page_alloc.c revision 3ee9a4f086716d792219c021e8509f91165a4128
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 61#include <asm/tlbflush.h> 62#include <asm/div64.h> 63#include "internal.h" 64 65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 66DEFINE_PER_CPU(int, numa_node); 67EXPORT_PER_CPU_SYMBOL(numa_node); 68#endif 69 70#ifdef CONFIG_HAVE_MEMORYLESS_NODES 71/* 72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 75 * defined in <linux/topology.h>. 76 */ 77DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 78EXPORT_PER_CPU_SYMBOL(_numa_mem_); 79#endif 80 81/* 82 * Array of node states. 83 */ 84nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 85 [N_POSSIBLE] = NODE_MASK_ALL, 86 [N_ONLINE] = { { [0] = 1UL } }, 87#ifndef CONFIG_NUMA 88 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 89#ifdef CONFIG_HIGHMEM 90 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 91#endif 92 [N_CPU] = { { [0] = 1UL } }, 93#endif /* NUMA */ 94}; 95EXPORT_SYMBOL(node_states); 96 97unsigned long totalram_pages __read_mostly; 98unsigned long totalreserve_pages __read_mostly; 99int percpu_pagelist_fraction; 100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 101 102#ifdef CONFIG_PM_SLEEP 103/* 104 * The following functions are used by the suspend/hibernate code to temporarily 105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 106 * while devices are suspended. To avoid races with the suspend/hibernate code, 107 * they should always be called with pm_mutex held (gfp_allowed_mask also should 108 * only be modified with pm_mutex held, unless the suspend/hibernate code is 109 * guaranteed not to run in parallel with that modification). 110 */ 111 112static gfp_t saved_gfp_mask; 113 114void pm_restore_gfp_mask(void) 115{ 116 WARN_ON(!mutex_is_locked(&pm_mutex)); 117 if (saved_gfp_mask) { 118 gfp_allowed_mask = saved_gfp_mask; 119 saved_gfp_mask = 0; 120 } 121} 122 123void pm_restrict_gfp_mask(void) 124{ 125 WARN_ON(!mutex_is_locked(&pm_mutex)); 126 WARN_ON(saved_gfp_mask); 127 saved_gfp_mask = gfp_allowed_mask; 128 gfp_allowed_mask &= ~GFP_IOFS; 129} 130#endif /* CONFIG_PM_SLEEP */ 131 132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 133int pageblock_order __read_mostly; 134#endif 135 136static void __free_pages_ok(struct page *page, unsigned int order); 137 138/* 139 * results with 256, 32 in the lowmem_reserve sysctl: 140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 141 * 1G machine -> (16M dma, 784M normal, 224M high) 142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 145 * 146 * TBD: should special case ZONE_DMA32 machines here - in those we normally 147 * don't need any ZONE_NORMAL reservation 148 */ 149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 150#ifdef CONFIG_ZONE_DMA 151 256, 152#endif 153#ifdef CONFIG_ZONE_DMA32 154 256, 155#endif 156#ifdef CONFIG_HIGHMEM 157 32, 158#endif 159 32, 160}; 161 162EXPORT_SYMBOL(totalram_pages); 163 164static char * const zone_names[MAX_NR_ZONES] = { 165#ifdef CONFIG_ZONE_DMA 166 "DMA", 167#endif 168#ifdef CONFIG_ZONE_DMA32 169 "DMA32", 170#endif 171 "Normal", 172#ifdef CONFIG_HIGHMEM 173 "HighMem", 174#endif 175 "Movable", 176}; 177 178int min_free_kbytes = 1024; 179 180static unsigned long __meminitdata nr_kernel_pages; 181static unsigned long __meminitdata nr_all_pages; 182static unsigned long __meminitdata dma_reserve; 183 184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 185 /* 186 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 187 * ranges of memory (RAM) that may be registered with add_active_range(). 188 * Ranges passed to add_active_range() will be merged if possible 189 * so the number of times add_active_range() can be called is 190 * related to the number of nodes and the number of holes 191 */ 192 #ifdef CONFIG_MAX_ACTIVE_REGIONS 193 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 194 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 195 #else 196 #if MAX_NUMNODES >= 32 197 /* If there can be many nodes, allow up to 50 holes per node */ 198 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 199 #else 200 /* By default, allow up to 256 distinct regions */ 201 #define MAX_ACTIVE_REGIONS 256 202 #endif 203 #endif 204 205 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 206 static int __meminitdata nr_nodemap_entries; 207 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 208 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 209 static unsigned long __initdata required_kernelcore; 210 static unsigned long __initdata required_movablecore; 211 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 212 213 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 214 int movable_zone; 215 EXPORT_SYMBOL(movable_zone); 216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 217 218#if MAX_NUMNODES > 1 219int nr_node_ids __read_mostly = MAX_NUMNODES; 220int nr_online_nodes __read_mostly = 1; 221EXPORT_SYMBOL(nr_node_ids); 222EXPORT_SYMBOL(nr_online_nodes); 223#endif 224 225int page_group_by_mobility_disabled __read_mostly; 226 227static void set_pageblock_migratetype(struct page *page, int migratetype) 228{ 229 230 if (unlikely(page_group_by_mobility_disabled)) 231 migratetype = MIGRATE_UNMOVABLE; 232 233 set_pageblock_flags_group(page, (unsigned long)migratetype, 234 PB_migrate, PB_migrate_end); 235} 236 237bool oom_killer_disabled __read_mostly; 238 239#ifdef CONFIG_DEBUG_VM 240static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 241{ 242 int ret = 0; 243 unsigned seq; 244 unsigned long pfn = page_to_pfn(page); 245 246 do { 247 seq = zone_span_seqbegin(zone); 248 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 249 ret = 1; 250 else if (pfn < zone->zone_start_pfn) 251 ret = 1; 252 } while (zone_span_seqretry(zone, seq)); 253 254 return ret; 255} 256 257static int page_is_consistent(struct zone *zone, struct page *page) 258{ 259 if (!pfn_valid_within(page_to_pfn(page))) 260 return 0; 261 if (zone != page_zone(page)) 262 return 0; 263 264 return 1; 265} 266/* 267 * Temporary debugging check for pages not lying within a given zone. 268 */ 269static int bad_range(struct zone *zone, struct page *page) 270{ 271 if (page_outside_zone_boundaries(zone, page)) 272 return 1; 273 if (!page_is_consistent(zone, page)) 274 return 1; 275 276 return 0; 277} 278#else 279static inline int bad_range(struct zone *zone, struct page *page) 280{ 281 return 0; 282} 283#endif 284 285static void bad_page(struct page *page) 286{ 287 static unsigned long resume; 288 static unsigned long nr_shown; 289 static unsigned long nr_unshown; 290 291 /* Don't complain about poisoned pages */ 292 if (PageHWPoison(page)) { 293 reset_page_mapcount(page); /* remove PageBuddy */ 294 return; 295 } 296 297 /* 298 * Allow a burst of 60 reports, then keep quiet for that minute; 299 * or allow a steady drip of one report per second. 300 */ 301 if (nr_shown == 60) { 302 if (time_before(jiffies, resume)) { 303 nr_unshown++; 304 goto out; 305 } 306 if (nr_unshown) { 307 printk(KERN_ALERT 308 "BUG: Bad page state: %lu messages suppressed\n", 309 nr_unshown); 310 nr_unshown = 0; 311 } 312 nr_shown = 0; 313 } 314 if (nr_shown++ == 0) 315 resume = jiffies + 60 * HZ; 316 317 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 318 current->comm, page_to_pfn(page)); 319 dump_page(page); 320 321 print_modules(); 322 dump_stack(); 323out: 324 /* Leave bad fields for debug, except PageBuddy could make trouble */ 325 reset_page_mapcount(page); /* remove PageBuddy */ 326 add_taint(TAINT_BAD_PAGE); 327} 328 329/* 330 * Higher-order pages are called "compound pages". They are structured thusly: 331 * 332 * The first PAGE_SIZE page is called the "head page". 333 * 334 * The remaining PAGE_SIZE pages are called "tail pages". 335 * 336 * All pages have PG_compound set. All pages have their ->private pointing at 337 * the head page (even the head page has this). 338 * 339 * The first tail page's ->lru.next holds the address of the compound page's 340 * put_page() function. Its ->lru.prev holds the order of allocation. 341 * This usage means that zero-order pages may not be compound. 342 */ 343 344static void free_compound_page(struct page *page) 345{ 346 __free_pages_ok(page, compound_order(page)); 347} 348 349void prep_compound_page(struct page *page, unsigned long order) 350{ 351 int i; 352 int nr_pages = 1 << order; 353 354 set_compound_page_dtor(page, free_compound_page); 355 set_compound_order(page, order); 356 __SetPageHead(page); 357 for (i = 1; i < nr_pages; i++) { 358 struct page *p = page + i; 359 360 __SetPageTail(p); 361 p->first_page = page; 362 } 363} 364 365/* update __split_huge_page_refcount if you change this function */ 366static int destroy_compound_page(struct page *page, unsigned long order) 367{ 368 int i; 369 int nr_pages = 1 << order; 370 int bad = 0; 371 372 if (unlikely(compound_order(page) != order) || 373 unlikely(!PageHead(page))) { 374 bad_page(page); 375 bad++; 376 } 377 378 __ClearPageHead(page); 379 380 for (i = 1; i < nr_pages; i++) { 381 struct page *p = page + i; 382 383 if (unlikely(!PageTail(p) || (p->first_page != page))) { 384 bad_page(page); 385 bad++; 386 } 387 __ClearPageTail(p); 388 } 389 390 return bad; 391} 392 393static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 394{ 395 int i; 396 397 /* 398 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 399 * and __GFP_HIGHMEM from hard or soft interrupt context. 400 */ 401 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 402 for (i = 0; i < (1 << order); i++) 403 clear_highpage(page + i); 404} 405 406static inline void set_page_order(struct page *page, int order) 407{ 408 set_page_private(page, order); 409 __SetPageBuddy(page); 410} 411 412static inline void rmv_page_order(struct page *page) 413{ 414 __ClearPageBuddy(page); 415 set_page_private(page, 0); 416} 417 418/* 419 * Locate the struct page for both the matching buddy in our 420 * pair (buddy1) and the combined O(n+1) page they form (page). 421 * 422 * 1) Any buddy B1 will have an order O twin B2 which satisfies 423 * the following equation: 424 * B2 = B1 ^ (1 << O) 425 * For example, if the starting buddy (buddy2) is #8 its order 426 * 1 buddy is #10: 427 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 428 * 429 * 2) Any buddy B will have an order O+1 parent P which 430 * satisfies the following equation: 431 * P = B & ~(1 << O) 432 * 433 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 434 */ 435static inline unsigned long 436__find_buddy_index(unsigned long page_idx, unsigned int order) 437{ 438 return page_idx ^ (1 << order); 439} 440 441/* 442 * This function checks whether a page is free && is the buddy 443 * we can do coalesce a page and its buddy if 444 * (a) the buddy is not in a hole && 445 * (b) the buddy is in the buddy system && 446 * (c) a page and its buddy have the same order && 447 * (d) a page and its buddy are in the same zone. 448 * 449 * For recording whether a page is in the buddy system, we set ->_mapcount -2. 450 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. 451 * 452 * For recording page's order, we use page_private(page). 453 */ 454static inline int page_is_buddy(struct page *page, struct page *buddy, 455 int order) 456{ 457 if (!pfn_valid_within(page_to_pfn(buddy))) 458 return 0; 459 460 if (page_zone_id(page) != page_zone_id(buddy)) 461 return 0; 462 463 if (PageBuddy(buddy) && page_order(buddy) == order) { 464 VM_BUG_ON(page_count(buddy) != 0); 465 return 1; 466 } 467 return 0; 468} 469 470/* 471 * Freeing function for a buddy system allocator. 472 * 473 * The concept of a buddy system is to maintain direct-mapped table 474 * (containing bit values) for memory blocks of various "orders". 475 * The bottom level table contains the map for the smallest allocatable 476 * units of memory (here, pages), and each level above it describes 477 * pairs of units from the levels below, hence, "buddies". 478 * At a high level, all that happens here is marking the table entry 479 * at the bottom level available, and propagating the changes upward 480 * as necessary, plus some accounting needed to play nicely with other 481 * parts of the VM system. 482 * At each level, we keep a list of pages, which are heads of continuous 483 * free pages of length of (1 << order) and marked with _mapcount -2. Page's 484 * order is recorded in page_private(page) field. 485 * So when we are allocating or freeing one, we can derive the state of the 486 * other. That is, if we allocate a small block, and both were 487 * free, the remainder of the region must be split into blocks. 488 * If a block is freed, and its buddy is also free, then this 489 * triggers coalescing into a block of larger size. 490 * 491 * -- wli 492 */ 493 494static inline void __free_one_page(struct page *page, 495 struct zone *zone, unsigned int order, 496 int migratetype) 497{ 498 unsigned long page_idx; 499 unsigned long combined_idx; 500 unsigned long uninitialized_var(buddy_idx); 501 struct page *buddy; 502 503 if (unlikely(PageCompound(page))) 504 if (unlikely(destroy_compound_page(page, order))) 505 return; 506 507 VM_BUG_ON(migratetype == -1); 508 509 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 510 511 VM_BUG_ON(page_idx & ((1 << order) - 1)); 512 VM_BUG_ON(bad_range(zone, page)); 513 514 while (order < MAX_ORDER-1) { 515 buddy_idx = __find_buddy_index(page_idx, order); 516 buddy = page + (buddy_idx - page_idx); 517 if (!page_is_buddy(page, buddy, order)) 518 break; 519 520 /* Our buddy is free, merge with it and move up one order. */ 521 list_del(&buddy->lru); 522 zone->free_area[order].nr_free--; 523 rmv_page_order(buddy); 524 combined_idx = buddy_idx & page_idx; 525 page = page + (combined_idx - page_idx); 526 page_idx = combined_idx; 527 order++; 528 } 529 set_page_order(page, order); 530 531 /* 532 * If this is not the largest possible page, check if the buddy 533 * of the next-highest order is free. If it is, it's possible 534 * that pages are being freed that will coalesce soon. In case, 535 * that is happening, add the free page to the tail of the list 536 * so it's less likely to be used soon and more likely to be merged 537 * as a higher order page 538 */ 539 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 540 struct page *higher_page, *higher_buddy; 541 combined_idx = buddy_idx & page_idx; 542 higher_page = page + (combined_idx - page_idx); 543 buddy_idx = __find_buddy_index(combined_idx, order + 1); 544 higher_buddy = page + (buddy_idx - combined_idx); 545 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 546 list_add_tail(&page->lru, 547 &zone->free_area[order].free_list[migratetype]); 548 goto out; 549 } 550 } 551 552 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 553out: 554 zone->free_area[order].nr_free++; 555} 556 557/* 558 * free_page_mlock() -- clean up attempts to free and mlocked() page. 559 * Page should not be on lru, so no need to fix that up. 560 * free_pages_check() will verify... 561 */ 562static inline void free_page_mlock(struct page *page) 563{ 564 __dec_zone_page_state(page, NR_MLOCK); 565 __count_vm_event(UNEVICTABLE_MLOCKFREED); 566} 567 568static inline int free_pages_check(struct page *page) 569{ 570 if (unlikely(page_mapcount(page) | 571 (page->mapping != NULL) | 572 (atomic_read(&page->_count) != 0) | 573 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | 574 (mem_cgroup_bad_page_check(page)))) { 575 bad_page(page); 576 return 1; 577 } 578 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 579 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 580 return 0; 581} 582 583/* 584 * Frees a number of pages from the PCP lists 585 * Assumes all pages on list are in same zone, and of same order. 586 * count is the number of pages to free. 587 * 588 * If the zone was previously in an "all pages pinned" state then look to 589 * see if this freeing clears that state. 590 * 591 * And clear the zone's pages_scanned counter, to hold off the "all pages are 592 * pinned" detection logic. 593 */ 594static void free_pcppages_bulk(struct zone *zone, int count, 595 struct per_cpu_pages *pcp) 596{ 597 int migratetype = 0; 598 int batch_free = 0; 599 int to_free = count; 600 601 spin_lock(&zone->lock); 602 zone->all_unreclaimable = 0; 603 zone->pages_scanned = 0; 604 605 while (to_free) { 606 struct page *page; 607 struct list_head *list; 608 609 /* 610 * Remove pages from lists in a round-robin fashion. A 611 * batch_free count is maintained that is incremented when an 612 * empty list is encountered. This is so more pages are freed 613 * off fuller lists instead of spinning excessively around empty 614 * lists 615 */ 616 do { 617 batch_free++; 618 if (++migratetype == MIGRATE_PCPTYPES) 619 migratetype = 0; 620 list = &pcp->lists[migratetype]; 621 } while (list_empty(list)); 622 623 /* This is the only non-empty list. Free them all. */ 624 if (batch_free == MIGRATE_PCPTYPES) 625 batch_free = to_free; 626 627 do { 628 page = list_entry(list->prev, struct page, lru); 629 /* must delete as __free_one_page list manipulates */ 630 list_del(&page->lru); 631 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 632 __free_one_page(page, zone, 0, page_private(page)); 633 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 634 } while (--to_free && --batch_free && !list_empty(list)); 635 } 636 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 637 spin_unlock(&zone->lock); 638} 639 640static void free_one_page(struct zone *zone, struct page *page, int order, 641 int migratetype) 642{ 643 spin_lock(&zone->lock); 644 zone->all_unreclaimable = 0; 645 zone->pages_scanned = 0; 646 647 __free_one_page(page, zone, order, migratetype); 648 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 649 spin_unlock(&zone->lock); 650} 651 652static bool free_pages_prepare(struct page *page, unsigned int order) 653{ 654 int i; 655 int bad = 0; 656 657 trace_mm_page_free_direct(page, order); 658 kmemcheck_free_shadow(page, order); 659 660 if (PageAnon(page)) 661 page->mapping = NULL; 662 for (i = 0; i < (1 << order); i++) 663 bad += free_pages_check(page + i); 664 if (bad) 665 return false; 666 667 if (!PageHighMem(page)) { 668 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 669 debug_check_no_obj_freed(page_address(page), 670 PAGE_SIZE << order); 671 } 672 arch_free_page(page, order); 673 kernel_map_pages(page, 1 << order, 0); 674 675 return true; 676} 677 678static void __free_pages_ok(struct page *page, unsigned int order) 679{ 680 unsigned long flags; 681 int wasMlocked = __TestClearPageMlocked(page); 682 683 if (!free_pages_prepare(page, order)) 684 return; 685 686 local_irq_save(flags); 687 if (unlikely(wasMlocked)) 688 free_page_mlock(page); 689 __count_vm_events(PGFREE, 1 << order); 690 free_one_page(page_zone(page), page, order, 691 get_pageblock_migratetype(page)); 692 local_irq_restore(flags); 693} 694 695/* 696 * permit the bootmem allocator to evade page validation on high-order frees 697 */ 698void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 699{ 700 if (order == 0) { 701 __ClearPageReserved(page); 702 set_page_count(page, 0); 703 set_page_refcounted(page); 704 __free_page(page); 705 } else { 706 int loop; 707 708 prefetchw(page); 709 for (loop = 0; loop < BITS_PER_LONG; loop++) { 710 struct page *p = &page[loop]; 711 712 if (loop + 1 < BITS_PER_LONG) 713 prefetchw(p + 1); 714 __ClearPageReserved(p); 715 set_page_count(p, 0); 716 } 717 718 set_page_refcounted(page); 719 __free_pages(page, order); 720 } 721} 722 723 724/* 725 * The order of subdivision here is critical for the IO subsystem. 726 * Please do not alter this order without good reasons and regression 727 * testing. Specifically, as large blocks of memory are subdivided, 728 * the order in which smaller blocks are delivered depends on the order 729 * they're subdivided in this function. This is the primary factor 730 * influencing the order in which pages are delivered to the IO 731 * subsystem according to empirical testing, and this is also justified 732 * by considering the behavior of a buddy system containing a single 733 * large block of memory acted on by a series of small allocations. 734 * This behavior is a critical factor in sglist merging's success. 735 * 736 * -- wli 737 */ 738static inline void expand(struct zone *zone, struct page *page, 739 int low, int high, struct free_area *area, 740 int migratetype) 741{ 742 unsigned long size = 1 << high; 743 744 while (high > low) { 745 area--; 746 high--; 747 size >>= 1; 748 VM_BUG_ON(bad_range(zone, &page[size])); 749 list_add(&page[size].lru, &area->free_list[migratetype]); 750 area->nr_free++; 751 set_page_order(&page[size], high); 752 } 753} 754 755/* 756 * This page is about to be returned from the page allocator 757 */ 758static inline int check_new_page(struct page *page) 759{ 760 if (unlikely(page_mapcount(page) | 761 (page->mapping != NULL) | 762 (atomic_read(&page->_count) != 0) | 763 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | 764 (mem_cgroup_bad_page_check(page)))) { 765 bad_page(page); 766 return 1; 767 } 768 return 0; 769} 770 771static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 772{ 773 int i; 774 775 for (i = 0; i < (1 << order); i++) { 776 struct page *p = page + i; 777 if (unlikely(check_new_page(p))) 778 return 1; 779 } 780 781 set_page_private(page, 0); 782 set_page_refcounted(page); 783 784 arch_alloc_page(page, order); 785 kernel_map_pages(page, 1 << order, 1); 786 787 if (gfp_flags & __GFP_ZERO) 788 prep_zero_page(page, order, gfp_flags); 789 790 if (order && (gfp_flags & __GFP_COMP)) 791 prep_compound_page(page, order); 792 793 return 0; 794} 795 796/* 797 * Go through the free lists for the given migratetype and remove 798 * the smallest available page from the freelists 799 */ 800static inline 801struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 802 int migratetype) 803{ 804 unsigned int current_order; 805 struct free_area * area; 806 struct page *page; 807 808 /* Find a page of the appropriate size in the preferred list */ 809 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 810 area = &(zone->free_area[current_order]); 811 if (list_empty(&area->free_list[migratetype])) 812 continue; 813 814 page = list_entry(area->free_list[migratetype].next, 815 struct page, lru); 816 list_del(&page->lru); 817 rmv_page_order(page); 818 area->nr_free--; 819 expand(zone, page, order, current_order, area, migratetype); 820 return page; 821 } 822 823 return NULL; 824} 825 826 827/* 828 * This array describes the order lists are fallen back to when 829 * the free lists for the desirable migrate type are depleted 830 */ 831static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 832 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 833 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 834 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 835 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 836}; 837 838/* 839 * Move the free pages in a range to the free lists of the requested type. 840 * Note that start_page and end_pages are not aligned on a pageblock 841 * boundary. If alignment is required, use move_freepages_block() 842 */ 843static int move_freepages(struct zone *zone, 844 struct page *start_page, struct page *end_page, 845 int migratetype) 846{ 847 struct page *page; 848 unsigned long order; 849 int pages_moved = 0; 850 851#ifndef CONFIG_HOLES_IN_ZONE 852 /* 853 * page_zone is not safe to call in this context when 854 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 855 * anyway as we check zone boundaries in move_freepages_block(). 