page_alloc.c revision 1f522509c77a5dea8dc384b735314f03908a6415
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/compiler.h> 25#include <linux/kernel.h> 26#include <linux/kmemcheck.h> 27#include <linux/module.h> 28#include <linux/suspend.h> 29#include <linux/pagevec.h> 30#include <linux/blkdev.h> 31#include <linux/slab.h> 32#include <linux/oom.h> 33#include <linux/notifier.h> 34#include <linux/topology.h> 35#include <linux/sysctl.h> 36#include <linux/cpu.h> 37#include <linux/cpuset.h> 38#include <linux/memory_hotplug.h> 39#include <linux/nodemask.h> 40#include <linux/vmalloc.h> 41#include <linux/mempolicy.h> 42#include <linux/stop_machine.h> 43#include <linux/sort.h> 44#include <linux/pfn.h> 45#include <linux/backing-dev.h> 46#include <linux/fault-inject.h> 47#include <linux/page-isolation.h> 48#include <linux/page_cgroup.h> 49#include <linux/debugobjects.h> 50#include <linux/kmemleak.h> 51#include <linux/memory.h> 52#include <linux/compaction.h> 53#include <trace/events/kmem.h> 54#include <linux/ftrace_event.h> 55 56#include <asm/tlbflush.h> 57#include <asm/div64.h> 58#include "internal.h" 59 60/* 61 * Array of node states. 62 */ 63nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 64 [N_POSSIBLE] = NODE_MASK_ALL, 65 [N_ONLINE] = { { [0] = 1UL } }, 66#ifndef CONFIG_NUMA 67 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 68#ifdef CONFIG_HIGHMEM 69 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 70#endif 71 [N_CPU] = { { [0] = 1UL } }, 72#endif /* NUMA */ 73}; 74EXPORT_SYMBOL(node_states); 75 76unsigned long totalram_pages __read_mostly; 77unsigned long totalreserve_pages __read_mostly; 78int percpu_pagelist_fraction; 79gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 80 81#ifdef CONFIG_PM_SLEEP 82/* 83 * The following functions are used by the suspend/hibernate code to temporarily 84 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 85 * while devices are suspended. To avoid races with the suspend/hibernate code, 86 * they should always be called with pm_mutex held (gfp_allowed_mask also should 87 * only be modified with pm_mutex held, unless the suspend/hibernate code is 88 * guaranteed not to run in parallel with that modification). 89 */ 90void set_gfp_allowed_mask(gfp_t mask) 91{ 92 WARN_ON(!mutex_is_locked(&pm_mutex)); 93 gfp_allowed_mask = mask; 94} 95 96gfp_t clear_gfp_allowed_mask(gfp_t mask) 97{ 98 gfp_t ret = gfp_allowed_mask; 99 100 WARN_ON(!mutex_is_locked(&pm_mutex)); 101 gfp_allowed_mask &= ~mask; 102 return ret; 103} 104#endif /* CONFIG_PM_SLEEP */ 105 106#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 107int pageblock_order __read_mostly; 108#endif 109 110static void __free_pages_ok(struct page *page, unsigned int order); 111 112/* 113 * results with 256, 32 in the lowmem_reserve sysctl: 114 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 115 * 1G machine -> (16M dma, 784M normal, 224M high) 116 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 117 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 118 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 119 * 120 * TBD: should special case ZONE_DMA32 machines here - in those we normally 121 * don't need any ZONE_NORMAL reservation 122 */ 123int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 124#ifdef CONFIG_ZONE_DMA 125 256, 126#endif 127#ifdef CONFIG_ZONE_DMA32 128 256, 129#endif 130#ifdef CONFIG_HIGHMEM 131 32, 132#endif 133 32, 134}; 135 136EXPORT_SYMBOL(totalram_pages); 137 138static char * const zone_names[MAX_NR_ZONES] = { 139#ifdef CONFIG_ZONE_DMA 140 "DMA", 141#endif 142#ifdef CONFIG_ZONE_DMA32 143 "DMA32", 144#endif 145 "Normal", 146#ifdef CONFIG_HIGHMEM 147 "HighMem", 148#endif 149 "Movable", 150}; 151 152int min_free_kbytes = 1024; 153 154static unsigned long __meminitdata nr_kernel_pages; 155static unsigned long __meminitdata nr_all_pages; 156static unsigned long __meminitdata dma_reserve; 157 158#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 159 /* 160 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 161 * ranges of memory (RAM) that may be registered with add_active_range(). 162 * Ranges passed to add_active_range() will be merged if possible 163 * so the number of times add_active_range() can be called is 164 * related to the number of nodes and the number of holes 165 */ 166 #ifdef CONFIG_MAX_ACTIVE_REGIONS 167 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 168 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 169 #else 170 #if MAX_NUMNODES >= 32 171 /* If there can be many nodes, allow up to 50 holes per node */ 172 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 173 #else 174 /* By default, allow up to 256 distinct regions */ 175 #define MAX_ACTIVE_REGIONS 256 176 #endif 177 #endif 178 179 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 180 static int __meminitdata nr_nodemap_entries; 181 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 182 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 183 static unsigned long __initdata required_kernelcore; 184 static unsigned long __initdata required_movablecore; 185 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 186 187 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 188 int movable_zone; 189 EXPORT_SYMBOL(movable_zone); 190#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 191 192#if MAX_NUMNODES > 1 193int nr_node_ids __read_mostly = MAX_NUMNODES; 194int nr_online_nodes __read_mostly = 1; 195EXPORT_SYMBOL(nr_node_ids); 196EXPORT_SYMBOL(nr_online_nodes); 197#endif 198 199int page_group_by_mobility_disabled __read_mostly; 200 201static void set_pageblock_migratetype(struct page *page, int migratetype) 202{ 203 204 if (unlikely(page_group_by_mobility_disabled)) 205 migratetype = MIGRATE_UNMOVABLE; 206 207 set_pageblock_flags_group(page, (unsigned long)migratetype, 208 PB_migrate, PB_migrate_end); 209} 210 211bool oom_killer_disabled __read_mostly; 212 213#ifdef CONFIG_DEBUG_VM 214static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 215{ 216 int ret = 0; 217 unsigned seq; 218 unsigned long pfn = page_to_pfn(page); 219 220 do { 221 seq = zone_span_seqbegin(zone); 222 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 223 ret = 1; 224 else if (pfn < zone->zone_start_pfn) 225 ret = 1; 226 } while (zone_span_seqretry(zone, seq)); 227 228 return ret; 229} 230 231static int page_is_consistent(struct zone *zone, struct page *page) 232{ 233 if (!pfn_valid_within(page_to_pfn(page))) 234 return 0; 235 if (zone != page_zone(page)) 236 return 0; 237 238 return 1; 239} 240/* 241 * Temporary debugging check for pages not lying within a given zone. 242 */ 243static int bad_range(struct zone *zone, struct page *page) 244{ 245 if (page_outside_zone_boundaries(zone, page)) 246 return 1; 247 if (!page_is_consistent(zone, page)) 248 return 1; 249 250 return 0; 251} 252#else 253static inline int bad_range(struct zone *zone, struct page *page) 254{ 255 return 0; 256} 257#endif 258 259static void bad_page(struct page *page) 260{ 261 static unsigned long resume; 262 static unsigned long nr_shown; 263 static unsigned long nr_unshown; 264 265 /* Don't complain about poisoned pages */ 266 if (PageHWPoison(page)) { 267 __ClearPageBuddy(page); 268 return; 269 } 270 271 /* 272 * Allow a burst of 60 reports, then keep quiet for that minute; 273 * or allow a steady drip of one report per second. 274 */ 275 if (nr_shown == 60) { 276 if (time_before(jiffies, resume)) { 277 nr_unshown++; 278 goto out; 279 } 280 if (nr_unshown) { 281 printk(KERN_ALERT 282 "BUG: Bad page state: %lu messages suppressed\n", 283 nr_unshown); 284 nr_unshown = 0; 285 } 286 nr_shown = 0; 287 } 288 if (nr_shown++ == 0) 289 resume = jiffies + 60 * HZ; 290 291 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 292 current->comm, page_to_pfn(page)); 293 dump_page(page); 294 295 dump_stack(); 296out: 297 /* Leave bad fields for debug, except PageBuddy could make trouble */ 298 __ClearPageBuddy(page); 299 add_taint(TAINT_BAD_PAGE); 300} 301 302/* 303 * Higher-order pages are called "compound pages". They are structured thusly: 304 * 305 * The first PAGE_SIZE page is called the "head page". 306 * 307 * The remaining PAGE_SIZE pages are called "tail pages". 308 * 309 * All pages have PG_compound set. All pages have their ->private pointing at 310 * the head page (even the head page has this). 311 * 312 * The first tail page's ->lru.next holds the address of the compound page's 313 * put_page() function. Its ->lru.prev holds the order of allocation. 314 * This usage means that zero-order pages may not be compound. 315 */ 316 317static void free_compound_page(struct page *page) 318{ 319 __free_pages_ok(page, compound_order(page)); 320} 321 322void prep_compound_page(struct page *page, unsigned long order) 323{ 324 int i; 325 int nr_pages = 1 << order; 326 327 set_compound_page_dtor(page, free_compound_page); 328 set_compound_order(page, order); 329 __SetPageHead(page); 330 for (i = 1; i < nr_pages; i++) { 331 struct page *p = page + i; 332 333 __SetPageTail(p); 334 p->first_page = page; 335 } 336} 337 338static int destroy_compound_page(struct page *page, unsigned long order) 339{ 340 int i; 341 int nr_pages = 1 << order; 342 int bad = 0; 343 344 if (unlikely(compound_order(page) != order) || 345 unlikely(!PageHead(page))) { 346 bad_page(page); 347 bad++; 348 } 349 350 __ClearPageHead(page); 351 352 for (i = 1; i < nr_pages; i++) { 353 struct page *p = page + i; 354 355 if (unlikely(!PageTail(p) || (p->first_page != page))) { 356 bad_page(page); 357 bad++; 358 } 359 __ClearPageTail(p); 360 } 361 362 return bad; 363} 364 365static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 366{ 367 int i; 368 369 /* 370 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 371 * and __GFP_HIGHMEM from hard or soft interrupt context. 372 */ 373 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 374 for (i = 0; i < (1 << order); i++) 375 clear_highpage(page + i); 376} 377 378static inline void set_page_order(struct page *page, int order) 379{ 380 set_page_private(page, order); 381 __SetPageBuddy(page); 382} 383 384static inline void rmv_page_order(struct page *page) 385{ 386 __ClearPageBuddy(page); 387 set_page_private(page, 0); 388} 389 390/* 391 * Locate the struct page for both the matching buddy in our 392 * pair (buddy1) and the combined O(n+1) page they form (page). 393 * 394 * 1) Any buddy B1 will have an order O twin B2 which satisfies 395 * the following equation: 396 * B2 = B1 ^ (1 << O) 397 * For example, if the starting buddy (buddy2) is #8 its order 398 * 1 buddy is #10: 399 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 400 * 401 * 2) Any buddy B will have an order O+1 parent P which 402 * satisfies the following equation: 403 * P = B & ~(1 << O) 404 * 405 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 406 */ 407static inline struct page * 408__page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 409{ 410 unsigned long buddy_idx = page_idx ^ (1 << order); 411 412 return page + (buddy_idx - page_idx); 413} 414 415static inline unsigned long 416__find_combined_index(unsigned long page_idx, unsigned int order) 417{ 418 return (page_idx & ~(1 << order)); 419} 420 421/* 422 * This function checks whether a page is free && is the buddy 423 * we can do coalesce a page and its buddy if 424 * (a) the buddy is not in a hole && 425 * (b) the buddy is in the buddy system && 426 * (c) a page and its buddy have the same order && 427 * (d) a page and its buddy are in the same zone. 428 * 429 * For recording whether a page is in the buddy system, we use PG_buddy. 430 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 431 * 432 * For recording page's order, we use page_private(page). 433 */ 434static inline int page_is_buddy(struct page *page, struct page *buddy, 435 int order) 436{ 437 if (!pfn_valid_within(page_to_pfn(buddy))) 438 return 0; 439 440 if (page_zone_id(page) != page_zone_id(buddy)) 441 return 0; 442 443 if (PageBuddy(buddy) && page_order(buddy) == order) { 444 VM_BUG_ON(page_count(buddy) != 0); 445 return 1; 446 } 447 return 0; 448} 449 450/* 451 * Freeing function for a buddy system allocator. 452 * 453 * The concept of a buddy system is to maintain direct-mapped table 454 * (containing bit values) for memory blocks of various "orders". 455 * The bottom level table contains the map for the smallest allocatable 456 * units of memory (here, pages), and each level above it describes 457 * pairs of units from the levels below, hence, "buddies". 458 * At a high level, all that happens here is marking the table entry 459 * at the bottom level available, and propagating the changes upward 460 * as necessary, plus some accounting needed to play nicely with other 461 * parts of the VM system. 462 * At each level, we keep a list of pages, which are heads of continuous 463 * free pages of length of (1 << order) and marked with PG_buddy. Page's 464 * order is recorded in page_private(page) field. 465 * So when we are allocating or freeing one, we can derive the state of the 466 * other. That is, if we allocate a small block, and both were 467 * free, the remainder of the region must be split into blocks. 468 * If a block is freed, and its buddy is also free, then this 469 * triggers coalescing into a block of larger size. 470 * 471 * -- wli 472 */ 473 474static inline void __free_one_page(struct page *page, 475 struct zone *zone, unsigned int order, 476 int migratetype) 477{ 478 unsigned long page_idx; 479 unsigned long combined_idx; 480 struct page *buddy; 481 482 if (unlikely(PageCompound(page))) 483 if (unlikely(destroy_compound_page(page, order))) 484 return; 485 486 VM_BUG_ON(migratetype == -1); 487 488 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 489 490 VM_BUG_ON(page_idx & ((1 << order) - 1)); 491 VM_BUG_ON(bad_range(zone, page)); 492 493 while (order < MAX_ORDER-1) { 494 buddy = __page_find_buddy(page, page_idx, order); 495 if (!page_is_buddy(page, buddy, order)) 496 break; 497 498 /* Our buddy is free, merge with it and move up one order. */ 499 list_del(&buddy->lru); 500 zone->free_area[order].nr_free--; 501 rmv_page_order(buddy); 502 combined_idx = __find_combined_index(page_idx, order); 503 page = page + (combined_idx - page_idx); 504 page_idx = combined_idx; 505 order++; 506 } 507 set_page_order(page, order); 508 509 /* 510 * If this is not the largest possible page, check if the buddy 511 * of the next-highest order is free. If it is, it's possible 512 * that pages are being freed that will coalesce soon. In case, 513 * that is happening, add the free page to the tail of the list 514 * so it's less likely to be used soon and more likely to be merged 515 * as a higher order page 516 */ 517 if ((order < MAX_ORDER-1) && pfn_valid_within(page_to_pfn(buddy))) { 518 struct page *higher_page, *higher_buddy; 519 combined_idx = __find_combined_index(page_idx, order); 520 higher_page = page + combined_idx - page_idx; 521 higher_buddy = __page_find_buddy(higher_page, combined_idx, order + 1); 522 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 523 list_add_tail(&page->lru, 524 &zone->free_area[order].free_list[migratetype]); 525 goto out; 526 } 527 } 528 529 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 530out: 531 zone->free_area[order].nr_free++; 532} 533 534/* 535 * free_page_mlock() -- clean up attempts to free and mlocked() page. 536 * Page should not be on lru, so no need to fix that up. 537 * free_pages_check() will verify... 538 */ 539static inline void free_page_mlock(struct page *page) 540{ 541 __dec_zone_page_state(page, NR_MLOCK); 542 __count_vm_event(UNEVICTABLE_MLOCKFREED); 543} 544 545static inline int free_pages_check(struct page *page) 546{ 547 if (unlikely(page_mapcount(page) | 548 (page->mapping != NULL) | 549 (atomic_read(&page->_count) != 0) | 550 (page->flags & PAGE_FLAGS_CHECK_AT_FREE))) { 551 bad_page(page); 552 return 1; 553 } 554 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 555 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 556 return 0; 557} 558 559/* 560 * Frees a number of pages from the PCP lists 561 * Assumes all pages on list are in same zone, and of same order. 562 * count is the number of pages to free. 563 * 564 * If the zone was previously in an "all pages pinned" state then look to 565 * see if this freeing clears that state. 566 * 567 * And clear the zone's pages_scanned counter, to hold off the "all pages are 568 * pinned" detection logic. 569 */ 570static void free_pcppages_bulk(struct zone *zone, int count, 571 struct per_cpu_pages *pcp) 572{ 573 int migratetype = 0; 574 int batch_free = 0; 575 576 spin_lock(&zone->lock); 577 zone->all_unreclaimable = 0; 578 zone->pages_scanned = 0; 579 580 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 581 while (count) { 582 struct page *page; 583 struct list_head *list; 584 585 /* 586 * Remove pages from lists in a round-robin fashion. A 587 * batch_free count is maintained that is incremented when an 588 * empty list is encountered. This is so more pages are freed 589 * off fuller lists instead of spinning excessively around empty 590 * lists 591 */ 592 do { 593 batch_free++; 594 if (++migratetype == MIGRATE_PCPTYPES) 595 migratetype = 0; 596 list = &pcp->lists[migratetype]; 597 } while (list_empty(list)); 598 599 do { 600 page = list_entry(list->prev, struct page, lru); 601 /* must delete as __free_one_page list manipulates */ 602 list_del(&page->lru); 603 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 604 __free_one_page(page, zone, 0, page_private(page)); 605 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 606 } while (--count && --batch_free && !list_empty(list)); 607 } 608 spin_unlock(&zone->lock); 609} 610 611static void free_one_page(struct zone *zone, struct page *page, int order, 612 int migratetype) 613{ 614 spin_lock(&zone->lock); 615 zone->all_unreclaimable = 0; 616 zone->pages_scanned = 0; 617 618 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 619 __free_one_page(page, zone, order, migratetype); 620 spin_unlock(&zone->lock); 621} 622 623static bool free_pages_prepare(struct page *page, unsigned int order) 624{ 625 int i; 626 int bad = 0; 627 628 trace_mm_page_free_direct(page, order); 629 kmemcheck_free_shadow(page, order); 630 631 for (i = 0; i < (1 << order); i++) { 632 struct page *pg = page + i; 633 634 if (PageAnon(pg)) 635 pg->mapping = NULL; 636 bad += free_pages_check(pg); 637 } 638 if (bad) 639 return false; 640 641 if (!PageHighMem(page)) { 642 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 643 debug_check_no_obj_freed(page_address(page), 644 PAGE_SIZE << order); 645 } 646 arch_free_page(page, order); 647 kernel_map_pages(page, 1 << order, 0); 648 649 return true; 650} 651 652static void __free_pages_ok(struct page *page, unsigned int order) 653{ 654 unsigned long flags; 655 int wasMlocked = __TestClearPageMlocked(page); 656 657 if (!free_pages_prepare(page, order)) 658 return; 659 660 local_irq_save(flags); 661 if (unlikely(wasMlocked)) 662 free_page_mlock(page); 663 __count_vm_events(PGFREE, 1 << order); 664 free_one_page(page_zone(page), page, order, 665 get_pageblock_migratetype(page)); 666 local_irq_restore(flags); 667} 668 669/* 670 * permit the bootmem allocator to evade page validation on high-order frees 671 */ 672void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 673{ 674 if (order == 0) { 675 __ClearPageReserved(page); 676 set_page_count(page, 0); 677 set_page_refcounted(page); 678 __free_page(page); 679 } else { 680 int loop; 681 682 prefetchw(page); 683 for (loop = 0; loop < BITS_PER_LONG; loop++) { 684 struct page *p = &page[loop]; 685 686 if (loop + 1 < BITS_PER_LONG) 687 prefetchw(p + 1); 688 __ClearPageReserved(p); 689 set_page_count(p, 0); 690 } 691 692 set_page_refcounted(page); 693 __free_pages(page, order); 694 } 695} 696 697 698/* 699 * The order of subdivision here is critical for the IO subsystem. 700 * Please do not alter this order without good reasons and regression 701 * testing. Specifically, as large blocks of memory are subdivided, 702 * the order in which smaller blocks are delivered depends on the order 703 * they're subdivided in this function. This is the primary factor 704 * influencing the order in which pages are delivered to the IO 705 * subsystem according to empirical testing, and this is also justified 706 * by considering the behavior of a buddy system containing a single 707 * large block of memory acted on by a series of small allocations. 708 * This behavior is a critical factor in sglist merging's success. 