unlock_irqrestore(&zone->lock, flags);
 680         }
 681
 682         if (page != NULL) {
 683                 BUG_ON(bad_range(zone, page));
 684                 mod_page_state_zone(zone, pgalloc, 1 << order);
 685                 prep_new_page(page, order);
 686
 687                 if (gfp_flags & __GFP_ZERO)
 688                         prep_zero_page(page, order, gfp_flags);
 689
 690                 if (order && (gfp_flags & __GFP_COMP))
 691                         prep_compound_page(page, order);
 692         }
 693         return page;
 694 }
 695
 696 /*
 697  * Return 1 if free pages are above 'mark'. This takes into account the order
 698  * of the allocation.
 699  */
 700 int zone_watermark_ok(struct zone *z, int order, unsigned long mark,
 701                       int classzone_idx, int can_try_harder, int gfp_high)
 702 {
 703         /* free_pages my go negative - that's OK */
 704         long min = mark, free_pages = z->free_pages - (1 << order) + 1;
 705         int o;
 706
 707         if (gfp_high)
 708                 min -= min / 2;
 709         if (can_try_harder)
 710                 min -= min / 4;
 711
 712         if (free_pages <= min + z->lowmem_reserve[classzone_idx])
 713                 return 0;
 714         for (o = 0; o < order; o++) {
 715                 /* At the next order, this order's pages become unavailable */
 716                 free_pages -= z->free_area[o].nr_free << o;
 717
 718                 /* Require fewer higher order pages to be free */
 719                 min >>= 1;
 720
 721                 if (free_pages <= min)
 722                         return 0;
 723         }
 724         return 1;
 725 }
 726
 727 /*
 728  * This is the 'heart' of the zoned buddy allocator.
 729  */
 730 struct page * fastcall
 731 __alloc_pages(unsigned int __nocast gfp_mask, unsigned int order,
 732                 struct zonelist *zonelist)
 733 {
 734         const int wait = gfp_mask & __GFP_WAIT;
 735         struct zone **zones, *z;
 736         struct page *page;
 737         struct reclaim_state reclaim_state;
 738         struct task_struct *p = current;
 739         int i;
 740         int classzone_idx;
 741         int do_retry;
 742         int can_try_harder;
 743         int did_some_progress;
 744
 745         might_sleep_if(wait);
 746
 747         /*
 748          * The caller may dip into page reserves a bit more if the caller
 749          * cannot run direct reclaim, or is the caller has realtime scheduling
 750          * policy
 751          */
 752         can_try_harder = (unlikely(rt_task(p)) && !in_interrupt()) || !wait;
 753
 754         zones = zonelist->zones;  /* the list of zones suitable for gfp_mask */
 755
 756         if (unlikely(zones[0] == NULL)) {
 757                 /* Should this ever happen?? */
 758                 return NULL;
 759         }
 760
 761         classzone_idx = zone_idx(zones[0]);
 762
 763  restart:
 764         /* Go through the zonelist once, looking for a zone with enough free */
 765         for (i = 0; (z = zones[i]) != NULL; i++) {
 766
 767                 if (!zone_watermark_ok(z, order, z->pages_low,
 768                                        classzone_idx, 0, 0))
 769                         continue;
 770
 771                 if (!cpuset_zone_allowed(z))
 772                         continue;
 773
 774                 page = buffered_rmqueue(z, order, gfp_mask);
 775                 if (page)
 776                         goto got_pg;
 777         }
 778
 779         for (i = 0; (z = zones[i]) != NULL; i++)
 780                 wakeup_kswapd(z, order);
 781
 782         /*
 783          * Go through the zonelist again. Let __GFP_HIGH and allocations
 784          * coming from realtime tasks to go deeper into reserves
 785          *
 786          * This is the last chance, in general, before the goto nopage.
 787          * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
 788          */
 789         for (i = 0; (z = zones[i]) != NULL; i++) {
 790                 if (!zone_watermark_ok(z, order, z->pages_min,
 791                                        classzone_idx, can_try_harder,
 792                                        gfp_mask & __GFP_HIGH))
 793                         continue;
 794
 795                 if (wait && !cpuset_zone_allowed(z))
 796                         continue;
 797
 798                 page = buffered_rmqueue(z, order, gfp_mask);
 799                 if (page)
 800                         goto got_pg;
 801         }
 802
 803         /* This allocation should allow future memory freeing. */
 804
 805         if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE)))
 806                         && !in_interrupt()) {
 807                 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
 808                         /* go through the zonelist yet again, ignoring mins */
 809                         for (i = 0; (z = zones[i]) != NULL; i++) {
 810                                 if (!cpuset_zone_allowed(z))
 811                                         continue;
 812                                 page = buffered_rmqueue(z, order, gfp_mask);
 813                                 if (page)
 814                                         goto got_pg;
 815                         }
 816                 }
 817                 goto nopage;
 818         }
 819
 820         /* Atomic allocations - we can't balance anything */
 821         if (!wait)
 822                 goto nopage;
 823
 824 rebalance:
 825         cond_resched();
 826
 827         /* We now go into synchronous reclaim */
 828         p->flags |= PF_MEMALLOC;
 829         reclaim_state.reclaimed_slab = 0;
 830         p->reclaim_state = &reclaim_state;
 831
 832         did_some_progress = try_to_free_pages(zones, gfp_mask, order);
 833
 834         p->reclaim_state = NULL;
 835         p->flags &= ~PF_MEMALLOC;
 836
 837         cond_resched();
 838
 839         if (likely(did_some_progress)) {
 840                 /*
 841                  * Go through the zonelist yet one more time, keep
 842                  * very high watermark here, this is only to catch
 843                  * a parallel oom killing, we must fail if we're still
 844                  * under heavy pressure.
 845                  */
 846                 for (i = 0; (z = zones[i]) != NULL; i++) {
 847                         if (!zone_watermark_ok(z, order, z->pages_min,
 848                                                classzone_idx, can_try_harder,
 849                                                gfp_mask & __GFP_HIGH))
 850                                 continue;
 851
 852                         if (!cpuset_zone_allowed(z))
 853                                 continue;
 854
 855                         page = buffered_rmqueue(z, order, gfp_mask);
 856                         if (page)
 857                                 goto got_pg;
 858                 }
 859         } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
 860                 /*
 861                  * Go through the zonelist yet one more time, keep
 862                  * very high watermark here, this is only to catch
 863                  * a parallel oom killing, we must fail if we're still
 864                  * under heavy pressure.
 865                  */
 866                 for (i = 0; (z = zones[i]) != NULL; i++) {
 867                         if (!zone_watermark_ok(z, order, z->pages_high,
 868                                                classzone_idx, 0, 0))
 869                                 continue;
 870
 871                         if (!cpuset_zone_allowed(z))
 872                                 continue;
 873
 874                         page = buffered_rmqueue(z, order, gfp_mask);
 875                         if (page)
 876                                 goto got_pg;
 877                 }
 878
 879                 out_of_memory(gfp_mask);
 880                 goto restart;
 881         }
 882
 883         /*
 884          * Don't let big-order allocations loop unless the caller explicitly
 885          * requests that.  Wait for some write requests to complete then retry.
 886          *
 887          * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order
 888          * <= 3, but that may not be true in other implementations.
