page_alloc.c revision bdae58317df5f1a25de82ff684347badc76ba006
1/* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17#include <linux/stddef.h> 18#include <linux/mm.h> 19#include <linux/swap.h> 20#include <linux/interrupt.h> 21#include <linux/pagemap.h> 22#include <linux/jiffies.h> 23#include <linux/bootmem.h> 24#include <linux/memblock.h> 25#include <linux/compiler.h> 26#include <linux/kernel.h> 27#include <linux/kmemcheck.h> 28#include <linux/module.h> 29#include <linux/suspend.h> 30#include <linux/pagevec.h> 31#include <linux/blkdev.h> 32#include <linux/slab.h> 33#include <linux/ratelimit.h> 34#include <linux/oom.h> 35#include <linux/notifier.h> 36#include <linux/topology.h> 37#include <linux/sysctl.h> 38#include <linux/cpu.h> 39#include <linux/cpuset.h> 40#include <linux/memory_hotplug.h> 41#include <linux/nodemask.h> 42#include <linux/vmalloc.h> 43#include <linux/vmstat.h> 44#include <linux/mempolicy.h> 45#include <linux/stop_machine.h> 46#include <linux/sort.h> 47#include <linux/pfn.h> 48#include <linux/backing-dev.h> 49#include <linux/fault-inject.h> 50#include <linux/page-isolation.h> 51#include <linux/page_cgroup.h> 52#include <linux/debugobjects.h> 53#include <linux/kmemleak.h> 54#include <linux/memory.h> 55#include <linux/compaction.h> 56#include <trace/events/kmem.h> 57#include <linux/ftrace_event.h> 58#include <linux/memcontrol.h> 59#include <linux/prefetch.h> 60#include <linux/page-debug-flags.h> 61 62#include <asm/tlbflush.h> 63#include <asm/div64.h> 64#include "internal.h" 65 66#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 67DEFINE_PER_CPU(int, numa_node); 68EXPORT_PER_CPU_SYMBOL(numa_node); 69#endif 70 71#ifdef CONFIG_HAVE_MEMORYLESS_NODES 72/* 73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 76 * defined in <linux/topology.h>. 77 */ 78DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 79EXPORT_PER_CPU_SYMBOL(_numa_mem_); 80#endif 81 82/* 83 * Array of node states. 84 */ 85nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 86 [N_POSSIBLE] = NODE_MASK_ALL, 87 [N_ONLINE] = { { [0] = 1UL } }, 88#ifndef CONFIG_NUMA 89 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 90#ifdef CONFIG_HIGHMEM 91 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 92#endif 93 [N_CPU] = { { [0] = 1UL } }, 94#endif /* NUMA */ 95}; 96EXPORT_SYMBOL(node_states); 97 98unsigned long totalram_pages __read_mostly; 99unsigned long totalreserve_pages __read_mostly; 100/* 101 * When calculating the number of globally allowed dirty pages, there 102 * is a certain number of per-zone reserves that should not be 103 * considered dirtyable memory. This is the sum of those reserves 104 * over all existing zones that contribute dirtyable memory. 105 */ 106unsigned long dirty_balance_reserve __read_mostly; 107 108int percpu_pagelist_fraction; 109gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 110 111#ifdef CONFIG_PM_SLEEP 112/* 113 * The following functions are used by the suspend/hibernate code to temporarily 114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 115 * while devices are suspended. To avoid races with the suspend/hibernate code, 116 * they should always be called with pm_mutex held (gfp_allowed_mask also should 117 * only be modified with pm_mutex held, unless the suspend/hibernate code is 118 * guaranteed not to run in parallel with that modification). 119 */ 120 121static gfp_t saved_gfp_mask; 122 123void pm_restore_gfp_mask(void) 124{ 125 WARN_ON(!mutex_is_locked(&pm_mutex)); 126 if (saved_gfp_mask) { 127 gfp_allowed_mask = saved_gfp_mask; 128 saved_gfp_mask = 0; 129 } 130} 131 132void pm_restrict_gfp_mask(void) 133{ 134 WARN_ON(!mutex_is_locked(&pm_mutex)); 135 WARN_ON(saved_gfp_mask); 136 saved_gfp_mask = gfp_allowed_mask; 137 gfp_allowed_mask &= ~GFP_IOFS; 138} 139 140bool pm_suspended_storage(void) 141{ 142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) 143 return false; 144 return true; 145} 146#endif /* CONFIG_PM_SLEEP */ 147 148#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 149int pageblock_order __read_mostly; 150#endif 151 152static void __free_pages_ok(struct page *page, unsigned int order); 153 154/* 155 * results with 256, 32 in the lowmem_reserve sysctl: 156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 157 * 1G machine -> (16M dma, 784M normal, 224M high) 158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 161 * 162 * TBD: should special case ZONE_DMA32 machines here - in those we normally 163 * don't need any ZONE_NORMAL reservation 164 */ 165int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 166#ifdef CONFIG_ZONE_DMA 167 256, 168#endif 169#ifdef CONFIG_ZONE_DMA32 170 256, 171#endif 172#ifdef CONFIG_HIGHMEM 173 32, 174#endif 175 32, 176}; 177 178EXPORT_SYMBOL(totalram_pages); 179 180static char * const zone_names[MAX_NR_ZONES] = { 181#ifdef CONFIG_ZONE_DMA 182 "DMA", 183#endif 184#ifdef CONFIG_ZONE_DMA32 185 "DMA32", 186#endif 187 "Normal", 188#ifdef CONFIG_HIGHMEM 189 "HighMem", 190#endif 191 "Movable", 192}; 193 194int min_free_kbytes = 1024; 195int min_free_order_shift = 1; 196 197static unsigned long __meminitdata nr_kernel_pages; 198static unsigned long __meminitdata nr_all_pages; 199static unsigned long __meminitdata dma_reserve; 200 201#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 202static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 203static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 204static unsigned long __initdata required_kernelcore; 205static unsigned long __initdata required_movablecore; 206static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 207 208/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 209int movable_zone; 210EXPORT_SYMBOL(movable_zone); 211#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 212 213#if MAX_NUMNODES > 1 214int nr_node_ids __read_mostly = MAX_NUMNODES; 215int nr_online_nodes __read_mostly = 1; 216EXPORT_SYMBOL(nr_node_ids); 217EXPORT_SYMBOL(nr_online_nodes); 218#endif 219 220int page_group_by_mobility_disabled __read_mostly; 221 222static void set_pageblock_migratetype(struct page *page, int migratetype) 223{ 224 225 if (unlikely(page_group_by_mobility_disabled)) 226 migratetype = MIGRATE_UNMOVABLE; 227 228 set_pageblock_flags_group(page, (unsigned long)migratetype, 229 PB_migrate, PB_migrate_end); 230} 231 232bool oom_killer_disabled __read_mostly; 233 234#ifdef CONFIG_DEBUG_VM 235static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 236{ 237 int ret = 0; 238 unsigned seq; 239 unsigned long pfn = page_to_pfn(page); 240 241 do { 242 seq = zone_span_seqbegin(zone); 243 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 244 ret = 1; 245 else if (pfn < zone->zone_start_pfn) 246 ret = 1; 247 } while (zone_span_seqretry(zone, seq)); 248 249 return ret; 250} 251 252static int page_is_consistent(struct zone *zone, struct page *page) 253{ 254 if (!pfn_valid_within(page_to_pfn(page))) 255 return 0; 256 if (zone != page_zone(page)) 257 return 0; 258 259 return 1; 260} 261/* 262 * Temporary debugging check for pages not lying within a given zone. 263 */ 264static int bad_range(struct zone *zone, struct page *page) 265{ 266 if (page_outside_zone_boundaries(zone, page)) 267 return 1; 268 if (!page_is_consistent(zone, page)) 269 return 1; 270 271 return 0; 272} 273#else 274static inline int bad_range(struct zone *zone, struct page *page) 275{ 276 return 0; 277} 278#endif 279 280static void bad_page(struct page *page) 281{ 282 static unsigned long resume; 283 static unsigned long nr_shown; 284 static unsigned long nr_unshown; 285 286 /* Don't complain about poisoned pages */ 287 if (PageHWPoison(page)) { 288 reset_page_mapcount(page); /* remove PageBuddy */ 289 return; 290 } 291 292 /* 293 * Allow a burst of 60 reports, then keep quiet for that minute; 294 * or allow a steady drip of one report per second. 295 */ 296 if (nr_shown == 60) { 297 if (time_before(jiffies, resume)) { 298 nr_unshown++; 299 goto out; 300 } 301 if (nr_unshown) { 302 printk(KERN_ALERT 303 "BUG: Bad page state: %lu messages suppressed\n", 304 nr_unshown); 305 nr_unshown = 0; 306 } 307 nr_shown = 0; 308 } 309 if (nr_shown++ == 0) 310 resume = jiffies + 60 * HZ; 311 312 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 313 current->comm, page_to_pfn(page)); 314 dump_page(page); 315 316 print_modules(); 317 dump_stack(); 318out: 319 /* Leave bad fields for debug, except PageBuddy could make trouble */ 320 reset_page_mapcount(page); /* remove PageBuddy */ 321 add_taint(TAINT_BAD_PAGE); 322} 323 324/* 325 * Higher-order pages are called "compound pages". They are structured thusly: 326 * 327 * The first PAGE_SIZE page is called the "head page". 328 * 329 * The remaining PAGE_SIZE pages are called "tail pages". 330 * 331 * All pages have PG_compound set. All tail pages have their ->first_page 332 * pointing at the head page. 333 * 334 * The first tail page's ->lru.next holds the address of the compound page's 335 * put_page() function. Its ->lru.prev holds the order of allocation. 336 * This usage means that zero-order pages may not be compound. 337 */ 338 339static void free_compound_page(struct page *page) 340{ 341 __free_pages_ok(page, compound_order(page)); 342} 343 344void prep_compound_page(struct page *page, unsigned long order) 345{ 346 int i; 347 int nr_pages = 1 << order; 348 349 set_compound_page_dtor(page, free_compound_page); 350 set_compound_order(page, order); 351 __SetPageHead(page); 352 for (i = 1; i < nr_pages; i++) { 353 struct page *p = page + i; 354 __SetPageTail(p); 355 set_page_count(p, 0); 356 p->first_page = page; 357 } 358} 359 360/* update __split_huge_page_refcount if you change this function */ 361static int destroy_compound_page(struct page *page, unsigned long order) 362{ 363 int i; 364 int nr_pages = 1 << order; 365 int bad = 0; 366 367 if (unlikely(compound_order(page) != order) || 368 unlikely(!PageHead(page))) { 369 bad_page(page); 370 bad++; 371 } 372 373 __ClearPageHead(page); 374 375 for (i = 1; i < nr_pages; i++) { 376 struct page *p = page + i; 377 378 if (unlikely(!PageTail(p) || (p->first_page != page))) { 379 bad_page(page); 380 bad++; 381 } 382 __ClearPageTail(p); 383 } 384 385 return bad; 386} 387 388static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 389{ 390 int i; 391 392 /* 393 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 394 * and __GFP_HIGHMEM from hard or soft interrupt context. 395 */ 396 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 397 for (i = 0; i < (1 << order); i++) 398 clear_highpage(page + i); 399} 400 401#ifdef CONFIG_DEBUG_PAGEALLOC 402unsigned int _debug_guardpage_minorder; 403 404static int __init debug_guardpage_minorder_setup(char *buf) 405{ 406 unsigned long res; 407 408 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 409 printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); 410 return 0; 411 } 412 _debug_guardpage_minorder = res; 413 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); 414 return 0; 415} 416__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); 417 418static inline void set_page_guard_flag(struct page *page) 419{ 420 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 421} 422 423static inline void clear_page_guard_flag(struct page *page) 424{ 425 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 426} 427#else 428static inline void set_page_guard_flag(struct page *page) { } 429static inline void clear_page_guard_flag(struct page *page) { } 430#endif 431 432static inline void set_page_order(struct page *page, int order) 433{ 434 set_page_private(page, order); 435 __SetPageBuddy(page); 436} 437 438static inline void rmv_page_order(struct page *page) 439{ 440 __ClearPageBuddy(page); 441 set_page_private(page, 0); 442} 443 444/* 445 * Locate the struct page for both the matching buddy in our 446 * pair (buddy1) and the combined O(n+1) page they form (page). 447 * 448 * 1) Any buddy B1 will have an order O twin B2 which satisfies 449 * the following equation: 450 * B2 = B1 ^ (1 << O) 451 * For example, if the starting buddy (buddy2) is #8 its order 452 * 1 buddy is #10: 453 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 454 * 455 * 2) Any buddy B will have an order O+1 parent P which 456 * satisfies the following equation: 457 * P = B & ~(1 << O) 458 * 459 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 460 */ 461static inline unsigned long 462__find_buddy_index(unsigned long page_idx, unsigned int order) 463{ 464 return page_idx ^ (1 << order); 465} 466 467/* 468 * This function checks whether a page is free && is the buddy 469 * we can do coalesce a page and its buddy if 470 * (a) the buddy is not in a hole && 471 * (b) the buddy is in the buddy system && 472 * (c) a page and its buddy have the same order && 473 * (d) a page and its buddy are in the same zone. 474 * 475 * For recording whether a page is in the buddy system, we set ->_mapcount -2. 476 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. 477 * 478 * For recording page's order, we use page_private(page). 479 */ 480static inline int page_is_buddy(struct page *page, struct page *buddy, 481 int order) 482{ 483 if (!pfn_valid_within(page_to_pfn(buddy))) 484 return 0; 485 486 if (page_zone_id(page) != page_zone_id(buddy)) 487 return 0; 488 489 if (page_is_guard(buddy) && page_order(buddy) == order) { 490 VM_BUG_ON(page_count(buddy) != 0); 491 return 1; 492 } 493 494 if (PageBuddy(buddy) && page_order(buddy) == order) { 495 VM_BUG_ON(page_count(buddy) != 0); 496 return 1; 497 } 498 return 0; 499} 500 501/* 502 * Freeing function for a buddy system allocator. 503 * 504 * The concept of a buddy system is to maintain direct-mapped table 505 * (containing bit values) for memory blocks of various "orders". 506 * The bottom level table contains the map for the smallest allocatable 507 * units of memory (here, pages), and each level above it describes 508 * pairs of units from the levels below, hence, "buddies". 509 * At a high level, all that happens here is marking the table entry 510 * at the bottom level available, and propagating the changes upward 511 * as necessary, plus some accounting needed to play nicely with other 512 * parts of the VM system. 513 * At each level, we keep a list of pages, which are heads of continuous 514 * free pages of length of (1 << order) and marked with _mapcount -2. Page's 515 * order is recorded in page_private(page) field. 516 * So when we are allocating or freeing one, we can derive the state of the 517 * other. That is, if we allocate a small block, and both were 518 * free, the remainder of the region must be split into blocks. 519 * If a block is freed, and its buddy is also free, then this 520 * triggers coalescing into a block of larger size. 521 * 522 * -- wli 523 */ 524 525static inline void __free_one_page(struct page *page, 526 struct zone *zone, unsigned int order, 527 int migratetype) 528{ 529 unsigned long page_idx; 530 unsigned long combined_idx; 531 unsigned long uninitialized_var(buddy_idx); 532 struct page *buddy; 533 534 if (unlikely(PageCompound(page))) 535 if (unlikely(destroy_compound_page(page, order))) 536 return; 537 538 VM_BUG_ON(migratetype == -1); 539 540 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 541 542 VM_BUG_ON(page_idx & ((1 << order) - 1)); 543 VM_BUG_ON(bad_range(zone, page)); 544 545 while (order < MAX_ORDER-1) { 546 buddy_idx = __find_buddy_index(page_idx, order); 547 buddy = page + (buddy_idx - page_idx); 548 if (!page_is_buddy(page, buddy, order)) 549 break; 550 /* 551 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 552 * merge with it and move up one order. 553 */ 554 if (page_is_guard(buddy)) { 555 clear_page_guard_flag(buddy); 556 set_page_private(page, 0); 557 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 558 } else { 559 list_del(&buddy->lru); 560 zone->free_area[order].nr_free--; 561 rmv_page_order(buddy); 562 } 563 combined_idx = buddy_idx & page_idx; 564 page = page + (combined_idx - page_idx); 565 page_idx = combined_idx; 566 order++; 567 } 568 set_page_order(page, order); 569 570 /* 571 * If this is not the largest possible page, check if the buddy 572 * of the next-highest order is free. If it is, it's possible 573 * that pages are being freed that will coalesce soon. In case, 574 * that is happening, add the free page to the tail of the list 575 * so it's less likely to be used soon and more likely to be merged 576 * as a higher order page 577 */ 578 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 579 struct page *higher_page, *higher_buddy; 580 combined_idx = buddy_idx & page_idx; 581 higher_page = page + (combined_idx - page_idx); 582 buddy_idx = __find_buddy_index(combined_idx, order + 1); 583 higher_buddy = page + (buddy_idx - combined_idx); 584 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 585 list_add_tail(&page->lru, 586 &zone->free_area[order].free_list[migratetype]); 587 goto out; 588 } 589 } 590 591 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 592out: 593 zone->free_area[order].nr_free++; 594} 595 596/* 597 * free_page_mlock() -- clean up attempts to free and mlocked() page. 598 * Page should not be on lru, so no need to fix that up. 599 * free_pages_check() will verify... 600 */ 601static inline void free_page_mlock(struct page *page) 602{ 603 __dec_zone_page_state(page, NR_MLOCK); 604 __count_vm_event(UNEVICTABLE_MLOCKFREED); 605} 606 607static inline int free_pages_check(struct page *page) 608{ 609 if (unlikely(page_mapcount(page) | 610 (page->mapping != NULL) | 611 (atomic_read(&page->_count) != 0) | 612 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | 613 (mem_cgroup_bad_page_check(page)))) { 614 bad_page(page); 615 return 1; 616 } 617 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 618 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 619 return 0; 620} 621 622/* 623 * Frees a number of pages from the PCP lists 624 * Assumes all pages on list are in same zone, and of same order. 625 * count is the number of pages to free. 626 * 627 * If the zone was previously in an "all pages pinned" state then look to 628 * see if this freeing clears that state. 629 * 630 * And clear the zone's pages_scanned counter, to hold off the "all pages are 631 * pinned" detection logic. 632 */ 633static void free_pcppages_bulk(struct zone *zone, int count, 634 struct per_cpu_pages *pcp) 635{ 636 int migratetype = 0; 637 int batch_free = 0; 638 int to_free = count; 639 640 spin_lock(&zone->lock); 641 zone->all_unreclaimable = 0; 642 zone->pages_scanned = 0; 643 644 while (to_free) { 645 struct page *page; 646 struct list_head *list; 647 648 /* 649 * Remove pages from lists in a round-robin fashion. A 650 * batch_free count is maintained that is incremented when an 651 * empty list is encountered. This is so more pages are freed 652 * off fuller lists instead of spinning excessively around empty 653 * lists 654 */ 655 do { 656 batch_free++; 657 if (++migratetype == MIGRATE_PCPTYPES) 658 migratetype = 0; 659 list = &pcp->lists[migratetype]; 660 } while (list_empty(list)); 661 662 /* This is the only non-empty list. Free them all. */ 663 if (batch_free == MIGRATE_PCPTYPES) 664 batch_free = to_free; 665 666 do { 667 page = list_entry(list->prev, struct page, lru); 668 /* must delete as __free_one_page list manipulates */ 669 list_del(&page->lru); 670 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 671 __free_one_page(page, zone, 0, page_private(page)); 672 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 673 } while (--to_free && --batch_free && !list_empty(list)); 674 } 675 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 676 spin_unlock(&zone->lock); 677} 678 679static void free_one_page(struct zone *zone, struct page *page, int order, 680 int migratetype) 681{ 682 spin_lock(&zone->lock); 683 zone->all_unreclaimable = 0; 684 zone->pages_scanned = 0; 685 686 __free_one_page(page, zone, order, migratetype); 687 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 688 spin_unlock(&zone->lock); 689} 690 691static bool free_pages_prepare(struct page *page, unsigned int order) 692{ 693 int i; 694 int bad = 0; 695 696 trace_mm_page_free(page, order); 697 kmemcheck_free_shadow(page, order); 698 699 if (PageAnon(page)) 700 page->mapping = NULL; 701 for (i = 0; i < (1 << order); i++) 702 bad += free_pages_check(page + i); 703 if (bad) 704 return false; 705 706 if (!PageHighMem(page)) { 707 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 708 debug_check_no_obj_freed(page_address(page), 709 PAGE_SIZE << order); 710 } 711 arch_free_page(page, order); 712 kernel_map_pages(page, 1 << order, 0); 713 714 return true; 715} 716 717static void __free_pages_ok(struct page *page, unsigned int order) 718{ 719 unsigned long flags; 720 int wasMlocked = __TestClearPageMlocked(page); 721 722 if (!free_pages_prepare(page, order)) 723 return; 724 725 local_irq_save(flags); 726 if (unlikely(wasMlocked)) 727 free_page_mlock(page); 728 __count_vm_events(PGFREE, 1 << order); 729 free_one_page(page_zone(page), page, order, 730 get_pageblock_migratetype(page)); 731 local_irq_restore(flags); 732} 733 734void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 735{ 736 unsigned int nr_pages = 1 << order; 737 unsigned int loop; 738 739 prefetchw(page); 740 for (loop = 0; loop < nr_pages; loop++) { 741 struct page *p = &page[loop]; 742 743 if (loop + 1 < nr_pages) 744 prefetchw(p + 1); 745 __ClearPageReserved(p); 746 set_page_count(p, 0); 747 } 748 749 set_page_refcounted(page); 750 __free_pages(page, order); 751} 752 753 754/* 755 * The order of subdivision here is critical for the IO subsystem. 