1/* Modified by Broadcom Corp. Portions Copyright (c) Broadcom Corp, 2012. */ 2/* 3 * linux/mm/page_alloc.c 4 * 5 * Manages the free list, the system allocates free pages here. 6 * Note that kmalloc() lives in slab.c 7 * 8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 9 * Swap reorganised 29.12.95, Stephen Tweedie 10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 16 */ 17 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/jiffies.h> 24#include <linux/bootmem.h> 25#include <linux/compiler.h> 26#include <linux/kernel.h> 27#include <linux/kmemcheck.h> 28#include <linux/module.h> 29#include <linux/suspend.h> 30#include <linux/pagevec.h> 31#include <linux/blkdev.h> 32#include <linux/slab.h> 33#include <linux/oom.h> 34#include <linux/notifier.h> 35#include <linux/topology.h> 36#include <linux/sysctl.h> 37#include <linux/cpu.h> 38#include <linux/cpuset.h> 39#include <linux/memory_hotplug.h> 40#include <linux/nodemask.h> 41#include <linux/vmalloc.h> 42#include <linux/mempolicy.h> 43#include <linux/stop_machine.h> 44#include <linux/sort.h> 45#include <linux/pfn.h> 46#include <linux/backing-dev.h> 47#include <linux/fault-inject.h> 48#include <linux/page-isolation.h> 49#include <linux/page_cgroup.h> 50#include <linux/debugobjects.h> 51#include <linux/kmemleak.h> 52#include <linux/memory.h> 53#include <linux/compaction.h> 54#include <trace/events/kmem.h> 55#include <linux/ftrace_event.h> 56 57#include <asm/tlbflush.h> 58#include <asm/div64.h> 59#include "internal.h" 60 61#include <typedefs.h> 62#include <bcmdefs.h> 63 64#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 65DEFINE_PER_CPU(int, numa_node); 66EXPORT_PER_CPU_SYMBOL(numa_node); 67#endif 68 69#ifdef CONFIG_HAVE_MEMORYLESS_NODES 70/* 71 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 72 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 73 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 74 * defined in <linux/topology.h>. 75 */ 76DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 77EXPORT_PER_CPU_SYMBOL(_numa_mem_); 78#endif 79 80/* 81 * Array of node states. 82 */ 83nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 84 [N_POSSIBLE] = NODE_MASK_ALL, 85 [N_ONLINE] = { { [0] = 1UL } }, 86#ifndef CONFIG_NUMA 87 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 88#ifdef CONFIG_HIGHMEM 89 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 90#endif 91 [N_CPU] = { { [0] = 1UL } }, 92#endif /* NUMA */ 93}; 94EXPORT_SYMBOL(node_states); 95 96unsigned long totalram_pages __read_mostly; 97unsigned long totalreserve_pages __read_mostly; 98int percpu_pagelist_fraction; 99gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 100 101#ifdef CONFIG_PM_SLEEP 102/* 103 * The following functions are used by the suspend/hibernate code to temporarily 104 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 105 * while devices are suspended. To avoid races with the suspend/hibernate code, 106 * they should always be called with pm_mutex held (gfp_allowed_mask also should 107 * only be modified with pm_mutex held, unless the suspend/hibernate code is 108 * guaranteed not to run in parallel with that modification). 109 */ 110 111static gfp_t saved_gfp_mask; 112 113void pm_restore_gfp_mask(void) 114{ 115 WARN_ON(!mutex_is_locked(&pm_mutex)); 116 if (saved_gfp_mask) { 117 gfp_allowed_mask = saved_gfp_mask; 118 saved_gfp_mask = 0; 119 } 120} 121 122void pm_restrict_gfp_mask(void) 123{ 124 WARN_ON(!mutex_is_locked(&pm_mutex)); 125 WARN_ON(saved_gfp_mask); 126 saved_gfp_mask = gfp_allowed_mask; 127 gfp_allowed_mask &= ~GFP_IOFS; 128} 129#endif /* CONFIG_PM_SLEEP */ 130 131#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 132int pageblock_order __read_mostly; 133#endif 134 135static void __free_pages_ok(struct page *page, unsigned int order); 136 137/* 138 * results with 256, 32 in the lowmem_reserve sysctl: 139 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 140 * 1G machine -> (16M dma, 784M normal, 224M high) 141 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 142 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 143 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 144 * 145 * TBD: should special case ZONE_DMA32 machines here - in those we normally 146 * don't need any ZONE_NORMAL reservation 147 */ 148int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 149#ifdef CONFIG_ZONE_DMA 150 256, 151#endif 152#ifdef CONFIG_ZONE_DMA32 153 256, 154#endif 155#ifdef CONFIG_HIGHMEM 156 32, 157#endif 158 32, 159}; 160 161EXPORT_SYMBOL(totalram_pages); 162 163static char * const zone_names[MAX_NR_ZONES] = { 164#ifdef CONFIG_ZONE_DMA 165 "DMA", 166#endif 167#ifdef CONFIG_ZONE_DMA32 168 "DMA32", 169#endif 170 "Normal", 171#ifdef CONFIG_HIGHMEM 172 "HighMem", 173#endif 174 "Movable", 175}; 176 177int min_free_kbytes = 1024; 178 179static unsigned long __meminitdata nr_kernel_pages; 180static unsigned long __meminitdata nr_all_pages; 181static unsigned long __meminitdata dma_reserve; 182 183#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 184 /* 185 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 186 * ranges of memory (RAM) that may be registered with add_active_range(). 187 * Ranges passed to add_active_range() will be merged if possible 188 * so the number of times add_active_range() can be called is 189 * related to the number of nodes and the number of holes 190 */ 191 #ifdef CONFIG_MAX_ACTIVE_REGIONS 192 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 193 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 194 #else 195 #if MAX_NUMNODES >= 32 196 /* If there can be many nodes, allow up to 50 holes per node */ 197 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 198 #else 199 /* By default, allow up to 256 distinct regions */ 200 #define MAX_ACTIVE_REGIONS 256 201 #endif 202 #endif 203 204 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 205 static int __meminitdata nr_nodemap_entries; 206 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 207 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 208 static unsigned long __initdata required_kernelcore; 209 static unsigned long __initdata required_movablecore; 210 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 211 212 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 213 int movable_zone; 214 EXPORT_SYMBOL(movable_zone); 215#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 216 217#if MAX_NUMNODES > 1 218int nr_node_ids __read_mostly = MAX_NUMNODES; 219int nr_online_nodes __read_mostly = 1; 220EXPORT_SYMBOL(nr_node_ids); 221EXPORT_SYMBOL(nr_online_nodes); 222#endif 223 224int page_group_by_mobility_disabled __read_mostly; 225 226static void set_pageblock_migratetype(struct page *page, int migratetype) 227{ 228 229 if (unlikely(page_group_by_mobility_disabled)) 230 migratetype = MIGRATE_UNMOVABLE; 231 232 set_pageblock_flags_group(page, (unsigned long)migratetype, 233 PB_migrate, PB_migrate_end); 234} 235 236bool oom_killer_disabled __read_mostly; 237 238#ifdef CONFIG_DEBUG_VM 239static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 240{ 241 int ret = 0; 242 unsigned seq; 243 unsigned long pfn = page_to_pfn(page); 244 245 do { 246 seq = zone_span_seqbegin(zone); 247 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 248 ret = 1; 249 else if (pfn < zone->zone_start_pfn) 250 ret = 1; 251 } while (zone_span_seqretry(zone, seq)); 252 253 return ret; 254} 255 256static int page_is_consistent(struct zone *zone, struct page *page) 257{ 258 if (!pfn_valid_within(page_to_pfn(page))) 259 return 0; 260 if (zone != page_zone(page)) 261 return 0; 262 263 return 1; 264} 265/* 266 * Temporary debugging check for pages not lying within a given zone. 267 */ 268static int bad_range(struct zone *zone, struct page *page) 269{ 270 if (page_outside_zone_boundaries(zone, page)) 271 return 1; 272 if (!page_is_consistent(zone, page)) 273 return 1; 274 275 return 0; 276} 277#else 278static inline int bad_range(struct zone *zone, struct page *page) 279{ 280 return 0; 281} 282#endif 283 284static void bad_page(struct page *page) 285{ 286 static unsigned long resume; 287 static unsigned long nr_shown; 288 static unsigned long nr_unshown; 289 290 /* Don't complain about poisoned pages */ 291 if (PageHWPoison(page)) { 292 __ClearPageBuddy(page); 293 return; 294 } 295 296 /* 297 * Allow a burst of 60 reports, then keep quiet for that minute; 298 * or allow a steady drip of one report per second. 299 */ 300 if (nr_shown == 60) { 301 if (time_before(jiffies, resume)) { 302 nr_unshown++; 303 goto out; 304 } 305 if (nr_unshown) { 306 printk(KERN_ALERT 307 "BUG: Bad page state: %lu messages suppressed\n", 308 nr_unshown); 309 nr_unshown = 0; 310 } 311 nr_shown = 0; 312 } 313 if (nr_shown++ == 0) 314 resume = jiffies + 60 * HZ; 315 316 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 317 current->comm, page_to_pfn(page)); 318 dump_page(page); 319 320 dump_stack(); 321out: 322 /* Leave bad fields for debug, except PageBuddy could make trouble */ 323 __ClearPageBuddy(page); 324 add_taint(TAINT_BAD_PAGE); 325} 326 327/* 328 * Higher-order pages are called "compound pages". They are structured thusly: 329 * 330 * The first PAGE_SIZE page is called the "head page". 331 * 332 * The remaining PAGE_SIZE pages are called "tail pages". 333 * 334 * All pages have PG_compound set. All pages have their ->private pointing at 335 * the head page (even the head page has this). 336 * 337 * The first tail page's ->lru.next holds the address of the compound page's 338 * put_page() function. Its ->lru.prev holds the order of allocation. 339 * This usage means that zero-order pages may not be compound. 340 */ 341 342static void free_compound_page(struct page *page) 343{ 344 __free_pages_ok(page, compound_order(page)); 345} 346 347void prep_compound_page(struct page *page, unsigned long order) 348{ 349 int i; 350 int nr_pages = 1 << order; 351 352 set_compound_page_dtor(page, free_compound_page); 353 set_compound_order(page, order); 354 __SetPageHead(page); 355 for (i = 1; i < nr_pages; i++) { 356 struct page *p = page + i; 357 358 __SetPageTail(p); 359 p->first_page = page; 360 } 361} 362 363static int destroy_compound_page(struct page *page, unsigned long order) 364{ 365 int i; 366 int nr_pages = 1 << order; 367 int bad = 0; 368 369 if (unlikely(compound_order(page) != order) || 370 unlikely(!PageHead(page))) { 371 bad_page(page); 372 bad++; 373 } 374 375 __ClearPageHead(page); 376 377 for (i = 1; i < nr_pages; i++) { 378 struct page *p = page + i; 379 380 if (unlikely(!PageTail(p) || (p->first_page != page))) { 381 bad_page(page); 382 bad++; 383 } 384 __ClearPageTail(p); 385 } 386 387 return bad; 388} 389 390static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 391{ 392 int i; 393 394 /* 395 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 396 * and __GFP_HIGHMEM from hard or soft interrupt context. 397 */ 398 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 399 for (i = 0; i < (1 << order); i++) 400 clear_highpage(page + i); 401} 402 403static inline void set_page_order(struct page *page, int order) 404{ 405 set_page_private(page, order); 406 __SetPageBuddy(page); 407} 408 409static inline void rmv_page_order(struct page *page) 410{ 411 __ClearPageBuddy(page); 412 set_page_private(page, 0); 413} 414 415/* 416 * Locate the struct page for both the matching buddy in our 417 * pair (buddy1) and the combined O(n+1) page they form (page). 418 * 419 * 1) Any buddy B1 will have an order O twin B2 which satisfies 420 * the following equation: 421 * B2 = B1 ^ (1 << O) 422 * For example, if the starting buddy (buddy2) is #8 its order 423 * 1 buddy is #10: 424 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 425 * 426 * 2) Any buddy B will have an order O+1 parent P which 427 * satisfies the following equation: 428 * P = B & ~(1 << O) 429 * 430 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 431 */ 432static inline struct page * 433__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 434{ 435 unsigned long buddy_idx = page_idx ^ (1 << order); 436 437 return page + (buddy_idx - page_idx); 438} 439 440static inline unsigned long 441__find_combined_index(unsigned long page_idx, unsigned int order) 442{ 443 return (page_idx & ~(1 << order)); 444} 445 446/* 447 * This function checks whether a page is free && is the buddy 448 * we can do coalesce a page and its buddy if 449 * (a) the buddy is not in a hole && 450 * (b) the buddy is in the buddy system && 451 * (c) a page and its buddy have the same order && 452 * (d) a page and its buddy are in the same zone. 453 * 454 * For recording whether a page is in the buddy system, we use PG_buddy. 455 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 456 * 457 * For recording page's order, we use page_private(page). 458 */ 459static inline int page_is_buddy(struct page *page, struct page *buddy, 460 int order) 461{ 462 if (!pfn_valid_within(page_to_pfn(buddy))) 463 return 0; 464 465 if (page_zone_id(page) != page_zone_id(buddy)) 466 return 0; 467 468 if (PageBuddy(buddy) && page_order(buddy) == order) { 469 VM_BUG_ON(page_count(buddy) != 0); 470 return 1; 471 } 472 return 0; 473} 474 475/* 476 * Freeing function for a buddy system allocator. 477 * 478 * The concept of a buddy system is to maintain direct-mapped table 479 * (containing bit values) for memory blocks of various "orders". 480 * The bottom level table contains the map for the smallest allocatable 481 * units of memory (here, pages), and each level above it describes 482 * pairs of units from the levels below, hence, "buddies". 483 * At a high level, all that happens here is marking the table entry 484 * at the bottom level available, and propagating the changes upward 485 * as necessary, plus some accounting needed to play nicely with other 486 * parts of the VM system. 487 * At each level, we keep a list of pages, which are heads of continuous 488 * free pages of length of (1 << order) and marked with PG_buddy. Page's 489 * order is recorded in page_private(page) field. 490 * So when we are allocating or freeing one, we can derive the state of the 491 * other. That is, if we allocate a small block, and both were 492 * free, the remainder of the region must be split into blocks. 493 * If a block is freed, and its buddy is also free, then this 494 * triggers coalescing into a block of larger size. 495 * 496 * -- wli 497 */ 498 499static inline void __free_one_page(struct page *page, 500 struct zone *zone, unsigned int order, 501 int migratetype) 502{ 503 unsigned long page_idx; 504 unsigned long combined_idx; 505 struct page *buddy; 506 507 if (unlikely(PageCompound(page))) 508 if (unlikely(destroy_compound_page(page, order))) 509 return; 510 511 VM_BUG_ON(migratetype == -1); 512 513 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 514 515 VM_BUG_ON(page_idx & ((1 << order) - 1)); 516 VM_BUG_ON(bad_range(zone, page)); 517 518 while (order < MAX_ORDER-1) { 519 buddy = __page_find_buddy(page, page_idx, order); 520 if (!page_is_buddy(page, buddy, order)) 521 break; 522 523 /* Our buddy is free, merge with it and move up one order. */ 524 list_del(&buddy->lru); 525 zone->free_area[order].nr_free--; 526 rmv_page_order(buddy); 527 combined_idx = __find_combined_index(page_idx, order); 528 page = page + (combined_idx - page_idx); 529 page_idx = combined_idx; 530 order++; 531 } 532 set_page_order(page, order); 533 534 /* 535 * If this is not the largest possible page, check if the buddy 536 * of the next-highest order is free. If it is, it's possible 537 * that pages are being freed that will coalesce soon. In case, 538 * that is happening, add the free page to the tail of the list 539 * so it's less likely to be used soon and more likely to be merged 540 * as a higher order page 541 */ 542 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 543 struct page *higher_page, *higher_buddy; 544 combined_idx = __find_combined_index(page_idx, order); 545 higher_page = page + combined_idx - page_idx; 546 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1); 547 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 548 list_add_tail(&page->lru, 549 &zone->free_area[order].free_list[migratetype]); 550 goto out; 551 } 552 } 553 554 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 555out: 556 zone->free_area[order].nr_free++; 557} 558 559/* 560 * free_page_mlock() -- clean up attempts to free and mlocked() page. 561 * Page should not be on lru, so no need to fix that up. 562 * free_pages_check() will verify... 563 */ 564static inline void free_page_mlock(struct page *page) 565{ 566 __dec_zone_page_state(page, NR_MLOCK); 567 __count_vm_event(UNEVICTABLE_MLOCKFREED); 568} 569 570static inline int free_pages_check(struct page *page) 571{ 572 if (unlikely(page_mapcount(page) | 573 (page->mapping != NULL) | 574 (atomic_read(&page->_count) != 0) | 575 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 576 bad_page(page); 577 return 1; 578 } 579 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 580 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 581 return 0; 582} 583 584/* 585 * Frees a number of pages from the PCP lists 586 * Assumes all pages on list are in same zone, and of same order. 587 * count is the number of pages to free. 588 * 589 * If the zone was previously in an "all pages pinned" state then look to 590 * see if this freeing clears that state. 591 * 592 * And clear the zone's pages_scanned counter, to hold off the "all pages are 593 * pinned" detection logic. 594 */ 595static void free_pcppages_bulk(struct zone *zone, int count, 596 struct per_cpu_pages *pcp) 597{ 598 int migratetype = 0; 599 int batch_free = 0; 600 int to_free = count; 601 602 spin_lock(&zone->lock); 603 zone->all_unreclaimable = 0; 604 zone->pages_scanned = 0; 605 606 while (to_free) { 607 struct page *page; 608 struct list_head *list; 609 610 /* 611 * Remove pages from lists in a round-robin fashion. A 612 * batch_free count is maintained that is incremented when an 613 * empty list is encountered. This is so more pages are freed 614 * off fuller lists instead of spinning excessively around empty 615 * lists 616 */ 617 do { 618 batch_free++; 619 if (++migratetype == MIGRATE_PCPTYPES) 620 migratetype = 0; 621 list = &pcp->lists[migratetype]; 622 } while (list_empty(list)); 623 624 do { 625 page = list_entry(list->prev, struct page, lru); 626 /* must delete as __free_one_page list manipulates */ 627 list_del(&page->lru); 628 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 629 __free_one_page(page, zone, 0, page_private(page)); 630 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 631 } while (--to_free && --batch_free && !list_empty(list)); 632 } 633 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 634 spin_unlock(&zone->lock); 635} 636 637static void free_one_page(struct zone *zone, struct page *page, int order, 638 int migratetype) 639{ 640 spin_lock(&zone->lock); 641 zone->all_unreclaimable = 0; 642 zone->pages_scanned = 0; 643 644 __free_one_page(page, zone, order, migratetype); 645 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 646 spin_unlock(&zone->lock); 647} 648 649static bool free_pages_prepare(struct page *page, unsigned int order) 650{ 651 int i; 652 int bad = 0; 653 654 trace_mm_page_free_direct(page, order); 655 kmemcheck_free_shadow(page, order); 656 657 for (i = 0; i < (1 << order); i++) { 658 struct page *pg = page + i; 659 660 if (PageAnon(pg)) 661 pg->mapping = NULL; 662 bad += free_pages_check(pg); 663 } 664 if (bad) 665 return false; 666 667 if (!PageHighMem(page)) { 668 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 669 debug_check_no_obj_freed(page_address(page), 670 PAGE_SIZE << order); 671 } 672 arch_free_page(page, order); 673 kernel_map_pages(page, 1 << order, 0); 674 675 return true; 676} 677 678static void __free_pages_ok(struct page *page, unsigned int order) 679{ 680 unsigned long flags; 681 int wasMlocked = __TestClearPageMlocked(page); 682 683 if (!free_pages_prepare(page, order)) 684 return; 685 686 local_irq_save(flags); 687 if (unlikely(wasMlocked)) 688 free_page_mlock(page); 689 __count_vm_events(PGFREE, 1 << order); 690 free_one_page(page_zone(page), page, order, 691 get_pageblock_migratetype(page)); 692 local_irq_restore(flags); 693} 694 695/* 696 * permit the bootmem allocator to evade page validation on high-order frees 697 */ 698void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 699{ 700 if (order == 0) { 701 __ClearPageReserved(page); 702 set_page_count(page, 0); 703 set_page_refcounted(page); 704 __free_page(page); 705 } else { 706 int loop; 707 708 prefetchw(page); 709 for (loop = 0; loop < BITS_PER_LONG; loop++) { 710 struct page *p = &page[loop]; 711 712 if (loop + 1 < BITS_PER_LONG) 713 prefetchw(p + 1); 714 __ClearPageReserved(p); 715 set_page_count(p, 0); 716 } 717 718 set_page_refcounted(page); 719 __free_pages(page, order); 720 } 721} 722 723 724/* 725 * The order of subdivision here is critical for the IO subsystem. 