page_alloc.c revision c06b1fca18c3ad868bfcaca230146e3038583422
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/memblock.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#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 62DEFINE_PER_CPU(int, numa_node); 63EXPORT_PER_CPU_SYMBOL(numa_node); 64#endif 65 66#ifdef CONFIG_HAVE_MEMORYLESS_NODES 67/* 68 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 69 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 70 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 71 * defined in <linux/topology.h>. 72 */ 73DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 74EXPORT_PER_CPU_SYMBOL(_numa_mem_); 75#endif 76 77/* 78 * Array of node states. 79 */ 80nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 81 [N_POSSIBLE] = NODE_MASK_ALL, 82 [N_ONLINE] = { { [0] = 1UL } }, 83#ifndef CONFIG_NUMA 84 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 85#ifdef CONFIG_HIGHMEM 86 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 87#endif 88 [N_CPU] = { { [0] = 1UL } }, 89#endif /* NUMA */ 90}; 91EXPORT_SYMBOL(node_states); 92 93unsigned long totalram_pages __read_mostly; 94unsigned long totalreserve_pages __read_mostly; 95int percpu_pagelist_fraction; 96gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 97 98#ifdef CONFIG_PM_SLEEP 99/* 100 * The following functions are used by the suspend/hibernate code to temporarily 101 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 102 * while devices are suspended. To avoid races with the suspend/hibernate code, 103 * they should always be called with pm_mutex held (gfp_allowed_mask also should 104 * only be modified with pm_mutex held, unless the suspend/hibernate code is 105 * guaranteed not to run in parallel with that modification). 106 */ 107 108static gfp_t saved_gfp_mask; 109 110void pm_restore_gfp_mask(void) 111{ 112 WARN_ON(!mutex_is_locked(&pm_mutex)); 113 if (saved_gfp_mask) { 114 gfp_allowed_mask = saved_gfp_mask; 115 saved_gfp_mask = 0; 116 } 117} 118 119void pm_restrict_gfp_mask(void) 120{ 121 WARN_ON(!mutex_is_locked(&pm_mutex)); 122 WARN_ON(saved_gfp_mask); 123 saved_gfp_mask = gfp_allowed_mask; 124 gfp_allowed_mask &= ~GFP_IOFS; 125} 126#endif /* CONFIG_PM_SLEEP */ 127 128#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 129int pageblock_order __read_mostly; 130#endif 131 132static void __free_pages_ok(struct page *page, unsigned int order); 133 134/* 135 * results with 256, 32 in the lowmem_reserve sysctl: 136 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 137 * 1G machine -> (16M dma, 784M normal, 224M high) 138 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 139 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 140 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 141 * 142 * TBD: should special case ZONE_DMA32 machines here - in those we normally 143 * don't need any ZONE_NORMAL reservation 144 */ 145int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 146#ifdef CONFIG_ZONE_DMA 147 256, 148#endif 149#ifdef CONFIG_ZONE_DMA32 150 256, 151#endif 152#ifdef CONFIG_HIGHMEM 153 32, 154#endif 155 32, 156}; 157 158EXPORT_SYMBOL(totalram_pages); 159 160static char * const zone_names[MAX_NR_ZONES] = { 161#ifdef CONFIG_ZONE_DMA 162 "DMA", 163#endif 164#ifdef CONFIG_ZONE_DMA32 165 "DMA32", 166#endif 167 "Normal", 168#ifdef CONFIG_HIGHMEM 169 "HighMem", 170#endif 171 "Movable", 172}; 173 174int min_free_kbytes = 1024; 175 176static unsigned long __meminitdata nr_kernel_pages; 177static unsigned long __meminitdata nr_all_pages; 178static unsigned long __meminitdata dma_reserve; 179 180#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 181 /* 182 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 183 * ranges of memory (RAM) that may be registered with add_active_range(). 184 * Ranges passed to add_active_range() will be merged if possible 185 * so the number of times add_active_range() can be called is 186 * related to the number of nodes and the number of holes 187 */ 188 #ifdef CONFIG_MAX_ACTIVE_REGIONS 189 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 190 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 191 #else 192 #if MAX_NUMNODES >= 32 193 /* If there can be many nodes, allow up to 50 holes per node */ 194 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 195 #else 196 /* By default, allow up to 256 distinct regions */ 197 #define MAX_ACTIVE_REGIONS 256 198 #endif 199 #endif 200 201 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 202 static int __meminitdata nr_nodemap_entries; 203 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 204 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 205 static unsigned long __initdata required_kernelcore; 206 static unsigned long __initdata required_movablecore; 207 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 208 209 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 210 int movable_zone; 211 EXPORT_SYMBOL(movable_zone); 212#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 213 214#if MAX_NUMNODES > 1 215int nr_node_ids __read_mostly = MAX_NUMNODES; 216int nr_online_nodes __read_mostly = 1; 217EXPORT_SYMBOL(nr_node_ids); 218EXPORT_SYMBOL(nr_online_nodes); 219#endif 220 221int page_group_by_mobility_disabled __read_mostly; 222 223static void set_pageblock_migratetype(struct page *page, int migratetype) 224{ 225 226 if (unlikely(page_group_by_mobility_disabled)) 227 migratetype = MIGRATE_UNMOVABLE; 228 229 set_pageblock_flags_group(page, (unsigned long)migratetype, 230 PB_migrate, PB_migrate_end); 231} 232 233bool oom_killer_disabled __read_mostly; 234 235#ifdef CONFIG_DEBUG_VM 236static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 237{ 238 int ret = 0; 239 unsigned seq; 240 unsigned long pfn = page_to_pfn(page); 241 242 do { 243 seq = zone_span_seqbegin(zone); 244 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 245 ret = 1; 246 else if (pfn < zone->zone_start_pfn) 247 ret = 1; 248 } while (zone_span_seqretry(zone, seq)); 249 250 return ret; 251} 252 253static int page_is_consistent(struct zone *zone, struct page *page) 254{ 255 if (!pfn_valid_within(page_to_pfn(page))) 256 return 0; 257 if (zone != page_zone(page)) 258 return 0; 259 260 return 1; 261} 262/* 263 * Temporary debugging check for pages not lying within a given zone. 264 */ 265static int bad_range(struct zone *zone, struct page *page) 266{ 267 if (page_outside_zone_boundaries(zone, page)) 268 return 1; 269 if (!page_is_consistent(zone, page)) 270 return 1; 271 272 return 0; 273} 274#else 275static inline int bad_range(struct zone *zone, struct page *page) 276{ 277 return 0; 278} 279#endif 280 281static void bad_page(struct page *page) 282{ 283 static unsigned long resume; 284 static unsigned long nr_shown; 285 static unsigned long nr_unshown; 286 287 /* Don't complain about poisoned pages */ 288 if (PageHWPoison(page)) { 289 __ClearPageBuddy(page); 290 return; 291 } 292 293 /* 294 * Allow a burst of 60 reports, then keep quiet for that minute; 295 * or allow a steady drip of one report per second. 296 */ 297 if (nr_shown == 60) { 298 if (time_before(jiffies, resume)) { 299 nr_unshown++; 300 goto out; 301 } 302 if (nr_unshown) { 303 printk(KERN_ALERT 304 "BUG: Bad page state: %lu messages suppressed\n", 305 nr_unshown); 306 nr_unshown = 0; 307 } 308 nr_shown = 0; 309 } 310 if (nr_shown++ == 0) 311 resume = jiffies + 60 * HZ; 312 313 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 314 current->comm, page_to_pfn(page)); 315 dump_page(page); 316 317 dump_stack(); 318out: 319 /* Leave bad fields for debug, except PageBuddy could make trouble */ 320 __ClearPageBuddy(page); 321 add_taint(TAINT_BAD_PAGE); 322} 323 324/* 325 * Higher-order pages are called "compound pages". They are structured thusly: 326 * 327 * The first PAGE_SIZE page is called the "head page". 328 * 329 * The remaining PAGE_SIZE pages are called "tail pages". 330 * 331 * All pages have PG_compound set. All pages have their ->private pointing at 332 * the head page (even the head page has this). 333 * 334 * The first tail page's ->lru.next holds the address of the compound page's 335 * put_page() function. Its ->lru.prev holds the order of allocation. 336 * This usage means that zero-order pages may not be compound. 337 */ 338 339static void free_compound_page(struct page *page) 340{ 341 __free_pages_ok(page, compound_order(page)); 342} 343 344void prep_compound_page(struct page *page, unsigned long order) 345{ 346 int i; 347 int nr_pages = 1 << order; 348 349 set_compound_page_dtor(page, free_compound_page); 350 set_compound_order(page, order); 351 __SetPageHead(page); 352 for (i = 1; i < nr_pages; i++) { 353 struct page *p = page + i; 354 355 __SetPageTail(p); 356 p->first_page = page; 357 } 358} 359 360/* update __split_huge_page_refcount if you change this function */ 361static int destroy_compound_page(struct page *page, unsigned long order) 362{ 363 int i; 364 int nr_pages = 1 << order; 365 int bad = 0; 366 367 if (unlikely(compound_order(page) != order) || 368 unlikely(!PageHead(page))) { 369 bad_page(page); 370 bad++; 371 } 372 373 __ClearPageHead(page); 374 375 for (i = 1; i < nr_pages; i++) { 376 struct page *p = page + i; 377 378 if (unlikely(!PageTail(p) || (p->first_page != page))) { 379 bad_page(page); 380 bad++; 381 } 382 __ClearPageTail(p); 383 } 384 385 return bad; 386} 387 388static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 389{ 390 int i; 391 392 /* 393 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 394 * and __GFP_HIGHMEM from hard or soft interrupt context. 395 */ 396 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 397 for (i = 0; i < (1 << order); i++) 398 clear_highpage(page + i); 399} 400 401static inline void set_page_order(struct page *page, int order) 402{ 403 set_page_private(page, order); 404 __SetPageBuddy(page); 405} 406 407static inline void rmv_page_order(struct page *page) 408{ 409 __ClearPageBuddy(page); 410 set_page_private(page, 0); 411} 412 413/* 414 * Locate the struct page for both the matching buddy in our 415 * pair (buddy1) and the combined O(n+1) page they form (page). 416 * 417 * 1) Any buddy B1 will have an order O twin B2 which satisfies 418 * the following equation: 419 * B2 = B1 ^ (1 << O) 420 * For example, if the starting buddy (buddy2) is #8 its order 421 * 1 buddy is #10: 422 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 423 * 424 * 2) Any buddy B will have an order O+1 parent P which 425 * satisfies the following equation: 426 * P = B & ~(1 << O) 427 * 428 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 429 */ 430static inline unsigned long 431__find_buddy_index(unsigned long page_idx, unsigned int order) 432{ 433 return page_idx ^ (1 << order); 434} 435 436/* 437 * This function checks whether a page is free && is the buddy 438 * we can do coalesce a page and its buddy if 439 * (a) the buddy is not in a hole && 440 * (b) the buddy is in the buddy system && 441 * (c) a page and its buddy have the same order && 442 * (d) a page and its buddy are in the same zone. 443 * 444 * For recording whether a page is in the buddy system, we set ->_mapcount -2. 445 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. 446 * 447 * For recording page's order, we use page_private(page). 448 */ 449static inline int page_is_buddy(struct page *page, struct page *buddy, 450 int order) 451{ 452 if (!pfn_valid_within(page_to_pfn(buddy))) 453 return 0; 454 455 if (page_zone_id(page) != page_zone_id(buddy)) 456 return 0; 457 458 if (PageBuddy(buddy) && page_order(buddy) == order) { 459 VM_BUG_ON(page_count(buddy) != 0); 460 return 1; 461 } 462 return 0; 463} 464 465/* 466 * Freeing function for a buddy system allocator. 467 * 468 * The concept of a buddy system is to maintain direct-mapped table 469 * (containing bit values) for memory blocks of various "orders". 470 * The bottom level table contains the map for the smallest allocatable 471 * units of memory (here, pages), and each level above it describes 472 * pairs of units from the levels below, hence, "buddies". 473 * At a high level, all that happens here is marking the table entry 474 * at the bottom level available, and propagating the changes upward 475 * as necessary, plus some accounting needed to play nicely with other 476 * parts of the VM system. 477 * At each level, we keep a list of pages, which are heads of continuous 478 * free pages of length of (1 << order) and marked with _mapcount -2. Page's 479 * order is recorded in page_private(page) field. 480 * So when we are allocating or freeing one, we can derive the state of the 481 * other. That is, if we allocate a small block, and both were 482 * free, the remainder of the region must be split into blocks. 483 * If a block is freed, and its buddy is also free, then this 484 * triggers coalescing into a block of larger size. 485 * 486 * -- wli 487 */ 488 489static inline void __free_one_page(struct page *page, 490 struct zone *zone, unsigned int order, 491 int migratetype) 492{ 493 unsigned long page_idx; 494 unsigned long combined_idx; 495 unsigned long uninitialized_var(buddy_idx); 496 struct page *buddy; 497 498 if (unlikely(PageCompound(page))) 499 if (unlikely(destroy_compound_page(page, order))) 500 return; 501 502 VM_BUG_ON(migratetype == -1); 503 504 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 505 506 VM_BUG_ON(page_idx & ((1 << order) - 1)); 507 VM_BUG_ON(bad_range(zone, page)); 508 509 while (order < MAX_ORDER-1) { 510 buddy_idx = __find_buddy_index(page_idx, order); 511 buddy = page + (buddy_idx - page_idx); 512 if (!page_is_buddy(page, buddy, order)) 513 break; 514 515 /* Our buddy is free, merge with it and move up one order. */ 516 list_del(&buddy->lru); 517 zone->free_area[order].nr_free--; 518 rmv_page_order(buddy); 519 combined_idx = buddy_idx & page_idx; 520 page = page + (combined_idx - page_idx); 521 page_idx = combined_idx; 522 order++; 523 } 524 set_page_order(page, order); 525 526 /* 527 * If this is not the largest possible page, check if the buddy 528 * of the next-highest order is free. If it is, it's possible 529 * that pages are being freed that will coalesce soon. In case, 530 * that is happening, add the free page to the tail of the list 531 * so it's less likely to be used soon and more likely to be merged 532 * as a higher order page 533 */ 534 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 535 struct page *higher_page, *higher_buddy; 536 combined_idx = buddy_idx & page_idx; 537 higher_page = page + (combined_idx - page_idx); 538 buddy_idx = __find_buddy_index(combined_idx, order + 1); 539 higher_buddy = page + (buddy_idx - combined_idx); 540 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 541 list_add_tail(&page->lru, 542 &zone->free_area[order].free_list[migratetype]); 543 goto out; 544 } 545 } 546 547 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 548out: 549 zone->free_area[order].nr_free++; 550} 551 552/* 553 * free_page_mlock() -- clean up attempts to free and mlocked() page. 554 * Page should not be on lru, so no need to fix that up. 555 * free_pages_check() will verify... 556 */ 557static inline void free_page_mlock(struct page *page) 558{ 559 __dec_zone_page_state(page, NR_MLOCK); 560 __count_vm_event(UNEVICTABLE_MLOCKFREED); 561} 562 563static inline int free_pages_check(struct page *page) 564{ 565 if (unlikely(page_mapcount(page) | 566 (page->mapping != NULL) | 567 (atomic_read(&page->_count) != 0) | 568 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 569 bad_page(page); 570 return 1; 571 } 572 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 573 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 574 return 0; 575} 576 577/* 578 * Frees a number of pages from the PCP lists 579 * Assumes all pages on list are in same zone, and of same order. 580 * count is the number of pages to free. 581 * 582 * If the zone was previously in an "all pages pinned" state then look to 583 * see if this freeing clears that state. 584 * 585 * And clear the zone's pages_scanned counter, to hold off the "all pages are 586 * pinned" detection logic. 587 */ 588static void free_pcppages_bulk(struct zone *zone, int count, 589 struct per_cpu_pages *pcp) 590{ 591 int migratetype = 0; 592 int batch_free = 0; 593 int to_free = count; 594 595 spin_lock(&zone->lock); 596 zone->all_unreclaimable = 0; 597 zone->pages_scanned = 0; 598 599 while (to_free) { 600 struct page *page; 601 struct list_head *list; 602 603 /* 604 * Remove pages from lists in a round-robin fashion. A 605 * batch_free count is maintained that is incremented when an 606 * empty list is encountered. This is so more pages are freed 607 * off fuller lists instead of spinning excessively around empty 608 * lists 609 */ 610 do { 611 batch_free++; 612 if (++migratetype == MIGRATE_PCPTYPES) 613 migratetype = 0; 614 list = &pcp->lists[migratetype]; 615 } while (list_empty(list)); 616 617 do { 618 page = list_entry(list->prev, struct page, lru); 619 /* must delete as __free_one_page list manipulates */ 620 list_del(&page->lru); 621 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 622 __free_one_page(page, zone, 0, page_private(page)); 623 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 624 } while (--to_free && --batch_free && !list_empty(list)); 625 } 626 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 627 spin_unlock(&zone->lock); 628} 629 630static void free_one_page(struct zone *zone, struct page *page, int order, 631 int migratetype) 632{ 633 spin_lock(&zone->lock); 634 zone->all_unreclaimable = 0; 635 zone->pages_scanned = 0; 636 637 __free_one_page(page, zone, order, migratetype); 638 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 639 spin_unlock(&zone->lock); 640} 641 642static bool free_pages_prepare(struct page *page, unsigned int order) 643{ 644 int i; 645 int bad = 0; 646 647 trace_mm_page_free_direct(page, order); 648 kmemcheck_free_shadow(page, order); 649 650 if (PageAnon(page)) 651 page->mapping = NULL; 652 for (i = 0; i < (1 << order); i++) 653 bad += free_pages_check(page + i); 654 if (bad) 655 return false; 656 657 if (!PageHighMem(page)) { 658 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 659 debug_check_no_obj_freed(page_address(page), 660 PAGE_SIZE << order); 661 } 662 arch_free_page(page, order); 663 kernel_map_pages(page, 1 << order, 0); 664 665 return true; 666} 667 668static void __free_pages_ok(struct page *page, unsigned int order) 669{ 670 unsigned long flags; 671 int wasMlocked = __TestClearPageMlocked(page); 672 673 if (!free_pages_prepare(page, order)) 674 return; 675 676 local_irq_save(flags); 677 if (unlikely(wasMlocked)) 678 free_page_mlock(page); 679 __count_vm_events(PGFREE, 1 << order); 680 free_one_page(page_zone(page), page, order, 681 get_pageblock_migratetype(page)); 682 local_irq_restore(flags); 683} 684 685/* 686 * permit the bootmem allocator to evade page validation on high-order frees 687 */ 688void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 689{ 690 if (order == 0) { 691 __ClearPageReserved(page); 692 set_page_count(page, 0); 693 set_page_refcounted(page); 694 __free_page(page); 695 } else { 696 int loop; 697 698 prefetchw(page); 699 for (loop = 0; loop < BITS_PER_LONG; loop++) { 700 struct page *p = &page[loop]; 701 702 if (loop + 1 < BITS_PER_LONG) 703 prefetchw(p + 1); 704 __ClearPageReserved(p); 705 set_page_count(p, 0); 706 } 707 708 set_page_refcounted(page); 709 __free_pages(page, order); 710 } 711} 712 713 714/* 715 * The order of subdivision here is critical for the IO subsystem. 716 * Please do not alter this order without good reasons and regression 717 * testing. Specifically, as large blocks of memory are subdivided, 718 * the order in which smaller blocks are delivered depends on the order 719 * they're subdivided in this function. This is the primary factor 720 * influencing the order in which pages are delivered to the IO 721 * subsystem according to empirical testing, and this is also justified 722 * by considering the behavior of a buddy system containing a single 723 * large block of memory acted on by a series of small allocations. 724 * This behavior is a critical factor in sglist merging's success. 