709 * 710 * -- wli 711 */ 712static inline void expand(struct zone *zone, struct page *page, 713 int low, int high, struct free_area *area, 714 int migratetype) 715{ 716 unsigned long size = 1 << high; 717 718 while (high > low) { 719 area--; 720 high--; 721 size >>= 1; 722 VM_BUG_ON(bad_range(zone, &page[size])); 723 list_add(&page[size].lru, &area->free_list[migratetype]); 724 area->nr_free++; 725 set_page_order(&page[size], high); 726 } 727} 728 729/* 730 * This page is about to be returned from the page allocator 731 */ 732static inline int check_new_page(struct page *page) 733{ 734 if (unlikely(page_mapcount(page) | 735 (page->mapping != NULL) | 736 (atomic_read(&page->_count) != 0) | 737 (page->flags & PAGE_FLAGS_CHECK_AT_PREP))) { 738 bad_page(page); 739 return 1; 740 } 741 return 0; 742} 743 744static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 745{ 746 int i; 747 748 for (i = 0; i < (1 << order); i++) { 749 struct page *p = page + i; 750 if (unlikely(check_new_page(p))) 751 return 1; 752 } 753 754 set_page_private(page, 0); 755 set_page_refcounted(page); 756 757 arch_alloc_page(page, order); 758 kernel_map_pages(page, 1 << order, 1); 759 760 if (gfp_flags & __GFP_ZERO) 761 prep_zero_page(page, order, gfp_flags); 762 763 if (order && (gfp_flags & __GFP_COMP)) 764 prep_compound_page(page, order); 765 766 return 0; 767} 768 769/* 770 * Go through the free lists for the given migratetype and remove 771 * the smallest available page from the freelists 772 */ 773static inline 774struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 775 int migratetype) 776{ 777 unsigned int current_order; 778 struct free_area * area; 779 struct page *page; 780 781 /* Find a page of the appropriate size in the preferred list */ 782 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 783 area = &(zone->free_area[current_order]); 784 if (list_empty(&area->free_list[migratetype])) 785 continue; 786 787 page = list_entry(area->free_list[migratetype].next, 788 struct page, lru); 789 list_del(&page->lru); 790 rmv_page_order(page); 791 area->nr_free--; 792 expand(zone, page, order, current_order, area, migratetype); 793 return page; 794 } 795 796 return NULL; 797} 798 799 800/* 801 * This array describes the order lists are fallen back to when 802 * the free lists for the desirable migrate type are depleted 803 */ 804static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 805 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 806 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 807 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 808 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 809}; 810 811/* 812 * Move the free pages in a range to the free lists of the requested type. 813 * Note that start_page and end_pages are not aligned on a pageblock 814 * boundary. If alignment is required, use move_freepages_block() 815 */ 816static int move_freepages(struct zone *zone, 817 struct page *start_page, struct page *end_page, 818 int migratetype) 819{ 820 struct page *page; 821 unsigned long order; 822 int pages_moved = 0; 823 824#ifndef CONFIG_HOLES_IN_ZONE 825 /* 826 * page_zone is not safe to call in this context when 827 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 828 * anyway as we check zone boundaries in move_freepages_block(). 829 * Remove at a later date when no bug reports exist related to 830 * grouping pages by mobility 831 */ 832 BUG_ON(page_zone(start_page) != page_zone(end_page)); 833#endif 834 835 for (page = start_page; page <= end_page;) { 836 /* Make sure we are not inadvertently changing nodes */ 837 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 838 839 if (!pfn_valid_within(page_to_pfn(page))) { 840 page++; 841 continue; 842 } 843 844 if (!PageBuddy(page)) { 845 page++; 846 continue; 847 } 848 849 order = page_order(page); 850 list_del(&page->lru); 851 list_add(&page->lru, 852 &zone->free_area[order].free_list[migratetype]); 853 page += 1 << order; 854 pages_moved += 1 << order; 855 } 856 857 return pages_moved; 858} 859 860static int move_freepages_block(struct zone *zone, struct page *page, 861 int migratetype) 862{ 863 unsigned long start_pfn, end_pfn; 864 struct page *start_page, *end_page; 865 866 start_pfn = page_to_pfn(page); 867 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 868 start_page = pfn_to_page(start_pfn); 869 end_page = start_page + pageblock_nr_pages - 1; 870 end_pfn = start_pfn + pageblock_nr_pages - 1; 871 872 /* Do not cross zone boundaries */ 873 if (start_pfn < zone->zone_start_pfn) 874 start_page = page; 875 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 876 return 0; 877 878 return move_freepages(zone, start_page, end_page, migratetype); 879} 880 881static void change_pageblock_range(struct page *pageblock_page, 882 int start_order, int migratetype) 883{ 884 int nr_pageblocks = 1 << (start_order - pageblock_order); 885 886 while (nr_pageblocks--) { 887 set_pageblock_migratetype(pageblock_page, migratetype); 888 pageblock_page += pageblock_nr_pages; 889 } 890} 891 892/* Remove an element from the buddy allocator from the fallback list */ 893static inline struct page * 894__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 895{ 896 struct free_area * area; 897 int current_order; 898 struct page *page; 899 int migratetype, i; 900 901 /* Find the largest possible block of pages in the other list */ 902 for (current_order = MAX_ORDER-1; current_order >= order; 903 --current_order) { 904 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 905 migratetype = fallbacks[start_migratetype][i]; 906 907 /* MIGRATE_RESERVE handled later if necessary */ 908 if (migratetype == MIGRATE_RESERVE) 909 continue; 910 911 area = &(zone->free_area[current_order]); 912 if (list_empty(&area->free_list[migratetype])) 913 continue; 914 915 page = list_entry(area->free_list[migratetype].next, 916 struct page, lru); 917 area->nr_free--; 918 919 /* 920 * If breaking a large block of pages, move all free 921 * pages to the preferred allocation list. If falling 922 * back for a reclaimable kernel allocation, be more 923 * agressive about taking ownership of free pages 924 */ 925 if (unlikely(current_order >= (pageblock_order >> 1)) || 926 start_migratetype == MIGRATE_RECLAIMABLE || 927 page_group_by_mobility_disabled) { 928 unsigned long pages; 929 pages = move_freepages_block(zone, page, 930 start_migratetype); 931 932 /* Claim the whole block if over half of it is free */ 933 if (pages >= (1 << (pageblock_order-1)) || 934 page_group_by_mobility_disabled) 935 set_pageblock_migratetype(page, 936 start_migratetype); 937 938 migratetype = start_migratetype; 939 } 940 941 /* Remove the page from the freelists */ 942 list_del(&page->lru); 943 rmv_page_order(page); 944 945 /* Take ownership for orders >= pageblock_order */ 946 if (current_order >= pageblock_order) 947 change_pageblock_range(page, current_order, 948 start_migratetype); 949 950 expand(zone, page, order, current_order, area, migratetype); 951 952 trace_mm_page_alloc_extfrag(page, order, current_order, 953 start_migratetype, migratetype); 954 955 return page; 956 } 957 } 958 959 return NULL; 960} 961 962/* 963 * Do the hard work of removing an element from the buddy allocator. 964 * Call me with the zone->lock already held. 965 */ 966static struct page *__rmqueue(struct zone *zone, unsigned int order, 967 int migratetype) 968{ 969 struct page *page; 970 971retry_reserve: 972 page = __rmqueue_smallest(zone, order, migratetype); 973 974 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 975 page = __rmqueue_fallback(zone, order, migratetype); 976 977 /* 978 * Use MIGRATE_RESERVE rather than fail an allocation. goto 979 * is used because __rmqueue_smallest is an inline function 980 * and we want just one call site 981 */ 982 if (!page) { 983 migratetype = MIGRATE_RESERVE; 984 goto retry_reserve; 985 } 986 } 987 988 trace_mm_page_alloc_zone_locked(page, order, migratetype); 989 return page; 990} 991 992/* 993 * Obtain a specified number of elements from the buddy allocator, all under 994 * a single hold of the lock, for efficiency. Add them to the supplied list. 995 * Returns the number of new pages which were placed at *list. 996 */ 997static int rmqueue_bulk(struct zone *zone, unsigned int order, 998 unsigned long count, struct list_head *list, 999 int migratetype, int cold) 1000{ 1001 int i; 1002 1003 spin_lock(&zone->lock); 1004 for (i = 0; i < count; ++i) { 1005 struct page *page = __rmqueue(zone, order, migratetype); 1006 if (unlikely(page == NULL)) 1007 break; 1008 1009 /* 1010 * Split buddy pages returned by expand() are received here 1011 * in physical page order. The page is added to the callers and 1012 * list and the list head then moves forward. From the callers 1013 * perspective, the linked list is ordered by page number in 1014 * some conditions. This is useful for IO devices that can 1015 * merge IO requests if the physical pages are ordered 1016 * properly. 1017 */ 1018 if (likely(cold == 0)) 1019 list_add(&page->lru, list); 1020 else 1021 list_add_tail(&page->lru, list); 1022 set_page_private(page, migratetype); 1023 list = &page->lru; 1024 } 1025 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1026 spin_unlock(&zone->lock); 1027 return i; 1028} 1029 1030#ifdef CONFIG_NUMA 1031/* 1032 * Called from the vmstat counter updater to drain pagesets of this 1033 * currently executing processor on remote nodes after they have 1034 * expired. 1035 * 1036 * Note that this function must be called with the thread pinned to 1037 * a single processor. 1038 */ 1039void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1040{ 1041 unsigned long flags; 1042 int to_drain; 1043 1044 local_irq_save(flags); 1045 if (pcp->count >= pcp->batch) 1046 to_drain = pcp->batch; 1047 else 1048 to_drain = pcp->count; 1049 free_pcppages_bulk(zone, to_drain, pcp); 1050 pcp->count -= to_drain; 1051 local_irq_restore(flags); 1052} 1053#endif 1054 1055/* 1056 * Drain pages of the indicated processor. 1057 * 1058 * The processor must either be the current processor and the 1059 * thread pinned to the current processor or a processor that 1060 * is not online. 1061 */ 1062static void drain_pages(unsigned int cpu) 1063{ 1064 unsigned long flags; 1065 struct zone *zone; 1066 1067 for_each_populated_zone(zone) { 1068 struct per_cpu_pageset *pset; 1069 struct per_cpu_pages *pcp; 1070 1071 local_irq_save(flags); 1072 pset = per_cpu_ptr(zone->pageset, cpu); 1073 1074 pcp = &pset->pcp; 1075 free_pcppages_bulk(zone, pcp->count, pcp); 1076 pcp->count = 0; 1077 local_irq_restore(flags); 1078 } 1079} 1080 1081/* 1082 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1083 */ 1084void drain_local_pages(void *arg) 1085{ 1086 drain_pages(smp_processor_id()); 1087} 1088 1089/* 1090 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1091 */ 1092void drain_all_pages(void) 1093{ 1094 on_each_cpu(drain_local_pages, NULL, 1); 1095} 1096 1097#ifdef CONFIG_HIBERNATION 1098 1099void mark_free_pages(struct zone *zone) 1100{ 1101 unsigned long pfn, max_zone_pfn; 1102 unsigned long flags; 1103 int order, t; 1104 struct list_head *curr; 1105 1106 if (!zone->spanned_pages) 1107 return; 1108 1109 spin_lock_irqsave(&zone->lock, flags); 1110 1111 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1112 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1113 if (pfn_valid(pfn)) { 1114 struct page *page = pfn_to_page(pfn); 1115 1116 if (!swsusp_page_is_forbidden(page)) 1117 swsusp_unset_page_free(page); 1118 } 1119 1120 for_each_migratetype_order(order, t) { 1121 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1122 unsigned long i; 1123 1124 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1125 for (i = 0; i < (1UL << order); i++) 1126 swsusp_set_page_free(pfn_to_page(pfn + i)); 1127 } 1128 } 1129 spin_unlock_irqrestore(&zone->lock, flags); 1130} 1131#endif /* CONFIG_PM */ 1132 1133/* 1134 * Free a 0-order page 1135 * cold == 1 ? free a cold page : free a hot page 1136 */ 1137void free_hot_cold_page(struct page *page, int cold) 1138{ 1139 struct zone *zone = page_zone(page); 1140 struct per_cpu_pages *pcp; 1141 unsigned long flags; 1142 int migratetype; 1143 int wasMlocked = __TestClearPageMlocked(page); 1144 1145 if (!free_pages_prepare(page, 0)) 1146 return; 1147 1148 migratetype = get_pageblock_migratetype(page); 1149 set_page_private(page, migratetype); 1150 local_irq_save(flags); 1151 if (unlikely(wasMlocked)) 1152 free_page_mlock(page); 1153 __count_vm_event(PGFREE); 1154 1155 /* 1156 * We only track unmovable, reclaimable and movable on pcp lists. 1157 * Free ISOLATE pages back to the allocator because they are being 1158 * offlined but treat RESERVE as movable pages so we can get those 1159 * areas back if necessary. Otherwise, we may have to free 1160 * excessively into the page allocator 1161 */ 1162 if (migratetype >= MIGRATE_PCPTYPES) { 1163 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1164 free_one_page(zone, page, 0, migratetype); 1165 goto out; 1166 } 1167 migratetype = MIGRATE_MOVABLE; 1168 } 1169 1170 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1171 if (cold) 1172 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1173 else 1174 list_add(&page->lru, &pcp->lists[migratetype]); 1175 pcp->count++; 1176 if (pcp->count >= pcp->high) { 1177 free_pcppages_bulk(zone, pcp->batch, pcp); 1178 pcp->count -= pcp->batch; 1179 } 1180 1181out: 1182 local_irq_restore(flags); 1183} 1184 1185/* 1186 * split_page takes a non-compound higher-order page, and splits it into 1187 * n (1<<order) sub-pages: page[0..n] 1188 * Each sub-page must be freed individually. 1189 * 1190 * Note: this is probably too low level an operation for use in drivers. 1191 * Please consult with lkml before using this in your driver. 1192 */ 1193void split_page(struct page *page, unsigned int order) 1194{ 1195 int i; 1196 1197 VM_BUG_ON(PageCompound(page)); 1198 VM_BUG_ON(!page_count(page)); 1199 1200#ifdef CONFIG_KMEMCHECK 1201 /* 1202 * Split shadow pages too, because free(page[0]) would 1203 * otherwise free the whole shadow. 1204 */ 1205 if (kmemcheck_page_is_tracked(page)) 1206 split_page(virt_to_page(page[0].shadow), order); 1207#endif 1208 1209 for (i = 1; i < (1 << order); i++) 1210 set_page_refcounted(page + i); 1211} 1212 1213/* 1214 * Similar to split_page except the page is already free. As this is only 1215 * being used for migration, the migratetype of the block also changes. 1216 * As this is called with interrupts disabled, the caller is responsible 1217 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1218 * are enabled. 1219 * 1220 * Note: this is probably too low level an operation for use in drivers. 1221 * Please consult with lkml before using this in your driver. 1222 */ 1223int split_free_page(struct page *page) 1224{ 1225 unsigned int order; 1226 unsigned long watermark; 1227 struct zone *zone; 1228 1229 BUG_ON(!PageBuddy(page)); 1230 1231 zone = page_zone(page); 1232 order = page_order(page); 1233 1234 /* Obey watermarks as if the page was being allocated */ 1235 watermark = low_wmark_pages(zone) + (1 << order); 1236 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1237 return 0; 1238 1239 /* Remove page from free list */ 1240 list_del(&page->lru); 1241 zone->free_area[order].nr_free--; 1242 rmv_page_order(page); 1243 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1244 1245 /* Split into individual pages */ 1246 set_page_refcounted(page); 1247 split_page(page, order); 1248 1249 if (order >= pageblock_order - 1) { 1250 struct page *endpage = page + (1 << order) - 1; 1251 for (; page < endpage; page += pageblock_nr_pages) 1252 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1253 } 1254 1255 return 1 << order; 1256} 1257 1258/* 1259 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1260 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1261 * or two. 1262 */ 1263static inline 1264struct page *buffered_rmqueue(struct zone *preferred_zone, 1265 struct zone *zone, int order, gfp_t gfp_flags, 1266 int migratetype) 1267{ 1268 unsigned long flags; 1269 struct page *page; 1270 int cold = !!(gfp_flags & __GFP_COLD); 1271 1272again: 1273 if (likely(order == 0)) { 1274 struct per_cpu_pages *pcp; 1275 struct list_head *list; 1276 1277 local_irq_save(flags); 1278 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1279 list = &pcp->lists[migratetype]; 1280 if (list_empty(list)) { 1281 pcp->count += rmqueue_bulk(zone, 0, 1282 pcp->batch, list, 1283 migratetype, cold); 1284 if (unlikely(list_empty(list))) 1285 goto failed; 1286 } 1287 1288 if (cold) 1289 page = list_entry(list->prev, struct page, lru); 1290 else 1291 page = list_entry(list->next, struct page, lru); 1292 1293 list_del(&page->lru); 1294 pcp->count--; 1295 } else { 1296 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1297 /* 1298 * __GFP_NOFAIL is not to be used in new code. 1299 * 1300 * All __GFP_NOFAIL callers should be fixed so that they 1301 * properly detect and handle allocation failures. 1302 * 1303 * We most definitely don't want callers attempting to 1304 * allocate greater than order-1 page units with 1305 * __GFP_NOFAIL. 1306 */ 1307 WARN_ON_ONCE(order > 1); 1308 } 1309 spin_lock_irqsave(&zone->lock, flags); 1310 page = __rmqueue(zone, order, migratetype); 1311 spin_unlock(&zone->lock); 1312 if (!page) 1313 goto failed; 1314 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1315 } 1316 1317 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1318 zone_statistics(preferred_zone, zone); 1319 local_irq_restore(flags); 1320 1321 VM_BUG_ON(bad_range(zone, page)); 1322 if (prep_new_page(page, order, gfp_flags)) 1323 goto again; 1324 return page; 1325 1326failed: 1327 local_irq_restore(flags); 1328 return NULL; 1329} 1330 1331/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1332#define ALLOC_WMARK_MIN WMARK_MIN 1333#define ALLOC_WMARK_LOW WMARK_LOW 1334#define ALLOC_WMARK_HIGH WMARK_HIGH 1335#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1336 1337/* Mask to get the watermark bits */ 1338#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1339 1340#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1341#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1342#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1343 1344#ifdef CONFIG_FAIL_PAGE_ALLOC 1345 1346static struct fail_page_alloc_attr { 1347 struct fault_attr attr; 1348 1349 u32 ignore_gfp_highmem; 1350 u32 ignore_gfp_wait; 1351 u32 min_order; 1352 1353#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1354 1355 struct dentry *ignore_gfp_highmem_file; 1356 struct dentry *ignore_gfp_wait_file; 1357 struct dentry *min_order_file; 1358 1359#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1360 1361} fail_page_alloc = { 1362 .attr = FAULT_ATTR_INITIALIZER, 1363 .ignore_gfp_wait = 1, 1364 .ignore_gfp_highmem = 1, 1365 .min_order = 1, 1366}; 1367 1368static int __init setup_fail_page_alloc(char *str) 1369{ 1370 return setup_fault_attr(&fail_page_alloc.attr, str); 1371} 1372__setup("fail_page_alloc=", setup_fail_page_alloc); 1373 1374static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1375{ 1376 if (order < fail_page_alloc.min_order) 1377 return 0; 1378 if (gfp_mask & __GFP_NOFAIL) 1379 return 0; 1380 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1381 return 0; 1382 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1383 return 0; 1384 1385 return should_fail(&fail_page_alloc.attr, 1 << order); 1386} 1387 1388#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1389 1390static int __init fail_page_alloc_debugfs(void) 1391{ 1392 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1393 struct dentry *dir; 1394 int err; 1395 1396 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1397 "fail_page_alloc"); 1398 if (err) 1399 return err; 1400 dir = fail_page_alloc.attr.dentries.dir; 1401 1402 fail_page_alloc.ignore_gfp_wait_file = 1403 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1404 &fail_page_alloc.ignore_gfp_wait); 1405 1406 fail_page_alloc.ignore_gfp_highmem_file = 1407 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1408 &fail_page_alloc.ignore_gfp_highmem); 1409 fail_page_alloc.min_order_file = 1410 debugfs_create_u32("min-order", mode, dir, 1411 &fail_page_alloc.min_order); 1412 1413 if (!fail_page_alloc.ignore_gfp_wait_file || 1414 !fail_page_alloc.ignore_gfp_highmem_file || 1415 !fail_page_alloc.min_order_file) { 1416 err = -ENOMEM; 1417 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1418 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1419 debugfs_remove(fail_page_alloc.min_order_file); 1420 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1421 } 1422 1423 return err; 1424} 1425 1426late_initcall(fail_page_alloc_debugfs); 1427 1428#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1429 1430#else /* CONFIG_FAIL_PAGE_ALLOC */ 1431 1432static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1433{ 1434 return 0; 1435} 1436 1437#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1438 1439/* 1440 * Return 1 if free pages are above 'mark'. This takes into account the order 1441 * of the allocation. 