page_alloc.c revision b2588c4b4c3c075e9b45d61065d86c60de2b6441
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 dump_stack(); 322out: 323 /* Leave bad fields for debug, except PageBuddy could make trouble */ 324 reset_page_mapcount(page); /* remove PageBuddy */ 325 add_taint(TAINT_BAD_PAGE); 326} 327 328/* 329 * Higher-order pages are called "compound pages". They are structured thusly: 330 * 331 * The first PAGE_SIZE page is called the "head page". 332 * 333 * The remaining PAGE_SIZE pages are called "tail pages". 334 * 335 * All pages have PG_compound set. All pages have their ->private pointing at 336 * the head page (even the head page has this). 337 * 338 * The first tail page's ->lru.next holds the address of the compound page's 339 * put_page() function. Its ->lru.prev holds the order of allocation. 340 * This usage means that zero-order pages may not be compound. 341 */ 342 343static void free_compound_page(struct page *page) 344{ 345 __free_pages_ok(page, compound_order(page)); 346} 347 348void prep_compound_page(struct page *page, unsigned long order) 349{ 350 int i; 351 int nr_pages = 1 << order; 352 353 set_compound_page_dtor(page, free_compound_page); 354 set_compound_order(page, order); 355 __SetPageHead(page); 356 for (i = 1; i < nr_pages; i++) { 357 struct page *p = page + i; 358 359 __SetPageTail(p); 360 p->first_page = page; 361 } 362} 363 364/* update __split_huge_page_refcount if you change this function */ 365static int destroy_compound_page(struct page *page, unsigned long order) 366{ 367 int i; 368 int nr_pages = 1 << order; 369 int bad = 0; 370 371 if (unlikely(compound_order(page) != order) || 372 unlikely(!PageHead(page))) { 373 bad_page(page); 374 bad++; 375 } 376 377 __ClearPageHead(page); 378 379 for (i = 1; i < nr_pages; i++) { 380 struct page *p = page + i; 381 382 if (unlikely(!PageTail(p) || (p->first_page != page))) { 383 bad_page(page); 384 bad++; 385 } 386 __ClearPageTail(p); 387 } 388 389 return bad; 390} 391 392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 393{ 394 int i; 395 396 /* 397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 398 * and __GFP_HIGHMEM from hard or soft interrupt context. 399 */ 400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 401 for (i = 0; i < (1 << order); i++) 402 clear_highpage(page + i); 403} 404 405static inline void set_page_order(struct page *page, int order) 406{ 407 set_page_private(page, order); 408 __SetPageBuddy(page); 409} 410 411static inline void rmv_page_order(struct page *page) 412{ 413 __ClearPageBuddy(page); 414 set_page_private(page, 0); 415} 416 417/* 418 * Locate the struct page for both the matching buddy in our 419 * pair (buddy1) and the combined O(n+1) page they form (page). 420 * 421 * 1) Any buddy B1 will have an order O twin B2 which satisfies 422 * the following equation: 423 * B2 = B1 ^ (1 << O) 424 * For example, if the starting buddy (buddy2) is #8 its order 425 * 1 buddy is #10: 426 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 427 * 428 * 2) Any buddy B will have an order O+1 parent P which 429 * satisfies the following equation: 430 * P = B & ~(1 << O) 431 * 432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 433 */ 434static inline unsigned long 435__find_buddy_index(unsigned long page_idx, unsigned int order) 436{ 437 return page_idx ^ (1 << order); 438} 439 440/* 441 * This function checks whether a page is free && is the buddy 442 * we can do coalesce a page and its buddy if 443 * (a) the buddy is not in a hole && 444 * (b) the buddy is in the buddy system && 445 * (c) a page and its buddy have the same order && 446 * (d) a page and its buddy are in the same zone. 447 * 448 * For recording whether a page is in the buddy system, we set ->_mapcount -2. 449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. 450 * 451 * For recording page's order, we use page_private(page). 