756 * Please do not alter this order without good reasons and regression 757 * testing. Specifically, as large blocks of memory are subdivided, 758 * the order in which smaller blocks are delivered depends on the order 759 * they're subdivided in this function. This is the primary factor 760 * influencing the order in which pages are delivered to the IO 761 * subsystem according to empirical testing, and this is also justified 762 * by considering the behavior of a buddy system containing a single 763 * large block of memory acted on by a series of small allocations. 764 * This behavior is a critical factor in sglist merging's success. 765 * 766 * -- wli 767 */ 768static inline void expand(struct zone *zone, struct page *page, 769 int low, int high, struct free_area *area, 770 int migratetype) 771{ 772 unsigned long size = 1 << high; 773 774 while (high > low) { 775 area--; 776 high--; 777 size >>= 1; 778 VM_BUG_ON(bad_range(zone, &page[size])); 779 780#ifdef CONFIG_DEBUG_PAGEALLOC 781 if (high < debug_guardpage_minorder()) { 782 /* 783 * Mark as guard pages (or page), that will allow to 784 * merge back to allocator when buddy will be freed. 785 * Corresponding page table entries will not be touched, 786 * pages will stay not present in virtual address space 787 */ 788 INIT_LIST_HEAD(&page[size].lru); 789 set_page_guard_flag(&page[size]); 790 set_page_private(&page[size], high); 791 /* Guard pages are not available for any usage */ 792 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high)); 793 continue; 794 } 795#endif 796 list_add(&page[size].lru, &area->free_list[migratetype]); 797 area->nr_free++; 798 set_page_order(&page[size], high); 799 } 800} 801 802/* 803 * This page is about to be returned from the page allocator 804 */ 805static inline int check_new_page(struct page *page) 806{ 807 if (unlikely(page_mapcount(page) | 808 (page->mapping != NULL) | 809 (atomic_read(&page->_count) != 0) | 810 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | 811 (mem_cgroup_bad_page_check(page)))) { 812 bad_page(page); 813 return 1; 814 } 815 return 0; 816} 817 818static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 819{ 820 int i; 821 822 for (i = 0; i < (1 << order); i++) { 823 struct page *p = page + i; 824 if (unlikely(check_new_page(p))) 825 return 1; 826 } 827 828 set_page_private(page, 0); 829 set_page_refcounted(page); 830 831 arch_alloc_page(page, order); 832 kernel_map_pages(page, 1 << order, 1); 833 834 if (gfp_flags & __GFP_ZERO) 835 prep_zero_page(page, order, gfp_flags); 836 837 if (order && (gfp_flags & __GFP_COMP)) 838 prep_compound_page(page, order); 839 840 return 0; 841} 842 843/* 844 * Go through the free lists for the given migratetype and remove 845 * the smallest available page from the freelists 846 */ 847static inline 848struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 849 int migratetype) 850{ 851 unsigned int current_order; 852 struct free_area * area; 853 struct page *page; 854 855 /* Find a page of the appropriate size in the preferred list */ 856 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 857 area = &(zone->free_area[current_order]); 858 if (list_empty(&area->free_list[migratetype])) 859 continue; 860 861 page = list_entry(area->free_list[migratetype].next, 862 struct page, lru); 863 list_del(&page->lru); 864 rmv_page_order(page); 865 area->nr_free--; 866 expand(zone, page, order, current_order, area, migratetype); 867 return page; 868 } 869 870 return NULL; 871} 872 873 874/* 875 * This array describes the order lists are fallen back to when 876 * the free lists for the desirable migrate type are depleted 877 */ 878static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 879 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 880 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 881 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 882 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 883}; 884 885/* 886 * Move the free pages in a range to the free lists of the requested type. 887 * Note that start_page and end_pages are not aligned on a pageblock 888 * boundary. If alignment is required, use move_freepages_block() 889 */ 890static int move_freepages(struct zone *zone, 891 struct page *start_page, struct page *end_page, 892 int migratetype) 893{ 894 struct page *page; 895 unsigned long order; 896 int pages_moved = 0; 897 898#ifndef CONFIG_HOLES_IN_ZONE 899 /* 900 * page_zone is not safe to call in this context when 901 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 902 * anyway as we check zone boundaries in move_freepages_block(). 903 * Remove at a later date when no bug reports exist related to 904 * grouping pages by mobility 905 */ 906 BUG_ON(page_zone(start_page) != page_zone(end_page)); 907#endif 908 909 for (page = start_page; page <= end_page;) { 910 /* Make sure we are not inadvertently changing nodes */ 911 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 912 913 if (!pfn_valid_within(page_to_pfn(page))) { 914 page++; 915 continue; 916 } 917 918 if (!PageBuddy(page)) { 919 page++; 920 continue; 921 } 922 923 order = page_order(page); 924 list_move(&page->lru, 925 &zone->free_area[order].free_list[migratetype]); 926 page += 1 << order; 927 pages_moved += 1 << order; 928 } 929 930 return pages_moved; 931} 932 933static int move_freepages_block(struct zone *zone, struct page *page, 934 int migratetype) 935{ 936 unsigned long start_pfn, end_pfn; 937 struct page *start_page, *end_page; 938 939 start_pfn = page_to_pfn(page); 940 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 941 start_page = pfn_to_page(start_pfn); 942 end_page = start_page + pageblock_nr_pages - 1; 943 end_pfn = start_pfn + pageblock_nr_pages - 1; 944 945 /* Do not cross zone boundaries */ 946 if (start_pfn < zone->zone_start_pfn) 947 start_page = page; 948 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 949 return 0; 950 951 return move_freepages(zone, start_page, end_page, migratetype); 952} 953 954static void change_pageblock_range(struct page *pageblock_page, 955 int start_order, int migratetype) 956{ 957 int nr_pageblocks = 1 << (start_order - pageblock_order); 958 959 while (nr_pageblocks--) { 960 set_pageblock_migratetype(pageblock_page, migratetype); 961 pageblock_page += pageblock_nr_pages; 962 } 963} 964 965/* Remove an element from the buddy allocator from the fallback list */ 966static inline struct page * 967__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 968{ 969 struct free_area * area; 970 int current_order; 971 struct page *page; 972 int migratetype, i; 973 974 /* Find the largest possible block of pages in the other list */ 975 for (current_order = MAX_ORDER-1; current_order >= order; 976 --current_order) { 977 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 978 migratetype = fallbacks[start_migratetype][i]; 979 980 /* MIGRATE_RESERVE handled later if necessary */ 981 if (migratetype == MIGRATE_RESERVE) 982 continue; 983 984 area = &(zone->free_area[current_order]); 985 if (list_empty(&area->free_list[migratetype])) 986 continue; 987 988 page = list_entry(area->free_list[migratetype].next, 989 struct page, lru); 990 area->nr_free--; 991 992 /* 993 * If breaking a large block of pages, move all free 994 * pages to the preferred allocation list. If falling 995 * back for a reclaimable kernel allocation, be more 996 * aggressive about taking ownership of free pages 997 */ 998 if (unlikely(current_order >= (pageblock_order >> 1)) || 999 start_migratetype == MIGRATE_RECLAIMABLE || 1000 page_group_by_mobility_disabled) { 1001 unsigned long pages; 1002 pages = move_freepages_block(zone, page, 1003 start_migratetype); 1004 1005 /* Claim the whole block if over half of it is free */ 1006 if (pages >= (1 << (pageblock_order-1)) || 1007 page_group_by_mobility_disabled) 1008 set_pageblock_migratetype(page, 1009 start_migratetype); 1010 1011 migratetype = start_migratetype; 1012 } 1013 1014 /* Remove the page from the freelists */ 1015 list_del(&page->lru); 1016 rmv_page_order(page); 1017 1018 /* Take ownership for orders >= pageblock_order */ 1019 if (current_order >= pageblock_order) 1020 change_pageblock_range(page, current_order, 1021 start_migratetype); 1022 1023 expand(zone, page, order, current_order, area, migratetype); 1024 1025 trace_mm_page_alloc_extfrag(page, order, current_order, 1026 start_migratetype, migratetype); 1027 1028 return page; 1029 } 1030 } 1031 1032 return NULL; 1033} 1034 1035/* 1036 * Do the hard work of removing an element from the buddy allocator. 1037 * Call me with the zone->lock already held. 1038 */ 1039static struct page *__rmqueue(struct zone *zone, unsigned int order, 1040 int migratetype) 1041{ 1042 struct page *page; 1043 1044retry_reserve: 1045 page = __rmqueue_smallest(zone, order, migratetype); 1046 1047 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1048 page = __rmqueue_fallback(zone, order, migratetype); 1049 1050 /* 1051 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1052 * is used because __rmqueue_smallest is an inline function 1053 * and we want just one call site 1054 */ 1055 if (!page) { 1056 migratetype = MIGRATE_RESERVE; 1057 goto retry_reserve; 1058 } 1059 } 1060 1061 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1062 return page; 1063} 1064 1065/* 1066 * Obtain a specified number of elements from the buddy allocator, all under 1067 * a single hold of the lock, for efficiency. Add them to the supplied list. 1068 * Returns the number of new pages which were placed at *list. 1069 */ 1070static int rmqueue_bulk(struct zone *zone, unsigned int order, 1071 unsigned long count, struct list_head *list, 1072 int migratetype, int cold) 1073{ 1074 int i; 1075 1076 spin_lock(&zone->lock); 1077 for (i = 0; i < count; ++i) { 1078 struct page *page = __rmqueue(zone, order, migratetype); 1079 if (unlikely(page == NULL)) 1080 break; 1081 1082 /* 1083 * Split buddy pages returned by expand() are received here 1084 * in physical page order. The page is added to the callers and 1085 * list and the list head then moves forward. From the callers 1086 * perspective, the linked list is ordered by page number in 1087 * some conditions. This is useful for IO devices that can 1088 * merge IO requests if the physical pages are ordered 1089 * properly. 1090 */ 1091 if (likely(cold == 0)) 1092 list_add(&page->lru, list); 1093 else 1094 list_add_tail(&page->lru, list); 1095 set_page_private(page, migratetype); 1096 list = &page->lru; 1097 } 1098 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1099 spin_unlock(&zone->lock); 1100 return i; 1101} 1102 1103#ifdef CONFIG_NUMA 1104/* 1105 * Called from the vmstat counter updater to drain pagesets of this 1106 * currently executing processor on remote nodes after they have 1107 * expired. 1108 * 1109 * Note that this function must be called with the thread pinned to 1110 * a single processor. 1111 */ 1112void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1113{ 1114 unsigned long flags; 1115 int to_drain; 1116 1117 local_irq_save(flags); 1118 if (pcp->count >= pcp->batch) 1119 to_drain = pcp->batch; 1120 else 1121 to_drain = pcp->count; 1122 free_pcppages_bulk(zone, to_drain, pcp); 1123 pcp->count -= to_drain; 1124 local_irq_restore(flags); 1125} 1126#endif 1127 1128/* 1129 * Drain pages of the indicated processor. 1130 * 1131 * The processor must either be the current processor and the 1132 * thread pinned to the current processor or a processor that 1133 * is not online. 1134 */ 1135static void drain_pages(unsigned int cpu) 1136{ 1137 unsigned long flags; 1138 struct zone *zone; 1139 1140 for_each_populated_zone(zone) { 1141 struct per_cpu_pageset *pset; 1142 struct per_cpu_pages *pcp; 1143 1144 local_irq_save(flags); 1145 pset = per_cpu_ptr(zone->pageset, cpu); 1146 1147 pcp = &pset->pcp; 1148 if (pcp->count) { 1149 free_pcppages_bulk(zone, pcp->count, pcp); 1150 pcp->count = 0; 1151 } 1152 local_irq_restore(flags); 1153 } 1154} 1155 1156/* 1157 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1158 */ 1159void drain_local_pages(void *arg) 1160{ 1161 drain_pages(smp_processor_id()); 1162} 1163 1164/* 1165 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1166 */ 1167void drain_all_pages(void) 1168{ 1169 on_each_cpu(drain_local_pages, NULL, 1); 1170} 1171 1172#ifdef CONFIG_HIBERNATION 1173 1174void mark_free_pages(struct zone *zone) 1175{ 1176 unsigned long pfn, max_zone_pfn; 1177 unsigned long flags; 1178 int order, t; 1179 struct list_head *curr; 1180 1181 if (!zone->spanned_pages) 1182 return; 1183 1184 spin_lock_irqsave(&zone->lock, flags); 1185 1186 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1187 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1188 if (pfn_valid(pfn)) { 1189 struct page *page = pfn_to_page(pfn); 1190 1191 if (!swsusp_page_is_forbidden(page)) 1192 swsusp_unset_page_free(page); 1193 } 1194 1195 for_each_migratetype_order(order, t) { 1196 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1197 unsigned long i; 1198 1199 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1200 for (i = 0; i < (1UL << order); i++) 1201 swsusp_set_page_free(pfn_to_page(pfn + i)); 1202 } 1203 } 1204 spin_unlock_irqrestore(&zone->lock, flags); 1205} 1206#endif /* CONFIG_PM */ 1207 1208/* 1209 * Free a 0-order page 1210 * cold == 1 ? free a cold page : free a hot page 1211 */ 1212void free_hot_cold_page(struct page *page, int cold) 1213{ 1214 struct zone *zone = page_zone(page); 1215 struct per_cpu_pages *pcp; 1216 unsigned long flags; 1217 int migratetype; 1218 int wasMlocked = __TestClearPageMlocked(page); 1219 1220 if (!free_pages_prepare(page, 0)) 1221 return; 1222 1223 migratetype = get_pageblock_migratetype(page); 1224 set_page_private(page, migratetype); 1225 local_irq_save(flags); 1226 if (unlikely(wasMlocked)) 1227 free_page_mlock(page); 1228 __count_vm_event(PGFREE); 1229 1230 /* 1231 * We only track unmovable, reclaimable and movable on pcp lists. 1232 * Free ISOLATE pages back to the allocator because they are being 1233 * offlined but treat RESERVE as movable pages so we can get those 1234 * areas back if necessary. Otherwise, we may have to free 1235 * excessively into the page allocator 1236 */ 1237 if (migratetype >= MIGRATE_PCPTYPES) { 1238 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1239 free_one_page(zone, page, 0, migratetype); 1240 goto out; 1241 } 1242 migratetype = MIGRATE_MOVABLE; 1243 } 1244 1245 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1246 if (cold) 1247 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1248 else 1249 list_add(&page->lru, &pcp->lists[migratetype]); 1250 pcp->count++; 1251 if (pcp->count >= pcp->high) { 1252 free_pcppages_bulk(zone, pcp->batch, pcp); 1253 pcp->count -= pcp->batch; 1254 } 1255 1256out: 1257 local_irq_restore(flags); 1258} 1259 1260/* 1261 * Free a list of 0-order pages 1262 */ 1263void free_hot_cold_page_list(struct list_head *list, int cold) 1264{ 1265 struct page *page, *next; 1266 1267 list_for_each_entry_safe(page, next, list, lru) { 1268 trace_mm_page_free_batched(page, cold); 1269 free_hot_cold_page(page, cold); 1270 } 1271} 1272 1273/* 1274 * split_page takes a non-compound higher-order page, and splits it into 1275 * n (1<<order) sub-pages: page[0..n] 1276 * Each sub-page must be freed individually. 1277 * 1278 * Note: this is probably too low level an operation for use in drivers. 1279 * Please consult with lkml before using this in your driver. 1280 */ 1281void split_page(struct page *page, unsigned int order) 1282{ 1283 int i; 1284 1285 VM_BUG_ON(PageCompound(page)); 1286 VM_BUG_ON(!page_count(page)); 1287 1288#ifdef CONFIG_KMEMCHECK 1289 /* 1290 * Split shadow pages too, because free(page[0]) would 1291 * otherwise free the whole shadow. 1292 */ 1293 if (kmemcheck_page_is_tracked(page)) 1294 split_page(virt_to_page(page[0].shadow), order); 1295#endif 1296 1297 for (i = 1; i < (1 << order); i++) 1298 set_page_refcounted(page + i); 1299} 1300 1301/* 1302 * Similar to split_page except the page is already free. As this is only 1303 * being used for migration, the migratetype of the block also changes. 1304 * As this is called with interrupts disabled, the caller is responsible 1305 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1306 * are enabled. 1307 * 1308 * Note: this is probably too low level an operation for use in drivers. 1309 * Please consult with lkml before using this in your driver. 1310 */ 1311int split_free_page(struct page *page) 1312{ 1313 unsigned int order; 1314 unsigned long watermark; 1315 struct zone *zone; 1316 1317 BUG_ON(!PageBuddy(page)); 1318 1319 zone = page_zone(page); 1320 order = page_order(page); 1321 1322 /* Obey watermarks as if the page was being allocated */ 1323 watermark = low_wmark_pages(zone) + (1 << order); 1324 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1325 return 0; 1326 1327 /* Remove page from free list */ 1328 list_del(&page->lru); 1329 zone->free_area[order].nr_free--; 1330 rmv_page_order(page); 1331 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1332 1333 /* Split into individual pages */ 1334 set_page_refcounted(page); 1335 split_page(page, order); 1336 1337 if (order >= pageblock_order - 1) { 1338 struct page *endpage = page + (1 << order) - 1; 1339 for (; page < endpage; page += pageblock_nr_pages) 1340 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1341 } 1342 1343 return 1 << order; 1344} 1345 1346/* 1347 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1348 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1349 * or two. 1350 */ 1351static inline 1352struct page *buffered_rmqueue(struct zone *preferred_zone, 1353 struct zone *zone, int order, gfp_t gfp_flags, 1354 int migratetype) 1355{ 1356 unsigned long flags; 1357 struct page *page; 1358 int cold = !!(gfp_flags & __GFP_COLD); 1359 1360again: 1361 if (likely(order == 0)) { 1362 struct per_cpu_pages *pcp; 1363 struct list_head *list; 1364 1365 local_irq_save(flags); 1366 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1367 list = &pcp->lists[migratetype]; 1368 if (list_empty(list)) { 1369 pcp->count += rmqueue_bulk(zone, 0, 1370 pcp->batch, list, 1371 migratetype, cold); 1372 if (unlikely(list_empty(list))) 1373 goto failed; 1374 } 1375 1376 if (cold) 1377 page = list_entry(list->prev, struct page, lru); 1378 else 1379 page = list_entry(list->next, struct page, lru); 1380 1381 list_del(&page->lru); 1382 pcp->count--; 1383 } else { 1384 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1385 /* 1386 * __GFP_NOFAIL is not to be used in new code. 1387 * 1388 * All __GFP_NOFAIL callers should be fixed so that they 1389 * properly detect and handle allocation failures. 1390 * 1391 * We most definitely don't want callers attempting to 1392 * allocate greater than order-1 page units with 1393 * __GFP_NOFAIL. 1394 */ 1395 WARN_ON_ONCE(order > 1); 1396 } 1397 spin_lock_irqsave(&zone->lock, flags); 1398 page = __rmqueue(zone, order, migratetype); 1399 spin_unlock(&zone->lock); 1400 if (!page) 1401 goto failed; 1402 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1403 } 1404 1405 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1406 zone_statistics(preferred_zone, zone, gfp_flags); 1407 local_irq_restore(flags); 1408 1409 VM_BUG_ON(bad_range(zone, page)); 1410 if (prep_new_page(page, order, gfp_flags)) 1411 goto again; 1412 return page; 1413 1414failed: 1415 local_irq_restore(flags); 1416 return NULL; 1417} 1418 1419/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1420#define ALLOC_WMARK_MIN WMARK_MIN 1421#define ALLOC_WMARK_LOW WMARK_LOW 1422#define ALLOC_WMARK_HIGH WMARK_HIGH 1423#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1424 1425/* Mask to get the watermark bits */ 1426#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1427 1428#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1429#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1430#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1431 1432#ifdef CONFIG_FAIL_PAGE_ALLOC 1433 1434static struct { 1435 struct fault_attr attr; 1436 1437 u32 ignore_gfp_highmem; 1438 u32 ignore_gfp_wait; 1439 u32 min_order; 1440} fail_page_alloc = { 1441 .attr = FAULT_ATTR_INITIALIZER, 1442 .ignore_gfp_wait = 1, 1443 .ignore_gfp_highmem = 1, 1444 .min_order = 1, 1445}; 1446 1447static int __init setup_fail_page_alloc(char *str) 1448{ 1449 return setup_fault_attr(&fail_page_alloc.attr, str); 1450} 1451__setup("fail_page_alloc=", setup_fail_page_alloc); 1452 1453static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1454{ 1455 if (order < fail_page_alloc.min_order) 1456 return 0; 1457 if (gfp_mask & __GFP_NOFAIL) 1458 return 0; 1459 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1460 return 0; 1461 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1462 return 0; 1463 1464 return should_fail(&fail_page_alloc.attr, 1 << order); 1465} 1466 1467#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1468 1469static int __init fail_page_alloc_debugfs(void) 1470{ 1471 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1472 struct dentry *dir; 1473 1474 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 1475 &fail_page_alloc.attr); 1476 if (IS_ERR(dir)) 1477 return PTR_ERR(dir); 1478 1479 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 1480 &fail_page_alloc.ignore_gfp_wait)) 1481 goto fail; 1482 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1483 &fail_page_alloc.ignore_gfp_highmem)) 1484 goto fail; 1485 if (!debugfs_create_u32("min-order", mode, dir, 1486 &fail_page_alloc.min_order)) 1487 goto fail; 1488 1489 return 0; 1490fail: 1491 debugfs_remove_recursive(dir); 1492 1493 return -ENOMEM; 1494} 1495 1496late_initcall(fail_page_alloc_debugfs); 1497 1498#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1499 1500#else /* CONFIG_FAIL_PAGE_ALLOC */ 1501 1502static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1503{ 1504 return 0; 1505} 1506 1507#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1508 1509/* 1510 * Return true if free pages are above 'mark'. This takes into account the order 1511 * of the allocation. 