725 * 726 * -- wli 727 */ 728static inline void expand(struct zone *zone, struct page *page, 729 int low, int high, struct free_area *area, 730 int migratetype) 731{ 732 unsigned long size = 1 << high; 733 734 while (high > low) { 735 area--; 736 high--; 737 size >>= 1; 738 VM_BUG_ON(bad_range(zone, &page[size])); 739 list_add(&page[size].lru, &area->free_list[migratetype]); 740 area->nr_free++; 741 set_page_order(&page[size], high); 742 } 743} 744 745/* 746 * This page is about to be returned from the page allocator 747 */ 748static inline int check_new_page(struct page *page) 749{ 750 if (unlikely(page_mapcount(page) | 751 (page->mapping != NULL) | 752 (atomic_read(&page->_count) != 0) | 753 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 754 bad_page(page); 755 return 1; 756 } 757 return 0; 758} 759 760static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 761{ 762 int i; 763 764 for (i = 0; i < (1 << order); i++) { 765 struct page *p = page + i; 766 if (unlikely(check_new_page(p))) 767 return 1; 768 } 769 770 set_page_private(page, 0); 771 set_page_refcounted(page); 772 773 arch_alloc_page(page, order); 774 kernel_map_pages(page, 1 << order, 1); 775 776 if (gfp_flags & __GFP_ZERO) 777 prep_zero_page(page, order, gfp_flags); 778 779 if (order && (gfp_flags & __GFP_COMP)) 780 prep_compound_page(page, order); 781 782 return 0; 783} 784 785/* 786 * Go through the free lists for the given migratetype and remove 787 * the smallest available page from the freelists 788 */ 789static inline 790struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 791 int migratetype) 792{ 793 unsigned int current_order; 794 struct free_area * area; 795 struct page *page; 796 797 /* Find a page of the appropriate size in the preferred list */ 798 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 799 area = &(zone->free_area[current_order]); 800 if (list_empty(&area->free_list[migratetype])) 801 continue; 802 803 page = list_entry(area->free_list[migratetype].next, 804 struct page, lru); 805 list_del(&page->lru); 806 rmv_page_order(page); 807 area->nr_free--; 808 expand(zone, page, order, current_order, area, migratetype); 809 return page; 810 } 811 812 return NULL; 813} 814 815 816/* 817 * This array describes the order lists are fallen back to when 818 * the free lists for the desirable migrate type are depleted 819 */ 820static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 821 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 822 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 823 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 824 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 825}; 826 827/* 828 * Move the free pages in a range to the free lists of the requested type. 829 * Note that start_page and end_pages are not aligned on a pageblock 830 * boundary. If alignment is required, use move_freepages_block() 831 */ 832static int move_freepages(struct zone *zone, 833 struct page *start_page, struct page *end_page, 834 int migratetype) 835{ 836 struct page *page; 837 unsigned long order; 838 int pages_moved = 0; 839 840#ifndef CONFIG_HOLES_IN_ZONE 841 /* 842 * page_zone is not safe to call in this context when 843 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 844 * anyway as we check zone boundaries in move_freepages_block(). 845 * Remove at a later date when no bug reports exist related to 846 * grouping pages by mobility 847 */ 848 BUG_ON(page_zone(start_page) != page_zone(end_page)); 849#endif 850 851 for (page = start_page; page <= end_page;) { 852 /* Make sure we are not inadvertently changing nodes */ 853 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 854 855 if (!pfn_valid_within(page_to_pfn(page))) { 856 page++; 857 continue; 858 } 859 860 if (!PageBuddy(page)) { 861 page++; 862 continue; 863 } 864 865 order = page_order(page); 866 list_del(&page->lru); 867 list_add(&page->lru, 868 &zone->free_area[order].free_list[migratetype]); 869 page += 1 << order; 870 pages_moved += 1 << order; 871 } 872 873 return pages_moved; 874} 875 876static int move_freepages_block(struct zone *zone, struct page *page, 877 int migratetype) 878{ 879 unsigned long start_pfn, end_pfn; 880 struct page *start_page, *end_page; 881 882 start_pfn = page_to_pfn(page); 883 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 884 start_page = pfn_to_page(start_pfn); 885 end_page = start_page + pageblock_nr_pages - 1; 886 end_pfn = start_pfn + pageblock_nr_pages - 1; 887 888 /* Do not cross zone boundaries */ 889 if (start_pfn < zone->zone_start_pfn) 890 start_page = page; 891 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 892 return 0; 893 894 return move_freepages(zone, start_page, end_page, migratetype); 895} 896 897static void change_pageblock_range(struct page *pageblock_page, 898 int start_order, int migratetype) 899{ 900 int nr_pageblocks = 1 << (start_order - pageblock_order); 901 902 while (nr_pageblocks--) { 903 set_pageblock_migratetype(pageblock_page, migratetype); 904 pageblock_page += pageblock_nr_pages; 905 } 906} 907 908/* Remove an element from the buddy allocator from the fallback list */ 909static inline struct page * 910__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 911{ 912 struct free_area * area; 913 int current_order; 914 struct page *page; 915 int migratetype, i; 916 917 /* Find the largest possible block of pages in the other list */ 918 for (current_order = MAX_ORDER-1; current_order >= order; 919 --current_order) { 920 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 921 migratetype = fallbacks[start_migratetype][i]; 922 923 /* MIGRATE_RESERVE handled later if necessary */ 924 if (migratetype == MIGRATE_RESERVE) 925 continue; 926 927 area = &(zone->free_area[current_order]); 928 if (list_empty(&area->free_list[migratetype])) 929 continue; 930 931 page = list_entry(area->free_list[migratetype].next, 932 struct page, lru); 933 area->nr_free--; 934 935 /* 936 * If breaking a large block of pages, move all free 937 * pages to the preferred allocation list. If falling 938 * back for a reclaimable kernel allocation, be more 939 * agressive about taking ownership of free pages 940 */ 941 if (unlikely(current_order >= (pageblock_order >> 1)) || 942 start_migratetype == MIGRATE_RECLAIMABLE || 943 page_group_by_mobility_disabled) { 944 unsigned long pages; 945 pages = move_freepages_block(zone, page, 946 start_migratetype); 947 948 /* Claim the whole block if over half of it is free */ 949 if (pages >= (1 << (pageblock_order-1)) || 950 page_group_by_mobility_disabled) 951 set_pageblock_migratetype(page, 952 start_migratetype); 953 954 migratetype = start_migratetype; 955 } 956 957 /* Remove the page from the freelists */ 958 list_del(&page->lru); 959 rmv_page_order(page); 960 961 /* Take ownership for orders >= pageblock_order */ 962 if (current_order >= pageblock_order) 963 change_pageblock_range(page, current_order, 964 start_migratetype); 965 966 expand(zone, page, order, current_order, area, migratetype); 967 968 trace_mm_page_alloc_extfrag(page, order, current_order, 969 start_migratetype, migratetype); 970 971 return page; 972 } 973 } 974 975 return NULL; 976} 977 978/* 979 * Do the hard work of removing an element from the buddy allocator. 980 * Call me with the zone->lock already held. 981 */ 982static struct page *__rmqueue(struct zone *zone, unsigned int order, 983 int migratetype) 984{ 985 struct page *page; 986 987retry_reserve: 988 page = __rmqueue_smallest(zone, order, migratetype); 989 990 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 991 page = __rmqueue_fallback(zone, order, migratetype); 992 993 /* 994 * Use MIGRATE_RESERVE rather than fail an allocation. goto 995 * is used because __rmqueue_smallest is an inline function 996 * and we want just one call site 997 */ 998 if (!page) { 999 migratetype = MIGRATE_RESERVE; 1000 goto retry_reserve; 1001 } 1002 } 1003 1004 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1005 return page; 1006} 1007 1008/* 1009 * Obtain a specified number of elements from the buddy allocator, all under 1010 * a single hold of the lock, for efficiency. Add them to the supplied list. 1011 * Returns the number of new pages which were placed at *list. 1012 */ 1013static int rmqueue_bulk(struct zone *zone, unsigned int order, 1014 unsigned long count, struct list_head *list, 1015 int migratetype, int cold) 1016{ 1017 int i; 1018 1019 spin_lock(&zone->lock); 1020 for (i = 0; i < count; ++i) { 1021 struct page *page = __rmqueue(zone, order, migratetype); 1022 if (unlikely(page == NULL)) 1023 break; 1024 1025 /* 1026 * Split buddy pages returned by expand() are received here 1027 * in physical page order. The page is added to the callers and 1028 * list and the list head then moves forward. From the callers 1029 * perspective, the linked list is ordered by page number in 1030 * some conditions. This is useful for IO devices that can 1031 * merge IO requests if the physical pages are ordered 1032 * properly. 1033 */ 1034 if (likely(cold == 0)) 1035 list_add(&page->lru, list); 1036 else 1037 list_add_tail(&page->lru, list); 1038 set_page_private(page, migratetype); 1039 list = &page->lru; 1040 } 1041 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1042 spin_unlock(&zone->lock); 1043 return i; 1044} 1045 1046#ifdef CONFIG_NUMA 1047/* 1048 * Called from the vmstat counter updater to drain pagesets of this 1049 * currently executing processor on remote nodes after they have 1050 * expired. 1051 * 1052 * Note that this function must be called with the thread pinned to 1053 * a single processor. 1054 */ 1055void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1056{ 1057 unsigned long flags; 1058 int to_drain; 1059 1060 local_irq_save(flags); 1061 if (pcp->count >= pcp->batch) 1062 to_drain = pcp->batch; 1063 else 1064 to_drain = pcp->count; 1065 free_pcppages_bulk(zone, to_drain, pcp); 1066 pcp->count -= to_drain; 1067 local_irq_restore(flags); 1068} 1069#endif 1070 1071/* 1072 * Drain pages of the indicated processor. 1073 * 1074 * The processor must either be the current processor and the 1075 * thread pinned to the current processor or a processor that 1076 * is not online. 1077 */ 1078static void drain_pages(unsigned int cpu) 1079{ 1080 unsigned long flags; 1081 struct zone *zone; 1082 1083 for_each_populated_zone(zone) { 1084 struct per_cpu_pageset *pset; 1085 struct per_cpu_pages *pcp; 1086 1087 local_irq_save(flags); 1088 pset = per_cpu_ptr(zone->pageset, cpu); 1089 1090 pcp = &pset->pcp; 1091 free_pcppages_bulk(zone, pcp->count, pcp); 1092 pcp->count = 0; 1093 local_irq_restore(flags); 1094 } 1095} 1096 1097/* 1098 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1099 */ 1100void drain_local_pages(void *arg) 1101{ 1102 drain_pages(smp_processor_id()); 1103} 1104 1105/* 1106 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1107 */ 1108void drain_all_pages(void) 1109{ 1110 on_each_cpu(drain_local_pages, NULL, 1); 1111} 1112 1113#ifdef CONFIG_HIBERNATION 1114 1115void mark_free_pages(struct zone *zone) 1116{ 1117 unsigned long pfn, max_zone_pfn; 1118 unsigned long flags; 1119 int order, t; 1120 struct list_head *curr; 1121 1122 if (!zone->spanned_pages) 1123 return; 1124 1125 spin_lock_irqsave(&zone->lock, flags); 1126 1127 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1128 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1129 if (pfn_valid(pfn)) { 1130 struct page *page = pfn_to_page(pfn); 1131 1132 if (!swsusp_page_is_forbidden(page)) 1133 swsusp_unset_page_free(page); 1134 } 1135 1136 for_each_migratetype_order(order, t) { 1137 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1138 unsigned long i; 1139 1140 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1141 for (i = 0; i < (1UL << order); i++) 1142 swsusp_set_page_free(pfn_to_page(pfn + i)); 1143 } 1144 } 1145 spin_unlock_irqrestore(&zone->lock, flags); 1146} 1147#endif /* CONFIG_PM */ 1148 1149/* 1150 * Free a 0-order page 1151 * cold == 1 ? free a cold page : free a hot page 1152 */ 1153void free_hot_cold_page(struct page *page, int cold) 1154{ 1155 struct zone *zone = page_zone(page); 1156 struct per_cpu_pages *pcp; 1157 unsigned long flags; 1158 int migratetype; 1159 int wasMlocked = __TestClearPageMlocked(page); 1160 1161 if (!free_pages_prepare(page, 0)) 1162 return; 1163 1164 migratetype = get_pageblock_migratetype(page); 1165 set_page_private(page, migratetype); 1166 local_irq_save(flags); 1167 if (unlikely(wasMlocked)) 1168 free_page_mlock(page); 1169 __count_vm_event(PGFREE); 1170 1171 /* 1172 * We only track unmovable, reclaimable and movable on pcp lists. 1173 * Free ISOLATE pages back to the allocator because they are being 1174 * offlined but treat RESERVE as movable pages so we can get those 1175 * areas back if necessary. Otherwise, we may have to free 1176 * excessively into the page allocator 1177 */ 1178 if (migratetype >= MIGRATE_PCPTYPES) { 1179 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1180 free_one_page(zone, page, 0, migratetype); 1181 goto out; 1182 } 1183 migratetype = MIGRATE_MOVABLE; 1184 } 1185 1186 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1187 if (cold) 1188 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1189 else 1190 list_add(&page->lru, &pcp->lists[migratetype]); 1191 pcp->count++; 1192 if (pcp->count >= pcp->high) { 1193 free_pcppages_bulk(zone, pcp->batch, pcp); 1194 pcp->count -= pcp->batch; 1195 } 1196 1197out: 1198 local_irq_restore(flags); 1199} 1200 1201/* 1202 * split_page takes a non-compound higher-order page, and splits it into 1203 * n (1<<order) sub-pages: page[0..n] 1204 * Each sub-page must be freed individually. 1205 * 1206 * Note: this is probably too low level an operation for use in drivers. 1207 * Please consult with lkml before using this in your driver. 1208 */ 1209void split_page(struct page *page, unsigned int order) 1210{ 1211 int i; 1212 1213 VM_BUG_ON(PageCompound(page)); 1214 VM_BUG_ON(!page_count(page)); 1215 1216#ifdef CONFIG_KMEMCHECK 1217 /* 1218 * Split shadow pages too, because free(page[0]) would 1219 * otherwise free the whole shadow. 1220 */ 1221 if (kmemcheck_page_is_tracked(page)) 1222 split_page(virt_to_page(page[0].shadow), order); 1223#endif 1224 1225 for (i = 1; i < (1 << order); i++) 1226 set_page_refcounted(page + i); 1227} 1228 1229/* 1230 * Similar to split_page except the page is already free. As this is only 1231 * being used for migration, the migratetype of the block also changes. 1232 * As this is called with interrupts disabled, the caller is responsible 1233 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1234 * are enabled. 1235 * 1236 * Note: this is probably too low level an operation for use in drivers. 1237 * Please consult with lkml before using this in your driver. 1238 */ 1239int split_free_page(struct page *page) 1240{ 1241 unsigned int order; 1242 unsigned long watermark; 1243 struct zone *zone; 1244 1245 BUG_ON(!PageBuddy(page)); 1246 1247 zone = page_zone(page); 1248 order = page_order(page); 1249 1250 /* Obey watermarks as if the page was being allocated */ 1251 watermark = low_wmark_pages(zone) + (1 << order); 1252 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1253 return 0; 1254 1255 /* Remove page from free list */ 1256 list_del(&page->lru); 1257 zone->free_area[order].nr_free--; 1258 rmv_page_order(page); 1259 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1260 1261 /* Split into individual pages */ 1262 set_page_refcounted(page); 1263 split_page(page, order); 1264 1265 if (order >= pageblock_order - 1) { 1266 struct page *endpage = page + (1 << order) - 1; 1267 for (; page < endpage; page += pageblock_nr_pages) 1268 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1269 } 1270 1271 return 1 << order; 1272} 1273 1274/* 1275 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1276 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1277 * or two. 1278 */ 1279static inline 1280struct page *buffered_rmqueue(struct zone *preferred_zone, 1281 struct zone *zone, int order, gfp_t gfp_flags, 1282 int migratetype) 1283{ 1284 unsigned long flags; 1285 struct page *page; 1286 int cold = !!(gfp_flags & __GFP_COLD); 1287 1288again: 1289 if (likely(order == 0)) { 1290 struct per_cpu_pages *pcp; 1291 struct list_head *list; 1292 1293 local_irq_save(flags); 1294 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1295 list = &pcp->lists[migratetype]; 1296 if (list_empty(list)) { 1297 pcp->count += rmqueue_bulk(zone, 0, 1298 pcp->batch, list, 1299 migratetype, cold); 1300 if (unlikely(list_empty(list))) 1301 goto failed; 1302 } 1303 1304 if (cold) 1305 page = list_entry(list->prev, struct page, lru); 1306 else 1307 page = list_entry(list->next, struct page, lru); 1308 1309 list_del(&page->lru); 1310 pcp->count--; 1311 } else { 1312 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1313 /* 1314 * __GFP_NOFAIL is not to be used in new code. 1315 * 1316 * All __GFP_NOFAIL callers should be fixed so that they 1317 * properly detect and handle allocation failures. 1318 * 1319 * We most definitely don't want callers attempting to 1320 * allocate greater than order-1 page units with 1321 * __GFP_NOFAIL. 1322 */ 1323 WARN_ON_ONCE(order > 1); 1324 } 1325 spin_lock_irqsave(&zone->lock, flags); 1326 page = __rmqueue(zone, order, migratetype); 1327 spin_unlock(&zone->lock); 1328 if (!page) 1329 goto failed; 1330 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1331 } 1332 1333 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1334 zone_statistics(preferred_zone, zone); 1335 local_irq_restore(flags); 1336 1337 VM_BUG_ON(bad_range(zone, page)); 1338 if (prep_new_page(page, order, gfp_flags)) 1339 goto again; 1340 return page; 1341 1342failed: 1343 local_irq_restore(flags); 1344 return NULL; 1345} 1346 1347/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1348#define ALLOC_WMARK_MIN WMARK_MIN 1349#define ALLOC_WMARK_LOW WMARK_LOW 1350#define ALLOC_WMARK_HIGH WMARK_HIGH 1351#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1352 1353/* Mask to get the watermark bits */ 1354#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1355 1356#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1357#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1358#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1359 1360#ifdef CONFIG_FAIL_PAGE_ALLOC 1361 1362static struct fail_page_alloc_attr { 1363 struct fault_attr attr; 1364 1365 u32 ignore_gfp_highmem; 1366 u32 ignore_gfp_wait; 1367 u32 min_order; 1368 1369#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1370 1371 struct dentry *ignore_gfp_highmem_file; 1372 struct dentry *ignore_gfp_wait_file; 1373 struct dentry *min_order_file; 1374 1375#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1376 1377} fail_page_alloc = { 1378 .attr = FAULT_ATTR_INITIALIZER, 1379 .ignore_gfp_wait = 1, 1380 .ignore_gfp_highmem = 1, 1381 .min_order = 1, 1382}; 1383 1384static int __init setup_fail_page_alloc(char *str) 1385{ 1386 return setup_fault_attr(&fail_page_alloc.attr, str); 1387} 1388__setup("fail_page_alloc=", setup_fail_page_alloc); 1389 1390static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1391{ 1392 if (order < fail_page_alloc.min_order) 1393 return 0; 1394 if (gfp_mask & __GFP_NOFAIL) 1395 return 0; 1396 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1397 return 0; 1398 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1399 return 0; 1400 1401 return should_fail(&fail_page_alloc.attr, 1 << order); 1402} 1403 1404#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1405 1406static int __init fail_page_alloc_debugfs(void) 1407{ 1408 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1409 struct dentry *dir; 1410 int err; 1411 1412 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1413 "fail_page_alloc"); 1414 if (err) 1415 return err; 1416 dir = fail_page_alloc.attr.dentries.dir; 1417 1418 fail_page_alloc.ignore_gfp_wait_file = 1419 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1420 &fail_page_alloc.ignore_gfp_wait); 1421 1422 fail_page_alloc.ignore_gfp_highmem_file = 1423 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1424 &fail_page_alloc.ignore_gfp_highmem); 1425 fail_page_alloc.min_order_file = 1426 debugfs_create_u32("min-order", mode, dir, 1427 &fail_page_alloc.min_order); 1428 1429 if (!fail_page_alloc.ignore_gfp_wait_file || 1430 !fail_page_alloc.ignore_gfp_highmem_file || 1431 !fail_page_alloc.min_order_file) { 1432 err = -ENOMEM; 1433 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1434 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1435 debugfs_remove(fail_page_alloc.min_order_file); 1436 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1437 } 1438 1439 return err; 1440} 1441 1442late_initcall(fail_page_alloc_debugfs); 1443 1444#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1445 1446#else /* CONFIG_FAIL_PAGE_ALLOC */ 1447 1448static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1449{ 1450 return 0; 1451} 1452 1453#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1454 1455/* 1456 * Return true if free pages are above 'mark'. This takes into account the order 1457 * of the allocation. 1458 */ 1459static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1460 int classzone_idx, int alloc_flags, long free_pages) 1461{ 1462 /* free_pages my go negative - that's OK */ 1463 long min = mark; 1464 int o; 1465 1466 free_pages -= (1 << order) + 1; 1467 if (alloc_flags & ALLOC_HIGH) 1468 min -= min / 2; 1469 if (alloc_flags & ALLOC_HARDER) 1470 min -= min / 4; 1471 1472 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1473 return false; 1474 for (o = 0; o < order; o++) { 1475 /* At the next order, this order's pages become unavailable */ 1476 free_pages -= z->free_area[o].nr_free << o; 1477 1478 /* Require fewer higher order pages to be free */ 1479 min >>= 1; 1480 1481 if (free_pages <= min) 1482 return false; 1483 } 1484 return true; 1485} 1486 1487bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1488 int classzone_idx, int alloc_flags) 1489{ 1490 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1491 zone_page_state(z, NR_FREE_PAGES)); 1492} 1493 1494bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1495 int classzone_idx, int alloc_flags) 1496{ 1497 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1498 1499 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1500 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1501 1502 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1503 free_pages); 1504} 1505 1506#ifdef CONFIG_NUMA 1507/* 1508 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1509 * skip over zones that are not allowed by the cpuset, or that have 1510 * been recently (in last second) found to be nearly full. See further 1511 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1512 * that have to skip over a lot of full or unallowed zones. 