452 */ 453static inline int page_is_buddy(struct page *page, struct page *buddy, 454 int order) 455{ 456 if (!pfn_valid_within(page_to_pfn(buddy))) 457 return 0; 458 459 if (page_zone_id(page) != page_zone_id(buddy)) 460 return 0; 461 462 if (PageBuddy(buddy) && page_order(buddy) == order) { 463 VM_BUG_ON(page_count(buddy) != 0); 464 return 1; 465 } 466 return 0; 467} 468 469/* 470 * Freeing function for a buddy system allocator. 471 * 472 * The concept of a buddy system is to maintain direct-mapped table 473 * (containing bit values) for memory blocks of various "orders". 474 * The bottom level table contains the map for the smallest allocatable 475 * units of memory (here, pages), and each level above it describes 476 * pairs of units from the levels below, hence, "buddies". 477 * At a high level, all that happens here is marking the table entry 478 * at the bottom level available, and propagating the changes upward 479 * as necessary, plus some accounting needed to play nicely with other 480 * parts of the VM system. 481 * At each level, we keep a list of pages, which are heads of continuous 482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's 483 * order is recorded in page_private(page) field. 484 * So when we are allocating or freeing one, we can derive the state of the 485 * other. That is, if we allocate a small block, and both were 486 * free, the remainder of the region must be split into blocks. 487 * If a block is freed, and its buddy is also free, then this 488 * triggers coalescing into a block of larger size. 489 * 490 * -- wli 491 */ 492 493static inline void __free_one_page(struct page *page, 494 struct zone *zone, unsigned int order, 495 int migratetype) 496{ 497 unsigned long page_idx; 498 unsigned long combined_idx; 499 unsigned long uninitialized_var(buddy_idx); 500 struct page *buddy; 501 502 if (unlikely(PageCompound(page))) 503 if (unlikely(destroy_compound_page(page, order))) 504 return; 505 506 VM_BUG_ON(migratetype == -1); 507 508 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 509 510 VM_BUG_ON(page_idx & ((1 << order) - 1)); 511 VM_BUG_ON(bad_range(zone, page)); 512 513 while (order < MAX_ORDER-1) { 514 buddy_idx = __find_buddy_index(page_idx, order); 515 buddy = page + (buddy_idx - page_idx); 516 if (!page_is_buddy(page, buddy, order)) 517 break; 518 519 /* Our buddy is free, merge with it and move up one order. */ 520 list_del(&buddy->lru); 521 zone->free_area[order].nr_free--; 522 rmv_page_order(buddy); 523 combined_idx = buddy_idx & page_idx; 524 page = page + (combined_idx - page_idx); 525 page_idx = combined_idx; 526 order++; 527 } 528 set_page_order(page, order); 529 530 /* 531 * If this is not the largest possible page, check if the buddy 532 * of the next-highest order is free. If it is, it's possible 533 * that pages are being freed that will coalesce soon. In case, 534 * that is happening, add the free page to the tail of the list 535 * so it's less likely to be used soon and more likely to be merged 536 * as a higher order page 537 */ 538 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 539 struct page *higher_page, *higher_buddy; 540 combined_idx = buddy_idx & page_idx; 541 higher_page = page + (combined_idx - page_idx); 542 buddy_idx = __find_buddy_index(combined_idx, order + 1); 543 higher_buddy = page + (buddy_idx - combined_idx); 544 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 545 list_add_tail(&page->lru, 546 &zone->free_area[order].free_list[migratetype]); 547 goto out; 548 } 549 } 550 551 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 552out: 553 zone->free_area[order].nr_free++; 554} 555 556/* 557 * free_page_mlock() -- clean up attempts to free and mlocked() page. 558 * Page should not be on lru, so no need to fix that up. 559 * free_pages_check() will verify... 560 */ 561static inline void free_page_mlock(struct page *page) 562{ 563 __dec_zone_page_state(page, NR_MLOCK); 564 __count_vm_event(UNEVICTABLE_MLOCKFREED); 565} 566 567static inline int free_pages_check(struct page *page) 568{ 569 if (unlikely(page_mapcount(page) | 570 (page->mapping != NULL) | 571 (atomic_read(&page->_count) != 0) | 572 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | 573 (mem_cgroup_bad_page_check(page)))) { 574 bad_page(page); 575 return 1; 576 } 577 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 578 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 579 return 0; 580} 581 582/* 583 * Frees a number of pages from the PCP lists 584 * Assumes all pages on list are in same zone, and of same order. 