1442 */ 1443int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1444 int classzone_idx, int alloc_flags) 1445{ 1446 /* free_pages my go negative - that's OK */ 1447 long min = mark; 1448 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1449 int o; 1450 1451 if (alloc_flags & ALLOC_HIGH) 1452 min -= min / 2; 1453 if (alloc_flags & ALLOC_HARDER) 1454 min -= min / 4; 1455 1456 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1457 return 0; 1458 for (o = 0; o < order; o++) { 1459 /* At the next order, this order's pages become unavailable */ 1460 free_pages -= z->free_area[o].nr_free << o; 1461 1462 /* Require fewer higher order pages to be free */ 1463 min >>= 1; 1464 1465 if (free_pages <= min) 1466 return 0; 1467 } 1468 return 1; 1469} 1470 1471#ifdef CONFIG_NUMA 1472/* 1473 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1474 * skip over zones that are not allowed by the cpuset, or that have 1475 * been recently (in last second) found to be nearly full. See further 1476 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1477 * that have to skip over a lot of full or unallowed zones. 1478 * 1479 * If the zonelist cache is present in the passed in zonelist, then 1480 * returns a pointer to the allowed node mask (either the current 1481 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1482 * 1483 * If the zonelist cache is not available for this zonelist, does 1484 * nothing and returns NULL. 1485 * 1486 * If the fullzones BITMAP in the zonelist cache is stale (more than 1487 * a second since last zap'd) then we zap it out (clear its bits.) 1488 * 1489 * We hold off even calling zlc_setup, until after we've checked the 1490 * first zone in the zonelist, on the theory that most allocations will 1491 * be satisfied from that first zone, so best to examine that zone as 1492 * quickly as we can. 1493 */ 1494static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1495{ 1496 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1497 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1498 1499 zlc = zonelist->zlcache_ptr; 1500 if (!zlc) 1501 return NULL; 1502 1503 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1504 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1505 zlc->last_full_zap = jiffies; 1506 } 1507 1508 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1509 &cpuset_current_mems_allowed : 1510 &node_states[N_HIGH_MEMORY]; 1511 return allowednodes; 1512} 1513 1514/* 1515 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1516 * if it is worth looking at further for free memory: 1517 * 1) Check that the zone isn't thought to be full (doesn't have its 1518 * bit set in the zonelist_cache fullzones BITMAP). 1519 * 2) Check that the zones node (obtained from the zonelist_cache 1520 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1521 * Return true (non-zero) if zone is worth looking at further, or 1522 * else return false (zero) if it is not. 1523 * 1524 * This check -ignores- the distinction between various watermarks, 1525 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1526 * found to be full for any variation of these watermarks, it will 1527 * be considered full for up to one second by all requests, unless 1528 * we are so low on memory on all allowed nodes that we are forced 1529 * into the second scan of the zonelist. 1530 * 1531 * In the second scan we ignore this zonelist cache and exactly 1532 * apply the watermarks to all zones, even it is slower to do so. 1533 * We are low on memory in the second scan, and should leave no stone 1534 * unturned looking for a free page. 1535 */ 1536static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1537 nodemask_t *allowednodes) 1538{ 1539 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1540 int i; /* index of *z in zonelist zones */ 1541 int n; /* node that zone *z is on */ 1542 1543 zlc = zonelist->zlcache_ptr; 1544 if (!zlc) 1545 return 1; 1546 1547 i = z - zonelist->_zonerefs; 1548 n = zlc->z_to_n[i]; 1549 1550 /* This zone is worth trying if it is allowed but not full */ 1551 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1552} 1553 1554/* 1555 * Given 'z' scanning a zonelist, set the corresponding bit in 1556 * zlc->fullzones, so that subsequent attempts to allocate a page 1557 * from that zone don't waste time re-examining it. 1558 */ 1559static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1560{ 1561 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1562 int i; /* index of *z in zonelist zones */ 1563 1564 zlc = zonelist->zlcache_ptr; 1565 if (!zlc) 1566 return; 1567 1568 i = z - zonelist->_zonerefs; 1569 1570 set_bit(i, zlc->fullzones); 1571} 1572 1573#else /* CONFIG_NUMA */ 1574 1575static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1576{ 1577 return NULL; 1578} 1579 1580static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1581 nodemask_t *allowednodes) 1582{ 1583 return 1; 1584} 1585 1586static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1587{ 1588} 1589#endif /* CONFIG_NUMA */ 1590 1591/* 1592 * get_page_from_freelist goes through the zonelist trying to allocate 1593 * a page. 1594 */ 1595static struct page * 1596get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1597 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1598 struct zone *preferred_zone, int migratetype) 1599{ 1600 struct zoneref *z; 1601 struct page *page = NULL; 1602 int classzone_idx; 1603 struct zone *zone; 1604 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1605 int zlc_active = 0; /* set if using zonelist_cache */ 1606 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1607 1608 classzone_idx = zone_idx(preferred_zone); 1609zonelist_scan: 1610 /* 1611 * Scan zonelist, looking for a zone with enough free. 1612 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1613 */ 1614 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1615 high_zoneidx, nodemask) { 1616 if (NUMA_BUILD && zlc_active && 1617 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1618 continue; 1619 if ((alloc_flags & ALLOC_CPUSET) && 1620 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1621 goto try_next_zone; 1622 1623 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1624 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1625 unsigned long mark; 1626 int ret; 1627 1628 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1629 if (zone_watermark_ok(zone, order, mark, 1630 classzone_idx, alloc_flags)) 1631 goto try_this_zone; 1632 1633 if (zone_reclaim_mode == 0) 1634 goto this_zone_full; 1635 1636 ret = zone_reclaim(zone, gfp_mask, order); 1637 switch (ret) { 1638 case ZONE_RECLAIM_NOSCAN: 1639 /* did not scan */ 1640 goto try_next_zone; 1641 case ZONE_RECLAIM_FULL: 1642 /* scanned but unreclaimable */ 1643 goto this_zone_full; 1644 default: 1645 /* did we reclaim enough */ 1646 if (!zone_watermark_ok(zone, order, mark, 1647 classzone_idx, alloc_flags)) 1648 goto this_zone_full; 1649 } 1650 } 1651 1652try_this_zone: 1653 page = buffered_rmqueue(preferred_zone, zone, order, 1654 gfp_mask, migratetype); 1655 if (page) 1656 break; 1657this_zone_full: 1658 if (NUMA_BUILD) 1659 zlc_mark_zone_full(zonelist, z); 1660try_next_zone: 1661 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1662 /* 1663 * we do zlc_setup after the first zone is tried but only 1664 * if there are multiple nodes make it worthwhile 1665 */ 1666 allowednodes = zlc_setup(zonelist, alloc_flags); 1667 zlc_active = 1; 1668 did_zlc_setup = 1; 1669 } 1670 } 1671 1672 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1673 /* Disable zlc cache for second zonelist scan */ 1674 zlc_active = 0; 1675 goto zonelist_scan; 1676 } 1677 return page; 1678} 1679 1680static inline int 1681should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1682 unsigned long pages_reclaimed) 1683{ 1684 /* Do not loop if specifically requested */ 1685 if (gfp_mask & __GFP_NORETRY) 1686 return 0; 1687 1688 /* 1689 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1690 * means __GFP_NOFAIL, but that may not be true in other 1691 * implementations. 1692 */ 1693 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1694 return 1; 1695 1696 /* 1697 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1698 * specified, then we retry until we no longer reclaim any pages 1699 * (above), or we've reclaimed an order of pages at least as 1700 * large as the allocation's order. In both cases, if the 1701 * allocation still fails, we stop retrying. 1702 */ 1703 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1704 return 1; 1705 1706 /* 1707 * Don't let big-order allocations loop unless the caller 1708 * explicitly requests that. 1709 */ 1710 if (gfp_mask & __GFP_NOFAIL) 1711 return 1; 1712 1713 return 0; 1714} 1715 1716static inline struct page * 1717__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1718 struct zonelist *zonelist, enum zone_type high_zoneidx, 1719 nodemask_t *nodemask, struct zone *preferred_zone, 1720 int migratetype) 1721{ 1722 struct page *page; 1723 1724 /* Acquire the OOM killer lock for the zones in zonelist */ 1725 if (!try_set_zone_oom(zonelist, gfp_mask)) { 1726 schedule_timeout_uninterruptible(1); 1727 return NULL; 1728 } 1729 1730 /* 1731 * Go through the zonelist yet one more time, keep very high watermark 1732 * here, this is only to catch a parallel oom killing, we must fail if 1733 * we're still under heavy pressure. 1734 */ 1735 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1736 order, zonelist, high_zoneidx, 1737 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1738 preferred_zone, migratetype); 1739 if (page) 1740 goto out; 1741 1742 if (!(gfp_mask & __GFP_NOFAIL)) { 1743 /* The OOM killer will not help higher order allocs */ 1744 if (order > PAGE_ALLOC_COSTLY_ORDER) 1745 goto out; 1746 /* 1747 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1748 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1749 * The caller should handle page allocation failure by itself if 1750 * it specifies __GFP_THISNODE. 1751 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1752 */ 1753 if (gfp_mask & __GFP_THISNODE) 1754 goto out; 1755 } 1756 /* Exhausted what can be done so it's blamo time */ 1757 out_of_memory(zonelist, gfp_mask, order, nodemask); 1758 1759out: 1760 clear_zonelist_oom(zonelist, gfp_mask); 1761 return page; 1762} 1763 1764#ifdef CONFIG_COMPACTION 1765/* Try memory compaction for high-order allocations before reclaim */ 1766static struct page * 1767__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1768 struct zonelist *zonelist, enum zone_type high_zoneidx, 1769 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1770 int migratetype, unsigned long *did_some_progress) 1771{ 1772 struct page *page; 1773 1774 if (!order || compaction_deferred(preferred_zone)) 1775 return NULL; 1776 1777 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 1778 nodemask); 1779 if (*did_some_progress != COMPACT_SKIPPED) { 1780 1781 /* Page migration frees to the PCP lists but we want merging */ 1782 drain_pages(get_cpu()); 1783 put_cpu(); 1784 1785 page = get_page_from_freelist(gfp_mask, nodemask, 1786 order, zonelist, high_zoneidx, 1787 alloc_flags, preferred_zone, 1788 migratetype); 1789 if (page) { 1790 preferred_zone->compact_considered = 0; 1791 preferred_zone->compact_defer_shift = 0; 1792 count_vm_event(COMPACTSUCCESS); 1793 return page; 1794 } 1795 1796 /* 1797 * It's bad if compaction run occurs and fails. 1798 * The most likely reason is that pages exist, 1799 * but not enough to satisfy watermarks. 1800 */ 1801 count_vm_event(COMPACTFAIL); 1802 defer_compaction(preferred_zone); 1803 1804 cond_resched(); 1805 } 1806 1807 return NULL; 1808} 1809#else 1810static inline struct page * 1811__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1812 struct zonelist *zonelist, enum zone_type high_zoneidx, 1813 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1814 int migratetype, unsigned long *did_some_progress) 1815{ 1816 return NULL; 1817} 1818#endif /* CONFIG_COMPACTION */ 1819 1820/* The really slow allocator path where we enter direct reclaim */ 1821static inline struct page * 1822__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1823 struct zonelist *zonelist, enum zone_type high_zoneidx, 1824 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1825 int migratetype, unsigned long *did_some_progress) 1826{ 1827 struct page *page = NULL; 1828 struct reclaim_state reclaim_state; 1829 struct task_struct *p = current; 1830 1831 cond_resched(); 1832 1833 /* We now go into synchronous reclaim */ 1834 cpuset_memory_pressure_bump(); 1835 p->flags |= PF_MEMALLOC; 1836 lockdep_set_current_reclaim_state(gfp_mask); 1837 reclaim_state.reclaimed_slab = 0; 1838 p->reclaim_state = &reclaim_state; 1839 1840 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1841 1842 p->reclaim_state = NULL; 1843 lockdep_clear_current_reclaim_state(); 1844 p->flags &= ~PF_MEMALLOC; 1845 1846 cond_resched(); 1847 1848 if (order != 0) 1849 drain_all_pages(); 1850 1851 if (likely(*did_some_progress)) 1852 page = get_page_from_freelist(gfp_mask, nodemask, order, 1853 zonelist, high_zoneidx, 1854 alloc_flags, preferred_zone, 1855 migratetype); 1856 return page; 1857} 1858 1859/* 1860 * This is called in the allocator slow-path if the allocation request is of 1861 * sufficient urgency to ignore watermarks and take other desperate measures 1862 */ 1863static inline struct page * 1864__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1865 struct zonelist *zonelist, enum zone_type high_zoneidx, 1866 nodemask_t *nodemask, struct zone *preferred_zone, 1867 int migratetype) 1868{ 1869 struct page *page; 1870 1871 do { 1872 page = get_page_from_freelist(gfp_mask, nodemask, order, 1873 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 1874 preferred_zone, migratetype); 1875 1876 if (!page && gfp_mask & __GFP_NOFAIL) 1877 congestion_wait(BLK_RW_ASYNC, HZ/50); 1878 } while (!page && (gfp_mask & __GFP_NOFAIL)); 1879 1880 return page; 1881} 1882 1883static inline 1884void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 1885 enum zone_type high_zoneidx) 1886{ 1887 struct zoneref *z; 1888 struct zone *zone; 1889 1890 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 1891 wakeup_kswapd(zone, order); 1892} 1893 1894static inline int 1895gfp_to_alloc_flags(gfp_t gfp_mask) 1896{ 1897 struct task_struct *p = current; 1898 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 1899 const gfp_t wait = gfp_mask & __GFP_WAIT; 1900 1901 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 1902 BUILD_BUG_ON(__GFP_HIGH != ALLOC_HIGH); 1903 1904 /* 1905 * The caller may dip into page reserves a bit more if the caller 1906 * cannot run direct reclaim, or if the caller has realtime scheduling 1907 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1908 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1909 */ 1910 alloc_flags |= (gfp_mask & __GFP_HIGH); 1911 1912 if (!wait) { 1913 alloc_flags |= ALLOC_HARDER; 1914 /* 1915 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1916 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1917 */ 1918 alloc_flags &= ~ALLOC_CPUSET; 1919 } else if (unlikely(rt_task(p)) && !in_interrupt()) 1920 alloc_flags |= ALLOC_HARDER; 1921 1922 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 1923 if (!in_interrupt() && 1924 ((p->flags & PF_MEMALLOC) || 1925 unlikely(test_thread_flag(TIF_MEMDIE)))) 1926 alloc_flags |= ALLOC_NO_WATERMARKS; 1927 } 1928 1929 return alloc_flags; 1930} 1931 1932static inline struct page * 1933__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 1934 struct zonelist *zonelist, enum zone_type high_zoneidx, 1935 nodemask_t *nodemask, struct zone *preferred_zone, 1936 int migratetype) 1937{ 1938 const gfp_t wait = gfp_mask & __GFP_WAIT; 1939 struct page *page = NULL; 1940 int alloc_flags; 1941 unsigned long pages_reclaimed = 0; 1942 unsigned long did_some_progress; 1943 struct task_struct *p = current; 1944 1945 /* 1946 * In the slowpath, we sanity check order to avoid ever trying to 1947 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 1948 * be using allocators in order of preference for an area that is 1949 * too large. 1950 */ 1951 if (order >= MAX_ORDER) { 1952 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 1953 return NULL; 1954 } 1955 1956 /* 1957 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1958 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1959 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1960 * using a larger set of nodes after it has established that the 1961 * allowed per node queues are empty and that nodes are 1962 * over allocated. 1963 */ 1964 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1965 goto nopage; 1966 1967restart: 1968 wake_all_kswapd(order, zonelist, high_zoneidx); 1969 1970 /* 1971 * OK, we're below the kswapd watermark and have kicked background 1972 * reclaim. Now things get more complex, so set up alloc_flags according 1973 * to how we want to proceed. 1974 */ 1975 alloc_flags = gfp_to_alloc_flags(gfp_mask); 1976 1977 /* This is the last chance, in general, before the goto nopage. */ 1978 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 1979 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 1980 preferred_zone, migratetype); 1981 if (page) 1982 goto got_pg; 1983 1984rebalance: 1985 /* Allocate without watermarks if the context allows */ 1986 if (alloc_flags & ALLOC_NO_WATERMARKS) { 1987 page = __alloc_pages_high_priority(gfp_mask, order, 1988 zonelist, high_zoneidx, nodemask, 1989 preferred_zone, migratetype); 1990 if (page) 1991 goto got_pg; 1992 } 1993 1994 /* Atomic allocations - we can't balance anything */ 1995 if (!wait) 1996 goto nopage; 1997 1998 /* Avoid recursion of direct reclaim */ 1999 if (p->flags & PF_MEMALLOC) 2000 goto nopage; 2001 2002 /* Avoid allocations with no watermarks from looping endlessly */ 2003 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2004 goto nopage; 2005 2006 /* Try direct compaction */ 2007 page = __alloc_pages_direct_compact(gfp_mask, order, 2008 zonelist, high_zoneidx, 2009 nodemask, 2010 alloc_flags, preferred_zone, 2011 migratetype, &did_some_progress); 2012 if (page) 2013 goto got_pg; 2014 2015 /* Try direct reclaim and then allocating */ 2016 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2017 zonelist, high_zoneidx, 2018 nodemask, 2019 alloc_flags, preferred_zone, 2020 migratetype, &did_some_progress); 2021 if (page) 2022 goto got_pg; 2023 2024 /* 2025 * If we failed to make any progress reclaiming, then we are 2026 * running out of options and have to consider going OOM 2027 */ 2028 if (!did_some_progress) { 2029 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2030 if (oom_killer_disabled) 2031 goto nopage; 2032 page = __alloc_pages_may_oom(gfp_mask, order, 2033 zonelist, high_zoneidx, 2034 nodemask, preferred_zone, 2035 migratetype); 2036 if (page) 2037 goto got_pg; 2038 2039 /* 2040 * The OOM killer does not trigger for high-order 2041 * ~__GFP_NOFAIL allocations so if no progress is being 2042 * made, there are no other options and retrying is 2043 * unlikely to help. 2044 */ 2045 if (order > PAGE_ALLOC_COSTLY_ORDER && 2046 !(gfp_mask & __GFP_NOFAIL)) 2047 goto nopage; 2048 2049 goto restart; 2050 } 2051 } 2052 2053 /* Check if we should retry the allocation */ 2054 pages_reclaimed += did_some_progress; 2055 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 2056 /* Wait for some write requests to complete then retry */ 2057 congestion_wait(BLK_RW_ASYNC, HZ/50); 2058 goto rebalance; 2059 } 2060 2061nopage: 2062 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 2063 printk(KERN_WARNING "%s: page allocation failure." 2064 " order:%d, mode:0x%x\n", 2065 p->comm, order, gfp_mask); 2066 dump_stack(); 2067 show_mem(); 2068 } 2069 return page; 2070got_pg: 2071 if (kmemcheck_enabled) 2072 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2073 return page; 2074 2075} 2076 2077/* 2078 * This is the 'heart' of the zoned buddy allocator. 