1513 * 1514 * If the zonelist cache is present in the passed in zonelist, then 1515 * returns a pointer to the allowed node mask (either the current 1516 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1517 * 1518 * If the zonelist cache is not available for this zonelist, does 1519 * nothing and returns NULL. 1520 * 1521 * If the fullzones BITMAP in the zonelist cache is stale (more than 1522 * a second since last zap'd) then we zap it out (clear its bits.) 1523 * 1524 * We hold off even calling zlc_setup, until after we've checked the 1525 * first zone in the zonelist, on the theory that most allocations will 1526 * be satisfied from that first zone, so best to examine that zone as 1527 * quickly as we can. 1528 */ 1529static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1530{ 1531 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1532 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1533 1534 zlc = zonelist->zlcache_ptr; 1535 if (!zlc) 1536 return NULL; 1537 1538 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1539 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1540 zlc->last_full_zap = jiffies; 1541 } 1542 1543 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1544 &cpuset_current_mems_allowed : 1545 &node_states[N_HIGH_MEMORY]; 1546 return allowednodes; 1547} 1548 1549/* 1550 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1551 * if it is worth looking at further for free memory: 1552 * 1) Check that the zone isn't thought to be full (doesn't have its 1553 * bit set in the zonelist_cache fullzones BITMAP). 1554 * 2) Check that the zones node (obtained from the zonelist_cache 1555 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1556 * Return true (non-zero) if zone is worth looking at further, or 1557 * else return false (zero) if it is not. 1558 * 1559 * This check -ignores- the distinction between various watermarks, 1560 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1561 * found to be full for any variation of these watermarks, it will 1562 * be considered full for up to one second by all requests, unless 1563 * we are so low on memory on all allowed nodes that we are forced 1564 * into the second scan of the zonelist. 1565 * 1566 * In the second scan we ignore this zonelist cache and exactly 1567 * apply the watermarks to all zones, even it is slower to do so. 1568 * We are low on memory in the second scan, and should leave no stone 1569 * unturned looking for a free page. 1570 */ 1571static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1572 nodemask_t *allowednodes) 1573{ 1574 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1575 int i; /* index of *z in zonelist zones */ 1576 int n; /* node that zone *z is on */ 1577 1578 zlc = zonelist->zlcache_ptr; 1579 if (!zlc) 1580 return 1; 1581 1582 i = z - zonelist->_zonerefs; 1583 n = zlc->z_to_n[i]; 1584 1585 /* This zone is worth trying if it is allowed but not full */ 1586 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1587} 1588 1589/* 1590 * Given 'z' scanning a zonelist, set the corresponding bit in 1591 * zlc->fullzones, so that subsequent attempts to allocate a page 1592 * from that zone don't waste time re-examining it. 1593 */ 1594static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1595{ 1596 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1597 int i; /* index of *z in zonelist zones */ 1598 1599 zlc = zonelist->zlcache_ptr; 1600 if (!zlc) 1601 return; 1602 1603 i = z - zonelist->_zonerefs; 1604 1605 set_bit(i, zlc->fullzones); 1606} 1607 1608#else /* CONFIG_NUMA */ 1609 1610static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1611{ 1612 return NULL; 1613} 1614 1615static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1616 nodemask_t *allowednodes) 1617{ 1618 return 1; 1619} 1620 1621static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1622{ 1623} 1624#endif /* CONFIG_NUMA */ 1625 1626/* 1627 * get_page_from_freelist goes through the zonelist trying to allocate 1628 * a page. 1629 */ 1630static struct page * 1631get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1632 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1633 struct zone *preferred_zone, int migratetype) 1634{ 1635 struct zoneref *z; 1636 struct page *page = NULL; 1637 int classzone_idx; 1638 struct zone *zone; 1639 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1640 int zlc_active = 0; /* set if using zonelist_cache */ 1641 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1642 1643 classzone_idx = zone_idx(preferred_zone); 1644zonelist_scan: 1645 /* 1646 * Scan zonelist, looking for a zone with enough free. 1647 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1648 */ 1649 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1650 high_zoneidx, nodemask) { 1651 if (NUMA_BUILD && zlc_active && 1652 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1653 continue; 1654 if ((alloc_flags & ALLOC_CPUSET) && 1655 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1656 goto try_next_zone; 1657 1658 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1659 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1660 unsigned long mark; 1661 int ret; 1662 1663 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1664 if (zone_watermark_ok(zone, order, mark, 1665 classzone_idx, alloc_flags)) 1666 goto try_this_zone; 1667 1668 if (zone_reclaim_mode == 0) 1669 goto this_zone_full; 1670 1671 ret = zone_reclaim(zone, gfp_mask, order); 1672 switch (ret) { 1673 case ZONE_RECLAIM_NOSCAN: 1674 /* did not scan */ 1675 goto try_next_zone; 1676 case ZONE_RECLAIM_FULL: 1677 /* scanned but unreclaimable */ 1678 goto this_zone_full; 1679 default: 1680 /* did we reclaim enough */ 1681 if (!zone_watermark_ok(zone, order, mark, 1682 classzone_idx, alloc_flags)) 1683 goto this_zone_full; 1684 } 1685 } 1686 1687try_this_zone: 1688 page = buffered_rmqueue(preferred_zone, zone, order, 1689 gfp_mask, migratetype); 1690 if (page) 1691 break; 1692this_zone_full: 1693 if (NUMA_BUILD) 1694 zlc_mark_zone_full(zonelist, z); 1695try_next_zone: 1696 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1697 /* 1698 * we do zlc_setup after the first zone is tried but only 1699 * if there are multiple nodes make it worthwhile 1700 */ 1701 allowednodes = zlc_setup(zonelist, alloc_flags); 1702 zlc_active = 1; 1703 did_zlc_setup = 1; 1704 } 1705 } 1706 1707 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1708 /* Disable zlc cache for second zonelist scan */ 1709 zlc_active = 0; 1710 goto zonelist_scan; 1711 } 1712 return page; 1713} 1714 1715static inline int 1716should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1717 unsigned long pages_reclaimed) 1718{ 1719 /* Do not loop if specifically requested */ 1720 if (gfp_mask & __GFP_NORETRY) 1721 return 0; 1722 1723 /* 1724 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1725 * means __GFP_NOFAIL, but that may not be true in other 1726 * implementations. 1727 */ 1728 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1729 return 1; 1730 1731 /* 1732 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1733 * specified, then we retry until we no longer reclaim any pages 1734 * (above), or we've reclaimed an order of pages at least as 1735 * large as the allocation's order. In both cases, if the 1736 * allocation still fails, we stop retrying. 1737 */ 1738 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1739 return 1; 1740 1741 /* 1742 * Don't let big-order allocations loop unless the caller 1743 * explicitly requests that. 1744 */ 1745 if (gfp_mask & __GFP_NOFAIL) 1746 return 1; 1747 1748 return 0; 1749} 1750 1751static inline struct page * 1752__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1753 struct zonelist *zonelist, enum zone_type high_zoneidx, 1754 nodemask_t *nodemask, struct zone *preferred_zone, 1755 int migratetype) 1756{ 1757 struct page *page; 1758 1759 /* Acquire the OOM killer lock for the zones in zonelist */ 1760 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1761 schedule_timeout_uninterruptible(1); 1762 return NULL; 1763 } 1764 1765 /* 1766 * Go through the zonelist yet one more time, keep very high watermark 1767 * here, this is only to catch a parallel oom killing, we must fail if 1768 * we're still under heavy pressure. 1769 */ 1770 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1771 order, zonelist, high_zoneidx, 1772 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1773 preferred_zone, migratetype); 1774 if (page) 1775 goto out; 1776 1777 if (!(gfp_mask & __GFP_NOFAIL)) { 1778 /* The OOM killer will not help higher order allocs */ 1779 if (order > PAGE_ALLOC_COSTLY_ORDER) 1780 goto out; 1781 /* The OOM killer does not needlessly kill tasks for lowmem */ 1782 if (high_zoneidx < ZONE_NORMAL) 1783 goto out; 1784 /* 1785 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1786 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1787 * The caller should handle page allocation failure by itself if 1788 * it specifies __GFP_THISNODE. 1789 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1790 */ 1791 if (gfp_mask & __GFP_THISNODE) 1792 goto out; 1793 } 1794 /* Exhausted what can be done so it's blamo time */ 1795 out_of_memory(zonelist, gfp_mask, order, nodemask); 1796 1797out: 1798 clear_zonelist_oom(zonelist, gfp_mask); 1799 return page; 1800} 1801 1802#ifdef CONFIG_COMPACTION 1803/* Try memory compaction for high-order allocations before reclaim */ 1804static struct page * 1805__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1806 struct zonelist *zonelist, enum zone_type high_zoneidx, 1807 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1808 int migratetype, unsigned long *did_some_progress, 1809 bool sync_migration) 1810{ 1811 struct page *page; 1812 1813 if (!order || compaction_deferred(preferred_zone)) 1814 return NULL; 1815 1816 current->flags |= PF_MEMALLOC; 1817 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 1818 nodemask, sync_migration); 1819 current->flags &= ~PF_MEMALLOC; 1820 if (*did_some_progress != COMPACT_SKIPPED) { 1821 1822 /* Page migration frees to the PCP lists but we want merging */ 1823 drain_pages(get_cpu()); 1824 put_cpu(); 1825 1826 page = get_page_from_freelist(gfp_mask, nodemask, 1827 order, zonelist, high_zoneidx, 1828 alloc_flags, preferred_zone, 1829 migratetype); 1830 if (page) { 1831 preferred_zone->compact_considered = 0; 1832 preferred_zone->compact_defer_shift = 0; 1833 count_vm_event(COMPACTSUCCESS); 1834 return page; 1835 } 1836 1837 /* 1838 * It's bad if compaction run occurs and fails. 1839 * The most likely reason is that pages exist, 1840 * but not enough to satisfy watermarks. 1841 */ 1842 count_vm_event(COMPACTFAIL); 1843 defer_compaction(preferred_zone); 1844 1845 cond_resched(); 1846 } 1847 1848 return NULL; 1849} 1850#else 1851static inline struct page * 1852__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1853 struct zonelist *zonelist, enum zone_type high_zoneidx, 1854 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1855 int migratetype, unsigned long *did_some_progress, 1856 bool sync_migration) 1857{ 1858 return NULL; 1859} 1860#endif /* CONFIG_COMPACTION */ 1861 1862/* The really slow allocator path where we enter direct reclaim */ 1863static inline struct page * 1864__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1865 struct zonelist *zonelist, enum zone_type high_zoneidx, 1866 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1867 int migratetype, unsigned long *did_some_progress) 1868{ 1869 struct page *page = NULL; 1870 struct reclaim_state reclaim_state; 1871 bool drained = false; 1872 1873 cond_resched(); 1874 1875 /* We now go into synchronous reclaim */ 1876 cpuset_memory_pressure_bump(); 1877 current->flags |= PF_MEMALLOC; 1878 lockdep_set_current_reclaim_state(gfp_mask); 1879 reclaim_state.reclaimed_slab = 0; 1880 current->reclaim_state = &reclaim_state; 1881 1882 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1883 1884 current->reclaim_state = NULL; 1885 lockdep_clear_current_reclaim_state(); 1886 current->flags &= ~PF_MEMALLOC; 1887 1888 cond_resched(); 1889 1890 if (unlikely(!(*did_some_progress))) 1891 return NULL; 1892 1893retry: 1894 page = get_page_from_freelist(gfp_mask, nodemask, order, 1895 zonelist, high_zoneidx, 1896 alloc_flags, preferred_zone, 1897 migratetype); 1898 1899 /* 1900 * If an allocation failed after direct reclaim, it could be because 1901 * pages are pinned on the per-cpu lists. Drain them and try again 1902 */ 1903 if (!page && !drained) { 1904 drain_all_pages(); 1905 drained = true; 1906 goto retry; 1907 } 1908 1909 return page; 1910} 1911 1912/* 1913 * This is called in the allocator slow-path if the allocation request is of 1914 * sufficient urgency to ignore watermarks and take other desperate measures 1915 */ 1916static inline struct page * 1917__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1918 struct zonelist *zonelist, enum zone_type high_zoneidx, 1919 nodemask_t *nodemask, struct zone *preferred_zone, 1920 int migratetype) 1921{ 1922 struct page *page; 1923 1924 do { 1925 page = get_page_from_freelist(gfp_mask, nodemask, order, 1926 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1927 preferred_zone, migratetype); 1928 1929 if (!page && gfp_mask & __GFP_NOFAIL) 1930 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 1931 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1932 1933 return page; 1934} 1935 1936static inline 1937void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1938 enum zone_type high_zoneidx, 1939 enum zone_type classzone_idx) 1940{ 1941 struct zoneref *z; 1942 struct zone *zone; 1943 1944 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1945 wakeup_kswapd(zone, order, classzone_idx); 1946} 1947 1948static inline int 1949gfp_to_alloc_flags(gfp_t gfp_mask) 1950{ 1951 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1952 const gfp_t wait = gfp_mask & __GFP_WAIT; 1953 1954 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1955 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 1956 1957 /* 1958 * The caller may dip into page reserves a bit more if the caller 1959 * cannot run direct reclaim, or if the caller has realtime scheduling 1960 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1961 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1962 */ 1963 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 1964 1965 if (!wait) { 1966 /* 1967 * Not worth trying to allocate harder for 1968 * __GFP_NOMEMALLOC even if it can't schedule. 1969 */ 1970 if (!(gfp_mask & __GFP_NOMEMALLOC)) 1971 alloc_flags |= ALLOC_HARDER; 1972 /* 1973 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1974 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1975 */ 1976 alloc_flags &= ~ALLOC_CPUSET; 1977 } else if (unlikely(rt_task(current)) && !in_interrupt()) 1978 alloc_flags |= ALLOC_HARDER; 1979 1980 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1981 if (!in_interrupt() && 1982 ((current->flags & PF_MEMALLOC) || 1983 unlikely(test_thread_flag(TIF_MEMDIE)))) 1984 alloc_flags |= ALLOC_NO_WATERMARKS; 1985 } 1986 1987 return alloc_flags; 1988} 1989 1990static inline struct page * 1991__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1992 struct zonelist *zonelist, enum zone_type high_zoneidx, 1993 nodemask_t *nodemask, struct zone *preferred_zone, 1994 int migratetype) 1995{ 1996 const gfp_t wait = gfp_mask & __GFP_WAIT; 1997 struct page *page = NULL; 1998 int alloc_flags; 1999 unsigned long pages_reclaimed = 0; 2000 unsigned long did_some_progress; 2001 bool sync_migration = false; 2002 2003 /* 2004 * In the slowpath, we sanity check order to avoid ever trying to 2005 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2006 * be using allocators in order of preference for an area that is 2007 * too large. 2008 */ 2009 if (order >= MAX_ORDER) { 2010 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2011 return NULL; 2012 } 2013 2014 /* 2015 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2016 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2017 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2018 * using a larger set of nodes after it has established that the 2019 * allowed per node queues are empty and that nodes are 2020 * over allocated. 2021 */ 2022 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2023 goto nopage; 2024 2025restart: 2026 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2027 wake_all_kswapd(order, zonelist, high_zoneidx, 2028 zone_idx(preferred_zone)); 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 (current->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 /* 2067 * Try direct compaction. The first pass is asynchronous. Subsequent 2068 * attempts after direct reclaim are synchronous 2069 */ 2070 page = __alloc_pages_direct_compact(gfp_mask, order, 2071 zonelist, high_zoneidx, 2072 nodemask, 2073 alloc_flags, preferred_zone, 2074 migratetype, &did_some_progress, 2075 sync_migration); 2076 if (page) 2077 goto got_pg; 2078 sync_migration = true; 2079 2080 /* Try direct reclaim and then allocating */ 2081 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2082 zonelist, high_zoneidx, 2083 nodemask, 2084 alloc_flags, preferred_zone, 2085 migratetype, &did_some_progress); 2086 if (page) 2087 goto got_pg; 2088 2089 /* 2090 * If we failed to make any progress reclaiming, then we are 2091 * running out of options and have to consider going OOM 2092 */ 2093 if (!did_some_progress) { 2094 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2095 if (oom_killer_disabled) 2096 goto nopage; 2097 page = __alloc_pages_may_oom(gfp_mask, order, 2098 zonelist, high_zoneidx, 2099 nodemask, preferred_zone, 2100 migratetype); 2101 if (page) 2102 goto got_pg; 2103 2104 if (!(gfp_mask & __GFP_NOFAIL)) { 2105 /* 2106 * The oom killer is not called for high-order 2107 * allocations that may fail, so if no progress 2108 * is being made, there are no other options and 2109 * retrying is unlikely to help. 2110 */ 2111 if (order > PAGE_ALLOC_COSTLY_ORDER) 2112 goto nopage; 2113 /* 2114 * The oom killer is not called for lowmem 2115 * allocations to prevent needlessly killing 2116 * innocent tasks. 2117 */ 2118 if (high_zoneidx < ZONE_NORMAL) 2119 goto nopage; 2120 } 2121 2122 goto restart; 2123 } 2124 } 2125 2126 /* Check if we should retry the allocation */ 2127 pages_reclaimed += did_some_progress; 2128 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 2129 /* Wait for some write requests to complete then retry */ 2130 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2131 goto rebalance; 2132 } else { 2133 /* 2134 * High-order allocations do not necessarily loop after 2135 * direct reclaim and reclaim/compaction depends on compaction 2136 * being called after reclaim so call directly if necessary 2137 */ 2138 page = __alloc_pages_direct_compact(gfp_mask, order, 2139 zonelist, high_zoneidx, 2140 nodemask, 2141 alloc_flags, preferred_zone, 2142 migratetype, &did_some_progress, 2143 sync_migration); 2144 if (page) 2145 goto got_pg; 2146 } 2147 2148nopage: 2149 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 2150 printk(KERN_WARNING "%s: page allocation failure." 2151 " order:%d, mode:0x%x\n", 2152 current->comm, order, gfp_mask); 2153 dump_stack(); 2154 show_mem(); 2155 } 2156 return page; 2157got_pg: 2158 if (kmemcheck_enabled) 2159 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2160 return page; 2161 2162} 2163 2164/* 2165 * This is the 'heart' of the zoned buddy allocator. 2166 */ 2167struct page * 2168__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2169 struct zonelist *zonelist, nodemask_t *nodemask) 2170{ 2171 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2172 struct zone *preferred_zone; 2173 struct page *page; 2174 int migratetype = allocflags_to_migratetype(gfp_mask); 2175 2176 gfp_mask &= gfp_allowed_mask; 2177 2178 lockdep_trace_alloc(gfp_mask); 2179 2180 might_sleep_if(gfp_mask & __GFP_WAIT); 2181 2182 if (should_fail_alloc_page(gfp_mask, order)) 2183 return NULL; 2184 2185 /* 2186 * Check the zones suitable for the gfp_mask contain at least one 2187 * valid zone. It's possible to have an empty zonelist as a result 2188 * of GFP_THISNODE and a memoryless node 2189 */ 2190 if (unlikely(!zonelist->_zonerefs->zone)) 2191 return NULL; 2192 2193 get_mems_allowed(); 2194 /* The preferred zone is used for statistics later */ 2195 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 2196 if (!preferred_zone) { 2197 put_mems_allowed(); 2198 return NULL; 2199 } 2200 2201 /* First allocation attempt */ 2202 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2203 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2204 preferred_zone, migratetype); 2205 if (unlikely(!page)) 2206 page = __alloc_pages_slowpath(gfp_mask, order, 2207 zonelist, high_zoneidx, nodemask, 2208 preferred_zone, migratetype); 2209 put_mems_allowed(); 2210 2211 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2212 return page; 2213} 2214EXPORT_SYMBOL(__alloc_pages_nodemask); 2215 2216/* 2217 * Common helper functions. 