856 * Remove at a later date when no bug reports exist related to 857 * grouping pages by mobility 858 */ 859 BUG_ON(page_zone(start_page) != page_zone(end_page)); 860#endif 861 862 for (page = start_page; page <= end_page;) { 863 /* Make sure we are not inadvertently changing nodes */ 864 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 865 866 if (!pfn_valid_within(page_to_pfn(page))) { 867 page++; 868 continue; 869 } 870 871 if (!PageBuddy(page)) { 872 page++; 873 continue; 874 } 875 876 order = page_order(page); 877 list_move(&page->lru, 878 &zone->free_area[order].free_list[migratetype]); 879 page += 1 << order; 880 pages_moved += 1 << order; 881 } 882 883 return pages_moved; 884} 885 886static int move_freepages_block(struct zone *zone, struct page *page, 887 int migratetype) 888{ 889 unsigned long start_pfn, end_pfn; 890 struct page *start_page, *end_page; 891 892 start_pfn = page_to_pfn(page); 893 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 894 start_page = pfn_to_page(start_pfn); 895 end_page = start_page + pageblock_nr_pages - 1; 896 end_pfn = start_pfn + pageblock_nr_pages - 1; 897 898 /* Do not cross zone boundaries */ 899 if (start_pfn < zone->zone_start_pfn) 900 start_page = page; 901 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 902 return 0; 903 904 return move_freepages(zone, start_page, end_page, migratetype); 905} 906 907static void change_pageblock_range(struct page *pageblock_page, 908 int start_order, int migratetype) 909{ 910 int nr_pageblocks = 1 << (start_order - pageblock_order); 911 912 while (nr_pageblocks--) { 913 set_pageblock_migratetype(pageblock_page, migratetype); 914 pageblock_page += pageblock_nr_pages; 915 } 916} 917 918/* Remove an element from the buddy allocator from the fallback list */ 919static inline struct page * 920__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 921{ 922 struct free_area * area; 923 int current_order; 924 struct page *page; 925 int migratetype, i; 926 927 /* Find the largest possible block of pages in the other list */ 928 for (current_order = MAX_ORDER-1; current_order >= order; 929 --current_order) { 930 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 931 migratetype = fallbacks[start_migratetype][i]; 932 933 /* MIGRATE_RESERVE handled later if necessary */ 934 if (migratetype == MIGRATE_RESERVE) 935 continue; 936 937 area = &(zone->free_area[current_order]); 938 if (list_empty(&area->free_list[migratetype])) 939 continue; 940 941 page = list_entry(area->free_list[migratetype].next, 942 struct page, lru); 943 area->nr_free--; 944 945 /* 946 * If breaking a large block of pages, move all free 947 * pages to the preferred allocation list. If falling 948 * back for a reclaimable kernel allocation, be more 949 * aggressive about taking ownership of free pages 950 */ 951 if (unlikely(current_order >= (pageblock_order >> 1)) || 952 start_migratetype == MIGRATE_RECLAIMABLE || 953 page_group_by_mobility_disabled) { 954 unsigned long pages; 955 pages = move_freepages_block(zone, page, 956 start_migratetype); 957 958 /* Claim the whole block if over half of it is free */ 959 if (pages >= (1 << (pageblock_order-1)) || 960 page_group_by_mobility_disabled) 961 set_pageblock_migratetype(page, 962 start_migratetype); 963 964 migratetype = start_migratetype; 965 } 966 967 /* Remove the page from the freelists */ 968 list_del(&page->lru); 969 rmv_page_order(page); 970 971 /* Take ownership for orders >= pageblock_order */ 972 if (current_order >= pageblock_order) 973 change_pageblock_range(page, current_order, 974 start_migratetype); 975 976 expand(zone, page, order, current_order, area, migratetype); 977 978 trace_mm_page_alloc_extfrag(page, order, current_order, 979 start_migratetype, migratetype); 980 981 return page; 982 } 983 } 984 985 return NULL; 986} 987 988/* 989 * Do the hard work of removing an element from the buddy allocator. 990 * Call me with the zone->lock already held. 991 */ 992static struct page *__rmqueue(struct zone *zone, unsigned int order, 993 int migratetype) 994{ 995 struct page *page; 996 997retry_reserve: 998 page = __rmqueue_smallest(zone, order, migratetype); 999 1000 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1001 page = __rmqueue_fallback(zone, order, migratetype); 1002 1003 /* 1004 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1005 * is used because __rmqueue_smallest is an inline function 1006 * and we want just one call site 1007 */ 1008 if (!page) { 1009 migratetype = MIGRATE_RESERVE; 1010 goto retry_reserve; 1011 } 1012 } 1013 1014 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1015 return page; 1016} 1017 1018/* 1019 * Obtain a specified number of elements from the buddy allocator, all under 1020 * a single hold of the lock, for efficiency. Add them to the supplied list. 1021 * Returns the number of new pages which were placed at *list. 1022 */ 1023static int rmqueue_bulk(struct zone *zone, unsigned int order, 1024 unsigned long count, struct list_head *list, 1025 int migratetype, int cold) 1026{ 1027 int i; 1028 1029 spin_lock(&zone->lock); 1030 for (i = 0; i < count; ++i) { 1031 struct page *page = __rmqueue(zone, order, migratetype); 1032 if (unlikely(page == NULL)) 1033 break; 1034 1035 /* 1036 * Split buddy pages returned by expand() are received here 1037 * in physical page order. The page is added to the callers and 1038 * list and the list head then moves forward. From the callers 1039 * perspective, the linked list is ordered by page number in 1040 * some conditions. This is useful for IO devices that can 1041 * merge IO requests if the physical pages are ordered 1042 * properly. 1043 */ 1044 if (likely(cold == 0)) 1045 list_add(&page->lru, list); 1046 else 1047 list_add_tail(&page->lru, list); 1048 set_page_private(page, migratetype); 1049 list = &page->lru; 1050 } 1051 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1052 spin_unlock(&zone->lock); 1053 return i; 1054} 1055 1056#ifdef CONFIG_NUMA 1057/* 1058 * Called from the vmstat counter updater to drain pagesets of this 1059 * currently executing processor on remote nodes after they have 1060 * expired. 1061 * 1062 * Note that this function must be called with the thread pinned to 1063 * a single processor. 1064 */ 1065void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1066{ 1067 unsigned long flags; 1068 int to_drain; 1069 1070 local_irq_save(flags); 1071 if (pcp->count >= pcp->batch) 1072 to_drain = pcp->batch; 1073 else 1074 to_drain = pcp->count; 1075 free_pcppages_bulk(zone, to_drain, pcp); 1076 pcp->count -= to_drain; 1077 local_irq_restore(flags); 1078} 1079#endif 1080 1081/* 1082 * Drain pages of the indicated processor. 1083 * 1084 * The processor must either be the current processor and the 1085 * thread pinned to the current processor or a processor that 1086 * is not online. 1087 */ 1088static void drain_pages(unsigned int cpu) 1089{ 1090 unsigned long flags; 1091 struct zone *zone; 1092 1093 for_each_populated_zone(zone) { 1094 struct per_cpu_pageset *pset; 1095 struct per_cpu_pages *pcp; 1096 1097 local_irq_save(flags); 1098 pset = per_cpu_ptr(zone->pageset, cpu); 1099 1100 pcp = &pset->pcp; 1101 if (pcp->count) { 1102 free_pcppages_bulk(zone, pcp->count, pcp); 1103 pcp->count = 0; 1104 } 1105 local_irq_restore(flags); 1106 } 1107} 1108 1109/* 1110 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1111 */ 1112void drain_local_pages(void *arg) 1113{ 1114 drain_pages(smp_processor_id()); 1115} 1116 1117/* 1118 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1119 */ 1120void drain_all_pages(void) 1121{ 1122 on_each_cpu(drain_local_pages, NULL, 1); 1123} 1124 1125#ifdef CONFIG_HIBERNATION 1126 1127void mark_free_pages(struct zone *zone) 1128{ 1129 unsigned long pfn, max_zone_pfn; 1130 unsigned long flags; 1131 int order, t; 1132 struct list_head *curr; 1133 1134 if (!zone->spanned_pages) 1135 return; 1136 1137 spin_lock_irqsave(&zone->lock, flags); 1138 1139 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1140 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1141 if (pfn_valid(pfn)) { 1142 struct page *page = pfn_to_page(pfn); 1143 1144 if (!swsusp_page_is_forbidden(page)) 1145 swsusp_unset_page_free(page); 1146 } 1147 1148 for_each_migratetype_order(order, t) { 1149 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1150 unsigned long i; 1151 1152 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1153 for (i = 0; i < (1UL << order); i++) 1154 swsusp_set_page_free(pfn_to_page(pfn + i)); 1155 } 1156 } 1157 spin_unlock_irqrestore(&zone->lock, flags); 1158} 1159#endif /* CONFIG_PM */ 1160 1161/* 1162 * Free a 0-order page 1163 * cold == 1 ? free a cold page : free a hot page 1164 */ 1165void free_hot_cold_page(struct page *page, int cold) 1166{ 1167 struct zone *zone = page_zone(page); 1168 struct per_cpu_pages *pcp; 1169 unsigned long flags; 1170 int migratetype; 1171 int wasMlocked = __TestClearPageMlocked(page); 1172 1173 if (!free_pages_prepare(page, 0)) 1174 return; 1175 1176 migratetype = get_pageblock_migratetype(page); 1177 set_page_private(page, migratetype); 1178 local_irq_save(flags); 1179 if (unlikely(wasMlocked)) 1180 free_page_mlock(page); 1181 __count_vm_event(PGFREE); 1182 1183 /* 1184 * We only track unmovable, reclaimable and movable on pcp lists. 1185 * Free ISOLATE pages back to the allocator because they are being 1186 * offlined but treat RESERVE as movable pages so we can get those 1187 * areas back if necessary. Otherwise, we may have to free 1188 * excessively into the page allocator 1189 */ 1190 if (migratetype >= MIGRATE_PCPTYPES) { 1191 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1192 free_one_page(zone, page, 0, migratetype); 1193 goto out; 1194 } 1195 migratetype = MIGRATE_MOVABLE; 1196 } 1197 1198 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1199 if (cold) 1200 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1201 else 1202 list_add(&page->lru, &pcp->lists[migratetype]); 1203 pcp->count++; 1204 if (pcp->count >= pcp->high) { 1205 free_pcppages_bulk(zone, pcp->batch, pcp); 1206 pcp->count -= pcp->batch; 1207 } 1208 1209out: 1210 local_irq_restore(flags); 1211} 1212 1213/* 1214 * split_page takes a non-compound higher-order page, and splits it into 1215 * n (1<<order) sub-pages: page[0..n] 1216 * Each sub-page must be freed individually. 1217 * 1218 * Note: this is probably too low level an operation for use in drivers. 1219 * Please consult with lkml before using this in your driver. 1220 */ 1221void split_page(struct page *page, unsigned int order) 1222{ 1223 int i; 1224 1225 VM_BUG_ON(PageCompound(page)); 1226 VM_BUG_ON(!page_count(page)); 1227 1228#ifdef CONFIG_KMEMCHECK 1229 /* 1230 * Split shadow pages too, because free(page[0]) would 1231 * otherwise free the whole shadow. 1232 */ 1233 if (kmemcheck_page_is_tracked(page)) 1234 split_page(virt_to_page(page[0].shadow), order); 1235#endif 1236 1237 for (i = 1; i < (1 << order); i++) 1238 set_page_refcounted(page + i); 1239} 1240 1241/* 1242 * Similar to split_page except the page is already free. As this is only 1243 * being used for migration, the migratetype of the block also changes. 1244 * As this is called with interrupts disabled, the caller is responsible 1245 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1246 * are enabled. 1247 * 1248 * Note: this is probably too low level an operation for use in drivers. 1249 * Please consult with lkml before using this in your driver. 1250 */ 1251int split_free_page(struct page *page) 1252{ 1253 unsigned int order; 1254 unsigned long watermark; 1255 struct zone *zone; 1256 1257 BUG_ON(!PageBuddy(page)); 1258 1259 zone = page_zone(page); 1260 order = page_order(page); 1261 1262 /* Obey watermarks as if the page was being allocated */ 1263 watermark = low_wmark_pages(zone) + (1 << order); 1264 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1265 return 0; 1266 1267 /* Remove page from free list */ 1268 list_del(&page->lru); 1269 zone->free_area[order].nr_free--; 1270 rmv_page_order(page); 1271 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1272 1273 /* Split into individual pages */ 1274 set_page_refcounted(page); 1275 split_page(page, order); 1276 1277 if (order >= pageblock_order - 1) { 1278 struct page *endpage = page + (1 << order) - 1; 1279 for (; page < endpage; page += pageblock_nr_pages) 1280 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1281 } 1282 1283 return 1 << order; 1284} 1285 1286/* 1287 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1288 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1289 * or two. 1290 */ 1291static inline 1292struct page *buffered_rmqueue(struct zone *preferred_zone, 1293 struct zone *zone, int order, gfp_t gfp_flags, 1294 int migratetype) 1295{ 1296 unsigned long flags; 1297 struct page *page; 1298 int cold = !!(gfp_flags & __GFP_COLD); 1299 1300again: 1301 if (likely(order == 0)) { 1302 struct per_cpu_pages *pcp; 1303 struct list_head *list; 1304 1305 local_irq_save(flags); 1306 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1307 list = &pcp->lists[migratetype]; 1308 if (list_empty(list)) { 1309 pcp->count += rmqueue_bulk(zone, 0, 1310 pcp->batch, list, 1311 migratetype, cold); 1312 if (unlikely(list_empty(list))) 1313 goto failed; 1314 } 1315 1316 if (cold) 1317 page = list_entry(list->prev, struct page, lru); 1318 else 1319 page = list_entry(list->next, struct page, lru); 1320 1321 list_del(&page->lru); 1322 pcp->count--; 1323 } else { 1324 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1325 /* 1326 * __GFP_NOFAIL is not to be used in new code. 1327 * 1328 * All __GFP_NOFAIL callers should be fixed so that they 1329 * properly detect and handle allocation failures. 1330 * 1331 * We most definitely don't want callers attempting to 1332 * allocate greater than order-1 page units with 1333 * __GFP_NOFAIL. 1334 */ 1335 WARN_ON_ONCE(order > 1); 1336 } 1337 spin_lock_irqsave(&zone->lock, flags); 1338 page = __rmqueue(zone, order, migratetype); 1339 spin_unlock(&zone->lock); 1340 if (!page) 1341 goto failed; 1342 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1343 } 1344 1345 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1346 zone_statistics(preferred_zone, zone, gfp_flags); 1347 local_irq_restore(flags); 1348 1349 VM_BUG_ON(bad_range(zone, page)); 1350 if (prep_new_page(page, order, gfp_flags)) 1351 goto again; 1352 return page; 1353 1354failed: 1355 local_irq_restore(flags); 1356 return NULL; 1357} 1358 1359/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1360#define ALLOC_WMARK_MIN WMARK_MIN 1361#define ALLOC_WMARK_LOW WMARK_LOW 1362#define ALLOC_WMARK_HIGH WMARK_HIGH 1363#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1364 1365/* Mask to get the watermark bits */ 1366#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1367 1368#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1369#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1370#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1371 1372#ifdef CONFIG_FAIL_PAGE_ALLOC 1373 1374static struct { 1375 struct fault_attr attr; 1376 1377 u32 ignore_gfp_highmem; 1378 u32 ignore_gfp_wait; 1379 u32 min_order; 1380} fail_page_alloc = { 1381 .attr = FAULT_ATTR_INITIALIZER, 1382 .ignore_gfp_wait = 1, 1383 .ignore_gfp_highmem = 1, 1384 .min_order = 1, 1385}; 1386 1387static int __init setup_fail_page_alloc(char *str) 1388{ 1389 return setup_fault_attr(&fail_page_alloc.attr, str); 1390} 1391__setup("fail_page_alloc=", setup_fail_page_alloc); 1392 1393static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1394{ 1395 if (order < fail_page_alloc.min_order) 1396 return 0; 1397 if (gfp_mask & __GFP_NOFAIL) 1398 return 0; 1399 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1400 return 0; 1401 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1402 return 0; 1403 1404 return should_fail(&fail_page_alloc.attr, 1 << order); 1405} 1406 1407#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1408 1409static int __init fail_page_alloc_debugfs(void) 1410{ 1411 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1412 struct dentry *dir; 1413 1414 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 1415 &fail_page_alloc.attr); 1416 if (IS_ERR(dir)) 1417 return PTR_ERR(dir); 1418 1419 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 1420 &fail_page_alloc.ignore_gfp_wait)) 1421 goto fail; 1422 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1423 &fail_page_alloc.ignore_gfp_highmem)) 1424 goto fail; 1425 if (!debugfs_create_u32("min-order", mode, dir, 1426 &fail_page_alloc.min_order)) 1427 goto fail; 1428 1429 return 0; 1430fail: 1431 debugfs_remove_recursive(dir); 1432 1433 return -ENOMEM; 1434} 1435 1436late_initcall(fail_page_alloc_debugfs); 1437 1438#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1439 1440#else /* CONFIG_FAIL_PAGE_ALLOC */ 1441 1442static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1443{ 1444 return 0; 1445} 1446 1447#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1448 1449/* 1450 * Return true if free pages are above 'mark'. This takes into account the order 1451 * of the allocation. 1452 */ 1453static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1454 int classzone_idx, int alloc_flags, long free_pages) 1455{ 1456 /* free_pages my go negative - that's OK */ 1457 long min = mark; 1458 int o; 1459 1460 free_pages -= (1 << order) + 1; 1461 if (alloc_flags & ALLOC_HIGH) 1462 min -= min / 2; 1463 if (alloc_flags & ALLOC_HARDER) 1464 min -= min / 4; 1465 1466 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1467 return false; 1468 for (o = 0; o < order; o++) { 1469 /* At the next order, this order's pages become unavailable */ 1470 free_pages -= z->free_area[o].nr_free << o; 1471 1472 /* Require fewer higher order pages to be free */ 1473 min >>= 1; 1474 1475 if (free_pages <= min) 1476 return false; 1477 } 1478 return true; 1479} 1480 1481bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1482 int classzone_idx, int alloc_flags) 1483{ 1484 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1485 zone_page_state(z, NR_FREE_PAGES)); 1486} 1487 1488bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1489 int classzone_idx, int alloc_flags) 1490{ 1491 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1492 1493 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1494 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1495 1496 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1497 free_pages); 1498} 1499 1500#ifdef CONFIG_NUMA 1501/* 1502 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1503 * skip over zones that are not allowed by the cpuset, or that have 1504 * been recently (in last second) found to be nearly full. See further 1505 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1506 * that have to skip over a lot of full or unallowed zones. 1507 * 1508 * If the zonelist cache is present in the passed in zonelist, then 1509 * returns a pointer to the allowed node mask (either the current 1510 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1511 * 1512 * If the zonelist cache is not available for this zonelist, does 1513 * nothing and returns NULL. 1514 * 1515 * If the fullzones BITMAP in the zonelist cache is stale (more than 1516 * a second since last zap'd) then we zap it out (clear its bits.) 1517 * 1518 * We hold off even calling zlc_setup, until after we've checked the 1519 * first zone in the zonelist, on the theory that most allocations will 1520 * be satisfied from that first zone, so best to examine that zone as 1521 * quickly as we can. 1522 */ 1523static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1524{ 1525 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1526 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1527 1528 zlc = zonelist->zlcache_ptr; 1529 if (!zlc) 1530 return NULL; 1531 1532 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1533 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1534 zlc->last_full_zap = jiffies; 1535 } 1536 1537 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1538 &cpuset_current_mems_allowed : 1539 &node_states[N_HIGH_MEMORY]; 1540 return allowednodes; 1541} 1542 1543/* 1544 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1545 * if it is worth looking at further for free memory: 1546 * 1) Check that the zone isn't thought to be full (doesn't have its 1547 * bit set in the zonelist_cache fullzones BITMAP). 