2079 */ 2080struct page * 2081__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2082 struct zonelist *zonelist, nodemask_t *nodemask) 2083{ 2084 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2085 struct zone *preferred_zone; 2086 struct page *page; 2087 int migratetype = allocflags_to_migratetype(gfp_mask); 2088 2089 gfp_mask &= gfp_allowed_mask; 2090 2091 lockdep_trace_alloc(gfp_mask); 2092 2093 might_sleep_if(gfp_mask & __GFP_WAIT); 2094 2095 if (should_fail_alloc_page(gfp_mask, order)) 2096 return NULL; 2097 2098 /* 2099 * Check the zones suitable for the gfp_mask contain at least one 2100 * valid zone. It's possible to have an empty zonelist as a result 2101 * of GFP_THISNODE and a memoryless node 2102 */ 2103 if (unlikely(!zonelist->_zonerefs->zone)) 2104 return NULL; 2105 2106 get_mems_allowed(); 2107 /* The preferred zone is used for statistics later */ 2108 first_zones_zonelist(zonelist, high_zoneidx, nodemask, &preferred_zone); 2109 if (!preferred_zone) { 2110 put_mems_allowed(); 2111 return NULL; 2112 } 2113 2114 /* First allocation attempt */ 2115 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2116 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2117 preferred_zone, migratetype); 2118 if (unlikely(!page)) 2119 page = __alloc_pages_slowpath(gfp_mask, order, 2120 zonelist, high_zoneidx, nodemask, 2121 preferred_zone, migratetype); 2122 put_mems_allowed(); 2123 2124 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2125 return page; 2126} 2127EXPORT_SYMBOL(__alloc_pages_nodemask); 2128 2129/* 2130 * Common helper functions. 2131 */ 2132unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2133{ 2134 struct page *page; 2135 2136 /* 2137 * __get_free_pages() returns a 32-bit address, which cannot represent 2138 * a highmem page 2139 */ 2140 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2141 2142 page = alloc_pages(gfp_mask, order); 2143 if (!page) 2144 return 0; 2145 return (unsigned long) page_address(page); 2146} 2147EXPORT_SYMBOL(__get_free_pages); 2148 2149unsigned long get_zeroed_page(gfp_t gfp_mask) 2150{ 2151 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2152} 2153EXPORT_SYMBOL(get_zeroed_page); 2154 2155void __pagevec_free(struct pagevec *pvec) 2156{ 2157 int i = pagevec_count(pvec); 2158 2159 while (--i >= 0) { 2160 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2161 free_hot_cold_page(pvec->pages[i], pvec->cold); 2162 } 2163} 2164 2165void __free_pages(struct page *page, unsigned int order) 2166{ 2167 if (put_page_testzero(page)) { 2168 if (order == 0) 2169 free_hot_cold_page(page, 0); 2170 else 2171 __free_pages_ok(page, order); 2172 } 2173} 2174 2175EXPORT_SYMBOL(__free_pages); 2176 2177void free_pages(unsigned long addr, unsigned int order) 2178{ 2179 if (addr != 0) { 2180 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2181 __free_pages(virt_to_page((void *)addr), order); 2182 } 2183} 2184 2185EXPORT_SYMBOL(free_pages); 2186 2187/** 2188 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2189 * @size: the number of bytes to allocate 2190 * @gfp_mask: GFP flags for the allocation 2191 * 2192 * This function is similar to alloc_pages(), except that it allocates the 2193 * minimum number of pages to satisfy the request. alloc_pages() can only 2194 * allocate memory in power-of-two pages. 2195 * 2196 * This function is also limited by MAX_ORDER. 2197 * 2198 * Memory allocated by this function must be released by free_pages_exact(). 2199 */ 2200void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2201{ 2202 unsigned int order = get_order(size); 2203 unsigned long addr; 2204 2205 addr = __get_free_pages(gfp_mask, order); 2206 if (addr) { 2207 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2208 unsigned long used = addr + PAGE_ALIGN(size); 2209 2210 split_page(virt_to_page((void *)addr), order); 2211 while (used < alloc_end) { 2212 free_page(used); 2213 used += PAGE_SIZE; 2214 } 2215 } 2216 2217 return (void *)addr; 2218} 2219EXPORT_SYMBOL(alloc_pages_exact); 2220 2221/** 2222 * free_pages_exact - release memory allocated via alloc_pages_exact() 2223 * @virt: the value returned by alloc_pages_exact. 2224 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2225 * 2226 * Release the memory allocated by a previous call to alloc_pages_exact. 2227 */ 2228void free_pages_exact(void *virt, size_t size) 2229{ 2230 unsigned long addr = (unsigned long)virt; 2231 unsigned long end = addr + PAGE_ALIGN(size); 2232 2233 while (addr < end) { 2234 free_page(addr); 2235 addr += PAGE_SIZE; 2236 } 2237} 2238EXPORT_SYMBOL(free_pages_exact); 2239 2240static unsigned int nr_free_zone_pages(int offset) 2241{ 2242 struct zoneref *z; 2243 struct zone *zone; 2244 2245 /* Just pick one node, since fallback list is circular */ 2246 unsigned int sum = 0; 2247 2248 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2249 2250 for_each_zone_zonelist(zone, z, zonelist, offset) { 2251 unsigned long size = zone->present_pages; 2252 unsigned long high = high_wmark_pages(zone); 2253 if (size > high) 2254 sum += size - high; 2255 } 2256 2257 return sum; 2258} 2259 2260/* 2261 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2262 */ 2263unsigned int nr_free_buffer_pages(void) 2264{ 2265 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2266} 2267EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2268 2269/* 2270 * Amount of free RAM allocatable within all zones 2271 */ 2272unsigned int nr_free_pagecache_pages(void) 2273{ 2274 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2275} 2276 2277static inline void show_node(struct zone *zone) 2278{ 2279 if (NUMA_BUILD) 2280 printk("Node %d ", zone_to_nid(zone)); 2281} 2282 2283void si_meminfo(struct sysinfo *val) 2284{ 2285 val->totalram = totalram_pages; 2286 val->sharedram = 0; 2287 val->freeram = global_page_state(NR_FREE_PAGES); 2288 val->bufferram = nr_blockdev_pages(); 2289 val->totalhigh = totalhigh_pages; 2290 val->freehigh = nr_free_highpages(); 2291 val->mem_unit = PAGE_SIZE; 2292} 2293 2294EXPORT_SYMBOL(si_meminfo); 2295 2296#ifdef CONFIG_NUMA 2297void si_meminfo_node(struct sysinfo *val, int nid) 2298{ 2299 pg_data_t *pgdat = NODE_DATA(nid); 2300 2301 val->totalram = pgdat->node_present_pages; 2302 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2303#ifdef CONFIG_HIGHMEM 2304 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2305 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2306 NR_FREE_PAGES); 2307#else 2308 val->totalhigh = 0; 2309 val->freehigh = 0; 2310#endif 2311 val->mem_unit = PAGE_SIZE; 2312} 2313#endif 2314 2315#define K(x) ((x) << (PAGE_SHIFT-10)) 2316 2317/* 2318 * Show free area list (used inside shift_scroll-lock stuff) 2319 * We also calculate the percentage fragmentation. We do this by counting the 2320 * memory on each free list with the exception of the first item on the list. 2321 */ 2322void show_free_areas(void) 2323{ 2324 int cpu; 2325 struct zone *zone; 2326 2327 for_each_populated_zone(zone) { 2328 show_node(zone); 2329 printk("%s per-cpu:\n", zone->name); 2330 2331 for_each_online_cpu(cpu) { 2332 struct per_cpu_pageset *pageset; 2333 2334 pageset = per_cpu_ptr(zone->pageset, cpu); 2335 2336 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2337 cpu, pageset->pcp.high, 2338 pageset->pcp.batch, pageset->pcp.count); 2339 } 2340 } 2341 2342 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2343 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2344 " unevictable:%lu" 2345 " dirty:%lu writeback:%lu unstable:%lu\n" 2346 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2347 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2348 global_page_state(NR_ACTIVE_ANON), 2349 global_page_state(NR_INACTIVE_ANON), 2350 global_page_state(NR_ISOLATED_ANON), 2351 global_page_state(NR_ACTIVE_FILE), 2352 global_page_state(NR_INACTIVE_FILE), 2353 global_page_state(NR_ISOLATED_FILE), 2354 global_page_state(NR_UNEVICTABLE), 2355 global_page_state(NR_FILE_DIRTY), 2356 global_page_state(NR_WRITEBACK), 2357 global_page_state(NR_UNSTABLE_NFS), 2358 global_page_state(NR_FREE_PAGES), 2359 global_page_state(NR_SLAB_RECLAIMABLE), 2360 global_page_state(NR_SLAB_UNRECLAIMABLE), 2361 global_page_state(NR_FILE_MAPPED), 2362 global_page_state(NR_SHMEM), 2363 global_page_state(NR_PAGETABLE), 2364 global_page_state(NR_BOUNCE)); 2365 2366 for_each_populated_zone(zone) { 2367 int i; 2368 2369 show_node(zone); 2370 printk("%s" 2371 " free:%lukB" 2372 " min:%lukB" 2373 " low:%lukB" 2374 " high:%lukB" 2375 " active_anon:%lukB" 2376 " inactive_anon:%lukB" 2377 " active_file:%lukB" 2378 " inactive_file:%lukB" 2379 " unevictable:%lukB" 2380 " isolated(anon):%lukB" 2381 " isolated(file):%lukB" 2382 " present:%lukB" 2383 " mlocked:%lukB" 2384 " dirty:%lukB" 2385 " writeback:%lukB" 2386 " mapped:%lukB" 2387 " shmem:%lukB" 2388 " slab_reclaimable:%lukB" 2389 " slab_unreclaimable:%lukB" 2390 " kernel_stack:%lukB" 2391 " pagetables:%lukB" 2392 " unstable:%lukB" 2393 " bounce:%lukB" 2394 " writeback_tmp:%lukB" 2395 " pages_scanned:%lu" 2396 " all_unreclaimable? %s" 2397 "\n", 2398 zone->name, 2399 K(zone_page_state(zone, NR_FREE_PAGES)), 2400 K(min_wmark_pages(zone)), 2401 K(low_wmark_pages(zone)), 2402 K(high_wmark_pages(zone)), 2403 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2404 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2405 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2406 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2407 K(zone_page_state(zone, NR_UNEVICTABLE)), 2408 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2409 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2410 K(zone->present_pages), 2411 K(zone_page_state(zone, NR_MLOCK)), 2412 K(zone_page_state(zone, NR_FILE_DIRTY)), 2413 K(zone_page_state(zone, NR_WRITEBACK)), 2414 K(zone_page_state(zone, NR_FILE_MAPPED)), 2415 K(zone_page_state(zone, NR_SHMEM)), 2416 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2417 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2418 zone_page_state(zone, NR_KERNEL_STACK) * 2419 THREAD_SIZE / 1024, 2420 K(zone_page_state(zone, NR_PAGETABLE)), 2421 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2422 K(zone_page_state(zone, NR_BOUNCE)), 2423 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2424 zone->pages_scanned, 2425 (zone->all_unreclaimable ? "yes" : "no") 2426 ); 2427 printk("lowmem_reserve[]:"); 2428 for (i = 0; i < MAX_NR_ZONES; i++) 2429 printk(" %lu", zone->lowmem_reserve[i]); 2430 printk("\n"); 2431 } 2432 2433 for_each_populated_zone(zone) { 2434 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2435 2436 show_node(zone); 2437 printk("%s: ", zone->name); 2438 2439 spin_lock_irqsave(&zone->lock, flags); 2440 for (order = 0; order < MAX_ORDER; order++) { 2441 nr[order] = zone->free_area[order].nr_free; 2442 total += nr[order] << order; 2443 } 2444 spin_unlock_irqrestore(&zone->lock, flags); 2445 for (order = 0; order < MAX_ORDER; order++) 2446 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2447 printk("= %lukB\n", K(total)); 2448 } 2449 2450 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2451 2452 show_swap_cache_info(); 2453} 2454 2455static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2456{ 2457 zoneref->zone = zone; 2458 zoneref->zone_idx = zone_idx(zone); 2459} 2460 2461/* 2462 * Builds allocation fallback zone lists. 2463 * 2464 * Add all populated zones of a node to the zonelist. 2465 */ 2466static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2467 int nr_zones, enum zone_type zone_type) 2468{ 2469 struct zone *zone; 2470 2471 BUG_ON(zone_type >= MAX_NR_ZONES); 2472 zone_type++; 2473 2474 do { 2475 zone_type--; 2476 zone = pgdat->node_zones + zone_type; 2477 if (populated_zone(zone)) { 2478 zoneref_set_zone(zone, 2479 &zonelist->_zonerefs[nr_zones++]); 2480 check_highest_zone(zone_type); 2481 } 2482 2483 } while (zone_type); 2484 return nr_zones; 2485} 2486 2487 2488/* 2489 * zonelist_order: 2490 * 0 = automatic detection of better ordering. 2491 * 1 = order by ([node] distance, -zonetype) 2492 * 2 = order by (-zonetype, [node] distance) 2493 * 2494 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2495 * the same zonelist. So only NUMA can configure this param. 2496 */ 2497#define ZONELIST_ORDER_DEFAULT 0 2498#define ZONELIST_ORDER_NODE 1 2499#define ZONELIST_ORDER_ZONE 2 2500 2501/* zonelist order in the kernel. 2502 * set_zonelist_order() will set this to NODE or ZONE. 2503 */ 2504static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2505static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2506 2507 2508#ifdef CONFIG_NUMA 2509/* The value user specified ....changed by config */ 2510static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2511/* string for sysctl */ 2512#define NUMA_ZONELIST_ORDER_LEN 16 2513char numa_zonelist_order[16] = "default"; 2514 2515/* 2516 * interface for configure zonelist ordering. 2517 * command line option "numa_zonelist_order" 2518 * = "[dD]efault - default, automatic configuration. 2519 * = "[nN]ode - order by node locality, then by zone within node 2520 * = "[zZ]one - order by zone, then by locality within zone 2521 */ 2522 2523static int __parse_numa_zonelist_order(char *s) 2524{ 2525 if (*s == 'd' || *s == 'D') { 2526 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2527 } else if (*s == 'n' || *s == 'N') { 2528 user_zonelist_order = ZONELIST_ORDER_NODE; 2529 } else if (*s == 'z' || *s == 'Z') { 2530 user_zonelist_order = ZONELIST_ORDER_ZONE; 2531 } else { 2532 printk(KERN_WARNING 2533 "Ignoring invalid numa_zonelist_order value: " 2534 "%s\n", s); 2535 return -EINVAL; 2536 } 2537 return 0; 2538} 2539 2540static __init int setup_numa_zonelist_order(char *s) 2541{ 2542 if (s) 2543 return __parse_numa_zonelist_order(s); 2544 return 0; 2545} 2546early_param("numa_zonelist_order", setup_numa_zonelist_order); 2547 2548/* 2549 * sysctl handler for numa_zonelist_order 2550 */ 2551int numa_zonelist_order_handler(ctl_table *table, int write, 2552 void __user *buffer, size_t *length, 2553 loff_t *ppos) 2554{ 2555 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2556 int ret; 2557 static DEFINE_MUTEX(zl_order_mutex); 2558 2559 mutex_lock(&zl_order_mutex); 2560 if (write) 2561 strcpy(saved_string, (char*)table->data); 2562 ret = proc_dostring(table, write, buffer, length, ppos); 2563 if (ret) 2564 goto out; 2565 if (write) { 2566 int oldval = user_zonelist_order; 2567 if (__parse_numa_zonelist_order((char*)table->data)) { 2568 /* 2569 * bogus value. restore saved string 2570 */ 2571 strncpy((char*)table->data, saved_string, 2572 NUMA_ZONELIST_ORDER_LEN); 2573 user_zonelist_order = oldval; 2574 } else if (oldval != user_zonelist_order) 2575 build_all_zonelists(NULL); 2576 } 2577out: 2578 mutex_unlock(&zl_order_mutex); 2579 return ret; 2580} 2581 2582 2583#define MAX_NODE_LOAD (nr_online_nodes) 2584static int node_load[MAX_NUMNODES]; 2585 2586/** 2587 * find_next_best_node - find the next node that should appear in a given node's fallback list 2588 * @node: node whose fallback list we're appending 2589 * @used_node_mask: nodemask_t of already used nodes 2590 * 2591 * We use a number of factors to determine which is the next node that should 2592 * appear on a given node's fallback list. The node should not have appeared 2593 * already in @node's fallback list, and it should be the next closest node 2594 * according to the distance array (which contains arbitrary distance values 2595 * from each node to each node in the system), and should also prefer nodes 2596 * with no CPUs, since presumably they'll have very little allocation pressure 2597 * on them otherwise. 2598 * It returns -1 if no node is found. 2599 */ 2600static int find_next_best_node(int node, nodemask_t *used_node_mask) 2601{ 2602 int n, val; 2603 int min_val = INT_MAX; 2604 int best_node = -1; 2605 const struct cpumask *tmp = cpumask_of_node(0); 2606 2607 /* Use the local node if we haven't already */ 2608 if (!node_isset(node, *used_node_mask)) { 2609 node_set(node, *used_node_mask); 2610 return node; 2611 } 2612 2613 for_each_node_state(n, N_HIGH_MEMORY) { 2614 2615 /* Don't want a node to appear more than once */ 2616 if (node_isset(n, *used_node_mask)) 2617 continue; 2618 2619 /* Use the distance array to find the distance */ 2620 val = node_distance(node, n); 2621 2622 /* Penalize nodes under us ("prefer the next node") */ 2623 val += (n < node); 2624 2625 /* Give preference to headless and unused nodes */ 2626 tmp = cpumask_of_node(n); 2627 if (!cpumask_empty(tmp)) 2628 val += PENALTY_FOR_NODE_WITH_CPUS; 2629 2630 /* Slight preference for less loaded node */ 2631 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2632 val += node_load[n]; 2633 2634 if (val < min_val) { 2635 min_val = val; 2636 best_node = n; 2637 } 2638 } 2639 2640 if (best_node >= 0) 2641 node_set(best_node, *used_node_mask); 2642 2643 return best_node; 2644} 2645 2646 2647/* 2648 * Build zonelists ordered by node and zones within node. 2649 * This results in maximum locality--normal zone overflows into local 2650 * DMA zone, if any--but risks exhausting DMA zone. 2651 */ 2652static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2653{ 2654 int j; 2655 struct zonelist *zonelist; 2656 2657 zonelist = &pgdat->node_zonelists[0]; 2658 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2659 ; 2660 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2661 MAX_NR_ZONES - 1); 2662 zonelist->_zonerefs[j].zone = NULL; 2663 zonelist->_zonerefs[j].zone_idx = 0; 2664} 2665 2666/* 2667 * Build gfp_thisnode zonelists 2668 */ 2669static void build_thisnode_zonelists(pg_data_t *pgdat) 2670{ 2671 int j; 2672 struct zonelist *zonelist; 2673 2674 zonelist = &pgdat->node_zonelists[1]; 2675 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2676 zonelist->_zonerefs[j].zone = NULL; 2677 zonelist->_zonerefs[j].zone_idx = 0; 2678} 2679 2680/* 2681 * Build zonelists ordered by zone and nodes within zones. 2682 * This results in conserving DMA zone[s] until all Normal memory is 2683 * exhausted, but results in overflowing to remote node while memory 2684 * may still exist in local DMA zone. 2685 */ 2686static int node_order[MAX_NUMNODES]; 2687 2688static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2689{ 2690 int pos, j, node; 2691 int zone_type; /* needs to be signed */ 2692 struct zone *z; 2693 struct zonelist *zonelist; 2694 2695 zonelist = &pgdat->node_zonelists[0]; 2696 pos = 0; 2697 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2698 for (j = 0; j < nr_nodes; j++) { 2699 node = node_order[j]; 2700 z = &NODE_DATA(node)->node_zones[zone_type]; 2701 if (populated_zone(z)) { 2702 zoneref_set_zone(z, 2703 &zonelist->_zonerefs[pos++]); 2704 check_highest_zone(zone_type); 2705 } 2706 } 2707 } 2708 zonelist->_zonerefs[pos].zone = NULL; 2709 zonelist->_zonerefs[pos].zone_idx = 0; 2710} 2711 2712static int default_zonelist_order(void) 2713{ 2714 int nid, zone_type; 2715 unsigned long low_kmem_size,total_size; 2716 struct zone *z; 2717 int average_size; 2718 /* 2719 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 2720 * If they are really small and used heavily, the system can fall 2721 * into OOM very easily. 2722 * This function detect ZONE_DMA/DMA32 size and configures zone order. 2723 */ 2724 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2725 low_kmem_size = 0; 2726 total_size = 0; 2727 for_each_online_node(nid) { 2728 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2729 z = &NODE_DATA(nid)->node_zones[zone_type]; 2730 if (populated_zone(z)) { 2731 if (zone_type < ZONE_NORMAL) 2732 low_kmem_size += z->present_pages; 2733 total_size += z->present_pages; 2734 } else if (zone_type == ZONE_NORMAL) { 2735 /* 2736 * If any node has only lowmem, then node order 2737 * is preferred to allow kernel allocations 2738 * locally; otherwise, they can easily infringe 2739 * on other nodes when there is an abundance of 2740 * lowmem available to allocate from. 2741 */ 2742 return ZONELIST_ORDER_NODE; 2743 } 2744 } 2745 } 2746 if (!low_kmem_size || /* there are no DMA area. */ 2747 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2748 return ZONELIST_ORDER_NODE; 2749 /* 2750 * look into each node's config. 2751 * If there is a node whose DMA/DMA32 memory is very big area on 2752 * local memory, NODE_ORDER may be suitable. 