 889          */
 890         do_retry = 0;
 891         if (!(gfp_mask & __GFP_NORETRY)) {
 892                 if ((order <= 3) || (gfp_mask & __GFP_REPEAT))
 893                         do_retry = 1;
 894                 if (gfp_mask & __GFP_NOFAIL)
 895                         do_retry = 1;
 896         }
 897         if (do_retry) {
 898                 blk_congestion_wait(WRITE, HZ/50);
 899                 goto rebalance;
 900         }
 901
 902 nopage:
 903         if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) {
 904                 printk(KERN_WARNING "%s: page allocation failure."
 905                         " order:%d, mode:0x%x\n",
 906                         p->comm, order, gfp_mask);
 907                 dump_stack();
 908         }
 909         return NULL;
 910 got_pg:
 911         zone_statistics(zonelist, z);
 912         return page;
 913 }
 914
 915 EXPORT_SYMBOL(__alloc_pages);
 916
 917 /*
 918  * Common helper functions.
 919  */
 920 fastcall unsigned long __get_free_pages(unsigned int __nocast gfp_mask, unsigned int order)
 921 {
 922         struct page * page;
 923         page = alloc_pages(gfp_mask, order);
 924         if (!page)
 925                 return 0;
 926         return (unsigned long) page_address(page);
 927 }
 928
 929 EXPORT_SYMBOL(__get_free_pages);
 930
 931 fastcall unsigned long get_zeroed_page(unsigned int __nocast gfp_mask)
 932 {
 933         struct page * page;
 934
 935         /*
 936          * get_zeroed_page() returns a 32-bit address, which cannot represent
 937          * a highmem page
 938          */
 939         BUG_ON(gfp_mask & __GFP_HIGHMEM);
 940
 941         page = alloc_pages(gfp_mask | __GFP_ZERO, 0);
 942         if (page)
 943                 return (unsigned long) page_address(page);
 944         return 0;
 945 }
 946
 947 EXPORT_SYMBOL(get_zeroed_page);
 948
 949 void __pagevec_free(struct pagevec *pvec)
 950 {
 951         int i = pagevec_count(pvec);
 952
 953         while (--i >= 0)
 954                 free_hot_cold_page(pvec->pages[i], pvec->cold);
 955 }
 956
 957 fastcall void __free_pages(struct page *page, unsigned int order)
 958 {
 959         if (!PageReserved(page) && put_page_testzero(page)) {
 960                 if (order == 0)
 961                         free_hot_page(page);
 962                 else
 963                         __free_pages_ok(page, order);
 964         }
 965 }
 966
 967 EXPORT_SYMBOL(__free_pages);
 968
 969 fastcall void free_pages(unsigned long addr, unsigned int order)
 970 {
 971         if (addr != 0) {
 972                 BUG_ON(!virt_addr_valid((void *)addr));
 973                 __free_pages(virt_to_page((void *)addr), order);
 974         }
 975 }
 976
 977 EXPORT_SYMBOL(free_pages);
 978
 979 /*
 980  * Total amount of free (allocatable) RAM:
 981  */
 982 unsigned int nr_free_pages(void)
 983 {
 984         unsigned int sum = 0;
 985         struct zone *zone;
 986
 987         for_each_zone(zone)
 988                 sum += zone->free_pages;
 989
 990         return sum;
 991 }
 992
 993 EXPORT_SYMBOL(nr_free_pages);
 994
 995 #ifdef CONFIG_NUMA
 996 unsigned int nr_free_pages_pgdat(pg_data_t *pgdat)
 997 {
 998         unsigned int i, sum = 0;
 999
1000         for (i = 0; i < MAX_NR_ZONES; i++)
1001                 sum += pgdat->node_zones[i].free_pages;
1002
1003         return sum;
1004 }
1005 #endif
1006
1007 static unsigned int nr_free_zone_pages(int offset)
1008 {
1009         pg_data_t *pgdat;
1010         unsigned int sum = 0;
1011
1012         for_each_pgdat(pgdat) {
1013                 struct zonelist *zonelist = pgdat->node_zonelists + offset;
1014                 struct zone **zonep = zonelist->zones;
1015                 struct zone *zone;
1016
1017                 for (zone = *zonep++; zone; zone = *zonep++) {
1018                         unsigned long size = zone->present_pages;
1019                         unsigned long high = zone->pages_high;
1020                         if (size > high)
1021                                 sum += size - high;
1022                 }
1023         }
1024
1025         return sum;
1026 }
1027
1028 /*
1029  * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
1030  */
1031 unsigned int nr_free_buffer_pages(void)
1032 {
1033         return nr_free_zone_pages(GFP_USER & GFP_ZONEMASK);
1034 }
1035
1036 /*
1037  * Amount of free RAM allocatable within all zones
1038  */
1039 unsigned int nr_free_pagecache_pages(void)
1040 {
1041         return nr_free_zone_pages(GFP_HIGHUSER & GFP_ZONEMASK);
1042 }
1043
1044 #ifdef CONFIG_HIGHMEM
1045 unsigned int nr_free_highpages (void)
1046 {
1047         pg_data_t *pgdat;
1048         unsigned int pages = 0;
1049
1050         for_each_pgdat(pgdat)
1051                 pages += pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1052
1053         return pages;
1054 }
1055 #endif
1056
1057 #ifdef CONFIG_NUMA
1058 static void show_node(struct zone *zone)
1059 {
1060         printk("Node %d ", zone->zone_pgdat->node_id);
1061 }
1062 #else
1063 #define show_node(zone) do { } while (0)
1064 #endif
1065
1066 /*
1067  * Accumulate the page_state information across all CPUs.
1068  * The result is unavoidably approximate - it can change
1069  * during and after execution of this function.