1548 * 2) Check that the zones node (obtained from the zonelist_cache 1549 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1550 * Return true (non-zero) if zone is worth looking at further, or 1551 * else return false (zero) if it is not. 1552 * 1553 * This check -ignores- the distinction between various watermarks, 1554 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1555 * found to be full for any variation of these watermarks, it will 1556 * be considered full for up to one second by all requests, unless 1557 * we are so low on memory on all allowed nodes that we are forced 1558 * into the second scan of the zonelist. 1559 * 1560 * In the second scan we ignore this zonelist cache and exactly 1561 * apply the watermarks to all zones, even it is slower to do so. 1562 * We are low on memory in the second scan, and should leave no stone 1563 * unturned looking for a free page. 1564 */ 1565static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1566 nodemask_t *allowednodes) 1567{ 1568 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1569 int i; /* index of *z in zonelist zones */ 1570 int n; /* node that zone *z is on */ 1571 1572 zlc = zonelist->zlcache_ptr; 1573 if (!zlc) 1574 return 1; 1575 1576 i = z - zonelist->_zonerefs; 1577 n = zlc->z_to_n[i]; 1578 1579 /* This zone is worth trying if it is allowed but not full */ 1580 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1581} 1582 1583/* 1584 * Given 'z' scanning a zonelist, set the corresponding bit in 1585 * zlc->fullzones, so that subsequent attempts to allocate a page 1586 * from that zone don't waste time re-examining it. 1587 */ 1588static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1589{ 1590 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1591 int i; /* index of *z in zonelist zones */ 1592 1593 zlc = zonelist->zlcache_ptr; 1594 if (!zlc) 1595 return; 1596 1597 i = z - zonelist->_zonerefs; 1598 1599 set_bit(i, zlc->fullzones); 1600} 1601 1602/* 1603 * clear all zones full, called after direct reclaim makes progress so that 1604 * a zone that was recently full is not skipped over for up to a second 1605 */ 1606static void zlc_clear_zones_full(struct zonelist *zonelist) 1607{ 1608 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1609 1610 zlc = zonelist->zlcache_ptr; 1611 if (!zlc) 1612 return; 1613 1614 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1615} 1616 1617#else /* CONFIG_NUMA */ 1618 1619static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1620{ 1621 return NULL; 1622} 1623 1624static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1625 nodemask_t *allowednodes) 1626{ 1627 return 1; 1628} 1629 1630static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1631{ 1632} 1633 1634static void zlc_clear_zones_full(struct zonelist *zonelist) 1635{ 1636} 1637#endif /* CONFIG_NUMA */ 1638 1639/* 1640 * get_page_from_freelist goes through the zonelist trying to allocate 1641 * a page. 1642 */ 1643static struct page * 1644get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1645 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1646 struct zone *preferred_zone, int migratetype) 1647{ 1648 struct zoneref *z; 1649 struct page *page = NULL; 1650 int classzone_idx; 1651 struct zone *zone; 1652 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1653 int zlc_active = 0; /* set if using zonelist_cache */ 1654 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1655 1656 classzone_idx = zone_idx(preferred_zone); 1657zonelist_scan: 1658 /* 1659 * Scan zonelist, looking for a zone with enough free. 1660 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1661 */ 1662 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1663 high_zoneidx, nodemask) { 1664 if (NUMA_BUILD && zlc_active && 1665 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1666 continue; 1667 if ((alloc_flags & ALLOC_CPUSET) && 1668 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1669 continue; 1670 1671 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1672 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1673 unsigned long mark; 1674 int ret; 1675 1676 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1677 if (zone_watermark_ok(zone, order, mark, 1678 classzone_idx, alloc_flags)) 1679 goto try_this_zone; 1680 1681 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1682 /* 1683 * we do zlc_setup if there are multiple nodes 1684 * and before considering the first zone allowed 1685 * by the cpuset. 1686 */ 1687 allowednodes = zlc_setup(zonelist, alloc_flags); 1688 zlc_active = 1; 1689 did_zlc_setup = 1; 1690 } 1691 1692 if (zone_reclaim_mode == 0) 1693 goto this_zone_full; 1694 1695 /* 1696 * As we may have just activated ZLC, check if the first 1697 * eligible zone has failed zone_reclaim recently. 1698 */ 1699 if (NUMA_BUILD && zlc_active && 1700 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1701 continue; 1702 1703 ret = zone_reclaim(zone, gfp_mask, order); 1704 switch (ret) { 1705 case ZONE_RECLAIM_NOSCAN: 1706 /* did not scan */ 1707 continue; 1708 case ZONE_RECLAIM_FULL: 1709 /* scanned but unreclaimable */ 1710 continue; 1711 default: 1712 /* did we reclaim enough */ 1713 if (!zone_watermark_ok(zone, order, mark, 1714 classzone_idx, alloc_flags)) 1715 goto this_zone_full; 1716 } 1717 } 1718 1719try_this_zone: 1720 page = buffered_rmqueue(preferred_zone, zone, order, 1721 gfp_mask, migratetype); 1722 if (page) 1723 break; 1724this_zone_full: 1725 if (NUMA_BUILD) 1726 zlc_mark_zone_full(zonelist, z); 1727 } 1728 1729 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1730 /* Disable zlc cache for second zonelist scan */ 1731 zlc_active = 0; 1732 goto zonelist_scan; 1733 } 1734 return page; 1735} 1736 1737/* 1738 * Large machines with many possible nodes should not always dump per-node 1739 * meminfo in irq context. 1740 */ 1741static inline bool should_suppress_show_mem(void) 1742{ 1743 bool ret = false; 1744 1745#if NODES_SHIFT > 8 1746 ret = in_interrupt(); 1747#endif 1748 return ret; 1749} 1750 1751static DEFINE_RATELIMIT_STATE(nopage_rs, 1752 DEFAULT_RATELIMIT_INTERVAL, 1753 DEFAULT_RATELIMIT_BURST); 1754 1755void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 1756{ 1757 unsigned int filter = SHOW_MEM_FILTER_NODES; 1758 1759 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs)) 1760 return; 1761 1762 /* 1763 * This documents exceptions given to allocations in certain 1764 * contexts that are allowed to allocate outside current's set 1765 * of allowed nodes. 1766 */ 1767 if (!(gfp_mask & __GFP_NOMEMALLOC)) 1768 if (test_thread_flag(TIF_MEMDIE) || 1769 (current->flags & (PF_MEMALLOC | PF_EXITING))) 1770 filter &= ~SHOW_MEM_FILTER_NODES; 1771 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 1772 filter &= ~SHOW_MEM_FILTER_NODES; 1773 1774 if (fmt) { 1775 struct va_format vaf; 1776 va_list args; 1777 1778 va_start(args, fmt); 1779 1780 vaf.fmt = fmt; 1781 vaf.va = &args; 1782 1783 pr_warn("%pV", &vaf); 1784 1785 va_end(args); 1786 } 1787 1788 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 1789 current->comm, order, gfp_mask); 1790 1791 dump_stack(); 1792 if (!should_suppress_show_mem()) 1793 show_mem(filter); 1794} 1795 1796static inline int 1797should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1798 unsigned long pages_reclaimed) 1799{ 1800 /* Do not loop if specifically requested */ 1801 if (gfp_mask & __GFP_NORETRY) 1802 return 0; 1803 1804 /* 1805 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1806 * means __GFP_NOFAIL, but that may not be true in other 1807 * implementations. 1808 */ 1809 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1810 return 1; 1811 1812 /* 1813 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1814 * specified, then we retry until we no longer reclaim any pages 1815 * (above), or we've reclaimed an order of pages at least as 1816 * large as the allocation's order. In both cases, if the 1817 * allocation still fails, we stop retrying. 1818 */ 1819 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1820 return 1; 1821 1822 /* 1823 * Don't let big-order allocations loop unless the caller 1824 * explicitly requests that. 1825 */ 1826 if (gfp_mask & __GFP_NOFAIL) 1827 return 1; 1828 1829 return 0; 1830} 1831 1832static inline struct page * 1833__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1834 struct zonelist *zonelist, enum zone_type high_zoneidx, 1835 nodemask_t *nodemask, struct zone *preferred_zone, 1836 int migratetype) 1837{ 1838 struct page *page; 1839 1840 /* Acquire the OOM killer lock for the zones in zonelist */ 1841 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1842 schedule_timeout_uninterruptible(1); 1843 return NULL; 1844 } 1845 1846 /* 1847 * Go through the zonelist yet one more time, keep very high watermark 1848 * here, this is only to catch a parallel oom killing, we must fail if 1849 * we're still under heavy pressure. 1850 */ 1851 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1852 order, zonelist, high_zoneidx, 1853 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1854 preferred_zone, migratetype); 1855 if (page) 1856 goto out; 1857 1858 if (!(gfp_mask & __GFP_NOFAIL)) { 1859 /* The OOM killer will not help higher order allocs */ 1860 if (order > PAGE_ALLOC_COSTLY_ORDER) 1861 goto out; 1862 /* The OOM killer does not needlessly kill tasks for lowmem */ 1863 if (high_zoneidx < ZONE_NORMAL) 1864 goto out; 1865 /* 1866 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1867 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1868 * The caller should handle page allocation failure by itself if 1869 * it specifies __GFP_THISNODE. 1870 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1871 */ 1872 if (gfp_mask & __GFP_THISNODE) 1873 goto out; 1874 } 1875 /* Exhausted what can be done so it's blamo time */ 1876 out_of_memory(zonelist, gfp_mask, order, nodemask); 1877 1878out: 1879 clear_zonelist_oom(zonelist, gfp_mask); 1880 return page; 1881} 1882 1883#ifdef CONFIG_COMPACTION 1884/* Try memory compaction for high-order allocations before reclaim */ 1885static struct page * 1886__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1887 struct zonelist *zonelist, enum zone_type high_zoneidx, 1888 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1889 int migratetype, unsigned long *did_some_progress, 1890 bool sync_migration) 1891{ 1892 struct page *page; 1893 1894 if (!order || compaction_deferred(preferred_zone)) 1895 return NULL; 1896 1897 current->flags |= PF_MEMALLOC; 1898 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 1899 nodemask, sync_migration); 1900 current->flags &= ~PF_MEMALLOC; 1901 if (*did_some_progress != COMPACT_SKIPPED) { 1902 1903 /* Page migration frees to the PCP lists but we want merging */ 1904 drain_pages(get_cpu()); 1905 put_cpu(); 1906 1907 page = get_page_from_freelist(gfp_mask, nodemask, 1908 order, zonelist, high_zoneidx, 1909 alloc_flags, preferred_zone, 1910 migratetype); 1911 if (page) { 1912 preferred_zone->compact_considered = 0; 1913 preferred_zone->compact_defer_shift = 0; 1914 count_vm_event(COMPACTSUCCESS); 1915 return page; 1916 } 1917 1918 /* 1919 * It's bad if compaction run occurs and fails. 1920 * The most likely reason is that pages exist, 1921 * but not enough to satisfy watermarks. 1922 */ 1923 count_vm_event(COMPACTFAIL); 1924 defer_compaction(preferred_zone); 1925 1926 cond_resched(); 1927 } 1928 1929 return NULL; 1930} 1931#else 1932static inline struct page * 1933__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1934 struct zonelist *zonelist, enum zone_type high_zoneidx, 1935 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1936 int migratetype, unsigned long *did_some_progress, 1937 bool sync_migration) 1938{ 1939 return NULL; 1940} 1941#endif /* CONFIG_COMPACTION */ 1942 1943/* The really slow allocator path where we enter direct reclaim */ 1944static inline struct page * 1945__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1946 struct zonelist *zonelist, enum zone_type high_zoneidx, 1947 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1948 int migratetype, unsigned long *did_some_progress) 1949{ 1950 struct page *page = NULL; 1951 struct reclaim_state reclaim_state; 1952 bool drained = false; 1953 1954 cond_resched(); 1955 1956 /* We now go into synchronous reclaim */ 1957 cpuset_memory_pressure_bump(); 1958 current->flags |= PF_MEMALLOC; 1959 lockdep_set_current_reclaim_state(gfp_mask); 1960 reclaim_state.reclaimed_slab = 0; 1961 current->reclaim_state = &reclaim_state; 1962 1963 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1964 1965 current->reclaim_state = NULL; 1966 lockdep_clear_current_reclaim_state(); 1967 current->flags &= ~PF_MEMALLOC; 1968 1969 cond_resched(); 1970 1971 if (unlikely(!(*did_some_progress))) 1972 return NULL; 1973 1974 /* After successful reclaim, reconsider all zones for allocation */ 1975 if (NUMA_BUILD) 1976 zlc_clear_zones_full(zonelist); 1977 1978retry: 1979 page = get_page_from_freelist(gfp_mask, nodemask, order, 1980 zonelist, high_zoneidx, 1981 alloc_flags, preferred_zone, 1982 migratetype); 1983 1984 /* 1985 * If an allocation failed after direct reclaim, it could be because 1986 * pages are pinned on the per-cpu lists. Drain them and try again 1987 */ 1988 if (!page && !drained) { 1989 drain_all_pages(); 1990 drained = true; 1991 goto retry; 1992 } 1993 1994 return page; 1995} 1996 1997/* 1998 * This is called in the allocator slow-path if the allocation request is of 1999 * sufficient urgency to ignore watermarks and take other desperate measures 2000 */ 2001static inline struct page * 2002__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2003 struct zonelist *zonelist, enum zone_type high_zoneidx, 2004 nodemask_t *nodemask, struct zone *preferred_zone, 2005 int migratetype) 2006{ 2007 struct page *page; 2008 2009 do { 2010 page = get_page_from_freelist(gfp_mask, nodemask, order, 2011 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 2012 preferred_zone, migratetype); 2013 2014 if (!page && gfp_mask & __GFP_NOFAIL) 2015 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2016 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2017 2018 return page; 2019} 2020 2021static inline 2022void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 2023 enum zone_type high_zoneidx, 2024 enum zone_type classzone_idx) 2025{ 2026 struct zoneref *z; 2027 struct zone *zone; 2028 2029 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 2030 wakeup_kswapd(zone, order, classzone_idx); 2031} 2032 2033static inline int 2034gfp_to_alloc_flags(gfp_t gfp_mask) 2035{ 2036 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2037 const gfp_t wait = gfp_mask & __GFP_WAIT; 2038 2039 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2040 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2041 2042 /* 2043 * The caller may dip into page reserves a bit more if the caller 2044 * cannot run direct reclaim, or if the caller has realtime scheduling 2045 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2046 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 2047 */ 2048 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2049 2050 if (!wait) { 2051 /* 2052 * Not worth trying to allocate harder for 2053 * __GFP_NOMEMALLOC even if it can't schedule. 2054 */ 2055 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2056 alloc_flags |= ALLOC_HARDER; 2057 /* 2058 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 2059 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 2060 */ 2061 alloc_flags &= ~ALLOC_CPUSET; 2062 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2063 alloc_flags |= ALLOC_HARDER; 2064 2065 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2066 if (!in_interrupt() && 2067 ((current->flags & PF_MEMALLOC) || 2068 unlikely(test_thread_flag(TIF_MEMDIE)))) 2069 alloc_flags |= ALLOC_NO_WATERMARKS; 2070 } 2071 2072 return alloc_flags; 2073} 2074 2075static inline struct page * 2076__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2077 struct zonelist *zonelist, enum zone_type high_zoneidx, 2078 nodemask_t *nodemask, struct zone *preferred_zone, 2079 int migratetype) 2080{ 2081 const gfp_t wait = gfp_mask & __GFP_WAIT; 2082 struct page *page = NULL; 2083 int alloc_flags; 2084 unsigned long pages_reclaimed = 0; 2085 unsigned long did_some_progress; 2086 bool sync_migration = false; 2087 2088 /* 2089 * In the slowpath, we sanity check order to avoid ever trying to 2090 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2091 * be using allocators in order of preference for an area that is 2092 * too large. 2093 */ 2094 if (order >= MAX_ORDER) { 2095 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2096 return NULL; 2097 } 2098 2099 /* 2100 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2101 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2102 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2103 * using a larger set of nodes after it has established that the 2104 * allowed per node queues are empty and that nodes are 2105 * over allocated. 2106 */ 2107 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2108 goto nopage; 2109 2110restart: 2111 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2112 wake_all_kswapd(order, zonelist, high_zoneidx, 2113 zone_idx(preferred_zone)); 2114 2115 /* 2116 * OK, we're below the kswapd watermark and have kicked background 2117 * reclaim. Now things get more complex, so set up alloc_flags according 2118 * to how we want to proceed. 2119 */ 2120 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2121 2122 /* 2123 * Find the true preferred zone if the allocation is unconstrained by 2124 * cpusets. 2125 */ 2126 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) 2127 first_zones_zonelist(zonelist, high_zoneidx, NULL, 2128 &preferred_zone); 2129 2130rebalance: 2131 /* This is the last chance, in general, before the goto nopage. */ 2132 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2133 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2134 preferred_zone, migratetype); 2135 if (page) 2136 goto got_pg; 2137 2138 /* Allocate without watermarks if the context allows */ 2139 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2140 page = __alloc_pages_high_priority(gfp_mask, order, 2141 zonelist, high_zoneidx, nodemask, 2142 preferred_zone, migratetype); 2143 if (page) 2144 goto got_pg; 2145 } 2146 2147 /* Atomic allocations - we can't balance anything */ 2148 if (!wait) 2149 goto nopage; 2150 2151 /* Avoid recursion of direct reclaim */ 2152 if (current->flags & PF_MEMALLOC) 2153 goto nopage; 2154 2155 /* Avoid allocations with no watermarks from looping endlessly */ 2156 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2157 goto nopage; 2158 2159 /* 2160 * Try direct compaction. The first pass is asynchronous. Subsequent 2161 * attempts after direct reclaim are synchronous 2162 */ 2163 page = __alloc_pages_direct_compact(gfp_mask, order, 2164 zonelist, high_zoneidx, 2165 nodemask, 2166 alloc_flags, preferred_zone, 2167 migratetype, &did_some_progress, 2168 sync_migration); 2169 if (page) 2170 goto got_pg; 2171 sync_migration = true; 2172 2173 /* Try direct reclaim and then allocating */ 2174 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2175 zonelist, high_zoneidx, 2176 nodemask, 2177 alloc_flags, preferred_zone, 2178 migratetype, &did_some_progress); 2179 if (page) 2180 goto got_pg; 2181 2182 /* 2183 * If we failed to make any progress reclaiming, then we are 2184 * running out of options and have to consider going OOM 2185 */ 2186 if (!did_some_progress) { 2187 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2188 if (oom_killer_disabled) 2189 goto nopage; 2190 page = __alloc_pages_may_oom(gfp_mask, order, 2191 zonelist, high_zoneidx, 2192 nodemask, preferred_zone, 2193 migratetype); 2194 if (page) 2195 goto got_pg; 2196 2197 if (!(gfp_mask & __GFP_NOFAIL)) { 2198 /* 2199 * The oom killer is not called for high-order 2200 * allocations that may fail, so if no progress 2201 * is being made, there are no other options and 2202 * retrying is unlikely to help. 2203 */ 2204 if (order > PAGE_ALLOC_COSTLY_ORDER) 2205 goto nopage; 2206 /* 2207 * The oom killer is not called for lowmem 2208 * allocations to prevent needlessly killing 2209 * innocent tasks. 2210 */ 2211 if (high_zoneidx < ZONE_NORMAL) 2212 goto nopage; 2213 } 2214 2215 goto restart; 2216 } 2217 } 2218 2219 /* Check if we should retry the allocation */ 2220 pages_reclaimed += did_some_progress; 2221 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 2222 /* Wait for some write requests to complete then retry */ 2223 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2224 goto rebalance; 2225 } else { 2226 /* 2227 * High-order allocations do not necessarily loop after 2228 * direct reclaim and reclaim/compaction depends on compaction 2229 * being called after reclaim so call directly if necessary 2230 */ 2231 page = __alloc_pages_direct_compact(gfp_mask, order, 2232 zonelist, high_zoneidx, 2233 nodemask, 2234 alloc_flags, preferred_zone, 2235 migratetype, &did_some_progress, 2236 sync_migration); 2237 if (page) 2238 goto got_pg; 2239 } 2240 2241nopage: 2242 warn_alloc_failed(gfp_mask, order, NULL); 2243 return page; 2244got_pg: 2245 if (kmemcheck_enabled) 2246 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2247 return page; 2248 2249} 2250 2251/* 2252 * This is the 'heart' of the zoned buddy allocator. 