726 * Please do not alter this order without good reasons and regression 727 * testing. Specifically, as large blocks of memory are subdivided, 728 * the order in which smaller blocks are delivered depends on the order 729 * they're subdivided in this function. This is the primary factor 730 * influencing the order in which pages are delivered to the IO 731 * subsystem according to empirical testing, and this is also justified 732 * by considering the behavior of a buddy system containing a single 733 * large block of memory acted on by a series of small allocations. 734 * This behavior is a critical factor in sglist merging's success. 735 * 736 * -- wli 737 */ 738static inline void expand(struct zone *zone, struct page *page, 739 int low, int high, struct free_area *area, 740 int migratetype) 741{ 742 unsigned long size = 1 << high; 743 744 while (high > low) { 745 area--; 746 high--; 747 size >>= 1; 748 VM_BUG_ON(bad_range(zone, &page[size])); 749 list_add(&page[size].lru, &area->free_list[migratetype]); 750 area->nr_free++; 751 set_page_order(&page[size], high); 752 } 753} 754 755/* 756 * This page is about to be returned from the page allocator 757 */ 758static inline int check_new_page(struct page *page) 759{ 760 if (unlikely(page_mapcount(page) | 761 (page->mapping != NULL) | 762 (atomic_read(&page->_count) != 0) | 763 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 764 bad_page(page); 765 return 1; 766 } 767 return 0; 768} 769 770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 771{ 772 int i; 773 774 for (i = 0; i < (1 << order); i++) { 775 struct page *p = page + i; 776 if (unlikely(check_new_page(p))) 777 return 1; 778 } 779 780 set_page_private(page, 0); 781 set_page_refcounted(page); 782 783 arch_alloc_page(page, order); 784 kernel_map_pages(page, 1 << order, 1); 785 786 if (gfp_flags & __GFP_ZERO) 787 prep_zero_page(page, order, gfp_flags); 788 789 if (order && (gfp_flags & __GFP_COMP)) 790 prep_compound_page(page, order); 791 792 return 0; 793} 794 795/* 796 * Go through the free lists for the given migratetype and remove 797 * the smallest available page from the freelists 798 */ 799static inline 800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 801 int migratetype) 802{ 803 unsigned int current_order; 804 struct free_area * area; 805 struct page *page; 806 807 /* Find a page of the appropriate size in the preferred list */ 808 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 809 area = &(zone->free_area[current_order]); 810 if (list_empty(&area->free_list[migratetype])) 811 continue; 812 813 page = list_entry(area->free_list[migratetype].next, 814 struct page, lru); 815 list_del(&page->lru); 816 rmv_page_order(page); 817 area->nr_free--; 818 expand(zone, page, order, current_order, area, migratetype); 819 return page; 820 } 821 822 return NULL; 823} 824 825 826/* 827 * This array describes the order lists are fallen back to when 828 * the free lists for the desirable migrate type are depleted 829 */ 830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 831 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 832 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 833 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 834 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 835}; 836 837/* 838 * Move the free pages in a range to the free lists of the requested type. 839 * Note that start_page and end_pages are not aligned on a pageblock 840 * boundary. If alignment is required, use move_freepages_block() 841 */ 842static int move_freepages(struct zone *zone, 843 struct page *start_page, struct page *end_page, 844 int migratetype) 845{ 846 struct page *page; 847 unsigned long order; 848 int pages_moved = 0; 849 850#ifndef CONFIG_HOLES_IN_ZONE 851 /* 852 * page_zone is not safe to call in this context when 853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 854 * anyway as we check zone boundaries in move_freepages_block(). 855 * Remove at a later date when no bug reports exist related to 856 * grouping pages by mobility 857 */ 858 BUG_ON(page_zone(start_page) != page_zone(end_page)); 859#endif 860 861 for (page = start_page; page <= end_page;) { 862 /* Make sure we are not inadvertently changing nodes */ 863 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 864 865 if (!pfn_valid_within(page_to_pfn(page))) { 866 page++; 867 continue; 868 } 869 870 if (!PageBuddy(page)) { 871 page++; 872 continue; 873 } 874 875 order = page_order(page); 876 list_del(&page->lru); 877 list_add(&page->lru, 878 &zone->free_area[order].free_list[migratetype]); 879 page += 1 << order; 880 pages_moved += 1 << order; 881 } 882 883 return pages_moved; 884} 885 886static int move_freepages_block(struct zone *zone, struct page *page, 887 int migratetype) 888{ 889 unsigned long start_pfn, end_pfn; 890 struct page *start_page, *end_page; 891 892 start_pfn = page_to_pfn(page); 893 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 894 start_page = pfn_to_page(start_pfn); 895 end_page = start_page + pageblock_nr_pages - 1; 896 end_pfn = start_pfn + pageblock_nr_pages - 1; 897 898 /* Do not cross zone boundaries */ 899 if (start_pfn < zone->zone_start_pfn) 900 start_page = page; 901 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 902 return 0; 903 904 return move_freepages(zone, start_page, end_page, migratetype); 905} 906 907static void change_pageblock_range(struct page *pageblock_page, 908 int start_order, int migratetype) 909{ 910 int nr_pageblocks = 1 << (start_order - pageblock_order); 911 912 while (nr_pageblocks--) { 913 set_pageblock_migratetype(pageblock_page, migratetype); 914 pageblock_page += pageblock_nr_pages; 915 } 916} 917 918/* Remove an element from the buddy allocator from the fallback list */ 919static inline struct page * 920__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 921{ 922 struct free_area * area; 923 int current_order; 924 struct page *page; 925 int migratetype, i; 926 927 /* Find the largest possible block of pages in the other list */ 928 for (current_order = MAX_ORDER-1; current_order >= order; 929 --current_order) { 930 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 931 migratetype = fallbacks[start_migratetype][i]; 932 933 /* MIGRATE_RESERVE handled later if necessary */ 934 if (migratetype == MIGRATE_RESERVE) 935 continue; 936 937 area = &(zone->free_area[current_order]); 938 if (list_empty(&area->free_list[migratetype])) 939 continue; 940 941 page = list_entry(area->free_list[migratetype].next, 942 struct page, lru); 943 area->nr_free--; 944 945 /* 946 * If breaking a large block of pages, move all free 947 * pages to the preferred allocation list. If falling 948 * back for a reclaimable kernel allocation, be more 949 * agressive about taking ownership of free pages 950 */ 951 if (unlikely(current_order >= (pageblock_order >> 1)) || 952 start_migratetype == MIGRATE_RECLAIMABLE || 953 page_group_by_mobility_disabled) { 954 unsigned long pages; 955 pages = move_freepages_block(zone, page, 956 start_migratetype); 957 958 /* Claim the whole block if over half of it is free */ 959 if (pages >= (1 << (pageblock_order-1)) || 960 page_group_by_mobility_disabled) 961 set_pageblock_migratetype(page, 962 start_migratetype); 963 964 migratetype = start_migratetype; 965 } 966 967 /* Remove the page from the freelists */ 968 list_del(&page->lru); 969 rmv_page_order(page); 970 971 /* Take ownership for orders >= pageblock_order */ 972 if (current_order >= pageblock_order) 973 change_pageblock_range(page, current_order, 974 start_migratetype); 975 976 expand(zone, page, order, current_order, area, migratetype); 977 978 trace_mm_page_alloc_extfrag(page, order, current_order, 979 start_migratetype, migratetype); 980 981 return page; 982 } 983 } 984 985 return NULL; 986} 987 988/* 989 * Do the hard work of removing an element from the buddy allocator. 990 * Call me with the zone->lock already held. 991 */ 992static struct page *__rmqueue(struct zone *zone, unsigned int order, 993 int migratetype) 994{ 995 struct page *page; 996 997retry_reserve: 998 page = __rmqueue_smallest(zone, order, migratetype); 999 1000 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1001 page = __rmqueue_fallback(zone, order, migratetype); 1002 1003 /* 1004 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1005 * is used because __rmqueue_smallest is an inline function 1006 * and we want just one call site 1007 */ 1008 if (!page) { 1009 migratetype = MIGRATE_RESERVE; 1010 goto retry_reserve; 1011 } 1012 } 1013 1014 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1015 return page; 1016} 1017 1018/* 1019 * Obtain a specified number of elements from the buddy allocator, all under 1020 * a single hold of the lock, for efficiency. Add them to the supplied list. 1021 * Returns the number of new pages which were placed at *list. 1022 */ 1023static int rmqueue_bulk(struct zone *zone, unsigned int order, 1024 unsigned long count, struct list_head *list, 1025 int migratetype, int cold) 1026{ 1027 int i; 1028 1029 spin_lock(&zone->lock); 1030 for (i = 0; i < count; ++i) { 1031 struct page *page = __rmqueue(zone, order, migratetype); 1032 if (unlikely(page == NULL)) 1033 break; 1034 1035 /* 1036 * Split buddy pages returned by expand() are received here 1037 * in physical page order. The page is added to the callers and 1038 * list and the list head then moves forward. From the callers 1039 * perspective, the linked list is ordered by page number in 1040 * some conditions. This is useful for IO devices that can 1041 * merge IO requests if the physical pages are ordered 1042 * properly. 1043 */ 1044 if (likely(cold == 0)) 1045 list_add(&page->lru, list); 1046 else 1047 list_add_tail(&page->lru, list); 1048 set_page_private(page, migratetype); 1049 list = &page->lru; 1050 } 1051 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1052 spin_unlock(&zone->lock); 1053 return i; 1054} 1055 1056#ifdef CONFIG_NUMA 1057/* 1058 * Called from the vmstat counter updater to drain pagesets of this 1059 * currently executing processor on remote nodes after they have 1060 * expired. 1061 * 1062 * Note that this function must be called with the thread pinned to 1063 * a single processor. 1064 */ 1065void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1066{ 1067 unsigned long flags; 1068 int to_drain; 1069 1070 local_irq_save(flags); 1071 if (pcp->count >= pcp->batch) 1072 to_drain = pcp->batch; 1073 else 1074 to_drain = pcp->count; 1075 free_pcppages_bulk(zone, to_drain, pcp); 1076 pcp->count -= to_drain; 1077 local_irq_restore(flags); 1078} 1079#endif 1080 1081/* 1082 * Drain pages of the indicated processor. 1083 * 1084 * The processor must either be the current processor and the 1085 * thread pinned to the current processor or a processor that 1086 * is not online. 1087 */ 1088static void drain_pages(unsigned int cpu) 1089{ 1090 unsigned long flags; 1091 struct zone *zone; 1092 1093 for_each_populated_zone(zone) { 1094 struct per_cpu_pageset *pset; 1095 struct per_cpu_pages *pcp; 1096 1097 local_irq_save(flags); 1098 pset = per_cpu_ptr(zone->pageset, cpu); 1099 1100 pcp = &pset->pcp; 1101 free_pcppages_bulk(zone, pcp->count, pcp); 1102 pcp->count = 0; 1103 local_irq_restore(flags); 1104 } 1105} 1106 1107/* 1108 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1109 */ 1110void drain_local_pages(void *arg) 1111{ 1112 drain_pages(smp_processor_id()); 1113} 1114 1115/* 1116 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1117 */ 1118void drain_all_pages(void) 1119{ 1120 on_each_cpu(drain_local_pages, NULL, 1); 1121} 1122 1123#ifdef CONFIG_HIBERNATION 1124 1125void mark_free_pages(struct zone *zone) 1126{ 1127 unsigned long pfn, max_zone_pfn; 1128 unsigned long flags; 1129 int order, t; 1130 struct list_head *curr; 1131 1132 if (!zone->spanned_pages) 1133 return; 1134 1135 spin_lock_irqsave(&zone->lock, flags); 1136 1137 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1138 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1139 if (pfn_valid(pfn)) { 1140 struct page *page = pfn_to_page(pfn); 1141 1142 if (!swsusp_page_is_forbidden(page)) 1143 swsusp_unset_page_free(page); 1144 } 1145 1146 for_each_migratetype_order(order, t) { 1147 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1148 unsigned long i; 1149 1150 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1151 for (i = 0; i < (1UL << order); i++) 1152 swsusp_set_page_free(pfn_to_page(pfn + i)); 1153 } 1154 } 1155 spin_unlock_irqrestore(&zone->lock, flags); 1156} 1157#endif /* CONFIG_PM */ 1158 1159/* 1160 * Free a 0-order page 1161 * cold == 1 ? free a cold page : free a hot page 1162 */ 1163void free_hot_cold_page(struct page *page, int cold) 1164{ 1165 struct zone *zone = page_zone(page); 1166 struct per_cpu_pages *pcp; 1167 unsigned long flags; 1168 int migratetype; 1169 int wasMlocked = __TestClearPageMlocked(page); 1170 1171 if (!free_pages_prepare(page, 0)) 1172 return; 1173 1174 migratetype = get_pageblock_migratetype(page); 1175 set_page_private(page, migratetype); 1176 local_irq_save(flags); 1177 if (unlikely(wasMlocked)) 1178 free_page_mlock(page); 1179 __count_vm_event(PGFREE); 1180 1181 /* 1182 * We only track unmovable, reclaimable and movable on pcp lists. 1183 * Free ISOLATE pages back to the allocator because they are being 1184 * offlined but treat RESERVE as movable pages so we can get those 1185 * areas back if necessary. Otherwise, we may have to free 1186 * excessively into the page allocator 1187 */ 1188 if (migratetype >= MIGRATE_PCPTYPES) { 1189 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1190 free_one_page(zone, page, 0, migratetype); 1191 goto out; 1192 } 1193 migratetype = MIGRATE_MOVABLE; 1194 } 1195 1196 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1197 if (cold) 1198 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1199 else 1200 list_add(&page->lru, &pcp->lists[migratetype]); 1201 pcp->count++; 1202 if (pcp->count >= pcp->high) { 1203 free_pcppages_bulk(zone, pcp->batch, pcp); 1204 pcp->count -= pcp->batch; 1205 } 1206 1207out: 1208 local_irq_restore(flags); 1209} 1210 1211/* 1212 * split_page takes a non-compound higher-order page, and splits it into 1213 * n (1<<order) sub-pages: page[0..n] 1214 * Each sub-page must be freed individually. 1215 * 1216 * Note: this is probably too low level an operation for use in drivers. 1217 * Please consult with lkml before using this in your driver. 1218 */ 1219void split_page(struct page *page, unsigned int order) 1220{ 1221 int i; 1222 1223 VM_BUG_ON(PageCompound(page)); 1224 VM_BUG_ON(!page_count(page)); 1225 1226#ifdef CONFIG_KMEMCHECK 1227 /* 1228 * Split shadow pages too, because free(page[0]) would 1229 * otherwise free the whole shadow. 1230 */ 1231 if (kmemcheck_page_is_tracked(page)) 1232 split_page(virt_to_page(page[0].shadow), order); 1233#endif 1234 1235 for (i = 1; i < (1 << order); i++) 1236 set_page_refcounted(page + i); 1237} 1238 1239/* 1240 * Similar to split_page except the page is already free. As this is only 1241 * being used for migration, the migratetype of the block also changes. 1242 * As this is called with interrupts disabled, the caller is responsible 1243 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1244 * are enabled. 1245 * 1246 * Note: this is probably too low level an operation for use in drivers. 1247 * Please consult with lkml before using this in your driver. 1248 */ 1249int split_free_page(struct page *page) 1250{ 1251 unsigned int order; 1252 unsigned long watermark; 1253 struct zone *zone; 1254 1255 BUG_ON(!PageBuddy(page)); 1256 1257 zone = page_zone(page); 1258 order = page_order(page); 1259 1260 /* Obey watermarks as if the page was being allocated */ 1261 watermark = low_wmark_pages(zone) + (1 << order); 1262 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1263 return 0; 1264 1265 /* Remove page from free list */ 1266 list_del(&page->lru); 1267 zone->free_area[order].nr_free--; 1268 rmv_page_order(page); 1269 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1270 1271 /* Split into individual pages */ 1272 set_page_refcounted(page); 1273 split_page(page, order); 1274 1275 if (order >= pageblock_order - 1) { 1276 struct page *endpage = page + (1 << order) - 1; 1277 for (; page < endpage; page += pageblock_nr_pages) 1278 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1279 } 1280 1281 return 1 << order; 1282} 1283 1284/* 1285 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1286 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1287 * or two. 1288 */ 1289static inline 1290struct page *buffered_rmqueue(struct zone *preferred_zone, 1291 struct zone *zone, int order, gfp_t gfp_flags, 1292 int migratetype) 1293{ 1294 unsigned long flags; 1295 struct page *page; 1296 int cold = !!(gfp_flags & __GFP_COLD); 1297 1298again: 1299 if (likely(order == 0)) { 1300 struct per_cpu_pages *pcp; 1301 struct list_head *list; 1302 1303 local_irq_save(flags); 1304 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1305 list = &pcp->lists[migratetype]; 1306 if (list_empty(list)) { 1307 pcp->count += rmqueue_bulk(zone, 0, 1308 pcp->batch, list, 1309 migratetype, cold); 1310 if (unlikely(list_empty(list))) 1311 goto failed; 1312 } 1313 1314 if (cold) 1315 page = list_entry(list->prev, struct page, lru); 1316 else 1317 page = list_entry(list->next, struct page, lru); 1318 1319 list_del(&page->lru); 1320 pcp->count--; 1321 } else { 1322 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1323 /* 1324 * __GFP_NOFAIL is not to be used in new code. 1325 * 1326 * All __GFP_NOFAIL callers should be fixed so that they 1327 * properly detect and handle allocation failures. 1328 * 1329 * We most definitely don't want callers attempting to 1330 * allocate greater than order-1 page units with 1331 * __GFP_NOFAIL. 1332 */ 1333 WARN_ON_ONCE(order > 1); 1334 } 1335 spin_lock_irqsave(&zone->lock, flags); 1336 page = __rmqueue(zone, order, migratetype); 1337 spin_unlock(&zone->lock); 1338 if (!page) 1339 goto failed; 1340 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1341 } 1342 1343 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1344 zone_statistics(preferred_zone, zone); 1345 local_irq_restore(flags); 1346 1347 VM_BUG_ON(bad_range(zone, page)); 1348 if (prep_new_page(page, order, gfp_flags)) 1349 goto again; 1350 return page; 1351 1352failed: 1353 local_irq_restore(flags); 1354 return NULL; 1355} 1356 1357/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1358#define ALLOC_WMARK_MIN WMARK_MIN 1359#define ALLOC_WMARK_LOW WMARK_LOW 1360#define ALLOC_WMARK_HIGH WMARK_HIGH 1361#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1362 1363/* Mask to get the watermark bits */ 1364#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1365 1366#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1367#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1368#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1369 1370#ifdef CONFIG_FAIL_PAGE_ALLOC 1371 1372static struct fail_page_alloc_attr { 1373 struct fault_attr attr; 1374 1375 u32 ignore_gfp_highmem; 1376 u32 ignore_gfp_wait; 1377 u32 min_order; 1378 1379#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1380 1381 struct dentry *ignore_gfp_highmem_file; 1382 struct dentry *ignore_gfp_wait_file; 1383 struct dentry *min_order_file; 1384 1385#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1386 1387} fail_page_alloc = { 1388 .attr = FAULT_ATTR_INITIALIZER, 1389 .ignore_gfp_wait = 1, 1390 .ignore_gfp_highmem = 1, 1391 .min_order = 1, 1392}; 1393 1394static int __init setup_fail_page_alloc(char *str) 1395{ 1396 return setup_fault_attr(&fail_page_alloc.attr, str); 1397} 1398__setup("fail_page_alloc=", setup_fail_page_alloc); 1399 1400static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1401{ 1402 if (order < fail_page_alloc.min_order) 1403 return 0; 1404 if (gfp_mask & __GFP_NOFAIL) 1405 return 0; 1406 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1407 return 0; 1408 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1409 return 0; 1410 1411 return should_fail(&fail_page_alloc.attr, 1 << order); 1412} 1413 1414#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1415 1416static int __init fail_page_alloc_debugfs(void) 1417{ 1418 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1419 struct dentry *dir; 1420 int err; 1421 1422 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1423 "fail_page_alloc"); 1424 if (err) 1425 return err; 1426 dir = fail_page_alloc.attr.dentries.dir; 1427 1428 fail_page_alloc.ignore_gfp_wait_file = 1429 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1430 &fail_page_alloc.ignore_gfp_wait); 1431 1432 fail_page_alloc.ignore_gfp_highmem_file = 1433 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1434 &fail_page_alloc.ignore_gfp_highmem); 1435 fail_page_alloc.min_order_file = 1436 debugfs_create_u32("min-order", mode, dir, 1437 &fail_page_alloc.min_order); 1438 1439 if (!fail_page_alloc.ignore_gfp_wait_file || 1440 !fail_page_alloc.ignore_gfp_highmem_file || 1441 !fail_page_alloc.min_order_file) { 1442 err = -ENOMEM; 1443 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1444 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1445 debugfs_remove(fail_page_alloc.min_order_file); 1446 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1447 } 1448 1449 return err; 1450} 1451 1452late_initcall(fail_page_alloc_debugfs); 1453 1454#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1455 1456#else /* CONFIG_FAIL_PAGE_ALLOC */ 1457 1458static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1459{ 1460 return 0; 1461} 1462 1463#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1464 1465/* 1466 * Return true if free pages are above 'mark'. This takes into account the order 1467 * of the allocation. 