1070  */
1071 static DEFINE_PER_CPU(struct page_state, page_states) = {0};
1072
1073 atomic_t nr_pagecache = ATOMIC_INIT(0);
1074 EXPORT_SYMBOL(nr_pagecache);
1075 #ifdef CONFIG_SMP
1076 DEFINE_PER_CPU(long, nr_pagecache_local) = 0;
1077 #endif
1078
1079 void __get_page_state(struct page_state *ret, int nr)
1080 {
1081         int cpu = 0;
1082
1083         memset(ret, 0, sizeof(*ret));
1084
1085         cpu = first_cpu(cpu_online_map);
1086         while (cpu < NR_CPUS) {
1087                 unsigned long *in, *out, off;
1088
1089                 in = (unsigned long *)&per_cpu(page_states, cpu);
1090
1091                 cpu = next_cpu(cpu, cpu_online_map);
1092
1093                 if (cpu < NR_CPUS)
1094                         prefetch(&per_cpu(page_states, cpu));
1095
1096                 out = (unsigned long *)ret;
1097                 for (off = 0; off < nr; off++)
1098                         *out++ += *in++;
1099         }
1100 }
1101
1102 void get_page_state(struct page_state *ret)
1103 {
1104         int nr;
1105
1106         nr = offsetof(struct page_state, GET_PAGE_STATE_LAST);
1107         nr /= sizeof(unsigned long);
1108
1109         __get_page_state(ret, nr + 1);
1110 }
1111
1112 void get_full_page_state(struct page_state *ret)
1113 {
1114         __get_page_state(ret, sizeof(*ret) / sizeof(unsigned long));
1115 }
1116
1117 unsigned long __read_page_state(unsigned offset)
1118 {
1119         unsigned long ret = 0;
1120         int cpu;
1121
1122         for_each_online_cpu(cpu) {
1123                 unsigned long in;
1124
1125                 in = (unsigned long)&per_cpu(page_states, cpu) + offset;
1126                 ret += *((unsigned long *)in);
1127         }
1128         return ret;
1129 }
1130
1131 void __mod_page_state(unsigned offset, unsigned long delta)
1132 {
1133         unsigned long flags;
1134         void* ptr;
1135
1136         local_irq_save(flags);
1137         ptr = &__get_cpu_var(page_states);
1138         *(unsigned long*)(ptr + offset) += delta;
1139         local_irq_restore(flags);
1140 }
1141
1142 EXPORT_SYMBOL(__mod_page_state);
1143
1144 void __get_zone_counts(unsigned long *active, unsigned long *inactive,
1145                         unsigned long *free, struct pglist_data *pgdat)
1146 {
1147         struct zone *zones = pgdat->node_zones;
1148         int i;
1149
1150         *active = 0;
1151         *inactive = 0;
1152         *free = 0;
1153         for (i = 0; i < MAX_NR_ZONES; i++) {
1154                 *active += zones[i].nr_active;
1155                 *inactive += zones[i].nr_inactive;
1156                 *free += zones[i].free_pages;
1157         }
1158 }
1159
1160 void get_zone_counts(unsigned long *active,
1161                 unsigned long *inactive, unsigned long *free)
1162 {
1163         struct pglist_data *pgdat;
1164
1165         *active = 0;
1166         *inactive = 0;
1167         *free = 0;
1168         for_each_pgdat(pgdat) {
1169                 unsigned long l, m, n;
1170                 __get_zone_counts(&l, &m, &n, pgdat);
1171                 *active += l;
1172                 *inactive += m;
1173                 *free += n;
1174         }
1175 }
1176
1177 void si_meminfo(struct sysinfo *val)
1178 {
1179         val->totalram = totalram_pages;
1180         val->sharedram = 0;
1181         val->freeram = nr_free_pages();
1182         val->bufferram = nr_blockdev_pages();
1183 #ifdef CONFIG_HIGHMEM
1184         val->totalhigh = totalhigh_pages;
1185         val->freehigh = nr_free_highpages();
1186 #else
1187         val->totalhigh = 0;
1188         val->freehigh = 0;
1189 #endif
1190         val->mem_unit = PAGE_SIZE;
1191 }
1192
1193 EXPORT_SYMBOL(si_meminfo);
1194
1195 #ifdef CONFIG_NUMA
1196 void si_meminfo_node(struct sysinfo *val, int nid)
1197 {
1198         pg_data_t *pgdat = NODE_DATA(nid);
1199
1200         val->totalram = pgdat->node_present_pages;
1201         val->freeram = nr_free_pages_pgdat(pgdat);
1202         val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
1203         val->freehigh = pgdat->node_zones[ZONE_HIGHMEM].free_pages;
1204         val->mem_unit = PAGE_SIZE;
1205 }
1206 #endif
1207
1208 #define K(x) ((x) << (PAGE_SHIFT-10))
1209
1210 /*
1211  * Show free area list (used inside shift_scroll-lock stuff)
1212  * We also calculate the percentage fragmentation. We do this by counting the
1213  * memory on each free list with the exception of the first item on the list.
1214  */
1215 void show_free_areas(void)
1216 {
1217         struct page_state ps;
1218         int cpu, temperature;
1219         unsigned long active;
1220         unsigned long inactive;
1221         unsigned long free;
1222         struct zone *zone;
1223
1224         for_each_zone(zone) {
1225                 show_node(zone);
1226                 printk("%s per-cpu:", zone->name);
1227
1228                 if (!zone->present_pages) {
1229                         printk(" empty\n");
1230                         continue;
1231                 } else
1232                         printk("\n");
1233
1234                 for (cpu = 0; cpu < NR_CPUS; ++cpu) {
1235                         struct per_cpu_pageset *pageset;
1236
1237                         if (!cpu_possible(cpu))
1238                                 continue;
1239
1240                         pageset = zone->pageset + cpu;
1241
1242                         for (temperature = 0; temperature < 2; temperature++)
1243                                 printk("cpu %d %s: low %d, high %d, batch %d\n",
1244                                         cpu,
1245                                         temperature ? "cold" : "hot",
1246                                         pageset->pcp[temperature].low,
1247                                         pageset->pcp[temperature].high,
1248                                         pageset->pcp[temperature].batch);
1249                 }
1250         }
1251
1252         get_page_state(&ps);
1253         get_zone_counts(&active, &inactive, &free);
1254
1255         printk("\nFree pages: %11ukB (%ukB HighMem)\n",
1256                 K(nr_free_pages()),
1257                 K(nr_free_highpages()));
1258
1259         printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu "
1260                 "unstable:%lu free:%u slab:%lu mapped:%lu pagetables:%lu\n",
1261                 active,
1262                 inactive,
1263                 ps.nr_dirty,
1264                 ps.nr_writeback,
1265                 ps.nr_unstable,
1266                 nr_free_pages(),
1267                 ps.nr_slab,
1268                 ps.nr_mapped,
1269                 ps.nr_page_table_pages);
1270
1271         for_each_zone(zone) {
1272                 int i;
1273
1274                 show_node(zone);
1275                 printk("%s"
1276                         " free:%lukB"
1277                         " min:%lukB"
1278                         " low:%lukB"
1279                         " high:%lukB"
1280                         " active:%lukB"
1281                         " inactive:%lukB"
1282                         " present:%lukB"
1283                         " pages_scanned:%lu"
1284                         " all_unreclaimable? %s"
1285                         "\n",
1286                         zone->name,
1287                         K(zone->free_pages),
1288                         K(zone->pages_min),
1289                         K(zone->pages_low),
1290                         K(zone->pages_high),
1291                         K(zone->nr_active),
1292                         K(zone->nr_inactive),
1293                         K(zone->present_pages),
1294                         zone->pages_scanned,
1295                         (zone->all_unreclaimable ? "yes" : "no")
1296                         );
1297                 printk("lowmem_reserve[]:");
1298                 for (i = 0; i < MAX_NR_ZONES; i++)
1299                         printk(" %lu", zone->lowmem_reserve[i]);
1300                 printk("\n");
1301         }
1302
1303         for_each_zone(zone) {
1304                 unsigned long nr, flags, order, total = 0;
1305
1306                 show_node(zone);
1307                 printk("%s: ", zone->name);
1308                 if (!zone->present_pages) {
1309                         printk("empty\n");
1310                         continue;
1311                 }
1312
1313                 spin_lock_irqsave(&zone->lock, flags);
1314                 for (order = 0; order < MAX_ORDER; order++) {
1315                         nr = zone->free_area[order].nr_free;
1316                         total += nr << order;
1317                         printk("%lu*%lukB ", nr, K(1UL) << order);
1318                 }
1319                 spin_unlock_irqrestore(&zone->lock, flags);
1320                 printk("= %lukB\n", K(total));
1321         }
1322
1323         show_swap_cache_info();
1324 }
1325
1326 /*
1327  * Builds allocation fallback zone lists.