1512 */ 1513static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1514 int classzone_idx, int alloc_flags, long free_pages) 1515{ 1516 /* free_pages my go negative - that's OK */ 1517 long min = mark; 1518 int o; 1519 1520 free_pages -= (1 << order) - 1; 1521 if (alloc_flags & ALLOC_HIGH) 1522 min -= min / 2; 1523 if (alloc_flags & ALLOC_HARDER) 1524 min -= min / 4; 1525 1526 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1527 return false; 1528 for (o = 0; o < order; o++) { 1529 /* At the next order, this order's pages become unavailable */ 1530 free_pages -= z->free_area[o].nr_free << o; 1531 1532 /* Require fewer higher order pages to be free */ 1533 min >>= min_free_order_shift; 1534 1535 if (free_pages <= min) 1536 return false; 1537 } 1538 return true; 1539} 1540 1541bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1542 int classzone_idx, int alloc_flags) 1543{ 1544 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1545 zone_page_state(z, NR_FREE_PAGES)); 1546} 1547 1548bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1549 int classzone_idx, int alloc_flags) 1550{ 1551 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1552 1553 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1554 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1555 1556 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1557 free_pages); 1558} 1559 1560#ifdef CONFIG_NUMA 1561/* 1562 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1563 * skip over zones that are not allowed by the cpuset, or that have 1564 * been recently (in last second) found to be nearly full. See further 1565 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1566 * that have to skip over a lot of full or unallowed zones. 1567 * 1568 * If the zonelist cache is present in the passed in zonelist, then 1569 * returns a pointer to the allowed node mask (either the current 1570 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1571 * 1572 * If the zonelist cache is not available for this zonelist, does 1573 * nothing and returns NULL. 1574 * 1575 * If the fullzones BITMAP in the zonelist cache is stale (more than 1576 * a second since last zap'd) then we zap it out (clear its bits.) 1577 * 1578 * We hold off even calling zlc_setup, until after we've checked the 1579 * first zone in the zonelist, on the theory that most allocations will 1580 * be satisfied from that first zone, so best to examine that zone as 1581 * quickly as we can. 1582 */ 1583static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1584{ 1585 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1586 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1587 1588 zlc = zonelist->zlcache_ptr; 1589 if (!zlc) 1590 return NULL; 1591 1592 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1593 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1594 zlc->last_full_zap = jiffies; 1595 } 1596 1597 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1598 &cpuset_current_mems_allowed : 1599 &node_states[N_HIGH_MEMORY]; 1600 return allowednodes; 1601} 1602 1603/* 1604 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1605 * if it is worth looking at further for free memory: 1606 * 1) Check that the zone isn't thought to be full (doesn't have its 1607 * bit set in the zonelist_cache fullzones BITMAP). 1608 * 2) Check that the zones node (obtained from the zonelist_cache 1609 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1610 * Return true (non-zero) if zone is worth looking at further, or 1611 * else return false (zero) if it is not. 1612 * 1613 * This check -ignores- the distinction between various watermarks, 1614 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1615 * found to be full for any variation of these watermarks, it will 1616 * be considered full for up to one second by all requests, unless 1617 * we are so low on memory on all allowed nodes that we are forced 1618 * into the second scan of the zonelist. 1619 * 1620 * In the second scan we ignore this zonelist cache and exactly 1621 * apply the watermarks to all zones, even it is slower to do so. 1622 * We are low on memory in the second scan, and should leave no stone 1623 * unturned looking for a free page. 1624 */ 1625static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1626 nodemask_t *allowednodes) 1627{ 1628 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1629 int i; /* index of *z in zonelist zones */ 1630 int n; /* node that zone *z is on */ 1631 1632 zlc = zonelist->zlcache_ptr; 1633 if (!zlc) 1634 return 1; 1635 1636 i = z - zonelist->_zonerefs; 1637 n = zlc->z_to_n[i]; 1638 1639 /* This zone is worth trying if it is allowed but not full */ 1640 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1641} 1642 1643/* 1644 * Given 'z' scanning a zonelist, set the corresponding bit in 1645 * zlc->fullzones, so that subsequent attempts to allocate a page 1646 * from that zone don't waste time re-examining it. 1647 */ 1648static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1649{ 1650 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1651 int i; /* index of *z in zonelist zones */ 1652 1653 zlc = zonelist->zlcache_ptr; 1654 if (!zlc) 1655 return; 1656 1657 i = z - zonelist->_zonerefs; 1658 1659 set_bit(i, zlc->fullzones); 1660} 1661 1662/* 1663 * clear all zones full, called after direct reclaim makes progress so that 1664 * a zone that was recently full is not skipped over for up to a second 1665 */ 1666static void zlc_clear_zones_full(struct zonelist *zonelist) 1667{ 1668 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1669 1670 zlc = zonelist->zlcache_ptr; 1671 if (!zlc) 1672 return; 1673 1674 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1675} 1676 1677#else /* CONFIG_NUMA */ 1678 1679static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1680{ 1681 return NULL; 1682} 1683 1684static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1685 nodemask_t *allowednodes) 1686{ 1687 return 1; 1688} 1689 1690static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1691{ 1692} 1693 1694static void zlc_clear_zones_full(struct zonelist *zonelist) 1695{ 1696} 1697#endif /* CONFIG_NUMA */ 1698 1699/* 1700 * get_page_from_freelist goes through the zonelist trying to allocate 1701 * a page. 1702 */ 1703static struct page * 1704get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1705 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1706 struct zone *preferred_zone, int migratetype) 1707{ 1708 struct zoneref *z; 1709 struct page *page = NULL; 1710 int classzone_idx; 1711 struct zone *zone; 1712 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1713 int zlc_active = 0; /* set if using zonelist_cache */ 1714 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1715 1716 classzone_idx = zone_idx(preferred_zone); 1717zonelist_scan: 1718 /* 1719 * Scan zonelist, looking for a zone with enough free. 1720 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1721 */ 1722 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1723 high_zoneidx, nodemask) { 1724 if (NUMA_BUILD && zlc_active && 1725 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1726 continue; 1727 if ((alloc_flags & ALLOC_CPUSET) && 1728 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1729 continue; 1730 /* 1731 * When allocating a page cache page for writing, we 1732 * want to get it from a zone that is within its dirty 1733 * limit, such that no single zone holds more than its 1734 * proportional share of globally allowed dirty pages. 1735 * The dirty limits take into account the zone's 1736 * lowmem reserves and high watermark so that kswapd 1737 * should be able to balance it without having to 1738 * write pages from its LRU list. 1739 * 1740 * This may look like it could increase pressure on 1741 * lower zones by failing allocations in higher zones 1742 * before they are full. But the pages that do spill 1743 * over are limited as the lower zones are protected 1744 * by this very same mechanism. It should not become 1745 * a practical burden to them. 1746 * 1747 * XXX: For now, allow allocations to potentially 1748 * exceed the per-zone dirty limit in the slowpath 1749 * (ALLOC_WMARK_LOW unset) before going into reclaim, 1750 * which is important when on a NUMA setup the allowed 1751 * zones are together not big enough to reach the 1752 * global limit. The proper fix for these situations 1753 * will require awareness of zones in the 1754 * dirty-throttling and the flusher threads. 1755 */ 1756 if ((alloc_flags & ALLOC_WMARK_LOW) && 1757 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone)) 1758 goto this_zone_full; 1759 1760 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1761 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1762 unsigned long mark; 1763 int ret; 1764 1765 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1766 if (zone_watermark_ok(zone, order, mark, 1767 classzone_idx, alloc_flags)) 1768 goto try_this_zone; 1769 1770 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1771 /* 1772 * we do zlc_setup if there are multiple nodes 1773 * and before considering the first zone allowed 1774 * by the cpuset. 1775 */ 1776 allowednodes = zlc_setup(zonelist, alloc_flags); 1777 zlc_active = 1; 1778 did_zlc_setup = 1; 1779 } 1780 1781 if (zone_reclaim_mode == 0) 1782 goto this_zone_full; 1783 1784 /* 1785 * As we may have just activated ZLC, check if the first 1786 * eligible zone has failed zone_reclaim recently. 1787 */ 1788 if (NUMA_BUILD && zlc_active && 1789 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1790 continue; 1791 1792 ret = zone_reclaim(zone, gfp_mask, order); 1793 switch (ret) { 1794 case ZONE_RECLAIM_NOSCAN: 1795 /* did not scan */ 1796 continue; 1797 case ZONE_RECLAIM_FULL: 1798 /* scanned but unreclaimable */ 1799 continue; 1800 default: 1801 /* did we reclaim enough */ 1802 if (!zone_watermark_ok(zone, order, mark, 1803 classzone_idx, alloc_flags)) 1804 goto this_zone_full; 1805 } 1806 } 1807 1808try_this_zone: 1809 page = buffered_rmqueue(preferred_zone, zone, order, 1810 gfp_mask, migratetype); 1811 if (page) 1812 break; 1813this_zone_full: 1814 if (NUMA_BUILD) 1815 zlc_mark_zone_full(zonelist, z); 1816 } 1817 1818 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1819 /* Disable zlc cache for second zonelist scan */ 1820 zlc_active = 0; 1821 goto zonelist_scan; 1822 } 1823 return page; 1824} 1825 1826/* 1827 * Large machines with many possible nodes should not always dump per-node 1828 * meminfo in irq context. 1829 */ 1830static inline bool should_suppress_show_mem(void) 1831{ 1832 bool ret = false; 1833 1834#if NODES_SHIFT > 8 1835 ret = in_interrupt(); 1836#endif 1837 return ret; 1838} 1839 1840static DEFINE_RATELIMIT_STATE(nopage_rs, 1841 DEFAULT_RATELIMIT_INTERVAL, 1842 DEFAULT_RATELIMIT_BURST); 1843 1844void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 1845{ 1846 unsigned int filter = SHOW_MEM_FILTER_NODES; 1847 1848 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 1849 debug_guardpage_minorder() > 0) 1850 return; 1851 1852 /* 1853 * This documents exceptions given to allocations in certain 1854 * contexts that are allowed to allocate outside current's set 1855 * of allowed nodes. 1856 */ 1857 if (!(gfp_mask & __GFP_NOMEMALLOC)) 1858 if (test_thread_flag(TIF_MEMDIE) || 1859 (current->flags & (PF_MEMALLOC | PF_EXITING))) 1860 filter &= ~SHOW_MEM_FILTER_NODES; 1861 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 1862 filter &= ~SHOW_MEM_FILTER_NODES; 1863 1864 if (fmt) { 1865 struct va_format vaf; 1866 va_list args; 1867 1868 va_start(args, fmt); 1869 1870 vaf.fmt = fmt; 1871 vaf.va = &args; 1872 1873 pr_warn("%pV", &vaf); 1874 1875 va_end(args); 1876 } 1877 1878 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 1879 current->comm, order, gfp_mask); 1880 1881 dump_stack(); 1882 if (!should_suppress_show_mem()) 1883 show_mem(filter); 1884} 1885 1886static inline int 1887should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1888 unsigned long did_some_progress, 1889 unsigned long pages_reclaimed) 1890{ 1891 /* Do not loop if specifically requested */ 1892 if (gfp_mask & __GFP_NORETRY) 1893 return 0; 1894 1895 /* Always retry if specifically requested */ 1896 if (gfp_mask & __GFP_NOFAIL) 1897 return 1; 1898 1899 /* 1900 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim 1901 * making forward progress without invoking OOM. Suspend also disables 1902 * storage devices so kswapd will not help. Bail if we are suspending. 1903 */ 1904 if (!did_some_progress && pm_suspended_storage()) 1905 return 0; 1906 1907 /* 1908 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1909 * means __GFP_NOFAIL, but that may not be true in other 1910 * implementations. 1911 */ 1912 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1913 return 1; 1914 1915 /* 1916 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1917 * specified, then we retry until we no longer reclaim any pages 1918 * (above), or we've reclaimed an order of pages at least as 1919 * large as the allocation's order. In both cases, if the 1920 * allocation still fails, we stop retrying. 1921 */ 1922 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1923 return 1; 1924 1925 return 0; 1926} 1927 1928static inline struct page * 1929__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1930 struct zonelist *zonelist, enum zone_type high_zoneidx, 1931 nodemask_t *nodemask, struct zone *preferred_zone, 1932 int migratetype) 1933{ 1934 struct page *page; 1935 1936 /* Acquire the OOM killer lock for the zones in zonelist */ 1937 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1938 schedule_timeout_uninterruptible(1); 1939 return NULL; 1940 } 1941 1942 /* 1943 * Go through the zonelist yet one more time, keep very high watermark 1944 * here, this is only to catch a parallel oom killing, we must fail if 1945 * we're still under heavy pressure. 1946 */ 1947 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1948 order, zonelist, high_zoneidx, 1949 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1950 preferred_zone, migratetype); 1951 if (page) 1952 goto out; 1953 1954 if (!(gfp_mask & __GFP_NOFAIL)) { 1955 /* The OOM killer will not help higher order allocs */ 1956 if (order > PAGE_ALLOC_COSTLY_ORDER) 1957 goto out; 1958 /* The OOM killer does not needlessly kill tasks for lowmem */ 1959 if (high_zoneidx < ZONE_NORMAL) 1960 goto out; 1961 /* 1962 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1963 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1964 * The caller should handle page allocation failure by itself if 1965 * it specifies __GFP_THISNODE. 1966 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1967 */ 1968 if (gfp_mask & __GFP_THISNODE) 1969 goto out; 1970 } 1971 /* Exhausted what can be done so it's blamo time */ 1972 out_of_memory(zonelist, gfp_mask, order, nodemask); 1973 1974out: 1975 clear_zonelist_oom(zonelist, gfp_mask); 1976 return page; 1977} 1978 1979#ifdef CONFIG_COMPACTION 1980/* Try memory compaction for high-order allocations before reclaim */ 1981static struct page * 1982__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1983 struct zonelist *zonelist, enum zone_type high_zoneidx, 1984 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1985 int migratetype, bool sync_migration, 1986 bool *deferred_compaction, 1987 unsigned long *did_some_progress) 1988{ 1989 struct page *page; 1990 1991 if (!order) 1992 return NULL; 1993 1994 if (compaction_deferred(preferred_zone)) { 1995 *deferred_compaction = true; 1996 return NULL; 1997 } 1998 1999 current->flags |= PF_MEMALLOC; 2000 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 2001 nodemask, sync_migration); 2002 current->flags &= ~PF_MEMALLOC; 2003 if (*did_some_progress != COMPACT_SKIPPED) { 2004 2005 /* Page migration frees to the PCP lists but we want merging */ 2006 drain_pages(get_cpu()); 2007 put_cpu(); 2008 2009 page = get_page_from_freelist(gfp_mask, nodemask, 2010 order, zonelist, high_zoneidx, 2011 alloc_flags, preferred_zone, 2012 migratetype); 2013 if (page) { 2014 preferred_zone->compact_considered = 0; 2015 preferred_zone->compact_defer_shift = 0; 2016 count_vm_event(COMPACTSUCCESS); 2017 return page; 2018 } 2019 2020 /* 2021 * It's bad if compaction run occurs and fails. 2022 * The most likely reason is that pages exist, 2023 * but not enough to satisfy watermarks. 2024 */ 2025 count_vm_event(COMPACTFAIL); 2026 2027 /* 2028 * As async compaction considers a subset of pageblocks, only 2029 * defer if the failure was a sync compaction failure. 2030 */ 2031 if (sync_migration) 2032 defer_compaction(preferred_zone); 2033 2034 cond_resched(); 2035 } 2036 2037 return NULL; 2038} 2039#else 2040static inline struct page * 2041__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2042 struct zonelist *zonelist, enum zone_type high_zoneidx, 2043 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2044 int migratetype, bool sync_migration, 2045 bool *deferred_compaction, 2046 unsigned long *did_some_progress) 2047{ 2048 return NULL; 2049} 2050#endif /* CONFIG_COMPACTION */ 2051 2052/* The really slow allocator path where we enter direct reclaim */ 2053static inline struct page * 2054__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 2055 struct zonelist *zonelist, enum zone_type high_zoneidx, 2056 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2057 int migratetype, unsigned long *did_some_progress) 2058{ 2059 struct page *page = NULL; 2060 struct reclaim_state reclaim_state; 2061 bool drained = false; 2062 2063 cond_resched(); 2064 2065 /* We now go into synchronous reclaim */ 2066 cpuset_memory_pressure_bump(); 2067 current->flags |= PF_MEMALLOC; 2068 lockdep_set_current_reclaim_state(gfp_mask); 2069 reclaim_state.reclaimed_slab = 0; 2070 current->reclaim_state = &reclaim_state; 2071 2072 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 2073 2074 current->reclaim_state = NULL; 2075 lockdep_clear_current_reclaim_state(); 2076 current->flags &= ~PF_MEMALLOC; 2077 2078 cond_resched(); 2079 2080 if (unlikely(!(*did_some_progress))) 2081 return NULL; 2082 2083 /* After successful reclaim, reconsider all zones for allocation */ 2084 if (NUMA_BUILD) 2085 zlc_clear_zones_full(zonelist); 2086 2087retry: 2088 page = get_page_from_freelist(gfp_mask, nodemask, order, 2089 zonelist, high_zoneidx, 2090 alloc_flags, preferred_zone, 2091 migratetype); 2092 2093 /* 2094 * If an allocation failed after direct reclaim, it could be because 2095 * pages are pinned on the per-cpu lists. Drain them and try again 2096 */ 2097 if (!page && !drained) { 2098 drain_all_pages(); 2099 drained = true; 2100 goto retry; 2101 } 2102 2103 return page; 2104} 2105 2106/* 2107 * This is called in the allocator slow-path if the allocation request is of 2108 * sufficient urgency to ignore watermarks and take other desperate measures 2109 */ 2110static inline struct page * 2111__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2112 struct zonelist *zonelist, enum zone_type high_zoneidx, 2113 nodemask_t *nodemask, struct zone *preferred_zone, 2114 int migratetype) 2115{ 2116 struct page *page; 2117 2118 do { 2119 page = get_page_from_freelist(gfp_mask, nodemask, order, 2120 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 2121 preferred_zone, migratetype); 2122 2123 if (!page && gfp_mask & __GFP_NOFAIL) 2124 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2125 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2126 2127 return page; 2128} 2129 2130static inline 2131void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 2132 enum zone_type high_zoneidx, 2133 enum zone_type classzone_idx) 2134{ 2135 struct zoneref *z; 2136 struct zone *zone; 2137 2138 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 2139 wakeup_kswapd(zone, order, classzone_idx); 2140} 2141 2142static inline int 2143gfp_to_alloc_flags(gfp_t gfp_mask) 2144{ 2145 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2146 const gfp_t wait = gfp_mask & __GFP_WAIT; 2147 2148 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2149 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2150 2151 /* 2152 * The caller may dip into page reserves a bit more if the caller 2153 * cannot run direct reclaim, or if the caller has realtime scheduling 2154 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2155 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 2156 */ 2157 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2158 2159 if (!wait) { 2160 /* 2161 * Not worth trying to allocate harder for 2162 * __GFP_NOMEMALLOC even if it can't schedule. 