2253 */ 2254struct page * 2255__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2256 struct zonelist *zonelist, nodemask_t *nodemask) 2257{ 2258 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2259 struct zone *preferred_zone; 2260 struct page *page; 2261 int migratetype = allocflags_to_migratetype(gfp_mask); 2262 2263 gfp_mask &= gfp_allowed_mask; 2264 2265 lockdep_trace_alloc(gfp_mask); 2266 2267 might_sleep_if(gfp_mask & __GFP_WAIT); 2268 2269 if (should_fail_alloc_page(gfp_mask, order)) 2270 return NULL; 2271 2272 /* 2273 * Check the zones suitable for the gfp_mask contain at least one 2274 * valid zone. It's possible to have an empty zonelist as a result 2275 * of GFP_THISNODE and a memoryless node 2276 */ 2277 if (unlikely(!zonelist->_zonerefs->zone)) 2278 return NULL; 2279 2280 get_mems_allowed(); 2281 /* The preferred zone is used for statistics later */ 2282 first_zones_zonelist(zonelist, high_zoneidx, 2283 nodemask ? : &cpuset_current_mems_allowed, 2284 &preferred_zone); 2285 if (!preferred_zone) { 2286 put_mems_allowed(); 2287 return NULL; 2288 } 2289 2290 /* First allocation attempt */ 2291 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2292 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2293 preferred_zone, migratetype); 2294 if (unlikely(!page)) 2295 page = __alloc_pages_slowpath(gfp_mask, order, 2296 zonelist, high_zoneidx, nodemask, 2297 preferred_zone, migratetype); 2298 put_mems_allowed(); 2299 2300 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2301 return page; 2302} 2303EXPORT_SYMBOL(__alloc_pages_nodemask); 2304 2305/* 2306 * Common helper functions. 2307 */ 2308unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2309{ 2310 struct page *page; 2311 2312 /* 2313 * __get_free_pages() returns a 32-bit address, which cannot represent 2314 * a highmem page 2315 */ 2316 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2317 2318 page = alloc_pages(gfp_mask, order); 2319 if (!page) 2320 return 0; 2321 return (unsigned long) page_address(page); 2322} 2323EXPORT_SYMBOL(__get_free_pages); 2324 2325unsigned long get_zeroed_page(gfp_t gfp_mask) 2326{ 2327 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2328} 2329EXPORT_SYMBOL(get_zeroed_page); 2330 2331void __pagevec_free(struct pagevec *pvec) 2332{ 2333 int i = pagevec_count(pvec); 2334 2335 while (--i >= 0) { 2336 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2337 free_hot_cold_page(pvec->pages[i], pvec->cold); 2338 } 2339} 2340 2341void __free_pages(struct page *page, unsigned int order) 2342{ 2343 if (put_page_testzero(page)) { 2344 if (order == 0) 2345 free_hot_cold_page(page, 0); 2346 else 2347 __free_pages_ok(page, order); 2348 } 2349} 2350 2351EXPORT_SYMBOL(__free_pages); 2352 2353void free_pages(unsigned long addr, unsigned int order) 2354{ 2355 if (addr != 0) { 2356 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2357 __free_pages(virt_to_page((void *)addr), order); 2358 } 2359} 2360 2361EXPORT_SYMBOL(free_pages); 2362 2363static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 2364{ 2365 if (addr) { 2366 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2367 unsigned long used = addr + PAGE_ALIGN(size); 2368 2369 split_page(virt_to_page((void *)addr), order); 2370 while (used < alloc_end) { 2371 free_page(used); 2372 used += PAGE_SIZE; 2373 } 2374 } 2375 return (void *)addr; 2376} 2377 2378/** 2379 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2380 * @size: the number of bytes to allocate 2381 * @gfp_mask: GFP flags for the allocation 2382 * 2383 * This function is similar to alloc_pages(), except that it allocates the 2384 * minimum number of pages to satisfy the request. alloc_pages() can only 2385 * allocate memory in power-of-two pages. 2386 * 2387 * This function is also limited by MAX_ORDER. 2388 * 2389 * Memory allocated by this function must be released by free_pages_exact(). 2390 */ 2391void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2392{ 2393 unsigned int order = get_order(size); 2394 unsigned long addr; 2395 2396 addr = __get_free_pages(gfp_mask, order); 2397 return make_alloc_exact(addr, order, size); 2398} 2399EXPORT_SYMBOL(alloc_pages_exact); 2400 2401/** 2402 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 2403 * pages on a node. 2404 * @nid: the preferred node ID where memory should be allocated 2405 * @size: the number of bytes to allocate 2406 * @gfp_mask: GFP flags for the allocation 2407 * 2408 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 2409 * back. 2410 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 2411 * but is not exact. 2412 */ 2413void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 2414{ 2415 unsigned order = get_order(size); 2416 struct page *p = alloc_pages_node(nid, gfp_mask, order); 2417 if (!p) 2418 return NULL; 2419 return make_alloc_exact((unsigned long)page_address(p), order, size); 2420} 2421EXPORT_SYMBOL(alloc_pages_exact_nid); 2422 2423/** 2424 * free_pages_exact - release memory allocated via alloc_pages_exact() 2425 * @virt: the value returned by alloc_pages_exact. 2426 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2427 * 2428 * Release the memory allocated by a previous call to alloc_pages_exact. 2429 */ 2430void free_pages_exact(void *virt, size_t size) 2431{ 2432 unsigned long addr = (unsigned long)virt; 2433 unsigned long end = addr + PAGE_ALIGN(size); 2434 2435 while (addr < end) { 2436 free_page(addr); 2437 addr += PAGE_SIZE; 2438 } 2439} 2440EXPORT_SYMBOL(free_pages_exact); 2441 2442static unsigned int nr_free_zone_pages(int offset) 2443{ 2444 struct zoneref *z; 2445 struct zone *zone; 2446 2447 /* Just pick one node, since fallback list is circular */ 2448 unsigned int sum = 0; 2449 2450 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2451 2452 for_each_zone_zonelist(zone, z, zonelist, offset) { 2453 unsigned long size = zone->present_pages; 2454 unsigned long high = high_wmark_pages(zone); 2455 if (size > high) 2456 sum += size - high; 2457 } 2458 2459 return sum; 2460} 2461 2462/* 2463 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2464 */ 2465unsigned int nr_free_buffer_pages(void) 2466{ 2467 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2468} 2469EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2470 2471/* 2472 * Amount of free RAM allocatable within all zones 2473 */ 2474unsigned int nr_free_pagecache_pages(void) 2475{ 2476 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2477} 2478 2479static inline void show_node(struct zone *zone) 2480{ 2481 if (NUMA_BUILD) 2482 printk("Node %d ", zone_to_nid(zone)); 2483} 2484 2485void si_meminfo(struct sysinfo *val) 2486{ 2487 val->totalram = totalram_pages; 2488 val->sharedram = 0; 2489 val->freeram = global_page_state(NR_FREE_PAGES); 2490 val->bufferram = nr_blockdev_pages(); 2491 val->totalhigh = totalhigh_pages; 2492 val->freehigh = nr_free_highpages(); 2493 val->mem_unit = PAGE_SIZE; 2494} 2495 2496EXPORT_SYMBOL(si_meminfo); 2497 2498#ifdef CONFIG_NUMA 2499void si_meminfo_node(struct sysinfo *val, int nid) 2500{ 2501 pg_data_t *pgdat = NODE_DATA(nid); 2502 2503 val->totalram = pgdat->node_present_pages; 2504 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2505#ifdef CONFIG_HIGHMEM 2506 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2507 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2508 NR_FREE_PAGES); 2509#else 2510 val->totalhigh = 0; 2511 val->freehigh = 0; 2512#endif 2513 val->mem_unit = PAGE_SIZE; 2514} 2515#endif 2516 2517/* 2518 * Determine whether the node should be displayed or not, depending on whether 2519 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 2520 */ 2521bool skip_free_areas_node(unsigned int flags, int nid) 2522{ 2523 bool ret = false; 2524 2525 if (!(flags & SHOW_MEM_FILTER_NODES)) 2526 goto out; 2527 2528 get_mems_allowed(); 2529 ret = !node_isset(nid, cpuset_current_mems_allowed); 2530 put_mems_allowed(); 2531out: 2532 return ret; 2533} 2534 2535#define K(x) ((x) << (PAGE_SHIFT-10)) 2536 2537/* 2538 * Show free area list (used inside shift_scroll-lock stuff) 2539 * We also calculate the percentage fragmentation. We do this by counting the 2540 * memory on each free list with the exception of the first item on the list. 2541 * Suppresses nodes that are not allowed by current's cpuset if 2542 * SHOW_MEM_FILTER_NODES is passed. 2543 */ 2544void show_free_areas(unsigned int filter) 2545{ 2546 int cpu; 2547 struct zone *zone; 2548 2549 for_each_populated_zone(zone) { 2550 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2551 continue; 2552 show_node(zone); 2553 printk("%s per-cpu:\n", zone->name); 2554 2555 for_each_online_cpu(cpu) { 2556 struct per_cpu_pageset *pageset; 2557 2558 pageset = per_cpu_ptr(zone->pageset, cpu); 2559 2560 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2561 cpu, pageset->pcp.high, 2562 pageset->pcp.batch, pageset->pcp.count); 2563 } 2564 } 2565 2566 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2567 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2568 " unevictable:%lu" 2569 " dirty:%lu writeback:%lu unstable:%lu\n" 2570 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2571 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2572 global_page_state(NR_ACTIVE_ANON), 2573 global_page_state(NR_INACTIVE_ANON), 2574 global_page_state(NR_ISOLATED_ANON), 2575 global_page_state(NR_ACTIVE_FILE), 2576 global_page_state(NR_INACTIVE_FILE), 2577 global_page_state(NR_ISOLATED_FILE), 2578 global_page_state(NR_UNEVICTABLE), 2579 global_page_state(NR_FILE_DIRTY), 2580 global_page_state(NR_WRITEBACK), 2581 global_page_state(NR_UNSTABLE_NFS), 2582 global_page_state(NR_FREE_PAGES), 2583 global_page_state(NR_SLAB_RECLAIMABLE), 2584 global_page_state(NR_SLAB_UNRECLAIMABLE), 2585 global_page_state(NR_FILE_MAPPED), 2586 global_page_state(NR_SHMEM), 2587 global_page_state(NR_PAGETABLE), 2588 global_page_state(NR_BOUNCE)); 2589 2590 for_each_populated_zone(zone) { 2591 int i; 2592 2593 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2594 continue; 2595 show_node(zone); 2596 printk("%s" 2597 " free:%lukB" 2598 " min:%lukB" 2599 " low:%lukB" 2600 " high:%lukB" 2601 " active_anon:%lukB" 2602 " inactive_anon:%lukB" 2603 " active_file:%lukB" 2604 " inactive_file:%lukB" 2605 " unevictable:%lukB" 2606 " isolated(anon):%lukB" 2607 " isolated(file):%lukB" 2608 " present:%lukB" 2609 " mlocked:%lukB" 2610 " dirty:%lukB" 2611 " writeback:%lukB" 2612 " mapped:%lukB" 2613 " shmem:%lukB" 2614 " slab_reclaimable:%lukB" 2615 " slab_unreclaimable:%lukB" 2616 " kernel_stack:%lukB" 2617 " pagetables:%lukB" 2618 " unstable:%lukB" 2619 " bounce:%lukB" 2620 " writeback_tmp:%lukB" 2621 " pages_scanned:%lu" 2622 " all_unreclaimable? %s" 2623 "\n", 2624 zone->name, 2625 K(zone_page_state(zone, NR_FREE_PAGES)), 2626 K(min_wmark_pages(zone)), 2627 K(low_wmark_pages(zone)), 2628 K(high_wmark_pages(zone)), 2629 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2630 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2631 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2632 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2633 K(zone_page_state(zone, NR_UNEVICTABLE)), 2634 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2635 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2636 K(zone->present_pages), 2637 K(zone_page_state(zone, NR_MLOCK)), 2638 K(zone_page_state(zone, NR_FILE_DIRTY)), 2639 K(zone_page_state(zone, NR_WRITEBACK)), 2640 K(zone_page_state(zone, NR_FILE_MAPPED)), 2641 K(zone_page_state(zone, NR_SHMEM)), 2642 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2643 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2644 zone_page_state(zone, NR_KERNEL_STACK) * 2645 THREAD_SIZE / 1024, 2646 K(zone_page_state(zone, NR_PAGETABLE)), 2647 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2648 K(zone_page_state(zone, NR_BOUNCE)), 2649 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2650 zone->pages_scanned, 2651 (zone->all_unreclaimable ? "yes" : "no") 2652 ); 2653 printk("lowmem_reserve[]:"); 2654 for (i = 0; i < MAX_NR_ZONES; i++) 2655 printk(" %lu", zone->lowmem_reserve[i]); 2656 printk("\n"); 2657 } 2658 2659 for_each_populated_zone(zone) { 2660 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2661 2662 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2663 continue; 2664 show_node(zone); 2665 printk("%s: ", zone->name); 2666 2667 spin_lock_irqsave(&zone->lock, flags); 2668 for (order = 0; order < MAX_ORDER; order++) { 2669 nr[order] = zone->free_area[order].nr_free; 2670 total += nr[order] << order; 2671 } 2672 spin_unlock_irqrestore(&zone->lock, flags); 2673 for (order = 0; order < MAX_ORDER; order++) 2674 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2675 printk("= %lukB\n", K(total)); 2676 } 2677 2678 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2679 2680 show_swap_cache_info(); 2681} 2682 2683static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2684{ 2685 zoneref->zone = zone; 2686 zoneref->zone_idx = zone_idx(zone); 2687} 2688 2689/* 2690 * Builds allocation fallback zone lists. 2691 * 2692 * Add all populated zones of a node to the zonelist. 2693 */ 2694static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2695 int nr_zones, enum zone_type zone_type) 2696{ 2697 struct zone *zone; 2698 2699 BUG_ON(zone_type >= MAX_NR_ZONES); 2700 zone_type++; 2701 2702 do { 2703 zone_type--; 2704 zone = pgdat->node_zones + zone_type; 2705 if (populated_zone(zone)) { 2706 zoneref_set_zone(zone, 2707 &zonelist->_zonerefs[nr_zones++]); 2708 check_highest_zone(zone_type); 2709 } 2710 2711 } while (zone_type); 2712 return nr_zones; 2713} 2714 2715 2716/* 2717 * zonelist_order: 2718 * 0 = automatic detection of better ordering. 2719 * 1 = order by ([node] distance, -zonetype) 2720 * 2 = order by (-zonetype, [node] distance) 2721 * 2722 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2723 * the same zonelist. So only NUMA can configure this param. 2724 */ 2725#define ZONELIST_ORDER_DEFAULT 0 2726#define ZONELIST_ORDER_NODE 1 2727#define ZONELIST_ORDER_ZONE 2 2728 2729/* zonelist order in the kernel. 2730 * set_zonelist_order() will set this to NODE or ZONE. 2731 */ 2732static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2733static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2734 2735 2736#ifdef CONFIG_NUMA 2737/* The value user specified ....changed by config */ 2738static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2739/* string for sysctl */ 2740#define NUMA_ZONELIST_ORDER_LEN 16 2741char numa_zonelist_order[16] = "default"; 2742 2743/* 2744 * interface for configure zonelist ordering. 2745 * command line option "numa_zonelist_order" 2746 * = "[dD]efault - default, automatic configuration. 2747 * = "[nN]ode - order by node locality, then by zone within node 2748 * = "[zZ]one - order by zone, then by locality within zone 2749 */ 2750 2751static int __parse_numa_zonelist_order(char *s) 2752{ 2753 if (*s == 'd' || *s == 'D') { 2754 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2755 } else if (*s == 'n' || *s == 'N') { 2756 user_zonelist_order = ZONELIST_ORDER_NODE; 2757 } else if (*s == 'z' || *s == 'Z') { 2758 user_zonelist_order = ZONELIST_ORDER_ZONE; 2759 } else { 2760 printk(KERN_WARNING 2761 "Ignoring invalid numa_zonelist_order value: " 2762 "%s\n", s); 2763 return -EINVAL; 2764 } 2765 return 0; 2766} 2767 2768static __init int setup_numa_zonelist_order(char *s) 2769{ 2770 int ret; 2771 2772 if (!s) 2773 return 0; 2774 2775 ret = __parse_numa_zonelist_order(s); 2776 if (ret == 0) 2777 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 2778 2779 return ret; 2780} 2781early_param("numa_zonelist_order", setup_numa_zonelist_order); 2782 2783/* 2784 * sysctl handler for numa_zonelist_order 2785 */ 2786int numa_zonelist_order_handler(ctl_table *table, int write, 2787 void __user *buffer, size_t *length, 2788 loff_t *ppos) 2789{ 2790 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2791 int ret; 2792 static DEFINE_MUTEX(zl_order_mutex); 2793 2794 mutex_lock(&zl_order_mutex); 2795 if (write) 2796 strcpy(saved_string, (char*)table->data); 2797 ret = proc_dostring(table, write, buffer, length, ppos); 2798 if (ret) 2799 goto out; 2800 if (write) { 2801 int oldval = user_zonelist_order; 2802 if (__parse_numa_zonelist_order((char*)table->data)) { 2803 /* 2804 * bogus value. restore saved string 2805 */ 2806 strncpy((char*)table->data, saved_string, 2807 NUMA_ZONELIST_ORDER_LEN); 2808 user_zonelist_order = oldval; 2809 } else if (oldval != user_zonelist_order) { 2810 mutex_lock(&zonelists_mutex); 2811 build_all_zonelists(NULL); 2812 mutex_unlock(&zonelists_mutex); 2813 } 2814 } 2815out: 2816 mutex_unlock(&zl_order_mutex); 2817 return ret; 2818} 2819 2820 2821#define MAX_NODE_LOAD (nr_online_nodes) 2822static int node_load[MAX_NUMNODES]; 2823 2824/** 2825 * find_next_best_node - find the next node that should appear in a given node's fallback list 2826 * @node: node whose fallback list we're appending 2827 * @used_node_mask: nodemask_t of already used nodes 2828 * 2829 * We use a number of factors to determine which is the next node that should 2830 * appear on a given node's fallback list. The node should not have appeared 2831 * already in @node's fallback list, and it should be the next closest node 2832 * according to the distance array (which contains arbitrary distance values 2833 * from each node to each node in the system), and should also prefer nodes 2834 * with no CPUs, since presumably they'll have very little allocation pressure 2835 * on them otherwise. 2836 * It returns -1 if no node is found. 2837 */ 2838static int find_next_best_node(int node, nodemask_t *used_node_mask) 2839{ 2840 int n, val; 2841 int min_val = INT_MAX; 2842 int best_node = -1; 2843 const struct cpumask *tmp = cpumask_of_node(0); 2844 2845 /* Use the local node if we haven't already */ 2846 if (!node_isset(node, *used_node_mask)) { 2847 node_set(node, *used_node_mask); 2848 return node; 2849 } 2850 2851 for_each_node_state(n, N_HIGH_MEMORY) { 2852 2853 /* Don't want a node to appear more than once */ 2854 if (node_isset(n, *used_node_mask)) 2855 continue; 2856 2857 /* Use the distance array to find the distance */ 2858 val = node_distance(node, n); 2859 2860 /* Penalize nodes under us ("prefer the next node") */ 2861 val += (n < node); 2862 2863 /* Give preference to headless and unused nodes */ 2864 tmp = cpumask_of_node(n); 2865 if (!cpumask_empty(tmp)) 2866 val += PENALTY_FOR_NODE_WITH_CPUS; 2867 2868 /* Slight preference for less loaded node */ 2869 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2870 val += node_load[n]; 2871 2872 if (val < min_val) { 2873 min_val = val; 2874 best_node = n; 2875 } 2876 } 2877 2878 if (best_node >= 0) 2879 node_set(best_node, *used_node_mask); 2880 2881 return best_node; 2882} 2883 2884 2885/* 2886 * Build zonelists ordered by node and zones within node. 2887 * This results in maximum locality--normal zone overflows into local 2888 * DMA zone, if any--but risks exhausting DMA zone. 2889 */ 2890static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2891{ 2892 int j; 2893 struct zonelist *zonelist; 2894 2895 zonelist = &pgdat->node_zonelists[0]; 2896 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2897 ; 2898 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2899 MAX_NR_ZONES - 1); 2900 zonelist->_zonerefs[j].zone = NULL; 2901 zonelist->_zonerefs[j].zone_idx = 0; 2902} 2903 2904/* 2905 * Build gfp_thisnode zonelists 2906 */ 2907static void build_thisnode_zonelists(pg_data_t *pgdat) 2908{ 2909 int j; 2910 struct zonelist *zonelist; 2911 2912 zonelist = &pgdat->node_zonelists[1]; 2913 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2914 zonelist->_zonerefs[j].zone = NULL; 2915 zonelist->_zonerefs[j].zone_idx = 0; 2916} 2917 2918/* 2919 * Build zonelists ordered by zone and nodes within zones. 2920 * This results in conserving DMA zone[s] until all Normal memory is 2921 * exhausted, but results in overflowing to remote node while memory 2922 * may still exist in local DMA zone. 2923 */ 2924static int node_order[MAX_NUMNODES]; 2925 2926static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2927{ 2928 int pos, j, node; 2929 int zone_type; /* needs to be signed */ 2930 struct zone *z; 2931 struct zonelist *zonelist; 2932 2933 zonelist = &pgdat->node_zonelists[0]; 2934 pos = 0; 2935 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2936 for (j = 0; j < nr_nodes; j++) { 2937 node = node_order[j]; 2938 z = &NODE_DATA(node)->node_zones[zone_type]; 2939 if (populated_zone(z)) { 2940 zoneref_set_zone(z, 2941 &zonelist->_zonerefs[pos++]); 2942 check_highest_zone(zone_type); 2943 } 2944 } 2945 } 2946 zonelist->_zonerefs[pos].zone = NULL; 2947 zonelist->_zonerefs[pos].zone_idx = 0; 2948} 2949 2950static int default_zonelist_order(void) 2951{ 2952 int nid, zone_type; 2953 unsigned long low_kmem_size,total_size; 2954 struct zone *z; 2955 int average_size; 2956 /* 2957 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 2958 * If they are really small and used heavily, the system can fall 2959 * into OOM very easily. 2960 * This function detect ZONE_DMA/DMA32 size and configures zone order. 2961 */ 2962 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2963 low_kmem_size = 0; 2964 total_size = 0; 2965 for_each_online_node(nid) { 2966 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2967 z = &NODE_DATA(nid)->node_zones[zone_type]; 2968 if (populated_zone(z)) { 2969 if (zone_type < ZONE_NORMAL) 2970 low_kmem_size += z->present_pages; 2971 total_size += z->present_pages; 2972 } else if (zone_type == ZONE_NORMAL) { 2973 /* 2974 * If any node has only lowmem, then node order 2975 * is preferred to allow kernel allocations 2976 * locally; otherwise, they can easily infringe 2977 * on other nodes when there is an abundance of 2978 * lowmem available to allocate from. 2979 */ 2980 return ZONELIST_ORDER_NODE; 2981 } 2982 } 2983 } 2984 if (!low_kmem_size || /* there are no DMA area. */ 2985 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2986 return ZONELIST_ORDER_NODE; 2987 /* 2988 * look into each node's config. 2989 * If there is a node whose DMA/DMA32 memory is very big area on 2990 * local memory, NODE_ORDER may be suitable. 