1328  */
1329 static int __init build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, int j, int k)
1330 {
1331         switch (k) {
1332                 struct zone *zone;
1333         default:
1334                 BUG();
1335         case ZONE_HIGHMEM:
1336                 zone = pgdat->node_zones + ZONE_HIGHMEM;
1337                 if (zone->present_pages) {
1338 #ifndef CONFIG_HIGHMEM
1339                         BUG();
1340 #endif
1341                         zonelist->zones[j++] = zone;
1342                 }
1343         case ZONE_NORMAL:
1344                 zone = pgdat->node_zones + ZONE_NORMAL;
1345                 if (zone->present_pages)
1346                         zonelist->zones[j++] = zone;
1347         case ZONE_DMA:
1348                 zone = pgdat->node_zones + ZONE_DMA;
1349                 if (zone->present_pages)
1350                         zonelist->zones[j++] = zone;
1351         }
1352
1353         return j;
1354 }
1355
1356 #ifdef CONFIG_NUMA
1357 #define MAX_NODE_LOAD (num_online_nodes())
1358 static int __initdata node_load[MAX_NUMNODES];
1359 /**
1360  * find_next_best_node - find the next node that should appear in a given node's fallback list
1361  * @node: node whose fallback list we're appending
1362  * @used_node_mask: nodemask_t of already used nodes
1363  *
1364  * We use a number of factors to determine which is the next node that should
1365  * appear on a given node's fallback list.  The node should not have appeared
1366  * already in @node's fallback list, and it should be the next closest node
1367  * according to the distance array (which contains arbitrary distance values
1368  * from each node to each node in the system), and should also prefer nodes
1369  * with no CPUs, since presumably they'll have very little allocation pressure
1370  * on them otherwise.
1371  * It returns -1 if no node is found.
1372  */
1373 static int __init find_next_best_node(int node, nodemask_t *used_node_mask)
1374 {
1375         int i, n, val;
1376         int min_val = INT_MAX;
1377         int best_node = -1;
1378
1379         for_each_online_node(i) {
1380                 cpumask_t tmp;
1381
1382                 /* Start from local node */
1383                 n = (node+i) % num_online_nodes();
1384
1385                 /* Don't want a node to appear more than once */
1386                 if (node_isset(n, *used_node_mask))
1387                         continue;
1388
1389                 /* Use the local node if we haven't already */
1390                 if (!node_isset(node, *used_node_mask)) {
1391                         best_node = node;
1392                         break;
1393                 }
1394
1395                 /* Use the distance array to find the distance */
1396                 val = node_distance(node, n);
1397
1398                 /* Give preference to headless and unused nodes */
1399                 tmp = node_to_cpumask(n);
1400                 if (!cpus_empty(tmp))
1401                         val += PENALTY_FOR_NODE_WITH_CPUS;
1402
1403                 /* Slight preference for less loaded node */
1404                 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
1405                 val += node_load[n];
1406
1407                 if (val < min_val) {
1408                         min_val = val;
1409                         best_node = n;
1410                 }
1411         }
1412
1413         if (best_node >= 0)
1414                 node_set(best_node, *used_node_mask);
1415
1416         return best_node;
1417 }
1418
1419 static void __init build_zonelists(pg_data_t *pgdat)
1420 {
1421         int i, j, k, node, local_node;
1422         int prev_node, load;
1423         struct zonelist *zonelist;
1424         nodemask_t used_mask;
1425
1426         /* initialize zonelists */
1427         for (i = 0; i < GFP_ZONETYPES; i++) {
1428                 zonelist = pgdat->node_zonelists + i;
1429                 zonelist->zones[0] = NULL;
1430         }
1431
1432         /* NUMA-aware ordering of nodes */
1433         local_node = pgdat->node_id;
1434         load = num_online_nodes();
1435         prev_node = local_node;
1436         nodes_clear(used_mask);
1437         while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
1438                 /*
1439                  * We don't want to pressure a particular node.
1440                  * So adding penalty to the first node in same
1441                  * distance group to make it round-robin.
1442                  */
1443                 if (node_distance(local_node, node) !=
1444                                 node_distance(local_node, prev_node))
1445                         node_load[node] += load;
1446                 prev_node = node;
1447                 load--;
1448                 for (i = 0; i < GFP_ZONETYPES; i++) {
1449                         zonelist = pgdat->node_zonelists + i;
1450                         for (j = 0; zonelist->zones[j] != NULL; j++);
1451
1452                         k = ZONE_NORMAL;
1453                         if (i & __GFP_HIGHMEM)
1454                                 k = ZONE_HIGHMEM;
1455                         if (i & __GFP_DMA)
1456                                 k = ZONE_DMA;
1457
1458                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1459                         zonelist->zones[j] = NULL;
1460                 }
1461         }
1462 }
1463
1464 #else   /* CONFIG_NUMA */
1465
1466 static void __init build_zonelists(pg_data_t *pgdat)
1467 {
1468         int i, j, k, node, local_node;
1469
1470         local_node = pgdat->node_id;
1471         for (i = 0; i < GFP_ZONETYPES; i++) {
1472                 struct zonelist *zonelist;
1473
1474                 zonelist = pgdat->node_zonelists + i;
1475
1476                 j = 0;
1477                 k = ZONE_NORMAL;
1478                 if (i & __GFP_HIGHMEM)
1479                         k = ZONE_HIGHMEM;
1480                 if (i & __GFP_DMA)
1481                         k = ZONE_DMA;
1482
1483                 j = build_zonelists_node(pgdat, zonelist, j, k);
1484                 /*
1485                  * Now we build the zonelist so that it contains the zones
1486                  * of all the other nodes.
1487                  * We don't want to pressure a particular node, so when
1488                  * building the zones for node N, we make sure that the
1489                  * zones coming right after the local ones are those from
1490                  * node N+1 (modulo N)
1491                  */
1492                 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
1493                         if (!node_online(node))
1494                                 continue;
1495                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1496                 }
1497                 for (node = 0; node < local_node; node++) {
1498                         if (!node_online(node))
1499                                 continue;
1500                         j = build_zonelists_node(NODE_DATA(node), zonelist, j, k);
1501                 }
1502
1503                 zonelist->zones[j] = NULL;
1504         }
1505 }
1506
1507 #endif  /* CONFIG_NUMA */
1508
1509 void __init build_all_zonelists(void)
1510 {
1511         int i;
1512
1513         for_each_online_node(i)
1514                 build_zonelists(NODE_DATA(i));
1515         printk("Built %i zonelists\n", num_online_nodes());
1516         cpuset_init_current_mems_allowed();
1517 }
1518
1519 /*
1520  * Helper functions to size the waitqueue hash table.
1521  * Essentially these want to choose hash table sizes sufficiently
1522  * large so that collisions trying to wait on pages are rare.
1523  * But in fact, the number of active page waitqueues on typical
1524  * systems is ridiculously low, less than 200. So this is even
1525  * conservative, even though it seems large.
1526  *
1527  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
1528  * waitqueues, i.e. the size of the waitq table given the number of pages.
1529  */
1530 #define PAGES_PER_WAITQUEUE     256
1531
1532 static inline unsigned long wait_table_size(unsigned long pages)
1533 {
1534         unsigned long size = 1;
1535
1536         pages /= PAGES_PER_WAITQUEUE;
1537
1538         while (size < pages)
1539                 size <<= 1;
1540
1541         /*
1542          * Once we have dozens or even hundreds of threads sleeping
1543          * on IO we've got bigger problems than wait queue collision.