1468 */ 1469static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1470 int classzone_idx, int alloc_flags, long free_pages) 1471{ 1472 /* free_pages my go negative - that's OK */ 1473 long min = mark; 1474 int o; 1475 1476 free_pages -= (1 << order) + 1; 1477 if (alloc_flags & ALLOC_HIGH) 1478 min -= min / 2; 1479 if (alloc_flags & ALLOC_HARDER) 1480 min -= min / 4; 1481 1482 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1483 return false; 1484 for (o = 0; o < order; o++) { 1485 /* At the next order, this order's pages become unavailable */ 1486 free_pages -= z->free_area[o].nr_free << o; 1487 1488 /* Require fewer higher order pages to be free */ 1489 min >>= 1; 1490 1491 if (free_pages <= min) 1492 return false; 1493 } 1494 return true; 1495} 1496 1497bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1498 int classzone_idx, int alloc_flags) 1499{ 1500 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1501 zone_page_state(z, NR_FREE_PAGES)); 1502} 1503 1504bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1505 int classzone_idx, int alloc_flags) 1506{ 1507 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1508 1509 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1510 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1511 1512 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1513 free_pages); 1514} 1515 1516#ifdef CONFIG_NUMA 1517/* 1518 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1519 * skip over zones that are not allowed by the cpuset, or that have 1520 * been recently (in last second) found to be nearly full. See further 1521 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1522 * that have to skip over a lot of full or unallowed zones. 1523 * 1524 * If the zonelist cache is present in the passed in zonelist, then 1525 * returns a pointer to the allowed node mask (either the current 1526 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1527 * 1528 * If the zonelist cache is not available for this zonelist, does 1529 * nothing and returns NULL. 1530 * 1531 * If the fullzones BITMAP in the zonelist cache is stale (more than 1532 * a second since last zap'd) then we zap it out (clear its bits.) 1533 * 1534 * We hold off even calling zlc_setup, until after we've checked the 1535 * first zone in the zonelist, on the theory that most allocations will 1536 * be satisfied from that first zone, so best to examine that zone as 1537 * quickly as we can. 1538 */ 1539static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1540{ 1541 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1542 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1543 1544 zlc = zonelist->zlcache_ptr; 1545 if (!zlc) 1546 return NULL; 1547 1548 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1549 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1550 zlc->last_full_zap = jiffies; 1551 } 1552 1553 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1554 &cpuset_current_mems_allowed : 1555 &node_states[N_HIGH_MEMORY]; 1556 return allowednodes; 1557} 1558 1559/* 1560 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1561 * if it is worth looking at further for free memory: 1562 * 1) Check that the zone isn't thought to be full (doesn't have its 1563 * bit set in the zonelist_cache fullzones BITMAP). 1564 * 2) Check that the zones node (obtained from the zonelist_cache 1565 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1566 * Return true (non-zero) if zone is worth looking at further, or 1567 * else return false (zero) if it is not. 1568 * 1569 * This check -ignores- the distinction between various watermarks, 1570 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1571 * found to be full for any variation of these watermarks, it will 1572 * be considered full for up to one second by all requests, unless 1573 * we are so low on memory on all allowed nodes that we are forced 1574 * into the second scan of the zonelist. 1575 * 1576 * In the second scan we ignore this zonelist cache and exactly 1577 * apply the watermarks to all zones, even it is slower to do so. 1578 * We are low on memory in the second scan, and should leave no stone 1579 * unturned looking for a free page. 1580 */ 1581static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1582 nodemask_t *allowednodes) 1583{ 1584 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1585 int i; /* index of *z in zonelist zones */ 1586 int n; /* node that zone *z is on */ 1587 1588 zlc = zonelist->zlcache_ptr; 1589 if (!zlc) 1590 return 1; 1591 1592 i = z - zonelist->_zonerefs; 1593 n = zlc->z_to_n[i]; 1594 1595 /* This zone is worth trying if it is allowed but not full */ 1596 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1597} 1598 1599/* 1600 * Given 'z' scanning a zonelist, set the corresponding bit in 1601 * zlc->fullzones, so that subsequent attempts to allocate a page 1602 * from that zone don't waste time re-examining it. 1603 */ 1604static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1605{ 1606 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1607 int i; /* index of *z in zonelist zones */ 1608 1609 zlc = zonelist->zlcache_ptr; 1610 if (!zlc) 1611 return; 1612 1613 i = z - zonelist->_zonerefs; 1614 1615 set_bit(i, zlc->fullzones); 1616} 1617 1618#else /* CONFIG_NUMA */ 1619 1620static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1621{ 1622 return NULL; 1623} 1624 1625static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1626 nodemask_t *allowednodes) 1627{ 1628 return 1; 1629} 1630 1631static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1632{ 1633} 1634#endif /* CONFIG_NUMA */ 1635 1636/* 1637 * get_page_from_freelist goes through the zonelist trying to allocate 1638 * a page. 1639 */ 1640static struct page * BCMFASTPATH_HOST 1641get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1642 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1643 struct zone *preferred_zone, int migratetype) 1644{ 1645 struct zoneref *z; 1646 struct page *page = NULL; 1647 int classzone_idx; 1648 struct zone *zone; 1649 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1650 int zlc_active = 0; /* set if using zonelist_cache */ 1651 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1652 1653 classzone_idx = zone_idx(preferred_zone); 1654zonelist_scan: 1655 /* 1656 * Scan zonelist, looking for a zone with enough free. 1657 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1658 */ 1659 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1660 high_zoneidx, nodemask) { 1661 if (NUMA_BUILD && zlc_active && 1662 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1663 continue; 1664 if ((alloc_flags & ALLOC_CPUSET) && 1665 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1666 goto try_next_zone; 1667 1668 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1669 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1670 unsigned long mark; 1671 int ret; 1672 1673 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1674 if (zone_watermark_ok(zone, order, mark, 1675 classzone_idx, alloc_flags)) 1676 goto try_this_zone; 1677 1678 if (zone_reclaim_mode == 0) 1679 goto this_zone_full; 1680 1681 ret = zone_reclaim(zone, gfp_mask, order); 1682 switch (ret) { 1683 case ZONE_RECLAIM_NOSCAN: 1684 /* did not scan */ 1685 goto try_next_zone; 1686 case ZONE_RECLAIM_FULL: 1687 /* scanned but unreclaimable */ 1688 goto this_zone_full; 1689 default: 1690 /* did we reclaim enough */ 1691 if (!zone_watermark_ok(zone, order, mark, 1692 classzone_idx, alloc_flags)) 1693 goto this_zone_full; 1694 } 1695 } 1696 1697try_this_zone: 1698 page = buffered_rmqueue(preferred_zone, zone, order, 1699 gfp_mask, migratetype); 1700 if (page) 1701 break; 1702this_zone_full: 1703 if (NUMA_BUILD) 1704 zlc_mark_zone_full(zonelist, z); 1705try_next_zone: 1706 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1707 /* 1708 * we do zlc_setup after the first zone is tried but only 1709 * if there are multiple nodes make it worthwhile 1710 */ 1711 allowednodes = zlc_setup(zonelist, alloc_flags); 1712 zlc_active = 1; 1713 did_zlc_setup = 1; 1714 } 1715 } 1716 1717 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1718 /* Disable zlc cache for second zonelist scan */ 1719 zlc_active = 0; 1720 goto zonelist_scan; 1721 } 1722 return page; 1723} 1724 1725static inline int 1726should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1727 unsigned long pages_reclaimed) 1728{ 1729 /* Do not loop if specifically requested */ 1730 if (gfp_mask & __GFP_NORETRY) 1731 return 0; 1732 1733 /* 1734 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1735 * means __GFP_NOFAIL, but that may not be true in other 1736 * implementations. 1737 */ 1738 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1739 return 1; 1740 1741 /* 1742 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1743 * specified, then we retry until we no longer reclaim any pages 1744 * (above), or we've reclaimed an order of pages at least as 1745 * large as the allocation's order. In both cases, if the 1746 * allocation still fails, we stop retrying. 1747 */ 1748 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1749 return 1; 1750 1751 /* 1752 * Don't let big-order allocations loop unless the caller 1753 * explicitly requests that. 1754 */ 1755 if (gfp_mask & __GFP_NOFAIL) 1756 return 1; 1757 1758 return 0; 1759} 1760 1761static inline struct page * 1762__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1763 struct zonelist *zonelist, enum zone_type high_zoneidx, 1764 nodemask_t *nodemask, struct zone *preferred_zone, 1765 int migratetype) 1766{ 1767 struct page *page; 1768 1769 /* Acquire the OOM killer lock for the zones in zonelist */ 1770 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1771 schedule_timeout_uninterruptible(1); 1772 return NULL; 1773 } 1774 1775 /* 1776 * Go through the zonelist yet one more time, keep very high watermark 1777 * here, this is only to catch a parallel oom killing, we must fail if 1778 * we're still under heavy pressure. 1779 */ 1780 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1781 order, zonelist, high_zoneidx, 1782 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1783 preferred_zone, migratetype); 1784 if (page) 1785 goto out; 1786 1787 if (!(gfp_mask & __GFP_NOFAIL)) { 1788 /* The OOM killer will not help higher order allocs */ 1789 if (order > PAGE_ALLOC_COSTLY_ORDER) 1790 goto out; 1791 /* The OOM killer does not needlessly kill tasks for lowmem */ 1792 if (high_zoneidx < ZONE_NORMAL) 1793 goto out; 1794 /* 1795 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1796 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1797 * The caller should handle page allocation failure by itself if 1798 * it specifies __GFP_THISNODE. 1799 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1800 */ 1801 if (gfp_mask & __GFP_THISNODE) 1802 goto out; 1803 } 1804 /* Exhausted what can be done so it's blamo time */ 1805 out_of_memory(zonelist, gfp_mask, order, nodemask); 1806 1807out: 1808 clear_zonelist_oom(zonelist, gfp_mask); 1809 return page; 1810} 1811 1812#ifdef CONFIG_COMPACTION 1813/* Try memory compaction for high-order allocations before reclaim */ 1814static struct page * 1815__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1816 struct zonelist *zonelist, enum zone_type high_zoneidx, 1817 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1818 int migratetype, unsigned long *did_some_progress) 1819{ 1820 struct page *page; 1821 1822 if (!order || compaction_deferred(preferred_zone)) 1823 return NULL; 1824 1825 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 1826 nodemask); 1827 if (*did_some_progress != COMPACT_SKIPPED) { 1828 1829 /* Page migration frees to the PCP lists but we want merging */ 1830 drain_pages(get_cpu()); 1831 put_cpu(); 1832 1833 page = get_page_from_freelist(gfp_mask, nodemask, 1834 order, zonelist, high_zoneidx, 1835 alloc_flags, preferred_zone, 1836 migratetype); 1837 if (page) { 1838 preferred_zone->compact_considered = 0; 1839 preferred_zone->compact_defer_shift = 0; 1840 count_vm_event(COMPACTSUCCESS); 1841 return page; 1842 } 1843 1844 /* 1845 * It's bad if compaction run occurs and fails. 1846 * The most likely reason is that pages exist, 1847 * but not enough to satisfy watermarks. 1848 */ 1849 count_vm_event(COMPACTFAIL); 1850 defer_compaction(preferred_zone); 1851 1852 cond_resched(); 1853 } 1854 1855 return NULL; 1856} 1857#else 1858static inline struct page * 1859__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1860 struct zonelist *zonelist, enum zone_type high_zoneidx, 1861 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1862 int migratetype, unsigned long *did_some_progress) 1863{ 1864 return NULL; 1865} 1866#endif /* CONFIG_COMPACTION */ 1867 1868/* The really slow allocator path where we enter direct reclaim */ 1869static inline struct page * 1870__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1871 struct zonelist *zonelist, enum zone_type high_zoneidx, 1872 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1873 int migratetype, unsigned long *did_some_progress) 1874{ 1875 struct page *page = NULL; 1876 struct reclaim_state reclaim_state; 1877 struct task_struct *p = current; 1878 bool drained = false; 1879 1880 cond_resched(); 1881 1882 /* We now go into synchronous reclaim */ 1883 cpuset_memory_pressure_bump(); 1884 p->flags |= PF_MEMALLOC; 1885 lockdep_set_current_reclaim_state(gfp_mask); 1886 reclaim_state.reclaimed_slab = 0; 1887 p->reclaim_state = &reclaim_state; 1888 1889 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1890 1891 p->reclaim_state = NULL; 1892 lockdep_clear_current_reclaim_state(); 1893 p->flags &= ~PF_MEMALLOC; 1894 1895 cond_resched(); 1896 1897 if (unlikely(!(*did_some_progress))) 1898 return NULL; 1899 1900retry: 1901 page = get_page_from_freelist(gfp_mask, nodemask, order, 1902 zonelist, high_zoneidx, 1903 alloc_flags, preferred_zone, 1904 migratetype); 1905 1906 /* 1907 * If an allocation failed after direct reclaim, it could be because 1908 * pages are pinned on the per-cpu lists. Drain them and try again 1909 */ 1910 if (!page && !drained) { 1911 drain_all_pages(); 1912 drained = true; 1913 goto retry; 1914 } 1915 1916 return page; 1917} 1918 1919/* 1920 * This is called in the allocator slow-path if the allocation request is of 1921 * sufficient urgency to ignore watermarks and take other desperate measures 1922 */ 1923static inline struct page * 1924__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1925 struct zonelist *zonelist, enum zone_type high_zoneidx, 1926 nodemask_t *nodemask, struct zone *preferred_zone, 1927 int migratetype) 1928{ 1929 struct page *page; 1930 1931 do { 1932 page = get_page_from_freelist(gfp_mask, nodemask, order, 1933 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1934 preferred_zone, migratetype); 1935 1936 if (!page && gfp_mask & __GFP_NOFAIL) 1937 congestion_wait(BLK_RW_ASYNC, HZ/50); 1938 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1939 1940 return page; 1941} 1942 1943static inline 1944void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1945 enum zone_type high_zoneidx) 1946{ 1947 struct zoneref *z; 1948 struct zone *zone; 1949 1950 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1951 wakeup_kswapd(zone, order); 1952} 1953 1954static inline int 1955gfp_to_alloc_flags(gfp_t gfp_mask) 1956{ 1957 struct task_struct *p = current; 1958 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1959 const gfp_t wait = gfp_mask & __GFP_WAIT; 1960 1961 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1962 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1963 1964 /* 1965 * The caller may dip into page reserves a bit more if the caller 1966 * cannot run direct reclaim, or if the caller has realtime scheduling 1967 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1968 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1969 */ 1970 alloc_flags |= (gfp_mask & __GFP_HIGH); 1971 1972 if (!wait) { 1973 alloc_flags |= ALLOC_HARDER; 1974 /* 1975 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1976 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1977 */ 1978 alloc_flags &= ~ALLOC_CPUSET; 1979 } else if (unlikely(rt_task(p)) && !in_interrupt()) 1980 alloc_flags |= ALLOC_HARDER; 1981 1982 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1983 if (!in_interrupt() && 1984 ((p->flags & PF_MEMALLOC) || 1985 unlikely(test_thread_flag(TIF_MEMDIE)))) 1986 alloc_flags |= ALLOC_NO_WATERMARKS; 1987 } 1988 1989 return alloc_flags; 1990} 1991 1992static inline struct page * 1993__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1994 struct zonelist *zonelist, enum zone_type high_zoneidx, 1995 nodemask_t *nodemask, struct zone *preferred_zone, 1996 int migratetype) 1997{ 1998 const gfp_t wait = gfp_mask & __GFP_WAIT; 1999 struct page *page = NULL; 2000 int alloc_flags; 2001 unsigned long pages_reclaimed = 0; 2002 unsigned long did_some_progress; 2003 struct task_struct *p = current; 2004 2005 /* 2006 * In the slowpath, we sanity check order to avoid ever trying to 2007 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2008 * be using allocators in order of preference for an area that is 2009 * too large. 2010 */ 2011 if (order >= MAX_ORDER) { 2012 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2013 return NULL; 2014 } 2015 2016 /* 2017 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2018 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2019 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2020 * using a larger set of nodes after it has established that the 2021 * allowed per node queues are empty and that nodes are 2022 * over allocated. 2023 */ 2024 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2025 goto nopage; 2026 2027restart: 2028 wake_all_kswapd(order, zonelist, high_zoneidx); 2029 2030 /* 2031 * OK, we're below the kswapd watermark and have kicked background 2032 * reclaim. Now things get more complex, so set up alloc_flags according 2033 * to how we want to proceed. 2034 */ 2035 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2036 2037 /* This is the last chance, in general, before the goto nopage. */ 2038 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2039 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2040 preferred_zone, migratetype); 2041 if (page) 2042 goto got_pg; 2043 2044rebalance: 2045 /* Allocate without watermarks if the context allows */ 2046 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2047 page = __alloc_pages_high_priority(gfp_mask, order, 2048 zonelist, high_zoneidx, nodemask, 2049 preferred_zone, migratetype); 2050 if (page) 2051 goto got_pg; 2052 } 2053 2054 /* Atomic allocations - we can't balance anything */ 2055 if (!wait) 2056 goto nopage; 2057 2058 /* Avoid recursion of direct reclaim */ 2059 if (p->flags & PF_MEMALLOC) 2060 goto nopage; 2061 2062 /* Avoid allocations with no watermarks from looping endlessly */ 2063 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2064 goto nopage; 2065 2066 /* Try direct compaction */ 2067 page = __alloc_pages_direct_compact(gfp_mask, order, 2068 zonelist, high_zoneidx, 2069 nodemask, 2070 alloc_flags, preferred_zone, 2071 migratetype, &did_some_progress); 2072 if (page) 2073 goto got_pg; 2074 2075 /* Try direct reclaim and then allocating */ 2076 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2077 zonelist, high_zoneidx, 2078 nodemask, 2079 alloc_flags, preferred_zone, 2080 migratetype, &did_some_progress); 2081 if (page) 2082 goto got_pg; 2083 2084 /* 2085 * If we failed to make any progress reclaiming, then we are 2086 * running out of options and have to consider going OOM 2087 */ 2088 if (!did_some_progress) { 2089 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2090 if (oom_killer_disabled) 2091 goto nopage; 2092 page = __alloc_pages_may_oom(gfp_mask, order, 2093 zonelist, high_zoneidx, 2094 nodemask, preferred_zone, 2095 migratetype); 2096 if (page) 2097 goto got_pg; 2098 2099 if (!(gfp_mask & __GFP_NOFAIL)) { 2100 /* 2101 * The oom killer is not called for high-order 2102 * allocations that may fail, so if no progress 2103 * is being made, there are no other options and 2104 * retrying is unlikely to help. 2105 */ 2106 if (order > PAGE_ALLOC_COSTLY_ORDER) 2107 goto nopage; 2108 /* 2109 * The oom killer is not called for lowmem 2110 * allocations to prevent needlessly killing 2111 * innocent tasks. 