2991 */ 2992 average_size = total_size / 2993 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2994 for_each_online_node(nid) { 2995 low_kmem_size = 0; 2996 total_size = 0; 2997 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2998 z = &NODE_DATA(nid)->node_zones[zone_type]; 2999 if (populated_zone(z)) { 3000 if (zone_type < ZONE_NORMAL) 3001 low_kmem_size += z->present_pages; 3002 total_size += z->present_pages; 3003 } 3004 } 3005 if (low_kmem_size && 3006 total_size > average_size && /* ignore small node */ 3007 low_kmem_size > total_size * 70/100) 3008 return ZONELIST_ORDER_NODE; 3009 } 3010 return ZONELIST_ORDER_ZONE; 3011} 3012 3013static void set_zonelist_order(void) 3014{ 3015 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 3016 current_zonelist_order = default_zonelist_order(); 3017 else 3018 current_zonelist_order = user_zonelist_order; 3019} 3020 3021static void build_zonelists(pg_data_t *pgdat) 3022{ 3023 int j, node, load; 3024 enum zone_type i; 3025 nodemask_t used_mask; 3026 int local_node, prev_node; 3027 struct zonelist *zonelist; 3028 int order = current_zonelist_order; 3029 3030 /* initialize zonelists */ 3031 for (i = 0; i < MAX_ZONELISTS; i++) { 3032 zonelist = pgdat->node_zonelists + i; 3033 zonelist->_zonerefs[0].zone = NULL; 3034 zonelist->_zonerefs[0].zone_idx = 0; 3035 } 3036 3037 /* NUMA-aware ordering of nodes */ 3038 local_node = pgdat->node_id; 3039 load = nr_online_nodes; 3040 prev_node = local_node; 3041 nodes_clear(used_mask); 3042 3043 memset(node_order, 0, sizeof(node_order)); 3044 j = 0; 3045 3046 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 3047 int distance = node_distance(local_node, node); 3048 3049 /* 3050 * If another node is sufficiently far away then it is better 3051 * to reclaim pages in a zone before going off node. 3052 */ 3053 if (distance > RECLAIM_DISTANCE) 3054 zone_reclaim_mode = 1; 3055 3056 /* 3057 * We don't want to pressure a particular node. 3058 * So adding penalty to the first node in same 3059 * distance group to make it round-robin. 3060 */ 3061 if (distance != node_distance(local_node, prev_node)) 3062 node_load[node] = load; 3063 3064 prev_node = node; 3065 load--; 3066 if (order == ZONELIST_ORDER_NODE) 3067 build_zonelists_in_node_order(pgdat, node); 3068 else 3069 node_order[j++] = node; /* remember order */ 3070 } 3071 3072 if (order == ZONELIST_ORDER_ZONE) { 3073 /* calculate node order -- i.e., DMA last! */ 3074 build_zonelists_in_zone_order(pgdat, j); 3075 } 3076 3077 build_thisnode_zonelists(pgdat); 3078} 3079 3080/* Construct the zonelist performance cache - see further mmzone.h */ 3081static void build_zonelist_cache(pg_data_t *pgdat) 3082{ 3083 struct zonelist *zonelist; 3084 struct zonelist_cache *zlc; 3085 struct zoneref *z; 3086 3087 zonelist = &pgdat->node_zonelists[0]; 3088 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 3089 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 3090 for (z = zonelist->_zonerefs; z->zone; z++) 3091 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 3092} 3093 3094#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3095/* 3096 * Return node id of node used for "local" allocations. 3097 * I.e., first node id of first zone in arg node's generic zonelist. 3098 * Used for initializing percpu 'numa_mem', which is used primarily 3099 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 3100 */ 3101int local_memory_node(int node) 3102{ 3103 struct zone *zone; 3104 3105 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 3106 gfp_zone(GFP_KERNEL), 3107 NULL, 3108 &zone); 3109 return zone->node; 3110} 3111#endif 3112 3113#else /* CONFIG_NUMA */ 3114 3115static void set_zonelist_order(void) 3116{ 3117 current_zonelist_order = ZONELIST_ORDER_ZONE; 3118} 3119 3120static void build_zonelists(pg_data_t *pgdat) 3121{ 3122 int node, local_node; 3123 enum zone_type j; 3124 struct zonelist *zonelist; 3125 3126 local_node = pgdat->node_id; 3127 3128 zonelist = &pgdat->node_zonelists[0]; 3129 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3130 3131 /* 3132 * Now we build the zonelist so that it contains the zones 3133 * of all the other nodes. 3134 * We don't want to pressure a particular node, so when 3135 * building the zones for node N, we make sure that the 3136 * zones coming right after the local ones are those from 3137 * node N+1 (modulo N) 3138 */ 3139 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 3140 if (!node_online(node)) 3141 continue; 3142 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3143 MAX_NR_ZONES - 1); 3144 } 3145 for (node = 0; node < local_node; node++) { 3146 if (!node_online(node)) 3147 continue; 3148 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3149 MAX_NR_ZONES - 1); 3150 } 3151 3152 zonelist->_zonerefs[j].zone = NULL; 3153 zonelist->_zonerefs[j].zone_idx = 0; 3154} 3155 3156/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3157static void build_zonelist_cache(pg_data_t *pgdat) 3158{ 3159 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3160} 3161 3162#endif /* CONFIG_NUMA */ 3163 3164/* 3165 * Boot pageset table. One per cpu which is going to be used for all 3166 * zones and all nodes. The parameters will be set in such a way 3167 * that an item put on a list will immediately be handed over to 3168 * the buddy list. This is safe since pageset manipulation is done 3169 * with interrupts disabled. 3170 * 3171 * The boot_pagesets must be kept even after bootup is complete for 3172 * unused processors and/or zones. They do play a role for bootstrapping 3173 * hotplugged processors. 3174 * 3175 * zoneinfo_show() and maybe other functions do 3176 * not check if the processor is online before following the pageset pointer. 3177 * Other parts of the kernel may not check if the zone is available. 3178 */ 3179static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3180static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3181static void setup_zone_pageset(struct zone *zone); 3182 3183/* 3184 * Global mutex to protect against size modification of zonelists 3185 * as well as to serialize pageset setup for the new populated zone. 3186 */ 3187DEFINE_MUTEX(zonelists_mutex); 3188 3189/* return values int ....just for stop_machine() */ 3190static __init_refok int __build_all_zonelists(void *data) 3191{ 3192 int nid; 3193 int cpu; 3194 3195#ifdef CONFIG_NUMA 3196 memset(node_load, 0, sizeof(node_load)); 3197#endif 3198 for_each_online_node(nid) { 3199 pg_data_t *pgdat = NODE_DATA(nid); 3200 3201 build_zonelists(pgdat); 3202 build_zonelist_cache(pgdat); 3203 } 3204 3205 /* 3206 * Initialize the boot_pagesets that are going to be used 3207 * for bootstrapping processors. The real pagesets for 3208 * each zone will be allocated later when the per cpu 3209 * allocator is available. 3210 * 3211 * boot_pagesets are used also for bootstrapping offline 3212 * cpus if the system is already booted because the pagesets 3213 * are needed to initialize allocators on a specific cpu too. 3214 * F.e. the percpu allocator needs the page allocator which 3215 * needs the percpu allocator in order to allocate its pagesets 3216 * (a chicken-egg dilemma). 3217 */ 3218 for_each_possible_cpu(cpu) { 3219 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3220 3221#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3222 /* 3223 * We now know the "local memory node" for each node-- 3224 * i.e., the node of the first zone in the generic zonelist. 3225 * Set up numa_mem percpu variable for on-line cpus. During 3226 * boot, only the boot cpu should be on-line; we'll init the 3227 * secondary cpus' numa_mem as they come on-line. During 3228 * node/memory hotplug, we'll fixup all on-line cpus. 3229 */ 3230 if (cpu_online(cpu)) 3231 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3232#endif 3233 } 3234 3235 return 0; 3236} 3237 3238/* 3239 * Called with zonelists_mutex held always 3240 * unless system_state == SYSTEM_BOOTING. 3241 */ 3242void __ref build_all_zonelists(void *data) 3243{ 3244 set_zonelist_order(); 3245 3246 if (system_state == SYSTEM_BOOTING) { 3247 __build_all_zonelists(NULL); 3248 mminit_verify_zonelist(); 3249 cpuset_init_current_mems_allowed(); 3250 } else { 3251 /* we have to stop all cpus to guarantee there is no user 3252 of zonelist */ 3253#ifdef CONFIG_MEMORY_HOTPLUG 3254 if (data) 3255 setup_zone_pageset((struct zone *)data); 3256#endif 3257 stop_machine(__build_all_zonelists, NULL, NULL); 3258 /* cpuset refresh routine should be here */ 3259 } 3260 vm_total_pages = nr_free_pagecache_pages(); 3261 /* 3262 * Disable grouping by mobility if the number of pages in the 3263 * system is too low to allow the mechanism to work. It would be 3264 * more accurate, but expensive to check per-zone. This check is 3265 * made on memory-hotadd so a system can start with mobility 3266 * disabled and enable it later 3267 */ 3268 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3269 page_group_by_mobility_disabled = 1; 3270 else 3271 page_group_by_mobility_disabled = 0; 3272 3273 printk("Built %i zonelists in %s order, mobility grouping %s. " 3274 "Total pages: %ld\n", 3275 nr_online_nodes, 3276 zonelist_order_name[current_zonelist_order], 3277 page_group_by_mobility_disabled ? "off" : "on", 3278 vm_total_pages); 3279#ifdef CONFIG_NUMA 3280 printk("Policy zone: %s\n", zone_names[policy_zone]); 3281#endif 3282} 3283 3284/* 3285 * Helper functions to size the waitqueue hash table. 3286 * Essentially these want to choose hash table sizes sufficiently 3287 * large so that collisions trying to wait on pages are rare. 3288 * But in fact, the number of active page waitqueues on typical 3289 * systems is ridiculously low, less than 200. So this is even 3290 * conservative, even though it seems large. 3291 * 3292 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3293 * waitqueues, i.e. the size of the waitq table given the number of pages. 3294 */ 3295#define PAGES_PER_WAITQUEUE 256 3296 3297#ifndef CONFIG_MEMORY_HOTPLUG 3298static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3299{ 3300 unsigned long size = 1; 3301 3302 pages /= PAGES_PER_WAITQUEUE; 3303 3304 while (size < pages) 3305 size <<= 1; 3306 3307 /* 3308 * Once we have dozens or even hundreds of threads sleeping 3309 * on IO we've got bigger problems than wait queue collision. 3310 * Limit the size of the wait table to a reasonable size. 3311 */ 3312 size = min(size, 4096UL); 3313 3314 return max(size, 4UL); 3315} 3316#else 3317/* 3318 * A zone's size might be changed by hot-add, so it is not possible to determine 3319 * a suitable size for its wait_table. So we use the maximum size now. 3320 * 3321 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3322 * 3323 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3324 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3325 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3326 * 3327 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3328 * or more by the traditional way. (See above). It equals: 3329 * 3330 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3331 * ia64(16K page size) : = ( 8G + 4M)byte. 3332 * powerpc (64K page size) : = (32G +16M)byte. 3333 */ 3334static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3335{ 3336 return 4096UL; 3337} 3338#endif 3339 3340/* 3341 * This is an integer logarithm so that shifts can be used later 3342 * to extract the more random high bits from the multiplicative 3343 * hash function before the remainder is taken. 3344 */ 3345static inline unsigned long wait_table_bits(unsigned long size) 3346{ 3347 return ffz(~size); 3348} 3349 3350#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3351 3352/* 3353 * Check if a pageblock contains reserved pages 3354 */ 3355static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 3356{ 3357 unsigned long pfn; 3358 3359 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3360 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 3361 return 1; 3362 } 3363 return 0; 3364} 3365 3366/* 3367 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3368 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3369 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3370 * higher will lead to a bigger reserve which will get freed as contiguous 3371 * blocks as reclaim kicks in 3372 */ 3373static void setup_zone_migrate_reserve(struct zone *zone) 3374{ 3375 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 3376 struct page *page; 3377 unsigned long block_migratetype; 3378 int reserve; 3379 3380 /* Get the start pfn, end pfn and the number of blocks to reserve */ 3381 start_pfn = zone->zone_start_pfn; 3382 end_pfn = start_pfn + zone->spanned_pages; 3383 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3384 pageblock_order; 3385 3386 /* 3387 * Reserve blocks are generally in place to help high-order atomic 3388 * allocations that are short-lived. A min_free_kbytes value that 3389 * would result in more than 2 reserve blocks for atomic allocations 3390 * is assumed to be in place to help anti-fragmentation for the 3391 * future allocation of hugepages at runtime. 3392 */ 3393 reserve = min(2, reserve); 3394 3395 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3396 if (!pfn_valid(pfn)) 3397 continue; 3398 page = pfn_to_page(pfn); 3399 3400 /* Watch out for overlapping nodes */ 3401 if (page_to_nid(page) != zone_to_nid(zone)) 3402 continue; 3403 3404 /* Blocks with reserved pages will never free, skip them. */ 3405 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 3406 if (pageblock_is_reserved(pfn, block_end_pfn)) 3407 continue; 3408 3409 block_migratetype = get_pageblock_migratetype(page); 3410 3411 /* If this block is reserved, account for it */ 3412 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 3413 reserve--; 3414 continue; 3415 } 3416 3417 /* Suitable for reserving if this block is movable */ 3418 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 3419 set_pageblock_migratetype(page, MIGRATE_RESERVE); 3420 move_freepages_block(zone, page, MIGRATE_RESERVE); 3421 reserve--; 3422 continue; 3423 } 3424 3425 /* 3426 * If the reserve is met and this is a previous reserved block, 3427 * take it back 3428 */ 3429 if (block_migratetype == MIGRATE_RESERVE) { 3430 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3431 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3432 } 3433 } 3434} 3435 3436/* 3437 * Initially all pages are reserved - free ones are freed 3438 * up by free_all_bootmem() once the early boot process is 3439 * done. Non-atomic initialization, single-pass. 3440 */ 3441void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3442 unsigned long start_pfn, enum memmap_context context) 3443{ 3444 struct page *page; 3445 unsigned long end_pfn = start_pfn + size; 3446 unsigned long pfn; 3447 struct zone *z; 3448 3449 if (highest_memmap_pfn < end_pfn - 1) 3450 highest_memmap_pfn = end_pfn - 1; 3451 3452 z = &NODE_DATA(nid)->node_zones[zone]; 3453 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3454 /* 3455 * There can be holes in boot-time mem_map[]s 3456 * handed to this function. They do not 3457 * exist on hotplugged memory. 3458 */ 3459 if (context == MEMMAP_EARLY) { 3460 if (!early_pfn_valid(pfn)) 3461 continue; 3462 if (!early_pfn_in_nid(pfn, nid)) 3463 continue; 3464 } 3465 page = pfn_to_page(pfn); 3466 set_page_links(page, zone, nid, pfn); 3467 mminit_verify_page_links(page, zone, nid, pfn); 3468 init_page_count(page); 3469 reset_page_mapcount(page); 3470 SetPageReserved(page); 3471 /* 3472 * Mark the block movable so that blocks are reserved for 3473 * movable at startup. This will force kernel allocations 3474 * to reserve their blocks rather than leaking throughout 3475 * the address space during boot when many long-lived 3476 * kernel allocations are made. Later some blocks near 3477 * the start are marked MIGRATE_RESERVE by 3478 * setup_zone_migrate_reserve() 3479 * 3480 * bitmap is created for zone's valid pfn range. but memmap 3481 * can be created for invalid pages (for alignment) 3482 * check here not to call set_pageblock_migratetype() against 3483 * pfn out of zone. 3484 */ 3485 if ((z->zone_start_pfn <= pfn) 3486 && (pfn < z->zone_start_pfn + z->spanned_pages) 3487 && !(pfn & (pageblock_nr_pages - 1))) 3488 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3489 3490 INIT_LIST_HEAD(&page->lru); 3491#ifdef WANT_PAGE_VIRTUAL 3492 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3493 if (!is_highmem_idx(zone)) 3494 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3495#endif 3496 } 3497} 3498 3499static void __meminit zone_init_free_lists(struct zone *zone) 3500{ 3501 int order, t; 3502 for_each_migratetype_order(order, t) { 3503 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3504 zone->free_area[order].nr_free = 0; 3505 } 3506} 3507 3508#ifndef __HAVE_ARCH_MEMMAP_INIT 3509#define memmap_init(size, nid, zone, start_pfn) \ 3510 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3511#endif 3512 3513static int zone_batchsize(struct zone *zone) 3514{ 3515#ifdef CONFIG_MMU 3516 int batch; 3517 3518 /* 3519 * The per-cpu-pages pools are set to around 1000th of the 3520 * size of the zone. But no more than 1/2 of a meg. 3521 * 3522 * OK, so we don't know how big the cache is. So guess. 3523 */ 3524 batch = zone->present_pages / 1024; 3525 if (batch * PAGE_SIZE > 512 * 1024) 3526 batch = (512 * 1024) / PAGE_SIZE; 3527 batch /= 4; /* We effectively *= 4 below */ 3528 if (batch < 1) 3529 batch = 1; 3530 3531 /* 3532 * Clamp the batch to a 2^n - 1 value. Having a power 3533 * of 2 value was found to be more likely to have 3534 * suboptimal cache aliasing properties in some cases. 3535 * 3536 * For example if 2 tasks are alternately allocating 3537 * batches of pages, one task can end up with a lot 3538 * of pages of one half of the possible page colors 3539 * and the other with pages of the other colors. 3540 */ 3541 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3542 3543 return batch; 3544 3545#else 3546 /* The deferral and batching of frees should be suppressed under NOMMU 3547 * conditions. 3548 * 3549 * The problem is that NOMMU needs to be able to allocate large chunks 3550 * of contiguous memory as there's no hardware page translation to 3551 * assemble apparent contiguous memory from discontiguous pages. 3552 * 3553 * Queueing large contiguous runs of pages for batching, however, 3554 * causes the pages to actually be freed in smaller chunks. As there 3555 * can be a significant delay between the individual batches being 3556 * recycled, this leads to the once large chunks of space being 3557 * fragmented and becoming unavailable for high-order allocations. 3558 */ 3559 return 0; 3560#endif 3561} 3562 3563static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3564{ 3565 struct per_cpu_pages *pcp; 3566 int migratetype; 3567 3568 memset(p, 0, sizeof(*p)); 3569 3570 pcp = &p->pcp; 3571 pcp->count = 0; 3572 pcp->high = 6 * batch; 3573 pcp->batch = max(1UL, 1 * batch); 3574 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3575 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3576} 3577 3578/* 3579 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3580 * to the value high for the pageset p. 3581 */ 3582 3583static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3584 unsigned long high) 3585{ 3586 struct per_cpu_pages *pcp; 3587 3588 pcp = &p->pcp; 3589 pcp->high = high; 3590 pcp->batch = max(1UL, high/4); 3591 if ((high/4) > (PAGE_SHIFT * 8)) 3592 pcp->batch = PAGE_SHIFT * 8; 3593} 3594 3595static void setup_zone_pageset(struct zone *zone) 3596{ 3597 int cpu; 3598 3599 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3600 3601 for_each_possible_cpu(cpu) { 3602 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3603 3604 setup_pageset(pcp, zone_batchsize(zone)); 3605 3606 if (percpu_pagelist_fraction) 3607 setup_pagelist_highmark(pcp, 3608 (zone->present_pages / 3609 percpu_pagelist_fraction)); 3610 } 3611} 3612 3613/* 3614 * Allocate per cpu pagesets and initialize them. 3615 * Before this call only boot pagesets were available. 3616 */ 3617void __init setup_per_cpu_pageset(void) 3618{ 3619 struct zone *zone; 3620 3621 for_each_populated_zone(zone) 3622 setup_zone_pageset(zone); 3623} 3624 3625static noinline __init_refok 3626int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3627{ 3628 int i; 3629 struct pglist_data *pgdat = zone->zone_pgdat; 3630 size_t alloc_size; 3631 3632 /* 3633 * The per-page waitqueue mechanism uses hashed waitqueues 3634 * per zone. 3635 */ 3636 zone->wait_table_hash_nr_entries = 3637 wait_table_hash_nr_entries(zone_size_pages); 3638 zone->wait_table_bits = 3639 wait_table_bits(zone->wait_table_hash_nr_entries); 3640 alloc_size = zone->wait_table_hash_nr_entries 3641 * sizeof(wait_queue_head_t); 3642 3643 if (!slab_is_available()) { 3644 zone->wait_table = (wait_queue_head_t *) 3645 alloc_bootmem_node_nopanic(pgdat, alloc_size); 3646 } else { 3647 /* 3648 * This case means that a zone whose size was 0 gets new memory 3649 * via memory hot-add. 3650 * But it may be the case that a new node was hot-added. In 3651 * this case vmalloc() will not be able to use this new node's 3652 * memory - this wait_table must be initialized to use this new 3653 * node itself as well. 3654 * To use this new node's memory, further consideration will be 3655 * necessary. 