1544          * Limit the size of the wait table to a reasonable size.
1545          */
1546         size = min(size, 4096UL);
1547
1548         return max(size, 4UL);
1549 }
1550
1551 /*
1552  * This is an integer logarithm so that shifts can be used later
1553  * to extract the more random high bits from the multiplicative
1554  * hash function before the remainder is taken.
1555  */
1556 static inline unsigned long wait_table_bits(unsigned long size)
1557 {
1558         return ffz(~size);
1559 }
1560
1561 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
1562
1563 static void __init calculate_zone_totalpages(struct pglist_data *pgdat,
1564                 unsigned long *zones_size, unsigned long *zholes_size)
1565 {
1566         unsigned long realtotalpages, totalpages = 0;
1567         int i;
1568
1569         for (i = 0; i < MAX_NR_ZONES; i++)
1570                 totalpages += zones_size[i];
1571         pgdat->node_spanned_pages = totalpages;
1572
1573         realtotalpages = totalpages;
1574         if (zholes_size)
1575                 for (i = 0; i < MAX_NR_ZONES; i++)
1576                         realtotalpages -= zholes_size[i];
1577         pgdat->node_present_pages = realtotalpages;
1578         printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
1579 }
1580
1581
1582 /*
1583  * Initially all pages are reserved - free ones are freed
1584  * up by free_all_bootmem() once the early boot process is
1585  * done. Non-atomic initialization, single-pass.
1586  */
1587 void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
1588                 unsigned long start_pfn)
1589 {
1590         struct page *start = pfn_to_page(start_pfn);
1591         struct page *page;
1592
1593         for (page = start; page < (start + size); page++) {
1594                 set_page_zone(page, NODEZONE(nid, zone));
1595                 set_page_count(page, 0);
1596                 reset_page_mapcount(page);
1597                 SetPageReserved(page);
1598                 INIT_LIST_HEAD(&page->lru);
1599 #ifdef WANT_PAGE_VIRTUAL
1600                 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1601                 if (!is_highmem_idx(zone))
1602                         set_page_address(page, __va(start_pfn << PAGE_SHIFT));
1603 #endif
1604                 start_pfn++;
1605         }
1606 }
1607
1608 void zone_init_free_lists(struct pglist_data *pgdat, struct zone *zone,
1609                                 unsigned long size)
1610 {
1611         int order;
1612         for (order = 0; order < MAX_ORDER ; order++) {
1613                 INIT_LIST_HEAD(&zone->free_area[order].free_list);
1614                 zone->free_area[order].nr_free = 0;
1615         }
1616 }
1617
1618 #ifndef __HAVE_ARCH_MEMMAP_INIT
1619 #define memmap_init(size, nid, zone, start_pfn) \
1620         memmap_init_zone((size), (nid), (zone), (start_pfn))
1621 #endif
1622
1623 /*
1624  * Set up the zone data structures:
1625  *   - mark all pages reserved
1626  *   - mark all memory queues empty
1627  *   - clear the memory bitmaps
1628  */
1629 static void __init free_area_init_core(struct pglist_data *pgdat,
1630                 unsigned long *zones_size, unsigned long *zholes_size)
1631 {
1632         unsigned long i, j;
1633         const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
1634         int cpu, nid = pgdat->node_id;
1635         unsigned long zone_start_pfn = pgdat->node_start_pfn;
1636
1637         pgdat->nr_zones = 0;
1638         init_waitqueue_head(&pgdat->kswapd_wait);
1639         pgdat->kswapd_max_order = 0;
1640
1641         for (j = 0; j < MAX_NR_ZONES; j++) {
1642                 struct zone *zone = pgdat->node_zones + j;
1643                 unsigned long size, realsize;
1644                 unsigned long batch;
1645
1646                 zone_table[NODEZONE(nid, j)] = zone;
1647                 realsize = size = zones_size[j];
1648                 if (zholes_size)
1649                         realsize -= zholes_size[j];
1650
1651                 if (j == ZONE_DMA || j == ZONE_NORMAL)
1652                         nr_kernel_pages += realsize;
1653                 nr_all_pages += realsize;
1654
1655                 zone->spanned_pages = size;
1656                 zone->present_pages = realsize;
1657                 zone->name = zone_names[j];
1658                 spin_lock_init(&zone->lock);
1659                 spin_lock_init(&zone->lru_lock);
1660                 zone->zone_pgdat = pgdat;
1661                 zone->free_pages = 0;
1662
1663                 zone->temp_priority = zone->prev_priority = DEF_PRIORITY;
1664
1665                 /*
1666                  * The per-cpu-pages pools are set to around 1000th of the
1667                  * size of the zone.  But no more than 1/4 of a meg - there's
1668                  * no point in going beyond the size of L2 cache.
1669                  *
1670                  * OK, so we don't know how big the cache is.  So guess.
1671                  */
1672                 batch = zone->present_pages / 1024;
1673                 if (batch * PAGE_SIZE > 256 * 1024)
1674                         batch = (256 * 1024) / PAGE_SIZE;
1675                 batch /= 4;             /* We effectively *= 4 below */
1676                 if (batch < 1)
1677                         batch = 1;
1678
1679                 /*
1680                  * Clamp the batch to a 2^n - 1 value. Having a power
1681                  * of 2 value was found to be more likely to have
1682                  * suboptimal cache aliasing properties in some cases.
1683                  *
1684                  * For example if 2 tasks are alternately allocating
1685                  * batches of pages, one task can end up with a lot
1686                  * of pages of one half of the possible page colors
1687                  * and the other with pages of the other colors.
1688                  */
1689                 batch = (1 << fls(batch + batch/2)) - 1;
1690
1691                 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1692                         struct per_cpu_pages *pcp;
1693
1694                         pcp = &zone->pageset[cpu].pcp[0];       /* hot */
1695                         pcp->count = 0;
1696                         pcp->low = 2 * batch;
1697                         pcp->high = 6 * batch;
1698                         pcp->batch = 1 * batch;
1699                         INIT_LIST_HEAD(&pcp->list);
1700
1701                         pcp = &zone->pageset[cpu].pcp[1];       /* cold */
1702                         pcp->count = 0;
1703                         pcp->low = 0;
1704                         pcp->high = 2 * batch;
1705                         pcp->batch = 1 * batch;
1706                         INIT_LIST_HEAD(&pcp->list);
1707                 }
1708                 printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%lu\n",
1709                                 zone_names[j], realsize, batch);
1710                 INIT_LIST_HEAD(&zone->active_list);
1711                 INIT_LIST_HEAD(&zone->inactive_list);
1712                 zone->nr_scan_active = 0;
1713                 zone->nr_scan_inactive = 0;
1714                 zone->nr_active = 0;
1715                 zone->nr_inactive = 0;
1716                 if (!size)
1717                         continue;
1718
1719                 /*
1720                  * The per-page waitqueue mechanism uses hashed waitqueues
1721                  * per zone.