2163 */ 2164 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2165 alloc_flags |= ALLOC_HARDER; 2166 /* 2167 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 2168 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 2169 */ 2170 alloc_flags &= ~ALLOC_CPUSET; 2171 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2172 alloc_flags |= ALLOC_HARDER; 2173 2174 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2175 if (!in_interrupt() && 2176 ((current->flags & PF_MEMALLOC) || 2177 unlikely(test_thread_flag(TIF_MEMDIE)))) 2178 alloc_flags |= ALLOC_NO_WATERMARKS; 2179 } 2180 2181 return alloc_flags; 2182} 2183 2184static inline struct page * 2185__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2186 struct zonelist *zonelist, enum zone_type high_zoneidx, 2187 nodemask_t *nodemask, struct zone *preferred_zone, 2188 int migratetype) 2189{ 2190 const gfp_t wait = gfp_mask & __GFP_WAIT; 2191 struct page *page = NULL; 2192 int alloc_flags; 2193 unsigned long pages_reclaimed = 0; 2194 unsigned long did_some_progress; 2195 bool sync_migration = false; 2196 bool deferred_compaction = false; 2197 2198 /* 2199 * In the slowpath, we sanity check order to avoid ever trying to 2200 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2201 * be using allocators in order of preference for an area that is 2202 * too large. 2203 */ 2204 if (order >= MAX_ORDER) { 2205 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2206 return NULL; 2207 } 2208 2209 /* 2210 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2211 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2212 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2213 * using a larger set of nodes after it has established that the 2214 * allowed per node queues are empty and that nodes are 2215 * over allocated. 2216 */ 2217 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2218 goto nopage; 2219 2220restart: 2221 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2222 wake_all_kswapd(order, zonelist, high_zoneidx, 2223 zone_idx(preferred_zone)); 2224 2225 /* 2226 * OK, we're below the kswapd watermark and have kicked background 2227 * reclaim. Now things get more complex, so set up alloc_flags according 2228 * to how we want to proceed. 2229 */ 2230 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2231 2232 /* 2233 * Find the true preferred zone if the allocation is unconstrained by 2234 * cpusets. 2235 */ 2236 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) 2237 first_zones_zonelist(zonelist, high_zoneidx, NULL, 2238 &preferred_zone); 2239 2240rebalance: 2241 /* This is the last chance, in general, before the goto nopage. */ 2242 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2243 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2244 preferred_zone, migratetype); 2245 if (page) 2246 goto got_pg; 2247 2248 /* Allocate without watermarks if the context allows */ 2249 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2250 page = __alloc_pages_high_priority(gfp_mask, order, 2251 zonelist, high_zoneidx, nodemask, 2252 preferred_zone, migratetype); 2253 if (page) 2254 goto got_pg; 2255 } 2256 2257 /* Atomic allocations - we can't balance anything */ 2258 if (!wait) 2259 goto nopage; 2260 2261 /* Avoid recursion of direct reclaim */ 2262 if (current->flags & PF_MEMALLOC) 2263 goto nopage; 2264 2265 /* Avoid allocations with no watermarks from looping endlessly */ 2266 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2267 goto nopage; 2268 2269 /* 2270 * Try direct compaction. The first pass is asynchronous. Subsequent 2271 * attempts after direct reclaim are synchronous 2272 */ 2273 page = __alloc_pages_direct_compact(gfp_mask, order, 2274 zonelist, high_zoneidx, 2275 nodemask, 2276 alloc_flags, preferred_zone, 2277 migratetype, sync_migration, 2278 &deferred_compaction, 2279 &did_some_progress); 2280 if (page) 2281 goto got_pg; 2282 sync_migration = true; 2283 2284 /* 2285 * If compaction is deferred for high-order allocations, it is because 2286 * sync compaction recently failed. In this is the case and the caller 2287 * has requested the system not be heavily disrupted, fail the 2288 * allocation now instead of entering direct reclaim 2289 */ 2290 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD)) 2291 goto nopage; 2292 2293 /* Try direct reclaim and then allocating */ 2294 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2295 zonelist, high_zoneidx, 2296 nodemask, 2297 alloc_flags, preferred_zone, 2298 migratetype, &did_some_progress); 2299 if (page) 2300 goto got_pg; 2301 2302 /* 2303 * If we failed to make any progress reclaiming, then we are 2304 * running out of options and have to consider going OOM 2305 */ 2306 if (!did_some_progress) { 2307 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2308 if (oom_killer_disabled) 2309 goto nopage; 2310 page = __alloc_pages_may_oom(gfp_mask, order, 2311 zonelist, high_zoneidx, 2312 nodemask, preferred_zone, 2313 migratetype); 2314 if (page) 2315 goto got_pg; 2316 2317 if (!(gfp_mask & __GFP_NOFAIL)) { 2318 /* 2319 * The oom killer is not called for high-order 2320 * allocations that may fail, so if no progress 2321 * is being made, there are no other options and 2322 * retrying is unlikely to help. 2323 */ 2324 if (order > PAGE_ALLOC_COSTLY_ORDER) 2325 goto nopage; 2326 /* 2327 * The oom killer is not called for lowmem 2328 * allocations to prevent needlessly killing 2329 * innocent tasks. 2330 */ 2331 if (high_zoneidx < ZONE_NORMAL) 2332 goto nopage; 2333 } 2334 2335 goto restart; 2336 } 2337 } 2338 2339 /* Check if we should retry the allocation */ 2340 pages_reclaimed += did_some_progress; 2341 if (should_alloc_retry(gfp_mask, order, did_some_progress, 2342 pages_reclaimed)) { 2343 /* Wait for some write requests to complete then retry */ 2344 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2345 goto rebalance; 2346 } else { 2347 /* 2348 * High-order allocations do not necessarily loop after 2349 * direct reclaim and reclaim/compaction depends on compaction 2350 * being called after reclaim so call directly if necessary 2351 */ 2352 page = __alloc_pages_direct_compact(gfp_mask, order, 2353 zonelist, high_zoneidx, 2354 nodemask, 2355 alloc_flags, preferred_zone, 2356 migratetype, sync_migration, 2357 &deferred_compaction, 2358 &did_some_progress); 2359 if (page) 2360 goto got_pg; 2361 } 2362 2363nopage: 2364 warn_alloc_failed(gfp_mask, order, NULL); 2365 return page; 2366got_pg: 2367 if (kmemcheck_enabled) 2368 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2369 return page; 2370 2371} 2372 2373/* 2374 * This is the 'heart' of the zoned buddy allocator. 2375 */ 2376struct page * 2377__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2378 struct zonelist *zonelist, nodemask_t *nodemask) 2379{ 2380 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2381 struct zone *preferred_zone; 2382 struct page *page; 2383 int migratetype = allocflags_to_migratetype(gfp_mask); 2384 2385 gfp_mask &= gfp_allowed_mask; 2386 2387 lockdep_trace_alloc(gfp_mask); 2388 2389 might_sleep_if(gfp_mask & __GFP_WAIT); 2390 2391 if (should_fail_alloc_page(gfp_mask, order)) 2392 return NULL; 2393 2394 /* 2395 * Check the zones suitable for the gfp_mask contain at least one 2396 * valid zone. It's possible to have an empty zonelist as a result 2397 * of GFP_THISNODE and a memoryless node 2398 */ 2399 if (unlikely(!zonelist->_zonerefs->zone)) 2400 return NULL; 2401 2402 get_mems_allowed(); 2403 /* The preferred zone is used for statistics later */ 2404 first_zones_zonelist(zonelist, high_zoneidx, 2405 nodemask ? : &cpuset_current_mems_allowed, 2406 &preferred_zone); 2407 if (!preferred_zone) { 2408 put_mems_allowed(); 2409 return NULL; 2410 } 2411 2412 /* First allocation attempt */ 2413 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2414 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2415 preferred_zone, migratetype); 2416 if (unlikely(!page)) 2417 page = __alloc_pages_slowpath(gfp_mask, order, 2418 zonelist, high_zoneidx, nodemask, 2419 preferred_zone, migratetype); 2420 put_mems_allowed(); 2421 2422 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2423 return page; 2424} 2425EXPORT_SYMBOL(__alloc_pages_nodemask); 2426 2427/* 2428 * Common helper functions. 2429 */ 2430unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2431{ 2432 struct page *page; 2433 2434 /* 2435 * __get_free_pages() returns a 32-bit address, which cannot represent 2436 * a highmem page 2437 */ 2438 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2439 2440 page = alloc_pages(gfp_mask, order); 2441 if (!page) 2442 return 0; 2443 return (unsigned long) page_address(page); 2444} 2445EXPORT_SYMBOL(__get_free_pages); 2446 2447unsigned long get_zeroed_page(gfp_t gfp_mask) 2448{ 2449 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2450} 2451EXPORT_SYMBOL(get_zeroed_page); 2452 2453void __free_pages(struct page *page, unsigned int order) 2454{ 2455 if (put_page_testzero(page)) { 2456 if (order == 0) 2457 free_hot_cold_page(page, 0); 2458 else 2459 __free_pages_ok(page, order); 2460 } 2461} 2462 2463EXPORT_SYMBOL(__free_pages); 2464 2465void free_pages(unsigned long addr, unsigned int order) 2466{ 2467 if (addr != 0) { 2468 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2469 __free_pages(virt_to_page((void *)addr), order); 2470 } 2471} 2472 2473EXPORT_SYMBOL(free_pages); 2474 2475static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 2476{ 2477 if (addr) { 2478 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2479 unsigned long used = addr + PAGE_ALIGN(size); 2480 2481 split_page(virt_to_page((void *)addr), order); 2482 while (used < alloc_end) { 2483 free_page(used); 2484 used += PAGE_SIZE; 2485 } 2486 } 2487 return (void *)addr; 2488} 2489 2490/** 2491 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2492 * @size: the number of bytes to allocate 2493 * @gfp_mask: GFP flags for the allocation 2494 * 2495 * This function is similar to alloc_pages(), except that it allocates the 2496 * minimum number of pages to satisfy the request. alloc_pages() can only 2497 * allocate memory in power-of-two pages. 2498 * 2499 * This function is also limited by MAX_ORDER. 2500 * 2501 * Memory allocated by this function must be released by free_pages_exact(). 2502 */ 2503void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2504{ 2505 unsigned int order = get_order(size); 2506 unsigned long addr; 2507 2508 addr = __get_free_pages(gfp_mask, order); 2509 return make_alloc_exact(addr, order, size); 2510} 2511EXPORT_SYMBOL(alloc_pages_exact); 2512 2513/** 2514 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 2515 * pages on a node. 2516 * @nid: the preferred node ID where memory should be allocated 2517 * @size: the number of bytes to allocate 2518 * @gfp_mask: GFP flags for the allocation 2519 * 2520 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 2521 * back. 2522 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 2523 * but is not exact. 2524 */ 2525void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 2526{ 2527 unsigned order = get_order(size); 2528 struct page *p = alloc_pages_node(nid, gfp_mask, order); 2529 if (!p) 2530 return NULL; 2531 return make_alloc_exact((unsigned long)page_address(p), order, size); 2532} 2533EXPORT_SYMBOL(alloc_pages_exact_nid); 2534 2535/** 2536 * free_pages_exact - release memory allocated via alloc_pages_exact() 2537 * @virt: the value returned by alloc_pages_exact. 2538 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2539 * 2540 * Release the memory allocated by a previous call to alloc_pages_exact. 2541 */ 2542void free_pages_exact(void *virt, size_t size) 2543{ 2544 unsigned long addr = (unsigned long)virt; 2545 unsigned long end = addr + PAGE_ALIGN(size); 2546 2547 while (addr < end) { 2548 free_page(addr); 2549 addr += PAGE_SIZE; 2550 } 2551} 2552EXPORT_SYMBOL(free_pages_exact); 2553 2554static unsigned int nr_free_zone_pages(int offset) 2555{ 2556 struct zoneref *z; 2557 struct zone *zone; 2558 2559 /* Just pick one node, since fallback list is circular */ 2560 unsigned int sum = 0; 2561 2562 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2563 2564 for_each_zone_zonelist(zone, z, zonelist, offset) { 2565 unsigned long size = zone->present_pages; 2566 unsigned long high = high_wmark_pages(zone); 2567 if (size > high) 2568 sum += size - high; 2569 } 2570 2571 return sum; 2572} 2573 2574/* 2575 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2576 */ 2577unsigned int nr_free_buffer_pages(void) 2578{ 2579 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2580} 2581EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2582 2583/* 2584 * Amount of free RAM allocatable within all zones 2585 */ 2586unsigned int nr_free_pagecache_pages(void) 2587{ 2588 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2589} 2590 2591static inline void show_node(struct zone *zone) 2592{ 2593 if (NUMA_BUILD) 2594 printk("Node %d ", zone_to_nid(zone)); 2595} 2596 2597void si_meminfo(struct sysinfo *val) 2598{ 2599 val->totalram = totalram_pages; 2600 val->sharedram = 0; 2601 val->freeram = global_page_state(NR_FREE_PAGES); 2602 val->bufferram = nr_blockdev_pages(); 2603 val->totalhigh = totalhigh_pages; 2604 val->freehigh = nr_free_highpages(); 2605 val->mem_unit = PAGE_SIZE; 2606} 2607 2608EXPORT_SYMBOL(si_meminfo); 2609 2610#ifdef CONFIG_NUMA 2611void si_meminfo_node(struct sysinfo *val, int nid) 2612{ 2613 pg_data_t *pgdat = NODE_DATA(nid); 2614 2615 val->totalram = pgdat->node_present_pages; 2616 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2617#ifdef CONFIG_HIGHMEM 2618 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2619 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2620 NR_FREE_PAGES); 2621#else 2622 val->totalhigh = 0; 2623 val->freehigh = 0; 2624#endif 2625 val->mem_unit = PAGE_SIZE; 2626} 2627#endif 2628 2629/* 2630 * Determine whether the node should be displayed or not, depending on whether 2631 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 2632 */ 2633bool skip_free_areas_node(unsigned int flags, int nid) 2634{ 2635 bool ret = false; 2636 2637 if (!(flags & SHOW_MEM_FILTER_NODES)) 2638 goto out; 2639 2640 get_mems_allowed(); 2641 ret = !node_isset(nid, cpuset_current_mems_allowed); 2642 put_mems_allowed(); 2643out: 2644 return ret; 2645} 2646 2647#define K(x) ((x) << (PAGE_SHIFT-10)) 2648 2649/* 2650 * Show free area list (used inside shift_scroll-lock stuff) 2651 * We also calculate the percentage fragmentation. We do this by counting the 2652 * memory on each free list with the exception of the first item on the list. 2653 * Suppresses nodes that are not allowed by current's cpuset if 2654 * SHOW_MEM_FILTER_NODES is passed. 2655 */ 2656void show_free_areas(unsigned int filter) 2657{ 2658 int cpu; 2659 struct zone *zone; 2660 2661 for_each_populated_zone(zone) { 2662 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2663 continue; 2664 show_node(zone); 2665 printk("%s per-cpu:\n", zone->name); 2666 2667 for_each_online_cpu(cpu) { 2668 struct per_cpu_pageset *pageset; 2669 2670 pageset = per_cpu_ptr(zone->pageset, cpu); 2671 2672 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2673 cpu, pageset->pcp.high, 2674 pageset->pcp.batch, pageset->pcp.count); 2675 } 2676 } 2677 2678 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2679 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2680 " unevictable:%lu" 2681 " dirty:%lu writeback:%lu unstable:%lu\n" 2682 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2683 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2684 global_page_state(NR_ACTIVE_ANON), 2685 global_page_state(NR_INACTIVE_ANON), 2686 global_page_state(NR_ISOLATED_ANON), 2687 global_page_state(NR_ACTIVE_FILE), 2688 global_page_state(NR_INACTIVE_FILE), 2689 global_page_state(NR_ISOLATED_FILE), 2690 global_page_state(NR_UNEVICTABLE), 2691 global_page_state(NR_FILE_DIRTY), 2692 global_page_state(NR_WRITEBACK), 2693 global_page_state(NR_UNSTABLE_NFS), 2694 global_page_state(NR_FREE_PAGES), 2695 global_page_state(NR_SLAB_RECLAIMABLE), 2696 global_page_state(NR_SLAB_UNRECLAIMABLE), 2697 global_page_state(NR_FILE_MAPPED), 2698 global_page_state(NR_SHMEM), 2699 global_page_state(NR_PAGETABLE), 2700 global_page_state(NR_BOUNCE)); 2701 2702 for_each_populated_zone(zone) { 2703 int i; 2704 2705 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2706 continue; 2707 show_node(zone); 2708 printk("%s" 2709 " free:%lukB" 2710 " min:%lukB" 2711 " low:%lukB" 2712 " high:%lukB" 2713 " active_anon:%lukB" 2714 " inactive_anon:%lukB" 2715 " active_file:%lukB" 2716 " inactive_file:%lukB" 2717 " unevictable:%lukB" 2718 " isolated(anon):%lukB" 2719 " isolated(file):%lukB" 2720 " present:%lukB" 2721 " mlocked:%lukB" 2722 " dirty:%lukB" 2723 " writeback:%lukB" 2724 " mapped:%lukB" 2725 " shmem:%lukB" 2726 " slab_reclaimable:%lukB" 2727 " slab_unreclaimable:%lukB" 2728 " kernel_stack:%lukB" 2729 " pagetables:%lukB" 2730 " unstable:%lukB" 2731 " bounce:%lukB" 2732 " writeback_tmp:%lukB" 2733 " pages_scanned:%lu" 2734 " all_unreclaimable? %s" 2735 "\n", 2736 zone->name, 2737 K(zone_page_state(zone, NR_FREE_PAGES)), 2738 K(min_wmark_pages(zone)), 2739 K(low_wmark_pages(zone)), 2740 K(high_wmark_pages(zone)), 2741 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2742 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2743 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2744 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2745 K(zone_page_state(zone, NR_UNEVICTABLE)), 2746 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2747 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2748 K(zone->present_pages), 2749 K(zone_page_state(zone, NR_MLOCK)), 2750 K(zone_page_state(zone, NR_FILE_DIRTY)), 2751 K(zone_page_state(zone, NR_WRITEBACK)), 2752 K(zone_page_state(zone, NR_FILE_MAPPED)), 2753 K(zone_page_state(zone, NR_SHMEM)), 2754 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2755 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2756 zone_page_state(zone, NR_KERNEL_STACK) * 2757 THREAD_SIZE / 1024, 2758 K(zone_page_state(zone, NR_PAGETABLE)), 2759 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2760 K(zone_page_state(zone, NR_BOUNCE)), 2761 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2762 zone->pages_scanned, 2763 (zone->all_unreclaimable ? "yes" : "no") 2764 ); 2765 printk("lowmem_reserve[]:"); 2766 for (i = 0; i < MAX_NR_ZONES; i++) 2767 printk(" %lu", zone->lowmem_reserve[i]); 2768 printk("\n"); 2769 } 2770 2771 for_each_populated_zone(zone) { 2772 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2773 2774 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2775 continue; 2776 show_node(zone); 2777 printk("%s: ", zone->name); 2778 2779 spin_lock_irqsave(&zone->lock, flags); 2780 for (order = 0; order < MAX_ORDER; order++) { 2781 nr[order] = zone->free_area[order].nr_free; 2782 total += nr[order] << order; 2783 } 2784 spin_unlock_irqrestore(&zone->lock, flags); 2785 for (order = 0; order < MAX_ORDER; order++) 2786 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2787 printk("= %lukB\n", K(total)); 2788 } 2789 2790 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2791 2792 show_swap_cache_info(); 2793} 2794 2795static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2796{ 2797 zoneref->zone = zone; 2798 zoneref->zone_idx = zone_idx(zone); 2799} 2800 2801/* 2802 * Builds allocation fallback zone lists. 2803 * 2804 * Add all populated zones of a node to the zonelist. 2805 */ 2806static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2807 int nr_zones, enum zone_type zone_type) 2808{ 2809 struct zone *zone; 2810 2811 BUG_ON(zone_type >= MAX_NR_ZONES); 2812 zone_type++; 2813 2814 do { 2815 zone_type--; 2816 zone = pgdat->node_zones + zone_type; 2817 if (populated_zone(zone)) { 2818 zoneref_set_zone(zone, 2819 &zonelist->_zonerefs[nr_zones++]); 2820 check_highest_zone(zone_type); 2821 } 2822 2823 } while (zone_type); 2824 return nr_zones; 2825} 2826 2827 2828/* 2829 * zonelist_order: 2830 * 0 = automatic detection of better ordering. 2831 * 1 = order by ([node] distance, -zonetype) 2832 * 2 = order by (-zonetype, [node] distance) 2833 * 2834 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2835 * the same zonelist. So only NUMA can configure this param. 2836 */ 2837#define ZONELIST_ORDER_DEFAULT 0 2838#define ZONELIST_ORDER_NODE 1 2839#define ZONELIST_ORDER_ZONE 2 2840 2841/* zonelist order in the kernel. 2842 * set_zonelist_order() will set this to NODE or ZONE. 2843 */ 2844static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2845static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2846 2847 2848#ifdef CONFIG_NUMA 2849/* The value user specified ....changed by config */ 2850static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2851/* string for sysctl */ 2852#define NUMA_ZONELIST_ORDER_LEN 16 2853char numa_zonelist_order[16] = "default"; 2854 2855/* 2856 * interface for configure zonelist ordering. 2857 * command line option "numa_zonelist_order" 2858 * = "[dD]efault - default, automatic configuration. 