2218 */ 2219unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2220{ 2221 struct page *page; 2222 2223 /* 2224 * __get_free_pages() returns a 32-bit address, which cannot represent 2225 * a highmem page 2226 */ 2227 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2228 2229 page = alloc_pages(gfp_mask, order); 2230 if (!page) 2231 return 0; 2232 return (unsigned long) page_address(page); 2233} 2234EXPORT_SYMBOL(__get_free_pages); 2235 2236unsigned long get_zeroed_page(gfp_t gfp_mask) 2237{ 2238 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2239} 2240EXPORT_SYMBOL(get_zeroed_page); 2241 2242void __pagevec_free(struct pagevec *pvec) 2243{ 2244 int i = pagevec_count(pvec); 2245 2246 while (--i >= 0) { 2247 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2248 free_hot_cold_page(pvec->pages[i], pvec->cold); 2249 } 2250} 2251 2252void __free_pages(struct page *page, unsigned int order) 2253{ 2254 if (put_page_testzero(page)) { 2255 if (order == 0) 2256 free_hot_cold_page(page, 0); 2257 else 2258 __free_pages_ok(page, order); 2259 } 2260} 2261 2262EXPORT_SYMBOL(__free_pages); 2263 2264void free_pages(unsigned long addr, unsigned int order) 2265{ 2266 if (addr != 0) { 2267 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2268 __free_pages(virt_to_page((void *)addr), order); 2269 } 2270} 2271 2272EXPORT_SYMBOL(free_pages); 2273 2274/** 2275 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2276 * @size: the number of bytes to allocate 2277 * @gfp_mask: GFP flags for the allocation 2278 * 2279 * This function is similar to alloc_pages(), except that it allocates the 2280 * minimum number of pages to satisfy the request. alloc_pages() can only 2281 * allocate memory in power-of-two pages. 2282 * 2283 * This function is also limited by MAX_ORDER. 2284 * 2285 * Memory allocated by this function must be released by free_pages_exact(). 2286 */ 2287void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2288{ 2289 unsigned int order = get_order(size); 2290 unsigned long addr; 2291 2292 addr = __get_free_pages(gfp_mask, order); 2293 if (addr) { 2294 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2295 unsigned long used = addr + PAGE_ALIGN(size); 2296 2297 split_page(virt_to_page((void *)addr), order); 2298 while (used < alloc_end) { 2299 free_page(used); 2300 used += PAGE_SIZE; 2301 } 2302 } 2303 2304 return (void *)addr; 2305} 2306EXPORT_SYMBOL(alloc_pages_exact); 2307 2308/** 2309 * free_pages_exact - release memory allocated via alloc_pages_exact() 2310 * @virt: the value returned by alloc_pages_exact. 2311 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2312 * 2313 * Release the memory allocated by a previous call to alloc_pages_exact. 2314 */ 2315void free_pages_exact(void *virt, size_t size) 2316{ 2317 unsigned long addr = (unsigned long)virt; 2318 unsigned long end = addr + PAGE_ALIGN(size); 2319 2320 while (addr < end) { 2321 free_page(addr); 2322 addr += PAGE_SIZE; 2323 } 2324} 2325EXPORT_SYMBOL(free_pages_exact); 2326 2327static unsigned int nr_free_zone_pages(int offset) 2328{ 2329 struct zoneref *z; 2330 struct zone *zone; 2331 2332 /* Just pick one node, since fallback list is circular */ 2333 unsigned int sum = 0; 2334 2335 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2336 2337 for_each_zone_zonelist(zone, z, zonelist, offset) { 2338 unsigned long size = zone->present_pages; 2339 unsigned long high = high_wmark_pages(zone); 2340 if (size > high) 2341 sum += size - high; 2342 } 2343 2344 return sum; 2345} 2346 2347/* 2348 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2349 */ 2350unsigned int nr_free_buffer_pages(void) 2351{ 2352 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2353} 2354EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2355 2356/* 2357 * Amount of free RAM allocatable within all zones 2358 */ 2359unsigned int nr_free_pagecache_pages(void) 2360{ 2361 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2362} 2363 2364static inline void show_node(struct zone *zone) 2365{ 2366 if (NUMA_BUILD) 2367 printk("Node %d ", zone_to_nid(zone)); 2368} 2369 2370void si_meminfo(struct sysinfo *val) 2371{ 2372 val->totalram = totalram_pages; 2373 val->sharedram = 0; 2374 val->freeram = global_page_state(NR_FREE_PAGES); 2375 val->bufferram = nr_blockdev_pages(); 2376 val->totalhigh = totalhigh_pages; 2377 val->freehigh = nr_free_highpages(); 2378 val->mem_unit = PAGE_SIZE; 2379} 2380 2381EXPORT_SYMBOL(si_meminfo); 2382 2383#ifdef CONFIG_NUMA 2384void si_meminfo_node(struct sysinfo *val, int nid) 2385{ 2386 pg_data_t *pgdat = NODE_DATA(nid); 2387 2388 val->totalram = pgdat->node_present_pages; 2389 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2390#ifdef CONFIG_HIGHMEM 2391 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2392 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2393 NR_FREE_PAGES); 2394#else 2395 val->totalhigh = 0; 2396 val->freehigh = 0; 2397#endif 2398 val->mem_unit = PAGE_SIZE; 2399} 2400#endif 2401 2402#define K(x) ((x) << (PAGE_SHIFT-10)) 2403 2404/* 2405 * Show free area list (used inside shift_scroll-lock stuff) 2406 * We also calculate the percentage fragmentation. We do this by counting the 2407 * memory on each free list with the exception of the first item on the list. 2408 */ 2409void show_free_areas(void) 2410{ 2411 int cpu; 2412 struct zone *zone; 2413 2414 for_each_populated_zone(zone) { 2415 show_node(zone); 2416 printk("%s per-cpu:\n", zone->name); 2417 2418 for_each_online_cpu(cpu) { 2419 struct per_cpu_pageset *pageset; 2420 2421 pageset = per_cpu_ptr(zone->pageset, cpu); 2422 2423 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2424 cpu, pageset->pcp.high, 2425 pageset->pcp.batch, pageset->pcp.count); 2426 } 2427 } 2428 2429 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2430 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2431 " unevictable:%lu" 2432 " dirty:%lu writeback:%lu unstable:%lu\n" 2433 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2434 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2435 global_page_state(NR_ACTIVE_ANON), 2436 global_page_state(NR_INACTIVE_ANON), 2437 global_page_state(NR_ISOLATED_ANON), 2438 global_page_state(NR_ACTIVE_FILE), 2439 global_page_state(NR_INACTIVE_FILE), 2440 global_page_state(NR_ISOLATED_FILE), 2441 global_page_state(NR_UNEVICTABLE), 2442 global_page_state(NR_FILE_DIRTY), 2443 global_page_state(NR_WRITEBACK), 2444 global_page_state(NR_UNSTABLE_NFS), 2445 global_page_state(NR_FREE_PAGES), 2446 global_page_state(NR_SLAB_RECLAIMABLE), 2447 global_page_state(NR_SLAB_UNRECLAIMABLE), 2448 global_page_state(NR_FILE_MAPPED), 2449 global_page_state(NR_SHMEM), 2450 global_page_state(NR_PAGETABLE), 2451 global_page_state(NR_BOUNCE)); 2452 2453 for_each_populated_zone(zone) { 2454 int i; 2455 2456 show_node(zone); 2457 printk("%s" 2458 " free:%lukB" 2459 " min:%lukB" 2460 " low:%lukB" 2461 " high:%lukB" 2462 " active_anon:%lukB" 2463 " inactive_anon:%lukB" 2464 " active_file:%lukB" 2465 " inactive_file:%lukB" 2466 " unevictable:%lukB" 2467 " isolated(anon):%lukB" 2468 " isolated(file):%lukB" 2469 " present:%lukB" 2470 " mlocked:%lukB" 2471 " dirty:%lukB" 2472 " writeback:%lukB" 2473 " mapped:%lukB" 2474 " shmem:%lukB" 2475 " slab_reclaimable:%lukB" 2476 " slab_unreclaimable:%lukB" 2477 " kernel_stack:%lukB" 2478 " pagetables:%lukB" 2479 " unstable:%lukB" 2480 " bounce:%lukB" 2481 " writeback_tmp:%lukB" 2482 " pages_scanned:%lu" 2483 " all_unreclaimable? %s" 2484 "\n", 2485 zone->name, 2486 K(zone_page_state(zone, NR_FREE_PAGES)), 2487 K(min_wmark_pages(zone)), 2488 K(low_wmark_pages(zone)), 2489 K(high_wmark_pages(zone)), 2490 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2491 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2492 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2493 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2494 K(zone_page_state(zone, NR_UNEVICTABLE)), 2495 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2496 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2497 K(zone->present_pages), 2498 K(zone_page_state(zone, NR_MLOCK)), 2499 K(zone_page_state(zone, NR_FILE_DIRTY)), 2500 K(zone_page_state(zone, NR_WRITEBACK)), 2501 K(zone_page_state(zone, NR_FILE_MAPPED)), 2502 K(zone_page_state(zone, NR_SHMEM)), 2503 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2504 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2505 zone_page_state(zone, NR_KERNEL_STACK) * 2506 THREAD_SIZE / 1024, 2507 K(zone_page_state(zone, NR_PAGETABLE)), 2508 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2509 K(zone_page_state(zone, NR_BOUNCE)), 2510 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2511 zone->pages_scanned, 2512 (zone->all_unreclaimable ? "yes" : "no") 2513 ); 2514 printk("lowmem_reserve[]:"); 2515 for (i = 0; i < MAX_NR_ZONES; i++) 2516 printk(" %lu", zone->lowmem_reserve[i]); 2517 printk("\n"); 2518 } 2519 2520 for_each_populated_zone(zone) { 2521 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2522 2523 show_node(zone); 2524 printk("%s: ", zone->name); 2525 2526 spin_lock_irqsave(&zone->lock, flags); 2527 for (order = 0; order < MAX_ORDER; order++) { 2528 nr[order] = zone->free_area[order].nr_free; 2529 total += nr[order] << order; 2530 } 2531 spin_unlock_irqrestore(&zone->lock, flags); 2532 for (order = 0; order < MAX_ORDER; order++) 2533 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2534 printk("= %lukB\n", K(total)); 2535 } 2536 2537 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2538 2539 show_swap_cache_info(); 2540} 2541 2542static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2543{ 2544 zoneref->zone = zone; 2545 zoneref->zone_idx = zone_idx(zone); 2546} 2547 2548/* 2549 * Builds allocation fallback zone lists. 2550 * 2551 * Add all populated zones of a node to the zonelist. 2552 */ 2553static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2554 int nr_zones, enum zone_type zone_type) 2555{ 2556 struct zone *zone; 2557 2558 BUG_ON(zone_type >= MAX_NR_ZONES); 2559 zone_type++; 2560 2561 do { 2562 zone_type--; 2563 zone = pgdat->node_zones + zone_type; 2564 if (populated_zone(zone)) { 2565 zoneref_set_zone(zone, 2566 &zonelist->_zonerefs[nr_zones++]); 2567 check_highest_zone(zone_type); 2568 } 2569 2570 } while (zone_type); 2571 return nr_zones; 2572} 2573 2574 2575/* 2576 * zonelist_order: 2577 * 0 = automatic detection of better ordering. 2578 * 1 = order by ([node] distance, -zonetype) 2579 * 2 = order by (-zonetype, [node] distance) 2580 * 2581 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2582 * the same zonelist. So only NUMA can configure this param. 2583 */ 2584#define ZONELIST_ORDER_DEFAULT 0 2585#define ZONELIST_ORDER_NODE 1 2586#define ZONELIST_ORDER_ZONE 2 2587 2588/* zonelist order in the kernel. 2589 * set_zonelist_order() will set this to NODE or ZONE. 2590 */ 2591static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2592static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2593 2594 2595#ifdef CONFIG_NUMA 2596/* The value user specified ....changed by config */ 2597static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2598/* string for sysctl */ 2599#define NUMA_ZONELIST_ORDER_LEN 16 2600char numa_zonelist_order[16] = "default"; 2601 2602/* 2603 * interface for configure zonelist ordering. 2604 * command line option "numa_zonelist_order" 2605 * = "[dD]efault - default, automatic configuration. 2606 * = "[nN]ode - order by node locality, then by zone within node 2607 * = "[zZ]one - order by zone, then by locality within zone 2608 */ 2609 2610static int __parse_numa_zonelist_order(char *s) 2611{ 2612 if (*s == 'd' || *s == 'D') { 2613 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2614 } else if (*s == 'n' || *s == 'N') { 2615 user_zonelist_order = ZONELIST_ORDER_NODE; 2616 } else if (*s == 'z' || *s == 'Z') { 2617 user_zonelist_order = ZONELIST_ORDER_ZONE; 2618 } else { 2619 printk(KERN_WARNING 2620 "Ignoring invalid numa_zonelist_order value: " 2621 "%s\n", s); 2622 return -EINVAL; 2623 } 2624 return 0; 2625} 2626 2627static __init int setup_numa_zonelist_order(char *s) 2628{ 2629 int ret; 2630 2631 if (!s) 2632 return 0; 2633 2634 ret = __parse_numa_zonelist_order(s); 2635 if (ret == 0) 2636 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 2637 2638 return ret; 2639} 2640early_param("numa_zonelist_order", setup_numa_zonelist_order); 2641 2642/* 2643 * sysctl handler for numa_zonelist_order 2644 */ 2645int numa_zonelist_order_handler(ctl_table *table, int write, 2646 void __user *buffer, size_t *length, 2647 loff_t *ppos) 2648{ 2649 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2650 int ret; 2651 static DEFINE_MUTEX(zl_order_mutex); 2652 2653 mutex_lock(&zl_order_mutex); 2654 if (write) 2655 strcpy(saved_string, (char*)table->data); 2656 ret = proc_dostring(table, write, buffer, length, ppos); 2657 if (ret) 2658 goto out; 2659 if (write) { 2660 int oldval = user_zonelist_order; 2661 if (__parse_numa_zonelist_order((char*)table->data)) { 2662 /* 2663 * bogus value. restore saved string 2664 */ 2665 strncpy((char*)table->data, saved_string, 2666 NUMA_ZONELIST_ORDER_LEN); 2667 user_zonelist_order = oldval; 2668 } else if (oldval != user_zonelist_order) { 2669 mutex_lock(&zonelists_mutex); 2670 build_all_zonelists(NULL); 2671 mutex_unlock(&zonelists_mutex); 2672 } 2673 } 2674out: 2675 mutex_unlock(&zl_order_mutex); 2676 return ret; 2677} 2678 2679 2680#define MAX_NODE_LOAD (nr_online_nodes) 2681static int node_load[MAX_NUMNODES]; 2682 2683/** 2684 * find_next_best_node - find the next node that should appear in a given node's fallback list 2685 * @node: node whose fallback list we're appending 2686 * @used_node_mask: nodemask_t of already used nodes 2687 * 2688 * We use a number of factors to determine which is the next node that should 2689 * appear on a given node's fallback list. The node should not have appeared 2690 * already in @node's fallback list, and it should be the next closest node 2691 * according to the distance array (which contains arbitrary distance values 2692 * from each node to each node in the system), and should also prefer nodes 2693 * with no CPUs, since presumably they'll have very little allocation pressure 2694 * on them otherwise. 2695 * It returns -1 if no node is found. 2696 */ 2697static int find_next_best_node(int node, nodemask_t *used_node_mask) 2698{ 2699 int n, val; 2700 int min_val = INT_MAX; 2701 int best_node = -1; 2702 const struct cpumask *tmp = cpumask_of_node(0); 2703 2704 /* Use the local node if we haven't already */ 2705 if (!node_isset(node, *used_node_mask)) { 2706 node_set(node, *used_node_mask); 2707 return node; 2708 } 2709 2710 for_each_node_state(n, N_HIGH_MEMORY) { 2711 2712 /* Don't want a node to appear more than once */ 2713 if (node_isset(n, *used_node_mask)) 2714 continue; 2715 2716 /* Use the distance array to find the distance */ 2717 val = node_distance(node, n); 2718 2719 /* Penalize nodes under us ("prefer the next node") */ 2720 val += (n < node); 2721 2722 /* Give preference to headless and unused nodes */ 2723 tmp = cpumask_of_node(n); 2724 if (!cpumask_empty(tmp)) 2725 val += PENALTY_FOR_NODE_WITH_CPUS; 2726 2727 /* Slight preference for less loaded node */ 2728 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2729 val += node_load[n]; 2730 2731 if (val < min_val) { 2732 min_val = val; 2733 best_node = n; 2734 } 2735 } 2736 2737 if (best_node >= 0) 2738 node_set(best_node, *used_node_mask); 2739 2740 return best_node; 2741} 2742 2743 2744/* 2745 * Build zonelists ordered by node and zones within node. 2746 * This results in maximum locality--normal zone overflows into local 2747 * DMA zone, if any--but risks exhausting DMA zone. 2748 */ 2749static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2750{ 2751 int j; 2752 struct zonelist *zonelist; 2753 2754 zonelist = &pgdat->node_zonelists[0]; 2755 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2756 ; 2757 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2758 MAX_NR_ZONES - 1); 2759 zonelist->_zonerefs[j].zone = NULL; 2760 zonelist->_zonerefs[j].zone_idx = 0; 2761} 2762 2763/* 2764 * Build gfp_thisnode zonelists 2765 */ 2766static void build_thisnode_zonelists(pg_data_t *pgdat) 2767{ 2768 int j; 2769 struct zonelist *zonelist; 2770 2771 zonelist = &pgdat->node_zonelists[1]; 2772 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2773 zonelist->_zonerefs[j].zone = NULL; 2774 zonelist->_zonerefs[j].zone_idx = 0; 2775} 2776 2777/* 2778 * Build zonelists ordered by zone and nodes within zones. 2779 * This results in conserving DMA zone[s] until all Normal memory is 2780 * exhausted, but results in overflowing to remote node while memory 2781 * may still exist in local DMA zone. 2782 */ 2783static int node_order[MAX_NUMNODES]; 2784 2785static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2786{ 2787 int pos, j, node; 2788 int zone_type; /* needs to be signed */ 2789 struct zone *z; 2790 struct zonelist *zonelist; 2791 2792 zonelist = &pgdat->node_zonelists[0]; 2793 pos = 0; 2794 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2795 for (j = 0; j < nr_nodes; j++) { 2796 node = node_order[j]; 2797 z = &NODE_DATA(node)->node_zones[zone_type]; 2798 if (populated_zone(z)) { 2799 zoneref_set_zone(z, 2800 &zonelist->_zonerefs[pos++]); 2801 check_highest_zone(zone_type); 2802 } 2803 } 2804 } 2805 zonelist->_zonerefs[pos].zone = NULL; 2806 zonelist->_zonerefs[pos].zone_idx = 0; 2807} 2808 2809static int default_zonelist_order(void) 2810{ 2811 int nid, zone_type; 2812 unsigned long low_kmem_size,total_size; 2813 struct zone *z; 2814 int average_size; 2815 /* 2816 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 2817 * If they are really small and used heavily, the system can fall 2818 * into OOM very easily. 2819 * This function detect ZONE_DMA/DMA32 size and configures zone order. 2820 */ 2821 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2822 low_kmem_size = 0; 2823 total_size = 0; 2824 for_each_online_node(nid) { 2825 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2826 z = &NODE_DATA(nid)->node_zones[zone_type]; 2827 if (populated_zone(z)) { 2828 if (zone_type < ZONE_NORMAL) 2829 low_kmem_size += z->present_pages; 2830 total_size += z->present_pages; 2831 } else if (zone_type == ZONE_NORMAL) { 2832 /* 2833 * If any node has only lowmem, then node order 2834 * is preferred to allow kernel allocations 2835 * locally; otherwise, they can easily infringe 2836 * on other nodes when there is an abundance of 2837 * lowmem available to allocate from. 2838 */ 2839 return ZONELIST_ORDER_NODE; 2840 } 2841 } 2842 } 2843 if (!low_kmem_size || /* there are no DMA area. */ 2844 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2845 return ZONELIST_ORDER_NODE; 2846 /* 2847 * look into each node's config. 2848 * If there is a node whose DMA/DMA32 memory is very big area on 2849 * local memory, NODE_ORDER may be suitable. 2850 */ 2851 average_size = total_size / 2852 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2853 for_each_online_node(nid) { 2854 low_kmem_size = 0; 2855 total_size = 0; 2856 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2857 z = &NODE_DATA(nid)->node_zones[zone_type]; 2858 if (populated_zone(z)) { 2859 if (zone_type < ZONE_NORMAL) 2860 low_kmem_size += z->present_pages; 2861 total_size += z->present_pages; 2862 } 2863 } 2864 if (low_kmem_size && 2865 total_size > average_size && /* ignore small node */ 2866 low_kmem_size > total_size * 70/100) 2867 return ZONELIST_ORDER_NODE; 2868 } 2869 return ZONELIST_ORDER_ZONE; 2870} 2871 2872static void set_zonelist_order(void) 2873{ 2874 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2875 current_zonelist_order = default_zonelist_order(); 2876 else 2877 current_zonelist_order = user_zonelist_order; 2878} 2879 2880static void build_zonelists(pg_data_t *pgdat) 2881{ 2882 int j, node, load; 2883 enum zone_type i; 2884 nodemask_t used_mask; 2885 int local_node, prev_node; 2886 struct zonelist *zonelist; 2887 int order = current_zonelist_order; 2888 2889 /* initialize zonelists */ 2890 for (i = 0; i < MAX_ZONELISTS; i++) { 2891 zonelist = pgdat->node_zonelists + i; 2892 zonelist->_zonerefs[0].zone = NULL; 2893 zonelist->_zonerefs[0].zone_idx = 0; 2894 } 2895 2896 /* NUMA-aware ordering of nodes */ 2897 local_node = pgdat->node_id; 2898 load = nr_online_nodes; 2899 prev_node = local_node; 2900 nodes_clear(used_mask); 2901 2902 memset(node_order, 0, sizeof(node_order)); 2903 j = 0; 2904 2905 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2906 int distance = node_distance(local_node, node); 2907 2908 /* 2909 * If another node is sufficiently far away then it is better 2910 * to reclaim pages in a zone before going off node. 2911 */ 2912 if (distance > RECLAIM_DISTANCE) 2913 zone_reclaim_mode = 1; 2914 2915 /* 2916 * We don't want to pressure a particular node. 2917 * So adding penalty to the first node in same 2918 * distance group to make it round-robin. 