1722                  */
1723                 zone->wait_table_size = wait_table_size(size);
1724                 zone->wait_table_bits =
1725                         wait_table_bits(zone->wait_table_size);
1726                 zone->wait_table = (wait_queue_head_t *)
1727                         alloc_bootmem_node(pgdat, zone->wait_table_size
1728                                                 * sizeof(wait_queue_head_t));
1729
1730                 for(i = 0; i < zone->wait_table_size; ++i)
1731                         init_waitqueue_head(zone->wait_table + i);
1732
1733                 pgdat->nr_zones = j+1;
1734
1735                 zone->zone_mem_map = pfn_to_page(zone_start_pfn);
1736                 zone->zone_start_pfn = zone_start_pfn;
1737
1738                 if ((zone_start_pfn) & (zone_required_alignment-1))
1739                         printk(KERN_CRIT "BUG: wrong zone alignment, it will crash\n");
1740
1741                 memmap_init(size, nid, j, zone_start_pfn);
1742
1743                 zone_start_pfn += size;
1744
1745                 zone_init_free_lists(pgdat, zone, zone->spanned_pages);
1746         }
1747 }
1748
1749 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
1750 {
1751         unsigned long size;
1752
1753         /* Skip empty nodes */
1754         if (!pgdat->node_spanned_pages)
1755                 return;
1756
1757         /* ia64 gets its own node_mem_map, before this, without bootmem */
1758         if (!pgdat->node_mem_map) {
1759                 size = (pgdat->node_spanned_pages + 1) * sizeof(struct page);
1760                 pgdat->node_mem_map = alloc_bootmem_node(pgdat, size);
1761         }
1762 #ifndef CONFIG_DISCONTIGMEM
1763         /*
1764          * With no DISCONTIG, the global mem_map is just set as node 0's
1765          */
1766         if (pgdat == NODE_DATA(0))
1767                 mem_map = NODE_DATA(0)->node_mem_map;
1768 #endif
1769 }
1770
1771 void __init free_area_init_node(int nid, struct pglist_data *pgdat,
1772                 unsigned long *zones_size, unsigned long node_start_pfn,
1773                 unsigned long *zholes_size)
1774 {
1775         pgdat->node_id = nid;
1776         pgdat->node_start_pfn = node_start_pfn;
1777         calculate_zone_totalpages(pgdat, zones_size, zholes_size);
1778
1779         alloc_node_mem_map(pgdat);
1780
1781         free_area_init_core(pgdat, zones_size, zholes_size);
1782 }
1783
1784 #ifndef CONFIG_DISCONTIGMEM
1785 static bootmem_data_t contig_bootmem_data;
1786 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data };
1787
1788 EXPORT_SYMBOL(contig_page_data);
1789
1790 void __init free_area_init(unsigned long *zones_size)
1791 {
1792         free_area_init_node(0, &contig_page_data, zones_size,
1793                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
1794 }
1795 #endif
1796
1797 #ifdef CONFIG_PROC_FS
1798
1799 #include <linux/seq_file.h>
1800
1801 static void *frag_start(struct seq_file *m, loff_t *pos)
1802 {
1803         pg_data_t *pgdat;
1804         loff_t node = *pos;
1805
1806         for (pgdat = pgdat_list; pgdat && node; pgdat = pgdat->pgdat_next)
1807                 --node;
1808
1809         return pgdat;
1810 }
1811
1812 static void *frag_next(struct seq_file *m, void *arg, loff_t *pos)
1813 {
1814         pg_data_t *pgdat = (pg_data_t *)arg;
1815
1816         (*pos)++;
1817         return pgdat->pgdat_next;
1818 }
1819
1820 static void frag_stop(struct seq_file *m, void *arg)
1821 {
1822 }
1823
1824 /*
1825  * This walks the free areas for each zone.
1826  */
1827 static int frag_show(struct seq_file *m, void *arg)
1828 {
1829         pg_data_t *pgdat = (pg_data_t *)arg;
1830         struct zone *zone;
1831         struct zone *node_zones = pgdat->node_zones;
1832         unsigned long flags;
1833         int order;
1834
1835         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
1836                 if (!zone->present_pages)
1837                         continue;
1838
1839                 spin_lock_irqsave(&zone->lock, flags);
1840                 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
1841                 for (order = 0; order < MAX_ORDER; ++order)
1842                         seq_printf(m, "%6lu ", zone->free_area[order].nr_free);
1843                 spin_unlock_irqrestore(&zone->lock, flags);
1844                 seq_putc(m, '\n');
1845         }
1846         return 0;
1847 }
1848
1849 struct seq_operations fragmentation_op = {
1850         .start  = frag_start,
1851         .next   = frag_next,
1852         .stop   = frag_stop,
1853         .show   = frag_show,
1854 };
1855
1856 /*
1857  * Output information about zones in @pgdat.
1858  */
1859 static int zoneinfo_show(struct seq_file *m, void *arg)
1860 {
1861         pg_data_t *pgdat = arg;
1862         struct zone *zone;
1863         struct zone *node_zones = pgdat->node_zones;
1864         unsigned long flags;
1865
1866         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; zone++) {
1867                 int i;
1868
1869                 if (!zone->present_pages)
1870                         continue;
1871
1872                 spin_lock_irqsave(&zone->lock, flags);
1873                 seq_printf(m, "Node %d, zone %8s", pgdat->node_id, zone->name);
1874                 seq_printf(m,
1875                            "\n  pages free     %lu"
1876                            "\n        min      %lu"
1877                            "\n        low      %lu"
1878                            "\n        high     %lu"
1879                            "\n        active   %lu"
1880                            "\n        inactive %lu"
1881                            "\n        scanned  %lu (a: %lu i: %lu)"
1882                            "\n        spanned  %lu"
1883                            "\n        present  %lu",
1884                            zone->free_pages,
1885                            zone->pages_min,
1886                            zone->pages_low,
1887                            zone->pages_high,
1888                            zone->nr_active,
1889                            zone->nr_inactive,
1890                            zone->pages_scanned,
1891                            zone->nr_scan_active, zone->nr_scan_inactive,
1892                            zone->spanned_pages,
1893                            zone->present_pages);
1894                 seq_printf(m,
1895                            "\n        protection: (%lu",
1896                            zone->lowmem_reserve[0]);
1897                 for (i = 1; i < ARRAY_SIZE(zone->lowmem_reserve); i++)
1898                         seq_printf(m, ", %lu", zone->lowmem_reserve[i]);
1899                 seq_printf(m,
1900                            ")"
1901                            "\n  pagesets");
1902                 for (i = 0; i < ARRAY_SIZE(zone->pageset); i++) {
1903                         struct per_cpu_pageset *pageset;
1904                         int j;
1905
1906                         pageset = &zone->pageset[i];
1907                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
1908                                 if (pageset->pcp[j].count)
1909                                         break;
1910                         }
1911                         if (j == ARRAY_SIZE(pageset->pcp))
1912                                 continue;
1913                         for (j = 0; j < ARRAY_SIZE(pageset->pcp); j++) {
1914                                 seq_printf(m,
1915                                            "\n    cpu: %i pcp: %i"
1916                                            "\n              count: %i"
1917                                            "\n              low:   %i"
1918                                            "\n              high:  %i"
1919                                            "\n              batch: %i",