2859 * = "[nN]ode - order by node locality, then by zone within node 2860 * = "[zZ]one - order by zone, then by locality within zone 2861 */ 2862 2863static int __parse_numa_zonelist_order(char *s) 2864{ 2865 if (*s == 'd' || *s == 'D') { 2866 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2867 } else if (*s == 'n' || *s == 'N') { 2868 user_zonelist_order = ZONELIST_ORDER_NODE; 2869 } else if (*s == 'z' || *s == 'Z') { 2870 user_zonelist_order = ZONELIST_ORDER_ZONE; 2871 } else { 2872 printk(KERN_WARNING 2873 "Ignoring invalid numa_zonelist_order value: " 2874 "%s\n", s); 2875 return -EINVAL; 2876 } 2877 return 0; 2878} 2879 2880static __init int setup_numa_zonelist_order(char *s) 2881{ 2882 int ret; 2883 2884 if (!s) 2885 return 0; 2886 2887 ret = __parse_numa_zonelist_order(s); 2888 if (ret == 0) 2889 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 2890 2891 return ret; 2892} 2893early_param("numa_zonelist_order", setup_numa_zonelist_order); 2894 2895/* 2896 * sysctl handler for numa_zonelist_order 2897 */ 2898int numa_zonelist_order_handler(ctl_table *table, int write, 2899 void __user *buffer, size_t *length, 2900 loff_t *ppos) 2901{ 2902 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2903 int ret; 2904 static DEFINE_MUTEX(zl_order_mutex); 2905 2906 mutex_lock(&zl_order_mutex); 2907 if (write) 2908 strcpy(saved_string, (char*)table->data); 2909 ret = proc_dostring(table, write, buffer, length, ppos); 2910 if (ret) 2911 goto out; 2912 if (write) { 2913 int oldval = user_zonelist_order; 2914 if (__parse_numa_zonelist_order((char*)table->data)) { 2915 /* 2916 * bogus value. restore saved string 2917 */ 2918 strncpy((char*)table->data, saved_string, 2919 NUMA_ZONELIST_ORDER_LEN); 2920 user_zonelist_order = oldval; 2921 } else if (oldval != user_zonelist_order) { 2922 mutex_lock(&zonelists_mutex); 2923 build_all_zonelists(NULL); 2924 mutex_unlock(&zonelists_mutex); 2925 } 2926 } 2927out: 2928 mutex_unlock(&zl_order_mutex); 2929 return ret; 2930} 2931 2932 2933#define MAX_NODE_LOAD (nr_online_nodes) 2934static int node_load[MAX_NUMNODES]; 2935 2936/** 2937 * find_next_best_node - find the next node that should appear in a given node's fallback list 2938 * @node: node whose fallback list we're appending 2939 * @used_node_mask: nodemask_t of already used nodes 2940 * 2941 * We use a number of factors to determine which is the next node that should 2942 * appear on a given node's fallback list. The node should not have appeared 2943 * already in @node's fallback list, and it should be the next closest node 2944 * according to the distance array (which contains arbitrary distance values 2945 * from each node to each node in the system), and should also prefer nodes 2946 * with no CPUs, since presumably they'll have very little allocation pressure 2947 * on them otherwise. 2948 * It returns -1 if no node is found. 2949 */ 2950static int find_next_best_node(int node, nodemask_t *used_node_mask) 2951{ 2952 int n, val; 2953 int min_val = INT_MAX; 2954 int best_node = -1; 2955 const struct cpumask *tmp = cpumask_of_node(0); 2956 2957 /* Use the local node if we haven't already */ 2958 if (!node_isset(node, *used_node_mask)) { 2959 node_set(node, *used_node_mask); 2960 return node; 2961 } 2962 2963 for_each_node_state(n, N_HIGH_MEMORY) { 2964 2965 /* Don't want a node to appear more than once */ 2966 if (node_isset(n, *used_node_mask)) 2967 continue; 2968 2969 /* Use the distance array to find the distance */ 2970 val = node_distance(node, n); 2971 2972 /* Penalize nodes under us ("prefer the next node") */ 2973 val += (n < node); 2974 2975 /* Give preference to headless and unused nodes */ 2976 tmp = cpumask_of_node(n); 2977 if (!cpumask_empty(tmp)) 2978 val += PENALTY_FOR_NODE_WITH_CPUS; 2979 2980 /* Slight preference for less loaded node */ 2981 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2982 val += node_load[n]; 2983 2984 if (val < min_val) { 2985 min_val = val; 2986 best_node = n; 2987 } 2988 } 2989 2990 if (best_node >= 0) 2991 node_set(best_node, *used_node_mask); 2992 2993 return best_node; 2994} 2995 2996 2997/* 2998 * Build zonelists ordered by node and zones within node. 2999 * This results in maximum locality--normal zone overflows into local 3000 * DMA zone, if any--but risks exhausting DMA zone. 3001 */ 3002static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 3003{ 3004 int j; 3005 struct zonelist *zonelist; 3006 3007 zonelist = &pgdat->node_zonelists[0]; 3008 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 3009 ; 3010 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3011 MAX_NR_ZONES - 1); 3012 zonelist->_zonerefs[j].zone = NULL; 3013 zonelist->_zonerefs[j].zone_idx = 0; 3014} 3015 3016/* 3017 * Build gfp_thisnode zonelists 3018 */ 3019static void build_thisnode_zonelists(pg_data_t *pgdat) 3020{ 3021 int j; 3022 struct zonelist *zonelist; 3023 3024 zonelist = &pgdat->node_zonelists[1]; 3025 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3026 zonelist->_zonerefs[j].zone = NULL; 3027 zonelist->_zonerefs[j].zone_idx = 0; 3028} 3029 3030/* 3031 * Build zonelists ordered by zone and nodes within zones. 3032 * This results in conserving DMA zone[s] until all Normal memory is 3033 * exhausted, but results in overflowing to remote node while memory 3034 * may still exist in local DMA zone. 3035 */ 3036static int node_order[MAX_NUMNODES]; 3037 3038static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 3039{ 3040 int pos, j, node; 3041 int zone_type; /* needs to be signed */ 3042 struct zone *z; 3043 struct zonelist *zonelist; 3044 3045 zonelist = &pgdat->node_zonelists[0]; 3046 pos = 0; 3047 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 3048 for (j = 0; j < nr_nodes; j++) { 3049 node = node_order[j]; 3050 z = &NODE_DATA(node)->node_zones[zone_type]; 3051 if (populated_zone(z)) { 3052 zoneref_set_zone(z, 3053 &zonelist->_zonerefs[pos++]); 3054 check_highest_zone(zone_type); 3055 } 3056 } 3057 } 3058 zonelist->_zonerefs[pos].zone = NULL; 3059 zonelist->_zonerefs[pos].zone_idx = 0; 3060} 3061 3062static int default_zonelist_order(void) 3063{ 3064 int nid, zone_type; 3065 unsigned long low_kmem_size,total_size; 3066 struct zone *z; 3067 int average_size; 3068 /* 3069 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 3070 * If they are really small and used heavily, the system can fall 3071 * into OOM very easily. 3072 * This function detect ZONE_DMA/DMA32 size and configures zone order. 3073 */ 3074 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 3075 low_kmem_size = 0; 3076 total_size = 0; 3077 for_each_online_node(nid) { 3078 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 3079 z = &NODE_DATA(nid)->node_zones[zone_type]; 3080 if (populated_zone(z)) { 3081 if (zone_type < ZONE_NORMAL) 3082 low_kmem_size += z->present_pages; 3083 total_size += z->present_pages; 3084 } else if (zone_type == ZONE_NORMAL) { 3085 /* 3086 * If any node has only lowmem, then node order 3087 * is preferred to allow kernel allocations 3088 * locally; otherwise, they can easily infringe 3089 * on other nodes when there is an abundance of 3090 * lowmem available to allocate from. 3091 */ 3092 return ZONELIST_ORDER_NODE; 3093 } 3094 } 3095 } 3096 if (!low_kmem_size || /* there are no DMA area. */ 3097 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 3098 return ZONELIST_ORDER_NODE; 3099 /* 3100 * look into each node's config. 3101 * If there is a node whose DMA/DMA32 memory is very big area on 3102 * local memory, NODE_ORDER may be suitable. 3103 */ 3104 average_size = total_size / 3105 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 3106 for_each_online_node(nid) { 3107 low_kmem_size = 0; 3108 total_size = 0; 3109 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 3110 z = &NODE_DATA(nid)->node_zones[zone_type]; 3111 if (populated_zone(z)) { 3112 if (zone_type < ZONE_NORMAL) 3113 low_kmem_size += z->present_pages; 3114 total_size += z->present_pages; 3115 } 3116 } 3117 if (low_kmem_size && 3118 total_size > average_size && /* ignore small node */ 3119 low_kmem_size > total_size * 70/100) 3120 return ZONELIST_ORDER_NODE; 3121 } 3122 return ZONELIST_ORDER_ZONE; 3123} 3124 3125static void set_zonelist_order(void) 3126{ 3127 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 3128 current_zonelist_order = default_zonelist_order(); 3129 else 3130 current_zonelist_order = user_zonelist_order; 3131} 3132 3133static void build_zonelists(pg_data_t *pgdat) 3134{ 3135 int j, node, load; 3136 enum zone_type i; 3137 nodemask_t used_mask; 3138 int local_node, prev_node; 3139 struct zonelist *zonelist; 3140 int order = current_zonelist_order; 3141 3142 /* initialize zonelists */ 3143 for (i = 0; i < MAX_ZONELISTS; i++) { 3144 zonelist = pgdat->node_zonelists + i; 3145 zonelist->_zonerefs[0].zone = NULL; 3146 zonelist->_zonerefs[0].zone_idx = 0; 3147 } 3148 3149 /* NUMA-aware ordering of nodes */ 3150 local_node = pgdat->node_id; 3151 load = nr_online_nodes; 3152 prev_node = local_node; 3153 nodes_clear(used_mask); 3154 3155 memset(node_order, 0, sizeof(node_order)); 3156 j = 0; 3157 3158 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 3159 int distance = node_distance(local_node, node); 3160 3161 /* 3162 * If another node is sufficiently far away then it is better 3163 * to reclaim pages in a zone before going off node. 3164 */ 3165 if (distance > RECLAIM_DISTANCE) 3166 zone_reclaim_mode = 1; 3167 3168 /* 3169 * We don't want to pressure a particular node. 3170 * So adding penalty to the first node in same 3171 * distance group to make it round-robin. 3172 */ 3173 if (distance != node_distance(local_node, prev_node)) 3174 node_load[node] = load; 3175 3176 prev_node = node; 3177 load--; 3178 if (order == ZONELIST_ORDER_NODE) 3179 build_zonelists_in_node_order(pgdat, node); 3180 else 3181 node_order[j++] = node; /* remember order */ 3182 } 3183 3184 if (order == ZONELIST_ORDER_ZONE) { 3185 /* calculate node order -- i.e., DMA last! */ 3186 build_zonelists_in_zone_order(pgdat, j); 3187 } 3188 3189 build_thisnode_zonelists(pgdat); 3190} 3191 3192/* Construct the zonelist performance cache - see further mmzone.h */ 3193static void build_zonelist_cache(pg_data_t *pgdat) 3194{ 3195 struct zonelist *zonelist; 3196 struct zonelist_cache *zlc; 3197 struct zoneref *z; 3198 3199 zonelist = &pgdat->node_zonelists[0]; 3200 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 3201 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 3202 for (z = zonelist->_zonerefs; z->zone; z++) 3203 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 3204} 3205 3206#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3207/* 3208 * Return node id of node used for "local" allocations. 3209 * I.e., first node id of first zone in arg node's generic zonelist. 3210 * Used for initializing percpu 'numa_mem', which is used primarily 3211 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 3212 */ 3213int local_memory_node(int node) 3214{ 3215 struct zone *zone; 3216 3217 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 3218 gfp_zone(GFP_KERNEL), 3219 NULL, 3220 &zone); 3221 return zone->node; 3222} 3223#endif 3224 3225#else /* CONFIG_NUMA */ 3226 3227static void set_zonelist_order(void) 3228{ 3229 current_zonelist_order = ZONELIST_ORDER_ZONE; 3230} 3231 3232static void build_zonelists(pg_data_t *pgdat) 3233{ 3234 int node, local_node; 3235 enum zone_type j; 3236 struct zonelist *zonelist; 3237 3238 local_node = pgdat->node_id; 3239 3240 zonelist = &pgdat->node_zonelists[0]; 3241 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3242 3243 /* 3244 * Now we build the zonelist so that it contains the zones 3245 * of all the other nodes. 3246 * We don't want to pressure a particular node, so when 3247 * building the zones for node N, we make sure that the 3248 * zones coming right after the local ones are those from 3249 * node N+1 (modulo N) 3250 */ 3251 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 3252 if (!node_online(node)) 3253 continue; 3254 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3255 MAX_NR_ZONES - 1); 3256 } 3257 for (node = 0; node < local_node; node++) { 3258 if (!node_online(node)) 3259 continue; 3260 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3261 MAX_NR_ZONES - 1); 3262 } 3263 3264 zonelist->_zonerefs[j].zone = NULL; 3265 zonelist->_zonerefs[j].zone_idx = 0; 3266} 3267 3268/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3269static void build_zonelist_cache(pg_data_t *pgdat) 3270{ 3271 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3272} 3273 3274#endif /* CONFIG_NUMA */ 3275 3276/* 3277 * Boot pageset table. One per cpu which is going to be used for all 3278 * zones and all nodes. The parameters will be set in such a way 3279 * that an item put on a list will immediately be handed over to 3280 * the buddy list. This is safe since pageset manipulation is done 3281 * with interrupts disabled. 3282 * 3283 * The boot_pagesets must be kept even after bootup is complete for 3284 * unused processors and/or zones. They do play a role for bootstrapping 3285 * hotplugged processors. 3286 * 3287 * zoneinfo_show() and maybe other functions do 3288 * not check if the processor is online before following the pageset pointer. 3289 * Other parts of the kernel may not check if the zone is available. 3290 */ 3291static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3292static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3293static void setup_zone_pageset(struct zone *zone); 3294 3295/* 3296 * Global mutex to protect against size modification of zonelists 3297 * as well as to serialize pageset setup for the new populated zone. 3298 */ 3299DEFINE_MUTEX(zonelists_mutex); 3300 3301/* return values int ....just for stop_machine() */ 3302static __init_refok int __build_all_zonelists(void *data) 3303{ 3304 int nid; 3305 int cpu; 3306 3307#ifdef CONFIG_NUMA 3308 memset(node_load, 0, sizeof(node_load)); 3309#endif 3310 for_each_online_node(nid) { 3311 pg_data_t *pgdat = NODE_DATA(nid); 3312 3313 build_zonelists(pgdat); 3314 build_zonelist_cache(pgdat); 3315 } 3316 3317 /* 3318 * Initialize the boot_pagesets that are going to be used 3319 * for bootstrapping processors. The real pagesets for 3320 * each zone will be allocated later when the per cpu 3321 * allocator is available. 3322 * 3323 * boot_pagesets are used also for bootstrapping offline 3324 * cpus if the system is already booted because the pagesets 3325 * are needed to initialize allocators on a specific cpu too. 3326 * F.e. the percpu allocator needs the page allocator which 3327 * needs the percpu allocator in order to allocate its pagesets 3328 * (a chicken-egg dilemma). 3329 */ 3330 for_each_possible_cpu(cpu) { 3331 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3332 3333#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3334 /* 3335 * We now know the "local memory node" for each node-- 3336 * i.e., the node of the first zone in the generic zonelist. 3337 * Set up numa_mem percpu variable for on-line cpus. During 3338 * boot, only the boot cpu should be on-line; we'll init the 3339 * secondary cpus' numa_mem as they come on-line. During 3340 * node/memory hotplug, we'll fixup all on-line cpus. 3341 */ 3342 if (cpu_online(cpu)) 3343 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3344#endif 3345 } 3346 3347 return 0; 3348} 3349 3350/* 3351 * Called with zonelists_mutex held always 3352 * unless system_state == SYSTEM_BOOTING. 3353 */ 3354void __ref build_all_zonelists(void *data) 3355{ 3356 set_zonelist_order(); 3357 3358 if (system_state == SYSTEM_BOOTING) { 3359 __build_all_zonelists(NULL); 3360 mminit_verify_zonelist(); 3361 cpuset_init_current_mems_allowed(); 3362 } else { 3363 /* we have to stop all cpus to guarantee there is no user 3364 of zonelist */ 3365#ifdef CONFIG_MEMORY_HOTPLUG 3366 if (data) 3367 setup_zone_pageset((struct zone *)data); 3368#endif 3369 stop_machine(__build_all_zonelists, NULL, NULL); 3370 /* cpuset refresh routine should be here */ 3371 } 3372 vm_total_pages = nr_free_pagecache_pages(); 3373 /* 3374 * Disable grouping by mobility if the number of pages in the 3375 * system is too low to allow the mechanism to work. It would be 3376 * more accurate, but expensive to check per-zone. This check is 3377 * made on memory-hotadd so a system can start with mobility 3378 * disabled and enable it later 3379 */ 3380 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3381 page_group_by_mobility_disabled = 1; 3382 else 3383 page_group_by_mobility_disabled = 0; 3384 3385 printk("Built %i zonelists in %s order, mobility grouping %s. " 3386 "Total pages: %ld\n", 3387 nr_online_nodes, 3388 zonelist_order_name[current_zonelist_order], 3389 page_group_by_mobility_disabled ? "off" : "on", 3390 vm_total_pages); 3391#ifdef CONFIG_NUMA 3392 printk("Policy zone: %s\n", zone_names[policy_zone]); 3393#endif 3394} 3395 3396/* 3397 * Helper functions to size the waitqueue hash table. 3398 * Essentially these want to choose hash table sizes sufficiently 3399 * large so that collisions trying to wait on pages are rare. 3400 * But in fact, the number of active page waitqueues on typical 3401 * systems is ridiculously low, less than 200. So this is even 3402 * conservative, even though it seems large. 3403 * 3404 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3405 * waitqueues, i.e. the size of the waitq table given the number of pages. 3406 */ 3407#define PAGES_PER_WAITQUEUE 256 3408 3409#ifndef CONFIG_MEMORY_HOTPLUG 3410static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3411{ 3412 unsigned long size = 1; 3413 3414 pages /= PAGES_PER_WAITQUEUE; 3415 3416 while (size < pages) 3417 size <<= 1; 3418 3419 /* 3420 * Once we have dozens or even hundreds of threads sleeping 3421 * on IO we've got bigger problems than wait queue collision. 3422 * Limit the size of the wait table to a reasonable size. 3423 */ 3424 size = min(size, 4096UL); 3425 3426 return max(size, 4UL); 3427} 3428#else 3429/* 3430 * A zone's size might be changed by hot-add, so it is not possible to determine 3431 * a suitable size for its wait_table. So we use the maximum size now. 3432 * 3433 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3434 * 3435 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3436 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3437 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3438 * 3439 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3440 * or more by the traditional way. (See above). It equals: 3441 * 3442 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3443 * ia64(16K page size) : = ( 8G + 4M)byte. 3444 * powerpc (64K page size) : = (32G +16M)byte. 3445 */ 3446static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3447{ 3448 return 4096UL; 3449} 3450#endif 3451 3452/* 3453 * This is an integer logarithm so that shifts can be used later 3454 * to extract the more random high bits from the multiplicative 3455 * hash function before the remainder is taken. 3456 */ 3457static inline unsigned long wait_table_bits(unsigned long size) 3458{ 3459 return ffz(~size); 3460} 3461 3462#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3463 3464/* 3465 * Check if a pageblock contains reserved pages 3466 */ 3467static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 3468{ 3469 unsigned long pfn; 3470 3471 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3472 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 3473 return 1; 3474 } 3475 return 0; 3476} 3477 3478/* 3479 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3480 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3481 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3482 * higher will lead to a bigger reserve which will get freed as contiguous 3483 * blocks as reclaim kicks in 3484 */ 3485static void setup_zone_migrate_reserve(struct zone *zone) 3486{ 3487 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 3488 struct page *page; 3489 unsigned long block_migratetype; 3490 int reserve; 3491 3492 /* 3493 * Get the start pfn, end pfn and the number of blocks to reserve 3494 * We have to be careful to be aligned to pageblock_nr_pages to 3495 * make sure that we always check pfn_valid for the first page in 3496 * the block. 3497 */ 3498 start_pfn = zone->zone_start_pfn; 3499 end_pfn = start_pfn + zone->spanned_pages; 3500 start_pfn = roundup(start_pfn, pageblock_nr_pages); 3501 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3502 pageblock_order; 3503 3504 /* 3505 * Reserve blocks are generally in place to help high-order atomic 3506 * allocations that are short-lived. A min_free_kbytes value that 3507 * would result in more than 2 reserve blocks for atomic allocations 3508 * is assumed to be in place to help anti-fragmentation for the 3509 * future allocation of hugepages at runtime. 3510 */ 3511 reserve = min(2, reserve); 3512 3513 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3514 if (!pfn_valid(pfn)) 3515 continue; 3516 page = pfn_to_page(pfn); 3517 3518 /* Watch out for overlapping nodes */ 3519 if (page_to_nid(page) != zone_to_nid(zone)) 3520 continue; 3521 3522 block_migratetype = get_pageblock_migratetype(page); 3523 3524 /* Only test what is necessary when the reserves are not met */ 3525 if (reserve > 0) { 3526 /* 3527 * Blocks with reserved pages will never free, skip 3528 * them. 3529 */ 3530 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 3531 if (pageblock_is_reserved(pfn, block_end_pfn)) 3532 continue; 3533 3534 /* If this block is reserved, account for it */ 3535 if (block_migratetype == MIGRATE_RESERVE) { 3536 reserve--; 3537 continue; 3538 } 3539 3540 /* Suitable for reserving if this block is movable */ 3541 if (block_migratetype == MIGRATE_MOVABLE) { 3542 set_pageblock_migratetype(page, 3543 MIGRATE_RESERVE); 3544 move_freepages_block(zone, page, 3545 MIGRATE_RESERVE); 3546 reserve--; 3547 continue; 3548 } 3549 } 3550 3551 /* 3552 * If the reserve is met and this is a previous reserved block, 3553 * take it back 3554 */ 3555 if (block_migratetype == MIGRATE_RESERVE) { 3556 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3557 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3558 } 3559 } 3560} 3561 3562/* 3563 * Initially all pages are reserved - free ones are freed 3564 * up by free_all_bootmem() once the early boot process is 3565 * done. Non-atomic initialization, single-pass. 3566 */ 3567void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3568 unsigned long start_pfn, enum memmap_context context) 3569{ 3570 struct page *page; 3571 unsigned long end_pfn = start_pfn + size; 3572 unsigned long pfn; 3573 struct zone *z; 3574 3575 if (highest_memmap_pfn < end_pfn - 1) 3576 highest_memmap_pfn = end_pfn - 1; 3577 3578 z = &NODE_DATA(nid)->node_zones[zone]; 3579 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3580 /* 3581 * There can be holes in boot-time mem_map[]s 3582 * handed to this function. They do not 3583 * exist on hotplugged memory. 