2919 */ 2920 if (distance != node_distance(local_node, prev_node)) 2921 node_load[node] = load; 2922 2923 prev_node = node; 2924 load--; 2925 if (order == ZONELIST_ORDER_NODE) 2926 build_zonelists_in_node_order(pgdat, node); 2927 else 2928 node_order[j++] = node; /* remember order */ 2929 } 2930 2931 if (order == ZONELIST_ORDER_ZONE) { 2932 /* calculate node order -- i.e., DMA last! */ 2933 build_zonelists_in_zone_order(pgdat, j); 2934 } 2935 2936 build_thisnode_zonelists(pgdat); 2937} 2938 2939/* Construct the zonelist performance cache - see further mmzone.h */ 2940static void build_zonelist_cache(pg_data_t *pgdat) 2941{ 2942 struct zonelist *zonelist; 2943 struct zonelist_cache *zlc; 2944 struct zoneref *z; 2945 2946 zonelist = &pgdat->node_zonelists[0]; 2947 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2948 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2949 for (z = zonelist->_zonerefs; z->zone; z++) 2950 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2951} 2952 2953#ifdef CONFIG_HAVE_MEMORYLESS_NODES 2954/* 2955 * Return node id of node used for "local" allocations. 2956 * I.e., first node id of first zone in arg node's generic zonelist. 2957 * Used for initializing percpu 'numa_mem', which is used primarily 2958 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 2959 */ 2960int local_memory_node(int node) 2961{ 2962 struct zone *zone; 2963 2964 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 2965 gfp_zone(GFP_KERNEL), 2966 NULL, 2967 &zone); 2968 return zone->node; 2969} 2970#endif 2971 2972#else /* CONFIG_NUMA */ 2973 2974static void set_zonelist_order(void) 2975{ 2976 current_zonelist_order = ZONELIST_ORDER_ZONE; 2977} 2978 2979static void build_zonelists(pg_data_t *pgdat) 2980{ 2981 int node, local_node; 2982 enum zone_type j; 2983 struct zonelist *zonelist; 2984 2985 local_node = pgdat->node_id; 2986 2987 zonelist = &pgdat->node_zonelists[0]; 2988 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2989 2990 /* 2991 * Now we build the zonelist so that it contains the zones 2992 * of all the other nodes. 2993 * We don't want to pressure a particular node, so when 2994 * building the zones for node N, we make sure that the 2995 * zones coming right after the local ones are those from 2996 * node N+1 (modulo N) 2997 */ 2998 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2999 if (!node_online(node)) 3000 continue; 3001 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3002 MAX_NR_ZONES - 1); 3003 } 3004 for (node = 0; node < local_node; node++) { 3005 if (!node_online(node)) 3006 continue; 3007 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3008 MAX_NR_ZONES - 1); 3009 } 3010 3011 zonelist->_zonerefs[j].zone = NULL; 3012 zonelist->_zonerefs[j].zone_idx = 0; 3013} 3014 3015/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3016static void build_zonelist_cache(pg_data_t *pgdat) 3017{ 3018 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3019} 3020 3021#endif /* CONFIG_NUMA */ 3022 3023/* 3024 * Boot pageset table. One per cpu which is going to be used for all 3025 * zones and all nodes. The parameters will be set in such a way 3026 * that an item put on a list will immediately be handed over to 3027 * the buddy list. This is safe since pageset manipulation is done 3028 * with interrupts disabled. 3029 * 3030 * The boot_pagesets must be kept even after bootup is complete for 3031 * unused processors and/or zones. They do play a role for bootstrapping 3032 * hotplugged processors. 3033 * 3034 * zoneinfo_show() and maybe other functions do 3035 * not check if the processor is online before following the pageset pointer. 3036 * Other parts of the kernel may not check if the zone is available. 3037 */ 3038static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3039static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3040static void setup_zone_pageset(struct zone *zone); 3041 3042/* 3043 * Global mutex to protect against size modification of zonelists 3044 * as well as to serialize pageset setup for the new populated zone. 3045 */ 3046DEFINE_MUTEX(zonelists_mutex); 3047 3048/* return values int ....just for stop_machine() */ 3049static __init_refok int __build_all_zonelists(void *data) 3050{ 3051 int nid; 3052 int cpu; 3053 3054#ifdef CONFIG_NUMA 3055 memset(node_load, 0, sizeof(node_load)); 3056#endif 3057 for_each_online_node(nid) { 3058 pg_data_t *pgdat = NODE_DATA(nid); 3059 3060 build_zonelists(pgdat); 3061 build_zonelist_cache(pgdat); 3062 } 3063 3064 /* 3065 * Initialize the boot_pagesets that are going to be used 3066 * for bootstrapping processors. The real pagesets for 3067 * each zone will be allocated later when the per cpu 3068 * allocator is available. 3069 * 3070 * boot_pagesets are used also for bootstrapping offline 3071 * cpus if the system is already booted because the pagesets 3072 * are needed to initialize allocators on a specific cpu too. 3073 * F.e. the percpu allocator needs the page allocator which 3074 * needs the percpu allocator in order to allocate its pagesets 3075 * (a chicken-egg dilemma). 3076 */ 3077 for_each_possible_cpu(cpu) { 3078 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3079 3080#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3081 /* 3082 * We now know the "local memory node" for each node-- 3083 * i.e., the node of the first zone in the generic zonelist. 3084 * Set up numa_mem percpu variable for on-line cpus. During 3085 * boot, only the boot cpu should be on-line; we'll init the 3086 * secondary cpus' numa_mem as they come on-line. During 3087 * node/memory hotplug, we'll fixup all on-line cpus. 3088 */ 3089 if (cpu_online(cpu)) 3090 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3091#endif 3092 } 3093 3094 return 0; 3095} 3096 3097/* 3098 * Called with zonelists_mutex held always 3099 * unless system_state == SYSTEM_BOOTING. 3100 */ 3101void build_all_zonelists(void *data) 3102{ 3103 set_zonelist_order(); 3104 3105 if (system_state == SYSTEM_BOOTING) { 3106 __build_all_zonelists(NULL); 3107 mminit_verify_zonelist(); 3108 cpuset_init_current_mems_allowed(); 3109 } else { 3110 /* we have to stop all cpus to guarantee there is no user 3111 of zonelist */ 3112#ifdef CONFIG_MEMORY_HOTPLUG 3113 if (data) 3114 setup_zone_pageset((struct zone *)data); 3115#endif 3116 stop_machine(__build_all_zonelists, NULL, NULL); 3117 /* cpuset refresh routine should be here */ 3118 } 3119 vm_total_pages = nr_free_pagecache_pages(); 3120 /* 3121 * Disable grouping by mobility if the number of pages in the 3122 * system is too low to allow the mechanism to work. It would be 3123 * more accurate, but expensive to check per-zone. This check is 3124 * made on memory-hotadd so a system can start with mobility 3125 * disabled and enable it later 3126 */ 3127 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3128 page_group_by_mobility_disabled = 1; 3129 else 3130 page_group_by_mobility_disabled = 0; 3131 3132 printk("Built %i zonelists in %s order, mobility grouping %s. " 3133 "Total pages: %ld\n", 3134 nr_online_nodes, 3135 zonelist_order_name[current_zonelist_order], 3136 page_group_by_mobility_disabled ? "off" : "on", 3137 vm_total_pages); 3138#ifdef CONFIG_NUMA 3139 printk("Policy zone: %s\n", zone_names[policy_zone]); 3140#endif 3141} 3142 3143/* 3144 * Helper functions to size the waitqueue hash table. 3145 * Essentially these want to choose hash table sizes sufficiently 3146 * large so that collisions trying to wait on pages are rare. 3147 * But in fact, the number of active page waitqueues on typical 3148 * systems is ridiculously low, less than 200. So this is even 3149 * conservative, even though it seems large. 3150 * 3151 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3152 * waitqueues, i.e. the size of the waitq table given the number of pages. 3153 */ 3154#define PAGES_PER_WAITQUEUE 256 3155 3156#ifndef CONFIG_MEMORY_HOTPLUG 3157static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3158{ 3159 unsigned long size = 1; 3160 3161 pages /= PAGES_PER_WAITQUEUE; 3162 3163 while (size < pages) 3164 size <<= 1; 3165 3166 /* 3167 * Once we have dozens or even hundreds of threads sleeping 3168 * on IO we've got bigger problems than wait queue collision. 3169 * Limit the size of the wait table to a reasonable size. 3170 */ 3171 size = min(size, 4096UL); 3172 3173 return max(size, 4UL); 3174} 3175#else 3176/* 3177 * A zone's size might be changed by hot-add, so it is not possible to determine 3178 * a suitable size for its wait_table. So we use the maximum size now. 3179 * 3180 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3181 * 3182 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3183 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3184 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3185 * 3186 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3187 * or more by the traditional way. (See above). It equals: 3188 * 3189 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3190 * ia64(16K page size) : = ( 8G + 4M)byte. 3191 * powerpc (64K page size) : = (32G +16M)byte. 3192 */ 3193static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3194{ 3195 return 4096UL; 3196} 3197#endif 3198 3199/* 3200 * This is an integer logarithm so that shifts can be used later 3201 * to extract the more random high bits from the multiplicative 3202 * hash function before the remainder is taken. 3203 */ 3204static inline unsigned long wait_table_bits(unsigned long size) 3205{ 3206 return ffz(~size); 3207} 3208 3209#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3210 3211/* 3212 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3213 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3214 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3215 * higher will lead to a bigger reserve which will get freed as contiguous 3216 * blocks as reclaim kicks in 3217 */ 3218static void setup_zone_migrate_reserve(struct zone *zone) 3219{ 3220 unsigned long start_pfn, pfn, end_pfn; 3221 struct page *page; 3222 unsigned long block_migratetype; 3223 int reserve; 3224 3225 /* Get the start pfn, end pfn and the number of blocks to reserve */ 3226 start_pfn = zone->zone_start_pfn; 3227 end_pfn = start_pfn + zone->spanned_pages; 3228 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3229 pageblock_order; 3230 3231 /* 3232 * Reserve blocks are generally in place to help high-order atomic 3233 * allocations that are short-lived. A min_free_kbytes value that 3234 * would result in more than 2 reserve blocks for atomic allocations 3235 * is assumed to be in place to help anti-fragmentation for the 3236 * future allocation of hugepages at runtime. 3237 */ 3238 reserve = min(2, reserve); 3239 3240 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3241 if (!pfn_valid(pfn)) 3242 continue; 3243 page = pfn_to_page(pfn); 3244 3245 /* Watch out for overlapping nodes */ 3246 if (page_to_nid(page) != zone_to_nid(zone)) 3247 continue; 3248 3249 /* Blocks with reserved pages will never free, skip them. */ 3250 if (PageReserved(page)) 3251 continue; 3252 3253 block_migratetype = get_pageblock_migratetype(page); 3254 3255 /* If this block is reserved, account for it */ 3256 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 3257 reserve--; 3258 continue; 3259 } 3260 3261 /* Suitable for reserving if this block is movable */ 3262 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 3263 set_pageblock_migratetype(page, MIGRATE_RESERVE); 3264 move_freepages_block(zone, page, MIGRATE_RESERVE); 3265 reserve--; 3266 continue; 3267 } 3268 3269 /* 3270 * If the reserve is met and this is a previous reserved block, 3271 * take it back 3272 */ 3273 if (block_migratetype == MIGRATE_RESERVE) { 3274 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3275 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3276 } 3277 } 3278} 3279 3280/* 3281 * Initially all pages are reserved - free ones are freed 3282 * up by free_all_bootmem() once the early boot process is 3283 * done. Non-atomic initialization, single-pass. 3284 */ 3285void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3286 unsigned long start_pfn, enum memmap_context context) 3287{ 3288 struct page *page; 3289 unsigned long end_pfn = start_pfn + size; 3290 unsigned long pfn; 3291 struct zone *z; 3292 3293 if (highest_memmap_pfn < end_pfn - 1) 3294 highest_memmap_pfn = end_pfn - 1; 3295 3296 z = &NODE_DATA(nid)->node_zones[zone]; 3297 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3298 /* 3299 * There can be holes in boot-time mem_map[]s 3300 * handed to this function. They do not 3301 * exist on hotplugged memory. 3302 */ 3303 if (context == MEMMAP_EARLY) { 3304 if (!early_pfn_valid(pfn)) 3305 continue; 3306 if (!early_pfn_in_nid(pfn, nid)) 3307 continue; 3308 } 3309 page = pfn_to_page(pfn); 3310 set_page_links(page, zone, nid, pfn); 3311 mminit_verify_page_links(page, zone, nid, pfn); 3312 init_page_count(page); 3313 reset_page_mapcount(page); 3314 SetPageReserved(page); 3315 /* 3316 * Mark the block movable so that blocks are reserved for 3317 * movable at startup. This will force kernel allocations 3318 * to reserve their blocks rather than leaking throughout 3319 * the address space during boot when many long-lived 3320 * kernel allocations are made. Later some blocks near 3321 * the start are marked MIGRATE_RESERVE by 3322 * setup_zone_migrate_reserve() 3323 * 3324 * bitmap is created for zone's valid pfn range. but memmap 3325 * can be created for invalid pages (for alignment) 3326 * check here not to call set_pageblock_migratetype() against 3327 * pfn out of zone. 3328 */ 3329 if ((z->zone_start_pfn <= pfn) 3330 && (pfn < z->zone_start_pfn + z->spanned_pages) 3331 && !(pfn & (pageblock_nr_pages - 1))) 3332 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3333 3334 INIT_LIST_HEAD(&page->lru); 3335#ifdef WANT_PAGE_VIRTUAL 3336 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3337 if (!is_highmem_idx(zone)) 3338 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3339#endif 3340 } 3341} 3342 3343static void __meminit zone_init_free_lists(struct zone *zone) 3344{ 3345 int order, t; 3346 for_each_migratetype_order(order, t) { 3347 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3348 zone->free_area[order].nr_free = 0; 3349 } 3350} 3351 3352#ifndef __HAVE_ARCH_MEMMAP_INIT 3353#define memmap_init(size, nid, zone, start_pfn) \ 3354 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3355#endif 3356 3357static int zone_batchsize(struct zone *zone) 3358{ 3359#ifdef CONFIG_MMU 3360 int batch; 3361 3362 /* 3363 * The per-cpu-pages pools are set to around 1000th of the 3364 * size of the zone. But no more than 1/2 of a meg. 3365 * 3366 * OK, so we don't know how big the cache is. So guess. 3367 */ 3368 batch = zone->present_pages / 1024; 3369 if (batch * PAGE_SIZE > 512 * 1024) 3370 batch = (512 * 1024) / PAGE_SIZE; 3371 batch /= 4; /* We effectively *= 4 below */ 3372 if (batch < 1) 3373 batch = 1; 3374 3375 /* 3376 * Clamp the batch to a 2^n - 1 value. Having a power 3377 * of 2 value was found to be more likely to have 3378 * suboptimal cache aliasing properties in some cases. 3379 * 3380 * For example if 2 tasks are alternately allocating 3381 * batches of pages, one task can end up with a lot 3382 * of pages of one half of the possible page colors 3383 * and the other with pages of the other colors. 3384 */ 3385 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3386 3387 return batch; 3388 3389#else 3390 /* The deferral and batching of frees should be suppressed under NOMMU 3391 * conditions. 3392 * 3393 * The problem is that NOMMU needs to be able to allocate large chunks 3394 * of contiguous memory as there's no hardware page translation to 3395 * assemble apparent contiguous memory from discontiguous pages. 3396 * 3397 * Queueing large contiguous runs of pages for batching, however, 3398 * causes the pages to actually be freed in smaller chunks. As there 3399 * can be a significant delay between the individual batches being 3400 * recycled, this leads to the once large chunks of space being 3401 * fragmented and becoming unavailable for high-order allocations. 3402 */ 3403 return 0; 3404#endif 3405} 3406 3407static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3408{ 3409 struct per_cpu_pages *pcp; 3410 int migratetype; 3411 3412 memset(p, 0, sizeof(*p)); 3413 3414 pcp = &p->pcp; 3415 pcp->count = 0; 3416 pcp->high = 6 * batch; 3417 pcp->batch = max(1UL, 1 * batch); 3418 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3419 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3420} 3421 3422/* 3423 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3424 * to the value high for the pageset p. 3425 */ 3426 3427static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3428 unsigned long high) 3429{ 3430 struct per_cpu_pages *pcp; 3431 3432 pcp = &p->pcp; 3433 pcp->high = high; 3434 pcp->batch = max(1UL, high/4); 3435 if ((high/4) > (PAGE_SHIFT * 8)) 3436 pcp->batch = PAGE_SHIFT * 8; 3437} 3438 3439static __meminit void setup_zone_pageset(struct zone *zone) 3440{ 3441 int cpu; 3442 3443 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3444 3445 for_each_possible_cpu(cpu) { 3446 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3447 3448 setup_pageset(pcp, zone_batchsize(zone)); 3449 3450 if (percpu_pagelist_fraction) 3451 setup_pagelist_highmark(pcp, 3452 (zone->present_pages / 3453 percpu_pagelist_fraction)); 3454 } 3455} 3456 3457/* 3458 * Allocate per cpu pagesets and initialize them. 3459 * Before this call only boot pagesets were available. 3460 */ 3461void __init setup_per_cpu_pageset(void) 3462{ 3463 struct zone *zone; 3464 3465 for_each_populated_zone(zone) 3466 setup_zone_pageset(zone); 3467} 3468 3469static noinline __init_refok 3470int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3471{ 3472 int i; 3473 struct pglist_data *pgdat = zone->zone_pgdat; 3474 size_t alloc_size; 3475 3476 /* 3477 * The per-page waitqueue mechanism uses hashed waitqueues 3478 * per zone. 3479 */ 3480 zone->wait_table_hash_nr_entries = 3481 wait_table_hash_nr_entries(zone_size_pages); 3482 zone->wait_table_bits = 3483 wait_table_bits(zone->wait_table_hash_nr_entries); 3484 alloc_size = zone->wait_table_hash_nr_entries 3485 * sizeof(wait_queue_head_t); 3486 3487 if (!slab_is_available()) { 3488 zone->wait_table = (wait_queue_head_t *) 3489 alloc_bootmem_node(pgdat, alloc_size); 3490 } else { 3491 /* 3492 * This case means that a zone whose size was 0 gets new memory 3493 * via memory hot-add. 3494 * But it may be the case that a new node was hot-added. In 3495 * this case vmalloc() will not be able to use this new node's 3496 * memory - this wait_table must be initialized to use this new 3497 * node itself as well. 3498 * To use this new node's memory, further consideration will be 3499 * necessary. 