2753 */ 2754 average_size = total_size / 2755 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2756 for_each_online_node(nid) { 2757 low_kmem_size = 0; 2758 total_size = 0; 2759 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2760 z = &NODE_DATA(nid)->node_zones[zone_type]; 2761 if (populated_zone(z)) { 2762 if (zone_type < ZONE_NORMAL) 2763 low_kmem_size += z->present_pages; 2764 total_size += z->present_pages; 2765 } 2766 } 2767 if (low_kmem_size && 2768 total_size > average_size && /* ignore small node */ 2769 low_kmem_size > total_size * 70/100) 2770 return ZONELIST_ORDER_NODE; 2771 } 2772 return ZONELIST_ORDER_ZONE; 2773} 2774 2775static void set_zonelist_order(void) 2776{ 2777 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2778 current_zonelist_order = default_zonelist_order(); 2779 else 2780 current_zonelist_order = user_zonelist_order; 2781} 2782 2783static void build_zonelists(pg_data_t *pgdat) 2784{ 2785 int j, node, load; 2786 enum zone_type i; 2787 nodemask_t used_mask; 2788 int local_node, prev_node; 2789 struct zonelist *zonelist; 2790 int order = current_zonelist_order; 2791 2792 /* initialize zonelists */ 2793 for (i = 0; i < MAX_ZONELISTS; i++) { 2794 zonelist = pgdat->node_zonelists + i; 2795 zonelist->_zonerefs[0].zone = NULL; 2796 zonelist->_zonerefs[0].zone_idx = 0; 2797 } 2798 2799 /* NUMA-aware ordering of nodes */ 2800 local_node = pgdat->node_id; 2801 load = nr_online_nodes; 2802 prev_node = local_node; 2803 nodes_clear(used_mask); 2804 2805 memset(node_order, 0, sizeof(node_order)); 2806 j = 0; 2807 2808 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2809 int distance = node_distance(local_node, node); 2810 2811 /* 2812 * If another node is sufficiently far away then it is better 2813 * to reclaim pages in a zone before going off node. 2814 */ 2815 if (distance > RECLAIM_DISTANCE) 2816 zone_reclaim_mode = 1; 2817 2818 /* 2819 * We don't want to pressure a particular node. 2820 * So adding penalty to the first node in same 2821 * distance group to make it round-robin. 2822 */ 2823 if (distance != node_distance(local_node, prev_node)) 2824 node_load[node] = load; 2825 2826 prev_node = node; 2827 load--; 2828 if (order == ZONELIST_ORDER_NODE) 2829 build_zonelists_in_node_order(pgdat, node); 2830 else 2831 node_order[j++] = node; /* remember order */ 2832 } 2833 2834 if (order == ZONELIST_ORDER_ZONE) { 2835 /* calculate node order -- i.e., DMA last! */ 2836 build_zonelists_in_zone_order(pgdat, j); 2837 } 2838 2839 build_thisnode_zonelists(pgdat); 2840} 2841 2842/* Construct the zonelist performance cache - see further mmzone.h */ 2843static void build_zonelist_cache(pg_data_t *pgdat) 2844{ 2845 struct zonelist *zonelist; 2846 struct zonelist_cache *zlc; 2847 struct zoneref *z; 2848 2849 zonelist = &pgdat->node_zonelists[0]; 2850 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2851 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2852 for (z = zonelist->_zonerefs; z->zone; z++) 2853 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 2854} 2855 2856 2857#else /* CONFIG_NUMA */ 2858 2859static void set_zonelist_order(void) 2860{ 2861 current_zonelist_order = ZONELIST_ORDER_ZONE; 2862} 2863 2864static void build_zonelists(pg_data_t *pgdat) 2865{ 2866 int node, local_node; 2867 enum zone_type j; 2868 struct zonelist *zonelist; 2869 2870 local_node = pgdat->node_id; 2871 2872 zonelist = &pgdat->node_zonelists[0]; 2873 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2874 2875 /* 2876 * Now we build the zonelist so that it contains the zones 2877 * of all the other nodes. 2878 * We don't want to pressure a particular node, so when 2879 * building the zones for node N, we make sure that the 2880 * zones coming right after the local ones are those from 2881 * node N+1 (modulo N) 2882 */ 2883 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2884 if (!node_online(node)) 2885 continue; 2886 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2887 MAX_NR_ZONES - 1); 2888 } 2889 for (node = 0; node < local_node; node++) { 2890 if (!node_online(node)) 2891 continue; 2892 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2893 MAX_NR_ZONES - 1); 2894 } 2895 2896 zonelist->_zonerefs[j].zone = NULL; 2897 zonelist->_zonerefs[j].zone_idx = 0; 2898} 2899 2900/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2901static void build_zonelist_cache(pg_data_t *pgdat) 2902{ 2903 pgdat->node_zonelists[0].zlcache_ptr = NULL; 2904} 2905 2906#endif /* CONFIG_NUMA */ 2907 2908/* 2909 * Boot pageset table. One per cpu which is going to be used for all 2910 * zones and all nodes. The parameters will be set in such a way 2911 * that an item put on a list will immediately be handed over to 2912 * the buddy list. This is safe since pageset manipulation is done 2913 * with interrupts disabled. 2914 * 2915 * The boot_pagesets must be kept even after bootup is complete for 2916 * unused processors and/or zones. They do play a role for bootstrapping 2917 * hotplugged processors. 2918 * 2919 * zoneinfo_show() and maybe other functions do 2920 * not check if the processor is online before following the pageset pointer. 2921 * Other parts of the kernel may not check if the zone is available. 2922 */ 2923static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 2924static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 2925static void setup_zone_pageset(struct zone *zone); 2926 2927/* return values int ....just for stop_machine() */ 2928static __init_refok int __build_all_zonelists(void *data) 2929{ 2930 int nid; 2931 int cpu; 2932 2933#ifdef CONFIG_NUMA 2934 memset(node_load, 0, sizeof(node_load)); 2935#endif 2936 for_each_online_node(nid) { 2937 pg_data_t *pgdat = NODE_DATA(nid); 2938 2939 build_zonelists(pgdat); 2940 build_zonelist_cache(pgdat); 2941 } 2942 2943#ifdef CONFIG_MEMORY_HOTPLUG 2944 /* Setup real pagesets for the new zone */ 2945 if (data) { 2946 struct zone *zone = data; 2947 setup_zone_pageset(zone); 2948 } 2949#endif 2950 2951 /* 2952 * Initialize the boot_pagesets that are going to be used 2953 * for bootstrapping processors. The real pagesets for 2954 * each zone will be allocated later when the per cpu 2955 * allocator is available. 2956 * 2957 * boot_pagesets are used also for bootstrapping offline 2958 * cpus if the system is already booted because the pagesets 2959 * are needed to initialize allocators on a specific cpu too. 2960 * F.e. the percpu allocator needs the page allocator which 2961 * needs the percpu allocator in order to allocate its pagesets 2962 * (a chicken-egg dilemma). 2963 */ 2964 for_each_possible_cpu(cpu) 2965 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 2966 2967 return 0; 2968} 2969 2970void build_all_zonelists(void *data) 2971{ 2972 set_zonelist_order(); 2973 2974 if (system_state == SYSTEM_BOOTING) { 2975 __build_all_zonelists(NULL); 2976 mminit_verify_zonelist(); 2977 cpuset_init_current_mems_allowed(); 2978 } else { 2979 /* we have to stop all cpus to guarantee there is no user 2980 of zonelist */ 2981 stop_machine(__build_all_zonelists, data, NULL); 2982 /* cpuset refresh routine should be here */ 2983 } 2984 vm_total_pages = nr_free_pagecache_pages(); 2985 /* 2986 * Disable grouping by mobility if the number of pages in the 2987 * system is too low to allow the mechanism to work. It would be 2988 * more accurate, but expensive to check per-zone. This check is 2989 * made on memory-hotadd so a system can start with mobility 2990 * disabled and enable it later 2991 */ 2992 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2993 page_group_by_mobility_disabled = 1; 2994 else 2995 page_group_by_mobility_disabled = 0; 2996 2997 printk("Built %i zonelists in %s order, mobility grouping %s. " 2998 "Total pages: %ld\n", 2999 nr_online_nodes, 3000 zonelist_order_name[current_zonelist_order], 3001 page_group_by_mobility_disabled ? "off" : "on", 3002 vm_total_pages); 3003#ifdef CONFIG_NUMA 3004 printk("Policy zone: %s\n", zone_names[policy_zone]); 3005#endif 3006} 3007 3008/* 3009 * Helper functions to size the waitqueue hash table. 3010 * Essentially these want to choose hash table sizes sufficiently 3011 * large so that collisions trying to wait on pages are rare. 3012 * But in fact, the number of active page waitqueues on typical 3013 * systems is ridiculously low, less than 200. So this is even 3014 * conservative, even though it seems large. 3015 * 3016 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3017 * waitqueues, i.e. the size of the waitq table given the number of pages. 3018 */ 3019#define PAGES_PER_WAITQUEUE 256 3020 3021#ifndef CONFIG_MEMORY_HOTPLUG 3022static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3023{ 3024 unsigned long size = 1; 3025 3026 pages /= PAGES_PER_WAITQUEUE; 3027 3028 while (size < pages) 3029 size <<= 1; 3030 3031 /* 3032 * Once we have dozens or even hundreds of threads sleeping 3033 * on IO we've got bigger problems than wait queue collision. 3034 * Limit the size of the wait table to a reasonable size. 3035 */ 3036 size = min(size, 4096UL); 3037 3038 return max(size, 4UL); 3039} 3040#else 3041/* 3042 * A zone's size might be changed by hot-add, so it is not possible to determine 3043 * a suitable size for its wait_table. So we use the maximum size now. 3044 * 3045 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3046 * 3047 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3048 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3049 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3050 * 3051 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3052 * or more by the traditional way. (See above). It equals: 3053 * 3054 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3055 * ia64(16K page size) : = ( 8G + 4M)byte. 3056 * powerpc (64K page size) : = (32G +16M)byte. 3057 */ 3058static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3059{ 3060 return 4096UL; 3061} 3062#endif 3063 3064/* 3065 * This is an integer logarithm so that shifts can be used later 3066 * to extract the more random high bits from the multiplicative 3067 * hash function before the remainder is taken. 3068 */ 3069static inline unsigned long wait_table_bits(unsigned long size) 3070{ 3071 return ffz(~size); 3072} 3073 3074#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3075 3076/* 3077 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3078 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3079 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3080 * higher will lead to a bigger reserve which will get freed as contiguous 3081 * blocks as reclaim kicks in 3082 */ 3083static void setup_zone_migrate_reserve(struct zone *zone) 3084{ 3085 unsigned long start_pfn, pfn, end_pfn; 3086 struct page *page; 3087 unsigned long block_migratetype; 3088 int reserve; 3089 3090 /* Get the start pfn, end pfn and the number of blocks to reserve */ 3091 start_pfn = zone->zone_start_pfn; 3092 end_pfn = start_pfn + zone->spanned_pages; 3093 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3094 pageblock_order; 3095 3096 /* 3097 * Reserve blocks are generally in place to help high-order atomic 3098 * allocations that are short-lived. A min_free_kbytes value that 3099 * would result in more than 2 reserve blocks for atomic allocations 3100 * is assumed to be in place to help anti-fragmentation for the 3101 * future allocation of hugepages at runtime. 3102 */ 3103 reserve = min(2, reserve); 3104 3105 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3106 if (!pfn_valid(pfn)) 3107 continue; 3108 page = pfn_to_page(pfn); 3109 3110 /* Watch out for overlapping nodes */ 3111 if (page_to_nid(page) != zone_to_nid(zone)) 3112 continue; 3113 3114 /* Blocks with reserved pages will never free, skip them. */ 3115 if (PageReserved(page)) 3116 continue; 3117 3118 block_migratetype = get_pageblock_migratetype(page); 3119 3120 /* If this block is reserved, account for it */ 3121 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 3122 reserve--; 3123 continue; 3124 } 3125 3126 /* Suitable for reserving if this block is movable */ 3127 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 3128 set_pageblock_migratetype(page, MIGRATE_RESERVE); 3129 move_freepages_block(zone, page, MIGRATE_RESERVE); 3130 reserve--; 3131 continue; 3132 } 3133 3134 /* 3135 * If the reserve is met and this is a previous reserved block, 3136 * take it back 3137 */ 3138 if (block_migratetype == MIGRATE_RESERVE) { 3139 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3140 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3141 } 3142 } 3143} 3144 3145/* 3146 * Initially all pages are reserved - free ones are freed 3147 * up by free_all_bootmem() once the early boot process is 3148 * done. Non-atomic initialization, single-pass. 3149 */ 3150void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3151 unsigned long start_pfn, enum memmap_context context) 3152{ 3153 struct page *page; 3154 unsigned long end_pfn = start_pfn + size; 3155 unsigned long pfn; 3156 struct zone *z; 3157 3158 if (highest_memmap_pfn < end_pfn - 1) 3159 highest_memmap_pfn = end_pfn - 1; 3160 3161 z = &NODE_DATA(nid)->node_zones[zone]; 3162 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3163 /* 3164 * There can be holes in boot-time mem_map[]s 3165 * handed to this function. They do not 3166 * exist on hotplugged memory. 3167 */ 3168 if (context == MEMMAP_EARLY) { 3169 if (!early_pfn_valid(pfn)) 3170 continue; 3171 if (!early_pfn_in_nid(pfn, nid)) 3172 continue; 3173 } 3174 page = pfn_to_page(pfn); 3175 set_page_links(page, zone, nid, pfn); 3176 mminit_verify_page_links(page, zone, nid, pfn); 3177 init_page_count(page); 3178 reset_page_mapcount(page); 3179 SetPageReserved(page); 3180 /* 3181 * Mark the block movable so that blocks are reserved for 3182 * movable at startup. This will force kernel allocations 3183 * to reserve their blocks rather than leaking throughout 3184 * the address space during boot when many long-lived 3185 * kernel allocations are made. Later some blocks near 3186 * the start are marked MIGRATE_RESERVE by 3187 * setup_zone_migrate_reserve() 3188 * 3189 * bitmap is created for zone's valid pfn range. but memmap 3190 * can be created for invalid pages (for alignment) 3191 * check here not to call set_pageblock_migratetype() against 3192 * pfn out of zone. 3193 */ 3194 if ((z->zone_start_pfn <= pfn) 3195 && (pfn < z->zone_start_pfn + z->spanned_pages) 3196 && !(pfn & (pageblock_nr_pages - 1))) 3197 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3198 3199 INIT_LIST_HEAD(&page->lru); 3200#ifdef WANT_PAGE_VIRTUAL 3201 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3202 if (!is_highmem_idx(zone)) 3203 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3204#endif 3205 } 3206} 3207 3208static void __meminit zone_init_free_lists(struct zone *zone) 3209{ 3210 int order, t; 3211 for_each_migratetype_order(order, t) { 3212 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3213 zone->free_area[order].nr_free = 0; 3214 } 3215} 3216 3217#ifndef __HAVE_ARCH_MEMMAP_INIT 3218#define memmap_init(size, nid, zone, start_pfn) \ 3219 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3220#endif 3221 3222static int zone_batchsize(struct zone *zone) 3223{ 3224#ifdef CONFIG_MMU 3225 int batch; 3226 3227 /* 3228 * The per-cpu-pages pools are set to around 1000th of the 3229 * size of the zone. But no more than 1/2 of a meg. 3230 * 3231 * OK, so we don't know how big the cache is. So guess. 3232 */ 3233 batch = zone->present_pages / 1024; 3234 if (batch * PAGE_SIZE > 512 * 1024) 3235 batch = (512 * 1024) / PAGE_SIZE; 3236 batch /= 4; /* We effectively *= 4 below */ 3237 if (batch < 1) 3238 batch = 1; 3239 3240 /* 3241 * Clamp the batch to a 2^n - 1 value. Having a power 3242 * of 2 value was found to be more likely to have 3243 * suboptimal cache aliasing properties in some cases. 3244 * 3245 * For example if 2 tasks are alternately allocating 3246 * batches of pages, one task can end up with a lot 3247 * of pages of one half of the possible page colors 3248 * and the other with pages of the other colors. 3249 */ 3250 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3251 3252 return batch; 3253 3254#else 3255 /* The deferral and batching of frees should be suppressed under NOMMU 3256 * conditions. 3257 * 3258 * The problem is that NOMMU needs to be able to allocate large chunks 3259 * of contiguous memory as there's no hardware page translation to 3260 * assemble apparent contiguous memory from discontiguous pages. 3261 * 3262 * Queueing large contiguous runs of pages for batching, however, 3263 * causes the pages to actually be freed in smaller chunks. As there 3264 * can be a significant delay between the individual batches being 3265 * recycled, this leads to the once large chunks of space being 3266 * fragmented and becoming unavailable for high-order allocations. 3267 */ 3268 return 0; 3269#endif 3270} 3271 3272static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3273{ 3274 struct per_cpu_pages *pcp; 3275 int migratetype; 3276 3277 memset(p, 0, sizeof(*p)); 3278 3279 pcp = &p->pcp; 3280 pcp->count = 0; 3281 pcp->high = 6 * batch; 3282 pcp->batch = max(1UL, 1 * batch); 3283 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3284 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3285} 3286 3287/* 3288 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3289 * to the value high for the pageset p. 3290 */ 3291 3292static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3293 unsigned long high) 3294{ 3295 struct per_cpu_pages *pcp; 3296 3297 pcp = &p->pcp; 3298 pcp->high = high; 3299 pcp->batch = max(1UL, high/4); 3300 if ((high/4) > (PAGE_SHIFT * 8)) 3301 pcp->batch = PAGE_SHIFT * 8; 3302} 3303 3304static __meminit void setup_zone_pageset(struct zone *zone) 3305{ 3306 int cpu; 3307 3308 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3309 3310 for_each_possible_cpu(cpu) { 3311 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3312 3313 setup_pageset(pcp, zone_batchsize(zone)); 3314 3315 if (percpu_pagelist_fraction) 3316 setup_pagelist_highmark(pcp, 3317 (zone->present_pages / 3318 percpu_pagelist_fraction)); 3319 } 3320} 3321 3322/* 3323 * Allocate per cpu pagesets and initialize them. 3324 * Before this call only boot pagesets were available. 3325 */ 3326void __init setup_per_cpu_pageset(void) 3327{ 3328 struct zone *zone; 3329 3330 for_each_populated_zone(zone) 3331 setup_zone_pageset(zone); 3332} 3333 3334static noinline __init_refok 3335int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3336{ 3337 int i; 3338 struct pglist_data *pgdat = zone->zone_pgdat; 3339 size_t alloc_size; 3340 3341 /* 3342 * The per-page waitqueue mechanism uses hashed waitqueues 3343 * per zone. 3344 */ 3345 zone->wait_table_hash_nr_entries = 3346 wait_table_hash_nr_entries(zone_size_pages); 3347 zone->wait_table_bits = 3348 wait_table_bits(zone->wait_table_hash_nr_entries); 3349 alloc_size = zone->wait_table_hash_nr_entries 3350 * sizeof(wait_queue_head_t); 3351 3352 if (!slab_is_available()) { 3353 zone->wait_table = (wait_queue_head_t *) 3354 alloc_bootmem_node(pgdat, alloc_size); 3355 } else { 3356 /* 3357 * This case means that a zone whose size was 0 gets new memory 3358 * via memory hot-add. 3359 * But it may be the case that a new node was hot-added. In 3360 * this case vmalloc() will not be able to use this new node's 3361 * memory - this wait_table must be initialized to use this new 3362 * node itself as well. 3363 * To use this new node's memory, further consideration will be 3364 * necessary. 