3584 */ 3585 if (context == MEMMAP_EARLY) { 3586 if (!early_pfn_valid(pfn)) 3587 continue; 3588 if (!early_pfn_in_nid(pfn, nid)) 3589 continue; 3590 } 3591 page = pfn_to_page(pfn); 3592 set_page_links(page, zone, nid, pfn); 3593 mminit_verify_page_links(page, zone, nid, pfn); 3594 init_page_count(page); 3595 reset_page_mapcount(page); 3596 SetPageReserved(page); 3597 /* 3598 * Mark the block movable so that blocks are reserved for 3599 * movable at startup. This will force kernel allocations 3600 * to reserve their blocks rather than leaking throughout 3601 * the address space during boot when many long-lived 3602 * kernel allocations are made. Later some blocks near 3603 * the start are marked MIGRATE_RESERVE by 3604 * setup_zone_migrate_reserve() 3605 * 3606 * bitmap is created for zone's valid pfn range. but memmap 3607 * can be created for invalid pages (for alignment) 3608 * check here not to call set_pageblock_migratetype() against 3609 * pfn out of zone. 3610 */ 3611 if ((z->zone_start_pfn <= pfn) 3612 && (pfn < z->zone_start_pfn + z->spanned_pages) 3613 && !(pfn & (pageblock_nr_pages - 1))) 3614 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3615 3616 INIT_LIST_HEAD(&page->lru); 3617#ifdef WANT_PAGE_VIRTUAL 3618 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3619 if (!is_highmem_idx(zone)) 3620 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3621#endif 3622 } 3623} 3624 3625static void __meminit zone_init_free_lists(struct zone *zone) 3626{ 3627 int order, t; 3628 for_each_migratetype_order(order, t) { 3629 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3630 zone->free_area[order].nr_free = 0; 3631 } 3632} 3633 3634#ifndef __HAVE_ARCH_MEMMAP_INIT 3635#define memmap_init(size, nid, zone, start_pfn) \ 3636 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3637#endif 3638 3639static int zone_batchsize(struct zone *zone) 3640{ 3641#ifdef CONFIG_MMU 3642 int batch; 3643 3644 /* 3645 * The per-cpu-pages pools are set to around 1000th of the 3646 * size of the zone. But no more than 1/2 of a meg. 3647 * 3648 * OK, so we don't know how big the cache is. So guess. 3649 */ 3650 batch = zone->present_pages / 1024; 3651 if (batch * PAGE_SIZE > 512 * 1024) 3652 batch = (512 * 1024) / PAGE_SIZE; 3653 batch /= 4; /* We effectively *= 4 below */ 3654 if (batch < 1) 3655 batch = 1; 3656 3657 /* 3658 * Clamp the batch to a 2^n - 1 value. Having a power 3659 * of 2 value was found to be more likely to have 3660 * suboptimal cache aliasing properties in some cases. 3661 * 3662 * For example if 2 tasks are alternately allocating 3663 * batches of pages, one task can end up with a lot 3664 * of pages of one half of the possible page colors 3665 * and the other with pages of the other colors. 3666 */ 3667 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3668 3669 return batch; 3670 3671#else 3672 /* The deferral and batching of frees should be suppressed under NOMMU 3673 * conditions. 3674 * 3675 * The problem is that NOMMU needs to be able to allocate large chunks 3676 * of contiguous memory as there's no hardware page translation to 3677 * assemble apparent contiguous memory from discontiguous pages. 3678 * 3679 * Queueing large contiguous runs of pages for batching, however, 3680 * causes the pages to actually be freed in smaller chunks. As there 3681 * can be a significant delay between the individual batches being 3682 * recycled, this leads to the once large chunks of space being 3683 * fragmented and becoming unavailable for high-order allocations. 3684 */ 3685 return 0; 3686#endif 3687} 3688 3689static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3690{ 3691 struct per_cpu_pages *pcp; 3692 int migratetype; 3693 3694 memset(p, 0, sizeof(*p)); 3695 3696 pcp = &p->pcp; 3697 pcp->count = 0; 3698 pcp->high = 6 * batch; 3699 pcp->batch = max(1UL, 1 * batch); 3700 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3701 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3702} 3703 3704/* 3705 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3706 * to the value high for the pageset p. 3707 */ 3708 3709static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3710 unsigned long high) 3711{ 3712 struct per_cpu_pages *pcp; 3713 3714 pcp = &p->pcp; 3715 pcp->high = high; 3716 pcp->batch = max(1UL, high/4); 3717 if ((high/4) > (PAGE_SHIFT * 8)) 3718 pcp->batch = PAGE_SHIFT * 8; 3719} 3720 3721static void setup_zone_pageset(struct zone *zone) 3722{ 3723 int cpu; 3724 3725 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3726 3727 for_each_possible_cpu(cpu) { 3728 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3729 3730 setup_pageset(pcp, zone_batchsize(zone)); 3731 3732 if (percpu_pagelist_fraction) 3733 setup_pagelist_highmark(pcp, 3734 (zone->present_pages / 3735 percpu_pagelist_fraction)); 3736 } 3737} 3738 3739/* 3740 * Allocate per cpu pagesets and initialize them. 3741 * Before this call only boot pagesets were available. 3742 */ 3743void __init setup_per_cpu_pageset(void) 3744{ 3745 struct zone *zone; 3746 3747 for_each_populated_zone(zone) 3748 setup_zone_pageset(zone); 3749} 3750 3751static noinline __init_refok 3752int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3753{ 3754 int i; 3755 struct pglist_data *pgdat = zone->zone_pgdat; 3756 size_t alloc_size; 3757 3758 /* 3759 * The per-page waitqueue mechanism uses hashed waitqueues 3760 * per zone. 3761 */ 3762 zone->wait_table_hash_nr_entries = 3763 wait_table_hash_nr_entries(zone_size_pages); 3764 zone->wait_table_bits = 3765 wait_table_bits(zone->wait_table_hash_nr_entries); 3766 alloc_size = zone->wait_table_hash_nr_entries 3767 * sizeof(wait_queue_head_t); 3768 3769 if (!slab_is_available()) { 3770 zone->wait_table = (wait_queue_head_t *) 3771 alloc_bootmem_node_nopanic(pgdat, alloc_size); 3772 } else { 3773 /* 3774 * This case means that a zone whose size was 0 gets new memory 3775 * via memory hot-add. 3776 * But it may be the case that a new node was hot-added. In 3777 * this case vmalloc() will not be able to use this new node's 3778 * memory - this wait_table must be initialized to use this new 3779 * node itself as well. 3780 * To use this new node's memory, further consideration will be 3781 * necessary. 3782 */ 3783 zone->wait_table = vmalloc(alloc_size); 3784 } 3785 if (!zone->wait_table) 3786 return -ENOMEM; 3787 3788 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3789 init_waitqueue_head(zone->wait_table + i); 3790 3791 return 0; 3792} 3793 3794static int __zone_pcp_update(void *data) 3795{ 3796 struct zone *zone = data; 3797 int cpu; 3798 unsigned long batch = zone_batchsize(zone), flags; 3799 3800 for_each_possible_cpu(cpu) { 3801 struct per_cpu_pageset *pset; 3802 struct per_cpu_pages *pcp; 3803 3804 pset = per_cpu_ptr(zone->pageset, cpu); 3805 pcp = &pset->pcp; 3806 3807 local_irq_save(flags); 3808 free_pcppages_bulk(zone, pcp->count, pcp); 3809 setup_pageset(pset, batch); 3810 local_irq_restore(flags); 3811 } 3812 return 0; 3813} 3814 3815void zone_pcp_update(struct zone *zone) 3816{ 3817 stop_machine(__zone_pcp_update, zone, NULL); 3818} 3819 3820static __meminit void zone_pcp_init(struct zone *zone) 3821{ 3822 /* 3823 * per cpu subsystem is not up at this point. The following code 3824 * relies on the ability of the linker to provide the 3825 * offset of a (static) per cpu variable into the per cpu area. 3826 */ 3827 zone->pageset = &boot_pageset; 3828 3829 if (zone->present_pages) 3830 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3831 zone->name, zone->present_pages, 3832 zone_batchsize(zone)); 3833} 3834 3835__meminit int init_currently_empty_zone(struct zone *zone, 3836 unsigned long zone_start_pfn, 3837 unsigned long size, 3838 enum memmap_context context) 3839{ 3840 struct pglist_data *pgdat = zone->zone_pgdat; 3841 int ret; 3842 ret = zone_wait_table_init(zone, size); 3843 if (ret) 3844 return ret; 3845 pgdat->nr_zones = zone_idx(zone) + 1; 3846 3847 zone->zone_start_pfn = zone_start_pfn; 3848 3849 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3850 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3851 pgdat->node_id, 3852 (unsigned long)zone_idx(zone), 3853 zone_start_pfn, (zone_start_pfn + size)); 3854 3855 zone_init_free_lists(zone); 3856 3857 return 0; 3858} 3859 3860#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 3861#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3862/* 3863 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3864 * Architectures may implement their own version but if add_active_range() 3865 * was used and there are no special requirements, this is a convenient 3866 * alternative 3867 */ 3868int __meminit __early_pfn_to_nid(unsigned long pfn) 3869{ 3870 unsigned long start_pfn, end_pfn; 3871 int i, nid; 3872 3873 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 3874 if (start_pfn <= pfn && pfn < end_pfn) 3875 return nid; 3876 /* This is a memory hole */ 3877 return -1; 3878} 3879#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3880 3881int __meminit early_pfn_to_nid(unsigned long pfn) 3882{ 3883 int nid; 3884 3885 nid = __early_pfn_to_nid(pfn); 3886 if (nid >= 0) 3887 return nid; 3888 /* just returns 0 */ 3889 return 0; 3890} 3891 3892#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3893bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3894{ 3895 int nid; 3896 3897 nid = __early_pfn_to_nid(pfn); 3898 if (nid >= 0 && nid != node) 3899 return false; 3900 return true; 3901} 3902#endif 3903 3904/** 3905 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3906 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3907 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3908 * 3909 * If an architecture guarantees that all ranges registered with 3910 * add_active_ranges() contain no holes and may be freed, this 3911 * this function may be used instead of calling free_bootmem() manually. 3912 */ 3913void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 3914{ 3915 unsigned long start_pfn, end_pfn; 3916 int i, this_nid; 3917 3918 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 3919 start_pfn = min(start_pfn, max_low_pfn); 3920 end_pfn = min(end_pfn, max_low_pfn); 3921 3922 if (start_pfn < end_pfn) 3923 free_bootmem_node(NODE_DATA(this_nid), 3924 PFN_PHYS(start_pfn), 3925 (end_pfn - start_pfn) << PAGE_SHIFT); 3926 } 3927} 3928 3929int __init add_from_early_node_map(struct range *range, int az, 3930 int nr_range, int nid) 3931{ 3932 unsigned long start_pfn, end_pfn; 3933 int i; 3934 3935 /* need to go over early_node_map to find out good range for node */ 3936 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) 3937 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn); 3938 return nr_range; 3939} 3940 3941/** 3942 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3943 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3944 * 3945 * If an architecture guarantees that all ranges registered with 3946 * add_active_ranges() contain no holes and may be freed, this 3947 * function may be used instead of calling memory_present() manually. 3948 */ 3949void __init sparse_memory_present_with_active_regions(int nid) 3950{ 3951 unsigned long start_pfn, end_pfn; 3952 int i, this_nid; 3953 3954 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 3955 memory_present(this_nid, start_pfn, end_pfn); 3956} 3957 3958/** 3959 * get_pfn_range_for_nid - Return the start and end page frames for a node 3960 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3961 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3962 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3963 * 3964 * It returns the start and end page frame of a node based on information 3965 * provided by an arch calling add_active_range(). If called for a node 3966 * with no available memory, a warning is printed and the start and end 3967 * PFNs will be 0. 3968 */ 3969void __meminit get_pfn_range_for_nid(unsigned int nid, 3970 unsigned long *start_pfn, unsigned long *end_pfn) 3971{ 3972 unsigned long this_start_pfn, this_end_pfn; 3973 int i; 3974 3975 *start_pfn = -1UL; 3976 *end_pfn = 0; 3977 3978 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 3979 *start_pfn = min(*start_pfn, this_start_pfn); 3980 *end_pfn = max(*end_pfn, this_end_pfn); 3981 } 3982 3983 if (*start_pfn == -1UL) 3984 *start_pfn = 0; 3985} 3986 3987/* 3988 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3989 * assumption is made that zones within a node are ordered in monotonic 3990 * increasing memory addresses so that the "highest" populated zone is used 3991 */ 3992static void __init find_usable_zone_for_movable(void) 3993{ 3994 int zone_index; 3995 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3996 if (zone_index == ZONE_MOVABLE) 3997 continue; 3998 3999 if (arch_zone_highest_possible_pfn[zone_index] > 4000 arch_zone_lowest_possible_pfn[zone_index]) 4001 break; 4002 } 4003 4004 VM_BUG_ON(zone_index == -1); 4005 movable_zone = zone_index; 4006} 4007 4008/* 4009 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 4010 * because it is sized independent of architecture. Unlike the other zones, 4011 * the starting point for ZONE_MOVABLE is not fixed. It may be different 4012 * in each node depending on the size of each node and how evenly kernelcore 4013 * is distributed. This helper function adjusts the zone ranges 4014 * provided by the architecture for a given node by using the end of the 4015 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 4016 * zones within a node are in order of monotonic increases memory addresses 4017 */ 4018static void __meminit adjust_zone_range_for_zone_movable(int nid, 4019 unsigned long zone_type, 4020 unsigned long node_start_pfn, 4021 unsigned long node_end_pfn, 4022 unsigned long *zone_start_pfn, 4023 unsigned long *zone_end_pfn) 4024{ 4025 /* Only adjust if ZONE_MOVABLE is on this node */ 4026 if (zone_movable_pfn[nid]) { 4027 /* Size ZONE_MOVABLE */ 4028 if (zone_type == ZONE_MOVABLE) { 4029 *zone_start_pfn = zone_movable_pfn[nid]; 4030 *zone_end_pfn = min(node_end_pfn, 4031 arch_zone_highest_possible_pfn[movable_zone]); 4032 4033 /* Adjust for ZONE_MOVABLE starting within this range */ 4034 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 4035 *zone_end_pfn > zone_movable_pfn[nid]) { 4036 *zone_end_pfn = zone_movable_pfn[nid]; 4037 4038 /* Check if this whole range is within ZONE_MOVABLE */ 4039 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 4040 *zone_start_pfn = *zone_end_pfn; 4041 } 4042} 4043 4044/* 4045 * Return the number of pages a zone spans in a node, including holes 4046 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 4047 */ 4048static unsigned long __meminit zone_spanned_pages_in_node(int nid, 4049 unsigned long zone_type, 4050 unsigned long *ignored) 4051{ 4052 unsigned long node_start_pfn, node_end_pfn; 4053 unsigned long zone_start_pfn, zone_end_pfn; 4054 4055 /* Get the start and end of the node and zone */ 4056 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4057 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 4058 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 4059 adjust_zone_range_for_zone_movable(nid, zone_type, 4060 node_start_pfn, node_end_pfn, 4061 &zone_start_pfn, &zone_end_pfn); 4062 4063 /* Check that this node has pages within the zone's required range */ 4064 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 4065 return 0; 4066 4067 /* Move the zone boundaries inside the node if necessary */ 4068 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 4069 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 4070 4071 /* Return the spanned pages */ 4072 return zone_end_pfn - zone_start_pfn; 4073} 4074 4075/* 4076 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 4077 * then all holes in the requested range will be accounted for. 4078 */ 4079unsigned long __meminit __absent_pages_in_range(int nid, 4080 unsigned long range_start_pfn, 4081 unsigned long range_end_pfn) 4082{ 4083 unsigned long nr_absent = range_end_pfn - range_start_pfn; 4084 unsigned long start_pfn, end_pfn; 4085 int i; 4086 4087 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 4088 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 4089 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 4090 nr_absent -= end_pfn - start_pfn; 4091 } 4092 return nr_absent; 4093} 4094 4095/** 4096 * absent_pages_in_range - Return number of page frames in holes within a range 4097 * @start_pfn: The start PFN to start searching for holes 4098 * @end_pfn: The end PFN to stop searching for holes 4099 * 4100 * It returns the number of pages frames in memory holes within a range. 4101 */ 4102unsigned long __init absent_pages_in_range(unsigned long start_pfn, 4103 unsigned long end_pfn) 4104{ 4105 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 4106} 4107 4108/* Return the number of page frames in holes in a zone on a node */ 4109static unsigned long __meminit zone_absent_pages_in_node(int nid, 4110 unsigned long zone_type, 4111 unsigned long *ignored) 4112{ 4113 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 4114 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 4115 unsigned long node_start_pfn, node_end_pfn; 4116 unsigned long zone_start_pfn, zone_end_pfn; 4117 4118 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4119 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 4120 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 4121 4122 adjust_zone_range_for_zone_movable(nid, zone_type, 4123 node_start_pfn, node_end_pfn, 4124 &zone_start_pfn, &zone_end_pfn); 4125 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 4126} 4127 4128#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4129static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 4130 unsigned long zone_type, 4131 unsigned long *zones_size) 4132{ 4133 return zones_size[zone_type]; 4134} 4135 4136static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4137 unsigned long zone_type, 4138 unsigned long *zholes_size) 4139{ 4140 if (!zholes_size) 4141 return 0; 4142 4143 return zholes_size[zone_type]; 4144} 4145 4146#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4147 4148static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4149 unsigned long *zones_size, unsigned long *zholes_size) 4150{ 4151 unsigned long realtotalpages, totalpages = 0; 4152 enum zone_type i; 4153 4154 for (i = 0; i < MAX_NR_ZONES; i++) 4155 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4156 zones_size); 4157 pgdat->node_spanned_pages = totalpages; 4158 4159 realtotalpages = totalpages; 4160 for (i = 0; i < MAX_NR_ZONES; i++) 4161 realtotalpages -= 4162 zone_absent_pages_in_node(pgdat->node_id, i, 4163 zholes_size); 4164 pgdat->node_present_pages = realtotalpages; 4165 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4166 realtotalpages); 4167} 4168 4169#ifndef CONFIG_SPARSEMEM 4170/* 4171 * Calculate the size of the zone->blockflags rounded to an unsigned long 4172 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4173 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4174 * round what is now in bits to nearest long in bits, then return it in 4175 * bytes. 