3500 */ 3501 zone->wait_table = vmalloc(alloc_size); 3502 } 3503 if (!zone->wait_table) 3504 return -ENOMEM; 3505 3506 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3507 init_waitqueue_head(zone->wait_table + i); 3508 3509 return 0; 3510} 3511 3512static int __zone_pcp_update(void *data) 3513{ 3514 struct zone *zone = data; 3515 int cpu; 3516 unsigned long batch = zone_batchsize(zone), flags; 3517 3518 for_each_possible_cpu(cpu) { 3519 struct per_cpu_pageset *pset; 3520 struct per_cpu_pages *pcp; 3521 3522 pset = per_cpu_ptr(zone->pageset, cpu); 3523 pcp = &pset->pcp; 3524 3525 local_irq_save(flags); 3526 free_pcppages_bulk(zone, pcp->count, pcp); 3527 setup_pageset(pset, batch); 3528 local_irq_restore(flags); 3529 } 3530 return 0; 3531} 3532 3533void zone_pcp_update(struct zone *zone) 3534{ 3535 stop_machine(__zone_pcp_update, zone, NULL); 3536} 3537 3538static __meminit void zone_pcp_init(struct zone *zone) 3539{ 3540 /* 3541 * per cpu subsystem is not up at this point. The following code 3542 * relies on the ability of the linker to provide the 3543 * offset of a (static) per cpu variable into the per cpu area. 3544 */ 3545 zone->pageset = &boot_pageset; 3546 3547 if (zone->present_pages) 3548 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3549 zone->name, zone->present_pages, 3550 zone_batchsize(zone)); 3551} 3552 3553__meminit int init_currently_empty_zone(struct zone *zone, 3554 unsigned long zone_start_pfn, 3555 unsigned long size, 3556 enum memmap_context context) 3557{ 3558 struct pglist_data *pgdat = zone->zone_pgdat; 3559 int ret; 3560 ret = zone_wait_table_init(zone, size); 3561 if (ret) 3562 return ret; 3563 pgdat->nr_zones = zone_idx(zone) + 1; 3564 3565 zone->zone_start_pfn = zone_start_pfn; 3566 3567 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3568 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3569 pgdat->node_id, 3570 (unsigned long)zone_idx(zone), 3571 zone_start_pfn, (zone_start_pfn + size)); 3572 3573 zone_init_free_lists(zone); 3574 3575 return 0; 3576} 3577 3578#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3579/* 3580 * Basic iterator support. Return the first range of PFNs for a node 3581 * Note: nid == MAX_NUMNODES returns first region regardless of node 3582 */ 3583static int __meminit first_active_region_index_in_nid(int nid) 3584{ 3585 int i; 3586 3587 for (i = 0; i < nr_nodemap_entries; i++) 3588 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3589 return i; 3590 3591 return -1; 3592} 3593 3594/* 3595 * Basic iterator support. Return the next active range of PFNs for a node 3596 * Note: nid == MAX_NUMNODES returns next region regardless of node 3597 */ 3598static int __meminit next_active_region_index_in_nid(int index, int nid) 3599{ 3600 for (index = index + 1; index < nr_nodemap_entries; index++) 3601 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3602 return index; 3603 3604 return -1; 3605} 3606 3607#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3608/* 3609 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3610 * Architectures may implement their own version but if add_active_range() 3611 * was used and there are no special requirements, this is a convenient 3612 * alternative 3613 */ 3614int __meminit __early_pfn_to_nid(unsigned long pfn) 3615{ 3616 int i; 3617 3618 for (i = 0; i < nr_nodemap_entries; i++) { 3619 unsigned long start_pfn = early_node_map[i].start_pfn; 3620 unsigned long end_pfn = early_node_map[i].end_pfn; 3621 3622 if (start_pfn <= pfn && pfn < end_pfn) 3623 return early_node_map[i].nid; 3624 } 3625 /* This is a memory hole */ 3626 return -1; 3627} 3628#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3629 3630int __meminit early_pfn_to_nid(unsigned long pfn) 3631{ 3632 int nid; 3633 3634 nid = __early_pfn_to_nid(pfn); 3635 if (nid >= 0) 3636 return nid; 3637 /* just returns 0 */ 3638 return 0; 3639} 3640 3641#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3642bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3643{ 3644 int nid; 3645 3646 nid = __early_pfn_to_nid(pfn); 3647 if (nid >= 0 && nid != node) 3648 return false; 3649 return true; 3650} 3651#endif 3652 3653/* Basic iterator support to walk early_node_map[] */ 3654#define for_each_active_range_index_in_nid(i, nid) \ 3655 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3656 i = next_active_region_index_in_nid(i, nid)) 3657 3658/** 3659 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3660 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3661 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3662 * 3663 * If an architecture guarantees that all ranges registered with 3664 * add_active_ranges() contain no holes and may be freed, this 3665 * this function may be used instead of calling free_bootmem() manually. 3666 */ 3667void __init free_bootmem_with_active_regions(int nid, 3668 unsigned long max_low_pfn) 3669{ 3670 int i; 3671 3672 for_each_active_range_index_in_nid(i, nid) { 3673 unsigned long size_pages = 0; 3674 unsigned long end_pfn = early_node_map[i].end_pfn; 3675 3676 if (early_node_map[i].start_pfn >= max_low_pfn) 3677 continue; 3678 3679 if (end_pfn > max_low_pfn) 3680 end_pfn = max_low_pfn; 3681 3682 size_pages = end_pfn - early_node_map[i].start_pfn; 3683 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3684 PFN_PHYS(early_node_map[i].start_pfn), 3685 size_pages << PAGE_SHIFT); 3686 } 3687} 3688 3689#ifdef CONFIG_HAVE_MEMBLOCK 3690u64 __init find_memory_core_early(int nid, u64 size, u64 align, 3691 u64 goal, u64 limit) 3692{ 3693 int i; 3694 3695 /* Need to go over early_node_map to find out good range for node */ 3696 for_each_active_range_index_in_nid(i, nid) { 3697 u64 addr; 3698 u64 ei_start, ei_last; 3699 u64 final_start, final_end; 3700 3701 ei_last = early_node_map[i].end_pfn; 3702 ei_last <<= PAGE_SHIFT; 3703 ei_start = early_node_map[i].start_pfn; 3704 ei_start <<= PAGE_SHIFT; 3705 3706 final_start = max(ei_start, goal); 3707 final_end = min(ei_last, limit); 3708 3709 if (final_start >= final_end) 3710 continue; 3711 3712 addr = memblock_find_in_range(final_start, final_end, size, align); 3713 3714 if (addr == MEMBLOCK_ERROR) 3715 continue; 3716 3717 return addr; 3718 } 3719 3720 return MEMBLOCK_ERROR; 3721} 3722#endif 3723 3724int __init add_from_early_node_map(struct range *range, int az, 3725 int nr_range, int nid) 3726{ 3727 int i; 3728 u64 start, end; 3729 3730 /* need to go over early_node_map to find out good range for node */ 3731 for_each_active_range_index_in_nid(i, nid) { 3732 start = early_node_map[i].start_pfn; 3733 end = early_node_map[i].end_pfn; 3734 nr_range = add_range(range, az, nr_range, start, end); 3735 } 3736 return nr_range; 3737} 3738 3739#ifdef CONFIG_NO_BOOTMEM 3740void * __init __alloc_memory_core_early(int nid, u64 size, u64 align, 3741 u64 goal, u64 limit) 3742{ 3743 void *ptr; 3744 u64 addr; 3745 3746 if (limit > memblock.current_limit) 3747 limit = memblock.current_limit; 3748 3749 addr = find_memory_core_early(nid, size, align, goal, limit); 3750 3751 if (addr == MEMBLOCK_ERROR) 3752 return NULL; 3753 3754 ptr = phys_to_virt(addr); 3755 memset(ptr, 0, size); 3756 memblock_x86_reserve_range(addr, addr + size, "BOOTMEM"); 3757 /* 3758 * The min_count is set to 0 so that bootmem allocated blocks 3759 * are never reported as leaks. 3760 */ 3761 kmemleak_alloc(ptr, size, 0, 0); 3762 return ptr; 3763} 3764#endif 3765 3766 3767void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3768{ 3769 int i; 3770 int ret; 3771 3772 for_each_active_range_index_in_nid(i, nid) { 3773 ret = work_fn(early_node_map[i].start_pfn, 3774 early_node_map[i].end_pfn, data); 3775 if (ret) 3776 break; 3777 } 3778} 3779/** 3780 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3781 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3782 * 3783 * If an architecture guarantees that all ranges registered with 3784 * add_active_ranges() contain no holes and may be freed, this 3785 * function may be used instead of calling memory_present() manually. 3786 */ 3787void __init sparse_memory_present_with_active_regions(int nid) 3788{ 3789 int i; 3790 3791 for_each_active_range_index_in_nid(i, nid) 3792 memory_present(early_node_map[i].nid, 3793 early_node_map[i].start_pfn, 3794 early_node_map[i].end_pfn); 3795} 3796 3797/** 3798 * get_pfn_range_for_nid - Return the start and end page frames for a node 3799 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3800 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3801 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3802 * 3803 * It returns the start and end page frame of a node based on information 3804 * provided by an arch calling add_active_range(). If called for a node 3805 * with no available memory, a warning is printed and the start and end 3806 * PFNs will be 0. 3807 */ 3808void __meminit get_pfn_range_for_nid(unsigned int nid, 3809 unsigned long *start_pfn, unsigned long *end_pfn) 3810{ 3811 int i; 3812 *start_pfn = -1UL; 3813 *end_pfn = 0; 3814 3815 for_each_active_range_index_in_nid(i, nid) { 3816 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3817 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3818 } 3819 3820 if (*start_pfn == -1UL) 3821 *start_pfn = 0; 3822} 3823 3824/* 3825 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3826 * assumption is made that zones within a node are ordered in monotonic 3827 * increasing memory addresses so that the "highest" populated zone is used 3828 */ 3829static void __init find_usable_zone_for_movable(void) 3830{ 3831 int zone_index; 3832 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3833 if (zone_index == ZONE_MOVABLE) 3834 continue; 3835 3836 if (arch_zone_highest_possible_pfn[zone_index] > 3837 arch_zone_lowest_possible_pfn[zone_index]) 3838 break; 3839 } 3840 3841 VM_BUG_ON(zone_index == -1); 3842 movable_zone = zone_index; 3843} 3844 3845/* 3846 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3847 * because it is sized independant of architecture. Unlike the other zones, 3848 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3849 * in each node depending on the size of each node and how evenly kernelcore 3850 * is distributed. This helper function adjusts the zone ranges 3851 * provided by the architecture for a given node by using the end of the 3852 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3853 * zones within a node are in order of monotonic increases memory addresses 3854 */ 3855static void __meminit adjust_zone_range_for_zone_movable(int nid, 3856 unsigned long zone_type, 3857 unsigned long node_start_pfn, 3858 unsigned long node_end_pfn, 3859 unsigned long *zone_start_pfn, 3860 unsigned long *zone_end_pfn) 3861{ 3862 /* Only adjust if ZONE_MOVABLE is on this node */ 3863 if (zone_movable_pfn[nid]) { 3864 /* Size ZONE_MOVABLE */ 3865 if (zone_type == ZONE_MOVABLE) { 3866 *zone_start_pfn = zone_movable_pfn[nid]; 3867 *zone_end_pfn = min(node_end_pfn, 3868 arch_zone_highest_possible_pfn[movable_zone]); 3869 3870 /* Adjust for ZONE_MOVABLE starting within this range */ 3871 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3872 *zone_end_pfn > zone_movable_pfn[nid]) { 3873 *zone_end_pfn = zone_movable_pfn[nid]; 3874 3875 /* Check if this whole range is within ZONE_MOVABLE */ 3876 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3877 *zone_start_pfn = *zone_end_pfn; 3878 } 3879} 3880 3881/* 3882 * Return the number of pages a zone spans in a node, including holes 3883 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3884 */ 3885static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3886 unsigned long zone_type, 3887 unsigned long *ignored) 3888{ 3889 unsigned long node_start_pfn, node_end_pfn; 3890 unsigned long zone_start_pfn, zone_end_pfn; 3891 3892 /* Get the start and end of the node and zone */ 3893 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3894 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3895 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3896 adjust_zone_range_for_zone_movable(nid, zone_type, 3897 node_start_pfn, node_end_pfn, 3898 &zone_start_pfn, &zone_end_pfn); 3899 3900 /* Check that this node has pages within the zone's required range */ 3901 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3902 return 0; 3903 3904 /* Move the zone boundaries inside the node if necessary */ 3905 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3906 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3907 3908 /* Return the spanned pages */ 3909 return zone_end_pfn - zone_start_pfn; 3910} 3911 3912/* 3913 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3914 * then all holes in the requested range will be accounted for. 3915 */ 3916unsigned long __meminit __absent_pages_in_range(int nid, 3917 unsigned long range_start_pfn, 3918 unsigned long range_end_pfn) 3919{ 3920 int i = 0; 3921 unsigned long prev_end_pfn = 0, hole_pages = 0; 3922 unsigned long start_pfn; 3923 3924 /* Find the end_pfn of the first active range of pfns in the node */ 3925 i = first_active_region_index_in_nid(nid); 3926 if (i == -1) 3927 return 0; 3928 3929 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3930 3931 /* Account for ranges before physical memory on this node */ 3932 if (early_node_map[i].start_pfn > range_start_pfn) 3933 hole_pages = prev_end_pfn - range_start_pfn; 3934 3935 /* Find all holes for the zone within the node */ 3936 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3937 3938 /* No need to continue if prev_end_pfn is outside the zone */ 3939 if (prev_end_pfn >= range_end_pfn) 3940 break; 3941 3942 /* Make sure the end of the zone is not within the hole */ 3943 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3944 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3945 3946 /* Update the hole size cound and move on */ 3947 if (start_pfn > range_start_pfn) { 3948 BUG_ON(prev_end_pfn > start_pfn); 3949 hole_pages += start_pfn - prev_end_pfn; 3950 } 3951 prev_end_pfn = early_node_map[i].end_pfn; 3952 } 3953 3954 /* Account for ranges past physical memory on this node */ 3955 if (range_end_pfn > prev_end_pfn) 3956 hole_pages += range_end_pfn - 3957 max(range_start_pfn, prev_end_pfn); 3958 3959 return hole_pages; 3960} 3961 3962/** 3963 * absent_pages_in_range - Return number of page frames in holes within a range 3964 * @start_pfn: The start PFN to start searching for holes 3965 * @end_pfn: The end PFN to stop searching for holes 3966 * 3967 * It returns the number of pages frames in memory holes within a range. 3968 */ 3969unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3970 unsigned long end_pfn) 3971{ 3972 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3973} 3974 3975/* Return the number of page frames in holes in a zone on a node */ 3976static unsigned long __meminit zone_absent_pages_in_node(int nid, 3977 unsigned long zone_type, 3978 unsigned long *ignored) 3979{ 3980 unsigned long node_start_pfn, node_end_pfn; 3981 unsigned long zone_start_pfn, zone_end_pfn; 3982 3983 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3984 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3985 node_start_pfn); 3986 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3987 node_end_pfn); 3988 3989 adjust_zone_range_for_zone_movable(nid, zone_type, 3990 node_start_pfn, node_end_pfn, 3991 &zone_start_pfn, &zone_end_pfn); 3992 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3993} 3994 3995#else 3996static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3997 unsigned long zone_type, 3998 unsigned long *zones_size) 3999{ 4000 return zones_size[zone_type]; 4001} 4002 4003static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4004 unsigned long zone_type, 4005 unsigned long *zholes_size) 4006{ 4007 if (!zholes_size) 4008 return 0; 4009 4010 return zholes_size[zone_type]; 4011} 4012 4013#endif 4014 4015static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4016 unsigned long *zones_size, unsigned long *zholes_size) 4017{ 4018 unsigned long realtotalpages, totalpages = 0; 4019 enum zone_type i; 4020 4021 for (i = 0; i < MAX_NR_ZONES; i++) 4022 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4023 zones_size); 4024 pgdat->node_spanned_pages = totalpages; 4025 4026 realtotalpages = totalpages; 4027 for (i = 0; i < MAX_NR_ZONES; i++) 4028 realtotalpages -= 4029 zone_absent_pages_in_node(pgdat->node_id, i, 4030 zholes_size); 4031 pgdat->node_present_pages = realtotalpages; 4032 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4033 realtotalpages); 4034} 4035 4036#ifndef CONFIG_SPARSEMEM 4037/* 4038 * Calculate the size of the zone->blockflags rounded to an unsigned long 4039 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4040 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4041 * round what is now in bits to nearest long in bits, then return it in 4042 * bytes. 4043 */ 4044static unsigned long __init usemap_size(unsigned long zonesize) 4045{ 4046 unsigned long usemapsize; 4047 4048 usemapsize = roundup(zonesize, pageblock_nr_pages); 4049 usemapsize = usemapsize >> pageblock_order; 4050 usemapsize *= NR_PAGEBLOCK_BITS; 4051 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4052 4053 return usemapsize / 8; 4054} 4055 4056static void __init setup_usemap(struct pglist_data *pgdat, 4057 struct zone *zone, unsigned long zonesize) 4058{ 4059 unsigned long usemapsize = usemap_size(zonesize); 4060 zone->pageblock_flags = NULL; 4061 if (usemapsize) 4062 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 4063} 4064#else 4065static inline void setup_usemap(struct pglist_data *pgdat, 4066 struct zone *zone, unsigned long zonesize) {} 4067#endif /* CONFIG_SPARSEMEM */ 4068 4069#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4070 4071/* Return a sensible default order for the pageblock size. */ 4072static inline int pageblock_default_order(void) 4073{ 4074 if (HPAGE_SHIFT > PAGE_SHIFT) 4075 return HUGETLB_PAGE_ORDER; 4076 4077 return MAX_ORDER-1; 4078} 4079 4080/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4081static inline void __init set_pageblock_order(unsigned int order) 4082{ 4083 /* Check that pageblock_nr_pages has not already been setup */ 4084 if (pageblock_order) 4085 return; 4086 4087 /* 4088 * Assume the largest contiguous order of interest is a huge page. 4089 * This value may be variable depending on boot parameters on IA64 4090 */ 4091 pageblock_order = order; 4092} 4093#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4094 4095/* 4096 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4097 * and pageblock_default_order() are unused as pageblock_order is set 4098 * at compile-time. See include/linux/pageblock-flags.h for the values of 4099 * pageblock_order based on the kernel config 4100 */ 4101static inline int pageblock_default_order(unsigned int order) 4102{ 4103 return MAX_ORDER-1; 4104} 4105#define set_pageblock_order(x) do {} while (0) 4106 4107#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4108 4109/* 4110 * Set up the zone data structures: 4111 * - mark all pages reserved 4112 * - mark all memory queues empty 4113 * - clear the memory bitmaps 4114 */ 4115static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4116 unsigned long *zones_size, unsigned long *zholes_size) 4117{ 4118 enum zone_type j; 4119 int nid = pgdat->node_id; 4120 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4121 int ret; 4122 4123 pgdat_resize_init(pgdat); 4124 pgdat->nr_zones = 0; 4125 init_waitqueue_head(&pgdat->kswapd_wait); 4126 pgdat->kswapd_max_order = 0; 4127 pgdat_page_cgroup_init(pgdat); 4128 4129 for (j = 0; j < MAX_NR_ZONES; j++) { 4130 struct zone *zone = pgdat->node_zones + j; 4131 unsigned long size, realsize, memmap_pages; 4132 enum lru_list l; 4133 4134 size = zone_spanned_pages_in_node(nid, j, zones_size); 4135 realsize = size - zone_absent_pages_in_node(nid, j, 4136 zholes_size); 4137 4138 /* 4139 * Adjust realsize so that it accounts for how much memory 4140 * is used by this zone for memmap. This affects the watermark 4141 * and per-cpu initialisations 4142 */ 4143 memmap_pages = 4144 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4145 if (realsize >= memmap_pages) { 4146 realsize -= memmap_pages; 4147 if (memmap_pages) 4148 printk(KERN_DEBUG 4149 " %s zone: %lu pages used for memmap\n", 4150 zone_names[j], memmap_pages); 4151 } else 4152 printk(KERN_WARNING 4153 " %s zone: %lu pages exceeds realsize %lu\n", 4154 zone_names[j], memmap_pages, realsize); 4155 4156 /* Account for reserved pages */ 4157 if (j == 0 && realsize > dma_reserve) { 4158 realsize -= dma_reserve; 4159 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4160 zone_names[0], dma_reserve); 4161 } 4162 4163 if (!is_highmem_idx(j)) 4164 nr_kernel_pages += realsize; 4165 nr_all_pages += realsize; 4166 4167 zone->spanned_pages = size; 4168 zone->present_pages = realsize; 4169#ifdef CONFIG_NUMA 4170 zone->node = nid; 4171 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4172 / 100; 4173 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4174#endif 4175 zone->name = zone_names[j]; 4176 spin_lock_init(&zone->lock); 4177 spin_lock_init(&zone->lru_lock); 4178 zone_seqlock_init(zone); 4179 zone->zone_pgdat = pgdat; 4180 4181 zone_pcp_init(zone); 4182 for_each_lru(l) { 4183 INIT_LIST_HEAD(&zone->lru[l].list); 4184 zone->reclaim_stat.nr_saved_scan[l] = 0; 4185 } 4186 zone->reclaim_stat.recent_rotated[0] = 0; 4187 zone->reclaim_stat.recent_rotated[1] = 0; 4188 zone->reclaim_stat.recent_scanned[0] = 0; 4189 zone->reclaim_stat.recent_scanned[1] = 0; 4190 zap_zone_vm_stats(zone); 4191 zone->flags = 0; 4192 if (!size) 4193 continue; 4194 4195 set_pageblock_order(pageblock_default_order()); 4196 setup_usemap(pgdat, zone, size); 4197 ret = init_currently_empty_zone(zone, zone_start_pfn, 4198 size, MEMMAP_EARLY); 4199 BUG_ON(ret); 4200 memmap_init(size, nid, j, zone_start_pfn); 4201 zone_start_pfn += size; 4202 } 4203} 4204 4205static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4206{ 4207 /* Skip empty nodes */ 4208 if (!pgdat->node_spanned_pages) 4209 return; 4210 4211#ifdef CONFIG_FLAT_NODE_MEM_MAP 4212 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4213 if (!pgdat->node_mem_map) { 4214 unsigned long size, start, end; 4215 struct page *map; 4216 4217 /* 4218 * The zone's endpoints aren't required to be MAX_ORDER 4219 * aligned but the node_mem_map endpoints must be in order 4220 * for the buddy allocator to function correctly. 