2112 */ 2113 if (high_zoneidx < ZONE_NORMAL) 2114 goto nopage; 2115 } 2116 2117 goto restart; 2118 } 2119 } 2120 2121 /* Check if we should retry the allocation */ 2122 pages_reclaimed += did_some_progress; 2123 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 2124 /* Wait for some write requests to complete then retry */ 2125 congestion_wait(BLK_RW_ASYNC, HZ/50); 2126 goto rebalance; 2127 } 2128 2129nopage: 2130 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 2131 printk(KERN_WARNING "%s: page allocation failure." 2132 " order:%d, mode:0x%x\n", 2133 p->comm, order, gfp_mask); 2134 dump_stack(); 2135 show_mem(); 2136 } 2137 return page; 2138got_pg: 2139 if (kmemcheck_enabled) 2140 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2141 return page; 2142 2143} 2144 2145/* 2146 * This is the 'heart' of the zoned buddy allocator. 2147 */ 2148struct page * 2149__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2150 struct zonelist *zonelist, nodemask_t *nodemask) 2151{ 2152 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2153 struct zone *preferred_zone; 2154 struct page *page; 2155 int migratetype = allocflags_to_migratetype(gfp_mask); 2156 2157 gfp_mask &= gfp_allowed_mask; 2158 2159 lockdep_trace_alloc(gfp_mask); 2160 2161 might_sleep_if(gfp_mask & __GFP_WAIT); 2162 2163 if (should_fail_alloc_page(gfp_mask, order)) 2164 return NULL; 2165 2166 /* 2167 * Check the zones suitable for the gfp_mask contain at least one 2168 * valid zone. It's possible to have an empty zonelist as a result 2169 * of GFP_THISNODE and a memoryless node 2170 */ 2171 if (unlikely(!zonelist->_zonerefs->zone)) 2172 return NULL; 2173 2174 get_mems_allowed(); 2175 /* The preferred zone is used for statistics later */ 2176 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 2177 if (!preferred_zone) { 2178 put_mems_allowed(); 2179 return NULL; 2180 } 2181 2182 /* First allocation attempt */ 2183 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2184 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2185 preferred_zone, migratetype); 2186 if (unlikely(!page)) 2187 page = __alloc_pages_slowpath(gfp_mask, order, 2188 zonelist, high_zoneidx, nodemask, 2189 preferred_zone, migratetype); 2190 put_mems_allowed(); 2191 2192 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2193 return page; 2194} 2195EXPORT_SYMBOL(__alloc_pages_nodemask); 2196 2197/* 2198 * Common helper functions. 2199 */ 2200unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2201{ 2202 struct page *page; 2203 2204 /* 2205 * __get_free_pages() returns a 32-bit address, which cannot represent 2206 * a highmem page 2207 */ 2208 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2209 2210 page = alloc_pages(gfp_mask, order); 2211 if (!page) 2212 return 0; 2213 return (unsigned long) page_address(page); 2214} 2215EXPORT_SYMBOL(__get_free_pages); 2216 2217unsigned long get_zeroed_page(gfp_t gfp_mask) 2218{ 2219 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2220} 2221EXPORT_SYMBOL(get_zeroed_page); 2222 2223void __pagevec_free(struct pagevec *pvec) 2224{ 2225 int i = pagevec_count(pvec); 2226 2227 while (--i >= 0) { 2228 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2229 free_hot_cold_page(pvec->pages[i], pvec->cold); 2230 } 2231} 2232 2233void __free_pages(struct page *page, unsigned int order) 2234{ 2235 if (put_page_testzero(page)) { 2236 if (order == 0) 2237 free_hot_cold_page(page, 0); 2238 else 2239 __free_pages_ok(page, order); 2240 } 2241} 2242 2243EXPORT_SYMBOL(__free_pages); 2244 2245void free_pages(unsigned long addr, unsigned int order) 2246{ 2247 if (addr != 0) { 2248 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2249 __free_pages(virt_to_page((void *)addr), order); 2250 } 2251} 2252 2253EXPORT_SYMBOL(free_pages); 2254 2255/** 2256 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2257 * @size: the number of bytes to allocate 2258 * @gfp_mask: GFP flags for the allocation 2259 * 2260 * This function is similar to alloc_pages(), except that it allocates the 2261 * minimum number of pages to satisfy the request. alloc_pages() can only 2262 * allocate memory in power-of-two pages. 2263 * 2264 * This function is also limited by MAX_ORDER. 2265 * 2266 * Memory allocated by this function must be released by free_pages_exact(). 2267 */ 2268void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2269{ 2270 unsigned int order = get_order(size); 2271 unsigned long addr; 2272 2273 addr = __get_free_pages(gfp_mask, order); 2274 if (addr) { 2275 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2276 unsigned long used = addr + PAGE_ALIGN(size); 2277 2278 split_page(virt_to_page((void *)addr), order); 2279 while (used < alloc_end) { 2280 free_page(used); 2281 used += PAGE_SIZE; 2282 } 2283 } 2284 2285 return (void *)addr; 2286} 2287EXPORT_SYMBOL(alloc_pages_exact); 2288 2289/** 2290 * free_pages_exact - release memory allocated via alloc_pages_exact() 2291 * @virt: the value returned by alloc_pages_exact. 2292 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2293 * 2294 * Release the memory allocated by a previous call to alloc_pages_exact. 2295 */ 2296void free_pages_exact(void *virt, size_t size) 2297{ 2298 unsigned long addr = (unsigned long)virt; 2299 unsigned long end = addr + PAGE_ALIGN(size); 2300 2301 while (addr < end) { 2302 free_page(addr); 2303 addr += PAGE_SIZE; 2304 } 2305} 2306EXPORT_SYMBOL(free_pages_exact); 2307 2308static unsigned int nr_free_zone_pages(int offset) 2309{ 2310 struct zoneref *z; 2311 struct zone *zone; 2312 2313 /* Just pick one node, since fallback list is circular */ 2314 unsigned int sum = 0; 2315 2316 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2317 2318 for_each_zone_zonelist(zone, z, zonelist, offset) { 2319 unsigned long size = zone->present_pages; 2320 unsigned long high = high_wmark_pages(zone); 2321 if (size > high) 2322 sum += size - high; 2323 } 2324 2325 return sum; 2326} 2327 2328/* 2329 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2330 */ 2331unsigned int nr_free_buffer_pages(void) 2332{ 2333 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2334} 2335EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2336 2337/* 2338 * Amount of free RAM allocatable within all zones 2339 */ 2340unsigned int nr_free_pagecache_pages(void) 2341{ 2342 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2343} 2344 2345static inline void show_node(struct zone *zone) 2346{ 2347 if (NUMA_BUILD) 2348 printk("Node %d ", zone_to_nid(zone)); 2349} 2350 2351void si_meminfo(struct sysinfo *val) 2352{ 2353 val->totalram = totalram_pages; 2354 val->sharedram = 0; 2355 val->freeram = global_page_state(NR_FREE_PAGES); 2356 val->bufferram = nr_blockdev_pages(); 2357 val->totalhigh = totalhigh_pages; 2358 val->freehigh = nr_free_highpages(); 2359 val->mem_unit = PAGE_SIZE; 2360} 2361 2362EXPORT_SYMBOL(si_meminfo); 2363 2364#ifdef CONFIG_NUMA 2365void si_meminfo_node(struct sysinfo *val, int nid) 2366{ 2367 pg_data_t *pgdat = NODE_DATA(nid); 2368 2369 val->totalram = pgdat->node_present_pages; 2370 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2371#ifdef CONFIG_HIGHMEM 2372 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2373 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2374 NR_FREE_PAGES); 2375#else 2376 val->totalhigh = 0; 2377 val->freehigh = 0; 2378#endif 2379 val->mem_unit = PAGE_SIZE; 2380} 2381#endif 2382 2383#define K(x) ((x) << (PAGE_SHIFT-10)) 2384 2385/* 2386 * Show free area list (used inside shift_scroll-lock stuff) 2387 * We also calculate the percentage fragmentation. We do this by counting the 2388 * memory on each free list with the exception of the first item on the list. 2389 */ 2390void show_free_areas(void) 2391{ 2392 int cpu; 2393 struct zone *zone; 2394 2395 for_each_populated_zone(zone) { 2396 show_node(zone); 2397 printk("%s per-cpu:\n", zone->name); 2398 2399 for_each_online_cpu(cpu) { 2400 struct per_cpu_pageset *pageset; 2401 2402 pageset = per_cpu_ptr(zone->pageset, cpu); 2403 2404 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2405 cpu, pageset->pcp.high, 2406 pageset->pcp.batch, pageset->pcp.count); 2407 } 2408 } 2409 2410 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2411 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2412 " unevictable:%lu" 2413 " dirty:%lu writeback:%lu unstable:%lu\n" 2414 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2415 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2416 global_page_state(NR_ACTIVE_ANON), 2417 global_page_state(NR_INACTIVE_ANON), 2418 global_page_state(NR_ISOLATED_ANON), 2419 global_page_state(NR_ACTIVE_FILE), 2420 global_page_state(NR_INACTIVE_FILE), 2421 global_page_state(NR_ISOLATED_FILE), 2422 global_page_state(NR_UNEVICTABLE), 2423 global_page_state(NR_FILE_DIRTY), 2424 global_page_state(NR_WRITEBACK), 2425 global_page_state(NR_UNSTABLE_NFS), 2426 global_page_state(NR_FREE_PAGES), 2427 global_page_state(NR_SLAB_RECLAIMABLE), 2428 global_page_state(NR_SLAB_UNRECLAIMABLE), 2429 global_page_state(NR_FILE_MAPPED), 2430 global_page_state(NR_SHMEM), 2431 global_page_state(NR_PAGETABLE), 2432 global_page_state(NR_BOUNCE)); 2433 2434 for_each_populated_zone(zone) { 2435 int i; 2436 2437 show_node(zone); 2438 printk("%s" 2439 " free:%lukB" 2440 " min:%lukB" 2441 " low:%lukB" 2442 " high:%lukB" 2443 " active_anon:%lukB" 2444 " inactive_anon:%lukB" 2445 " active_file:%lukB" 2446 " inactive_file:%lukB" 2447 " unevictable:%lukB" 2448 " isolated(anon):%lukB" 2449 " isolated(file):%lukB" 2450 " present:%lukB" 2451 " mlocked:%lukB" 2452 " dirty:%lukB" 2453 " writeback:%lukB" 2454 " mapped:%lukB" 2455 " shmem:%lukB" 2456 " slab_reclaimable:%lukB" 2457 " slab_unreclaimable:%lukB" 2458 " kernel_stack:%lukB" 2459 " pagetables:%lukB" 2460 " unstable:%lukB" 2461 " bounce:%lukB" 2462 " writeback_tmp:%lukB" 2463 " pages_scanned:%lu" 2464 " all_unreclaimable? %s" 2465 "\n", 2466 zone->name, 2467 K(zone_page_state(zone, NR_FREE_PAGES)), 2468 K(min_wmark_pages(zone)), 2469 K(low_wmark_pages(zone)), 2470 K(high_wmark_pages(zone)), 2471 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2472 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2473 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2474 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2475 K(zone_page_state(zone, NR_UNEVICTABLE)), 2476 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2477 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2478 K(zone->present_pages), 2479 K(zone_page_state(zone, NR_MLOCK)), 2480 K(zone_page_state(zone, NR_FILE_DIRTY)), 2481 K(zone_page_state(zone, NR_WRITEBACK)), 2482 K(zone_page_state(zone, NR_FILE_MAPPED)), 2483 K(zone_page_state(zone, NR_SHMEM)), 2484 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2485 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2486 zone_page_state(zone, NR_KERNEL_STACK) * 2487 THREAD_SIZE / 1024, 2488 K(zone_page_state(zone, NR_PAGETABLE)), 2489 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2490 K(zone_page_state(zone, NR_BOUNCE)), 2491 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2492 zone->pages_scanned, 2493 (zone->all_unreclaimable ? "yes" : "no") 2494 ); 2495 printk("lowmem_reserve[]:"); 2496 for (i = 0; i < MAX_NR_ZONES; i++) 2497 printk(" %lu", zone->lowmem_reserve[i]); 2498 printk("\n"); 2499 } 2500 2501 for_each_populated_zone(zone) { 2502 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2503 2504 show_node(zone); 2505 printk("%s: ", zone->name); 2506 2507 spin_lock_irqsave(&zone->lock, flags); 2508 for (order = 0; order < MAX_ORDER; order++) { 2509 nr[order] = zone->free_area[order].nr_free; 2510 total += nr[order] << order; 2511 } 2512 spin_unlock_irqrestore(&zone->lock, flags); 2513 for (order = 0; order < MAX_ORDER; order++) 2514 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2515 printk("= %lukB\n", K(total)); 2516 } 2517 2518 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2519 2520 show_swap_cache_info(); 2521} 2522 2523static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2524{ 2525 zoneref->zone = zone; 2526 zoneref->zone_idx = zone_idx(zone); 2527} 2528 2529/* 2530 * Builds allocation fallback zone lists. 2531 * 2532 * Add all populated zones of a node to the zonelist. 2533 */ 2534static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2535 int nr_zones, enum zone_type zone_type) 2536{ 2537 struct zone *zone; 2538 2539 BUG_ON(zone_type >= MAX_NR_ZONES); 2540 zone_type++; 2541 2542 do { 2543 zone_type--; 2544 zone = pgdat->node_zones + zone_type; 2545 if (populated_zone(zone)) { 2546 zoneref_set_zone(zone, 2547 &zonelist->_zonerefs[nr_zones++]); 2548 check_highest_zone(zone_type); 2549 } 2550 2551 } while (zone_type); 2552 return nr_zones; 2553} 2554 2555 2556/* 2557 * zonelist_order: 2558 * 0 = automatic detection of better ordering. 2559 * 1 = order by ([node] distance, -zonetype) 2560 * 2 = order by (-zonetype, [node] distance) 2561 * 2562 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2563 * the same zonelist. So only NUMA can configure this param. 2564 */ 2565#define ZONELIST_ORDER_DEFAULT 0 2566#define ZONELIST_ORDER_NODE 1 2567#define ZONELIST_ORDER_ZONE 2 2568 2569/* zonelist order in the kernel. 2570 * set_zonelist_order() will set this to NODE or ZONE. 2571 */ 2572static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2573static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2574 2575 2576#ifdef CONFIG_NUMA 2577/* The value user specified ....changed by config */ 2578static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2579/* string for sysctl */ 2580#define NUMA_ZONELIST_ORDER_LEN 16 2581char numa_zonelist_order[16] = "default"; 2582 2583/* 2584 * interface for configure zonelist ordering. 2585 * command line option "numa_zonelist_order" 2586 * = "[dD]efault - default, automatic configuration. 2587 * = "[nN]ode - order by node locality, then by zone within node 2588 * = "[zZ]one - order by zone, then by locality within zone 2589 */ 2590 2591static int __parse_numa_zonelist_order(char *s) 2592{ 2593 if (*s == 'd' || *s == 'D') { 2594 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2595 } else if (*s == 'n' || *s == 'N') { 2596 user_zonelist_order = ZONELIST_ORDER_NODE; 2597 } else if (*s == 'z' || *s == 'Z') { 2598 user_zonelist_order = ZONELIST_ORDER_ZONE; 2599 } else { 2600 printk(KERN_WARNING 2601 "Ignoring invalid numa_zonelist_order value: " 2602 "%s\n", s); 2603 return -EINVAL; 2604 } 2605 return 0; 2606} 2607 2608static __init int setup_numa_zonelist_order(char *s) 2609{ 2610 if (s) 2611 return __parse_numa_zonelist_order(s); 2612 return 0; 2613} 2614early_param("numa_zonelist_order", setup_numa_zonelist_order); 2615 2616/* 2617 * sysctl handler for numa_zonelist_order 2618 */ 2619int numa_zonelist_order_handler(ctl_table *table, int write, 2620 void __user *buffer, size_t *length, 2621 loff_t *ppos) 2622{ 2623 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2624 int ret; 2625 static DEFINE_MUTEX(zl_order_mutex); 2626 2627 mutex_lock(&zl_order_mutex); 2628 if (write) 2629 strcpy(saved_string, (char*)table->data); 2630 ret = proc_dostring(table, write, buffer, length, ppos); 2631 if (ret) 2632 goto out; 2633 if (write) { 2634 int oldval = user_zonelist_order; 2635 if (__parse_numa_zonelist_order((char*)table->data)) { 2636 /* 2637 * bogus value. restore saved string 2638 */ 2639 strncpy((char*)table->data, saved_string, 2640 NUMA_ZONELIST_ORDER_LEN); 2641 user_zonelist_order = oldval; 2642 } else if (oldval != user_zonelist_order) { 2643 mutex_lock(&zonelists_mutex); 2644 build_all_zonelists(NULL); 2645 mutex_unlock(&zonelists_mutex); 2646 } 2647 } 2648out: 2649 mutex_unlock(&zl_order_mutex); 2650 return ret; 2651} 2652 2653 2654#define MAX_NODE_LOAD (nr_online_nodes) 2655static int node_load[MAX_NUMNODES]; 2656 2657/** 2658 * find_next_best_node - find the next node that should appear in a given node's fallback list 2659 * @node: node whose fallback list we're appending 2660 * @used_node_mask: nodemask_t of already used nodes 2661 * 2662 * We use a number of factors to determine which is the next node that should 2663 * appear on a given node's fallback list. The node should not have appeared 2664 * already in @node's fallback list, and it should be the next closest node 2665 * according to the distance array (which contains arbitrary distance values 2666 * from each node to each node in the system), and should also prefer nodes 2667 * with no CPUs, since presumably they'll have very little allocation pressure 2668 * on them otherwise. 2669 * It returns -1 if no node is found. 2670 */ 2671static int find_next_best_node(int node, nodemask_t *used_node_mask) 2672{ 2673 int n, val; 2674 int min_val = INT_MAX; 2675 int best_node = -1; 2676 const struct cpumask *tmp = cpumask_of_node(0); 2677 2678 /* Use the local node if we haven't already */ 2679 if (!node_isset(node, *used_node_mask)) { 2680 node_set(node, *used_node_mask); 2681 return node; 2682 } 2683 2684 for_each_node_state(n, N_HIGH_MEMORY) { 2685 2686 /* Don't want a node to appear more than once */ 2687 if (node_isset(n, *used_node_mask)) 2688 continue; 2689 2690 /* Use the distance array to find the distance */ 2691 val = node_distance(node, n); 2692 2693 /* Penalize nodes under us ("prefer the next node") */ 2694 val += (n < node); 2695 2696 /* Give preference to headless and unused nodes */ 2697 tmp = cpumask_of_node(n); 2698 if (!cpumask_empty(tmp)) 2699 val += PENALTY_FOR_NODE_WITH_CPUS; 2700 2701 /* Slight preference for less loaded node */ 2702 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2703 val += node_load[n]; 2704 2705 if (val < min_val) { 2706 min_val = val; 2707 best_node = n; 2708 } 2709 } 2710 2711 if (best_node >= 0) 2712 node_set(best_node, *used_node_mask); 2713 2714 return best_node; 2715} 2716 2717 2718/* 2719 * Build zonelists ordered by node and zones within node. 2720 * This results in maximum locality--normal zone overflows into local 2721 * DMA zone, if any--but risks exhausting DMA zone. 2722 */ 2723static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2724{ 2725 int j; 2726 struct zonelist *zonelist; 2727 2728 zonelist = &pgdat->node_zonelists[0]; 2729 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2730 ; 2731 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2732 MAX_NR_ZONES - 1); 2733 zonelist->_zonerefs[j].zone = NULL; 2734 zonelist->_zonerefs[j].zone_idx = 0; 2735} 2736 2737/* 2738 * Build gfp_thisnode zonelists 2739 */ 2740static void build_thisnode_zonelists(pg_data_t *pgdat) 2741{ 2742 int j; 2743 struct zonelist *zonelist; 2744 2745 zonelist = &pgdat->node_zonelists[1]; 2746 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2747 zonelist->_zonerefs[j].zone = NULL; 2748 zonelist->_zonerefs[j].zone_idx = 0; 2749} 2750 2751/* 2752 * Build zonelists ordered by zone and nodes within zones. 2753 * This results in conserving DMA zone[s] until all Normal memory is 2754 * exhausted, but results in overflowing to remote node while memory 2755 * may still exist in local DMA zone. 2756 */ 2757static int node_order[MAX_NUMNODES]; 2758 2759static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2760{ 2761 int pos, j, node; 2762 int zone_type; /* needs to be signed */ 2763 struct zone *z; 2764 struct zonelist *zonelist; 2765 2766 zonelist = &pgdat->node_zonelists[0]; 2767 pos = 0; 2768 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2769 for (j = 0; j < nr_nodes; j++) { 2770 node = node_order[j]; 2771 z = &NODE_DATA(node)->node_zones[zone_type]; 2772 if (populated_zone(z)) { 2773 zoneref_set_zone(z, 2774 &zonelist->_zonerefs[pos++]); 2775 check_highest_zone(zone_type); 2776 } 2777 } 2778 } 2779 zonelist->_zonerefs[pos].zone = NULL; 2780 zonelist->_zonerefs[pos].zone_idx = 0; 2781} 2782 2783static int default_zonelist_order(void) 2784{ 2785 int nid, zone_type; 2786 unsigned long low_kmem_size,total_size; 2787 struct zone *z; 2788 int average_size; 2789 /* 2790 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 2791 * If they are really small and used heavily, the system can fall 2792 * into OOM very easily. 2793 * This function detect ZONE_DMA/DMA32 size and configures zone order. 2794 */ 2795 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2796 low_kmem_size = 0; 2797 total_size = 0; 2798 for_each_online_node(nid) { 2799 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2800 z = &NODE_DATA(nid)->node_zones[zone_type]; 2801 if (populated_zone(z)) { 2802 if (zone_type < ZONE_NORMAL) 2803 low_kmem_size += z->present_pages; 2804 total_size += z->present_pages; 2805 } else if (zone_type == ZONE_NORMAL) { 2806 /* 2807 * If any node has only lowmem, then node order 2808 * is preferred to allow kernel allocations 2809 * locally; otherwise, they can easily infringe 2810 * on other nodes when there is an abundance of 2811 * lowmem available to allocate from. 2812 */ 2813 return ZONELIST_ORDER_NODE; 2814 } 2815 } 2816 } 2817 if (!low_kmem_size || /* there are no DMA area. */ 2818 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2819 return ZONELIST_ORDER_NODE; 2820 /* 2821 * look into each node's config. 