4176 */ 4177static unsigned long __init usemap_size(unsigned long zonesize) 4178{ 4179 unsigned long usemapsize; 4180 4181 usemapsize = roundup(zonesize, pageblock_nr_pages); 4182 usemapsize = usemapsize >> pageblock_order; 4183 usemapsize *= NR_PAGEBLOCK_BITS; 4184 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4185 4186 return usemapsize / 8; 4187} 4188 4189static void __init setup_usemap(struct pglist_data *pgdat, 4190 struct zone *zone, unsigned long zonesize) 4191{ 4192 unsigned long usemapsize = usemap_size(zonesize); 4193 zone->pageblock_flags = NULL; 4194 if (usemapsize) 4195 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, 4196 usemapsize); 4197} 4198#else 4199static inline void setup_usemap(struct pglist_data *pgdat, 4200 struct zone *zone, unsigned long zonesize) {} 4201#endif /* CONFIG_SPARSEMEM */ 4202 4203#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4204 4205/* Return a sensible default order for the pageblock size. */ 4206static inline int pageblock_default_order(void) 4207{ 4208 if (HPAGE_SHIFT > PAGE_SHIFT) 4209 return HUGETLB_PAGE_ORDER; 4210 4211 return MAX_ORDER-1; 4212} 4213 4214/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4215static inline void __init set_pageblock_order(unsigned int order) 4216{ 4217 /* Check that pageblock_nr_pages has not already been setup */ 4218 if (pageblock_order) 4219 return; 4220 4221 /* 4222 * Assume the largest contiguous order of interest is a huge page. 4223 * This value may be variable depending on boot parameters on IA64 4224 */ 4225 pageblock_order = order; 4226} 4227#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4228 4229/* 4230 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4231 * and pageblock_default_order() are unused as pageblock_order is set 4232 * at compile-time. See include/linux/pageblock-flags.h for the values of 4233 * pageblock_order based on the kernel config 4234 */ 4235static inline int pageblock_default_order(unsigned int order) 4236{ 4237 return MAX_ORDER-1; 4238} 4239#define set_pageblock_order(x) do {} while (0) 4240 4241#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4242 4243/* 4244 * Set up the zone data structures: 4245 * - mark all pages reserved 4246 * - mark all memory queues empty 4247 * - clear the memory bitmaps 4248 */ 4249static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4250 unsigned long *zones_size, unsigned long *zholes_size) 4251{ 4252 enum zone_type j; 4253 int nid = pgdat->node_id; 4254 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4255 int ret; 4256 4257 pgdat_resize_init(pgdat); 4258 pgdat->nr_zones = 0; 4259 init_waitqueue_head(&pgdat->kswapd_wait); 4260 pgdat->kswapd_max_order = 0; 4261 pgdat_page_cgroup_init(pgdat); 4262 4263 for (j = 0; j < MAX_NR_ZONES; j++) { 4264 struct zone *zone = pgdat->node_zones + j; 4265 unsigned long size, realsize, memmap_pages; 4266 enum lru_list lru; 4267 4268 size = zone_spanned_pages_in_node(nid, j, zones_size); 4269 realsize = size - zone_absent_pages_in_node(nid, j, 4270 zholes_size); 4271 4272 /* 4273 * Adjust realsize so that it accounts for how much memory 4274 * is used by this zone for memmap. This affects the watermark 4275 * and per-cpu initialisations 4276 */ 4277 memmap_pages = 4278 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4279 if (realsize >= memmap_pages) { 4280 realsize -= memmap_pages; 4281 if (memmap_pages) 4282 printk(KERN_DEBUG 4283 " %s zone: %lu pages used for memmap\n", 4284 zone_names[j], memmap_pages); 4285 } else 4286 printk(KERN_WARNING 4287 " %s zone: %lu pages exceeds realsize %lu\n", 4288 zone_names[j], memmap_pages, realsize); 4289 4290 /* Account for reserved pages */ 4291 if (j == 0 && realsize > dma_reserve) { 4292 realsize -= dma_reserve; 4293 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4294 zone_names[0], dma_reserve); 4295 } 4296 4297 if (!is_highmem_idx(j)) 4298 nr_kernel_pages += realsize; 4299 nr_all_pages += realsize; 4300 4301 zone->spanned_pages = size; 4302 zone->present_pages = realsize; 4303#ifdef CONFIG_NUMA 4304 zone->node = nid; 4305 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4306 / 100; 4307 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4308#endif 4309 zone->name = zone_names[j]; 4310 spin_lock_init(&zone->lock); 4311 spin_lock_init(&zone->lru_lock); 4312 zone_seqlock_init(zone); 4313 zone->zone_pgdat = pgdat; 4314 4315 zone_pcp_init(zone); 4316 for_each_lru(lru) 4317 INIT_LIST_HEAD(&zone->lruvec.lists[lru]); 4318 zone->reclaim_stat.recent_rotated[0] = 0; 4319 zone->reclaim_stat.recent_rotated[1] = 0; 4320 zone->reclaim_stat.recent_scanned[0] = 0; 4321 zone->reclaim_stat.recent_scanned[1] = 0; 4322 zap_zone_vm_stats(zone); 4323 zone->flags = 0; 4324 if (!size) 4325 continue; 4326 4327 set_pageblock_order(pageblock_default_order()); 4328 setup_usemap(pgdat, zone, size); 4329 ret = init_currently_empty_zone(zone, zone_start_pfn, 4330 size, MEMMAP_EARLY); 4331 BUG_ON(ret); 4332 memmap_init(size, nid, j, zone_start_pfn); 4333 zone_start_pfn += size; 4334 } 4335} 4336 4337static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4338{ 4339 /* Skip empty nodes */ 4340 if (!pgdat->node_spanned_pages) 4341 return; 4342 4343#ifdef CONFIG_FLAT_NODE_MEM_MAP 4344 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4345 if (!pgdat->node_mem_map) { 4346 unsigned long size, start, end; 4347 struct page *map; 4348 4349 /* 4350 * The zone's endpoints aren't required to be MAX_ORDER 4351 * aligned but the node_mem_map endpoints must be in order 4352 * for the buddy allocator to function correctly. 4353 */ 4354 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4355 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4356 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4357 size = (end - start) * sizeof(struct page); 4358 map = alloc_remap(pgdat->node_id, size); 4359 if (!map) 4360 map = alloc_bootmem_node_nopanic(pgdat, size); 4361 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4362 } 4363#ifndef CONFIG_NEED_MULTIPLE_NODES 4364 /* 4365 * With no DISCONTIG, the global mem_map is just set as node 0's 4366 */ 4367 if (pgdat == NODE_DATA(0)) { 4368 mem_map = NODE_DATA(0)->node_mem_map; 4369#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4370 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4371 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4372#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4373 } 4374#endif 4375#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4376} 4377 4378void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4379 unsigned long node_start_pfn, unsigned long *zholes_size) 4380{ 4381 pg_data_t *pgdat = NODE_DATA(nid); 4382 4383 pgdat->node_id = nid; 4384 pgdat->node_start_pfn = node_start_pfn; 4385 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4386 4387 alloc_node_mem_map(pgdat); 4388#ifdef CONFIG_FLAT_NODE_MEM_MAP 4389 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4390 nid, (unsigned long)pgdat, 4391 (unsigned long)pgdat->node_mem_map); 4392#endif 4393 4394 free_area_init_core(pgdat, zones_size, zholes_size); 4395} 4396 4397#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4398 4399#if MAX_NUMNODES > 1 4400/* 4401 * Figure out the number of possible node ids. 4402 */ 4403static void __init setup_nr_node_ids(void) 4404{ 4405 unsigned int node; 4406 unsigned int highest = 0; 4407 4408 for_each_node_mask(node, node_possible_map) 4409 highest = node; 4410 nr_node_ids = highest + 1; 4411} 4412#else 4413static inline void setup_nr_node_ids(void) 4414{ 4415} 4416#endif 4417 4418/** 4419 * node_map_pfn_alignment - determine the maximum internode alignment 4420 * 4421 * This function should be called after node map is populated and sorted. 4422 * It calculates the maximum power of two alignment which can distinguish 4423 * all the nodes. 4424 * 4425 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 4426 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 4427 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 4428 * shifted, 1GiB is enough and this function will indicate so. 4429 * 4430 * This is used to test whether pfn -> nid mapping of the chosen memory 4431 * model has fine enough granularity to avoid incorrect mapping for the 4432 * populated node map. 4433 * 4434 * Returns the determined alignment in pfn's. 0 if there is no alignment 4435 * requirement (single node). 4436 */ 4437unsigned long __init node_map_pfn_alignment(void) 4438{ 4439 unsigned long accl_mask = 0, last_end = 0; 4440 unsigned long start, end, mask; 4441 int last_nid = -1; 4442 int i, nid; 4443 4444 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 4445 if (!start || last_nid < 0 || last_nid == nid) { 4446 last_nid = nid; 4447 last_end = end; 4448 continue; 4449 } 4450 4451 /* 4452 * Start with a mask granular enough to pin-point to the 4453 * start pfn and tick off bits one-by-one until it becomes 4454 * too coarse to separate the current node from the last. 4455 */ 4456 mask = ~((1 << __ffs(start)) - 1); 4457 while (mask && last_end <= (start & (mask << 1))) 4458 mask <<= 1; 4459 4460 /* accumulate all internode masks */ 4461 accl_mask |= mask; 4462 } 4463 4464 /* convert mask to number of pages */ 4465 return ~accl_mask + 1; 4466} 4467 4468/* Find the lowest pfn for a node */ 4469static unsigned long __init find_min_pfn_for_node(int nid) 4470{ 4471 unsigned long min_pfn = ULONG_MAX; 4472 unsigned long start_pfn; 4473 int i; 4474 4475 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 4476 min_pfn = min(min_pfn, start_pfn); 4477 4478 if (min_pfn == ULONG_MAX) { 4479 printk(KERN_WARNING 4480 "Could not find start_pfn for node %d\n", nid); 4481 return 0; 4482 } 4483 4484 return min_pfn; 4485} 4486 4487/** 4488 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4489 * 4490 * It returns the minimum PFN based on information provided via 4491 * add_active_range(). 4492 */ 4493unsigned long __init find_min_pfn_with_active_regions(void) 4494{ 4495 return find_min_pfn_for_node(MAX_NUMNODES); 4496} 4497 4498/* 4499 * early_calculate_totalpages() 4500 * Sum pages in active regions for movable zone. 4501 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4502 */ 4503static unsigned long __init early_calculate_totalpages(void) 4504{ 4505 unsigned long totalpages = 0; 4506 unsigned long start_pfn, end_pfn; 4507 int i, nid; 4508 4509 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 4510 unsigned long pages = end_pfn - start_pfn; 4511 4512 totalpages += pages; 4513 if (pages) 4514 node_set_state(nid, N_HIGH_MEMORY); 4515 } 4516 return totalpages; 4517} 4518 4519/* 4520 * Find the PFN the Movable zone begins in each node. Kernel memory 4521 * is spread evenly between nodes as long as the nodes have enough 4522 * memory. When they don't, some nodes will have more kernelcore than 4523 * others 4524 */ 4525static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4526{ 4527 int i, nid; 4528 unsigned long usable_startpfn; 4529 unsigned long kernelcore_node, kernelcore_remaining; 4530 /* save the state before borrow the nodemask */ 4531 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4532 unsigned long totalpages = early_calculate_totalpages(); 4533 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4534 4535 /* 4536 * If movablecore was specified, calculate what size of 4537 * kernelcore that corresponds so that memory usable for 4538 * any allocation type is evenly spread. If both kernelcore 4539 * and movablecore are specified, then the value of kernelcore 4540 * will be used for required_kernelcore if it's greater than 4541 * what movablecore would have allowed. 4542 */ 4543 if (required_movablecore) { 4544 unsigned long corepages; 4545 4546 /* 4547 * Round-up so that ZONE_MOVABLE is at least as large as what 4548 * was requested by the user 4549 */ 4550 required_movablecore = 4551 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4552 corepages = totalpages - required_movablecore; 4553 4554 required_kernelcore = max(required_kernelcore, corepages); 4555 } 4556 4557 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4558 if (!required_kernelcore) 4559 goto out; 4560 4561 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4562 find_usable_zone_for_movable(); 4563 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4564 4565restart: 4566 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4567 kernelcore_node = required_kernelcore / usable_nodes; 4568 for_each_node_state(nid, N_HIGH_MEMORY) { 4569 unsigned long start_pfn, end_pfn; 4570 4571 /* 4572 * Recalculate kernelcore_node if the division per node 4573 * now exceeds what is necessary to satisfy the requested 4574 * amount of memory for the kernel 4575 */ 4576 if (required_kernelcore < kernelcore_node) 4577 kernelcore_node = required_kernelcore / usable_nodes; 4578 4579 /* 4580 * As the map is walked, we track how much memory is usable 4581 * by the kernel using kernelcore_remaining. When it is 4582 * 0, the rest of the node is usable by ZONE_MOVABLE 4583 */ 4584 kernelcore_remaining = kernelcore_node; 4585 4586 /* Go through each range of PFNs within this node */ 4587 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 4588 unsigned long size_pages; 4589 4590 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 4591 if (start_pfn >= end_pfn) 4592 continue; 4593 4594 /* Account for what is only usable for kernelcore */ 4595 if (start_pfn < usable_startpfn) { 4596 unsigned long kernel_pages; 4597 kernel_pages = min(end_pfn, usable_startpfn) 4598 - start_pfn; 4599 4600 kernelcore_remaining -= min(kernel_pages, 4601 kernelcore_remaining); 4602 required_kernelcore -= min(kernel_pages, 4603 required_kernelcore); 4604 4605 /* Continue if range is now fully accounted */ 4606 if (end_pfn <= usable_startpfn) { 4607 4608 /* 4609 * Push zone_movable_pfn to the end so 4610 * that if we have to rebalance 4611 * kernelcore across nodes, we will 4612 * not double account here 4613 */ 4614 zone_movable_pfn[nid] = end_pfn; 4615 continue; 4616 } 4617 start_pfn = usable_startpfn; 4618 } 4619 4620 /* 4621 * The usable PFN range for ZONE_MOVABLE is from 4622 * start_pfn->end_pfn. Calculate size_pages as the 4623 * number of pages used as kernelcore 4624 */ 4625 size_pages = end_pfn - start_pfn; 4626 if (size_pages > kernelcore_remaining) 4627 size_pages = kernelcore_remaining; 4628 zone_movable_pfn[nid] = start_pfn + size_pages; 4629 4630 /* 4631 * Some kernelcore has been met, update counts and 4632 * break if the kernelcore for this node has been 4633 * satisified 4634 */ 4635 required_kernelcore -= min(required_kernelcore, 4636 size_pages); 4637 kernelcore_remaining -= size_pages; 4638 if (!kernelcore_remaining) 4639 break; 4640 } 4641 } 4642 4643 /* 4644 * If there is still required_kernelcore, we do another pass with one 4645 * less node in the count. This will push zone_movable_pfn[nid] further 4646 * along on the nodes that still have memory until kernelcore is 4647 * satisified 4648 */ 4649 usable_nodes--; 4650 if (usable_nodes && required_kernelcore > usable_nodes) 4651 goto restart; 4652 4653 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4654 for (nid = 0; nid < MAX_NUMNODES; nid++) 4655 zone_movable_pfn[nid] = 4656 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4657 4658out: 4659 /* restore the node_state */ 4660 node_states[N_HIGH_MEMORY] = saved_node_state; 4661} 4662 4663/* Any regular memory on that node ? */ 4664static void check_for_regular_memory(pg_data_t *pgdat) 4665{ 4666#ifdef CONFIG_HIGHMEM 4667 enum zone_type zone_type; 4668 4669 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4670 struct zone *zone = &pgdat->node_zones[zone_type]; 4671 if (zone->present_pages) { 4672 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4673 break; 4674 } 4675 } 4676#endif 4677} 4678 4679/** 4680 * free_area_init_nodes - Initialise all pg_data_t and zone data 4681 * @max_zone_pfn: an array of max PFNs for each zone 4682 * 4683 * This will call free_area_init_node() for each active node in the system. 4684 * Using the page ranges provided by add_active_range(), the size of each 4685 * zone in each node and their holes is calculated. If the maximum PFN 4686 * between two adjacent zones match, it is assumed that the zone is empty. 4687 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4688 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4689 * starts where the previous one ended. For example, ZONE_DMA32 starts 4690 * at arch_max_dma_pfn. 4691 */ 4692void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4693{ 4694 unsigned long start_pfn, end_pfn; 4695 int i, nid; 4696 4697 /* Record where the zone boundaries are */ 4698 memset(arch_zone_lowest_possible_pfn, 0, 4699 sizeof(arch_zone_lowest_possible_pfn)); 4700 memset(arch_zone_highest_possible_pfn, 0, 4701 sizeof(arch_zone_highest_possible_pfn)); 4702 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4703 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4704 for (i = 1; i < MAX_NR_ZONES; i++) { 4705 if (i == ZONE_MOVABLE) 4706 continue; 4707 arch_zone_lowest_possible_pfn[i] = 4708 arch_zone_highest_possible_pfn[i-1]; 4709 arch_zone_highest_possible_pfn[i] = 4710 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4711 } 4712 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4713 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4714 4715 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4716 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4717 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4718 4719 /* Print out the zone ranges */ 4720 printk("Zone PFN ranges:\n"); 4721 for (i = 0; i < MAX_NR_ZONES; i++) { 4722 if (i == ZONE_MOVABLE) 4723 continue; 4724 printk(" %-8s ", zone_names[i]); 4725 if (arch_zone_lowest_possible_pfn[i] == 4726 arch_zone_highest_possible_pfn[i]) 4727 printk("empty\n"); 4728 else 4729 printk("%0#10lx -> %0#10lx\n", 4730 arch_zone_lowest_possible_pfn[i], 4731 arch_zone_highest_possible_pfn[i]); 4732 } 4733 4734 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4735 printk("Movable zone start PFN for each node\n"); 4736 for (i = 0; i < MAX_NUMNODES; i++) { 4737 if (zone_movable_pfn[i]) 4738 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4739 } 4740 4741 /* Print out the early_node_map[] */ 4742 printk("Early memory PFN ranges\n"); 4743 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 4744 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn); 4745 4746 /* Initialise every node */ 4747 mminit_verify_pageflags_layout(); 4748 setup_nr_node_ids(); 4749 for_each_online_node(nid) { 4750 pg_data_t *pgdat = NODE_DATA(nid); 4751 free_area_init_node(nid, NULL, 4752 find_min_pfn_for_node(nid), NULL); 4753 4754 /* Any memory on that node */ 4755 if (pgdat->node_present_pages) 4756 node_set_state(nid, N_HIGH_MEMORY); 4757 check_for_regular_memory(pgdat); 4758 } 4759} 4760 4761static int __init cmdline_parse_core(char *p, unsigned long *core) 4762{ 4763 unsigned long long coremem; 4764 if (!p) 4765 return -EINVAL; 4766 4767 coremem = memparse(p, &p); 4768 *core = coremem >> PAGE_SHIFT; 4769 4770 /* Paranoid check that UL is enough for the coremem value */ 4771 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4772 4773 return 0; 4774} 4775 4776/* 4777 * kernelcore=size sets the amount of memory for use for allocations that 4778 * cannot be reclaimed or migrated. 4779 */ 4780static int __init cmdline_parse_kernelcore(char *p) 4781{ 4782 return cmdline_parse_core(p, &required_kernelcore); 4783} 4784 4785/* 4786 * movablecore=size sets the amount of memory for use for allocations that 4787 * can be reclaimed or migrated. 4788 */ 4789static int __init cmdline_parse_movablecore(char *p) 4790{ 4791 return cmdline_parse_core(p, &required_movablecore); 4792} 4793 4794early_param("kernelcore", cmdline_parse_kernelcore); 4795early_param("movablecore", cmdline_parse_movablecore); 4796 4797#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4798 4799/** 4800 * set_dma_reserve - set the specified number of pages reserved in the first zone 4801 * @new_dma_reserve: The number of pages to mark reserved 4802 * 4803 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4804 * In the DMA zone, a significant percentage may be consumed by kernel image 4805 * and other unfreeable allocations which can skew the watermarks badly. This 4806 * function may optionally be used to account for unfreeable pages in the 4807 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4808 * smaller per-cpu batchsize. 4809 */ 4810void __init set_dma_reserve(unsigned long new_dma_reserve) 4811{ 4812 dma_reserve = new_dma_reserve; 4813} 4814 4815void __init free_area_init(unsigned long *zones_size) 4816{ 4817 free_area_init_node(0, zones_size, 4818 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4819} 4820 4821static int page_alloc_cpu_notify(struct notifier_block *self, 4822 unsigned long action, void *hcpu) 4823{ 4824 int cpu = (unsigned long)hcpu; 4825 4826 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4827 drain_pages(cpu); 4828 4829 /* 4830 * Spill the event counters of the dead processor 4831 * into the current processors event counters. 4832 * This artificially elevates the count of the current 4833 * processor. 4834 */ 4835 vm_events_fold_cpu(cpu); 4836 4837 /* 4838 * Zero the differential counters of the dead processor 4839 * so that the vm statistics are consistent. 4840 * 4841 * This is only okay since the processor is dead and cannot 4842 * race with what we are doing. 4843 */ 4844 refresh_cpu_vm_stats(cpu); 4845 } 4846 return NOTIFY_OK; 4847} 4848 4849void __init page_alloc_init(void) 4850{ 4851 hotcpu_notifier(page_alloc_cpu_notify, 0); 4852} 4853 4854/* 4855 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4856 * or min_free_kbytes changes. 4857 */ 4858static void calculate_totalreserve_pages(void) 4859{ 4860 struct pglist_data *pgdat; 4861 unsigned long reserve_pages = 0; 4862 enum zone_type i, j; 4863 4864 for_each_online_pgdat(pgdat) { 4865 for (i = 0; i < MAX_NR_ZONES; i++) { 4866 struct zone *zone = pgdat->node_zones + i; 4867 unsigned long max = 0; 4868 4869 /* Find valid and maximum lowmem_reserve in the zone */ 4870 for (j = i; j < MAX_NR_ZONES; j++) { 4871 if (zone->lowmem_reserve[j] > max) 4872 max = zone->lowmem_reserve[j]; 4873 } 4874 4875 /* we treat the high watermark as reserved pages. */ 4876 max += high_wmark_pages(zone); 4877 4878 if (max > zone->present_pages) 4879 max = zone->present_pages; 4880 reserve_pages += max; 4881 /* 4882 * Lowmem reserves are not available to 4883 * GFP_HIGHUSER page cache allocations and 4884 * kswapd tries to balance zones to their high 4885 * watermark. As a result, neither should be 4886 * regarded as dirtyable memory, to prevent a 4887 * situation where reclaim has to clean pages 4888 * in order to balance the zones. 4889 */ 4890 zone->dirty_balance_reserve = max; 4891 } 4892 } 4893 dirty_balance_reserve = reserve_pages; 4894 totalreserve_pages = reserve_pages; 4895} 4896 4897/* 4898 * setup_per_zone_lowmem_reserve - called whenever 4899 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4900 * has a correct pages reserved value, so an adequate number of 4901 * pages are left in the zone after a successful __alloc_pages(). 