4221 */ 4222 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4223 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4224 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4225 size = (end - start) * sizeof(struct page); 4226 map = alloc_remap(pgdat->node_id, size); 4227 if (!map) 4228 map = alloc_bootmem_node(pgdat, size); 4229 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4230 } 4231#ifndef CONFIG_NEED_MULTIPLE_NODES 4232 /* 4233 * With no DISCONTIG, the global mem_map is just set as node 0's 4234 */ 4235 if (pgdat == NODE_DATA(0)) { 4236 mem_map = NODE_DATA(0)->node_mem_map; 4237#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4238 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4239 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4240#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4241 } 4242#endif 4243#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4244} 4245 4246void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4247 unsigned long node_start_pfn, unsigned long *zholes_size) 4248{ 4249 pg_data_t *pgdat = NODE_DATA(nid); 4250 4251 pgdat->node_id = nid; 4252 pgdat->node_start_pfn = node_start_pfn; 4253 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4254 4255 alloc_node_mem_map(pgdat); 4256#ifdef CONFIG_FLAT_NODE_MEM_MAP 4257 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4258 nid, (unsigned long)pgdat, 4259 (unsigned long)pgdat->node_mem_map); 4260#endif 4261 4262 free_area_init_core(pgdat, zones_size, zholes_size); 4263} 4264 4265#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4266 4267#if MAX_NUMNODES > 1 4268/* 4269 * Figure out the number of possible node ids. 4270 */ 4271static void __init setup_nr_node_ids(void) 4272{ 4273 unsigned int node; 4274 unsigned int highest = 0; 4275 4276 for_each_node_mask(node, node_possible_map) 4277 highest = node; 4278 nr_node_ids = highest + 1; 4279} 4280#else 4281static inline void setup_nr_node_ids(void) 4282{ 4283} 4284#endif 4285 4286/** 4287 * add_active_range - Register a range of PFNs backed by physical memory 4288 * @nid: The node ID the range resides on 4289 * @start_pfn: The start PFN of the available physical memory 4290 * @end_pfn: The end PFN of the available physical memory 4291 * 4292 * These ranges are stored in an early_node_map[] and later used by 4293 * free_area_init_nodes() to calculate zone sizes and holes. If the 4294 * range spans a memory hole, it is up to the architecture to ensure 4295 * the memory is not freed by the bootmem allocator. If possible 4296 * the range being registered will be merged with existing ranges. 4297 */ 4298void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4299 unsigned long end_pfn) 4300{ 4301 int i; 4302 4303 mminit_dprintk(MMINIT_TRACE, "memory_register", 4304 "Entering add_active_range(%d, %#lx, %#lx) " 4305 "%d entries of %d used\n", 4306 nid, start_pfn, end_pfn, 4307 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4308 4309 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4310 4311 /* Merge with existing active regions if possible */ 4312 for (i = 0; i < nr_nodemap_entries; i++) { 4313 if (early_node_map[i].nid != nid) 4314 continue; 4315 4316 /* Skip if an existing region covers this new one */ 4317 if (start_pfn >= early_node_map[i].start_pfn && 4318 end_pfn <= early_node_map[i].end_pfn) 4319 return; 4320 4321 /* Merge forward if suitable */ 4322 if (start_pfn <= early_node_map[i].end_pfn && 4323 end_pfn > early_node_map[i].end_pfn) { 4324 early_node_map[i].end_pfn = end_pfn; 4325 return; 4326 } 4327 4328 /* Merge backward if suitable */ 4329 if (start_pfn < early_node_map[i].start_pfn && 4330 end_pfn >= early_node_map[i].start_pfn) { 4331 early_node_map[i].start_pfn = start_pfn; 4332 return; 4333 } 4334 } 4335 4336 /* Check that early_node_map is large enough */ 4337 if (i >= MAX_ACTIVE_REGIONS) { 4338 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4339 MAX_ACTIVE_REGIONS); 4340 return; 4341 } 4342 4343 early_node_map[i].nid = nid; 4344 early_node_map[i].start_pfn = start_pfn; 4345 early_node_map[i].end_pfn = end_pfn; 4346 nr_nodemap_entries = i + 1; 4347} 4348 4349/** 4350 * remove_active_range - Shrink an existing registered range of PFNs 4351 * @nid: The node id the range is on that should be shrunk 4352 * @start_pfn: The new PFN of the range 4353 * @end_pfn: The new PFN of the range 4354 * 4355 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4356 * The map is kept near the end physical page range that has already been 4357 * registered. This function allows an arch to shrink an existing registered 4358 * range. 4359 */ 4360void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4361 unsigned long end_pfn) 4362{ 4363 int i, j; 4364 int removed = 0; 4365 4366 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4367 nid, start_pfn, end_pfn); 4368 4369 /* Find the old active region end and shrink */ 4370 for_each_active_range_index_in_nid(i, nid) { 4371 if (early_node_map[i].start_pfn >= start_pfn && 4372 early_node_map[i].end_pfn <= end_pfn) { 4373 /* clear it */ 4374 early_node_map[i].start_pfn = 0; 4375 early_node_map[i].end_pfn = 0; 4376 removed = 1; 4377 continue; 4378 } 4379 if (early_node_map[i].start_pfn < start_pfn && 4380 early_node_map[i].end_pfn > start_pfn) { 4381 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4382 early_node_map[i].end_pfn = start_pfn; 4383 if (temp_end_pfn > end_pfn) 4384 add_active_range(nid, end_pfn, temp_end_pfn); 4385 continue; 4386 } 4387 if (early_node_map[i].start_pfn >= start_pfn && 4388 early_node_map[i].end_pfn > end_pfn && 4389 early_node_map[i].start_pfn < end_pfn) { 4390 early_node_map[i].start_pfn = end_pfn; 4391 continue; 4392 } 4393 } 4394 4395 if (!removed) 4396 return; 4397 4398 /* remove the blank ones */ 4399 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4400 if (early_node_map[i].nid != nid) 4401 continue; 4402 if (early_node_map[i].end_pfn) 4403 continue; 4404 /* we found it, get rid of it */ 4405 for (j = i; j < nr_nodemap_entries - 1; j++) 4406 memcpy(&early_node_map[j], &early_node_map[j+1], 4407 sizeof(early_node_map[j])); 4408 j = nr_nodemap_entries - 1; 4409 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4410 nr_nodemap_entries--; 4411 } 4412} 4413 4414/** 4415 * remove_all_active_ranges - Remove all currently registered regions 4416 * 4417 * During discovery, it may be found that a table like SRAT is invalid 4418 * and an alternative discovery method must be used. This function removes 4419 * all currently registered regions. 4420 */ 4421void __init remove_all_active_ranges(void) 4422{ 4423 memset(early_node_map, 0, sizeof(early_node_map)); 4424 nr_nodemap_entries = 0; 4425} 4426 4427/* Compare two active node_active_regions */ 4428static int __init cmp_node_active_region(const void *a, const void *b) 4429{ 4430 struct node_active_region *arange = (struct node_active_region *)a; 4431 struct node_active_region *brange = (struct node_active_region *)b; 4432 4433 /* Done this way to avoid overflows */ 4434 if (arange->start_pfn > brange->start_pfn) 4435 return 1; 4436 if (arange->start_pfn < brange->start_pfn) 4437 return -1; 4438 4439 return 0; 4440} 4441 4442/* sort the node_map by start_pfn */ 4443void __init sort_node_map(void) 4444{ 4445 sort(early_node_map, (size_t)nr_nodemap_entries, 4446 sizeof(struct node_active_region), 4447 cmp_node_active_region, NULL); 4448} 4449 4450/* Find the lowest pfn for a node */ 4451static unsigned long __init find_min_pfn_for_node(int nid) 4452{ 4453 int i; 4454 unsigned long min_pfn = ULONG_MAX; 4455 4456 /* Assuming a sorted map, the first range found has the starting pfn */ 4457 for_each_active_range_index_in_nid(i, nid) 4458 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4459 4460 if (min_pfn == ULONG_MAX) { 4461 printk(KERN_WARNING 4462 "Could not find start_pfn for node %d\n", nid); 4463 return 0; 4464 } 4465 4466 return min_pfn; 4467} 4468 4469/** 4470 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4471 * 4472 * It returns the minimum PFN based on information provided via 4473 * add_active_range(). 4474 */ 4475unsigned long __init find_min_pfn_with_active_regions(void) 4476{ 4477 return find_min_pfn_for_node(MAX_NUMNODES); 4478} 4479 4480/* 4481 * early_calculate_totalpages() 4482 * Sum pages in active regions for movable zone. 4483 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4484 */ 4485static unsigned long __init early_calculate_totalpages(void) 4486{ 4487 int i; 4488 unsigned long totalpages = 0; 4489 4490 for (i = 0; i < nr_nodemap_entries; i++) { 4491 unsigned long pages = early_node_map[i].end_pfn - 4492 early_node_map[i].start_pfn; 4493 totalpages += pages; 4494 if (pages) 4495 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4496 } 4497 return totalpages; 4498} 4499 4500/* 4501 * Find the PFN the Movable zone begins in each node. Kernel memory 4502 * is spread evenly between nodes as long as the nodes have enough 4503 * memory. When they don't, some nodes will have more kernelcore than 4504 * others 4505 */ 4506static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4507{ 4508 int i, nid; 4509 unsigned long usable_startpfn; 4510 unsigned long kernelcore_node, kernelcore_remaining; 4511 /* save the state before borrow the nodemask */ 4512 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4513 unsigned long totalpages = early_calculate_totalpages(); 4514 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4515 4516 /* 4517 * If movablecore was specified, calculate what size of 4518 * kernelcore that corresponds so that memory usable for 4519 * any allocation type is evenly spread. If both kernelcore 4520 * and movablecore are specified, then the value of kernelcore 4521 * will be used for required_kernelcore if it's greater than 4522 * what movablecore would have allowed. 4523 */ 4524 if (required_movablecore) { 4525 unsigned long corepages; 4526 4527 /* 4528 * Round-up so that ZONE_MOVABLE is at least as large as what 4529 * was requested by the user 4530 */ 4531 required_movablecore = 4532 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4533 corepages = totalpages - required_movablecore; 4534 4535 required_kernelcore = max(required_kernelcore, corepages); 4536 } 4537 4538 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4539 if (!required_kernelcore) 4540 goto out; 4541 4542 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4543 find_usable_zone_for_movable(); 4544 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4545 4546restart: 4547 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4548 kernelcore_node = required_kernelcore / usable_nodes; 4549 for_each_node_state(nid, N_HIGH_MEMORY) { 4550 /* 4551 * Recalculate kernelcore_node if the division per node 4552 * now exceeds what is necessary to satisfy the requested 4553 * amount of memory for the kernel 4554 */ 4555 if (required_kernelcore < kernelcore_node) 4556 kernelcore_node = required_kernelcore / usable_nodes; 4557 4558 /* 4559 * As the map is walked, we track how much memory is usable 4560 * by the kernel using kernelcore_remaining. When it is 4561 * 0, the rest of the node is usable by ZONE_MOVABLE 4562 */ 4563 kernelcore_remaining = kernelcore_node; 4564 4565 /* Go through each range of PFNs within this node */ 4566 for_each_active_range_index_in_nid(i, nid) { 4567 unsigned long start_pfn, end_pfn; 4568 unsigned long size_pages; 4569 4570 start_pfn = max(early_node_map[i].start_pfn, 4571 zone_movable_pfn[nid]); 4572 end_pfn = early_node_map[i].end_pfn; 4573 if (start_pfn >= end_pfn) 4574 continue; 4575 4576 /* Account for what is only usable for kernelcore */ 4577 if (start_pfn < usable_startpfn) { 4578 unsigned long kernel_pages; 4579 kernel_pages = min(end_pfn, usable_startpfn) 4580 - start_pfn; 4581 4582 kernelcore_remaining -= min(kernel_pages, 4583 kernelcore_remaining); 4584 required_kernelcore -= min(kernel_pages, 4585 required_kernelcore); 4586 4587 /* Continue if range is now fully accounted */ 4588 if (end_pfn <= usable_startpfn) { 4589 4590 /* 4591 * Push zone_movable_pfn to the end so 4592 * that if we have to rebalance 4593 * kernelcore across nodes, we will 4594 * not double account here 4595 */ 4596 zone_movable_pfn[nid] = end_pfn; 4597 continue; 4598 } 4599 start_pfn = usable_startpfn; 4600 } 4601 4602 /* 4603 * The usable PFN range for ZONE_MOVABLE is from 4604 * start_pfn->end_pfn. Calculate size_pages as the 4605 * number of pages used as kernelcore 4606 */ 4607 size_pages = end_pfn - start_pfn; 4608 if (size_pages > kernelcore_remaining) 4609 size_pages = kernelcore_remaining; 4610 zone_movable_pfn[nid] = start_pfn + size_pages; 4611 4612 /* 4613 * Some kernelcore has been met, update counts and 4614 * break if the kernelcore for this node has been 4615 * satisified 4616 */ 4617 required_kernelcore -= min(required_kernelcore, 4618 size_pages); 4619 kernelcore_remaining -= size_pages; 4620 if (!kernelcore_remaining) 4621 break; 4622 } 4623 } 4624 4625 /* 4626 * If there is still required_kernelcore, we do another pass with one 4627 * less node in the count. This will push zone_movable_pfn[nid] further 4628 * along on the nodes that still have memory until kernelcore is 4629 * satisified 4630 */ 4631 usable_nodes--; 4632 if (usable_nodes && required_kernelcore > usable_nodes) 4633 goto restart; 4634 4635 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4636 for (nid = 0; nid < MAX_NUMNODES; nid++) 4637 zone_movable_pfn[nid] = 4638 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4639 4640out: 4641 /* restore the node_state */ 4642 node_states[N_HIGH_MEMORY] = saved_node_state; 4643} 4644 4645/* Any regular memory on that node ? */ 4646static void check_for_regular_memory(pg_data_t *pgdat) 4647{ 4648#ifdef CONFIG_HIGHMEM 4649 enum zone_type zone_type; 4650 4651 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4652 struct zone *zone = &pgdat->node_zones[zone_type]; 4653 if (zone->present_pages) 4654 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4655 } 4656#endif 4657} 4658 4659/** 4660 * free_area_init_nodes - Initialise all pg_data_t and zone data 4661 * @max_zone_pfn: an array of max PFNs for each zone 4662 * 4663 * This will call free_area_init_node() for each active node in the system. 4664 * Using the page ranges provided by add_active_range(), the size of each 4665 * zone in each node and their holes is calculated. If the maximum PFN 4666 * between two adjacent zones match, it is assumed that the zone is empty. 4667 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4668 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4669 * starts where the previous one ended. For example, ZONE_DMA32 starts 4670 * at arch_max_dma_pfn. 4671 */ 4672void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4673{ 4674 unsigned long nid; 4675 int i; 4676 4677 /* Sort early_node_map as initialisation assumes it is sorted */ 4678 sort_node_map(); 4679 4680 /* Record where the zone boundaries are */ 4681 memset(arch_zone_lowest_possible_pfn, 0, 4682 sizeof(arch_zone_lowest_possible_pfn)); 4683 memset(arch_zone_highest_possible_pfn, 0, 4684 sizeof(arch_zone_highest_possible_pfn)); 4685 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4686 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4687 for (i = 1; i < MAX_NR_ZONES; i++) { 4688 if (i == ZONE_MOVABLE) 4689 continue; 4690 arch_zone_lowest_possible_pfn[i] = 4691 arch_zone_highest_possible_pfn[i-1]; 4692 arch_zone_highest_possible_pfn[i] = 4693 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4694 } 4695 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4696 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4697 4698 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4699 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4700 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4701 4702 /* Print out the zone ranges */ 4703 printk("Zone PFN ranges:\n"); 4704 for (i = 0; i < MAX_NR_ZONES; i++) { 4705 if (i == ZONE_MOVABLE) 4706 continue; 4707 printk(" %-8s ", zone_names[i]); 4708 if (arch_zone_lowest_possible_pfn[i] == 4709 arch_zone_highest_possible_pfn[i]) 4710 printk("empty\n"); 4711 else 4712 printk("%0#10lx -> %0#10lx\n", 4713 arch_zone_lowest_possible_pfn[i], 4714 arch_zone_highest_possible_pfn[i]); 4715 } 4716 4717 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4718 printk("Movable zone start PFN for each node\n"); 4719 for (i = 0; i < MAX_NUMNODES; i++) { 4720 if (zone_movable_pfn[i]) 4721 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4722 } 4723 4724 /* Print out the early_node_map[] */ 4725 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4726 for (i = 0; i < nr_nodemap_entries; i++) 4727 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4728 early_node_map[i].start_pfn, 4729 early_node_map[i].end_pfn); 4730 4731 /* Initialise every node */ 4732 mminit_verify_pageflags_layout(); 4733 setup_nr_node_ids(); 4734 for_each_online_node(nid) { 4735 pg_data_t *pgdat = NODE_DATA(nid); 4736 free_area_init_node(nid, NULL, 4737 find_min_pfn_for_node(nid), NULL); 4738 4739 /* Any memory on that node */ 4740 if (pgdat->node_present_pages) 4741 node_set_state(nid, N_HIGH_MEMORY); 4742 check_for_regular_memory(pgdat); 4743 } 4744} 4745 4746static int __init cmdline_parse_core(char *p, unsigned long *core) 4747{ 4748 unsigned long long coremem; 4749 if (!p) 4750 return -EINVAL; 4751 4752 coremem = memparse(p, &p); 4753 *core = coremem >> PAGE_SHIFT; 4754 4755 /* Paranoid check that UL is enough for the coremem value */ 4756 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4757 4758 return 0; 4759} 4760 4761/* 4762 * kernelcore=size sets the amount of memory for use for allocations that 4763 * cannot be reclaimed or migrated. 4764 */ 4765static int __init cmdline_parse_kernelcore(char *p) 4766{ 4767 return cmdline_parse_core(p, &required_kernelcore); 4768} 4769 4770/* 4771 * movablecore=size sets the amount of memory for use for allocations that 4772 * can be reclaimed or migrated. 4773 */ 4774static int __init cmdline_parse_movablecore(char *p) 4775{ 4776 return cmdline_parse_core(p, &required_movablecore); 4777} 4778 4779early_param("kernelcore", cmdline_parse_kernelcore); 4780early_param("movablecore", cmdline_parse_movablecore); 4781 4782#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4783 4784/** 4785 * set_dma_reserve - set the specified number of pages reserved in the first zone 4786 * @new_dma_reserve: The number of pages to mark reserved 4787 * 4788 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4789 * In the DMA zone, a significant percentage may be consumed by kernel image 4790 * and other unfreeable allocations which can skew the watermarks badly. This 4791 * function may optionally be used to account for unfreeable pages in the 4792 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4793 * smaller per-cpu batchsize. 4794 */ 4795void __init set_dma_reserve(unsigned long new_dma_reserve) 4796{ 4797 dma_reserve = new_dma_reserve; 4798} 4799 4800#ifndef CONFIG_NEED_MULTIPLE_NODES 4801struct pglist_data __refdata contig_page_data = { 4802#ifndef CONFIG_NO_BOOTMEM 4803 .bdata = &bootmem_node_data[0] 4804#endif 4805 }; 4806EXPORT_SYMBOL(contig_page_data); 4807#endif 4808 4809void __init free_area_init(unsigned long *zones_size) 4810{ 4811 free_area_init_node(0, zones_size, 4812 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4813} 4814 4815static int page_alloc_cpu_notify(struct notifier_block *self, 4816 unsigned long action, void *hcpu) 4817{ 4818 int cpu = (unsigned long)hcpu; 4819 4820 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4821 drain_pages(cpu); 4822 4823 /* 4824 * Spill the event counters of the dead processor 4825 * into the current processors event counters. 4826 * This artificially elevates the count of the current 4827 * processor. 4828 */ 4829 vm_events_fold_cpu(cpu); 4830 4831 /* 4832 * Zero the differential counters of the dead processor 4833 * so that the vm statistics are consistent. 4834 * 4835 * This is only okay since the processor is dead and cannot 4836 * race with what we are doing. 4837 */ 4838 refresh_cpu_vm_stats(cpu); 4839 } 4840 return NOTIFY_OK; 4841} 4842 4843void __init page_alloc_init(void) 4844{ 4845 hotcpu_notifier(page_alloc_cpu_notify, 0); 4846} 4847 4848/* 4849 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4850 * or min_free_kbytes changes. 4851 */ 4852static void calculate_totalreserve_pages(void) 4853{ 4854 struct pglist_data *pgdat; 4855 unsigned long reserve_pages = 0; 4856 enum zone_type i, j; 4857 4858 for_each_online_pgdat(pgdat) { 4859 for (i = 0; i < MAX_NR_ZONES; i++) { 4860 struct zone *zone = pgdat->node_zones + i; 4861 unsigned long max = 0; 4862 4863 /* Find valid and maximum lowmem_reserve in the zone */ 4864 for (j = i; j < MAX_NR_ZONES; j++) { 4865 if (zone->lowmem_reserve[j] > max) 4866 max = zone->lowmem_reserve[j]; 4867 } 4868 4869 /* we treat the high watermark as reserved pages. */ 4870 max += high_wmark_pages(zone); 4871 4872 if (max > zone->present_pages) 4873 max = zone->present_pages; 4874 reserve_pages += max; 4875 } 4876 } 4877 totalreserve_pages = reserve_pages; 4878} 4879 4880/* 4881 * setup_per_zone_lowmem_reserve - called whenever 4882 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4883 * has a correct pages reserved value, so an adequate number of 4884 * pages are left in the zone after a successful __alloc_pages(). 