1920                                            i, j,
1921                                            pageset->pcp[j].count,
1922                                            pageset->pcp[j].low,
1923                                            pageset->pcp[j].high,
1924                                            pageset->pcp[j].batch);
1925                         }
1926 #ifdef CONFIG_NUMA
1927                         seq_printf(m,
1928                                    "\n            numa_hit:       %lu"
1929                                    "\n            numa_miss:      %lu"
1930                                    "\n            numa_foreign:   %lu"
1931                                    "\n            interleave_hit: %lu"
1932                                    "\n            local_node:     %lu"
1933                                    "\n            other_node:     %lu",
1934                                    pageset->numa_hit,
1935                                    pageset->numa_miss,
1936                                    pageset->numa_foreign,
1937                                    pageset->interleave_hit,
1938                                    pageset->local_node,
1939                                    pageset->other_node);
1940 #endif
1941                 }
1942                 seq_printf(m,
1943                            "\n  all_unreclaimable: %u"
1944                            "\n  prev_priority:     %i"
1945                            "\n  temp_priority:     %i"
1946                            "\n  start_pfn:         %lu",
1947                            zone->all_unreclaimable,
1948                            zone->prev_priority,
1949                            zone->temp_priority,
1950                            zone->zone_start_pfn);
1951                 spin_unlock_irqrestore(&zone->lock, flags);
1952                 seq_putc(m, '\n');
1953         }
1954         return 0;
1955 }
1956
1957 struct seq_operations zoneinfo_op = {
1958         .start  = frag_start, /* iterate over all zones. The same as in
1959                                * fragmentation. */
1960         .next   = frag_next,
1961         .stop   = frag_stop,
1962         .show   = zoneinfo_show,
1963 };
1964
1965 static char *vmstat_text[] = {
1966         "nr_dirty",
1967         "nr_writeback",
1968         "nr_unstable",
1969         "nr_page_table_pages",
1970         "nr_mapped",
1971         "nr_slab",
1972
1973         "pgpgin",
1974         "pgpgout",
1975         "pswpin",
1976         "pswpout",
1977         "pgalloc_high",
1978
1979         "pgalloc_normal",
1980         "pgalloc_dma",
1981         "pgfree",
1982         "pgactivate",
1983         "pgdeactivate",
1984
1985         "pgfault",
1986         "pgmajfault",
1987         "pgrefill_high",
1988         "pgrefill_normal",
1989         "pgrefill_dma",
1990
1991         "pgsteal_high",
1992         "pgsteal_normal",
1993         "pgsteal_dma",
1994         "pgscan_kswapd_high",
1995         "pgscan_kswapd_normal",
1996
1997         "pgscan_kswapd_dma",
1998         "pgscan_direct_high",
1999         "pgscan_direct_normal",
2000         "pgscan_direct_dma",
2001         "pginodesteal",
2002
2003         "slabs_scanned",
2004         "kswapd_steal",
2005         "kswapd_inodesteal",
2006         "pageoutrun",
2007         "allocstall",
2008
2009         "pgrotated",
2010         "nr_bounce",
2011 };
2012
2013 static void *vmstat_start(struct seq_file *m, loff_t *pos)
2014 {
2015         struct page_state *ps;
2016
2017         if (*pos >= ARRAY_SIZE(vmstat_text))
2018                 return NULL;
2019
2020         ps = kmalloc(sizeof(*ps), GFP_KERNEL);
2021         m->private = ps;
2022         if (!ps)
2023                 return ERR_PTR(-ENOMEM);
2024         get_full_page_state(ps);
2025         ps->pgpgin /= 2;                /* sectors -> kbytes */
2026         ps->pgpgout /= 2;
2027         return (unsigned long *)ps + *pos;
2028 }
2029
2030 static void *vmstat_next(struct seq_file *m, void *arg, loff_t *pos)
2031 {
2032         (*pos)++;
2033         if (*pos >= ARRAY_SIZE(vmstat_text))
2034                 return NULL;
2035         return (unsigned long *)m->private + *pos;
2036 }
2037
2038 static int vmstat_show(struct seq_file *m, void *arg)
2039 {
2040         unsigned long *l = arg;
2041         unsigned long off = l - (unsigned long *)m->private;
2042
2043         seq_printf(m, "%s %lu\n", vmstat_text[off], *l);
2044         return 0;
2045 }
2046
2047 static void vmstat_stop(struct seq_file *m, void *arg)
2048 {
2049         kfree(m->private);
2050         m->private = NULL;
2051 }
2052
2053 struct seq_operations vmstat_op = {
2054         .start  = vmstat_start,
2055         .next   = vmstat_next,
2056         .stop   = vmstat_stop,
2057         .show   = vmstat_show,
2058 };
2059
2060 #endif /* CONFIG_PROC_FS */
2061
2062 #ifdef CONFIG_HOTPLUG_CPU
2063 static int page_alloc_cpu_notify(struct notifier_block *self,
2064                                  unsigned long action, void *hcpu)
2065 {
2066         int cpu = (unsigned long)hcpu;
2067         long *count;
2068         unsigned long *src, *dest;
2069
2070         if (action == CPU_DEAD) {
2071                 int i;
2072
2073                 /* Drain local pagecache count. */
2074                 count = &per_cpu(nr_pagecache_local, cpu);
2075                 atomic_add(*count, &nr_pagecache);
2076                 *count = 0;
2077                 local_irq_disable();
2078                 __drain_pages(cpu);
2079
2080                 /* Add dead cpu's page_states to our own. */
2081                 dest = (unsigned long *)&__get_cpu_var(page_states);
2082                 src = (unsigned long *)&per_cpu(page_states, cpu);
2083
2084                 for (i = 0; i < sizeof(struct page_state)/sizeof(unsigned long);
2085                                 i++) {
2086                         dest[i] += src[i];
2087                         src[i] = 0;
2088                 }
2089
2090                 local_irq_enable();
2091         }
2092         return NOTIFY_OK;
2093 }
2094 #endif /* CONFIG_HOTPLUG_CPU */
2095
2096 void __init page_alloc_init(void)
2097 {
2098         hotcpu_notifier(page_alloc_cpu_notify, 0);
2099 }
2100
2101 /*
2102  * setup_per_zone_lowmem_reserve - called whenever
2103  *      sysctl_lower_zone_reserve_ratio changes.  Ensures that each zone
2104  *      has a correct pages reserved value, so an adequate number of
2105  *      pages are left in the zone after a successful __alloc_pages().
2106  */
2107 static void setup_per_zone_lowmem_reserve(void)
2108 {
2109         struct pglist_data *pgdat;
2110         int j, idx;
2111
2112         for_each_pgdat(pgdat) {
2113                 for (j = 0; j < MAX_NR_ZONES; j++) {
2114                         struct zone *zone = pgdat->node_zones + j;
2115                         unsigned long present_pages = zone->present_pages;
2116
2117                         zone->lowmem_reserve[j] = 0;
2118
2119                         for (idx = j-1; idx >= 0; idx--) {
2120                                 struct zone *lower_zone;
2121
2122                                 if (sysctl_lowmem_reserve_ratio[idx] < 1)
2123                                         sysctl_lowmem_reserve_ratio[idx] = 1;
2124
2125                                 lower_zone = pgdat->node_zones + idx;
2126                                 lower_zone->lowmem_reserve[j] = present_pages /
2127                                         sysctl_lowmem_reserve_ratio[idx];
2128                                 present_pages += lower_zone->present_pages;
2129                         }
2130                 }
2131         }
2132 }
2133
2134 /*
2135  * setup_per_zone_pages_min - called when min_free_kbytes changes.  Ensures
2136  *      that the pages_{min,low,high} values for each zone are set correctly
2137  *      with respect to min_free_kbytes.