4902 */ 4903static void setup_per_zone_lowmem_reserve(void) 4904{ 4905 struct pglist_data *pgdat; 4906 enum zone_type j, idx; 4907 4908 for_each_online_pgdat(pgdat) { 4909 for (j = 0; j < MAX_NR_ZONES; j++) { 4910 struct zone *zone = pgdat->node_zones + j; 4911 unsigned long present_pages = zone->present_pages; 4912 4913 zone->lowmem_reserve[j] = 0; 4914 4915 idx = j; 4916 while (idx) { 4917 struct zone *lower_zone; 4918 4919 idx--; 4920 4921 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4922 sysctl_lowmem_reserve_ratio[idx] = 1; 4923 4924 lower_zone = pgdat->node_zones + idx; 4925 lower_zone->lowmem_reserve[j] = present_pages / 4926 sysctl_lowmem_reserve_ratio[idx]; 4927 present_pages += lower_zone->present_pages; 4928 } 4929 } 4930 } 4931 4932 /* update totalreserve_pages */ 4933 calculate_totalreserve_pages(); 4934} 4935 4936/** 4937 * setup_per_zone_wmarks - called when min_free_kbytes changes 4938 * or when memory is hot-{added|removed} 4939 * 4940 * Ensures that the watermark[min,low,high] values for each zone are set 4941 * correctly with respect to min_free_kbytes. 4942 */ 4943void setup_per_zone_wmarks(void) 4944{ 4945 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4946 unsigned long lowmem_pages = 0; 4947 struct zone *zone; 4948 unsigned long flags; 4949 4950 /* Calculate total number of !ZONE_HIGHMEM pages */ 4951 for_each_zone(zone) { 4952 if (!is_highmem(zone)) 4953 lowmem_pages += zone->present_pages; 4954 } 4955 4956 for_each_zone(zone) { 4957 u64 tmp; 4958 4959 spin_lock_irqsave(&zone->lock, flags); 4960 tmp = (u64)pages_min * zone->present_pages; 4961 do_div(tmp, lowmem_pages); 4962 if (is_highmem(zone)) { 4963 /* 4964 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4965 * need highmem pages, so cap pages_min to a small 4966 * value here. 4967 * 4968 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4969 * deltas controls asynch page reclaim, and so should 4970 * not be capped for highmem. 4971 */ 4972 int min_pages; 4973 4974 min_pages = zone->present_pages / 1024; 4975 if (min_pages < SWAP_CLUSTER_MAX) 4976 min_pages = SWAP_CLUSTER_MAX; 4977 if (min_pages > 128) 4978 min_pages = 128; 4979 zone->watermark[WMARK_MIN] = min_pages; 4980 } else { 4981 /* 4982 * If it's a lowmem zone, reserve a number of pages 4983 * proportionate to the zone's size. 4984 */ 4985 zone->watermark[WMARK_MIN] = tmp; 4986 } 4987 4988 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4989 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4990 setup_zone_migrate_reserve(zone); 4991 spin_unlock_irqrestore(&zone->lock, flags); 4992 } 4993 4994 /* update totalreserve_pages */ 4995 calculate_totalreserve_pages(); 4996} 4997 4998/* 4999 * The inactive anon list should be small enough that the VM never has to 5000 * do too much work, but large enough that each inactive page has a chance 5001 * to be referenced again before it is swapped out. 5002 * 5003 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 5004 * INACTIVE_ANON pages on this zone's LRU, maintained by the 5005 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 5006 * the anonymous pages are kept on the inactive list. 5007 * 5008 * total target max 5009 * memory ratio inactive anon 5010 * ------------------------------------- 5011 * 10MB 1 5MB 5012 * 100MB 1 50MB 5013 * 1GB 3 250MB 5014 * 10GB 10 0.9GB 5015 * 100GB 31 3GB 5016 * 1TB 101 10GB 5017 * 10TB 320 32GB 5018 */ 5019static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 5020{ 5021 unsigned int gb, ratio; 5022 5023 /* Zone size in gigabytes */ 5024 gb = zone->present_pages >> (30 - PAGE_SHIFT); 5025 if (gb) 5026 ratio = int_sqrt(10 * gb); 5027 else 5028 ratio = 1; 5029 5030 zone->inactive_ratio = ratio; 5031} 5032 5033static void __meminit setup_per_zone_inactive_ratio(void) 5034{ 5035 struct zone *zone; 5036 5037 for_each_zone(zone) 5038 calculate_zone_inactive_ratio(zone); 5039} 5040 5041/* 5042 * Initialise min_free_kbytes. 5043 * 5044 * For small machines we want it small (128k min). For large machines 5045 * we want it large (64MB max). But it is not linear, because network 5046 * bandwidth does not increase linearly with machine size. We use 5047 * 5048 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5049 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5050 * 5051 * which yields 5052 * 5053 * 16MB: 512k 5054 * 32MB: 724k 5055 * 64MB: 1024k 5056 * 128MB: 1448k 5057 * 256MB: 2048k 5058 * 512MB: 2896k 5059 * 1024MB: 4096k 5060 * 2048MB: 5792k 5061 * 4096MB: 8192k 5062 * 8192MB: 11584k 5063 * 16384MB: 16384k 5064 */ 5065int __meminit init_per_zone_wmark_min(void) 5066{ 5067 unsigned long lowmem_kbytes; 5068 5069 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5070 5071 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5072 if (min_free_kbytes < 128) 5073 min_free_kbytes = 128; 5074 if (min_free_kbytes > 65536) 5075 min_free_kbytes = 65536; 5076 setup_per_zone_wmarks(); 5077 refresh_zone_stat_thresholds(); 5078 setup_per_zone_lowmem_reserve(); 5079 setup_per_zone_inactive_ratio(); 5080 return 0; 5081} 5082module_init(init_per_zone_wmark_min) 5083 5084/* 5085 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5086 * that we can call two helper functions whenever min_free_kbytes 5087 * changes. 5088 */ 5089int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5090 void __user *buffer, size_t *length, loff_t *ppos) 5091{ 5092 proc_dointvec(table, write, buffer, length, ppos); 5093 if (write) 5094 setup_per_zone_wmarks(); 5095 return 0; 5096} 5097 5098#ifdef CONFIG_NUMA 5099int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5100 void __user *buffer, size_t *length, loff_t *ppos) 5101{ 5102 struct zone *zone; 5103 int rc; 5104 5105 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5106 if (rc) 5107 return rc; 5108 5109 for_each_zone(zone) 5110 zone->min_unmapped_pages = (zone->present_pages * 5111 sysctl_min_unmapped_ratio) / 100; 5112 return 0; 5113} 5114 5115int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5116 void __user *buffer, size_t *length, loff_t *ppos) 5117{ 5118 struct zone *zone; 5119 int rc; 5120 5121 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5122 if (rc) 5123 return rc; 5124 5125 for_each_zone(zone) 5126 zone->min_slab_pages = (zone->present_pages * 5127 sysctl_min_slab_ratio) / 100; 5128 return 0; 5129} 5130#endif 5131 5132/* 5133 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5134 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5135 * whenever sysctl_lowmem_reserve_ratio changes. 5136 * 5137 * The reserve ratio obviously has absolutely no relation with the 5138 * minimum watermarks. The lowmem reserve ratio can only make sense 5139 * if in function of the boot time zone sizes. 5140 */ 5141int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5142 void __user *buffer, size_t *length, loff_t *ppos) 5143{ 5144 proc_dointvec_minmax(table, write, buffer, length, ppos); 5145 setup_per_zone_lowmem_reserve(); 5146 return 0; 5147} 5148 5149/* 5150 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5151 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5152 * can have before it gets flushed back to buddy allocator. 5153 */ 5154 5155int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5156 void __user *buffer, size_t *length, loff_t *ppos) 5157{ 5158 struct zone *zone; 5159 unsigned int cpu; 5160 int ret; 5161 5162 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5163 if (!write || (ret == -EINVAL)) 5164 return ret; 5165 for_each_populated_zone(zone) { 5166 for_each_possible_cpu(cpu) { 5167 unsigned long high; 5168 high = zone->present_pages / percpu_pagelist_fraction; 5169 setup_pagelist_highmark( 5170 per_cpu_ptr(zone->pageset, cpu), high); 5171 } 5172 } 5173 return 0; 5174} 5175 5176int hashdist = HASHDIST_DEFAULT; 5177 5178#ifdef CONFIG_NUMA 5179static int __init set_hashdist(char *str) 5180{ 5181 if (!str) 5182 return 0; 5183 hashdist = simple_strtoul(str, &str, 0); 5184 return 1; 5185} 5186__setup("hashdist=", set_hashdist); 5187#endif 5188 5189/* 5190 * allocate a large system hash table from bootmem 5191 * - it is assumed that the hash table must contain an exact power-of-2 5192 * quantity of entries 5193 * - limit is the number of hash buckets, not the total allocation size 5194 */ 5195void *__init alloc_large_system_hash(const char *tablename, 5196 unsigned long bucketsize, 5197 unsigned long numentries, 5198 int scale, 5199 int flags, 5200 unsigned int *_hash_shift, 5201 unsigned int *_hash_mask, 5202 unsigned long limit) 5203{ 5204 unsigned long long max = limit; 5205 unsigned long log2qty, size; 5206 void *table = NULL; 5207 5208 /* allow the kernel cmdline to have a say */ 5209 if (!numentries) { 5210 /* round applicable memory size up to nearest megabyte */ 5211 numentries = nr_kernel_pages; 5212 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5213 numentries >>= 20 - PAGE_SHIFT; 5214 numentries <<= 20 - PAGE_SHIFT; 5215 5216 /* limit to 1 bucket per 2^scale bytes of low memory */ 5217 if (scale > PAGE_SHIFT) 5218 numentries >>= (scale - PAGE_SHIFT); 5219 else 5220 numentries <<= (PAGE_SHIFT - scale); 5221 5222 /* Make sure we've got at least a 0-order allocation.. */ 5223 if (unlikely(flags & HASH_SMALL)) { 5224 /* Makes no sense without HASH_EARLY */ 5225 WARN_ON(!(flags & HASH_EARLY)); 5226 if (!(numentries >> *_hash_shift)) { 5227 numentries = 1UL << *_hash_shift; 5228 BUG_ON(!numentries); 5229 } 5230 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5231 numentries = PAGE_SIZE / bucketsize; 5232 } 5233 numentries = roundup_pow_of_two(numentries); 5234 5235 /* limit allocation size to 1/16 total memory by default */ 5236 if (max == 0) { 5237 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5238 do_div(max, bucketsize); 5239 } 5240 5241 if (numentries > max) 5242 numentries = max; 5243 5244 log2qty = ilog2(numentries); 5245 5246 do { 5247 size = bucketsize << log2qty; 5248 if (flags & HASH_EARLY) 5249 table = alloc_bootmem_nopanic(size); 5250 else if (hashdist) 5251 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5252 else { 5253 /* 5254 * If bucketsize is not a power-of-two, we may free 5255 * some pages at the end of hash table which 5256 * alloc_pages_exact() automatically does 5257 */ 5258 if (get_order(size) < MAX_ORDER) { 5259 table = alloc_pages_exact(size, GFP_ATOMIC); 5260 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5261 } 5262 } 5263 } while (!table && size > PAGE_SIZE && --log2qty); 5264 5265 if (!table) 5266 panic("Failed to allocate %s hash table\n", tablename); 5267 5268 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5269 tablename, 5270 (1UL << log2qty), 5271 ilog2(size) - PAGE_SHIFT, 5272 size); 5273 5274 if (_hash_shift) 5275 *_hash_shift = log2qty; 5276 if (_hash_mask) 5277 *_hash_mask = (1 << log2qty) - 1; 5278 5279 return table; 5280} 5281 5282/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5283static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5284 unsigned long pfn) 5285{ 5286#ifdef CONFIG_SPARSEMEM 5287 return __pfn_to_section(pfn)->pageblock_flags; 5288#else 5289 return zone->pageblock_flags; 5290#endif /* CONFIG_SPARSEMEM */ 5291} 5292 5293static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5294{ 5295#ifdef CONFIG_SPARSEMEM 5296 pfn &= (PAGES_PER_SECTION-1); 5297 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5298#else 5299 pfn = pfn - zone->zone_start_pfn; 5300 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5301#endif /* CONFIG_SPARSEMEM */ 5302} 5303 5304/** 5305 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5306 * @page: The page within the block of interest 5307 * @start_bitidx: The first bit of interest to retrieve 5308 * @end_bitidx: The last bit of interest 5309 * returns pageblock_bits flags 5310 */ 5311unsigned long get_pageblock_flags_group(struct page *page, 5312 int start_bitidx, int end_bitidx) 5313{ 5314 struct zone *zone; 5315 unsigned long *bitmap; 5316 unsigned long pfn, bitidx; 5317 unsigned long flags = 0; 5318 unsigned long value = 1; 5319 5320 zone = page_zone(page); 5321 pfn = page_to_pfn(page); 5322 bitmap = get_pageblock_bitmap(zone, pfn); 5323 bitidx = pfn_to_bitidx(zone, pfn); 5324 5325 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5326 if (test_bit(bitidx + start_bitidx, bitmap)) 5327 flags |= value; 5328 5329 return flags; 5330} 5331 5332/** 5333 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5334 * @page: The page within the block of interest 5335 * @start_bitidx: The first bit of interest 5336 * @end_bitidx: The last bit of interest 5337 * @flags: The flags to set 5338 */ 5339void set_pageblock_flags_group(struct page *page, unsigned long flags, 5340 int start_bitidx, int end_bitidx) 5341{ 5342 struct zone *zone; 5343 unsigned long *bitmap; 5344 unsigned long pfn, bitidx; 5345 unsigned long value = 1; 5346 5347 zone = page_zone(page); 5348 pfn = page_to_pfn(page); 5349 bitmap = get_pageblock_bitmap(zone, pfn); 5350 bitidx = pfn_to_bitidx(zone, pfn); 5351 VM_BUG_ON(pfn < zone->zone_start_pfn); 5352 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5353 5354 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5355 if (flags & value) 5356 __set_bit(bitidx + start_bitidx, bitmap); 5357 else 5358 __clear_bit(bitidx + start_bitidx, bitmap); 5359} 5360 5361/* 5362 * This is designed as sub function...plz see page_isolation.c also. 5363 * set/clear page block's type to be ISOLATE. 5364 * page allocater never alloc memory from ISOLATE block. 5365 */ 5366 5367static int 5368__count_immobile_pages(struct zone *zone, struct page *page, int count) 5369{ 5370 unsigned long pfn, iter, found; 5371 /* 5372 * For avoiding noise data, lru_add_drain_all() should be called 5373 * If ZONE_MOVABLE, the zone never contains immobile pages 5374 */ 5375 if (zone_idx(zone) == ZONE_MOVABLE) 5376 return true; 5377 5378 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE) 5379 return true; 5380 5381 pfn = page_to_pfn(page); 5382 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 5383 unsigned long check = pfn + iter; 5384 5385 if (!pfn_valid_within(check)) 5386 continue; 5387 5388 page = pfn_to_page(check); 5389 if (!page_count(page)) { 5390 if (PageBuddy(page)) 5391 iter += (1 << page_order(page)) - 1; 5392 continue; 5393 } 5394 if (!PageLRU(page)) 5395 found++; 5396 /* 5397 * If there are RECLAIMABLE pages, we need to check it. 5398 * But now, memory offline itself doesn't call shrink_slab() 5399 * and it still to be fixed. 5400 */ 5401 /* 5402 * If the page is not RAM, page_count()should be 0. 5403 * we don't need more check. This is an _used_ not-movable page. 5404 * 5405 * The problematic thing here is PG_reserved pages. PG_reserved 5406 * is set to both of a memory hole page and a _used_ kernel 5407 * page at boot. 5408 */ 5409 if (found > count) 5410 return false; 5411 } 5412 return true; 5413} 5414 5415bool is_pageblock_removable_nolock(struct page *page) 5416{ 5417 struct zone *zone; 5418 unsigned long pfn; 5419 5420 /* 5421 * We have to be careful here because we are iterating over memory 5422 * sections which are not zone aware so we might end up outside of 5423 * the zone but still within the section. 5424 * We have to take care about the node as well. If the node is offline 5425 * its NODE_DATA will be NULL - see page_zone. 5426 */ 5427 if (!node_online(page_to_nid(page))) 5428 return false; 5429 5430 zone = page_zone(page); 5431 pfn = page_to_pfn(page); 5432 if (zone->zone_start_pfn > pfn || 5433 zone->zone_start_pfn + zone->spanned_pages <= pfn) 5434 return false; 5435 5436 return __count_immobile_pages(zone, page, 0); 5437} 5438 5439int set_migratetype_isolate(struct page *page) 5440{ 5441 struct zone *zone; 5442 unsigned long flags, pfn; 5443 struct memory_isolate_notify arg; 5444 int notifier_ret; 5445 int ret = -EBUSY; 5446 5447 zone = page_zone(page); 5448 5449 spin_lock_irqsave(&zone->lock, flags); 5450 5451 pfn = page_to_pfn(page); 5452 arg.start_pfn = pfn; 5453 arg.nr_pages = pageblock_nr_pages; 5454 arg.pages_found = 0; 5455 5456 /* 5457 * It may be possible to isolate a pageblock even if the 5458 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5459 * notifier chain is used by balloon drivers to return the 5460 * number of pages in a range that are held by the balloon 5461 * driver to shrink memory. If all the pages are accounted for 5462 * by balloons, are free, or on the LRU, isolation can continue. 5463 * Later, for example, when memory hotplug notifier runs, these 5464 * pages reported as "can be isolated" should be isolated(freed) 5465 * by the balloon driver through the memory notifier chain. 5466 */ 5467 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5468 notifier_ret = notifier_to_errno(notifier_ret); 5469 if (notifier_ret) 5470 goto out; 5471 /* 5472 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. 5473 * We just check MOVABLE pages. 5474 */ 5475 if (__count_immobile_pages(zone, page, arg.pages_found)) 5476 ret = 0; 5477 5478 /* 5479 * immobile means "not-on-lru" paes. If immobile is larger than 5480 * removable-by-driver pages reported by notifier, we'll fail. 5481 */ 5482 5483out: 5484 if (!ret) { 5485 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5486 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5487 } 5488 5489 spin_unlock_irqrestore(&zone->lock, flags); 5490 if (!ret) 5491 drain_all_pages(); 5492 return ret; 5493} 5494 5495void unset_migratetype_isolate(struct page *page) 5496{ 5497 struct zone *zone; 5498 unsigned long flags; 5499 zone = page_zone(page); 5500 spin_lock_irqsave(&zone->lock, flags); 5501 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5502 goto out; 5503 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5504 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5505out: 5506 spin_unlock_irqrestore(&zone->lock, flags); 5507} 5508 5509#ifdef CONFIG_MEMORY_HOTREMOVE 5510/* 5511 * All pages in the range must be isolated before calling this. 5512 */ 5513void 5514__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5515{ 5516 struct page *page; 5517 struct zone *zone; 5518 int order, i; 5519 unsigned long pfn; 5520 unsigned long flags; 5521 /* find the first valid pfn */ 5522 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5523 if (pfn_valid(pfn)) 5524 break; 5525 if (pfn == end_pfn) 5526 return; 5527 zone = page_zone(pfn_to_page(pfn)); 5528 spin_lock_irqsave(&zone->lock, flags); 5529 pfn = start_pfn; 5530 while (pfn < end_pfn) { 5531 if (!pfn_valid(pfn)) { 5532 pfn++; 5533 continue; 5534 } 5535 page = pfn_to_page(pfn); 5536 BUG_ON(page_count(page)); 5537 BUG_ON(!PageBuddy(page)); 5538 order = page_order(page); 5539#ifdef CONFIG_DEBUG_VM 5540 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5541 pfn, 1 << order, end_pfn); 5542#endif 5543 list_del(&page->lru); 5544 rmv_page_order(page); 5545 zone->free_area[order].nr_free--; 5546 __mod_zone_page_state(zone, NR_FREE_PAGES, 5547 - (1UL << order)); 5548 for (i = 0; i < (1 << order); i++) 5549 SetPageReserved((page+i)); 5550 pfn += (1 << order); 5551 } 5552 spin_unlock_irqrestore(&zone->lock, flags); 5553} 5554#endif 5555 5556#ifdef CONFIG_MEMORY_FAILURE 5557bool is_free_buddy_page(struct page *page) 5558{ 5559 struct zone *zone = page_zone(page); 5560 unsigned long pfn = page_to_pfn(page); 5561 unsigned long flags; 5562 int order; 5563 5564 spin_lock_irqsave(&zone->lock, flags); 5565 for (order = 0; order < MAX_ORDER; order++) { 5566 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5567 5568 if (PageBuddy(page_head) && page_order(page_head) >= order) 5569 break; 5570 } 5571 spin_unlock_irqrestore(&zone->lock, flags); 5572 5573 return order < MAX_ORDER; 5574} 5575#endif 5576 5577static struct trace_print_flags pageflag_names[] = { 5578 {1UL << PG_locked, "locked" }, 5579 {1UL << PG_error, "error" }, 5580 {1UL << PG_referenced, "referenced" }, 5581 {1UL << PG_uptodate, "uptodate" }, 5582 {1UL << PG_dirty, "dirty" }, 5583 {1UL << PG_lru, "lru" }, 5584 {1UL << PG_active, "active" }, 5585 {1UL << PG_slab, "slab" }, 5586 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5587 {1UL << PG_arch_1, "arch_1" }, 5588 {1UL << PG_reserved, "reserved" }, 5589 {1UL << PG_private, "private" }, 5590 {1UL << PG_private_2, "private_2" }, 5591 {1UL << PG_writeback, "writeback" }, 5592#ifdef CONFIG_PAGEFLAGS_EXTENDED 5593 {1UL << PG_head, "head" }, 5594 {1UL << PG_tail, "tail" }, 5595#else 5596 {1UL << PG_compound, "compound" }, 5597#endif 5598 {1UL << PG_swapcache, "swapcache" }, 5599 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5600 {1UL << PG_reclaim, "reclaim" }, 5601 {1UL << PG_swapbacked, "swapbacked" }, 5602 {1UL << PG_unevictable, "unevictable" }, 5603#ifdef CONFIG_MMU 5604 {1UL << PG_mlocked, "mlocked" }, 5605#endif 5606#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5607 {1UL << PG_uncached, "uncached" }, 5608#endif 5609#ifdef CONFIG_MEMORY_FAILURE 5610 {1UL << PG_hwpoison, "hwpoison" }, 5611#endif 5612 {-1UL, NULL }, 5613}; 5614 5615static void dump_page_flags(unsigned long flags) 5616{ 5617 const char *delim = ""; 5618 unsigned long mask; 5619 int i; 5620 5621 printk(KERN_ALERT "page flags: %#lx(", flags); 5622 5623 /* remove zone id */ 5624 flags &= (1UL << NR_PAGEFLAGS) - 1; 5625 5626 for (i = 0; pageflag_names[i].name && flags; i++) { 5627 5628 mask = pageflag_names[i].mask; 5629 if ((flags & mask) != mask) 5630 continue; 5631 5632 flags &= ~mask; 5633 printk("%s%s", delim, pageflag_names[i].name); 5634 delim = "|"; 5635 } 5636 5637 /* check for left over flags */ 5638 if (flags) 5639 printk("%s%#lx", delim, flags); 5640 5641 printk(")\n"); 5642} 5643 5644void dump_page(struct page *page) 5645{ 5646 printk(KERN_ALERT 5647 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5648 page, atomic_read(&page->_count), page_mapcount(page), 5649 page->mapping, page->index); 5650 dump_page_flags(page->flags); 5651 mem_cgroup_print_bad_page(page); 5652} 5653