3365 */ 3366 zone->wait_table = vmalloc(alloc_size); 3367 } 3368 if (!zone->wait_table) 3369 return -ENOMEM; 3370 3371 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3372 init_waitqueue_head(zone->wait_table + i); 3373 3374 return 0; 3375} 3376 3377static int __zone_pcp_update(void *data) 3378{ 3379 struct zone *zone = data; 3380 int cpu; 3381 unsigned long batch = zone_batchsize(zone), flags; 3382 3383 for_each_possible_cpu(cpu) { 3384 struct per_cpu_pageset *pset; 3385 struct per_cpu_pages *pcp; 3386 3387 pset = per_cpu_ptr(zone->pageset, cpu); 3388 pcp = &pset->pcp; 3389 3390 local_irq_save(flags); 3391 free_pcppages_bulk(zone, pcp->count, pcp); 3392 setup_pageset(pset, batch); 3393 local_irq_restore(flags); 3394 } 3395 return 0; 3396} 3397 3398void zone_pcp_update(struct zone *zone) 3399{ 3400 stop_machine(__zone_pcp_update, zone, NULL); 3401} 3402 3403static __meminit void zone_pcp_init(struct zone *zone) 3404{ 3405 /* 3406 * per cpu subsystem is not up at this point. The following code 3407 * relies on the ability of the linker to provide the 3408 * offset of a (static) per cpu variable into the per cpu area. 3409 */ 3410 zone->pageset = &boot_pageset; 3411 3412 if (zone->present_pages) 3413 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3414 zone->name, zone->present_pages, 3415 zone_batchsize(zone)); 3416} 3417 3418__meminit int init_currently_empty_zone(struct zone *zone, 3419 unsigned long zone_start_pfn, 3420 unsigned long size, 3421 enum memmap_context context) 3422{ 3423 struct pglist_data *pgdat = zone->zone_pgdat; 3424 int ret; 3425 ret = zone_wait_table_init(zone, size); 3426 if (ret) 3427 return ret; 3428 pgdat->nr_zones = zone_idx(zone) + 1; 3429 3430 zone->zone_start_pfn = zone_start_pfn; 3431 3432 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3433 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3434 pgdat->node_id, 3435 (unsigned long)zone_idx(zone), 3436 zone_start_pfn, (zone_start_pfn + size)); 3437 3438 zone_init_free_lists(zone); 3439 3440 return 0; 3441} 3442 3443#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3444/* 3445 * Basic iterator support. Return the first range of PFNs for a node 3446 * Note: nid == MAX_NUMNODES returns first region regardless of node 3447 */ 3448static int __meminit first_active_region_index_in_nid(int nid) 3449{ 3450 int i; 3451 3452 for (i = 0; i < nr_nodemap_entries; i++) 3453 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3454 return i; 3455 3456 return -1; 3457} 3458 3459/* 3460 * Basic iterator support. Return the next active range of PFNs for a node 3461 * Note: nid == MAX_NUMNODES returns next region regardless of node 3462 */ 3463static int __meminit next_active_region_index_in_nid(int index, int nid) 3464{ 3465 for (index = index + 1; index < nr_nodemap_entries; index++) 3466 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3467 return index; 3468 3469 return -1; 3470} 3471 3472#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3473/* 3474 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3475 * Architectures may implement their own version but if add_active_range() 3476 * was used and there are no special requirements, this is a convenient 3477 * alternative 3478 */ 3479int __meminit __early_pfn_to_nid(unsigned long pfn) 3480{ 3481 int i; 3482 3483 for (i = 0; i < nr_nodemap_entries; i++) { 3484 unsigned long start_pfn = early_node_map[i].start_pfn; 3485 unsigned long end_pfn = early_node_map[i].end_pfn; 3486 3487 if (start_pfn <= pfn && pfn < end_pfn) 3488 return early_node_map[i].nid; 3489 } 3490 /* This is a memory hole */ 3491 return -1; 3492} 3493#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3494 3495int __meminit early_pfn_to_nid(unsigned long pfn) 3496{ 3497 int nid; 3498 3499 nid = __early_pfn_to_nid(pfn); 3500 if (nid >= 0) 3501 return nid; 3502 /* just returns 0 */ 3503 return 0; 3504} 3505 3506#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3507bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3508{ 3509 int nid; 3510 3511 nid = __early_pfn_to_nid(pfn); 3512 if (nid >= 0 && nid != node) 3513 return false; 3514 return true; 3515} 3516#endif 3517 3518/* Basic iterator support to walk early_node_map[] */ 3519#define for_each_active_range_index_in_nid(i, nid) \ 3520 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3521 i = next_active_region_index_in_nid(i, nid)) 3522 3523/** 3524 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3525 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3526 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3527 * 3528 * If an architecture guarantees that all ranges registered with 3529 * add_active_ranges() contain no holes and may be freed, this 3530 * this function may be used instead of calling free_bootmem() manually. 3531 */ 3532void __init free_bootmem_with_active_regions(int nid, 3533 unsigned long max_low_pfn) 3534{ 3535 int i; 3536 3537 for_each_active_range_index_in_nid(i, nid) { 3538 unsigned long size_pages = 0; 3539 unsigned long end_pfn = early_node_map[i].end_pfn; 3540 3541 if (early_node_map[i].start_pfn >= max_low_pfn) 3542 continue; 3543 3544 if (end_pfn > max_low_pfn) 3545 end_pfn = max_low_pfn; 3546 3547 size_pages = end_pfn - early_node_map[i].start_pfn; 3548 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3549 PFN_PHYS(early_node_map[i].start_pfn), 3550 size_pages << PAGE_SHIFT); 3551 } 3552} 3553 3554int __init add_from_early_node_map(struct range *range, int az, 3555 int nr_range, int nid) 3556{ 3557 int i; 3558 u64 start, end; 3559 3560 /* need to go over early_node_map to find out good range for node */ 3561 for_each_active_range_index_in_nid(i, nid) { 3562 start = early_node_map[i].start_pfn; 3563 end = early_node_map[i].end_pfn; 3564 nr_range = add_range(range, az, nr_range, start, end); 3565 } 3566 return nr_range; 3567} 3568 3569#ifdef CONFIG_NO_BOOTMEM 3570void * __init __alloc_memory_core_early(int nid, u64 size, u64 align, 3571 u64 goal, u64 limit) 3572{ 3573 int i; 3574 void *ptr; 3575 3576 /* need to go over early_node_map to find out good range for node */ 3577 for_each_active_range_index_in_nid(i, nid) { 3578 u64 addr; 3579 u64 ei_start, ei_last; 3580 3581 ei_last = early_node_map[i].end_pfn; 3582 ei_last <<= PAGE_SHIFT; 3583 ei_start = early_node_map[i].start_pfn; 3584 ei_start <<= PAGE_SHIFT; 3585 addr = find_early_area(ei_start, ei_last, 3586 goal, limit, size, align); 3587 3588 if (addr == -1ULL) 3589 continue; 3590 3591#if 0 3592 printk(KERN_DEBUG "alloc (nid=%d %llx - %llx) (%llx - %llx) %llx %llx => %llx\n", 3593 nid, 3594 ei_start, ei_last, goal, limit, size, 3595 align, addr); 3596#endif 3597 3598 ptr = phys_to_virt(addr); 3599 memset(ptr, 0, size); 3600 reserve_early_without_check(addr, addr + size, "BOOTMEM"); 3601 return ptr; 3602 } 3603 3604 return NULL; 3605} 3606#endif 3607 3608 3609void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3610{ 3611 int i; 3612 int ret; 3613 3614 for_each_active_range_index_in_nid(i, nid) { 3615 ret = work_fn(early_node_map[i].start_pfn, 3616 early_node_map[i].end_pfn, data); 3617 if (ret) 3618 break; 3619 } 3620} 3621/** 3622 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3623 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3624 * 3625 * If an architecture guarantees that all ranges registered with 3626 * add_active_ranges() contain no holes and may be freed, this 3627 * function may be used instead of calling memory_present() manually. 3628 */ 3629void __init sparse_memory_present_with_active_regions(int nid) 3630{ 3631 int i; 3632 3633 for_each_active_range_index_in_nid(i, nid) 3634 memory_present(early_node_map[i].nid, 3635 early_node_map[i].start_pfn, 3636 early_node_map[i].end_pfn); 3637} 3638 3639/** 3640 * get_pfn_range_for_nid - Return the start and end page frames for a node 3641 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3642 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3643 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3644 * 3645 * It returns the start and end page frame of a node based on information 3646 * provided by an arch calling add_active_range(). If called for a node 3647 * with no available memory, a warning is printed and the start and end 3648 * PFNs will be 0. 3649 */ 3650void __meminit get_pfn_range_for_nid(unsigned int nid, 3651 unsigned long *start_pfn, unsigned long *end_pfn) 3652{ 3653 int i; 3654 *start_pfn = -1UL; 3655 *end_pfn = 0; 3656 3657 for_each_active_range_index_in_nid(i, nid) { 3658 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3659 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3660 } 3661 3662 if (*start_pfn == -1UL) 3663 *start_pfn = 0; 3664} 3665 3666/* 3667 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3668 * assumption is made that zones within a node are ordered in monotonic 3669 * increasing memory addresses so that the "highest" populated zone is used 3670 */ 3671static void __init find_usable_zone_for_movable(void) 3672{ 3673 int zone_index; 3674 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3675 if (zone_index == ZONE_MOVABLE) 3676 continue; 3677 3678 if (arch_zone_highest_possible_pfn[zone_index] > 3679 arch_zone_lowest_possible_pfn[zone_index]) 3680 break; 3681 } 3682 3683 VM_BUG_ON(zone_index == -1); 3684 movable_zone = zone_index; 3685} 3686 3687/* 3688 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3689 * because it is sized independant of architecture. Unlike the other zones, 3690 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3691 * in each node depending on the size of each node and how evenly kernelcore 3692 * is distributed. This helper function adjusts the zone ranges 3693 * provided by the architecture for a given node by using the end of the 3694 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3695 * zones within a node are in order of monotonic increases memory addresses 3696 */ 3697static void __meminit adjust_zone_range_for_zone_movable(int nid, 3698 unsigned long zone_type, 3699 unsigned long node_start_pfn, 3700 unsigned long node_end_pfn, 3701 unsigned long *zone_start_pfn, 3702 unsigned long *zone_end_pfn) 3703{ 3704 /* Only adjust if ZONE_MOVABLE is on this node */ 3705 if (zone_movable_pfn[nid]) { 3706 /* Size ZONE_MOVABLE */ 3707 if (zone_type == ZONE_MOVABLE) { 3708 *zone_start_pfn = zone_movable_pfn[nid]; 3709 *zone_end_pfn = min(node_end_pfn, 3710 arch_zone_highest_possible_pfn[movable_zone]); 3711 3712 /* Adjust for ZONE_MOVABLE starting within this range */ 3713 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3714 *zone_end_pfn > zone_movable_pfn[nid]) { 3715 *zone_end_pfn = zone_movable_pfn[nid]; 3716 3717 /* Check if this whole range is within ZONE_MOVABLE */ 3718 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3719 *zone_start_pfn = *zone_end_pfn; 3720 } 3721} 3722 3723/* 3724 * Return the number of pages a zone spans in a node, including holes 3725 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3726 */ 3727static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3728 unsigned long zone_type, 3729 unsigned long *ignored) 3730{ 3731 unsigned long node_start_pfn, node_end_pfn; 3732 unsigned long zone_start_pfn, zone_end_pfn; 3733 3734 /* Get the start and end of the node and zone */ 3735 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3736 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3737 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3738 adjust_zone_range_for_zone_movable(nid, zone_type, 3739 node_start_pfn, node_end_pfn, 3740 &zone_start_pfn, &zone_end_pfn); 3741 3742 /* Check that this node has pages within the zone's required range */ 3743 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3744 return 0; 3745 3746 /* Move the zone boundaries inside the node if necessary */ 3747 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3748 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3749 3750 /* Return the spanned pages */ 3751 return zone_end_pfn - zone_start_pfn; 3752} 3753 3754/* 3755 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3756 * then all holes in the requested range will be accounted for. 3757 */ 3758unsigned long __meminit __absent_pages_in_range(int nid, 3759 unsigned long range_start_pfn, 3760 unsigned long range_end_pfn) 3761{ 3762 int i = 0; 3763 unsigned long prev_end_pfn = 0, hole_pages = 0; 3764 unsigned long start_pfn; 3765 3766 /* Find the end_pfn of the first active range of pfns in the node */ 3767 i = first_active_region_index_in_nid(nid); 3768 if (i == -1) 3769 return 0; 3770 3771 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3772 3773 /* Account for ranges before physical memory on this node */ 3774 if (early_node_map[i].start_pfn > range_start_pfn) 3775 hole_pages = prev_end_pfn - range_start_pfn; 3776 3777 /* Find all holes for the zone within the node */ 3778 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3779 3780 /* No need to continue if prev_end_pfn is outside the zone */ 3781 if (prev_end_pfn >= range_end_pfn) 3782 break; 3783 3784 /* Make sure the end of the zone is not within the hole */ 3785 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3786 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3787 3788 /* Update the hole size cound and move on */ 3789 if (start_pfn > range_start_pfn) { 3790 BUG_ON(prev_end_pfn > start_pfn); 3791 hole_pages += start_pfn - prev_end_pfn; 3792 } 3793 prev_end_pfn = early_node_map[i].end_pfn; 3794 } 3795 3796 /* Account for ranges past physical memory on this node */ 3797 if (range_end_pfn > prev_end_pfn) 3798 hole_pages += range_end_pfn - 3799 max(range_start_pfn, prev_end_pfn); 3800 3801 return hole_pages; 3802} 3803 3804/** 3805 * absent_pages_in_range - Return number of page frames in holes within a range 3806 * @start_pfn: The start PFN to start searching for holes 3807 * @end_pfn: The end PFN to stop searching for holes 3808 * 3809 * It returns the number of pages frames in memory holes within a range. 3810 */ 3811unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3812 unsigned long end_pfn) 3813{ 3814 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3815} 3816 3817/* Return the number of page frames in holes in a zone on a node */ 3818static unsigned long __meminit zone_absent_pages_in_node(int nid, 3819 unsigned long zone_type, 3820 unsigned long *ignored) 3821{ 3822 unsigned long node_start_pfn, node_end_pfn; 3823 unsigned long zone_start_pfn, zone_end_pfn; 3824 3825 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3826 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3827 node_start_pfn); 3828 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3829 node_end_pfn); 3830 3831 adjust_zone_range_for_zone_movable(nid, zone_type, 3832 node_start_pfn, node_end_pfn, 3833 &zone_start_pfn, &zone_end_pfn); 3834 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3835} 3836 3837#else 3838static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3839 unsigned long zone_type, 3840 unsigned long *zones_size) 3841{ 3842 return zones_size[zone_type]; 3843} 3844 3845static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3846 unsigned long zone_type, 3847 unsigned long *zholes_size) 3848{ 3849 if (!zholes_size) 3850 return 0; 3851 3852 return zholes_size[zone_type]; 3853} 3854 3855#endif 3856 3857static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3858 unsigned long *zones_size, unsigned long *zholes_size) 3859{ 3860 unsigned long realtotalpages, totalpages = 0; 3861 enum zone_type i; 3862 3863 for (i = 0; i < MAX_NR_ZONES; i++) 3864 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3865 zones_size); 3866 pgdat->node_spanned_pages = totalpages; 3867 3868 realtotalpages = totalpages; 3869 for (i = 0; i < MAX_NR_ZONES; i++) 3870 realtotalpages -= 3871 zone_absent_pages_in_node(pgdat->node_id, i, 3872 zholes_size); 3873 pgdat->node_present_pages = realtotalpages; 3874 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3875 realtotalpages); 3876} 3877 3878#ifndef CONFIG_SPARSEMEM 3879/* 3880 * Calculate the size of the zone->blockflags rounded to an unsigned long 3881 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3882 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3883 * round what is now in bits to nearest long in bits, then return it in 3884 * bytes. 3885 */ 3886static unsigned long __init usemap_size(unsigned long zonesize) 3887{ 3888 unsigned long usemapsize; 3889 3890 usemapsize = roundup(zonesize, pageblock_nr_pages); 3891 usemapsize = usemapsize >> pageblock_order; 3892 usemapsize *= NR_PAGEBLOCK_BITS; 3893 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3894 3895 return usemapsize / 8; 3896} 3897 3898static void __init setup_usemap(struct pglist_data *pgdat, 3899 struct zone *zone, unsigned long zonesize) 3900{ 3901 unsigned long usemapsize = usemap_size(zonesize); 3902 zone->pageblock_flags = NULL; 3903 if (usemapsize) 3904 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3905} 3906#else 3907static void inline setup_usemap(struct pglist_data *pgdat, 3908 struct zone *zone, unsigned long zonesize) {} 3909#endif /* CONFIG_SPARSEMEM */ 3910 3911#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3912 3913/* Return a sensible default order for the pageblock size. */ 3914static inline int pageblock_default_order(void) 3915{ 3916 if (HPAGE_SHIFT > PAGE_SHIFT) 3917 return HUGETLB_PAGE_ORDER; 3918 3919 return MAX_ORDER-1; 3920} 3921 3922/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3923static inline void __init set_pageblock_order(unsigned int order) 3924{ 3925 /* Check that pageblock_nr_pages has not already been setup */ 3926 if (pageblock_order) 3927 return; 3928 3929 /* 3930 * Assume the largest contiguous order of interest is a huge page. 3931 * This value may be variable depending on boot parameters on IA64 3932 */ 3933 pageblock_order = order; 3934} 3935#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3936 3937/* 3938 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3939 * and pageblock_default_order() are unused as pageblock_order is set 3940 * at compile-time. See include/linux/pageblock-flags.h for the values of 3941 * pageblock_order based on the kernel config 3942 */ 3943static inline int pageblock_default_order(unsigned int order) 3944{ 3945 return MAX_ORDER-1; 3946} 3947#define set_pageblock_order(x) do {} while (0) 3948 3949#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3950 3951/* 3952 * Set up the zone data structures: 3953 * - mark all pages reserved 3954 * - mark all memory queues empty 3955 * - clear the memory bitmaps 3956 */ 3957static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3958 unsigned long *zones_size, unsigned long *zholes_size) 3959{ 3960 enum zone_type j; 3961 int nid = pgdat->node_id; 3962 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3963 int ret; 3964 3965 pgdat_resize_init(pgdat); 3966 pgdat->nr_zones = 0; 3967 init_waitqueue_head(&pgdat->kswapd_wait); 3968 pgdat->kswapd_max_order = 0; 3969 pgdat_page_cgroup_init(pgdat); 3970 3971 for (j = 0; j < MAX_NR_ZONES; j++) { 3972 struct zone *zone = pgdat->node_zones + j; 3973 unsigned long size, realsize, memmap_pages; 3974 enum lru_list l; 3975 3976 size = zone_spanned_pages_in_node(nid, j, zones_size); 3977 realsize = size - zone_absent_pages_in_node(nid, j, 3978 zholes_size); 3979 3980 /* 3981 * Adjust realsize so that it accounts for how much memory 3982 * is used by this zone for memmap. This affects the watermark 3983 * and per-cpu initialisations 3984 */ 3985 memmap_pages = 3986 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 3987 if (realsize >= memmap_pages) { 3988 realsize -= memmap_pages; 3989 if (memmap_pages) 3990 printk(KERN_DEBUG 3991 " %s zone: %lu pages used for memmap\n", 3992 zone_names[j], memmap_pages); 3993 } else 3994 printk(KERN_WARNING 3995 " %s zone: %lu pages exceeds realsize %lu\n", 3996 zone_names[j], memmap_pages, realsize); 3997 3998 /* Account for reserved pages */ 3999 if (j == 0 && realsize > dma_reserve) { 4000 realsize -= dma_reserve; 4001 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4002 zone_names[0], dma_reserve); 4003 } 4004 4005 if (!is_highmem_idx(j)) 4006 nr_kernel_pages += realsize; 4007 nr_all_pages += realsize; 4008 4009 zone->spanned_pages = size; 4010 zone->present_pages = realsize; 4011#ifdef CONFIG_NUMA 4012 zone->node = nid; 4013 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4014 / 100; 4015 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4016#endif 4017 zone->name = zone_names[j]; 4018 spin_lock_init(&zone->lock); 4019 spin_lock_init(&zone->lru_lock); 4020 zone_seqlock_init(zone); 4021 zone->zone_pgdat = pgdat; 4022 4023 zone->prev_priority = DEF_PRIORITY; 4024 4025 zone_pcp_init(zone); 4026 for_each_lru(l) { 4027 INIT_LIST_HEAD(&zone->lru[l].list); 4028 zone->reclaim_stat.nr_saved_scan[l] = 0; 4029 } 4030 zone->reclaim_stat.recent_rotated[0] = 0; 4031 zone->reclaim_stat.recent_rotated[1] = 0; 4032 zone->reclaim_stat.recent_scanned[0] = 0; 4033 zone->reclaim_stat.recent_scanned[1] = 0; 4034 zap_zone_vm_stats(zone); 4035 zone->flags = 0; 4036 if (!size) 4037 continue; 4038 4039 set_pageblock_order(pageblock_default_order()); 4040 setup_usemap(pgdat, zone, size); 4041 ret = init_currently_empty_zone(zone, zone_start_pfn, 4042 size, MEMMAP_EARLY); 4043 BUG_ON(ret); 4044 memmap_init(size, nid, j, zone_start_pfn); 4045 zone_start_pfn += size; 4046 } 4047} 4048 4049static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4050{ 4051 /* Skip empty nodes */ 4052 if (!pgdat->node_spanned_pages) 4053 return; 4054 4055#ifdef CONFIG_FLAT_NODE_MEM_MAP 4056 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4057 if (!pgdat->node_mem_map) { 4058 unsigned long size, start, end; 4059 struct page *map; 4060 4061 /* 4062 * The zone's endpoints aren't required to be MAX_ORDER 4063 * aligned but the node_mem_map endpoints must be in order 4064 * for the buddy allocator to function correctly. 