2138  */
2139 static void setup_per_zone_pages_min(void)
2140 {
2141         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
2142         unsigned long lowmem_pages = 0;
2143         struct zone *zone;
2144         unsigned long flags;
2145
2146         /* Calculate total number of !ZONE_HIGHMEM pages */
2147         for_each_zone(zone) {
2148                 if (!is_highmem(zone))
2149                         lowmem_pages += zone->present_pages;
2150         }
2151
2152         for_each_zone(zone) {
2153                 spin_lock_irqsave(&zone->lru_lock, flags);
2154                 if (is_highmem(zone)) {
2155                         /*
2156                          * Often, highmem doesn't need to reserve any pages.
2157                          * But the pages_min/low/high values are also used for
2158                          * batching up page reclaim activity so we need a
2159                          * decent value here.
2160                          */
2161                         int min_pages;
2162
2163                         min_pages = zone->present_pages / 1024;
2164                         if (min_pages < SWAP_CLUSTER_MAX)
2165                                 min_pages = SWAP_CLUSTER_MAX;
2166                         if (min_pages > 128)
2167                                 min_pages = 128;
2168                         zone->pages_min = min_pages;
2169                 } else {
2170                         /* if it's a lowmem zone, reserve a number of pages
2171                          * proportionate to the zone's size.
2172                          */
2173                         zone->pages_min = (pages_min * zone->present_pages) /
2174                                            lowmem_pages;
2175                 }
2176
2177                 /*
2178                  * When interpreting these watermarks, just keep in mind that:
2179                  * zone->pages_min == (zone->pages_min * 4) / 4;
2180                  */
2181                 zone->pages_low   = (zone->pages_min * 5) / 4;
2182                 zone->pages_high  = (zone->pages_min * 6) / 4;
2183                 spin_unlock_irqrestore(&zone->lru_lock, flags);
2184         }
2185 }
2186
2187 /*
2188  * Initialise min_free_kbytes.
2189  *
2190  * For small machines we want it small (128k min).  For large machines
2191  * we want it large (64MB max).  But it is not linear, because network
2192  * bandwidth does not increase linearly with machine size.  We use
2193  *
2194  *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
2195  *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
2196  *
2197  * which yields
2198  *
2199  * 16MB:        512k
2200  * 32MB:        724k
2201  * 64MB:        1024k
2202  * 128MB:       1448k
2203  * 256MB:       2048k
2204  * 512MB:       2896k
2205  * 1024MB:      4096k
2206  * 2048MB:      5792k
2207  * 4096MB:      8192k
2208  * 8192MB:      11584k
2209  * 16384MB:     16384k
2210  */
2211 static int __init init_per_zone_pages_min(void)
2212 {
2213         unsigned long lowmem_kbytes;
2214
2215         lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
2216
2217         min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
2218         if (min_free_kbytes < 128)
2219                 min_free_kbytes = 128;
2220         if (min_free_kbytes > 65536)
2221                 min_free_kbytes = 65536;
2222         setup_per_zone_pages_min();
2223         setup_per_zone_lowmem_reserve();
2224         return 0;
2225 }
2226 module_init(init_per_zone_pages_min)
2227
2228 /*
2229  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
2230  *      that we can call two helper functions whenever min_free_kbytes
2231  *      changes.
2232  */
2233 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
2234         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2235 {
2236         proc_dointvec(table, write, file, buffer, length, ppos);
2237         setup_per_zone_pages_min();
2238         return 0;
2239 }
2240
2241 /*
2242  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
2243  *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
2244  *      whenever sysctl_lowmem_reserve_ratio changes.
2245  *
2246  * The reserve ratio obviously has absolutely no relation with the
2247  * pages_min watermarks. The lowmem reserve ratio can only make sense
2248  * if in function of the boot time zone sizes.
2249  */
2250 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
2251         struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
2252 {
2253         proc_dointvec_minmax(table, write, file, buffer, length, ppos);
2254         setup_per_zone_lowmem_reserve();
2255         return 0;
2256 }
2257
2258 __initdata int hashdist = HASHDIST_DEFAULT;
2259
2260 #ifdef CONFIG_NUMA
2261 static int __init set_hashdist(char *str)
2262 {
2263         if (!str)
2264                 return 0;
2265         hashdist = simple_strtoul(str, &str, 0);
2266         return 1;
2267 }
2268 __setup("hashdist=", set_hashdist);
2269 #endif
2270
2271 /*
2272  * allocate a large system hash table from bootmem
2273  * - it is assumed that the hash table must contain an exact power-of-2
2274  *   quantity of entries
2275  * - limit is the number of hash buckets, not the total allocation size
2276  */
2277 void *__init alloc_large_system_hash(const char *tablename,
2278                                      unsigned long bucketsize,
2279                                      unsigned long numentries,
2280                                      int scale,
2281                                      int flags,
2282                                      unsigned int *_hash_shift,
2283                                      unsigned int *_hash_mask,
2284                                      unsigned long limit)
2285 {
2286         unsigned long long max = limit;
2287         unsigned long log2qty, size;
2288         void *table = NULL;
2289
2290         /* allow the kernel cmdline to have a say */
2291         if (!numentries) {
2292                 /* round applicable memory size up to nearest megabyte */
2293                 numentries = (flags & HASH_HIGHMEM) ? nr_all_pages : nr_kernel_pages;
2294                 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
2295                 numentries >>= 20 - PAGE_SHIFT;
2296                 numentries <<= 20 - PAGE_SHIFT;
2297
2298                 /* limit to 1 bucket per 2^scale bytes of low memory */
2299                 if (scale > PAGE_SHIFT)
2300                         numentries >>= (scale - PAGE_SHIFT);
2301                 else
2302                         numentries <<= (PAGE_SHIFT - scale);
2303         }
2304         /* rounded up to nearest power of 2 in size */
2305         numentries = 1UL << (long_log2(numentries) + 1);
2306
2307         /* limit allocation size to 1/16 total memory by default */
2308         if (max == 0) {
2309                 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
2310                 do_div(max, bucketsize);
2311         }
2312
2313         if (numentries > max)
2314                 numentries = max;
2315
2316         log2qty = long_log2(numentries);
2317
2318         do {
2319                 size = bucketsize << log2qty;
2320                 if (flags & HASH_EARLY)
2321                         table = alloc_bootmem(size);
2322                 else if (hashdist)
2323                         table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
2324                 else {
2325                         unsigned long order;
2326                         for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++)
2327                                 ;
2328                         table = (void*) __get_free_pages(GFP_ATOMIC, order);
2329                 }
2330         } while (!table && size > PAGE_SIZE && --log2qty);
2331
2332         if (!table)
2333                 panic("Failed to allocate %s hash table\n", tablename);
2334
2335         printk("%s hash table entries: %d (order: %d, %lu bytes)\n",
2336                tablename,
2337                (1U << log2qty),
2338                long_log2(size) - PAGE_SHIFT,
2339                size);
2340
2341         if (_hash_shift)
2342                 *_hash_shift = log2qty;
2343         if (_hash_mask)
2344                 *_hash_mask = (1 << log2qty) - 1;
2345
2346         return table;
2347 }