3656 */ 3657 zone->wait_table = vmalloc(alloc_size); 3658 } 3659 if (!zone->wait_table) 3660 return -ENOMEM; 3661 3662 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3663 init_waitqueue_head(zone->wait_table + i); 3664 3665 return 0; 3666} 3667 3668static int __zone_pcp_update(void *data) 3669{ 3670 struct zone *zone = data; 3671 int cpu; 3672 unsigned long batch = zone_batchsize(zone), flags; 3673 3674 for_each_possible_cpu(cpu) { 3675 struct per_cpu_pageset *pset; 3676 struct per_cpu_pages *pcp; 3677 3678 pset = per_cpu_ptr(zone->pageset, cpu); 3679 pcp = &pset->pcp; 3680 3681 local_irq_save(flags); 3682 free_pcppages_bulk(zone, pcp->count, pcp); 3683 setup_pageset(pset, batch); 3684 local_irq_restore(flags); 3685 } 3686 return 0; 3687} 3688 3689void zone_pcp_update(struct zone *zone) 3690{ 3691 stop_machine(__zone_pcp_update, zone, NULL); 3692} 3693 3694static __meminit void zone_pcp_init(struct zone *zone) 3695{ 3696 /* 3697 * per cpu subsystem is not up at this point. The following code 3698 * relies on the ability of the linker to provide the 3699 * offset of a (static) per cpu variable into the per cpu area. 3700 */ 3701 zone->pageset = &boot_pageset; 3702 3703 if (zone->present_pages) 3704 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3705 zone->name, zone->present_pages, 3706 zone_batchsize(zone)); 3707} 3708 3709__meminit int init_currently_empty_zone(struct zone *zone, 3710 unsigned long zone_start_pfn, 3711 unsigned long size, 3712 enum memmap_context context) 3713{ 3714 struct pglist_data *pgdat = zone->zone_pgdat; 3715 int ret; 3716 ret = zone_wait_table_init(zone, size); 3717 if (ret) 3718 return ret; 3719 pgdat->nr_zones = zone_idx(zone) + 1; 3720 3721 zone->zone_start_pfn = zone_start_pfn; 3722 3723 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3724 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3725 pgdat->node_id, 3726 (unsigned long)zone_idx(zone), 3727 zone_start_pfn, (zone_start_pfn + size)); 3728 3729 zone_init_free_lists(zone); 3730 3731 return 0; 3732} 3733 3734#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3735/* 3736 * Basic iterator support. Return the first range of PFNs for a node 3737 * Note: nid == MAX_NUMNODES returns first region regardless of node 3738 */ 3739static int __meminit first_active_region_index_in_nid(int nid) 3740{ 3741 int i; 3742 3743 for (i = 0; i < nr_nodemap_entries; i++) 3744 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3745 return i; 3746 3747 return -1; 3748} 3749 3750/* 3751 * Basic iterator support. Return the next active range of PFNs for a node 3752 * Note: nid == MAX_NUMNODES returns next region regardless of node 3753 */ 3754static int __meminit next_active_region_index_in_nid(int index, int nid) 3755{ 3756 for (index = index + 1; index < nr_nodemap_entries; index++) 3757 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3758 return index; 3759 3760 return -1; 3761} 3762 3763#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3764/* 3765 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3766 * Architectures may implement their own version but if add_active_range() 3767 * was used and there are no special requirements, this is a convenient 3768 * alternative 3769 */ 3770int __meminit __early_pfn_to_nid(unsigned long pfn) 3771{ 3772 int i; 3773 3774 for (i = 0; i < nr_nodemap_entries; i++) { 3775 unsigned long start_pfn = early_node_map[i].start_pfn; 3776 unsigned long end_pfn = early_node_map[i].end_pfn; 3777 3778 if (start_pfn <= pfn && pfn < end_pfn) 3779 return early_node_map[i].nid; 3780 } 3781 /* This is a memory hole */ 3782 return -1; 3783} 3784#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3785 3786int __meminit early_pfn_to_nid(unsigned long pfn) 3787{ 3788 int nid; 3789 3790 nid = __early_pfn_to_nid(pfn); 3791 if (nid >= 0) 3792 return nid; 3793 /* just returns 0 */ 3794 return 0; 3795} 3796 3797#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3798bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3799{ 3800 int nid; 3801 3802 nid = __early_pfn_to_nid(pfn); 3803 if (nid >= 0 && nid != node) 3804 return false; 3805 return true; 3806} 3807#endif 3808 3809/* Basic iterator support to walk early_node_map[] */ 3810#define for_each_active_range_index_in_nid(i, nid) \ 3811 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3812 i = next_active_region_index_in_nid(i, nid)) 3813 3814/** 3815 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3816 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3817 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3818 * 3819 * If an architecture guarantees that all ranges registered with 3820 * add_active_ranges() contain no holes and may be freed, this 3821 * this function may be used instead of calling free_bootmem() manually. 3822 */ 3823void __init free_bootmem_with_active_regions(int nid, 3824 unsigned long max_low_pfn) 3825{ 3826 int i; 3827 3828 for_each_active_range_index_in_nid(i, nid) { 3829 unsigned long size_pages = 0; 3830 unsigned long end_pfn = early_node_map[i].end_pfn; 3831 3832 if (early_node_map[i].start_pfn >= max_low_pfn) 3833 continue; 3834 3835 if (end_pfn > max_low_pfn) 3836 end_pfn = max_low_pfn; 3837 3838 size_pages = end_pfn - early_node_map[i].start_pfn; 3839 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3840 PFN_PHYS(early_node_map[i].start_pfn), 3841 size_pages << PAGE_SHIFT); 3842 } 3843} 3844 3845#ifdef CONFIG_HAVE_MEMBLOCK 3846/* 3847 * Basic iterator support. Return the last range of PFNs for a node 3848 * Note: nid == MAX_NUMNODES returns last region regardless of node 3849 */ 3850static int __meminit last_active_region_index_in_nid(int nid) 3851{ 3852 int i; 3853 3854 for (i = nr_nodemap_entries - 1; i >= 0; i--) 3855 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3856 return i; 3857 3858 return -1; 3859} 3860 3861/* 3862 * Basic iterator support. Return the previous active range of PFNs for a node 3863 * Note: nid == MAX_NUMNODES returns next region regardless of node 3864 */ 3865static int __meminit previous_active_region_index_in_nid(int index, int nid) 3866{ 3867 for (index = index - 1; index >= 0; index--) 3868 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3869 return index; 3870 3871 return -1; 3872} 3873 3874#define for_each_active_range_index_in_nid_reverse(i, nid) \ 3875 for (i = last_active_region_index_in_nid(nid); i != -1; \ 3876 i = previous_active_region_index_in_nid(i, nid)) 3877 3878u64 __init find_memory_core_early(int nid, u64 size, u64 align, 3879 u64 goal, u64 limit) 3880{ 3881 int i; 3882 3883 /* Need to go over early_node_map to find out good range for node */ 3884 for_each_active_range_index_in_nid_reverse(i, nid) { 3885 u64 addr; 3886 u64 ei_start, ei_last; 3887 u64 final_start, final_end; 3888 3889 ei_last = early_node_map[i].end_pfn; 3890 ei_last <<= PAGE_SHIFT; 3891 ei_start = early_node_map[i].start_pfn; 3892 ei_start <<= PAGE_SHIFT; 3893 3894 final_start = max(ei_start, goal); 3895 final_end = min(ei_last, limit); 3896 3897 if (final_start >= final_end) 3898 continue; 3899 3900 addr = memblock_find_in_range(final_start, final_end, size, align); 3901 3902 if (addr == MEMBLOCK_ERROR) 3903 continue; 3904 3905 return addr; 3906 } 3907 3908 return MEMBLOCK_ERROR; 3909} 3910#endif 3911 3912int __init add_from_early_node_map(struct range *range, int az, 3913 int nr_range, int nid) 3914{ 3915 int i; 3916 u64 start, end; 3917 3918 /* need to go over early_node_map to find out good range for node */ 3919 for_each_active_range_index_in_nid(i, nid) { 3920 start = early_node_map[i].start_pfn; 3921 end = early_node_map[i].end_pfn; 3922 nr_range = add_range(range, az, nr_range, start, end); 3923 } 3924 return nr_range; 3925} 3926 3927void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3928{ 3929 int i; 3930 int ret; 3931 3932 for_each_active_range_index_in_nid(i, nid) { 3933 ret = work_fn(early_node_map[i].start_pfn, 3934 early_node_map[i].end_pfn, data); 3935 if (ret) 3936 break; 3937 } 3938} 3939/** 3940 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3941 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3942 * 3943 * If an architecture guarantees that all ranges registered with 3944 * add_active_ranges() contain no holes and may be freed, this 3945 * function may be used instead of calling memory_present() manually. 3946 */ 3947void __init sparse_memory_present_with_active_regions(int nid) 3948{ 3949 int i; 3950 3951 for_each_active_range_index_in_nid(i, nid) 3952 memory_present(early_node_map[i].nid, 3953 early_node_map[i].start_pfn, 3954 early_node_map[i].end_pfn); 3955} 3956 3957/** 3958 * get_pfn_range_for_nid - Return the start and end page frames for a node 3959 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3960 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3961 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3962 * 3963 * It returns the start and end page frame of a node based on information 3964 * provided by an arch calling add_active_range(). If called for a node 3965 * with no available memory, a warning is printed and the start and end 3966 * PFNs will be 0. 3967 */ 3968void __meminit get_pfn_range_for_nid(unsigned int nid, 3969 unsigned long *start_pfn, unsigned long *end_pfn) 3970{ 3971 int i; 3972 *start_pfn = -1UL; 3973 *end_pfn = 0; 3974 3975 for_each_active_range_index_in_nid(i, nid) { 3976 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3977 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3978 } 3979 3980 if (*start_pfn == -1UL) 3981 *start_pfn = 0; 3982} 3983 3984/* 3985 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3986 * assumption is made that zones within a node are ordered in monotonic 3987 * increasing memory addresses so that the "highest" populated zone is used 3988 */ 3989static void __init find_usable_zone_for_movable(void) 3990{ 3991 int zone_index; 3992 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3993 if (zone_index == ZONE_MOVABLE) 3994 continue; 3995 3996 if (arch_zone_highest_possible_pfn[zone_index] > 3997 arch_zone_lowest_possible_pfn[zone_index]) 3998 break; 3999 } 4000 4001 VM_BUG_ON(zone_index == -1); 4002 movable_zone = zone_index; 4003} 4004 4005/* 4006 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 4007 * because it is sized independent of architecture. Unlike the other zones, 4008 * the starting point for ZONE_MOVABLE is not fixed. It may be different 4009 * in each node depending on the size of each node and how evenly kernelcore 4010 * is distributed. This helper function adjusts the zone ranges 4011 * provided by the architecture for a given node by using the end of the 4012 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 4013 * zones within a node are in order of monotonic increases memory addresses 4014 */ 4015static void __meminit adjust_zone_range_for_zone_movable(int nid, 4016 unsigned long zone_type, 4017 unsigned long node_start_pfn, 4018 unsigned long node_end_pfn, 4019 unsigned long *zone_start_pfn, 4020 unsigned long *zone_end_pfn) 4021{ 4022 /* Only adjust if ZONE_MOVABLE is on this node */ 4023 if (zone_movable_pfn[nid]) { 4024 /* Size ZONE_MOVABLE */ 4025 if (zone_type == ZONE_MOVABLE) { 4026 *zone_start_pfn = zone_movable_pfn[nid]; 4027 *zone_end_pfn = min(node_end_pfn, 4028 arch_zone_highest_possible_pfn[movable_zone]); 4029 4030 /* Adjust for ZONE_MOVABLE starting within this range */ 4031 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 4032 *zone_end_pfn > zone_movable_pfn[nid]) { 4033 *zone_end_pfn = zone_movable_pfn[nid]; 4034 4035 /* Check if this whole range is within ZONE_MOVABLE */ 4036 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 4037 *zone_start_pfn = *zone_end_pfn; 4038 } 4039} 4040 4041/* 4042 * Return the number of pages a zone spans in a node, including holes 4043 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 4044 */ 4045static unsigned long __meminit zone_spanned_pages_in_node(int nid, 4046 unsigned long zone_type, 4047 unsigned long *ignored) 4048{ 4049 unsigned long node_start_pfn, node_end_pfn; 4050 unsigned long zone_start_pfn, zone_end_pfn; 4051 4052 /* Get the start and end of the node and zone */ 4053 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4054 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 4055 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 4056 adjust_zone_range_for_zone_movable(nid, zone_type, 4057 node_start_pfn, node_end_pfn, 4058 &zone_start_pfn, &zone_end_pfn); 4059 4060 /* Check that this node has pages within the zone's required range */ 4061 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 4062 return 0; 4063 4064 /* Move the zone boundaries inside the node if necessary */ 4065 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 4066 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 4067 4068 /* Return the spanned pages */ 4069 return zone_end_pfn - zone_start_pfn; 4070} 4071 4072/* 4073 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 4074 * then all holes in the requested range will be accounted for. 4075 */ 4076unsigned long __meminit __absent_pages_in_range(int nid, 4077 unsigned long range_start_pfn, 4078 unsigned long range_end_pfn) 4079{ 4080 int i = 0; 4081 unsigned long prev_end_pfn = 0, hole_pages = 0; 4082 unsigned long start_pfn; 4083 4084 /* Find the end_pfn of the first active range of pfns in the node */ 4085 i = first_active_region_index_in_nid(nid); 4086 if (i == -1) 4087 return 0; 4088 4089 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 4090 4091 /* Account for ranges before physical memory on this node */ 4092 if (early_node_map[i].start_pfn > range_start_pfn) 4093 hole_pages = prev_end_pfn - range_start_pfn; 4094 4095 /* Find all holes for the zone within the node */ 4096 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 4097 4098 /* No need to continue if prev_end_pfn is outside the zone */ 4099 if (prev_end_pfn >= range_end_pfn) 4100 break; 4101 4102 /* Make sure the end of the zone is not within the hole */ 4103 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 4104 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 4105 4106 /* Update the hole size cound and move on */ 4107 if (start_pfn > range_start_pfn) { 4108 BUG_ON(prev_end_pfn > start_pfn); 4109 hole_pages += start_pfn - prev_end_pfn; 4110 } 4111 prev_end_pfn = early_node_map[i].end_pfn; 4112 } 4113 4114 /* Account for ranges past physical memory on this node */ 4115 if (range_end_pfn > prev_end_pfn) 4116 hole_pages += range_end_pfn - 4117 max(range_start_pfn, prev_end_pfn); 4118 4119 return hole_pages; 4120} 4121 4122/** 4123 * absent_pages_in_range - Return number of page frames in holes within a range 4124 * @start_pfn: The start PFN to start searching for holes 4125 * @end_pfn: The end PFN to stop searching for holes 4126 * 4127 * It returns the number of pages frames in memory holes within a range. 4128 */ 4129unsigned long __init absent_pages_in_range(unsigned long start_pfn, 4130 unsigned long end_pfn) 4131{ 4132 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 4133} 4134 4135/* Return the number of page frames in holes in a zone on a node */ 4136static unsigned long __meminit zone_absent_pages_in_node(int nid, 4137 unsigned long zone_type, 4138 unsigned long *ignored) 4139{ 4140 unsigned long node_start_pfn, node_end_pfn; 4141 unsigned long zone_start_pfn, zone_end_pfn; 4142 4143 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4144 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 4145 node_start_pfn); 4146 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 4147 node_end_pfn); 4148 4149 adjust_zone_range_for_zone_movable(nid, zone_type, 4150 node_start_pfn, node_end_pfn, 4151 &zone_start_pfn, &zone_end_pfn); 4152 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 4153} 4154 4155#else 4156static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 4157 unsigned long zone_type, 4158 unsigned long *zones_size) 4159{ 4160 return zones_size[zone_type]; 4161} 4162 4163static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4164 unsigned long zone_type, 4165 unsigned long *zholes_size) 4166{ 4167 if (!zholes_size) 4168 return 0; 4169 4170 return zholes_size[zone_type]; 4171} 4172 4173#endif 4174 4175static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4176 unsigned long *zones_size, unsigned long *zholes_size) 4177{ 4178 unsigned long realtotalpages, totalpages = 0; 4179 enum zone_type i; 4180 4181 for (i = 0; i < MAX_NR_ZONES; i++) 4182 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4183 zones_size); 4184 pgdat->node_spanned_pages = totalpages; 4185 4186 realtotalpages = totalpages; 4187 for (i = 0; i < MAX_NR_ZONES; i++) 4188 realtotalpages -= 4189 zone_absent_pages_in_node(pgdat->node_id, i, 4190 zholes_size); 4191 pgdat->node_present_pages = realtotalpages; 4192 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4193 realtotalpages); 4194} 4195 4196#ifndef CONFIG_SPARSEMEM 4197/* 4198 * Calculate the size of the zone->blockflags rounded to an unsigned long 4199 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4200 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4201 * round what is now in bits to nearest long in bits, then return it in 4202 * bytes. 4203 */ 4204static unsigned long __init usemap_size(unsigned long zonesize) 4205{ 4206 unsigned long usemapsize; 4207 4208 usemapsize = roundup(zonesize, pageblock_nr_pages); 4209 usemapsize = usemapsize >> pageblock_order; 4210 usemapsize *= NR_PAGEBLOCK_BITS; 4211 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4212 4213 return usemapsize / 8; 4214} 4215 4216static void __init setup_usemap(struct pglist_data *pgdat, 4217 struct zone *zone, unsigned long zonesize) 4218{ 4219 unsigned long usemapsize = usemap_size(zonesize); 4220 zone->pageblock_flags = NULL; 4221 if (usemapsize) 4222 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, 4223 usemapsize); 4224} 4225#else 4226static inline void setup_usemap(struct pglist_data *pgdat, 4227 struct zone *zone, unsigned long zonesize) {} 4228#endif /* CONFIG_SPARSEMEM */ 4229 4230#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4231 4232/* Return a sensible default order for the pageblock size. */ 4233static inline int pageblock_default_order(void) 4234{ 4235 if (HPAGE_SHIFT > PAGE_SHIFT) 4236 return HUGETLB_PAGE_ORDER; 4237 4238 return MAX_ORDER-1; 4239} 4240 4241/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4242static inline void __init set_pageblock_order(unsigned int order) 4243{ 4244 /* Check that pageblock_nr_pages has not already been setup */ 4245 if (pageblock_order) 4246 return; 4247 4248 /* 4249 * Assume the largest contiguous order of interest is a huge page. 4250 * This value may be variable depending on boot parameters on IA64 4251 */ 4252 pageblock_order = order; 4253} 4254#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4255 4256/* 4257 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4258 * and pageblock_default_order() are unused as pageblock_order is set 4259 * at compile-time. See include/linux/pageblock-flags.h for the values of 4260 * pageblock_order based on the kernel config 4261 */ 4262static inline int pageblock_default_order(unsigned int order) 4263{ 4264 return MAX_ORDER-1; 4265} 4266#define set_pageblock_order(x) do {} while (0) 4267 4268#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4269 4270/* 4271 * Set up the zone data structures: 4272 * - mark all pages reserved 4273 * - mark all memory queues empty 4274 * - clear the memory bitmaps 4275 */ 4276static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4277 unsigned long *zones_size, unsigned long *zholes_size) 4278{ 4279 enum zone_type j; 4280 int nid = pgdat->node_id; 4281 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4282 int ret; 4283 4284 pgdat_resize_init(pgdat); 4285 pgdat->nr_zones = 0; 4286 init_waitqueue_head(&pgdat->kswapd_wait); 4287 pgdat->kswapd_max_order = 0; 4288 pgdat_page_cgroup_init(pgdat); 4289 4290 for (j = 0; j < MAX_NR_ZONES; j++) { 4291 struct zone *zone = pgdat->node_zones + j; 4292 unsigned long size, realsize, memmap_pages; 4293 enum lru_list l; 4294 4295 size = zone_spanned_pages_in_node(nid, j, zones_size); 4296 realsize = size - zone_absent_pages_in_node(nid, j, 4297 zholes_size); 4298 4299 /* 4300 * Adjust realsize so that it accounts for how much memory 4301 * is used by this zone for memmap. This affects the watermark 4302 * and per-cpu initialisations 4303 */ 4304 memmap_pages = 4305 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4306 if (realsize >= memmap_pages) { 4307 realsize -= memmap_pages; 4308 if (memmap_pages) 4309 printk(KERN_DEBUG 4310 " %s zone: %lu pages used for memmap\n", 4311 zone_names[j], memmap_pages); 4312 } else 4313 printk(KERN_WARNING 4314 " %s zone: %lu pages exceeds realsize %lu\n", 4315 zone_names[j], memmap_pages, realsize); 4316 4317 /* Account for reserved pages */ 4318 if (j == 0 && realsize > dma_reserve) { 4319 realsize -= dma_reserve; 4320 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4321 zone_names[0], dma_reserve); 4322 } 4323 4324 if (!is_highmem_idx(j)) 4325 nr_kernel_pages += realsize; 4326 nr_all_pages += realsize; 4327 4328 zone->spanned_pages = size; 4329 zone->present_pages = realsize; 4330#ifdef CONFIG_NUMA 4331 zone->node = nid; 4332 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4333 / 100; 4334 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4335#endif 4336 zone->name = zone_names[j]; 4337 spin_lock_init(&zone->lock); 4338 spin_lock_init(&zone->lru_lock); 4339 zone_seqlock_init(zone); 4340 zone->zone_pgdat = pgdat; 4341 4342 zone_pcp_init(zone); 4343 for_each_lru(l) 4344 INIT_LIST_HEAD(&zone->lru[l].list); 4345 zone->reclaim_stat.recent_rotated[0] = 0; 4346 zone->reclaim_stat.recent_rotated[1] = 0; 4347 zone->reclaim_stat.recent_scanned[0] = 0; 4348 zone->reclaim_stat.recent_scanned[1] = 0; 4349 zap_zone_vm_stats(zone); 4350 zone->flags = 0; 4351 if (!size) 4352 continue; 4353 4354 set_pageblock_order(pageblock_default_order()); 4355 setup_usemap(pgdat, zone, size); 4356 ret = init_currently_empty_zone(zone, zone_start_pfn, 4357 size, MEMMAP_EARLY); 4358 BUG_ON(ret); 4359 memmap_init(size, nid, j, zone_start_pfn); 4360 zone_start_pfn += size; 4361 } 4362} 4363 4364static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4365{ 4366 /* Skip empty nodes */ 4367 if (!pgdat->node_spanned_pages) 4368 return; 4369 4370#ifdef CONFIG_FLAT_NODE_MEM_MAP 4371 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4372 if (!pgdat->node_mem_map) { 4373 unsigned long size, start, end; 4374 struct page *map; 4375 4376 /* 4377 * The zone's endpoints aren't required to be MAX_ORDER 4378 * aligned but the node_mem_map endpoints must be in order 4379 * for the buddy allocator to function correctly. 