2822 * If there is a node whose DMA/DMA32 memory is very big area on 2823 * local memory, NODE_ORDER may be suitable. 2824 */ 2825 average_size = total_size / 2826 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2827 for_each_online_node(nid) { 2828 low_kmem_size = 0; 2829 total_size = 0; 2830 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2831 z = &NODE_DATA(nid)->node_zones[zone_type]; 2832 if (populated_zone(z)) { 2833 if (zone_type < ZONE_NORMAL) 2834 low_kmem_size += z->present_pages; 2835 total_size += z->present_pages; 2836 } 2837 } 2838 if (low_kmem_size && 2839 total_size > average_size && /* ignore small node */ 2840 low_kmem_size > total_size * 70/100) 2841 return ZONELIST_ORDER_NODE; 2842 } 2843 return ZONELIST_ORDER_ZONE; 2844} 2845 2846static void set_zonelist_order(void) 2847{ 2848 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2849 current_zonelist_order = default_zonelist_order(); 2850 else 2851 current_zonelist_order = user_zonelist_order; 2852} 2853 2854static void build_zonelists(pg_data_t *pgdat) 2855{ 2856 int j, node, load; 2857 enum zone_type i; 2858 nodemask_t used_mask; 2859 int local_node, prev_node; 2860 struct zonelist *zonelist; 2861 int order = current_zonelist_order; 2862 2863 /* initialize zonelists */ 2864 for (i = 0; i < MAX_ZONELISTS; i++) { 2865 zonelist = pgdat->node_zonelists + i; 2866 zonelist->_zonerefs[0].zone = NULL; 2867 zonelist->_zonerefs[0].zone_idx = 0; 2868 } 2869 2870 /* NUMA-aware ordering of nodes */ 2871 local_node = pgdat->node_id; 2872 load = nr_online_nodes; 2873 prev_node = local_node; 2874 nodes_clear(used_mask); 2875 2876 memset(node_order, 0, sizeof(node_order)); 2877 j = 0; 2878 2879 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2880 int distance = node_distance(local_node, node); 2881 2882 /* 2883 * If another node is sufficiently far away then it is better 2884 * to reclaim pages in a zone before going off node. 2885 */ 2886 if (distance > RECLAIM_DISTANCE) 2887 zone_reclaim_mode = 1; 2888 2889 /* 2890 * We don't want to pressure a particular node. 2891 * So adding penalty to the first node in same 2892 * distance group to make it round-robin. 2893 */ 2894 if (distance != node_distance(local_node, prev_node)) 2895 node_load[node] = load; 2896 2897 prev_node = node; 2898 load--; 2899 if (order == ZONELIST_ORDER_NODE) 2900 build_zonelists_in_node_order(pgdat, node); 2901 else 2902 node_order[j++] = node; /* remember order */ 2903 } 2904 2905 if (order == ZONELIST_ORDER_ZONE) { 2906 /* calculate node order -- i.e., DMA last! */ 2907 build_zonelists_in_zone_order(pgdat, j); 2908 } 2909 2910 build_thisnode_zonelists(pgdat); 2911} 2912 2913/* Construct the zonelist performance cache - see further mmzone.h */ 2914static void build_zonelist_cache(pg_data_t *pgdat) 2915{ 2916 struct zonelist *zonelist; 2917 struct zonelist_cache *zlc; 2918 struct zoneref *z; 2919 2920 zonelist = &pgdat->node_zonelists[0]; 2921 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2922 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2923 for (z = zonelist->_zonerefs; z->zone; z++) 2924 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2925} 2926 2927#ifdef CONFIG_HAVE_MEMORYLESS_NODES 2928/* 2929 * Return node id of node used for "local" allocations. 2930 * I.e., first node id of first zone in arg node's generic zonelist. 2931 * Used for initializing percpu 'numa_mem', which is used primarily 2932 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 2933 */ 2934int local_memory_node(int node) 2935{ 2936 struct zone *zone; 2937 2938 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 2939 gfp_zone(GFP_KERNEL), 2940 NULL, 2941 &zone); 2942 return zone->node; 2943} 2944#endif 2945 2946#else /* CONFIG_NUMA */ 2947 2948static void set_zonelist_order(void) 2949{ 2950 current_zonelist_order = ZONELIST_ORDER_ZONE; 2951} 2952 2953static void build_zonelists(pg_data_t *pgdat) 2954{ 2955 int node, local_node; 2956 enum zone_type j; 2957 struct zonelist *zonelist; 2958 2959 local_node = pgdat->node_id; 2960 2961 zonelist = &pgdat->node_zonelists[0]; 2962 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2963 2964 /* 2965 * Now we build the zonelist so that it contains the zones 2966 * of all the other nodes. 2967 * We don't want to pressure a particular node, so when 2968 * building the zones for node N, we make sure that the 2969 * zones coming right after the local ones are those from 2970 * node N+1 (modulo N) 2971 */ 2972 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2973 if (!node_online(node)) 2974 continue; 2975 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2976 MAX_NR_ZONES - 1); 2977 } 2978 for (node = 0; node < local_node; node++) { 2979 if (!node_online(node)) 2980 continue; 2981 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2982 MAX_NR_ZONES - 1); 2983 } 2984 2985 zonelist->_zonerefs[j].zone = NULL; 2986 zonelist->_zonerefs[j].zone_idx = 0; 2987} 2988 2989/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2990static void build_zonelist_cache(pg_data_t *pgdat) 2991{ 2992 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2993} 2994 2995#endif /* CONFIG_NUMA */ 2996 2997/* 2998 * Boot pageset table. One per cpu which is going to be used for all 2999 * zones and all nodes. The parameters will be set in such a way 3000 * that an item put on a list will immediately be handed over to 3001 * the buddy list. This is safe since pageset manipulation is done 3002 * with interrupts disabled. 3003 * 3004 * The boot_pagesets must be kept even after bootup is complete for 3005 * unused processors and/or zones. They do play a role for bootstrapping 3006 * hotplugged processors. 3007 * 3008 * zoneinfo_show() and maybe other functions do 3009 * not check if the processor is online before following the pageset pointer. 3010 * Other parts of the kernel may not check if the zone is available. 3011 */ 3012static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3013static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3014static void setup_zone_pageset(struct zone *zone); 3015 3016/* 3017 * Global mutex to protect against size modification of zonelists 3018 * as well as to serialize pageset setup for the new populated zone. 3019 */ 3020DEFINE_MUTEX(zonelists_mutex); 3021 3022/* return values int ....just for stop_machine() */ 3023static __init_refok int __build_all_zonelists(void *data) 3024{ 3025 int nid; 3026 int cpu; 3027 3028#ifdef CONFIG_NUMA 3029 memset(node_load, 0, sizeof(node_load)); 3030#endif 3031 for_each_online_node(nid) { 3032 pg_data_t *pgdat = NODE_DATA(nid); 3033 3034 build_zonelists(pgdat); 3035 build_zonelist_cache(pgdat); 3036 } 3037 3038#ifdef CONFIG_MEMORY_HOTPLUG 3039 /* Setup real pagesets for the new zone */ 3040 if (data) { 3041 struct zone *zone = data; 3042 setup_zone_pageset(zone); 3043 } 3044#endif 3045 3046 /* 3047 * Initialize the boot_pagesets that are going to be used 3048 * for bootstrapping processors. The real pagesets for 3049 * each zone will be allocated later when the per cpu 3050 * allocator is available. 3051 * 3052 * boot_pagesets are used also for bootstrapping offline 3053 * cpus if the system is already booted because the pagesets 3054 * are needed to initialize allocators on a specific cpu too. 3055 * F.e. the percpu allocator needs the page allocator which 3056 * needs the percpu allocator in order to allocate its pagesets 3057 * (a chicken-egg dilemma). 3058 */ 3059 for_each_possible_cpu(cpu) { 3060 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3061 3062#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3063 /* 3064 * We now know the "local memory node" for each node-- 3065 * i.e., the node of the first zone in the generic zonelist. 3066 * Set up numa_mem percpu variable for on-line cpus. During 3067 * boot, only the boot cpu should be on-line; we'll init the 3068 * secondary cpus' numa_mem as they come on-line. During 3069 * node/memory hotplug, we'll fixup all on-line cpus. 3070 */ 3071 if (cpu_online(cpu)) 3072 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3073#endif 3074 } 3075 3076 return 0; 3077} 3078 3079/* 3080 * Called with zonelists_mutex held always 3081 * unless system_state == SYSTEM_BOOTING. 3082 */ 3083void build_all_zonelists(void *data) 3084{ 3085 set_zonelist_order(); 3086 3087 if (system_state == SYSTEM_BOOTING) { 3088 __build_all_zonelists(NULL); 3089 mminit_verify_zonelist(); 3090 cpuset_init_current_mems_allowed(); 3091 } else { 3092 /* we have to stop all cpus to guarantee there is no user 3093 of zonelist */ 3094 stop_machine(__build_all_zonelists, data, NULL); 3095 /* cpuset refresh routine should be here */ 3096 } 3097 vm_total_pages = nr_free_pagecache_pages(); 3098 /* 3099 * Disable grouping by mobility if the number of pages in the 3100 * system is too low to allow the mechanism to work. It would be 3101 * more accurate, but expensive to check per-zone. This check is 3102 * made on memory-hotadd so a system can start with mobility 3103 * disabled and enable it later 3104 */ 3105 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3106 page_group_by_mobility_disabled = 1; 3107 else 3108 page_group_by_mobility_disabled = 0; 3109 3110 printk("Built %i zonelists in %s order, mobility grouping %s. " 3111 "Total pages: %ld\n", 3112 nr_online_nodes, 3113 zonelist_order_name[current_zonelist_order], 3114 page_group_by_mobility_disabled ? "off" : "on", 3115 vm_total_pages); 3116#ifdef CONFIG_NUMA 3117 printk("Policy zone: %s\n", zone_names[policy_zone]); 3118#endif 3119} 3120 3121/* 3122 * Helper functions to size the waitqueue hash table. 3123 * Essentially these want to choose hash table sizes sufficiently 3124 * large so that collisions trying to wait on pages are rare. 3125 * But in fact, the number of active page waitqueues on typical 3126 * systems is ridiculously low, less than 200. So this is even 3127 * conservative, even though it seems large. 3128 * 3129 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3130 * waitqueues, i.e. the size of the waitq table given the number of pages. 3131 */ 3132#define PAGES_PER_WAITQUEUE 256 3133 3134#ifndef CONFIG_MEMORY_HOTPLUG 3135static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3136{ 3137 unsigned long size = 1; 3138 3139 pages /= PAGES_PER_WAITQUEUE; 3140 3141 while (size < pages) 3142 size <<= 1; 3143 3144 /* 3145 * Once we have dozens or even hundreds of threads sleeping 3146 * on IO we've got bigger problems than wait queue collision. 3147 * Limit the size of the wait table to a reasonable size. 3148 */ 3149 size = min(size, 4096UL); 3150 3151 return max(size, 4UL); 3152} 3153#else 3154/* 3155 * A zone's size might be changed by hot-add, so it is not possible to determine 3156 * a suitable size for its wait_table. So we use the maximum size now. 3157 * 3158 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3159 * 3160 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3161 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3162 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3163 * 3164 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3165 * or more by the traditional way. (See above). It equals: 3166 * 3167 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3168 * ia64(16K page size) : = ( 8G + 4M)byte. 3169 * powerpc (64K page size) : = (32G +16M)byte. 3170 */ 3171static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3172{ 3173 return 4096UL; 3174} 3175#endif 3176 3177/* 3178 * This is an integer logarithm so that shifts can be used later 3179 * to extract the more random high bits from the multiplicative 3180 * hash function before the remainder is taken. 3181 */ 3182static inline unsigned long wait_table_bits(unsigned long size) 3183{ 3184 return ffz(~size); 3185} 3186 3187#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3188 3189/* 3190 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3191 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3192 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3193 * higher will lead to a bigger reserve which will get freed as contiguous 3194 * blocks as reclaim kicks in 3195 */ 3196static void setup_zone_migrate_reserve(struct zone *zone) 3197{ 3198 unsigned long start_pfn, pfn, end_pfn; 3199 struct page *page; 3200 unsigned long block_migratetype; 3201 int reserve; 3202 3203 /* Get the start pfn, end pfn and the number of blocks to reserve */ 3204 start_pfn = zone->zone_start_pfn; 3205 end_pfn = start_pfn + zone->spanned_pages; 3206 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3207 pageblock_order; 3208 3209 /* 3210 * Reserve blocks are generally in place to help high-order atomic 3211 * allocations that are short-lived. A min_free_kbytes value that 3212 * would result in more than 2 reserve blocks for atomic allocations 3213 * is assumed to be in place to help anti-fragmentation for the 3214 * future allocation of hugepages at runtime. 3215 */ 3216 reserve = min(2, reserve); 3217 3218 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3219 if (!pfn_valid(pfn)) 3220 continue; 3221 page = pfn_to_page(pfn); 3222 3223 /* Watch out for overlapping nodes */ 3224 if (page_to_nid(page) != zone_to_nid(zone)) 3225 continue; 3226 3227 /* Blocks with reserved pages will never free, skip them. */ 3228 if (PageReserved(page)) 3229 continue; 3230 3231 block_migratetype = get_pageblock_migratetype(page); 3232 3233 /* If this block is reserved, account for it */ 3234 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 3235 reserve--; 3236 continue; 3237 } 3238 3239 /* Suitable for reserving if this block is movable */ 3240 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 3241 set_pageblock_migratetype(page, MIGRATE_RESERVE); 3242 move_freepages_block(zone, page, MIGRATE_RESERVE); 3243 reserve--; 3244 continue; 3245 } 3246 3247 /* 3248 * If the reserve is met and this is a previous reserved block, 3249 * take it back 3250 */ 3251 if (block_migratetype == MIGRATE_RESERVE) { 3252 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3253 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3254 } 3255 } 3256} 3257 3258/* 3259 * Initially all pages are reserved - free ones are freed 3260 * up by free_all_bootmem() once the early boot process is 3261 * done. Non-atomic initialization, single-pass. 3262 */ 3263void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3264 unsigned long start_pfn, enum memmap_context context) 3265{ 3266 struct page *page; 3267 unsigned long end_pfn = start_pfn + size; 3268 unsigned long pfn; 3269 struct zone *z; 3270 3271 if (highest_memmap_pfn < end_pfn - 1) 3272 highest_memmap_pfn = end_pfn - 1; 3273 3274 z = &NODE_DATA(nid)->node_zones[zone]; 3275 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3276 /* 3277 * There can be holes in boot-time mem_map[]s 3278 * handed to this function. They do not 3279 * exist on hotplugged memory. 3280 */ 3281 if (context == MEMMAP_EARLY) { 3282 if (!early_pfn_valid(pfn)) 3283 continue; 3284 if (!early_pfn_in_nid(pfn, nid)) 3285 continue; 3286 } 3287 page = pfn_to_page(pfn); 3288 set_page_links(page, zone, nid, pfn); 3289 mminit_verify_page_links(page, zone, nid, pfn); 3290 init_page_count(page); 3291 reset_page_mapcount(page); 3292 SetPageReserved(page); 3293 /* 3294 * Mark the block movable so that blocks are reserved for 3295 * movable at startup. This will force kernel allocations 3296 * to reserve their blocks rather than leaking throughout 3297 * the address space during boot when many long-lived 3298 * kernel allocations are made. Later some blocks near 3299 * the start are marked MIGRATE_RESERVE by 3300 * setup_zone_migrate_reserve() 3301 * 3302 * bitmap is created for zone's valid pfn range. but memmap 3303 * can be created for invalid pages (for alignment) 3304 * check here not to call set_pageblock_migratetype() against 3305 * pfn out of zone. 3306 */ 3307 if ((z->zone_start_pfn <= pfn) 3308 && (pfn < z->zone_start_pfn + z->spanned_pages) 3309 && !(pfn & (pageblock_nr_pages - 1))) 3310 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3311 3312 INIT_LIST_HEAD(&page->lru); 3313#ifdef WANT_PAGE_VIRTUAL 3314 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3315 if (!is_highmem_idx(zone)) 3316 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3317#endif 3318 } 3319} 3320 3321static void __meminit zone_init_free_lists(struct zone *zone) 3322{ 3323 int order, t; 3324 for_each_migratetype_order(order, t) { 3325 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3326 zone->free_area[order].nr_free = 0; 3327 } 3328} 3329 3330#ifndef __HAVE_ARCH_MEMMAP_INIT 3331#define memmap_init(size, nid, zone, start_pfn) \ 3332 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3333#endif 3334 3335static int zone_batchsize(struct zone *zone) 3336{ 3337#ifdef CONFIG_MMU 3338 int batch; 3339 3340 /* 3341 * The per-cpu-pages pools are set to around 1000th of the 3342 * size of the zone. But no more than 1/2 of a meg. 3343 * 3344 * OK, so we don't know how big the cache is. So guess. 3345 */ 3346 batch = zone->present_pages / 1024; 3347 if (batch * PAGE_SIZE > 512 * 1024) 3348 batch = (512 * 1024) / PAGE_SIZE; 3349 batch /= 4; /* We effectively *= 4 below */ 3350 if (batch < 1) 3351 batch = 1; 3352 3353 /* 3354 * Clamp the batch to a 2^n - 1 value. Having a power 3355 * of 2 value was found to be more likely to have 3356 * suboptimal cache aliasing properties in some cases. 3357 * 3358 * For example if 2 tasks are alternately allocating 3359 * batches of pages, one task can end up with a lot 3360 * of pages of one half of the possible page colors 3361 * and the other with pages of the other colors. 3362 */ 3363 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3364 3365 return batch; 3366 3367#else 3368 /* The deferral and batching of frees should be suppressed under NOMMU 3369 * conditions. 3370 * 3371 * The problem is that NOMMU needs to be able to allocate large chunks 3372 * of contiguous memory as there's no hardware page translation to 3373 * assemble apparent contiguous memory from discontiguous pages. 3374 * 3375 * Queueing large contiguous runs of pages for batching, however, 3376 * causes the pages to actually be freed in smaller chunks. As there 3377 * can be a significant delay between the individual batches being 3378 * recycled, this leads to the once large chunks of space being 3379 * fragmented and becoming unavailable for high-order allocations. 3380 */ 3381 return 0; 3382#endif 3383} 3384 3385static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3386{ 3387 struct per_cpu_pages *pcp; 3388 int migratetype; 3389 3390 memset(p, 0, sizeof(*p)); 3391 3392 pcp = &p->pcp; 3393 pcp->count = 0; 3394 pcp->high = 6 * batch; 3395 pcp->batch = max(1UL, 1 * batch); 3396 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3397 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3398} 3399 3400/* 3401 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3402 * to the value high for the pageset p. 3403 */ 3404 3405static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3406 unsigned long high) 3407{ 3408 struct per_cpu_pages *pcp; 3409 3410 pcp = &p->pcp; 3411 pcp->high = high; 3412 pcp->batch = max(1UL, high/4); 3413 if ((high/4) > (PAGE_SHIFT * 8)) 3414 pcp->batch = PAGE_SHIFT * 8; 3415} 3416 3417static __meminit void setup_zone_pageset(struct zone *zone) 3418{ 3419 int cpu; 3420 3421 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3422 3423 for_each_possible_cpu(cpu) { 3424 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3425 3426 setup_pageset(pcp, zone_batchsize(zone)); 3427 3428 if (percpu_pagelist_fraction) 3429 setup_pagelist_highmark(pcp, 3430 (zone->present_pages / 3431 percpu_pagelist_fraction)); 3432 } 3433} 3434 3435/* 3436 * Allocate per cpu pagesets and initialize them. 3437 * Before this call only boot pagesets were available. 3438 */ 3439void __init setup_per_cpu_pageset(void) 3440{ 3441 struct zone *zone; 3442 3443 for_each_populated_zone(zone) 3444 setup_zone_pageset(zone); 3445} 3446 3447static noinline __init_refok 3448int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3449{ 3450 int i; 3451 struct pglist_data *pgdat = zone->zone_pgdat; 3452 size_t alloc_size; 3453 3454 /* 3455 * The per-page waitqueue mechanism uses hashed waitqueues 3456 * per zone. 3457 */ 3458 zone->wait_table_hash_nr_entries = 3459 wait_table_hash_nr_entries(zone_size_pages); 3460 zone->wait_table_bits = 3461 wait_table_bits(zone->wait_table_hash_nr_entries); 3462 alloc_size = zone->wait_table_hash_nr_entries 3463 * sizeof(wait_queue_head_t); 3464 3465 if (!slab_is_available()) { 3466 zone->wait_table = (wait_queue_head_t *) 3467 alloc_bootmem_node(pgdat, alloc_size); 3468 } else { 3469 /* 3470 * This case means that a zone whose size was 0 gets new memory 3471 * via memory hot-add. 3472 * But it may be the case that a new node was hot-added. In 3473 * this case vmalloc() will not be able to use this new node's 3474 * memory - this wait_table must be initialized to use this new 3475 * node itself as well. 3476 * To use this new node's memory, further consideration will be 3477 * necessary. 