4885 */ 4886static void setup_per_zone_lowmem_reserve(void) 4887{ 4888 struct pglist_data *pgdat; 4889 enum zone_type j, idx; 4890 4891 for_each_online_pgdat(pgdat) { 4892 for (j = 0; j < MAX_NR_ZONES; j++) { 4893 struct zone *zone = pgdat->node_zones + j; 4894 unsigned long present_pages = zone->present_pages; 4895 4896 zone->lowmem_reserve[j] = 0; 4897 4898 idx = j; 4899 while (idx) { 4900 struct zone *lower_zone; 4901 4902 idx--; 4903 4904 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4905 sysctl_lowmem_reserve_ratio[idx] = 1; 4906 4907 lower_zone = pgdat->node_zones + idx; 4908 lower_zone->lowmem_reserve[j] = present_pages / 4909 sysctl_lowmem_reserve_ratio[idx]; 4910 present_pages += lower_zone->present_pages; 4911 } 4912 } 4913 } 4914 4915 /* update totalreserve_pages */ 4916 calculate_totalreserve_pages(); 4917} 4918 4919/** 4920 * setup_per_zone_wmarks - called when min_free_kbytes changes 4921 * or when memory is hot-{added|removed} 4922 * 4923 * Ensures that the watermark[min,low,high] values for each zone are set 4924 * correctly with respect to min_free_kbytes. 4925 */ 4926void setup_per_zone_wmarks(void) 4927{ 4928 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4929 unsigned long lowmem_pages = 0; 4930 struct zone *zone; 4931 unsigned long flags; 4932 4933 /* Calculate total number of !ZONE_HIGHMEM pages */ 4934 for_each_zone(zone) { 4935 if (!is_highmem(zone)) 4936 lowmem_pages += zone->present_pages; 4937 } 4938 4939 for_each_zone(zone) { 4940 u64 tmp; 4941 4942 spin_lock_irqsave(&zone->lock, flags); 4943 tmp = (u64)pages_min * zone->present_pages; 4944 do_div(tmp, lowmem_pages); 4945 if (is_highmem(zone)) { 4946 /* 4947 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4948 * need highmem pages, so cap pages_min to a small 4949 * value here. 4950 * 4951 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4952 * deltas controls asynch page reclaim, and so should 4953 * not be capped for highmem. 4954 */ 4955 int min_pages; 4956 4957 min_pages = zone->present_pages / 1024; 4958 if (min_pages < SWAP_CLUSTER_MAX) 4959 min_pages = SWAP_CLUSTER_MAX; 4960 if (min_pages > 128) 4961 min_pages = 128; 4962 zone->watermark[WMARK_MIN] = min_pages; 4963 } else { 4964 /* 4965 * If it's a lowmem zone, reserve a number of pages 4966 * proportionate to the zone's size. 4967 */ 4968 zone->watermark[WMARK_MIN] = tmp; 4969 } 4970 4971 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4972 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4973 setup_zone_migrate_reserve(zone); 4974 spin_unlock_irqrestore(&zone->lock, flags); 4975 } 4976 4977 /* update totalreserve_pages */ 4978 calculate_totalreserve_pages(); 4979} 4980 4981/* 4982 * The inactive anon list should be small enough that the VM never has to 4983 * do too much work, but large enough that each inactive page has a chance 4984 * to be referenced again before it is swapped out. 4985 * 4986 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4987 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4988 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4989 * the anonymous pages are kept on the inactive list. 4990 * 4991 * total target max 4992 * memory ratio inactive anon 4993 * ------------------------------------- 4994 * 10MB 1 5MB 4995 * 100MB 1 50MB 4996 * 1GB 3 250MB 4997 * 10GB 10 0.9GB 4998 * 100GB 31 3GB 4999 * 1TB 101 10GB 5000 * 10TB 320 32GB 5001 */ 5002void calculate_zone_inactive_ratio(struct zone *zone) 5003{ 5004 unsigned int gb, ratio; 5005 5006 /* Zone size in gigabytes */ 5007 gb = zone->present_pages >> (30 - PAGE_SHIFT); 5008 if (gb) 5009 ratio = int_sqrt(10 * gb); 5010 else 5011 ratio = 1; 5012 5013 zone->inactive_ratio = ratio; 5014} 5015 5016static void __init setup_per_zone_inactive_ratio(void) 5017{ 5018 struct zone *zone; 5019 5020 for_each_zone(zone) 5021 calculate_zone_inactive_ratio(zone); 5022} 5023 5024/* 5025 * Initialise min_free_kbytes. 5026 * 5027 * For small machines we want it small (128k min). For large machines 5028 * we want it large (64MB max). But it is not linear, because network 5029 * bandwidth does not increase linearly with machine size. We use 5030 * 5031 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5032 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5033 * 5034 * which yields 5035 * 5036 * 16MB: 512k 5037 * 32MB: 724k 5038 * 64MB: 1024k 5039 * 128MB: 1448k 5040 * 256MB: 2048k 5041 * 512MB: 2896k 5042 * 1024MB: 4096k 5043 * 2048MB: 5792k 5044 * 4096MB: 8192k 5045 * 8192MB: 11584k 5046 * 16384MB: 16384k 5047 */ 5048static int __init init_per_zone_wmark_min(void) 5049{ 5050 unsigned long lowmem_kbytes; 5051 5052 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5053 5054 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5055 if (min_free_kbytes < 128) 5056 min_free_kbytes = 128; 5057 if (min_free_kbytes > 65536) 5058 min_free_kbytes = 65536; 5059 setup_per_zone_wmarks(); 5060 setup_per_zone_lowmem_reserve(); 5061 setup_per_zone_inactive_ratio(); 5062 return 0; 5063} 5064module_init(init_per_zone_wmark_min) 5065 5066/* 5067 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5068 * that we can call two helper functions whenever min_free_kbytes 5069 * changes. 5070 */ 5071int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5072 void __user *buffer, size_t *length, loff_t *ppos) 5073{ 5074 proc_dointvec(table, write, buffer, length, ppos); 5075 if (write) 5076 setup_per_zone_wmarks(); 5077 return 0; 5078} 5079 5080#ifdef CONFIG_NUMA 5081int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5082 void __user *buffer, size_t *length, loff_t *ppos) 5083{ 5084 struct zone *zone; 5085 int rc; 5086 5087 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5088 if (rc) 5089 return rc; 5090 5091 for_each_zone(zone) 5092 zone->min_unmapped_pages = (zone->present_pages * 5093 sysctl_min_unmapped_ratio) / 100; 5094 return 0; 5095} 5096 5097int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5098 void __user *buffer, size_t *length, loff_t *ppos) 5099{ 5100 struct zone *zone; 5101 int rc; 5102 5103 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5104 if (rc) 5105 return rc; 5106 5107 for_each_zone(zone) 5108 zone->min_slab_pages = (zone->present_pages * 5109 sysctl_min_slab_ratio) / 100; 5110 return 0; 5111} 5112#endif 5113 5114/* 5115 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5116 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5117 * whenever sysctl_lowmem_reserve_ratio changes. 5118 * 5119 * The reserve ratio obviously has absolutely no relation with the 5120 * minimum watermarks. The lowmem reserve ratio can only make sense 5121 * if in function of the boot time zone sizes. 5122 */ 5123int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5124 void __user *buffer, size_t *length, loff_t *ppos) 5125{ 5126 proc_dointvec_minmax(table, write, buffer, length, ppos); 5127 setup_per_zone_lowmem_reserve(); 5128 return 0; 5129} 5130 5131/* 5132 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5133 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5134 * can have before it gets flushed back to buddy allocator. 5135 */ 5136 5137int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5138 void __user *buffer, size_t *length, loff_t *ppos) 5139{ 5140 struct zone *zone; 5141 unsigned int cpu; 5142 int ret; 5143 5144 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5145 if (!write || (ret == -EINVAL)) 5146 return ret; 5147 for_each_populated_zone(zone) { 5148 for_each_possible_cpu(cpu) { 5149 unsigned long high; 5150 high = zone->present_pages / percpu_pagelist_fraction; 5151 setup_pagelist_highmark( 5152 per_cpu_ptr(zone->pageset, cpu), high); 5153 } 5154 } 5155 return 0; 5156} 5157 5158int hashdist = HASHDIST_DEFAULT; 5159 5160#ifdef CONFIG_NUMA 5161static int __init set_hashdist(char *str) 5162{ 5163 if (!str) 5164 return 0; 5165 hashdist = simple_strtoul(str, &str, 0); 5166 return 1; 5167} 5168__setup("hashdist=", set_hashdist); 5169#endif 5170 5171/* 5172 * allocate a large system hash table from bootmem 5173 * - it is assumed that the hash table must contain an exact power-of-2 5174 * quantity of entries 5175 * - limit is the number of hash buckets, not the total allocation size 5176 */ 5177void *__init alloc_large_system_hash(const char *tablename, 5178 unsigned long bucketsize, 5179 unsigned long numentries, 5180 int scale, 5181 int flags, 5182 unsigned int *_hash_shift, 5183 unsigned int *_hash_mask, 5184 unsigned long limit) 5185{ 5186 unsigned long long max = limit; 5187 unsigned long log2qty, size; 5188 void *table = NULL; 5189 5190 /* allow the kernel cmdline to have a say */ 5191 if (!numentries) { 5192 /* round applicable memory size up to nearest megabyte */ 5193 numentries = nr_kernel_pages; 5194 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5195 numentries >>= 20 - PAGE_SHIFT; 5196 numentries <<= 20 - PAGE_SHIFT; 5197 5198 /* limit to 1 bucket per 2^scale bytes of low memory */ 5199 if (scale > PAGE_SHIFT) 5200 numentries >>= (scale - PAGE_SHIFT); 5201 else 5202 numentries <<= (PAGE_SHIFT - scale); 5203 5204 /* Make sure we've got at least a 0-order allocation.. */ 5205 if (unlikely(flags & HASH_SMALL)) { 5206 /* Makes no sense without HASH_EARLY */ 5207 WARN_ON(!(flags & HASH_EARLY)); 5208 if (!(numentries >> *_hash_shift)) { 5209 numentries = 1UL << *_hash_shift; 5210 BUG_ON(!numentries); 5211 } 5212 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5213 numentries = PAGE_SIZE / bucketsize; 5214 } 5215 numentries = roundup_pow_of_two(numentries); 5216 5217 /* limit allocation size to 1/16 total memory by default */ 5218 if (max == 0) { 5219 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5220 do_div(max, bucketsize); 5221 } 5222 5223 if (numentries > max) 5224 numentries = max; 5225 5226 log2qty = ilog2(numentries); 5227 5228 do { 5229 size = bucketsize << log2qty; 5230 if (flags & HASH_EARLY) 5231 table = alloc_bootmem_nopanic(size); 5232 else if (hashdist) 5233 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5234 else { 5235 /* 5236 * If bucketsize is not a power-of-two, we may free 5237 * some pages at the end of hash table which 5238 * alloc_pages_exact() automatically does 5239 */ 5240 if (get_order(size) < MAX_ORDER) { 5241 table = alloc_pages_exact(size, GFP_ATOMIC); 5242 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5243 } 5244 } 5245 } while (!table && size > PAGE_SIZE && --log2qty); 5246 5247 if (!table) 5248 panic("Failed to allocate %s hash table\n", tablename); 5249 5250 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5251 tablename, 5252 (1UL << log2qty), 5253 ilog2(size) - PAGE_SHIFT, 5254 size); 5255 5256 if (_hash_shift) 5257 *_hash_shift = log2qty; 5258 if (_hash_mask) 5259 *_hash_mask = (1 << log2qty) - 1; 5260 5261 return table; 5262} 5263 5264/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5265static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5266 unsigned long pfn) 5267{ 5268#ifdef CONFIG_SPARSEMEM 5269 return __pfn_to_section(pfn)->pageblock_flags; 5270#else 5271 return zone->pageblock_flags; 5272#endif /* CONFIG_SPARSEMEM */ 5273} 5274 5275static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5276{ 5277#ifdef CONFIG_SPARSEMEM 5278 pfn &= (PAGES_PER_SECTION-1); 5279 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5280#else 5281 pfn = pfn - zone->zone_start_pfn; 5282 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5283#endif /* CONFIG_SPARSEMEM */ 5284} 5285 5286/** 5287 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5288 * @page: The page within the block of interest 5289 * @start_bitidx: The first bit of interest to retrieve 5290 * @end_bitidx: The last bit of interest 5291 * returns pageblock_bits flags 5292 */ 5293unsigned long get_pageblock_flags_group(struct page *page, 5294 int start_bitidx, int end_bitidx) 5295{ 5296 struct zone *zone; 5297 unsigned long *bitmap; 5298 unsigned long pfn, bitidx; 5299 unsigned long flags = 0; 5300 unsigned long value = 1; 5301 5302 zone = page_zone(page); 5303 pfn = page_to_pfn(page); 5304 bitmap = get_pageblock_bitmap(zone, pfn); 5305 bitidx = pfn_to_bitidx(zone, pfn); 5306 5307 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5308 if (test_bit(bitidx + start_bitidx, bitmap)) 5309 flags |= value; 5310 5311 return flags; 5312} 5313 5314/** 5315 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5316 * @page: The page within the block of interest 5317 * @start_bitidx: The first bit of interest 5318 * @end_bitidx: The last bit of interest 5319 * @flags: The flags to set 5320 */ 5321void set_pageblock_flags_group(struct page *page, unsigned long flags, 5322 int start_bitidx, int end_bitidx) 5323{ 5324 struct zone *zone; 5325 unsigned long *bitmap; 5326 unsigned long pfn, bitidx; 5327 unsigned long value = 1; 5328 5329 zone = page_zone(page); 5330 pfn = page_to_pfn(page); 5331 bitmap = get_pageblock_bitmap(zone, pfn); 5332 bitidx = pfn_to_bitidx(zone, pfn); 5333 VM_BUG_ON(pfn < zone->zone_start_pfn); 5334 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5335 5336 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5337 if (flags & value) 5338 __set_bit(bitidx + start_bitidx, bitmap); 5339 else 5340 __clear_bit(bitidx + start_bitidx, bitmap); 5341} 5342 5343/* 5344 * This is designed as sub function...plz see page_isolation.c also. 5345 * set/clear page block's type to be ISOLATE. 5346 * page allocater never alloc memory from ISOLATE block. 5347 */ 5348 5349static int 5350__count_immobile_pages(struct zone *zone, struct page *page, int count) 5351{ 5352 unsigned long pfn, iter, found; 5353 /* 5354 * For avoiding noise data, lru_add_drain_all() should be called 5355 * If ZONE_MOVABLE, the zone never contains immobile pages 5356 */ 5357 if (zone_idx(zone) == ZONE_MOVABLE) 5358 return true; 5359 5360 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE) 5361 return true; 5362 5363 pfn = page_to_pfn(page); 5364 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 5365 unsigned long check = pfn + iter; 5366 5367 if (!pfn_valid_within(check)) { 5368 iter++; 5369 continue; 5370 } 5371 page = pfn_to_page(check); 5372 if (!page_count(page)) { 5373 if (PageBuddy(page)) 5374 iter += (1 << page_order(page)) - 1; 5375 continue; 5376 } 5377 if (!PageLRU(page)) 5378 found++; 5379 /* 5380 * If there are RECLAIMABLE pages, we need to check it. 5381 * But now, memory offline itself doesn't call shrink_slab() 5382 * and it still to be fixed. 5383 */ 5384 /* 5385 * If the page is not RAM, page_count()should be 0. 5386 * we don't need more check. This is an _used_ not-movable page. 5387 * 5388 * The problematic thing here is PG_reserved pages. PG_reserved 5389 * is set to both of a memory hole page and a _used_ kernel 5390 * page at boot. 5391 */ 5392 if (found > count) 5393 return false; 5394 } 5395 return true; 5396} 5397 5398bool is_pageblock_removable_nolock(struct page *page) 5399{ 5400 struct zone *zone = page_zone(page); 5401 return __count_immobile_pages(zone, page, 0); 5402} 5403 5404int set_migratetype_isolate(struct page *page) 5405{ 5406 struct zone *zone; 5407 unsigned long flags, pfn; 5408 struct memory_isolate_notify arg; 5409 int notifier_ret; 5410 int ret = -EBUSY; 5411 int zone_idx; 5412 5413 zone = page_zone(page); 5414 zone_idx = zone_idx(zone); 5415 5416 spin_lock_irqsave(&zone->lock, flags); 5417 5418 pfn = page_to_pfn(page); 5419 arg.start_pfn = pfn; 5420 arg.nr_pages = pageblock_nr_pages; 5421 arg.pages_found = 0; 5422 5423 /* 5424 * It may be possible to isolate a pageblock even if the 5425 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5426 * notifier chain is used by balloon drivers to return the 5427 * number of pages in a range that are held by the balloon 5428 * driver to shrink memory. If all the pages are accounted for 5429 * by balloons, are free, or on the LRU, isolation can continue. 5430 * Later, for example, when memory hotplug notifier runs, these 5431 * pages reported as "can be isolated" should be isolated(freed) 5432 * by the balloon driver through the memory notifier chain. 5433 */ 5434 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5435 notifier_ret = notifier_to_errno(notifier_ret); 5436 if (notifier_ret) 5437 goto out; 5438 /* 5439 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. 5440 * We just check MOVABLE pages. 5441 */ 5442 if (__count_immobile_pages(zone, page, arg.pages_found)) 5443 ret = 0; 5444 5445 /* 5446 * immobile means "not-on-lru" paes. If immobile is larger than 5447 * removable-by-driver pages reported by notifier, we'll fail. 5448 */ 5449 5450out: 5451 if (!ret) { 5452 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5453 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5454 } 5455 5456 spin_unlock_irqrestore(&zone->lock, flags); 5457 if (!ret) 5458 drain_all_pages(); 5459 return ret; 5460} 5461 5462void unset_migratetype_isolate(struct page *page) 5463{ 5464 struct zone *zone; 5465 unsigned long flags; 5466 zone = page_zone(page); 5467 spin_lock_irqsave(&zone->lock, flags); 5468 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5469 goto out; 5470 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5471 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5472out: 5473 spin_unlock_irqrestore(&zone->lock, flags); 5474} 5475 5476#ifdef CONFIG_MEMORY_HOTREMOVE 5477/* 5478 * All pages in the range must be isolated before calling this. 5479 */ 5480void 5481__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5482{ 5483 struct page *page; 5484 struct zone *zone; 5485 int order, i; 5486 unsigned long pfn; 5487 unsigned long flags; 5488 /* find the first valid pfn */ 5489 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5490 if (pfn_valid(pfn)) 5491 break; 5492 if (pfn == end_pfn) 5493 return; 5494 zone = page_zone(pfn_to_page(pfn)); 5495 spin_lock_irqsave(&zone->lock, flags); 5496 pfn = start_pfn; 5497 while (pfn < end_pfn) { 5498 if (!pfn_valid(pfn)) { 5499 pfn++; 5500 continue; 5501 } 5502 page = pfn_to_page(pfn); 5503 BUG_ON(page_count(page)); 5504 BUG_ON(!PageBuddy(page)); 5505 order = page_order(page); 5506#ifdef CONFIG_DEBUG_VM 5507 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5508 pfn, 1 << order, end_pfn); 5509#endif 5510 list_del(&page->lru); 5511 rmv_page_order(page); 5512 zone->free_area[order].nr_free--; 5513 __mod_zone_page_state(zone, NR_FREE_PAGES, 5514 - (1UL << order)); 5515 for (i = 0; i < (1 << order); i++) 5516 SetPageReserved((page+i)); 5517 pfn += (1 << order); 5518 } 5519 spin_unlock_irqrestore(&zone->lock, flags); 5520} 5521#endif 5522 5523#ifdef CONFIG_MEMORY_FAILURE 5524bool is_free_buddy_page(struct page *page) 5525{ 5526 struct zone *zone = page_zone(page); 5527 unsigned long pfn = page_to_pfn(page); 5528 unsigned long flags; 5529 int order; 5530 5531 spin_lock_irqsave(&zone->lock, flags); 5532 for (order = 0; order < MAX_ORDER; order++) { 5533 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5534 5535 if (PageBuddy(page_head) && page_order(page_head) >= order) 5536 break; 5537 } 5538 spin_unlock_irqrestore(&zone->lock, flags); 5539 5540 return order < MAX_ORDER; 5541} 5542#endif 5543 5544static struct trace_print_flags pageflag_names[] = { 5545 {1UL << PG_locked, "locked" }, 5546 {1UL << PG_error, "error" }, 5547 {1UL << PG_referenced, "referenced" }, 5548 {1UL << PG_uptodate, "uptodate" }, 5549 {1UL << PG_dirty, "dirty" }, 5550 {1UL << PG_lru, "lru" }, 5551 {1UL << PG_active, "active" }, 5552 {1UL << PG_slab, "slab" }, 5553 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5554 {1UL << PG_arch_1, "arch_1" }, 5555 {1UL << PG_reserved, "reserved" }, 5556 {1UL << PG_private, "private" }, 5557 {1UL << PG_private_2, "private_2" }, 5558 {1UL << PG_writeback, "writeback" }, 5559#ifdef CONFIG_PAGEFLAGS_EXTENDED 5560 {1UL << PG_head, "head" }, 5561 {1UL << PG_tail, "tail" }, 5562#else 5563 {1UL << PG_compound, "compound" }, 5564#endif 5565 {1UL << PG_swapcache, "swapcache" }, 5566 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5567 {1UL << PG_reclaim, "reclaim" }, 5568 {1UL << PG_swapbacked, "swapbacked" }, 5569 {1UL << PG_unevictable, "unevictable" }, 5570#ifdef CONFIG_MMU 5571 {1UL << PG_mlocked, "mlocked" }, 5572#endif 5573#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5574 {1UL << PG_uncached, "uncached" }, 5575#endif 5576#ifdef CONFIG_MEMORY_FAILURE 5577 {1UL << PG_hwpoison, "hwpoison" }, 5578#endif 5579 {-1UL, NULL }, 5580}; 5581 5582static void dump_page_flags(unsigned long flags) 5583{ 5584 const char *delim = ""; 5585 unsigned long mask; 5586 int i; 5587 5588 printk(KERN_ALERT "page flags: %#lx(", flags); 5589 5590 /* remove zone id */ 5591 flags &= (1UL << NR_PAGEFLAGS) - 1; 5592 5593 for (i = 0; pageflag_names[i].name && flags; i++) { 5594 5595 mask = pageflag_names[i].mask; 5596 if ((flags & mask) != mask) 5597 continue; 5598 5599 flags &= ~mask; 5600 printk("%s%s", delim, pageflag_names[i].name); 5601 delim = "|"; 5602 } 5603 5604 /* check for left over flags */ 5605 if (flags) 5606 printk("%s%#lx", delim, flags); 5607 5608 printk(")\n"); 5609} 5610 5611void dump_page(struct page *page) 5612{ 5613 printk(KERN_ALERT 5614 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5615 page, atomic_read(&page->_count), page_mapcount(page), 5616 page->mapping, page->index); 5617 dump_page_flags(page->flags); 5618} 5619