4380 */ 4381 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4382 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4383 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4384 size = (end - start) * sizeof(struct page); 4385 map = alloc_remap(pgdat->node_id, size); 4386 if (!map) 4387 map = alloc_bootmem_node_nopanic(pgdat, size); 4388 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4389 } 4390#ifndef CONFIG_NEED_MULTIPLE_NODES 4391 /* 4392 * With no DISCONTIG, the global mem_map is just set as node 0's 4393 */ 4394 if (pgdat == NODE_DATA(0)) { 4395 mem_map = NODE_DATA(0)->node_mem_map; 4396#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4397 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4398 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4399#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4400 } 4401#endif 4402#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4403} 4404 4405void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4406 unsigned long node_start_pfn, unsigned long *zholes_size) 4407{ 4408 pg_data_t *pgdat = NODE_DATA(nid); 4409 4410 pgdat->node_id = nid; 4411 pgdat->node_start_pfn = node_start_pfn; 4412 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4413 4414 alloc_node_mem_map(pgdat); 4415#ifdef CONFIG_FLAT_NODE_MEM_MAP 4416 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4417 nid, (unsigned long)pgdat, 4418 (unsigned long)pgdat->node_mem_map); 4419#endif 4420 4421 free_area_init_core(pgdat, zones_size, zholes_size); 4422} 4423 4424#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4425 4426#if MAX_NUMNODES > 1 4427/* 4428 * Figure out the number of possible node ids. 4429 */ 4430static void __init setup_nr_node_ids(void) 4431{ 4432 unsigned int node; 4433 unsigned int highest = 0; 4434 4435 for_each_node_mask(node, node_possible_map) 4436 highest = node; 4437 nr_node_ids = highest + 1; 4438} 4439#else 4440static inline void setup_nr_node_ids(void) 4441{ 4442} 4443#endif 4444 4445/** 4446 * add_active_range - Register a range of PFNs backed by physical memory 4447 * @nid: The node ID the range resides on 4448 * @start_pfn: The start PFN of the available physical memory 4449 * @end_pfn: The end PFN of the available physical memory 4450 * 4451 * These ranges are stored in an early_node_map[] and later used by 4452 * free_area_init_nodes() to calculate zone sizes and holes. If the 4453 * range spans a memory hole, it is up to the architecture to ensure 4454 * the memory is not freed by the bootmem allocator. If possible 4455 * the range being registered will be merged with existing ranges. 4456 */ 4457void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4458 unsigned long end_pfn) 4459{ 4460 int i; 4461 4462 mminit_dprintk(MMINIT_TRACE, "memory_register", 4463 "Entering add_active_range(%d, %#lx, %#lx) " 4464 "%d entries of %d used\n", 4465 nid, start_pfn, end_pfn, 4466 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4467 4468 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4469 4470 /* Merge with existing active regions if possible */ 4471 for (i = 0; i < nr_nodemap_entries; i++) { 4472 if (early_node_map[i].nid != nid) 4473 continue; 4474 4475 /* Skip if an existing region covers this new one */ 4476 if (start_pfn >= early_node_map[i].start_pfn && 4477 end_pfn <= early_node_map[i].end_pfn) 4478 return; 4479 4480 /* Merge forward if suitable */ 4481 if (start_pfn <= early_node_map[i].end_pfn && 4482 end_pfn > early_node_map[i].end_pfn) { 4483 early_node_map[i].end_pfn = end_pfn; 4484 return; 4485 } 4486 4487 /* Merge backward if suitable */ 4488 if (start_pfn < early_node_map[i].start_pfn && 4489 end_pfn >= early_node_map[i].start_pfn) { 4490 early_node_map[i].start_pfn = start_pfn; 4491 return; 4492 } 4493 } 4494 4495 /* Check that early_node_map is large enough */ 4496 if (i >= MAX_ACTIVE_REGIONS) { 4497 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4498 MAX_ACTIVE_REGIONS); 4499 return; 4500 } 4501 4502 early_node_map[i].nid = nid; 4503 early_node_map[i].start_pfn = start_pfn; 4504 early_node_map[i].end_pfn = end_pfn; 4505 nr_nodemap_entries = i + 1; 4506} 4507 4508/** 4509 * remove_active_range - Shrink an existing registered range of PFNs 4510 * @nid: The node id the range is on that should be shrunk 4511 * @start_pfn: The new PFN of the range 4512 * @end_pfn: The new PFN of the range 4513 * 4514 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4515 * The map is kept near the end physical page range that has already been 4516 * registered. This function allows an arch to shrink an existing registered 4517 * range. 4518 */ 4519void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4520 unsigned long end_pfn) 4521{ 4522 int i, j; 4523 int removed = 0; 4524 4525 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4526 nid, start_pfn, end_pfn); 4527 4528 /* Find the old active region end and shrink */ 4529 for_each_active_range_index_in_nid(i, nid) { 4530 if (early_node_map[i].start_pfn >= start_pfn && 4531 early_node_map[i].end_pfn <= end_pfn) { 4532 /* clear it */ 4533 early_node_map[i].start_pfn = 0; 4534 early_node_map[i].end_pfn = 0; 4535 removed = 1; 4536 continue; 4537 } 4538 if (early_node_map[i].start_pfn < start_pfn && 4539 early_node_map[i].end_pfn > start_pfn) { 4540 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4541 early_node_map[i].end_pfn = start_pfn; 4542 if (temp_end_pfn > end_pfn) 4543 add_active_range(nid, end_pfn, temp_end_pfn); 4544 continue; 4545 } 4546 if (early_node_map[i].start_pfn >= start_pfn && 4547 early_node_map[i].end_pfn > end_pfn && 4548 early_node_map[i].start_pfn < end_pfn) { 4549 early_node_map[i].start_pfn = end_pfn; 4550 continue; 4551 } 4552 } 4553 4554 if (!removed) 4555 return; 4556 4557 /* remove the blank ones */ 4558 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4559 if (early_node_map[i].nid != nid) 4560 continue; 4561 if (early_node_map[i].end_pfn) 4562 continue; 4563 /* we found it, get rid of it */ 4564 for (j = i; j < nr_nodemap_entries - 1; j++) 4565 memcpy(&early_node_map[j], &early_node_map[j+1], 4566 sizeof(early_node_map[j])); 4567 j = nr_nodemap_entries - 1; 4568 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4569 nr_nodemap_entries--; 4570 } 4571} 4572 4573/** 4574 * remove_all_active_ranges - Remove all currently registered regions 4575 * 4576 * During discovery, it may be found that a table like SRAT is invalid 4577 * and an alternative discovery method must be used. This function removes 4578 * all currently registered regions. 4579 */ 4580void __init remove_all_active_ranges(void) 4581{ 4582 memset(early_node_map, 0, sizeof(early_node_map)); 4583 nr_nodemap_entries = 0; 4584} 4585 4586/* Compare two active node_active_regions */ 4587static int __init cmp_node_active_region(const void *a, const void *b) 4588{ 4589 struct node_active_region *arange = (struct node_active_region *)a; 4590 struct node_active_region *brange = (struct node_active_region *)b; 4591 4592 /* Done this way to avoid overflows */ 4593 if (arange->start_pfn > brange->start_pfn) 4594 return 1; 4595 if (arange->start_pfn < brange->start_pfn) 4596 return -1; 4597 4598 return 0; 4599} 4600 4601/* sort the node_map by start_pfn */ 4602void __init sort_node_map(void) 4603{ 4604 sort(early_node_map, (size_t)nr_nodemap_entries, 4605 sizeof(struct node_active_region), 4606 cmp_node_active_region, NULL); 4607} 4608 4609/** 4610 * node_map_pfn_alignment - determine the maximum internode alignment 4611 * 4612 * This function should be called after node map is populated and sorted. 4613 * It calculates the maximum power of two alignment which can distinguish 4614 * all the nodes. 4615 * 4616 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 4617 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 4618 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 4619 * shifted, 1GiB is enough and this function will indicate so. 4620 * 4621 * This is used to test whether pfn -> nid mapping of the chosen memory 4622 * model has fine enough granularity to avoid incorrect mapping for the 4623 * populated node map. 4624 * 4625 * Returns the determined alignment in pfn's. 0 if there is no alignment 4626 * requirement (single node). 4627 */ 4628unsigned long __init node_map_pfn_alignment(void) 4629{ 4630 unsigned long accl_mask = 0, last_end = 0; 4631 int last_nid = -1; 4632 int i; 4633 4634 for_each_active_range_index_in_nid(i, MAX_NUMNODES) { 4635 int nid = early_node_map[i].nid; 4636 unsigned long start = early_node_map[i].start_pfn; 4637 unsigned long end = early_node_map[i].end_pfn; 4638 unsigned long mask; 4639 4640 if (!start || last_nid < 0 || last_nid == nid) { 4641 last_nid = nid; 4642 last_end = end; 4643 continue; 4644 } 4645 4646 /* 4647 * Start with a mask granular enough to pin-point to the 4648 * start pfn and tick off bits one-by-one until it becomes 4649 * too coarse to separate the current node from the last. 4650 */ 4651 mask = ~((1 << __ffs(start)) - 1); 4652 while (mask && last_end <= (start & (mask << 1))) 4653 mask <<= 1; 4654 4655 /* accumulate all internode masks */ 4656 accl_mask |= mask; 4657 } 4658 4659 /* convert mask to number of pages */ 4660 return ~accl_mask + 1; 4661} 4662 4663/* Find the lowest pfn for a node */ 4664static unsigned long __init find_min_pfn_for_node(int nid) 4665{ 4666 int i; 4667 unsigned long min_pfn = ULONG_MAX; 4668 4669 /* Assuming a sorted map, the first range found has the starting pfn */ 4670 for_each_active_range_index_in_nid(i, nid) 4671 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4672 4673 if (min_pfn == ULONG_MAX) { 4674 printk(KERN_WARNING 4675 "Could not find start_pfn for node %d\n", nid); 4676 return 0; 4677 } 4678 4679 return min_pfn; 4680} 4681 4682/** 4683 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4684 * 4685 * It returns the minimum PFN based on information provided via 4686 * add_active_range(). 4687 */ 4688unsigned long __init find_min_pfn_with_active_regions(void) 4689{ 4690 return find_min_pfn_for_node(MAX_NUMNODES); 4691} 4692 4693/* 4694 * early_calculate_totalpages() 4695 * Sum pages in active regions for movable zone. 4696 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4697 */ 4698static unsigned long __init early_calculate_totalpages(void) 4699{ 4700 int i; 4701 unsigned long totalpages = 0; 4702 4703 for (i = 0; i < nr_nodemap_entries; i++) { 4704 unsigned long pages = early_node_map[i].end_pfn - 4705 early_node_map[i].start_pfn; 4706 totalpages += pages; 4707 if (pages) 4708 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4709 } 4710 return totalpages; 4711} 4712 4713/* 4714 * Find the PFN the Movable zone begins in each node. Kernel memory 4715 * is spread evenly between nodes as long as the nodes have enough 4716 * memory. When they don't, some nodes will have more kernelcore than 4717 * others 4718 */ 4719static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4720{ 4721 int i, nid; 4722 unsigned long usable_startpfn; 4723 unsigned long kernelcore_node, kernelcore_remaining; 4724 /* save the state before borrow the nodemask */ 4725 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4726 unsigned long totalpages = early_calculate_totalpages(); 4727 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4728 4729 /* 4730 * If movablecore was specified, calculate what size of 4731 * kernelcore that corresponds so that memory usable for 4732 * any allocation type is evenly spread. If both kernelcore 4733 * and movablecore are specified, then the value of kernelcore 4734 * will be used for required_kernelcore if it's greater than 4735 * what movablecore would have allowed. 4736 */ 4737 if (required_movablecore) { 4738 unsigned long corepages; 4739 4740 /* 4741 * Round-up so that ZONE_MOVABLE is at least as large as what 4742 * was requested by the user 4743 */ 4744 required_movablecore = 4745 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4746 corepages = totalpages - required_movablecore; 4747 4748 required_kernelcore = max(required_kernelcore, corepages); 4749 } 4750 4751 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4752 if (!required_kernelcore) 4753 goto out; 4754 4755 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4756 find_usable_zone_for_movable(); 4757 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4758 4759restart: 4760 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4761 kernelcore_node = required_kernelcore / usable_nodes; 4762 for_each_node_state(nid, N_HIGH_MEMORY) { 4763 /* 4764 * Recalculate kernelcore_node if the division per node 4765 * now exceeds what is necessary to satisfy the requested 4766 * amount of memory for the kernel 4767 */ 4768 if (required_kernelcore < kernelcore_node) 4769 kernelcore_node = required_kernelcore / usable_nodes; 4770 4771 /* 4772 * As the map is walked, we track how much memory is usable 4773 * by the kernel using kernelcore_remaining. When it is 4774 * 0, the rest of the node is usable by ZONE_MOVABLE 4775 */ 4776 kernelcore_remaining = kernelcore_node; 4777 4778 /* Go through each range of PFNs within this node */ 4779 for_each_active_range_index_in_nid(i, nid) { 4780 unsigned long start_pfn, end_pfn; 4781 unsigned long size_pages; 4782 4783 start_pfn = max(early_node_map[i].start_pfn, 4784 zone_movable_pfn[nid]); 4785 end_pfn = early_node_map[i].end_pfn; 4786 if (start_pfn >= end_pfn) 4787 continue; 4788 4789 /* Account for what is only usable for kernelcore */ 4790 if (start_pfn < usable_startpfn) { 4791 unsigned long kernel_pages; 4792 kernel_pages = min(end_pfn, usable_startpfn) 4793 - start_pfn; 4794 4795 kernelcore_remaining -= min(kernel_pages, 4796 kernelcore_remaining); 4797 required_kernelcore -= min(kernel_pages, 4798 required_kernelcore); 4799 4800 /* Continue if range is now fully accounted */ 4801 if (end_pfn <= usable_startpfn) { 4802 4803 /* 4804 * Push zone_movable_pfn to the end so 4805 * that if we have to rebalance 4806 * kernelcore across nodes, we will 4807 * not double account here 4808 */ 4809 zone_movable_pfn[nid] = end_pfn; 4810 continue; 4811 } 4812 start_pfn = usable_startpfn; 4813 } 4814 4815 /* 4816 * The usable PFN range for ZONE_MOVABLE is from 4817 * start_pfn->end_pfn. Calculate size_pages as the 4818 * number of pages used as kernelcore 4819 */ 4820 size_pages = end_pfn - start_pfn; 4821 if (size_pages > kernelcore_remaining) 4822 size_pages = kernelcore_remaining; 4823 zone_movable_pfn[nid] = start_pfn + size_pages; 4824 4825 /* 4826 * Some kernelcore has been met, update counts and 4827 * break if the kernelcore for this node has been 4828 * satisified 4829 */ 4830 required_kernelcore -= min(required_kernelcore, 4831 size_pages); 4832 kernelcore_remaining -= size_pages; 4833 if (!kernelcore_remaining) 4834 break; 4835 } 4836 } 4837 4838 /* 4839 * If there is still required_kernelcore, we do another pass with one 4840 * less node in the count. This will push zone_movable_pfn[nid] further 4841 * along on the nodes that still have memory until kernelcore is 4842 * satisified 4843 */ 4844 usable_nodes--; 4845 if (usable_nodes && required_kernelcore > usable_nodes) 4846 goto restart; 4847 4848 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4849 for (nid = 0; nid < MAX_NUMNODES; nid++) 4850 zone_movable_pfn[nid] = 4851 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4852 4853out: 4854 /* restore the node_state */ 4855 node_states[N_HIGH_MEMORY] = saved_node_state; 4856} 4857 4858/* Any regular memory on that node ? */ 4859static void check_for_regular_memory(pg_data_t *pgdat) 4860{ 4861#ifdef CONFIG_HIGHMEM 4862 enum zone_type zone_type; 4863 4864 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4865 struct zone *zone = &pgdat->node_zones[zone_type]; 4866 if (zone->present_pages) 4867 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4868 } 4869#endif 4870} 4871 4872/** 4873 * free_area_init_nodes - Initialise all pg_data_t and zone data 4874 * @max_zone_pfn: an array of max PFNs for each zone 4875 * 4876 * This will call free_area_init_node() for each active node in the system. 4877 * Using the page ranges provided by add_active_range(), the size of each 4878 * zone in each node and their holes is calculated. If the maximum PFN 4879 * between two adjacent zones match, it is assumed that the zone is empty. 4880 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4881 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4882 * starts where the previous one ended. For example, ZONE_DMA32 starts 4883 * at arch_max_dma_pfn. 4884 */ 4885void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4886{ 4887 unsigned long nid; 4888 int i; 4889 4890 /* Sort early_node_map as initialisation assumes it is sorted */ 4891 sort_node_map(); 4892 4893 /* Record where the zone boundaries are */ 4894 memset(arch_zone_lowest_possible_pfn, 0, 4895 sizeof(arch_zone_lowest_possible_pfn)); 4896 memset(arch_zone_highest_possible_pfn, 0, 4897 sizeof(arch_zone_highest_possible_pfn)); 4898 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4899 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4900 for (i = 1; i < MAX_NR_ZONES; i++) { 4901 if (i == ZONE_MOVABLE) 4902 continue; 4903 arch_zone_lowest_possible_pfn[i] = 4904 arch_zone_highest_possible_pfn[i-1]; 4905 arch_zone_highest_possible_pfn[i] = 4906 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4907 } 4908 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4909 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4910 4911 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4912 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4913 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4914 4915 /* Print out the zone ranges */ 4916 printk("Zone PFN ranges:\n"); 4917 for (i = 0; i < MAX_NR_ZONES; i++) { 4918 if (i == ZONE_MOVABLE) 4919 continue; 4920 printk(" %-8s ", zone_names[i]); 4921 if (arch_zone_lowest_possible_pfn[i] == 4922 arch_zone_highest_possible_pfn[i]) 4923 printk("empty\n"); 4924 else 4925 printk("%0#10lx -> %0#10lx\n", 4926 arch_zone_lowest_possible_pfn[i], 4927 arch_zone_highest_possible_pfn[i]); 4928 } 4929 4930 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4931 printk("Movable zone start PFN for each node\n"); 4932 for (i = 0; i < MAX_NUMNODES; i++) { 4933 if (zone_movable_pfn[i]) 4934 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4935 } 4936 4937 /* Print out the early_node_map[] */ 4938 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4939 for (i = 0; i < nr_nodemap_entries; i++) 4940 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4941 early_node_map[i].start_pfn, 4942 early_node_map[i].end_pfn); 4943 4944 /* Initialise every node */ 4945 mminit_verify_pageflags_layout(); 4946 setup_nr_node_ids(); 4947 for_each_online_node(nid) { 4948 pg_data_t *pgdat = NODE_DATA(nid); 4949 free_area_init_node(nid, NULL, 4950 find_min_pfn_for_node(nid), NULL); 4951 4952 /* Any memory on that node */ 4953 if (pgdat->node_present_pages) 4954 node_set_state(nid, N_HIGH_MEMORY); 4955 check_for_regular_memory(pgdat); 4956 } 4957} 4958 4959static int __init cmdline_parse_core(char *p, unsigned long *core) 4960{ 4961 unsigned long long coremem; 4962 if (!p) 4963 return -EINVAL; 4964 4965 coremem = memparse(p, &p); 4966 *core = coremem >> PAGE_SHIFT; 4967 4968 /* Paranoid check that UL is enough for the coremem value */ 4969 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4970 4971 return 0; 4972} 4973 4974/* 4975 * kernelcore=size sets the amount of memory for use for allocations that 4976 * cannot be reclaimed or migrated. 4977 */ 4978static int __init cmdline_parse_kernelcore(char *p) 4979{ 4980 return cmdline_parse_core(p, &required_kernelcore); 4981} 4982 4983/* 4984 * movablecore=size sets the amount of memory for use for allocations that 4985 * can be reclaimed or migrated. 4986 */ 4987static int __init cmdline_parse_movablecore(char *p) 4988{ 4989 return cmdline_parse_core(p, &required_movablecore); 4990} 4991 4992early_param("kernelcore", cmdline_parse_kernelcore); 4993early_param("movablecore", cmdline_parse_movablecore); 4994 4995#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4996 4997/** 4998 * set_dma_reserve - set the specified number of pages reserved in the first zone 4999 * @new_dma_reserve: The number of pages to mark reserved 5000 * 5001 * The per-cpu batchsize and zone watermarks are determined by present_pages. 5002 * In the DMA zone, a significant percentage may be consumed by kernel image 5003 * and other unfreeable allocations which can skew the watermarks badly. This 5004 * function may optionally be used to account for unfreeable pages in the 5005 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 5006 * smaller per-cpu batchsize. 5007 */ 5008void __init set_dma_reserve(unsigned long new_dma_reserve) 5009{ 5010 dma_reserve = new_dma_reserve; 5011} 5012 5013void __init free_area_init(unsigned long *zones_size) 5014{ 5015 free_area_init_node(0, zones_size, 5016 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 5017} 5018 5019static int page_alloc_cpu_notify(struct notifier_block *self, 5020 unsigned long action, void *hcpu) 5021{ 5022 int cpu = (unsigned long)hcpu; 5023 5024 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 5025 drain_pages(cpu); 5026 5027 /* 5028 * Spill the event counters of the dead processor 5029 * into the current processors event counters. 5030 * This artificially elevates the count of the current 5031 * processor. 5032 */ 5033 vm_events_fold_cpu(cpu); 5034 5035 /* 5036 * Zero the differential counters of the dead processor 5037 * so that the vm statistics are consistent. 5038 * 5039 * This is only okay since the processor is dead and cannot 5040 * race with what we are doing. 5041 */ 5042 refresh_cpu_vm_stats(cpu); 5043 } 5044 return NOTIFY_OK; 5045} 5046 5047void __init page_alloc_init(void) 5048{ 5049 hotcpu_notifier(page_alloc_cpu_notify, 0); 5050} 5051 5052/* 5053 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 5054 * or min_free_kbytes changes. 