585 * count is the number of pages to free. 586 * 587 * If the zone was previously in an "all pages pinned" state then look to 588 * see if this freeing clears that state. 589 * 590 * And clear the zone's pages_scanned counter, to hold off the "all pages are 591 * pinned" detection logic. 592 */ 593static void free_pcppages_bulk(struct zone *zone, int count, 594 struct per_cpu_pages *pcp) 595{ 596 int migratetype = 0; 597 int batch_free = 0; 598 int to_free = count; 599 600 spin_lock(&zone->lock); 601 zone->all_unreclaimable = 0; 602 zone->pages_scanned = 0; 603 604 while (to_free) { 605 struct page *page; 606 struct list_head *list; 607 608 /* 609 * Remove pages from lists in a round-robin fashion. A 610 * batch_free count is maintained that is incremented when an 611 * empty list is encountered. This is so more pages are freed 612 * off fuller lists instead of spinning excessively around empty 613 * lists 614 */ 615 do { 616 batch_free++; 617 if (++migratetype == MIGRATE_PCPTYPES) 618 migratetype = 0; 619 list = &pcp->lists[migratetype]; 620 } while (list_empty(list)); 621 622 /* This is the only non-empty list. Free them all. */ 623 if (batch_free == MIGRATE_PCPTYPES) 624 batch_free = to_free; 625 626 do { 627 page = list_entry(list->prev, struct page, lru); 628 /* must delete as __free_one_page list manipulates */ 629 list_del(&page->lru); 630 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 631 __free_one_page(page, zone, 0, page_private(page)); 632 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 633 } while (--to_free && --batch_free && !list_empty(list)); 634 } 635 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 636 spin_unlock(&zone->lock); 637} 638 639static void free_one_page(struct zone *zone, struct page *page, int order, 640 int migratetype) 641{ 642 spin_lock(&zone->lock); 643 zone->all_unreclaimable = 0; 644 zone->pages_scanned = 0; 645 646 __free_one_page(page, zone, order, migratetype); 647 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 648 spin_unlock(&zone->lock); 649} 650 651static bool free_pages_prepare(struct page *page, unsigned int order) 652{ 653 int i; 654 int bad = 0; 655 656 trace_mm_page_free_direct(page, order); 657 kmemcheck_free_shadow(page, order); 658 659 if (PageAnon(page)) 660 page->mapping = NULL; 661 for (i = 0; i < (1 << order); i++) 662 bad += free_pages_check(page + i); 663 if (bad) 664 return false; 665 666 if (!PageHighMem(page)) { 667 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 668 debug_check_no_obj_freed(page_address(page), 669 PAGE_SIZE << order); 670 } 671 arch_free_page(page, order); 672 kernel_map_pages(page, 1 << order, 0); 673 674 return true; 675} 676 677static void __free_pages_ok(struct page *page, unsigned int order) 678{ 679 unsigned long flags; 680 int wasMlocked = __TestClearPageMlocked(page); 681 682 if (!free_pages_prepare(page, order)) 683 return; 684 685 local_irq_save(flags); 686 if (unlikely(wasMlocked)) 687 free_page_mlock(page); 688 __count_vm_events(PGFREE, 1 << order); 689 free_one_page(page_zone(page), page, order, 690 get_pageblock_migratetype(page)); 691 local_irq_restore(flags); 692} 693 694/* 695 * permit the bootmem allocator to evade page validation on high-order frees 696 */ 697void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 698{ 699 if (order == 0) { 700 __ClearPageReserved(page); 701 set_page_count(page, 0); 702 set_page_refcounted(page); 703 __free_page(page); 704 } else { 705 int loop; 706 707 prefetchw(page); 708 for (loop = 0; loop < BITS_PER_LONG; loop++) { 709 struct page *p = &page[loop]; 710 711 if (loop + 1 < BITS_PER_LONG) 712 prefetchw(p + 1); 713 __ClearPageReserved(p); 714 set_page_count(p, 0); 715 } 716 717 set_page_refcounted(page); 718 __free_pages(page, order); 719 } 720} 721 722 723/* 724 * The order of subdivision here is critical for the IO subsystem. 