4065 */ 4066 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4067 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4068 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4069 size = (end - start) * sizeof(struct page); 4070 map = alloc_remap(pgdat->node_id, size); 4071 if (!map) 4072 map = alloc_bootmem_node(pgdat, size); 4073 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4074 } 4075#ifndef CONFIG_NEED_MULTIPLE_NODES 4076 /* 4077 * With no DISCONTIG, the global mem_map is just set as node 0's 4078 */ 4079 if (pgdat == NODE_DATA(0)) { 4080 mem_map = NODE_DATA(0)->node_mem_map; 4081#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4082 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4083 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4084#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4085 } 4086#endif 4087#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4088} 4089 4090void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4091 unsigned long node_start_pfn, unsigned long *zholes_size) 4092{ 4093 pg_data_t *pgdat = NODE_DATA(nid); 4094 4095 pgdat->node_id = nid; 4096 pgdat->node_start_pfn = node_start_pfn; 4097 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4098 4099 alloc_node_mem_map(pgdat); 4100#ifdef CONFIG_FLAT_NODE_MEM_MAP 4101 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4102 nid, (unsigned long)pgdat, 4103 (unsigned long)pgdat->node_mem_map); 4104#endif 4105 4106 free_area_init_core(pgdat, zones_size, zholes_size); 4107} 4108 4109#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4110 4111#if MAX_NUMNODES > 1 4112/* 4113 * Figure out the number of possible node ids. 4114 */ 4115static void __init setup_nr_node_ids(void) 4116{ 4117 unsigned int node; 4118 unsigned int highest = 0; 4119 4120 for_each_node_mask(node, node_possible_map) 4121 highest = node; 4122 nr_node_ids = highest + 1; 4123} 4124#else 4125static inline void setup_nr_node_ids(void) 4126{ 4127} 4128#endif 4129 4130/** 4131 * add_active_range - Register a range of PFNs backed by physical memory 4132 * @nid: The node ID the range resides on 4133 * @start_pfn: The start PFN of the available physical memory 4134 * @end_pfn: The end PFN of the available physical memory 4135 * 4136 * These ranges are stored in an early_node_map[] and later used by 4137 * free_area_init_nodes() to calculate zone sizes and holes. If the 4138 * range spans a memory hole, it is up to the architecture to ensure 4139 * the memory is not freed by the bootmem allocator. If possible 4140 * the range being registered will be merged with existing ranges. 4141 */ 4142void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4143 unsigned long end_pfn) 4144{ 4145 int i; 4146 4147 mminit_dprintk(MMINIT_TRACE, "memory_register", 4148 "Entering add_active_range(%d, %#lx, %#lx) " 4149 "%d entries of %d used\n", 4150 nid, start_pfn, end_pfn, 4151 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4152 4153 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4154 4155 /* Merge with existing active regions if possible */ 4156 for (i = 0; i < nr_nodemap_entries; i++) { 4157 if (early_node_map[i].nid != nid) 4158 continue; 4159 4160 /* Skip if an existing region covers this new one */ 4161 if (start_pfn >= early_node_map[i].start_pfn && 4162 end_pfn <= early_node_map[i].end_pfn) 4163 return; 4164 4165 /* Merge forward if suitable */ 4166 if (start_pfn <= early_node_map[i].end_pfn && 4167 end_pfn > early_node_map[i].end_pfn) { 4168 early_node_map[i].end_pfn = end_pfn; 4169 return; 4170 } 4171 4172 /* Merge backward if suitable */ 4173 if (start_pfn < early_node_map[i].start_pfn && 4174 end_pfn >= early_node_map[i].start_pfn) { 4175 early_node_map[i].start_pfn = start_pfn; 4176 return; 4177 } 4178 } 4179 4180 /* Check that early_node_map is large enough */ 4181 if (i >= MAX_ACTIVE_REGIONS) { 4182 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4183 MAX_ACTIVE_REGIONS); 4184 return; 4185 } 4186 4187 early_node_map[i].nid = nid; 4188 early_node_map[i].start_pfn = start_pfn; 4189 early_node_map[i].end_pfn = end_pfn; 4190 nr_nodemap_entries = i + 1; 4191} 4192 4193/** 4194 * remove_active_range - Shrink an existing registered range of PFNs 4195 * @nid: The node id the range is on that should be shrunk 4196 * @start_pfn: The new PFN of the range 4197 * @end_pfn: The new PFN of the range 4198 * 4199 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4200 * The map is kept near the end physical page range that has already been 4201 * registered. This function allows an arch to shrink an existing registered 4202 * range. 4203 */ 4204void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4205 unsigned long end_pfn) 4206{ 4207 int i, j; 4208 int removed = 0; 4209 4210 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4211 nid, start_pfn, end_pfn); 4212 4213 /* Find the old active region end and shrink */ 4214 for_each_active_range_index_in_nid(i, nid) { 4215 if (early_node_map[i].start_pfn >= start_pfn && 4216 early_node_map[i].end_pfn <= end_pfn) { 4217 /* clear it */ 4218 early_node_map[i].start_pfn = 0; 4219 early_node_map[i].end_pfn = 0; 4220 removed = 1; 4221 continue; 4222 } 4223 if (early_node_map[i].start_pfn < start_pfn && 4224 early_node_map[i].end_pfn > start_pfn) { 4225 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4226 early_node_map[i].end_pfn = start_pfn; 4227 if (temp_end_pfn > end_pfn) 4228 add_active_range(nid, end_pfn, temp_end_pfn); 4229 continue; 4230 } 4231 if (early_node_map[i].start_pfn >= start_pfn && 4232 early_node_map[i].end_pfn > end_pfn && 4233 early_node_map[i].start_pfn < end_pfn) { 4234 early_node_map[i].start_pfn = end_pfn; 4235 continue; 4236 } 4237 } 4238 4239 if (!removed) 4240 return; 4241 4242 /* remove the blank ones */ 4243 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4244 if (early_node_map[i].nid != nid) 4245 continue; 4246 if (early_node_map[i].end_pfn) 4247 continue; 4248 /* we found it, get rid of it */ 4249 for (j = i; j < nr_nodemap_entries - 1; j++) 4250 memcpy(&early_node_map[j], &early_node_map[j+1], 4251 sizeof(early_node_map[j])); 4252 j = nr_nodemap_entries - 1; 4253 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4254 nr_nodemap_entries--; 4255 } 4256} 4257 4258/** 4259 * remove_all_active_ranges - Remove all currently registered regions 4260 * 4261 * During discovery, it may be found that a table like SRAT is invalid 4262 * and an alternative discovery method must be used. This function removes 4263 * all currently registered regions. 4264 */ 4265void __init remove_all_active_ranges(void) 4266{ 4267 memset(early_node_map, 0, sizeof(early_node_map)); 4268 nr_nodemap_entries = 0; 4269} 4270 4271/* Compare two active node_active_regions */ 4272static int __init cmp_node_active_region(const void *a, const void *b) 4273{ 4274 struct node_active_region *arange = (struct node_active_region *)a; 4275 struct node_active_region *brange = (struct node_active_region *)b; 4276 4277 /* Done this way to avoid overflows */ 4278 if (arange->start_pfn > brange->start_pfn) 4279 return 1; 4280 if (arange->start_pfn < brange->start_pfn) 4281 return -1; 4282 4283 return 0; 4284} 4285 4286/* sort the node_map by start_pfn */ 4287void __init sort_node_map(void) 4288{ 4289 sort(early_node_map, (size_t)nr_nodemap_entries, 4290 sizeof(struct node_active_region), 4291 cmp_node_active_region, NULL); 4292} 4293 4294/* Find the lowest pfn for a node */ 4295static unsigned long __init find_min_pfn_for_node(int nid) 4296{ 4297 int i; 4298 unsigned long min_pfn = ULONG_MAX; 4299 4300 /* Assuming a sorted map, the first range found has the starting pfn */ 4301 for_each_active_range_index_in_nid(i, nid) 4302 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4303 4304 if (min_pfn == ULONG_MAX) { 4305 printk(KERN_WARNING 4306 "Could not find start_pfn for node %d\n", nid); 4307 return 0; 4308 } 4309 4310 return min_pfn; 4311} 4312 4313/** 4314 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4315 * 4316 * It returns the minimum PFN based on information provided via 4317 * add_active_range(). 4318 */ 4319unsigned long __init find_min_pfn_with_active_regions(void) 4320{ 4321 return find_min_pfn_for_node(MAX_NUMNODES); 4322} 4323 4324/* 4325 * early_calculate_totalpages() 4326 * Sum pages in active regions for movable zone. 4327 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4328 */ 4329static unsigned long __init early_calculate_totalpages(void) 4330{ 4331 int i; 4332 unsigned long totalpages = 0; 4333 4334 for (i = 0; i < nr_nodemap_entries; i++) { 4335 unsigned long pages = early_node_map[i].end_pfn - 4336 early_node_map[i].start_pfn; 4337 totalpages += pages; 4338 if (pages) 4339 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4340 } 4341 return totalpages; 4342} 4343 4344/* 4345 * Find the PFN the Movable zone begins in each node. Kernel memory 4346 * is spread evenly between nodes as long as the nodes have enough 4347 * memory. When they don't, some nodes will have more kernelcore than 4348 * others 4349 */ 4350static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4351{ 4352 int i, nid; 4353 unsigned long usable_startpfn; 4354 unsigned long kernelcore_node, kernelcore_remaining; 4355 /* save the state before borrow the nodemask */ 4356 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4357 unsigned long totalpages = early_calculate_totalpages(); 4358 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4359 4360 /* 4361 * If movablecore was specified, calculate what size of 4362 * kernelcore that corresponds so that memory usable for 4363 * any allocation type is evenly spread. If both kernelcore 4364 * and movablecore are specified, then the value of kernelcore 4365 * will be used for required_kernelcore if it's greater than 4366 * what movablecore would have allowed. 4367 */ 4368 if (required_movablecore) { 4369 unsigned long corepages; 4370 4371 /* 4372 * Round-up so that ZONE_MOVABLE is at least as large as what 4373 * was requested by the user 4374 */ 4375 required_movablecore = 4376 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4377 corepages = totalpages - required_movablecore; 4378 4379 required_kernelcore = max(required_kernelcore, corepages); 4380 } 4381 4382 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4383 if (!required_kernelcore) 4384 goto out; 4385 4386 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4387 find_usable_zone_for_movable(); 4388 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4389 4390restart: 4391 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4392 kernelcore_node = required_kernelcore / usable_nodes; 4393 for_each_node_state(nid, N_HIGH_MEMORY) { 4394 /* 4395 * Recalculate kernelcore_node if the division per node 4396 * now exceeds what is necessary to satisfy the requested 4397 * amount of memory for the kernel 4398 */ 4399 if (required_kernelcore < kernelcore_node) 4400 kernelcore_node = required_kernelcore / usable_nodes; 4401 4402 /* 4403 * As the map is walked, we track how much memory is usable 4404 * by the kernel using kernelcore_remaining. When it is 4405 * 0, the rest of the node is usable by ZONE_MOVABLE 4406 */ 4407 kernelcore_remaining = kernelcore_node; 4408 4409 /* Go through each range of PFNs within this node */ 4410 for_each_active_range_index_in_nid(i, nid) { 4411 unsigned long start_pfn, end_pfn; 4412 unsigned long size_pages; 4413 4414 start_pfn = max(early_node_map[i].start_pfn, 4415 zone_movable_pfn[nid]); 4416 end_pfn = early_node_map[i].end_pfn; 4417 if (start_pfn >= end_pfn) 4418 continue; 4419 4420 /* Account for what is only usable for kernelcore */ 4421 if (start_pfn < usable_startpfn) { 4422 unsigned long kernel_pages; 4423 kernel_pages = min(end_pfn, usable_startpfn) 4424 - start_pfn; 4425 4426 kernelcore_remaining -= min(kernel_pages, 4427 kernelcore_remaining); 4428 required_kernelcore -= min(kernel_pages, 4429 required_kernelcore); 4430 4431 /* Continue if range is now fully accounted */ 4432 if (end_pfn <= usable_startpfn) { 4433 4434 /* 4435 * Push zone_movable_pfn to the end so 4436 * that if we have to rebalance 4437 * kernelcore across nodes, we will 4438 * not double account here 4439 */ 4440 zone_movable_pfn[nid] = end_pfn; 4441 continue; 4442 } 4443 start_pfn = usable_startpfn; 4444 } 4445 4446 /* 4447 * The usable PFN range for ZONE_MOVABLE is from 4448 * start_pfn->end_pfn. Calculate size_pages as the 4449 * number of pages used as kernelcore 4450 */ 4451 size_pages = end_pfn - start_pfn; 4452 if (size_pages > kernelcore_remaining) 4453 size_pages = kernelcore_remaining; 4454 zone_movable_pfn[nid] = start_pfn + size_pages; 4455 4456 /* 4457 * Some kernelcore has been met, update counts and 4458 * break if the kernelcore for this node has been 4459 * satisified 4460 */ 4461 required_kernelcore -= min(required_kernelcore, 4462 size_pages); 4463 kernelcore_remaining -= size_pages; 4464 if (!kernelcore_remaining) 4465 break; 4466 } 4467 } 4468 4469 /* 4470 * If there is still required_kernelcore, we do another pass with one 4471 * less node in the count. This will push zone_movable_pfn[nid] further 4472 * along on the nodes that still have memory until kernelcore is 4473 * satisified 4474 */ 4475 usable_nodes--; 4476 if (usable_nodes && required_kernelcore > usable_nodes) 4477 goto restart; 4478 4479 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4480 for (nid = 0; nid < MAX_NUMNODES; nid++) 4481 zone_movable_pfn[nid] = 4482 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4483 4484out: 4485 /* restore the node_state */ 4486 node_states[N_HIGH_MEMORY] = saved_node_state; 4487} 4488 4489/* Any regular memory on that node ? */ 4490static void check_for_regular_memory(pg_data_t *pgdat) 4491{ 4492#ifdef CONFIG_HIGHMEM 4493 enum zone_type zone_type; 4494 4495 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4496 struct zone *zone = &pgdat->node_zones[zone_type]; 4497 if (zone->present_pages) 4498 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4499 } 4500#endif 4501} 4502 4503/** 4504 * free_area_init_nodes - Initialise all pg_data_t and zone data 4505 * @max_zone_pfn: an array of max PFNs for each zone 4506 * 4507 * This will call free_area_init_node() for each active node in the system. 4508 * Using the page ranges provided by add_active_range(), the size of each 4509 * zone in each node and their holes is calculated. If the maximum PFN 4510 * between two adjacent zones match, it is assumed that the zone is empty. 4511 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4512 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4513 * starts where the previous one ended. For example, ZONE_DMA32 starts 4514 * at arch_max_dma_pfn. 4515 */ 4516void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4517{ 4518 unsigned long nid; 4519 int i; 4520 4521 /* Sort early_node_map as initialisation assumes it is sorted */ 4522 sort_node_map(); 4523 4524 /* Record where the zone boundaries are */ 4525 memset(arch_zone_lowest_possible_pfn, 0, 4526 sizeof(arch_zone_lowest_possible_pfn)); 4527 memset(arch_zone_highest_possible_pfn, 0, 4528 sizeof(arch_zone_highest_possible_pfn)); 4529 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4530 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4531 for (i = 1; i < MAX_NR_ZONES; i++) { 4532 if (i == ZONE_MOVABLE) 4533 continue; 4534 arch_zone_lowest_possible_pfn[i] = 4535 arch_zone_highest_possible_pfn[i-1]; 4536 arch_zone_highest_possible_pfn[i] = 4537 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4538 } 4539 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4540 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4541 4542 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4543 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4544 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4545 4546 /* Print out the zone ranges */ 4547 printk("Zone PFN ranges:\n"); 4548 for (i = 0; i < MAX_NR_ZONES; i++) { 4549 if (i == ZONE_MOVABLE) 4550 continue; 4551 printk(" %-8s ", zone_names[i]); 4552 if (arch_zone_lowest_possible_pfn[i] == 4553 arch_zone_highest_possible_pfn[i]) 4554 printk("empty\n"); 4555 else 4556 printk("%0#10lx -> %0#10lx\n", 4557 arch_zone_lowest_possible_pfn[i], 4558 arch_zone_highest_possible_pfn[i]); 4559 } 4560 4561 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4562 printk("Movable zone start PFN for each node\n"); 4563 for (i = 0; i < MAX_NUMNODES; i++) { 4564 if (zone_movable_pfn[i]) 4565 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4566 } 4567 4568 /* Print out the early_node_map[] */ 4569 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4570 for (i = 0; i < nr_nodemap_entries; i++) 4571 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4572 early_node_map[i].start_pfn, 4573 early_node_map[i].end_pfn); 4574 4575 /* Initialise every node */ 4576 mminit_verify_pageflags_layout(); 4577 setup_nr_node_ids(); 4578 for_each_online_node(nid) { 4579 pg_data_t *pgdat = NODE_DATA(nid); 4580 free_area_init_node(nid, NULL, 4581 find_min_pfn_for_node(nid), NULL); 4582 4583 /* Any memory on that node */ 4584 if (pgdat->node_present_pages) 4585 node_set_state(nid, N_HIGH_MEMORY); 4586 check_for_regular_memory(pgdat); 4587 } 4588} 4589 4590static int __init cmdline_parse_core(char *p, unsigned long *core) 4591{ 4592 unsigned long long coremem; 4593 if (!p) 4594 return -EINVAL; 4595 4596 coremem = memparse(p, &p); 4597 *core = coremem >> PAGE_SHIFT; 4598 4599 /* Paranoid check that UL is enough for the coremem value */ 4600 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4601 4602 return 0; 4603} 4604 4605/* 4606 * kernelcore=size sets the amount of memory for use for allocations that 4607 * cannot be reclaimed or migrated. 4608 */ 4609static int __init cmdline_parse_kernelcore(char *p) 4610{ 4611 return cmdline_parse_core(p, &required_kernelcore); 4612} 4613 4614/* 4615 * movablecore=size sets the amount of memory for use for allocations that 4616 * can be reclaimed or migrated. 4617 */ 4618static int __init cmdline_parse_movablecore(char *p) 4619{ 4620 return cmdline_parse_core(p, &required_movablecore); 4621} 4622 4623early_param("kernelcore", cmdline_parse_kernelcore); 4624early_param("movablecore", cmdline_parse_movablecore); 4625 4626#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4627 4628/** 4629 * set_dma_reserve - set the specified number of pages reserved in the first zone 4630 * @new_dma_reserve: The number of pages to mark reserved 4631 * 4632 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4633 * In the DMA zone, a significant percentage may be consumed by kernel image 4634 * and other unfreeable allocations which can skew the watermarks badly. This 4635 * function may optionally be used to account for unfreeable pages in the 4636 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4637 * smaller per-cpu batchsize. 4638 */ 4639void __init set_dma_reserve(unsigned long new_dma_reserve) 4640{ 4641 dma_reserve = new_dma_reserve; 4642} 4643 4644#ifndef CONFIG_NEED_MULTIPLE_NODES 4645struct pglist_data __refdata contig_page_data = { 4646#ifndef CONFIG_NO_BOOTMEM 4647 .bdata = &bootmem_node_data[0] 4648#endif 4649 }; 4650EXPORT_SYMBOL(contig_page_data); 4651#endif 4652 4653void __init free_area_init(unsigned long *zones_size) 4654{ 4655 free_area_init_node(0, zones_size, 4656 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4657} 4658 4659static int page_alloc_cpu_notify(struct notifier_block *self, 4660 unsigned long action, void *hcpu) 4661{ 4662 int cpu = (unsigned long)hcpu; 4663 4664 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4665 drain_pages(cpu); 4666 4667 /* 4668 * Spill the event counters of the dead processor 4669 * into the current processors event counters. 4670 * This artificially elevates the count of the current 4671 * processor. 4672 */ 4673 vm_events_fold_cpu(cpu); 4674 4675 /* 4676 * Zero the differential counters of the dead processor 4677 * so that the vm statistics are consistent. 4678 * 4679 * This is only okay since the processor is dead and cannot 4680 * race with what we are doing. 4681 */ 4682 refresh_cpu_vm_stats(cpu); 4683 } 4684 return NOTIFY_OK; 4685} 4686 4687void __init page_alloc_init(void) 4688{ 4689 hotcpu_notifier(page_alloc_cpu_notify, 0); 4690} 4691 4692/* 4693 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4694 * or min_free_kbytes changes. 