3478 */ 3479 zone->wait_table = vmalloc(alloc_size); 3480 } 3481 if (!zone->wait_table) 3482 return -ENOMEM; 3483 3484 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3485 init_waitqueue_head(zone->wait_table + i); 3486 3487 return 0; 3488} 3489 3490static int __zone_pcp_update(void *data) 3491{ 3492 struct zone *zone = data; 3493 int cpu; 3494 unsigned long batch = zone_batchsize(zone), flags; 3495 3496 for_each_possible_cpu(cpu) { 3497 struct per_cpu_pageset *pset; 3498 struct per_cpu_pages *pcp; 3499 3500 pset = per_cpu_ptr(zone->pageset, cpu); 3501 pcp = &pset->pcp; 3502 3503 local_irq_save(flags); 3504 free_pcppages_bulk(zone, pcp->count, pcp); 3505 setup_pageset(pset, batch); 3506 local_irq_restore(flags); 3507 } 3508 return 0; 3509} 3510 3511void zone_pcp_update(struct zone *zone) 3512{ 3513 stop_machine(__zone_pcp_update, zone, NULL); 3514} 3515 3516static __meminit void zone_pcp_init(struct zone *zone) 3517{ 3518 /* 3519 * per cpu subsystem is not up at this point. The following code 3520 * relies on the ability of the linker to provide the 3521 * offset of a (static) per cpu variable into the per cpu area. 3522 */ 3523 zone->pageset = &boot_pageset; 3524 3525 if (zone->present_pages) 3526 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3527 zone->name, zone->present_pages, 3528 zone_batchsize(zone)); 3529} 3530 3531__meminit int init_currently_empty_zone(struct zone *zone, 3532 unsigned long zone_start_pfn, 3533 unsigned long size, 3534 enum memmap_context context) 3535{ 3536 struct pglist_data *pgdat = zone->zone_pgdat; 3537 int ret; 3538 ret = zone_wait_table_init(zone, size); 3539 if (ret) 3540 return ret; 3541 pgdat->nr_zones = zone_idx(zone) + 1; 3542 3543 zone->zone_start_pfn = zone_start_pfn; 3544 3545 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3546 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3547 pgdat->node_id, 3548 (unsigned long)zone_idx(zone), 3549 zone_start_pfn, (zone_start_pfn + size)); 3550 3551 zone_init_free_lists(zone); 3552 3553 return 0; 3554} 3555 3556#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3557/* 3558 * Basic iterator support. Return the first range of PFNs for a node 3559 * Note: nid == MAX_NUMNODES returns first region regardless of node 3560 */ 3561static int __meminit first_active_region_index_in_nid(int nid) 3562{ 3563 int i; 3564 3565 for (i = 0; i < nr_nodemap_entries; i++) 3566 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3567 return i; 3568 3569 return -1; 3570} 3571 3572/* 3573 * Basic iterator support. Return the next active range of PFNs for a node 3574 * Note: nid == MAX_NUMNODES returns next region regardless of node 3575 */ 3576static int __meminit next_active_region_index_in_nid(int index, int nid) 3577{ 3578 for (index = index + 1; index < nr_nodemap_entries; index++) 3579 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3580 return index; 3581 3582 return -1; 3583} 3584 3585#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3586/* 3587 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3588 * Architectures may implement their own version but if add_active_range() 3589 * was used and there are no special requirements, this is a convenient 3590 * alternative 3591 */ 3592int __meminit __early_pfn_to_nid(unsigned long pfn) 3593{ 3594 int i; 3595 3596 for (i = 0; i < nr_nodemap_entries; i++) { 3597 unsigned long start_pfn = early_node_map[i].start_pfn; 3598 unsigned long end_pfn = early_node_map[i].end_pfn; 3599 3600 if (start_pfn <= pfn && pfn < end_pfn) 3601 return early_node_map[i].nid; 3602 } 3603 /* This is a memory hole */ 3604 return -1; 3605} 3606#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3607 3608int __meminit early_pfn_to_nid(unsigned long pfn) 3609{ 3610 int nid; 3611 3612 nid = __early_pfn_to_nid(pfn); 3613 if (nid >= 0) 3614 return nid; 3615 /* just returns 0 */ 3616 return 0; 3617} 3618 3619#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3620bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3621{ 3622 int nid; 3623 3624 nid = __early_pfn_to_nid(pfn); 3625 if (nid >= 0 && nid != node) 3626 return false; 3627 return true; 3628} 3629#endif 3630 3631/* Basic iterator support to walk early_node_map[] */ 3632#define for_each_active_range_index_in_nid(i, nid) \ 3633 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3634 i = next_active_region_index_in_nid(i, nid)) 3635 3636/** 3637 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3638 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3639 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3640 * 3641 * If an architecture guarantees that all ranges registered with 3642 * add_active_ranges() contain no holes and may be freed, this 3643 * this function may be used instead of calling free_bootmem() manually. 3644 */ 3645void __init free_bootmem_with_active_regions(int nid, 3646 unsigned long max_low_pfn) 3647{ 3648 int i; 3649 3650 for_each_active_range_index_in_nid(i, nid) { 3651 unsigned long size_pages = 0; 3652 unsigned long end_pfn = early_node_map[i].end_pfn; 3653 3654 if (early_node_map[i].start_pfn >= max_low_pfn) 3655 continue; 3656 3657 if (end_pfn > max_low_pfn) 3658 end_pfn = max_low_pfn; 3659 3660 size_pages = end_pfn - early_node_map[i].start_pfn; 3661 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3662 PFN_PHYS(early_node_map[i].start_pfn), 3663 size_pages << PAGE_SHIFT); 3664 } 3665} 3666 3667int __init add_from_early_node_map(struct range *range, int az, 3668 int nr_range, int nid) 3669{ 3670 int i; 3671 u64 start, end; 3672 3673 /* need to go over early_node_map to find out good range for node */ 3674 for_each_active_range_index_in_nid(i, nid) { 3675 start = early_node_map[i].start_pfn; 3676 end = early_node_map[i].end_pfn; 3677 nr_range = add_range(range, az, nr_range, start, end); 3678 } 3679 return nr_range; 3680} 3681 3682#ifdef CONFIG_NO_BOOTMEM 3683void * __init __alloc_memory_core_early(int nid, u64 size, u64 align, 3684 u64 goal, u64 limit) 3685{ 3686 int i; 3687 void *ptr; 3688 3689 if (limit > get_max_mapped()) 3690 limit = get_max_mapped(); 3691 3692 /* need to go over early_node_map to find out good range for node */ 3693 for_each_active_range_index_in_nid(i, nid) { 3694 u64 addr; 3695 u64 ei_start, ei_last; 3696 3697 ei_last = early_node_map[i].end_pfn; 3698 ei_last <<= PAGE_SHIFT; 3699 ei_start = early_node_map[i].start_pfn; 3700 ei_start <<= PAGE_SHIFT; 3701 addr = find_early_area(ei_start, ei_last, 3702 goal, limit, size, align); 3703 3704 if (addr == -1ULL) 3705 continue; 3706 3707 3708 ptr = phys_to_virt(addr); 3709 memset(ptr, 0, size); 3710 reserve_early_without_check(addr, addr + size, "BOOTMEM"); 3711 /* 3712 * The min_count is set to 0 so that bootmem allocated blocks 3713 * are never reported as leaks. 3714 */ 3715 kmemleak_alloc(ptr, size, 0, 0); 3716 return ptr; 3717 } 3718 3719 return NULL; 3720} 3721#endif 3722 3723 3724void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3725{ 3726 int i; 3727 int ret; 3728 3729 for_each_active_range_index_in_nid(i, nid) { 3730 ret = work_fn(early_node_map[i].start_pfn, 3731 early_node_map[i].end_pfn, data); 3732 if (ret) 3733 break; 3734 } 3735} 3736/** 3737 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3738 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3739 * 3740 * If an architecture guarantees that all ranges registered with 3741 * add_active_ranges() contain no holes and may be freed, this 3742 * function may be used instead of calling memory_present() manually. 3743 */ 3744void __init sparse_memory_present_with_active_regions(int nid) 3745{ 3746 int i; 3747 3748 for_each_active_range_index_in_nid(i, nid) 3749 memory_present(early_node_map[i].nid, 3750 early_node_map[i].start_pfn, 3751 early_node_map[i].end_pfn); 3752} 3753 3754/** 3755 * get_pfn_range_for_nid - Return the start and end page frames for a node 3756 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3757 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3758 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3759 * 3760 * It returns the start and end page frame of a node based on information 3761 * provided by an arch calling add_active_range(). If called for a node 3762 * with no available memory, a warning is printed and the start and end 3763 * PFNs will be 0. 3764 */ 3765void __meminit get_pfn_range_for_nid(unsigned int nid, 3766 unsigned long *start_pfn, unsigned long *end_pfn) 3767{ 3768 int i; 3769 *start_pfn = -1UL; 3770 *end_pfn = 0; 3771 3772 for_each_active_range_index_in_nid(i, nid) { 3773 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3774 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3775 } 3776 3777 if (*start_pfn == -1UL) 3778 *start_pfn = 0; 3779} 3780 3781/* 3782 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3783 * assumption is made that zones within a node are ordered in monotonic 3784 * increasing memory addresses so that the "highest" populated zone is used 3785 */ 3786static void __init find_usable_zone_for_movable(void) 3787{ 3788 int zone_index; 3789 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3790 if (zone_index == ZONE_MOVABLE) 3791 continue; 3792 3793 if (arch_zone_highest_possible_pfn[zone_index] > 3794 arch_zone_lowest_possible_pfn[zone_index]) 3795 break; 3796 } 3797 3798 VM_BUG_ON(zone_index == -1); 3799 movable_zone = zone_index; 3800} 3801 3802/* 3803 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3804 * because it is sized independant of architecture. Unlike the other zones, 3805 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3806 * in each node depending on the size of each node and how evenly kernelcore 3807 * is distributed. This helper function adjusts the zone ranges 3808 * provided by the architecture for a given node by using the end of the 3809 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3810 * zones within a node are in order of monotonic increases memory addresses 3811 */ 3812static void __meminit adjust_zone_range_for_zone_movable(int nid, 3813 unsigned long zone_type, 3814 unsigned long node_start_pfn, 3815 unsigned long node_end_pfn, 3816 unsigned long *zone_start_pfn, 3817 unsigned long *zone_end_pfn) 3818{ 3819 /* Only adjust if ZONE_MOVABLE is on this node */ 3820 if (zone_movable_pfn[nid]) { 3821 /* Size ZONE_MOVABLE */ 3822 if (zone_type == ZONE_MOVABLE) { 3823 *zone_start_pfn = zone_movable_pfn[nid]; 3824 *zone_end_pfn = min(node_end_pfn, 3825 arch_zone_highest_possible_pfn[movable_zone]); 3826 3827 /* Adjust for ZONE_MOVABLE starting within this range */ 3828 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3829 *zone_end_pfn > zone_movable_pfn[nid]) { 3830 *zone_end_pfn = zone_movable_pfn[nid]; 3831 3832 /* Check if this whole range is within ZONE_MOVABLE */ 3833 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3834 *zone_start_pfn = *zone_end_pfn; 3835 } 3836} 3837 3838/* 3839 * Return the number of pages a zone spans in a node, including holes 3840 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3841 */ 3842static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3843 unsigned long zone_type, 3844 unsigned long *ignored) 3845{ 3846 unsigned long node_start_pfn, node_end_pfn; 3847 unsigned long zone_start_pfn, zone_end_pfn; 3848 3849 /* Get the start and end of the node and zone */ 3850 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3851 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3852 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3853 adjust_zone_range_for_zone_movable(nid, zone_type, 3854 node_start_pfn, node_end_pfn, 3855 &zone_start_pfn, &zone_end_pfn); 3856 3857 /* Check that this node has pages within the zone's required range */ 3858 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3859 return 0; 3860 3861 /* Move the zone boundaries inside the node if necessary */ 3862 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3863 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3864 3865 /* Return the spanned pages */ 3866 return zone_end_pfn - zone_start_pfn; 3867} 3868 3869/* 3870 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3871 * then all holes in the requested range will be accounted for. 3872 */ 3873unsigned long __meminit __absent_pages_in_range(int nid, 3874 unsigned long range_start_pfn, 3875 unsigned long range_end_pfn) 3876{ 3877 int i = 0; 3878 unsigned long prev_end_pfn = 0, hole_pages = 0; 3879 unsigned long start_pfn; 3880 3881 /* Find the end_pfn of the first active range of pfns in the node */ 3882 i = first_active_region_index_in_nid(nid); 3883 if (i == -1) 3884 return 0; 3885 3886 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3887 3888 /* Account for ranges before physical memory on this node */ 3889 if (early_node_map[i].start_pfn > range_start_pfn) 3890 hole_pages = prev_end_pfn - range_start_pfn; 3891 3892 /* Find all holes for the zone within the node */ 3893 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3894 3895 /* No need to continue if prev_end_pfn is outside the zone */ 3896 if (prev_end_pfn >= range_end_pfn) 3897 break; 3898 3899 /* Make sure the end of the zone is not within the hole */ 3900 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3901 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3902 3903 /* Update the hole size cound and move on */ 3904 if (start_pfn > range_start_pfn) { 3905 BUG_ON(prev_end_pfn > start_pfn); 3906 hole_pages += start_pfn - prev_end_pfn; 3907 } 3908 prev_end_pfn = early_node_map[i].end_pfn; 3909 } 3910 3911 /* Account for ranges past physical memory on this node */ 3912 if (range_end_pfn > prev_end_pfn) 3913 hole_pages += range_end_pfn - 3914 max(range_start_pfn, prev_end_pfn); 3915 3916 return hole_pages; 3917} 3918 3919/** 3920 * absent_pages_in_range - Return number of page frames in holes within a range 3921 * @start_pfn: The start PFN to start searching for holes 3922 * @end_pfn: The end PFN to stop searching for holes 3923 * 3924 * It returns the number of pages frames in memory holes within a range. 3925 */ 3926unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3927 unsigned long end_pfn) 3928{ 3929 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3930} 3931 3932/* Return the number of page frames in holes in a zone on a node */ 3933static unsigned long __meminit zone_absent_pages_in_node(int nid, 3934 unsigned long zone_type, 3935 unsigned long *ignored) 3936{ 3937 unsigned long node_start_pfn, node_end_pfn; 3938 unsigned long zone_start_pfn, zone_end_pfn; 3939 3940 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3941 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3942 node_start_pfn); 3943 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3944 node_end_pfn); 3945 3946 adjust_zone_range_for_zone_movable(nid, zone_type, 3947 node_start_pfn, node_end_pfn, 3948 &zone_start_pfn, &zone_end_pfn); 3949 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3950} 3951 3952#else 3953static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3954 unsigned long zone_type, 3955 unsigned long *zones_size) 3956{ 3957 return zones_size[zone_type]; 3958} 3959 3960static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3961 unsigned long zone_type, 3962 unsigned long *zholes_size) 3963{ 3964 if (!zholes_size) 3965 return 0; 3966 3967 return zholes_size[zone_type]; 3968} 3969 3970#endif 3971 3972static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3973 unsigned long *zones_size, unsigned long *zholes_size) 3974{ 3975 unsigned long realtotalpages, totalpages = 0; 3976 enum zone_type i; 3977 3978 for (i = 0; i < MAX_NR_ZONES; i++) 3979 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3980 zones_size); 3981 pgdat->node_spanned_pages = totalpages; 3982 3983 realtotalpages = totalpages; 3984 for (i = 0; i < MAX_NR_ZONES; i++) 3985 realtotalpages -= 3986 zone_absent_pages_in_node(pgdat->node_id, i, 3987 zholes_size); 3988 pgdat->node_present_pages = realtotalpages; 3989 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3990 realtotalpages); 3991} 3992 3993#ifndef CONFIG_SPARSEMEM 3994/* 3995 * Calculate the size of the zone->blockflags rounded to an unsigned long 3996 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3997 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3998 * round what is now in bits to nearest long in bits, then return it in 3999 * bytes. 4000 */ 4001static unsigned long __init usemap_size(unsigned long zonesize) 4002{ 4003 unsigned long usemapsize; 4004 4005 usemapsize = roundup(zonesize, pageblock_nr_pages); 4006 usemapsize = usemapsize >> pageblock_order; 4007 usemapsize *= NR_PAGEBLOCK_BITS; 4008 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4009 4010 return usemapsize / 8; 4011} 4012 4013static void __init setup_usemap(struct pglist_data *pgdat, 4014 struct zone *zone, unsigned long zonesize) 4015{ 4016 unsigned long usemapsize = usemap_size(zonesize); 4017 zone->pageblock_flags = NULL; 4018 if (usemapsize) 4019 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 4020} 4021#else 4022static void inline setup_usemap(struct pglist_data *pgdat, 4023 struct zone *zone, unsigned long zonesize) {} 4024#endif /* CONFIG_SPARSEMEM */ 4025 4026#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4027 4028/* Return a sensible default order for the pageblock size. */ 4029static inline int pageblock_default_order(void) 4030{ 4031 if (HPAGE_SHIFT > PAGE_SHIFT) 4032 return HUGETLB_PAGE_ORDER; 4033 4034 return MAX_ORDER-1; 4035} 4036 4037/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4038static inline void __init set_pageblock_order(unsigned int order) 4039{ 4040 /* Check that pageblock_nr_pages has not already been setup */ 4041 if (pageblock_order) 4042 return; 4043 4044 /* 4045 * Assume the largest contiguous order of interest is a huge page. 4046 * This value may be variable depending on boot parameters on IA64 4047 */ 4048 pageblock_order = order; 4049} 4050#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4051 4052/* 4053 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4054 * and pageblock_default_order() are unused as pageblock_order is set 4055 * at compile-time. See include/linux/pageblock-flags.h for the values of 4056 * pageblock_order based on the kernel config 4057 */ 4058static inline int pageblock_default_order(unsigned int order) 4059{ 4060 return MAX_ORDER-1; 4061} 4062#define set_pageblock_order(x) do {} while (0) 4063 4064#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4065 4066/* 4067 * Set up the zone data structures: 4068 * - mark all pages reserved 4069 * - mark all memory queues empty 4070 * - clear the memory bitmaps 4071 */ 4072static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4073 unsigned long *zones_size, unsigned long *zholes_size) 4074{ 4075 enum zone_type j; 4076 int nid = pgdat->node_id; 4077 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4078 int ret; 4079 4080 pgdat_resize_init(pgdat); 4081 pgdat->nr_zones = 0; 4082 init_waitqueue_head(&pgdat->kswapd_wait); 4083 pgdat->kswapd_max_order = 0; 4084 pgdat_page_cgroup_init(pgdat); 4085 4086 for (j = 0; j < MAX_NR_ZONES; j++) { 4087 struct zone *zone = pgdat->node_zones + j; 4088 unsigned long size, realsize, memmap_pages; 4089 enum lru_list l; 4090 4091 size = zone_spanned_pages_in_node(nid, j, zones_size); 4092 realsize = size - zone_absent_pages_in_node(nid, j, 4093 zholes_size); 4094 4095 /* 4096 * Adjust realsize so that it accounts for how much memory 4097 * is used by this zone for memmap. This affects the watermark 4098 * and per-cpu initialisations 4099 */ 4100 memmap_pages = 4101 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4102 if (realsize >= memmap_pages) { 4103 realsize -= memmap_pages; 4104 if (memmap_pages) 4105 printk(KERN_DEBUG 4106 " %s zone: %lu pages used for memmap\n", 4107 zone_names[j], memmap_pages); 4108 } else 4109 printk(KERN_WARNING 4110 " %s zone: %lu pages exceeds realsize %lu\n", 4111 zone_names[j], memmap_pages, realsize); 4112 4113 /* Account for reserved pages */ 4114 if (j == 0 && realsize > dma_reserve) { 4115 realsize -= dma_reserve; 4116 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4117 zone_names[0], dma_reserve); 4118 } 4119 4120 if (!is_highmem_idx(j)) 4121 nr_kernel_pages += realsize; 4122 nr_all_pages += realsize; 4123 4124 zone->spanned_pages = size; 4125 zone->present_pages = realsize; 4126#ifdef CONFIG_NUMA 4127 zone->node = nid; 4128 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4129 / 100; 4130 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4131#endif 4132 zone->name = zone_names[j]; 4133 spin_lock_init(&zone->lock); 4134 spin_lock_init(&zone->lru_lock); 4135 zone_seqlock_init(zone); 4136 zone->zone_pgdat = pgdat; 4137 4138 zone_pcp_init(zone); 4139 for_each_lru(l) { 4140 INIT_LIST_HEAD(&zone->lru[l].list); 4141 zone->reclaim_stat.nr_saved_scan[l] = 0; 4142 } 4143 zone->reclaim_stat.recent_rotated[0] = 0; 4144 zone->reclaim_stat.recent_rotated[1] = 0; 4145 zone->reclaim_stat.recent_scanned[0] = 0; 4146 zone->reclaim_stat.recent_scanned[1] = 0; 4147 zap_zone_vm_stats(zone); 4148 zone->flags = 0; 4149 if (!size) 4150 continue; 4151 4152 set_pageblock_order(pageblock_default_order()); 4153 setup_usemap(pgdat, zone, size); 4154 ret = init_currently_empty_zone(zone, zone_start_pfn, 4155 size, MEMMAP_EARLY); 4156 BUG_ON(ret); 4157 memmap_init(size, nid, j, zone_start_pfn); 4158 zone_start_pfn += size; 4159 } 4160} 4161 4162static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4163{ 4164 /* Skip empty nodes */ 4165 if (!pgdat->node_spanned_pages) 4166 return; 4167 4168#ifdef CONFIG_FLAT_NODE_MEM_MAP 4169 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4170 if (!pgdat->node_mem_map) { 4171 unsigned long size, start, end; 4172 struct page *map; 4173 4174 /* 4175 * The zone's endpoints aren't required to be MAX_ORDER 4176 * aligned but the node_mem_map endpoints must be in order 4177 * for the buddy allocator to function correctly. 