5055 */ 5056static void calculate_totalreserve_pages(void) 5057{ 5058 struct pglist_data *pgdat; 5059 unsigned long reserve_pages = 0; 5060 enum zone_type i, j; 5061 5062 for_each_online_pgdat(pgdat) { 5063 for (i = 0; i < MAX_NR_ZONES; i++) { 5064 struct zone *zone = pgdat->node_zones + i; 5065 unsigned long max = 0; 5066 5067 /* Find valid and maximum lowmem_reserve in the zone */ 5068 for (j = i; j < MAX_NR_ZONES; j++) { 5069 if (zone->lowmem_reserve[j] > max) 5070 max = zone->lowmem_reserve[j]; 5071 } 5072 5073 /* we treat the high watermark as reserved pages. */ 5074 max += high_wmark_pages(zone); 5075 5076 if (max > zone->present_pages) 5077 max = zone->present_pages; 5078 reserve_pages += max; 5079 } 5080 } 5081 totalreserve_pages = reserve_pages; 5082} 5083 5084/* 5085 * setup_per_zone_lowmem_reserve - called whenever 5086 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 5087 * has a correct pages reserved value, so an adequate number of 5088 * pages are left in the zone after a successful __alloc_pages(). 5089 */ 5090static void setup_per_zone_lowmem_reserve(void) 5091{ 5092 struct pglist_data *pgdat; 5093 enum zone_type j, idx; 5094 5095 for_each_online_pgdat(pgdat) { 5096 for (j = 0; j < MAX_NR_ZONES; j++) { 5097 struct zone *zone = pgdat->node_zones + j; 5098 unsigned long present_pages = zone->present_pages; 5099 5100 zone->lowmem_reserve[j] = 0; 5101 5102 idx = j; 5103 while (idx) { 5104 struct zone *lower_zone; 5105 5106 idx--; 5107 5108 if (sysctl_lowmem_reserve_ratio[idx] < 1) 5109 sysctl_lowmem_reserve_ratio[idx] = 1; 5110 5111 lower_zone = pgdat->node_zones + idx; 5112 lower_zone->lowmem_reserve[j] = present_pages / 5113 sysctl_lowmem_reserve_ratio[idx]; 5114 present_pages += lower_zone->present_pages; 5115 } 5116 } 5117 } 5118 5119 /* update totalreserve_pages */ 5120 calculate_totalreserve_pages(); 5121} 5122 5123/** 5124 * setup_per_zone_wmarks - called when min_free_kbytes changes 5125 * or when memory is hot-{added|removed} 5126 * 5127 * Ensures that the watermark[min,low,high] values for each zone are set 5128 * correctly with respect to min_free_kbytes. 5129 */ 5130void setup_per_zone_wmarks(void) 5131{ 5132 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 5133 unsigned long lowmem_pages = 0; 5134 struct zone *zone; 5135 unsigned long flags; 5136 5137 /* Calculate total number of !ZONE_HIGHMEM pages */ 5138 for_each_zone(zone) { 5139 if (!is_highmem(zone)) 5140 lowmem_pages += zone->present_pages; 5141 } 5142 5143 for_each_zone(zone) { 5144 u64 tmp; 5145 5146 spin_lock_irqsave(&zone->lock, flags); 5147 tmp = (u64)pages_min * zone->present_pages; 5148 do_div(tmp, lowmem_pages); 5149 if (is_highmem(zone)) { 5150 /* 5151 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 5152 * need highmem pages, so cap pages_min to a small 5153 * value here. 5154 * 5155 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 5156 * deltas controls asynch page reclaim, and so should 5157 * not be capped for highmem. 5158 */ 5159 int min_pages; 5160 5161 min_pages = zone->present_pages / 1024; 5162 if (min_pages < SWAP_CLUSTER_MAX) 5163 min_pages = SWAP_CLUSTER_MAX; 5164 if (min_pages > 128) 5165 min_pages = 128; 5166 zone->watermark[WMARK_MIN] = min_pages; 5167 } else { 5168 /* 5169 * If it's a lowmem zone, reserve a number of pages 5170 * proportionate to the zone's size. 5171 */ 5172 zone->watermark[WMARK_MIN] = tmp; 5173 } 5174 5175 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 5176 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 5177 setup_zone_migrate_reserve(zone); 5178 spin_unlock_irqrestore(&zone->lock, flags); 5179 } 5180 5181 /* update totalreserve_pages */ 5182 calculate_totalreserve_pages(); 5183} 5184 5185/* 5186 * The inactive anon list should be small enough that the VM never has to 5187 * do too much work, but large enough that each inactive page has a chance 5188 * to be referenced again before it is swapped out. 5189 * 5190 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 5191 * INACTIVE_ANON pages on this zone's LRU, maintained by the 5192 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 5193 * the anonymous pages are kept on the inactive list. 5194 * 5195 * total target max 5196 * memory ratio inactive anon 5197 * ------------------------------------- 5198 * 10MB 1 5MB 5199 * 100MB 1 50MB 5200 * 1GB 3 250MB 5201 * 10GB 10 0.9GB 5202 * 100GB 31 3GB 5203 * 1TB 101 10GB 5204 * 10TB 320 32GB 5205 */ 5206static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 5207{ 5208 unsigned int gb, ratio; 5209 5210 /* Zone size in gigabytes */ 5211 gb = zone->present_pages >> (30 - PAGE_SHIFT); 5212 if (gb) 5213 ratio = int_sqrt(10 * gb); 5214 else 5215 ratio = 1; 5216 5217 zone->inactive_ratio = ratio; 5218} 5219 5220static void __meminit setup_per_zone_inactive_ratio(void) 5221{ 5222 struct zone *zone; 5223 5224 for_each_zone(zone) 5225 calculate_zone_inactive_ratio(zone); 5226} 5227 5228/* 5229 * Initialise min_free_kbytes. 5230 * 5231 * For small machines we want it small (128k min). For large machines 5232 * we want it large (64MB max). But it is not linear, because network 5233 * bandwidth does not increase linearly with machine size. We use 5234 * 5235 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5236 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5237 * 5238 * which yields 5239 * 5240 * 16MB: 512k 5241 * 32MB: 724k 5242 * 64MB: 1024k 5243 * 128MB: 1448k 5244 * 256MB: 2048k 5245 * 512MB: 2896k 5246 * 1024MB: 4096k 5247 * 2048MB: 5792k 5248 * 4096MB: 8192k 5249 * 8192MB: 11584k 5250 * 16384MB: 16384k 5251 */ 5252int __meminit init_per_zone_wmark_min(void) 5253{ 5254 unsigned long lowmem_kbytes; 5255 5256 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5257 5258 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5259 if (min_free_kbytes < 128) 5260 min_free_kbytes = 128; 5261 if (min_free_kbytes > 65536) 5262 min_free_kbytes = 65536; 5263 setup_per_zone_wmarks(); 5264 refresh_zone_stat_thresholds(); 5265 setup_per_zone_lowmem_reserve(); 5266 setup_per_zone_inactive_ratio(); 5267 return 0; 5268} 5269module_init(init_per_zone_wmark_min) 5270 5271/* 5272 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5273 * that we can call two helper functions whenever min_free_kbytes 5274 * changes. 5275 */ 5276int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5277 void __user *buffer, size_t *length, loff_t *ppos) 5278{ 5279 proc_dointvec(table, write, buffer, length, ppos); 5280 if (write) 5281 setup_per_zone_wmarks(); 5282 return 0; 5283} 5284 5285#ifdef CONFIG_NUMA 5286int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5287 void __user *buffer, size_t *length, loff_t *ppos) 5288{ 5289 struct zone *zone; 5290 int rc; 5291 5292 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5293 if (rc) 5294 return rc; 5295 5296 for_each_zone(zone) 5297 zone->min_unmapped_pages = (zone->present_pages * 5298 sysctl_min_unmapped_ratio) / 100; 5299 return 0; 5300} 5301 5302int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5303 void __user *buffer, size_t *length, loff_t *ppos) 5304{ 5305 struct zone *zone; 5306 int rc; 5307 5308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5309 if (rc) 5310 return rc; 5311 5312 for_each_zone(zone) 5313 zone->min_slab_pages = (zone->present_pages * 5314 sysctl_min_slab_ratio) / 100; 5315 return 0; 5316} 5317#endif 5318 5319/* 5320 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5321 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5322 * whenever sysctl_lowmem_reserve_ratio changes. 5323 * 5324 * The reserve ratio obviously has absolutely no relation with the 5325 * minimum watermarks. The lowmem reserve ratio can only make sense 5326 * if in function of the boot time zone sizes. 5327 */ 5328int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5329 void __user *buffer, size_t *length, loff_t *ppos) 5330{ 5331 proc_dointvec_minmax(table, write, buffer, length, ppos); 5332 setup_per_zone_lowmem_reserve(); 5333 return 0; 5334} 5335 5336/* 5337 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5338 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5339 * can have before it gets flushed back to buddy allocator. 5340 */ 5341 5342int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5343 void __user *buffer, size_t *length, loff_t *ppos) 5344{ 5345 struct zone *zone; 5346 unsigned int cpu; 5347 int ret; 5348 5349 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5350 if (!write || (ret == -EINVAL)) 5351 return ret; 5352 for_each_populated_zone(zone) { 5353 for_each_possible_cpu(cpu) { 5354 unsigned long high; 5355 high = zone->present_pages / percpu_pagelist_fraction; 5356 setup_pagelist_highmark( 5357 per_cpu_ptr(zone->pageset, cpu), high); 5358 } 5359 } 5360 return 0; 5361} 5362 5363int hashdist = HASHDIST_DEFAULT; 5364 5365#ifdef CONFIG_NUMA 5366static int __init set_hashdist(char *str) 5367{ 5368 if (!str) 5369 return 0; 5370 hashdist = simple_strtoul(str, &str, 0); 5371 return 1; 5372} 5373__setup("hashdist=", set_hashdist); 5374#endif 5375 5376/* 5377 * allocate a large system hash table from bootmem 5378 * - it is assumed that the hash table must contain an exact power-of-2 5379 * quantity of entries 5380 * - limit is the number of hash buckets, not the total allocation size 5381 */ 5382void *__init alloc_large_system_hash(const char *tablename, 5383 unsigned long bucketsize, 5384 unsigned long numentries, 5385 int scale, 5386 int flags, 5387 unsigned int *_hash_shift, 5388 unsigned int *_hash_mask, 5389 unsigned long limit) 5390{ 5391 unsigned long long max = limit; 5392 unsigned long log2qty, size; 5393 void *table = NULL; 5394 5395 /* allow the kernel cmdline to have a say */ 5396 if (!numentries) { 5397 /* round applicable memory size up to nearest megabyte */ 5398 numentries = nr_kernel_pages; 5399 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5400 numentries >>= 20 - PAGE_SHIFT; 5401 numentries <<= 20 - PAGE_SHIFT; 5402 5403 /* limit to 1 bucket per 2^scale bytes of low memory */ 5404 if (scale > PAGE_SHIFT) 5405 numentries >>= (scale - PAGE_SHIFT); 5406 else 5407 numentries <<= (PAGE_SHIFT - scale); 5408 5409 /* Make sure we've got at least a 0-order allocation.. */ 5410 if (unlikely(flags & HASH_SMALL)) { 5411 /* Makes no sense without HASH_EARLY */ 5412 WARN_ON(!(flags & HASH_EARLY)); 5413 if (!(numentries >> *_hash_shift)) { 5414 numentries = 1UL << *_hash_shift; 5415 BUG_ON(!numentries); 5416 } 5417 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5418 numentries = PAGE_SIZE / bucketsize; 5419 } 5420 numentries = roundup_pow_of_two(numentries); 5421 5422 /* limit allocation size to 1/16 total memory by default */ 5423 if (max == 0) { 5424 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5425 do_div(max, bucketsize); 5426 } 5427 5428 if (numentries > max) 5429 numentries = max; 5430 5431 log2qty = ilog2(numentries); 5432 5433 do { 5434 size = bucketsize << log2qty; 5435 if (flags & HASH_EARLY) 5436 table = alloc_bootmem_nopanic(size); 5437 else if (hashdist) 5438 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5439 else { 5440 /* 5441 * If bucketsize is not a power-of-two, we may free 5442 * some pages at the end of hash table which 5443 * alloc_pages_exact() automatically does 5444 */ 5445 if (get_order(size) < MAX_ORDER) { 5446 table = alloc_pages_exact(size, GFP_ATOMIC); 5447 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5448 } 5449 } 5450 } while (!table && size > PAGE_SIZE && --log2qty); 5451 5452 if (!table) 5453 panic("Failed to allocate %s hash table\n", tablename); 5454 5455 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5456 tablename, 5457 (1UL << log2qty), 5458 ilog2(size) - PAGE_SHIFT, 5459 size); 5460 5461 if (_hash_shift) 5462 *_hash_shift = log2qty; 5463 if (_hash_mask) 5464 *_hash_mask = (1 << log2qty) - 1; 5465 5466 return table; 5467} 5468 5469/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5470static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5471 unsigned long pfn) 5472{ 5473#ifdef CONFIG_SPARSEMEM 5474 return __pfn_to_section(pfn)->pageblock_flags; 5475#else 5476 return zone->pageblock_flags; 5477#endif /* CONFIG_SPARSEMEM */ 5478} 5479 5480static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5481{ 5482#ifdef CONFIG_SPARSEMEM 5483 pfn &= (PAGES_PER_SECTION-1); 5484 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5485#else 5486 pfn = pfn - zone->zone_start_pfn; 5487 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5488#endif /* CONFIG_SPARSEMEM */ 5489} 5490 5491/** 5492 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5493 * @page: The page within the block of interest 5494 * @start_bitidx: The first bit of interest to retrieve 5495 * @end_bitidx: The last bit of interest 5496 * returns pageblock_bits flags 5497 */ 5498unsigned long get_pageblock_flags_group(struct page *page, 5499 int start_bitidx, int end_bitidx) 5500{ 5501 struct zone *zone; 5502 unsigned long *bitmap; 5503 unsigned long pfn, bitidx; 5504 unsigned long flags = 0; 5505 unsigned long value = 1; 5506 5507 zone = page_zone(page); 5508 pfn = page_to_pfn(page); 5509 bitmap = get_pageblock_bitmap(zone, pfn); 5510 bitidx = pfn_to_bitidx(zone, pfn); 5511 5512 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5513 if (test_bit(bitidx + start_bitidx, bitmap)) 5514 flags |= value; 5515 5516 return flags; 5517} 5518 5519/** 5520 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5521 * @page: The page within the block of interest 5522 * @start_bitidx: The first bit of interest 5523 * @end_bitidx: The last bit of interest 5524 * @flags: The flags to set 5525 */ 5526void set_pageblock_flags_group(struct page *page, unsigned long flags, 5527 int start_bitidx, int end_bitidx) 5528{ 5529 struct zone *zone; 5530 unsigned long *bitmap; 5531 unsigned long pfn, bitidx; 5532 unsigned long value = 1; 5533 5534 zone = page_zone(page); 5535 pfn = page_to_pfn(page); 5536 bitmap = get_pageblock_bitmap(zone, pfn); 5537 bitidx = pfn_to_bitidx(zone, pfn); 5538 VM_BUG_ON(pfn < zone->zone_start_pfn); 5539 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5540 5541 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5542 if (flags & value) 5543 __set_bit(bitidx + start_bitidx, bitmap); 5544 else 5545 __clear_bit(bitidx + start_bitidx, bitmap); 5546} 5547 5548/* 5549 * This is designed as sub function...plz see page_isolation.c also. 5550 * set/clear page block's type to be ISOLATE. 5551 * page allocater never alloc memory from ISOLATE block. 5552 */ 5553 5554static int 5555__count_immobile_pages(struct zone *zone, struct page *page, int count) 5556{ 5557 unsigned long pfn, iter, found; 5558 /* 5559 * For avoiding noise data, lru_add_drain_all() should be called 5560 * If ZONE_MOVABLE, the zone never contains immobile pages 5561 */ 5562 if (zone_idx(zone) == ZONE_MOVABLE) 5563 return true; 5564 5565 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE) 5566 return true; 5567 5568 pfn = page_to_pfn(page); 5569 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 5570 unsigned long check = pfn + iter; 5571 5572 if (!pfn_valid_within(check)) 5573 continue; 5574 5575 page = pfn_to_page(check); 5576 if (!page_count(page)) { 5577 if (PageBuddy(page)) 5578 iter += (1 << page_order(page)) - 1; 5579 continue; 5580 } 5581 if (!PageLRU(page)) 5582 found++; 5583 /* 5584 * If there are RECLAIMABLE pages, we need to check it. 5585 * But now, memory offline itself doesn't call shrink_slab() 5586 * and it still to be fixed. 5587 */ 5588 /* 5589 * If the page is not RAM, page_count()should be 0. 5590 * we don't need more check. This is an _used_ not-movable page. 5591 * 5592 * The problematic thing here is PG_reserved pages. PG_reserved 5593 * is set to both of a memory hole page and a _used_ kernel 5594 * page at boot. 5595 */ 5596 if (found > count) 5597 return false; 5598 } 5599 return true; 5600} 5601 5602bool is_pageblock_removable_nolock(struct page *page) 5603{ 5604 struct zone *zone = page_zone(page); 5605 return __count_immobile_pages(zone, page, 0); 5606} 5607 5608int set_migratetype_isolate(struct page *page) 5609{ 5610 struct zone *zone; 5611 unsigned long flags, pfn; 5612 struct memory_isolate_notify arg; 5613 int notifier_ret; 5614 int ret = -EBUSY; 5615 5616 zone = page_zone(page); 5617 5618 spin_lock_irqsave(&zone->lock, flags); 5619 5620 pfn = page_to_pfn(page); 5621 arg.start_pfn = pfn; 5622 arg.nr_pages = pageblock_nr_pages; 5623 arg.pages_found = 0; 5624 5625 /* 5626 * It may be possible to isolate a pageblock even if the 5627 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5628 * notifier chain is used by balloon drivers to return the 5629 * number of pages in a range that are held by the balloon 5630 * driver to shrink memory. If all the pages are accounted for 5631 * by balloons, are free, or on the LRU, isolation can continue. 5632 * Later, for example, when memory hotplug notifier runs, these 5633 * pages reported as "can be isolated" should be isolated(freed) 5634 * by the balloon driver through the memory notifier chain. 5635 */ 5636 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5637 notifier_ret = notifier_to_errno(notifier_ret); 5638 if (notifier_ret) 5639 goto out; 5640 /* 5641 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. 5642 * We just check MOVABLE pages. 5643 */ 5644 if (__count_immobile_pages(zone, page, arg.pages_found)) 5645 ret = 0; 5646 5647 /* 5648 * immobile means "not-on-lru" paes. If immobile is larger than 5649 * removable-by-driver pages reported by notifier, we'll fail. 5650 */ 5651 5652out: 5653 if (!ret) { 5654 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5655 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5656 } 5657 5658 spin_unlock_irqrestore(&zone->lock, flags); 5659 if (!ret) 5660 drain_all_pages(); 5661 return ret; 5662} 5663 5664void unset_migratetype_isolate(struct page *page) 5665{ 5666 struct zone *zone; 5667 unsigned long flags; 5668 zone = page_zone(page); 5669 spin_lock_irqsave(&zone->lock, flags); 5670 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5671 goto out; 5672 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5673 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5674out: 5675 spin_unlock_irqrestore(&zone->lock, flags); 5676} 5677 5678#ifdef CONFIG_MEMORY_HOTREMOVE 5679/* 5680 * All pages in the range must be isolated before calling this. 5681 */ 5682void 5683__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5684{ 5685 struct page *page; 5686 struct zone *zone; 5687 int order, i; 5688 unsigned long pfn; 5689 unsigned long flags; 5690 /* find the first valid pfn */ 5691 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5692 if (pfn_valid(pfn)) 5693 break; 5694 if (pfn == end_pfn) 5695 return; 5696 zone = page_zone(pfn_to_page(pfn)); 5697 spin_lock_irqsave(&zone->lock, flags); 5698 pfn = start_pfn; 5699 while (pfn < end_pfn) { 5700 if (!pfn_valid(pfn)) { 5701 pfn++; 5702 continue; 5703 } 5704 page = pfn_to_page(pfn); 5705 BUG_ON(page_count(page)); 5706 BUG_ON(!PageBuddy(page)); 5707 order = page_order(page); 5708#ifdef CONFIG_DEBUG_VM 5709 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5710 pfn, 1 << order, end_pfn); 5711#endif 5712 list_del(&page->lru); 5713 rmv_page_order(page); 5714 zone->free_area[order].nr_free--; 5715 __mod_zone_page_state(zone, NR_FREE_PAGES, 5716 - (1UL << order)); 5717 for (i = 0; i < (1 << order); i++) 5718 SetPageReserved((page+i)); 5719 pfn += (1 << order); 5720 } 5721 spin_unlock_irqrestore(&zone->lock, flags); 5722} 5723#endif 5724 5725#ifdef CONFIG_MEMORY_FAILURE 5726bool is_free_buddy_page(struct page *page) 5727{ 5728 struct zone *zone = page_zone(page); 5729 unsigned long pfn = page_to_pfn(page); 5730 unsigned long flags; 5731 int order; 5732 5733 spin_lock_irqsave(&zone->lock, flags); 5734 for (order = 0; order < MAX_ORDER; order++) { 5735 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5736 5737 if (PageBuddy(page_head) && page_order(page_head) >= order) 5738 break; 5739 } 5740 spin_unlock_irqrestore(&zone->lock, flags); 5741 5742 return order < MAX_ORDER; 5743} 5744#endif 5745 5746static struct trace_print_flags pageflag_names[] = { 5747 {1UL << PG_locked, "locked" }, 5748 {1UL << PG_error, "error" }, 5749 {1UL << PG_referenced, "referenced" }, 5750 {1UL << PG_uptodate, "uptodate" }, 5751 {1UL << PG_dirty, "dirty" }, 5752 {1UL << PG_lru, "lru" }, 5753 {1UL << PG_active, "active" }, 5754 {1UL << PG_slab, "slab" }, 5755 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5756 {1UL << PG_arch_1, "arch_1" }, 5757 {1UL << PG_reserved, "reserved" }, 5758 {1UL << PG_private, "private" }, 5759 {1UL << PG_private_2, "private_2" }, 5760 {1UL << PG_writeback, "writeback" }, 5761#ifdef CONFIG_PAGEFLAGS_EXTENDED 5762 {1UL << PG_head, "head" }, 5763 {1UL << PG_tail, "tail" }, 5764#else 5765 {1UL << PG_compound, "compound" }, 5766#endif 5767 {1UL << PG_swapcache, "swapcache" }, 5768 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5769 {1UL << PG_reclaim, "reclaim" }, 5770 {1UL << PG_swapbacked, "swapbacked" }, 5771 {1UL << PG_unevictable, "unevictable" }, 5772#ifdef CONFIG_MMU 5773 {1UL << PG_mlocked, "mlocked" }, 5774#endif 5775#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5776 {1UL << PG_uncached, "uncached" }, 5777#endif 5778#ifdef CONFIG_MEMORY_FAILURE 5779 {1UL << PG_hwpoison, "hwpoison" }, 5780#endif 5781 {-1UL, NULL }, 5782}; 5783 5784static void dump_page_flags(unsigned long flags) 5785{ 5786 const char *delim = ""; 5787 unsigned long mask; 5788 int i; 5789 5790 printk(KERN_ALERT "page flags: %#lx(", flags); 5791 5792 /* remove zone id */ 5793 flags &= (1UL << NR_PAGEFLAGS) - 1; 5794 5795 for (i = 0; pageflag_names[i].name && flags; i++) { 5796 5797 mask = pageflag_names[i].mask; 5798 if ((flags & mask) != mask) 5799 continue; 5800 5801 flags &= ~mask; 5802 printk("%s%s", delim, pageflag_names[i].name); 5803 delim = "|"; 5804 } 5805 5806 /* check for left over flags */ 5807 if (flags) 5808 printk("%s%#lx", delim, flags); 5809 5810 printk(")\n"); 5811} 5812 5813void dump_page(struct page *page) 5814{ 5815 printk(KERN_ALERT 5816 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5817 page, atomic_read(&page->_count), page_mapcount(page), 5818 page->mapping, page->index); 5819 dump_page_flags(page->flags); 5820 mem_cgroup_print_bad_page(page); 5821} 5822