4695 */ 4696static void calculate_totalreserve_pages(void) 4697{ 4698 struct pglist_data *pgdat; 4699 unsigned long reserve_pages = 0; 4700 enum zone_type i, j; 4701 4702 for_each_online_pgdat(pgdat) { 4703 for (i = 0; i < MAX_NR_ZONES; i++) { 4704 struct zone *zone = pgdat->node_zones + i; 4705 unsigned long max = 0; 4706 4707 /* Find valid and maximum lowmem_reserve in the zone */ 4708 for (j = i; j < MAX_NR_ZONES; j++) { 4709 if (zone->lowmem_reserve[j] > max) 4710 max = zone->lowmem_reserve[j]; 4711 } 4712 4713 /* we treat the high watermark as reserved pages. */ 4714 max += high_wmark_pages(zone); 4715 4716 if (max > zone->present_pages) 4717 max = zone->present_pages; 4718 reserve_pages += max; 4719 } 4720 } 4721 totalreserve_pages = reserve_pages; 4722} 4723 4724/* 4725 * setup_per_zone_lowmem_reserve - called whenever 4726 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4727 * has a correct pages reserved value, so an adequate number of 4728 * pages are left in the zone after a successful __alloc_pages(). 4729 */ 4730static void setup_per_zone_lowmem_reserve(void) 4731{ 4732 struct pglist_data *pgdat; 4733 enum zone_type j, idx; 4734 4735 for_each_online_pgdat(pgdat) { 4736 for (j = 0; j < MAX_NR_ZONES; j++) { 4737 struct zone *zone = pgdat->node_zones + j; 4738 unsigned long present_pages = zone->present_pages; 4739 4740 zone->lowmem_reserve[j] = 0; 4741 4742 idx = j; 4743 while (idx) { 4744 struct zone *lower_zone; 4745 4746 idx--; 4747 4748 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4749 sysctl_lowmem_reserve_ratio[idx] = 1; 4750 4751 lower_zone = pgdat->node_zones + idx; 4752 lower_zone->lowmem_reserve[j] = present_pages / 4753 sysctl_lowmem_reserve_ratio[idx]; 4754 present_pages += lower_zone->present_pages; 4755 } 4756 } 4757 } 4758 4759 /* update totalreserve_pages */ 4760 calculate_totalreserve_pages(); 4761} 4762 4763/** 4764 * setup_per_zone_wmarks - called when min_free_kbytes changes 4765 * or when memory is hot-{added|removed} 4766 * 4767 * Ensures that the watermark[min,low,high] values for each zone are set 4768 * correctly with respect to min_free_kbytes. 4769 */ 4770void setup_per_zone_wmarks(void) 4771{ 4772 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4773 unsigned long lowmem_pages = 0; 4774 struct zone *zone; 4775 unsigned long flags; 4776 4777 /* Calculate total number of !ZONE_HIGHMEM pages */ 4778 for_each_zone(zone) { 4779 if (!is_highmem(zone)) 4780 lowmem_pages += zone->present_pages; 4781 } 4782 4783 for_each_zone(zone) { 4784 u64 tmp; 4785 4786 spin_lock_irqsave(&zone->lock, flags); 4787 tmp = (u64)pages_min * zone->present_pages; 4788 do_div(tmp, lowmem_pages); 4789 if (is_highmem(zone)) { 4790 /* 4791 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4792 * need highmem pages, so cap pages_min to a small 4793 * value here. 4794 * 4795 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4796 * deltas controls asynch page reclaim, and so should 4797 * not be capped for highmem. 4798 */ 4799 int min_pages; 4800 4801 min_pages = zone->present_pages / 1024; 4802 if (min_pages < SWAP_CLUSTER_MAX) 4803 min_pages = SWAP_CLUSTER_MAX; 4804 if (min_pages > 128) 4805 min_pages = 128; 4806 zone->watermark[WMARK_MIN] = min_pages; 4807 } else { 4808 /* 4809 * If it's a lowmem zone, reserve a number of pages 4810 * proportionate to the zone's size. 4811 */ 4812 zone->watermark[WMARK_MIN] = tmp; 4813 } 4814 4815 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4816 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4817 setup_zone_migrate_reserve(zone); 4818 spin_unlock_irqrestore(&zone->lock, flags); 4819 } 4820 4821 /* update totalreserve_pages */ 4822 calculate_totalreserve_pages(); 4823} 4824 4825/* 4826 * The inactive anon list should be small enough that the VM never has to 4827 * do too much work, but large enough that each inactive page has a chance 4828 * to be referenced again before it is swapped out. 4829 * 4830 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 4831 * INACTIVE_ANON pages on this zone's LRU, maintained by the 4832 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 4833 * the anonymous pages are kept on the inactive list. 4834 * 4835 * total target max 4836 * memory ratio inactive anon 4837 * ------------------------------------- 4838 * 10MB 1 5MB 4839 * 100MB 1 50MB 4840 * 1GB 3 250MB 4841 * 10GB 10 0.9GB 4842 * 100GB 31 3GB 4843 * 1TB 101 10GB 4844 * 10TB 320 32GB 4845 */ 4846void calculate_zone_inactive_ratio(struct zone *zone) 4847{ 4848 unsigned int gb, ratio; 4849 4850 /* Zone size in gigabytes */ 4851 gb = zone->present_pages >> (30 - PAGE_SHIFT); 4852 if (gb) 4853 ratio = int_sqrt(10 * gb); 4854 else 4855 ratio = 1; 4856 4857 zone->inactive_ratio = ratio; 4858} 4859 4860static void __init setup_per_zone_inactive_ratio(void) 4861{ 4862 struct zone *zone; 4863 4864 for_each_zone(zone) 4865 calculate_zone_inactive_ratio(zone); 4866} 4867 4868/* 4869 * Initialise min_free_kbytes. 4870 * 4871 * For small machines we want it small (128k min). For large machines 4872 * we want it large (64MB max). But it is not linear, because network 4873 * bandwidth does not increase linearly with machine size. We use 4874 * 4875 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4876 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4877 * 4878 * which yields 4879 * 4880 * 16MB: 512k 4881 * 32MB: 724k 4882 * 64MB: 1024k 4883 * 128MB: 1448k 4884 * 256MB: 2048k 4885 * 512MB: 2896k 4886 * 1024MB: 4096k 4887 * 2048MB: 5792k 4888 * 4096MB: 8192k 4889 * 8192MB: 11584k 4890 * 16384MB: 16384k 4891 */ 4892static int __init init_per_zone_wmark_min(void) 4893{ 4894 unsigned long lowmem_kbytes; 4895 4896 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4897 4898 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4899 if (min_free_kbytes < 128) 4900 min_free_kbytes = 128; 4901 if (min_free_kbytes > 65536) 4902 min_free_kbytes = 65536; 4903 setup_per_zone_wmarks(); 4904 setup_per_zone_lowmem_reserve(); 4905 setup_per_zone_inactive_ratio(); 4906 return 0; 4907} 4908module_init(init_per_zone_wmark_min) 4909 4910/* 4911 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4912 * that we can call two helper functions whenever min_free_kbytes 4913 * changes. 4914 */ 4915int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4916 void __user *buffer, size_t *length, loff_t *ppos) 4917{ 4918 proc_dointvec(table, write, buffer, length, ppos); 4919 if (write) 4920 setup_per_zone_wmarks(); 4921 return 0; 4922} 4923 4924#ifdef CONFIG_NUMA 4925int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4926 void __user *buffer, size_t *length, loff_t *ppos) 4927{ 4928 struct zone *zone; 4929 int rc; 4930 4931 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4932 if (rc) 4933 return rc; 4934 4935 for_each_zone(zone) 4936 zone->min_unmapped_pages = (zone->present_pages * 4937 sysctl_min_unmapped_ratio) / 100; 4938 return 0; 4939} 4940 4941int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4942 void __user *buffer, size_t *length, loff_t *ppos) 4943{ 4944 struct zone *zone; 4945 int rc; 4946 4947 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 4948 if (rc) 4949 return rc; 4950 4951 for_each_zone(zone) 4952 zone->min_slab_pages = (zone->present_pages * 4953 sysctl_min_slab_ratio) / 100; 4954 return 0; 4955} 4956#endif 4957 4958/* 4959 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4960 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4961 * whenever sysctl_lowmem_reserve_ratio changes. 4962 * 4963 * The reserve ratio obviously has absolutely no relation with the 4964 * minimum watermarks. The lowmem reserve ratio can only make sense 4965 * if in function of the boot time zone sizes. 4966 */ 4967int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4968 void __user *buffer, size_t *length, loff_t *ppos) 4969{ 4970 proc_dointvec_minmax(table, write, buffer, length, ppos); 4971 setup_per_zone_lowmem_reserve(); 4972 return 0; 4973} 4974 4975/* 4976 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4977 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4978 * can have before it gets flushed back to buddy allocator. 4979 */ 4980 4981int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4982 void __user *buffer, size_t *length, loff_t *ppos) 4983{ 4984 struct zone *zone; 4985 unsigned int cpu; 4986 int ret; 4987 4988 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 4989 if (!write || (ret == -EINVAL)) 4990 return ret; 4991 for_each_populated_zone(zone) { 4992 for_each_possible_cpu(cpu) { 4993 unsigned long high; 4994 high = zone->present_pages / percpu_pagelist_fraction; 4995 setup_pagelist_highmark( 4996 per_cpu_ptr(zone->pageset, cpu), high); 4997 } 4998 } 4999 return 0; 5000} 5001 5002int hashdist = HASHDIST_DEFAULT; 5003 5004#ifdef CONFIG_NUMA 5005static int __init set_hashdist(char *str) 5006{ 5007 if (!str) 5008 return 0; 5009 hashdist = simple_strtoul(str, &str, 0); 5010 return 1; 5011} 5012__setup("hashdist=", set_hashdist); 5013#endif 5014 5015/* 5016 * allocate a large system hash table from bootmem 5017 * - it is assumed that the hash table must contain an exact power-of-2 5018 * quantity of entries 5019 * - limit is the number of hash buckets, not the total allocation size 5020 */ 5021void *__init alloc_large_system_hash(const char *tablename, 5022 unsigned long bucketsize, 5023 unsigned long numentries, 5024 int scale, 5025 int flags, 5026 unsigned int *_hash_shift, 5027 unsigned int *_hash_mask, 5028 unsigned long limit) 5029{ 5030 unsigned long long max = limit; 5031 unsigned long log2qty, size; 5032 void *table = NULL; 5033 5034 /* allow the kernel cmdline to have a say */ 5035 if (!numentries) { 5036 /* round applicable memory size up to nearest megabyte */ 5037 numentries = nr_kernel_pages; 5038 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5039 numentries >>= 20 - PAGE_SHIFT; 5040 numentries <<= 20 - PAGE_SHIFT; 5041 5042 /* limit to 1 bucket per 2^scale bytes of low memory */ 5043 if (scale > PAGE_SHIFT) 5044 numentries >>= (scale - PAGE_SHIFT); 5045 else 5046 numentries <<= (PAGE_SHIFT - scale); 5047 5048 /* Make sure we've got at least a 0-order allocation.. */ 5049 if (unlikely(flags & HASH_SMALL)) { 5050 /* Makes no sense without HASH_EARLY */ 5051 WARN_ON(!(flags & HASH_EARLY)); 5052 if (!(numentries >> *_hash_shift)) { 5053 numentries = 1UL << *_hash_shift; 5054 BUG_ON(!numentries); 5055 } 5056 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5057 numentries = PAGE_SIZE / bucketsize; 5058 } 5059 numentries = roundup_pow_of_two(numentries); 5060 5061 /* limit allocation size to 1/16 total memory by default */ 5062 if (max == 0) { 5063 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5064 do_div(max, bucketsize); 5065 } 5066 5067 if (numentries > max) 5068 numentries = max; 5069 5070 log2qty = ilog2(numentries); 5071 5072 do { 5073 size = bucketsize << log2qty; 5074 if (flags & HASH_EARLY) 5075 table = alloc_bootmem_nopanic(size); 5076 else if (hashdist) 5077 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5078 else { 5079 /* 5080 * If bucketsize is not a power-of-two, we may free 5081 * some pages at the end of hash table which 5082 * alloc_pages_exact() automatically does 5083 */ 5084 if (get_order(size) < MAX_ORDER) { 5085 table = alloc_pages_exact(size, GFP_ATOMIC); 5086 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5087 } 5088 } 5089 } while (!table && size > PAGE_SIZE && --log2qty); 5090 5091 if (!table) 5092 panic("Failed to allocate %s hash table\n", tablename); 5093 5094 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 5095 tablename, 5096 (1U << log2qty), 5097 ilog2(size) - PAGE_SHIFT, 5098 size); 5099 5100 if (_hash_shift) 5101 *_hash_shift = log2qty; 5102 if (_hash_mask) 5103 *_hash_mask = (1 << log2qty) - 1; 5104 5105 return table; 5106} 5107 5108/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5109static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5110 unsigned long pfn) 5111{ 5112#ifdef CONFIG_SPARSEMEM 5113 return __pfn_to_section(pfn)->pageblock_flags; 5114#else 5115 return zone->pageblock_flags; 5116#endif /* CONFIG_SPARSEMEM */ 5117} 5118 5119static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5120{ 5121#ifdef CONFIG_SPARSEMEM 5122 pfn &= (PAGES_PER_SECTION-1); 5123 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5124#else 5125 pfn = pfn - zone->zone_start_pfn; 5126 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5127#endif /* CONFIG_SPARSEMEM */ 5128} 5129 5130/** 5131 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5132 * @page: The page within the block of interest 5133 * @start_bitidx: The first bit of interest to retrieve 5134 * @end_bitidx: The last bit of interest 5135 * returns pageblock_bits flags 5136 */ 5137unsigned long get_pageblock_flags_group(struct page *page, 5138 int start_bitidx, int end_bitidx) 5139{ 5140 struct zone *zone; 5141 unsigned long *bitmap; 5142 unsigned long pfn, bitidx; 5143 unsigned long flags = 0; 5144 unsigned long value = 1; 5145 5146 zone = page_zone(page); 5147 pfn = page_to_pfn(page); 5148 bitmap = get_pageblock_bitmap(zone, pfn); 5149 bitidx = pfn_to_bitidx(zone, pfn); 5150 5151 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5152 if (test_bit(bitidx + start_bitidx, bitmap)) 5153 flags |= value; 5154 5155 return flags; 5156} 5157 5158/** 5159 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5160 * @page: The page within the block of interest 5161 * @start_bitidx: The first bit of interest 5162 * @end_bitidx: The last bit of interest 5163 * @flags: The flags to set 5164 */ 5165void set_pageblock_flags_group(struct page *page, unsigned long flags, 5166 int start_bitidx, int end_bitidx) 5167{ 5168 struct zone *zone; 5169 unsigned long *bitmap; 5170 unsigned long pfn, bitidx; 5171 unsigned long value = 1; 5172 5173 zone = page_zone(page); 5174 pfn = page_to_pfn(page); 5175 bitmap = get_pageblock_bitmap(zone, pfn); 5176 bitidx = pfn_to_bitidx(zone, pfn); 5177 VM_BUG_ON(pfn < zone->zone_start_pfn); 5178 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5179 5180 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5181 if (flags & value) 5182 __set_bit(bitidx + start_bitidx, bitmap); 5183 else 5184 __clear_bit(bitidx + start_bitidx, bitmap); 5185} 5186 5187/* 5188 * This is designed as sub function...plz see page_isolation.c also. 5189 * set/clear page block's type to be ISOLATE. 5190 * page allocater never alloc memory from ISOLATE block. 5191 */ 5192 5193int set_migratetype_isolate(struct page *page) 5194{ 5195 struct zone *zone; 5196 struct page *curr_page; 5197 unsigned long flags, pfn, iter; 5198 unsigned long immobile = 0; 5199 struct memory_isolate_notify arg; 5200 int notifier_ret; 5201 int ret = -EBUSY; 5202 int zone_idx; 5203 5204 zone = page_zone(page); 5205 zone_idx = zone_idx(zone); 5206 5207 spin_lock_irqsave(&zone->lock, flags); 5208 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE || 5209 zone_idx == ZONE_MOVABLE) { 5210 ret = 0; 5211 goto out; 5212 } 5213 5214 pfn = page_to_pfn(page); 5215 arg.start_pfn = pfn; 5216 arg.nr_pages = pageblock_nr_pages; 5217 arg.pages_found = 0; 5218 5219 /* 5220 * It may be possible to isolate a pageblock even if the 5221 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5222 * notifier chain is used by balloon drivers to return the 5223 * number of pages in a range that are held by the balloon 5224 * driver to shrink memory. If all the pages are accounted for 5225 * by balloons, are free, or on the LRU, isolation can continue. 5226 * Later, for example, when memory hotplug notifier runs, these 5227 * pages reported as "can be isolated" should be isolated(freed) 5228 * by the balloon driver through the memory notifier chain. 5229 */ 5230 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5231 notifier_ret = notifier_to_errno(notifier_ret); 5232 if (notifier_ret || !arg.pages_found) 5233 goto out; 5234 5235 for (iter = pfn; iter < (pfn + pageblock_nr_pages); iter++) { 5236 if (!pfn_valid_within(pfn)) 5237 continue; 5238 5239 curr_page = pfn_to_page(iter); 5240 if (!page_count(curr_page) || PageLRU(curr_page)) 5241 continue; 5242 5243 immobile++; 5244 } 5245 5246 if (arg.pages_found == immobile) 5247 ret = 0; 5248 5249out: 5250 if (!ret) { 5251 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5252 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5253 } 5254 5255 spin_unlock_irqrestore(&zone->lock, flags); 5256 if (!ret) 5257 drain_all_pages(); 5258 return ret; 5259} 5260 5261void unset_migratetype_isolate(struct page *page) 5262{ 5263 struct zone *zone; 5264 unsigned long flags; 5265 zone = page_zone(page); 5266 spin_lock_irqsave(&zone->lock, flags); 5267 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5268 goto out; 5269 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5270 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5271out: 5272 spin_unlock_irqrestore(&zone->lock, flags); 5273} 5274 5275#ifdef CONFIG_MEMORY_HOTREMOVE 5276/* 5277 * All pages in the range must be isolated before calling this. 5278 */ 5279void 5280__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5281{ 5282 struct page *page; 5283 struct zone *zone; 5284 int order, i; 5285 unsigned long pfn; 5286 unsigned long flags; 5287 /* find the first valid pfn */ 5288 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5289 if (pfn_valid(pfn)) 5290 break; 5291 if (pfn == end_pfn) 5292 return; 5293 zone = page_zone(pfn_to_page(pfn)); 5294 spin_lock_irqsave(&zone->lock, flags); 5295 pfn = start_pfn; 5296 while (pfn < end_pfn) { 5297 if (!pfn_valid(pfn)) { 5298 pfn++; 5299 continue; 5300 } 5301 page = pfn_to_page(pfn); 5302 BUG_ON(page_count(page)); 5303 BUG_ON(!PageBuddy(page)); 5304 order = page_order(page); 5305#ifdef CONFIG_DEBUG_VM 5306 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5307 pfn, 1 << order, end_pfn); 5308#endif 5309 list_del(&page->lru); 5310 rmv_page_order(page); 5311 zone->free_area[order].nr_free--; 5312 __mod_zone_page_state(zone, NR_FREE_PAGES, 5313 - (1UL << order)); 5314 for (i = 0; i < (1 << order); i++) 5315 SetPageReserved((page+i)); 5316 pfn += (1 << order); 5317 } 5318 spin_unlock_irqrestore(&zone->lock, flags); 5319} 5320#endif 5321 5322#ifdef CONFIG_MEMORY_FAILURE 5323bool is_free_buddy_page(struct page *page) 5324{ 5325 struct zone *zone = page_zone(page); 5326 unsigned long pfn = page_to_pfn(page); 5327 unsigned long flags; 5328 int order; 5329 5330 spin_lock_irqsave(&zone->lock, flags); 5331 for (order = 0; order < MAX_ORDER; order++) { 5332 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5333 5334 if (PageBuddy(page_head) && page_order(page_head) >= order) 5335 break; 5336 } 5337 spin_unlock_irqrestore(&zone->lock, flags); 5338 5339 return order < MAX_ORDER; 5340} 5341#endif 5342 5343static struct trace_print_flags pageflag_names[] = { 5344 {1UL << PG_locked, "locked" }, 5345 {1UL << PG_error, "error" }, 5346 {1UL << PG_referenced, "referenced" }, 5347 {1UL << PG_uptodate, "uptodate" }, 5348 {1UL << PG_dirty, "dirty" }, 5349 {1UL << PG_lru, "lru" }, 5350 {1UL << PG_active, "active" }, 5351 {1UL << PG_slab, "slab" }, 5352 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5353 {1UL << PG_arch_1, "arch_1" }, 5354 {1UL << PG_reserved, "reserved" }, 5355 {1UL << PG_private, "private" }, 5356 {1UL << PG_private_2, "private_2" }, 5357 {1UL << PG_writeback, "writeback" }, 5358#ifdef CONFIG_PAGEFLAGS_EXTENDED 5359 {1UL << PG_head, "head" }, 5360 {1UL << PG_tail, "tail" }, 5361#else 5362 {1UL << PG_compound, "compound" }, 5363#endif 5364 {1UL << PG_swapcache, "swapcache" }, 5365 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5366 {1UL << PG_reclaim, "reclaim" }, 5367 {1UL << PG_buddy, "buddy" }, 5368 {1UL << PG_swapbacked, "swapbacked" }, 5369 {1UL << PG_unevictable, "unevictable" }, 5370#ifdef CONFIG_MMU 5371 {1UL << PG_mlocked, "mlocked" }, 5372#endif 5373#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5374 {1UL << PG_uncached, "uncached" }, 5375#endif 5376#ifdef CONFIG_MEMORY_FAILURE 5377 {1UL << PG_hwpoison, "hwpoison" }, 5378#endif 5379 {-1UL, NULL }, 5380}; 5381 5382static void dump_page_flags(unsigned long flags) 5383{ 5384 const char *delim = ""; 5385 unsigned long mask; 5386 int i; 5387 5388 printk(KERN_ALERT "page flags: %#lx(", flags); 5389 5390 /* remove zone id */ 5391 flags &= (1UL << NR_PAGEFLAGS) - 1; 5392 5393 for (i = 0; pageflag_names[i].name && flags; i++) { 5394 5395 mask = pageflag_names[i].mask; 5396 if ((flags & mask) != mask) 5397 continue; 5398 5399 flags &= ~mask; 5400 printk("%s%s", delim, pageflag_names[i].name); 5401 delim = "|"; 5402 } 5403 5404 /* check for left over flags */ 5405 if (flags) 5406 printk("%s%#lx", delim, flags); 5407 5408 printk(")\n"); 5409} 5410 5411void dump_page(struct page *page) 5412{ 5413 printk(KERN_ALERT 5414 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5415 page, page_count(page), page_mapcount(page), 5416 page->mapping, page->index); 5417 dump_page_flags(page->flags); 5418} 5419