4178 */ 4179 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4180 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4181 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4182 size = (end - start) * sizeof(struct page); 4183 map = alloc_remap(pgdat->node_id, size); 4184 if (!map) 4185 map = alloc_bootmem_node(pgdat, size); 4186 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4187 } 4188#ifndef CONFIG_NEED_MULTIPLE_NODES 4189 /* 4190 * With no DISCONTIG, the global mem_map is just set as node 0's 4191 */ 4192 if (pgdat == NODE_DATA(0)) { 4193 mem_map = NODE_DATA(0)->node_mem_map; 4194#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4195 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4196 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4197#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4198 } 4199#endif 4200#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4201} 4202 4203void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4204 unsigned long node_start_pfn, unsigned long *zholes_size) 4205{ 4206 pg_data_t *pgdat = NODE_DATA(nid); 4207 4208 pgdat->node_id = nid; 4209 pgdat->node_start_pfn = node_start_pfn; 4210 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4211 4212 alloc_node_mem_map(pgdat); 4213#ifdef CONFIG_FLAT_NODE_MEM_MAP 4214 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4215 nid, (unsigned long)pgdat, 4216 (unsigned long)pgdat->node_mem_map); 4217#endif 4218 4219 free_area_init_core(pgdat, zones_size, zholes_size); 4220} 4221 4222#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4223 4224#if MAX_NUMNODES > 1 4225/* 4226 * Figure out the number of possible node ids. 4227 */ 4228static void __init setup_nr_node_ids(void) 4229{ 4230 unsigned int node; 4231 unsigned int highest = 0; 4232 4233 for_each_node_mask(node, node_possible_map) 4234 highest = node; 4235 nr_node_ids = highest + 1; 4236} 4237#else 4238static inline void setup_nr_node_ids(void) 4239{ 4240} 4241#endif 4242 4243/** 4244 * add_active_range - Register a range of PFNs backed by physical memory 4245 * @nid: The node ID the range resides on 4246 * @start_pfn: The start PFN of the available physical memory 4247 * @end_pfn: The end PFN of the available physical memory 4248 * 4249 * These ranges are stored in an early_node_map[] and later used by 4250 * free_area_init_nodes() to calculate zone sizes and holes. If the 4251 * range spans a memory hole, it is up to the architecture to ensure 4252 * the memory is not freed by the bootmem allocator. If possible 4253 * the range being registered will be merged with existing ranges. 4254 */ 4255void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4256 unsigned long end_pfn) 4257{ 4258 int i; 4259 4260 mminit_dprintk(MMINIT_TRACE, "memory_register", 4261 "Entering add_active_range(%d, %#lx, %#lx) " 4262 "%d entries of %d used\n", 4263 nid, start_pfn, end_pfn, 4264 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4265 4266 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4267 4268 /* Merge with existing active regions if possible */ 4269 for (i = 0; i < nr_nodemap_entries; i++) { 4270 if (early_node_map[i].nid != nid) 4271 continue; 4272 4273 /* Skip if an existing region covers this new one */ 4274 if (start_pfn >= early_node_map[i].start_pfn && 4275 end_pfn <= early_node_map[i].end_pfn) 4276 return; 4277 4278 /* Merge forward if suitable */ 4279 if (start_pfn <= early_node_map[i].end_pfn && 4280 end_pfn > early_node_map[i].end_pfn) { 4281 early_node_map[i].end_pfn = end_pfn; 4282 return; 4283 } 4284 4285 /* Merge backward if suitable */ 4286 if (start_pfn < early_node_map[i].start_pfn && 4287 end_pfn >= early_node_map[i].start_pfn) { 4288 early_node_map[i].start_pfn = start_pfn; 4289 return; 4290 } 4291 } 4292 4293 /* Check that early_node_map is large enough */ 4294 if (i >= MAX_ACTIVE_REGIONS) { 4295 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4296 MAX_ACTIVE_REGIONS); 4297 return; 4298 } 4299 4300 early_node_map[i].nid = nid; 4301 early_node_map[i].start_pfn = start_pfn; 4302 early_node_map[i].end_pfn = end_pfn; 4303 nr_nodemap_entries = i + 1; 4304} 4305 4306/** 4307 * remove_active_range - Shrink an existing registered range of PFNs 4308 * @nid: The node id the range is on that should be shrunk 4309 * @start_pfn: The new PFN of the range 4310 * @end_pfn: The new PFN of the range 4311 * 4312 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4313 * The map is kept near the end physical page range that has already been 4314 * registered. This function allows an arch to shrink an existing registered 4315 * range. 4316 */ 4317void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4318 unsigned long end_pfn) 4319{ 4320 int i, j; 4321 int removed = 0; 4322 4323 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4324 nid, start_pfn, end_pfn); 4325 4326 /* Find the old active region end and shrink */ 4327 for_each_active_range_index_in_nid(i, nid) { 4328 if (early_node_map[i].start_pfn >= start_pfn && 4329 early_node_map[i].end_pfn <= end_pfn) { 4330 /* clear it */ 4331 early_node_map[i].start_pfn = 0; 4332 early_node_map[i].end_pfn = 0; 4333 removed = 1; 4334 continue; 4335 } 4336 if (early_node_map[i].start_pfn < start_pfn && 4337 early_node_map[i].end_pfn > start_pfn) { 4338 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4339 early_node_map[i].end_pfn = start_pfn; 4340 if (temp_end_pfn > end_pfn) 4341 add_active_range(nid, end_pfn, temp_end_pfn); 4342 continue; 4343 } 4344 if (early_node_map[i].start_pfn >= start_pfn && 4345 early_node_map[i].end_pfn > end_pfn && 4346 early_node_map[i].start_pfn < end_pfn) { 4347 early_node_map[i].start_pfn = end_pfn; 4348 continue; 4349 } 4350 } 4351 4352 if (!removed) 4353 return; 4354 4355 /* remove the blank ones */ 4356 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4357 if (early_node_map[i].nid != nid) 4358 continue; 4359 if (early_node_map[i].end_pfn) 4360 continue; 4361 /* we found it, get rid of it */ 4362 for (j = i; j < nr_nodemap_entries - 1; j++) 4363 memcpy(&early_node_map[j], &early_node_map[j+1], 4364 sizeof(early_node_map[j])); 4365 j = nr_nodemap_entries - 1; 4366 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4367 nr_nodemap_entries--; 4368 } 4369} 4370 4371/** 4372 * remove_all_active_ranges - Remove all currently registered regions 4373 * 4374 * During discovery, it may be found that a table like SRAT is invalid 4375 * and an alternative discovery method must be used. This function removes 4376 * all currently registered regions. 4377 */ 4378void __init remove_all_active_ranges(void) 4379{ 4380 memset(early_node_map, 0, sizeof(early_node_map)); 4381 nr_nodemap_entries = 0; 4382} 4383 4384/* Compare two active node_active_regions */ 4385static int __init cmp_node_active_region(const void *a, const void *b) 4386{ 4387 struct node_active_region *arange = (struct node_active_region *)a; 4388 struct node_active_region *brange = (struct node_active_region *)b; 4389 4390 /* Done this way to avoid overflows */ 4391 if (arange->start_pfn > brange->start_pfn) 4392 return 1; 4393 if (arange->start_pfn < brange->start_pfn) 4394 return -1; 4395 4396 return 0; 4397} 4398 4399/* sort the node_map by start_pfn */ 4400void __init sort_node_map(void) 4401{ 4402 sort(early_node_map, (size_t)nr_nodemap_entries, 4403 sizeof(struct node_active_region), 4404 cmp_node_active_region, NULL); 4405} 4406 4407/* Find the lowest pfn for a node */ 4408static unsigned long __init find_min_pfn_for_node(int nid) 4409{ 4410 int i; 4411 unsigned long min_pfn = ULONG_MAX; 4412 4413 /* Assuming a sorted map, the first range found has the starting pfn */ 4414 for_each_active_range_index_in_nid(i, nid) 4415 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4416 4417 if (min_pfn == ULONG_MAX) { 4418 printk(KERN_WARNING 4419 "Could not find start_pfn for node %d\n", nid); 4420 return 0; 4421 } 4422 4423 return min_pfn; 4424} 4425 4426/** 4427 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4428 * 4429 * It returns the minimum PFN based on information provided via 4430 * add_active_range(). 4431 */ 4432unsigned long __init find_min_pfn_with_active_regions(void) 4433{ 4434 return find_min_pfn_for_node(MAX_NUMNODES); 4435} 4436 4437/* 4438 * early_calculate_totalpages() 4439 * Sum pages in active regions for movable zone. 4440 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4441 */ 4442static unsigned long __init early_calculate_totalpages(void) 4443{ 4444 int i; 4445 unsigned long totalpages = 0; 4446 4447 for (i = 0; i < nr_nodemap_entries; i++) { 4448 unsigned long pages = early_node_map[i].end_pfn - 4449 early_node_map[i].start_pfn; 4450 totalpages += pages; 4451 if (pages) 4452 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4453 } 4454 return totalpages; 4455} 4456 4457/* 4458 * Find the PFN the Movable zone begins in each node. Kernel memory 4459 * is spread evenly between nodes as long as the nodes have enough 4460 * memory. When they don't, some nodes will have more kernelcore than 4461 * others 4462 */ 4463static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4464{ 4465 int i, nid; 4466 unsigned long usable_startpfn; 4467 unsigned long kernelcore_node, kernelcore_remaining; 4468 /* save the state before borrow the nodemask */ 4469 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4470 unsigned long totalpages = early_calculate_totalpages(); 4471 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4472 4473 /* 4474 * If movablecore was specified, calculate what size of 4475 * kernelcore that corresponds so that memory usable for 4476 * any allocation type is evenly spread. If both kernelcore 4477 * and movablecore are specified, then the value of kernelcore 4478 * will be used for required_kernelcore if it's greater than 4479 * what movablecore would have allowed. 4480 */ 4481 if (required_movablecore) { 4482 unsigned long corepages; 4483 4484 /* 4485 * Round-up so that ZONE_MOVABLE is at least as large as what 4486 * was requested by the user 4487 */ 4488 required_movablecore = 4489 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4490 corepages = totalpages - required_movablecore; 4491 4492 required_kernelcore = max(required_kernelcore, corepages); 4493 } 4494 4495 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4496 if (!required_kernelcore) 4497 goto out; 4498 4499 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4500 find_usable_zone_for_movable(); 4501 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4502 4503restart: 4504 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4505 kernelcore_node = required_kernelcore / usable_nodes; 4506 for_each_node_state(nid, N_HIGH_MEMORY) { 4507 /* 4508 * Recalculate kernelcore_node if the division per node 4509 * now exceeds what is necessary to satisfy the requested 4510 * amount of memory for the kernel 4511 */ 4512 if (required_kernelcore < kernelcore_node) 4513 kernelcore_node = required_kernelcore / usable_nodes; 4514 4515 /* 4516 * As the map is walked, we track how much memory is usable 4517 * by the kernel using kernelcore_remaining. When it is 4518 * 0, the rest of the node is usable by ZONE_MOVABLE 4519 */ 4520 kernelcore_remaining = kernelcore_node; 4521 4522 /* Go through each range of PFNs within this node */ 4523 for_each_active_range_index_in_nid(i, nid) { 4524 unsigned long start_pfn, end_pfn; 4525 unsigned long size_pages; 4526 4527 start_pfn = max(early_node_map[i].start_pfn, 4528 zone_movable_pfn[nid]); 4529 end_pfn = early_node_map[i].end_pfn; 4530 if (start_pfn >= end_pfn) 4531 continue; 4532 4533 /* Account for what is only usable for kernelcore */ 4534 if (start_pfn < usable_startpfn) { 4535 unsigned long kernel_pages; 4536 kernel_pages = min(end_pfn, usable_startpfn) 4537 - start_pfn; 4538 4539 kernelcore_remaining -= min(kernel_pages, 4540 kernelcore_remaining); 4541 required_kernelcore -= min(kernel_pages, 4542 required_kernelcore); 4543 4544 /* Continue if range is now fully accounted */ 4545 if (end_pfn <= usable_startpfn) { 4546 4547 /* 4548 * Push zone_movable_pfn to the end so 4549 * that if we have to rebalance 4550 * kernelcore across nodes, we will 4551 * not double account here 4552 */ 4553 zone_movable_pfn[nid] = end_pfn; 4554 continue; 4555 } 4556 start_pfn = usable_startpfn; 4557 } 4558 4559 /* 4560 * The usable PFN range for ZONE_MOVABLE is from 4561 * start_pfn->end_pfn. Calculate size_pages as the 4562 * number of pages used as kernelcore 4563 */ 4564 size_pages = end_pfn - start_pfn; 4565 if (size_pages > kernelcore_remaining) 4566 size_pages = kernelcore_remaining; 4567 zone_movable_pfn[nid] = start_pfn + size_pages; 4568 4569 /* 4570 * Some kernelcore has been met, update counts and 4571 * break if the kernelcore for this node has been 4572 * satisified 4573 */ 4574 required_kernelcore -= min(required_kernelcore, 4575 size_pages); 4576 kernelcore_remaining -= size_pages; 4577 if (!kernelcore_remaining) 4578 break; 4579 } 4580 } 4581 4582 /* 4583 * If there is still required_kernelcore, we do another pass with one 4584 * less node in the count. This will push zone_movable_pfn[nid] further 4585 * along on the nodes that still have memory until kernelcore is 4586 * satisified 4587 */ 4588 usable_nodes--; 4589 if (usable_nodes && required_kernelcore > usable_nodes) 4590 goto restart; 4591 4592 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4593 for (nid = 0; nid < MAX_NUMNODES; nid++) 4594 zone_movable_pfn[nid] = 4595 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4596 4597out: 4598 /* restore the node_state */ 4599 node_states[N_HIGH_MEMORY] = saved_node_state; 4600} 4601 4602/* Any regular memory on that node ? */ 4603static void check_for_regular_memory(pg_data_t *pgdat) 4604{ 4605#ifdef CONFIG_HIGHMEM 4606 enum zone_type zone_type; 4607 4608 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4609 struct zone *zone = &pgdat->node_zones[zone_type]; 4610 if (zone->present_pages) 4611 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4612 } 4613#endif 4614} 4615 4616/** 4617 * free_area_init_nodes - Initialise all pg_data_t and zone data 4618 * @max_zone_pfn: an array of max PFNs for each zone 4619 * 4620 * This will call free_area_init_node() for each active node in the system. 4621 * Using the page ranges provided by add_active_range(), the size of each 4622 * zone in each node and their holes is calculated. If the maximum PFN 4623 * between two adjacent zones match, it is assumed that the zone is empty. 4624 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4625 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4626 * starts where the previous one ended. For example, ZONE_DMA32 starts 4627 * at arch_max_dma_pfn. 4628 */ 4629void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4630{ 4631 unsigned long nid; 4632 int i; 4633 4634 /* Sort early_node_map as initialisation assumes it is sorted */ 4635 sort_node_map(); 4636 4637 /* Record where the zone boundaries are */ 4638 memset(arch_zone_lowest_possible_pfn, 0, 4639 sizeof(arch_zone_lowest_possible_pfn)); 4640 memset(arch_zone_highest_possible_pfn, 0, 4641 sizeof(arch_zone_highest_possible_pfn)); 4642 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4643 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4644 for (i = 1; i < MAX_NR_ZONES; i++) { 4645 if (i == ZONE_MOVABLE) 4646 continue; 4647 arch_zone_lowest_possible_pfn[i] = 4648 arch_zone_highest_possible_pfn[i-1]; 4649 arch_zone_highest_possible_pfn[i] = 4650 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4651 } 4652 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4653 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4654 4655 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4656 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4657 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4658 4659 /* Print out the zone ranges */ 4660 printk("Zone PFN ranges:\n"); 4661 for (i = 0; i < MAX_NR_ZONES; i++) { 4662 if (i == ZONE_MOVABLE) 4663 continue; 4664 printk(" %-8s ", zone_names[i]); 4665 if (arch_zone_lowest_possible_pfn[i] == 4666 arch_zone_highest_possible_pfn[i]) 4667 printk("empty\n"); 4668 else 4669 printk("%0#10lx -> %0#10lx\n", 4670 arch_zone_lowest_possible_pfn[i], 4671 arch_zone_highest_possible_pfn[i]); 4672 } 4673 4674 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4675 printk("Movable zone start PFN for each node\n"); 4676 for (i = 0; i < MAX_NUMNODES; i++) { 4677 if (zone_movable_pfn[i]) 4678 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4679 } 4680 4681 /* Print out the early_node_map[] */ 4682 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4683 for (i = 0; i < nr_nodemap_entries; i++) 4684 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4685 early_node_map[i].start_pfn, 4686 early_node_map[i].end_pfn); 4687 4688 /* Initialise every node */ 4689 mminit_verify_pageflags_layout(); 4690 setup_nr_node_ids(); 4691 for_each_online_node(nid) { 4692 pg_data_t *pgdat = NODE_DATA(nid); 4693 free_area_init_node(nid, NULL, 4694 find_min_pfn_for_node(nid), NULL); 4695 4696 /* Any memory on that node */ 4697 if (pgdat->node_present_pages) 4698 node_set_state(nid, N_HIGH_MEMORY); 4699 check_for_regular_memory(pgdat); 4700 } 4701} 4702 4703static int __init cmdline_parse_core(char *p, unsigned long *core) 4704{ 4705 unsigned long long coremem; 4706 if (!p) 4707 return -EINVAL; 4708 4709 coremem = memparse(p, &p); 4710 *core = coremem >> PAGE_SHIFT; 4711 4712 /* Paranoid check that UL is enough for the coremem value */ 4713 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4714 4715 return 0; 4716} 4717 4718/* 4719 * kernelcore=size sets the amount of memory for use for allocations that 4720 * cannot be reclaimed or migrated. 4721 */ 4722static int __init cmdline_parse_kernelcore(char *p) 4723{ 4724 return cmdline_parse_core(p, &required_kernelcore); 4725} 4726 4727/* 4728 * movablecore=size sets the amount of memory for use for allocations that 4729 * can be reclaimed or migrated. 4730 */ 4731static int __init cmdline_parse_movablecore(char *p) 4732{ 4733 return cmdline_parse_core(p, &required_movablecore); 4734} 4735 4736early_param("kernelcore", cmdline_parse_kernelcore); 4737early_param("movablecore", cmdline_parse_movablecore); 4738 4739#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4740 4741/** 4742 * set_dma_reserve - set the specified number of pages reserved in the first zone 4743 * @new_dma_reserve: The number of pages to mark reserved 4744 * 4745 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4746 * In the DMA zone, a significant percentage may be consumed by kernel image 4747 * and other unfreeable allocations which can skew the watermarks badly. This 4748 * function may optionally be used to account for unfreeable pages in the 4749 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4750 * smaller per-cpu batchsize. 4751 */ 4752void __init set_dma_reserve(unsigned long new_dma_reserve) 4753{ 4754 dma_reserve = new_dma_reserve; 4755} 4756 4757#ifndef CONFIG_NEED_MULTIPLE_NODES 4758struct pglist_data __refdata contig_page_data = { 4759#ifndef CONFIG_NO_BOOTMEM 4760 .bdata = &bootmem_node_data[0] 4761#endif 4762 }; 4763EXPORT_SYMBOL(contig_page_data); 4764#endif 4765 4766void __init free_area_init(unsigned long *zones_size) 4767{ 4768 free_area_init_node(0, zones_size, 4769 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4770} 4771 4772static int page_alloc_cpu_notify(struct notifier_block *self, 4773 unsigned long action, void *hcpu) 4774{ 4775 int cpu = (unsigned long)hcpu; 4776 4777 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4778 drain_pages(cpu); 4779 4780 /* 4781 * Spill the event counters of the dead processor 4782 * into the current processors event counters. 4783 * This artificially elevates the count of the current 4784 * processor. 4785 */ 4786 vm_events_fold_cpu(cpu); 4787 4788 /* 4789 * Zero the differential counters of the dead processor 4790 * so that the vm statistics are consistent. 4791 * 4792 * This is only okay since the processor is dead and cannot 4793 * race with what we are doing. 4794 */ 4795 refresh_cpu_vm_stats(cpu); 4796 } 4797 return NOTIFY_OK; 4798} 4799 4800void __init page_alloc_init(void) 4801{ 4802 hotcpu_notifier(page_alloc_cpu_notify, 0); 4803} 4804 4805/* 4806 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4807 * or min_free_kbytes changes. 4808 */ 4809static void calculate_totalreserve_pages(void) 4810{ 4811 struct pglist_data *pgdat; 4812 unsigned long reserve_pages = 0; 4813 enum zone_type i, j; 4814 4815 for_each_online_pgdat(pgdat) { 4816 for (i = 0; i < MAX_NR_ZONES; i++) { 4817 struct zone *zone = pgdat->node_zones + i; 4818 unsigned long max = 0; 4819 4820 /* Find valid and maximum lowmem_reserve in the zone */ 4821 for (j = i; j < MAX_NR_ZONES; j++) { 4822 if (zone->lowmem_reserve[j] > max) 4823 max = zone->lowmem_reserve[j]; 4824 } 4825 4826 /* we treat the high watermark as reserved pages. */ 4827 max += high_wmark_pages(zone); 4828 4829 if (max > zone->present_pages) 4830 max = zone->present_pages; 4831 reserve_pages += max; 4832 } 4833 } 4834 totalreserve_pages = reserve_pages; 4835} 4836 4837/* 4838 * setup_per_zone_lowmem_reserve - called whenever 4839 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4840 * has a correct pages reserved value, so an adequate number of 4841 * pages are left in the zone after a successful __alloc_pages(). 4842 */ 4843static void setup_per_zone_lowmem_reserve(void) 4844{ 4845 struct pglist_data *pgdat; 4846 enum zone_type j, idx; 4847 4848 for_each_online_pgdat(pgdat) { 4849 for (j = 0; j < MAX_NR_ZONES; j++) { 4850 struct zone *zone = pgdat->node_zones + j; 4851 unsigned long present_pages = zone->present_pages; 4852 4853 zone->lowmem_reserve[j] = 0; 4854 4855 idx = j; 4856 while (idx) { 4857 struct zone *lower_zone; 4858 4859 idx--; 4860 4861 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4862 sysctl_lowmem_reserve_ratio[idx] = 1; 4863 4864 lower_zone = pgdat->node_zones + idx; 4865 lower_zone->lowmem_reserve[j] = present_pages / 4866 sysctl_lowmem_reserve_ratio[idx]; 4867 present_pages += lower_zone->present_pages; 4868 } 4869 } 4870 } 4871 4872 /* update totalreserve_pages */ 4873 calculate_totalreserve_pages(); 4874} 4875 4876/** 4877 * setup_per_zone_wmarks - called when min_free_kbytes changes 4878 * or when memory is hot-{added|removed} 4879 * 4880 * Ensures that the watermark[min,low,high] values for each zone are set 4881 * correctly with respect to min_free_kbytes. 4882 */ 4883void setup_per_zone_wmarks(void) 4884{ 4885 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4886 unsigned long lowmem_pages = 0; 4887 struct zone *zone; 4888 unsigned long flags; 4889 4890 /* Calculate total number of !ZONE_HIGHMEM pages */ 4891 for_each_zone(zone) { 4892 if (!is_highmem(zone)) 4893 lowmem_pages += zone->present_pages; 4894 } 4895 4896 for_each_zone(zone) { 4897 u64 tmp; 4898 4899 spin_lock_irqsave(&zone->lock, flags); 4900 tmp = (u64)pages_min * zone->present_pages; 4901 do_div(tmp, lowmem_pages); 4902 if (is_highmem(zone)) { 4903 /* 4904 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4905 * need highmem pages, so cap pages_min to a small 4906 * value here. 4907 * 4908 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4909 * deltas controls asynch page reclaim, and so should 4910 * not be capped for highmem. 4911 */ 4912 int min_pages; 4913 4914 min_pages = zone->present_pages / 1024; 4915 if (min_pages < SWAP_CLUSTER_MAX) 4916 min_pages = SWAP_CLUSTER_MAX; 4917 if (min_pages > 128) 4918 min_pages = 128; 4919 zone->watermark[WMARK_MIN] = min_pages; 4920 } else { 4921 /* 4922 * If it's a lowmem zone, reserve a number of pages 4923 * proportionate to the zone's size. 4924 */ 4925 zone->watermark[WMARK_MIN] = tmp; 4926 } 4927 4928 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4929 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4930 setup_zone_migrate_reserve(zone); 4931 spin_unlock_irqrestore(&zone->lock, flags); 4932 } 4933 4934 /* update totalreserve_pages */ 4935 calculate_totalreserve_pages(); 4936} 4937 4938/* 4939 * The inactive anon list should be small enough that the VM never has to 4940 * do too much work, but large enough that each inactive page has a chance 4941 * to be referenced again before it is swapped out. 4942 * 4943 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4944 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4945 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4946 * the anonymous pages are kept on the inactive list. 4947 * 4948 * total target max 4949 * memory ratio inactive anon 4950 * ------------------------------------- 4951 * 10MB 1 5MB 4952 * 100MB 1 50MB 4953 * 1GB 3 250MB 4954 * 10GB 10 0.9GB 4955 * 100GB 31 3GB 4956 * 1TB 101 10GB 4957 * 10TB 320 32GB 4958 */ 4959void calculate_zone_inactive_ratio(struct zone *zone) 4960{ 4961 unsigned int gb, ratio; 4962 4963 /* Zone size in gigabytes */ 4964 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4965 if (gb) 4966 ratio = int_sqrt(10 * gb); 4967 else 4968 ratio = 1; 4969 4970 zone->inactive_ratio = ratio; 4971} 4972 4973static void __init setup_per_zone_inactive_ratio(void) 4974{ 4975 struct zone *zone; 4976 4977 for_each_zone(zone) 4978 calculate_zone_inactive_ratio(zone); 4979} 4980 4981/* 4982 * Initialise min_free_kbytes. 4983 * 4984 * For small machines we want it small (128k min). For large machines 4985 * we want it large (64MB max). But it is not linear, because network 4986 * bandwidth does not increase linearly with machine size. We use 4987 * 4988 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4989 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4990 * 4991 * which yields 4992 * 4993 * 16MB: 512k 4994 * 32MB: 724k 4995 * 64MB: 1024k 4996 * 128MB: 1448k 4997 * 256MB: 2048k 4998 * 512MB: 2896k 4999 * 1024MB: 4096k 5000 * 2048MB: 5792k 5001 * 4096MB: 8192k 5002 * 8192MB: 11584k 5003 * 16384MB: 16384k 5004 */ 5005static int __init init_per_zone_wmark_min(void) 5006{ 5007 unsigned long lowmem_kbytes; 5008 5009 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5010 5011 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5012 if (min_free_kbytes < 128) 5013 min_free_kbytes = 128; 5014 if (min_free_kbytes > 65536) 5015 min_free_kbytes = 65536; 5016 setup_per_zone_wmarks(); 5017 setup_per_zone_lowmem_reserve(); 5018 setup_per_zone_inactive_ratio(); 5019 return 0; 5020} 5021module_init(init_per_zone_wmark_min) 5022 5023/* 5024 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5025 * that we can call two helper functions whenever min_free_kbytes 5026 * changes. 5027 */ 5028int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5029 void __user *buffer, size_t *length, loff_t *ppos) 5030{ 5031 proc_dointvec(table, write, buffer, length, ppos); 5032 if (write) 5033 setup_per_zone_wmarks(); 5034 return 0; 5035} 5036 5037#ifdef CONFIG_NUMA 5038int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5039 void __user *buffer, size_t *length, loff_t *ppos) 5040{ 5041 struct zone *zone; 5042 int rc; 5043 5044 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5045 if (rc) 5046 return rc; 5047 5048 for_each_zone(zone) 5049 zone->min_unmapped_pages = (zone->present_pages * 5050 sysctl_min_unmapped_ratio) / 100; 5051 return 0; 5052} 5053 5054int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5055 void __user *buffer, size_t *length, loff_t *ppos) 5056{ 5057 struct zone *zone; 5058 int rc; 5059 5060 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5061 if (rc) 5062 return rc; 5063 5064 for_each_zone(zone) 5065 zone->min_slab_pages = (zone->present_pages * 5066 sysctl_min_slab_ratio) / 100; 5067 return 0; 5068} 5069#endif 5070 5071/* 5072 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5073 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5074 * whenever sysctl_lowmem_reserve_ratio changes. 5075 * 5076 * The reserve ratio obviously has absolutely no relation with the 5077 * minimum watermarks. The lowmem reserve ratio can only make sense 5078 * if in function of the boot time zone sizes. 5079 */ 5080int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5081 void __user *buffer, size_t *length, loff_t *ppos) 5082{ 5083 proc_dointvec_minmax(table, write, buffer, length, ppos); 5084 setup_per_zone_lowmem_reserve(); 5085 return 0; 5086} 5087 5088/* 5089 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5090 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5091 * can have before it gets flushed back to buddy allocator. 5092 */ 5093 5094int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5095 void __user *buffer, size_t *length, loff_t *ppos) 5096{ 5097 struct zone *zone; 5098 unsigned int cpu; 5099 int ret; 5100 5101 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5102 if (!write || (ret == -EINVAL)) 5103 return ret; 5104 for_each_populated_zone(zone) { 5105 for_each_possible_cpu(cpu) { 5106 unsigned long high; 5107 high = zone->present_pages / percpu_pagelist_fraction; 5108 setup_pagelist_highmark( 5109 per_cpu_ptr(zone->pageset, cpu), high); 5110 } 5111 } 5112 return 0; 5113} 5114 5115int hashdist = HASHDIST_DEFAULT; 5116 5117#ifdef CONFIG_NUMA 5118static int __init set_hashdist(char *str) 5119{ 5120 if (!str) 5121 return 0; 5122 hashdist = simple_strtoul(str, &str, 0); 5123 return 1; 5124} 5125__setup("hashdist=", set_hashdist); 5126#endif 5127 5128/* 5129 * allocate a large system hash table from bootmem 5130 * - it is assumed that the hash table must contain an exact power-of-2 5131 * quantity of entries 5132 * - limit is the number of hash buckets, not the total allocation size 5133 */ 5134void *__init alloc_large_system_hash(const char *tablename, 5135 unsigned long bucketsize, 5136 unsigned long numentries, 5137 int scale, 5138 int flags, 5139 unsigned int *_hash_shift, 5140 unsigned int *_hash_mask, 5141 unsigned long limit) 5142{ 5143 unsigned long long max = limit; 5144 unsigned long log2qty, size; 5145 void *table = NULL; 5146 5147 /* allow the kernel cmdline to have a say */ 5148 if (!numentries) { 5149 /* round applicable memory size up to nearest megabyte */ 5150 numentries = nr_kernel_pages; 5151 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5152 numentries >>= 20 - PAGE_SHIFT; 5153 numentries <<= 20 - PAGE_SHIFT; 5154 5155 /* limit to 1 bucket per 2^scale bytes of low memory */ 5156 if (scale > PAGE_SHIFT) 5157 numentries >>= (scale - PAGE_SHIFT); 5158 else 5159 numentries <<= (PAGE_SHIFT - scale); 5160 5161 /* Make sure we've got at least a 0-order allocation.. */ 5162 if (unlikely(flags & HASH_SMALL)) { 5163 /* Makes no sense without HASH_EARLY */ 5164 WARN_ON(!(flags & HASH_EARLY)); 5165 if (!(numentries >> *_hash_shift)) { 5166 numentries = 1UL << *_hash_shift; 5167 BUG_ON(!numentries); 5168 } 5169 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5170 numentries = PAGE_SIZE / bucketsize; 5171 } 5172 numentries = roundup_pow_of_two(numentries); 5173 5174 /* limit allocation size to 1/16 total memory by default */ 5175 if (max == 0) { 5176 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5177 do_div(max, bucketsize); 5178 } 5179 5180 if (numentries > max) 5181 numentries = max; 5182 5183 log2qty = ilog2(numentries); 5184 5185 do { 5186 size = bucketsize << log2qty; 5187 if (flags & HASH_EARLY) 5188 table = alloc_bootmem_nopanic(size); 5189 else if (hashdist) 5190 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5191 else { 5192 /* 5193 * If bucketsize is not a power-of-two, we may free 5194 * some pages at the end of hash table which 5195 * alloc_pages_exact() automatically does 5196 */ 5197 if (get_order(size) < MAX_ORDER) { 5198 table = alloc_pages_exact(size, GFP_ATOMIC); 5199 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5200 } 5201 } 5202 } while (!table && size > PAGE_SIZE && --log2qty); 5203 5204 if (!table) 5205 panic("Failed to allocate %s hash table\n", tablename); 5206 5207 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5208 tablename, 5209 (1UL << log2qty), 5210 ilog2(size) - PAGE_SHIFT, 5211 size); 5212 5213 if (_hash_shift) 5214 *_hash_shift = log2qty; 5215 if (_hash_mask) 5216 *_hash_mask = (1 << log2qty) - 1; 5217 5218 return table; 5219} 5220 5221/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5222static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5223 unsigned long pfn) 5224{ 5225#ifdef CONFIG_SPARSEMEM 5226 return __pfn_to_section(pfn)->pageblock_flags; 5227#else 5228 return zone->pageblock_flags; 5229#endif /* CONFIG_SPARSEMEM */ 5230} 5231 5232static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5233{ 5234#ifdef CONFIG_SPARSEMEM 5235 pfn &= (PAGES_PER_SECTION-1); 5236 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5237#else 5238 pfn = pfn - zone->zone_start_pfn; 5239 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5240#endif /* CONFIG_SPARSEMEM */ 5241} 5242 5243/** 5244 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5245 * @page: The page within the block of interest 5246 * @start_bitidx: The first bit of interest to retrieve 5247 * @end_bitidx: The last bit of interest 5248 * returns pageblock_bits flags 5249 */ 5250unsigned long get_pageblock_flags_group(struct page *page, 5251 int start_bitidx, int end_bitidx) 5252{ 5253 struct zone *zone; 5254 unsigned long *bitmap; 5255 unsigned long pfn, bitidx; 5256 unsigned long flags = 0; 5257 unsigned long value = 1; 5258 5259 zone = page_zone(page); 5260 pfn = page_to_pfn(page); 5261 bitmap = get_pageblock_bitmap(zone, pfn); 5262 bitidx = pfn_to_bitidx(zone, pfn); 5263 5264 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5265 if (test_bit(bitidx + start_bitidx, bitmap)) 5266 flags |= value; 5267 5268 return flags; 5269} 5270 5271/** 5272 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5273 * @page: The page within the block of interest 5274 * @start_bitidx: The first bit of interest 5275 * @end_bitidx: The last bit of interest 5276 * @flags: The flags to set 5277 */ 5278void set_pageblock_flags_group(struct page *page, unsigned long flags, 5279 int start_bitidx, int end_bitidx) 5280{ 5281 struct zone *zone; 5282 unsigned long *bitmap; 5283 unsigned long pfn, bitidx; 5284 unsigned long value = 1; 5285 5286 zone = page_zone(page); 5287 pfn = page_to_pfn(page); 5288 bitmap = get_pageblock_bitmap(zone, pfn); 5289 bitidx = pfn_to_bitidx(zone, pfn); 5290 VM_BUG_ON(pfn < zone->zone_start_pfn); 5291 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5292 5293 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5294 if (flags & value) 5295 __set_bit(bitidx + start_bitidx, bitmap); 5296 else 5297 __clear_bit(bitidx + start_bitidx, bitmap); 5298} 5299 5300/* 5301 * This is designed as sub function...plz see page_isolation.c also. 5302 * set/clear page block's type to be ISOLATE. 5303 * page allocater never alloc memory from ISOLATE block. 5304 */ 5305 5306int set_migratetype_isolate(struct page *page) 5307{ 5308 struct zone *zone; 5309 struct page *curr_page; 5310 unsigned long flags, pfn, iter; 5311 unsigned long immobile = 0; 5312 struct memory_isolate_notify arg; 5313 int notifier_ret; 5314 int ret = -EBUSY; 5315 int zone_idx; 5316 5317 zone = page_zone(page); 5318 zone_idx = zone_idx(zone); 5319 5320 spin_lock_irqsave(&zone->lock, flags); 5321 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE || 5322 zone_idx == ZONE_MOVABLE) { 5323 ret = 0; 5324 goto out; 5325 } 5326 5327 pfn = page_to_pfn(page); 5328 arg.start_pfn = pfn; 5329 arg.nr_pages = pageblock_nr_pages; 5330 arg.pages_found = 0; 5331 5332 /* 5333 * It may be possible to isolate a pageblock even if the 5334 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5335 * notifier chain is used by balloon drivers to return the 5336 * number of pages in a range that are held by the balloon 5337 * driver to shrink memory. If all the pages are accounted for 5338 * by balloons, are free, or on the LRU, isolation can continue. 5339 * Later, for example, when memory hotplug notifier runs, these 5340 * pages reported as "can be isolated" should be isolated(freed) 5341 * by the balloon driver through the memory notifier chain. 5342 */ 5343 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5344 notifier_ret = notifier_to_errno(notifier_ret); 5345 if (notifier_ret || !arg.pages_found) 5346 goto out; 5347 5348 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) { 5349 if (!pfn_valid_within(pfn)) 5350 continue; 5351 5352 curr_page = pfn_to_page(iter); 5353 if (!page_count(curr_page) || PageLRU(curr_page)) 5354 continue; 5355 5356 immobile++; 5357 } 5358 5359 if (arg.pages_found == immobile) 5360 ret = 0; 5361 5362out: 5363 if (!ret) { 5364 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5365 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5366 } 5367 5368 spin_unlock_irqrestore(&zone->lock, flags); 5369 if (!ret) 5370 drain_all_pages(); 5371 return ret; 5372} 5373 5374void unset_migratetype_isolate(struct page *page) 5375{ 5376 struct zone *zone; 5377 unsigned long flags; 5378 zone = page_zone(page); 5379 spin_lock_irqsave(&zone->lock, flags); 5380 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5381 goto out; 5382 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5383 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5384out: 5385 spin_unlock_irqrestore(&zone->lock, flags); 5386} 5387 5388#ifdef CONFIG_MEMORY_HOTREMOVE 5389/* 5390 * All pages in the range must be isolated before calling this. 5391 */ 5392void 5393__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5394{ 5395 struct page *page; 5396 struct zone *zone; 5397 int order, i; 5398 unsigned long pfn; 5399 unsigned long flags; 5400 /* find the first valid pfn */ 5401 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5402 if (pfn_valid(pfn)) 5403 break; 5404 if (pfn == end_pfn) 5405 return; 5406 zone = page_zone(pfn_to_page(pfn)); 5407 spin_lock_irqsave(&zone->lock, flags); 5408 pfn = start_pfn; 5409 while (pfn < end_pfn) { 5410 if (!pfn_valid(pfn)) { 5411 pfn++; 5412 continue; 5413 } 5414 page = pfn_to_page(pfn); 5415 BUG_ON(page_count(page)); 5416 BUG_ON(!PageBuddy(page)); 5417 order = page_order(page); 5418#ifdef CONFIG_DEBUG_VM 5419 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5420 pfn, 1 << order, end_pfn); 5421#endif 5422 list_del(&page->lru); 5423 rmv_page_order(page); 5424 zone->free_area[order].nr_free--; 5425 __mod_zone_page_state(zone, NR_FREE_PAGES, 5426 - (1UL << order)); 5427 for (i = 0; i < (1 << order); i++) 5428 SetPageReserved((page+i)); 5429 pfn += (1 << order); 5430 } 5431 spin_unlock_irqrestore(&zone->lock, flags); 5432} 5433#endif 5434 5435#ifdef CONFIG_MEMORY_FAILURE 5436bool is_free_buddy_page(struct page *page) 5437{ 5438 struct zone *zone = page_zone(page); 5439 unsigned long pfn = page_to_pfn(page); 5440 unsigned long flags; 5441 int order; 5442 5443 spin_lock_irqsave(&zone->lock, flags); 5444 for (order = 0; order < MAX_ORDER; order++) { 5445 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5446 5447 if (PageBuddy(page_head) && page_order(page_head) >= order) 5448 break; 5449 } 5450 spin_unlock_irqrestore(&zone->lock, flags); 5451 5452 return order < MAX_ORDER; 5453} 5454#endif 5455 5456static struct trace_print_flags pageflag_names[] = { 5457 {1UL << PG_locked, "locked" }, 5458 {1UL << PG_error, "error" }, 5459 {1UL << PG_referenced, "referenced" }, 5460 {1UL << PG_uptodate, "uptodate" }, 5461 {1UL << PG_dirty, "dirty" }, 5462 {1UL << PG_lru, "lru" }, 5463 {1UL << PG_active, "active" }, 5464 {1UL << PG_slab, "slab" }, 5465 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5466 {1UL << PG_arch_1, "arch_1" }, 5467 {1UL << PG_reserved, "reserved" }, 5468 {1UL << PG_private, "private" }, 5469 {1UL << PG_private_2, "private_2" }, 5470 {1UL << PG_writeback, "writeback" }, 5471#ifdef CONFIG_PAGEFLAGS_EXTENDED 5472 {1UL << PG_head, "head" }, 5473 {1UL << PG_tail, "tail" }, 5474#else 5475 {1UL << PG_compound, "compound" }, 5476#endif 5477 {1UL << PG_swapcache, "swapcache" }, 5478 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5479 {1UL << PG_reclaim, "reclaim" }, 5480 {1UL << PG_buddy, "buddy" }, 5481 {1UL << PG_swapbacked, "swapbacked" }, 5482 {1UL << PG_unevictable, "unevictable" }, 5483#ifdef CONFIG_MMU 5484 {1UL << PG_mlocked, "mlocked" }, 5485#endif 5486#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5487 {1UL << PG_uncached, "uncached" }, 5488#endif 5489#ifdef CONFIG_MEMORY_FAILURE 5490 {1UL << PG_hwpoison, "hwpoison" }, 5491#endif 5492 {-1UL, NULL }, 5493}; 5494 5495static void dump_page_flags(unsigned long flags) 5496{ 5497 const char *delim = ""; 5498 unsigned long mask; 5499 int i; 5500 5501 printk(KERN_ALERT "page flags: %#lx(", flags); 5502 5503 /* remove zone id */ 5504 flags &= (1UL << NR_PAGEFLAGS) - 1; 5505 5506 for (i = 0; pageflag_names[i].name && flags; i++) { 5507 5508 mask = pageflag_names[i].mask; 5509 if ((flags & mask) != mask) 5510 continue; 5511 5512 flags &= ~mask; 5513 printk("%s%s", delim, pageflag_names[i].name); 5514 delim = "|"; 5515 } 5516 5517 /* check for left over flags */ 5518 if (flags) 5519 printk("%s%#lx", delim, flags); 5520 5521 printk(")\n"); 5522} 5523 5524void dump_page(struct page *page) 5525{ 5526 printk(KERN_ALERT 5527 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5528 page, page_count(page), page_mapcount(page), 5529 page->mapping, page->index); 5530 dump_page_flags(page->flags); 5531} 5532