725 * Please do not alter this order without good reasons and regression 726 * testing. Specifically, as large blocks of memory are subdivided, 727 * the order in which smaller blocks are delivered depends on the order 728 * they're subdivided in this function. This is the primary factor 729 * influencing the order in which pages are delivered to the IO 730 * subsystem according to empirical testing, and this is also justified 731 * by considering the behavior of a buddy system containing a single 732 * large block of memory acted on by a series of small allocations. 733 * This behavior is a critical factor in sglist merging's success. 734 * 735 * -- wli 736 */ 737static inline void expand(struct zone *zone, struct page *page, 738 int low, int high, struct free_area *area, 739 int migratetype) 740{ 741 unsigned long size = 1 << high; 742 743 while (high > low) { 744 area--; 745 high--; 746 size >>= 1; 747 VM_BUG_ON(bad_range(zone, &page[size])); 748 list_add(&page[size].lru, &area->free_list[migratetype]); 749 area->nr_free++; 750 set_page_order(&page[size], high); 751 } 752} 753 754/* 755 * This page is about to be returned from the page allocator 756 */ 757static inline int check_new_page(struct page *page) 758{ 759 if (unlikely(page_mapcount(page) | 760 (page->mapping != NULL) | 761 (atomic_read(&page->_count) != 0) | 762 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | 763 (mem_cgroup_bad_page_check(page)))) { 764 bad_page(page); 765 return 1; 766 } 767 return 0; 768} 769 770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 771{ 772 int i; 773 774 for (i = 0; i < (1 << order); i++) { 775 struct page *p = page + i; 776 if (unlikely(check_new_page(p))) 777 return 1; 778 } 779 780 set_page_private(page, 0); 781 set_page_refcounted(page); 782 783 arch_alloc_page(page, order); 784 kernel_map_pages(page, 1 << order, 1); 785 786 if (gfp_flags & __GFP_ZERO) 787 prep_zero_page(page, order, gfp_flags); 788 789 if (order && (gfp_flags & __GFP_COMP)) 790 prep_compound_page(page, order); 791 792 return 0; 793} 794 795/* 796 * Go through the free lists for the given migratetype and remove 797 * the smallest available page from the freelists 798 */ 799static inline 800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 801 int migratetype) 802{ 803 unsigned int current_order; 804 struct free_area * area; 805 struct page *page; 806 807 /* Find a page of the appropriate size in the preferred list */ 808 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 809 area = &(zone->free_area[current_order]); 810 if (list_empty(&area->free_list[migratetype])) 811 continue; 812 813 page = list_entry(area->free_list[migratetype].next, 814 struct page, lru); 815 list_del(&page->lru); 816 rmv_page_order(page); 817 area->nr_free--; 818 expand(zone, page, order, current_order, area, migratetype); 819 return page; 820 } 821 822 return NULL; 823} 824 825 826/* 827 * This array describes the order lists are fallen back to when 828 * the free lists for the desirable migrate type are depleted 829 */ 830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 831 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 832 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 833 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 834 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 835}; 836 837/* 838 * Move the free pages in a range to the free lists of the requested type. 839 * Note that start_page and end_pages are not aligned on a pageblock 840 * boundary. If alignment is required, use move_freepages_block() 841 */ 842static int move_freepages(struct zone *zone, 843 struct page *start_page, struct page *end_page, 844 int migratetype) 845{ 846 struct page *page; 847 unsigned long order; 848 int pages_moved = 0; 849 850#ifndef CONFIG_HOLES_IN_ZONE 851 /* 852 * page_zone is not safe to call in this context when 853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 854 * anyway as we check zone boundaries in move_freepages_block(). 855 * Remove at a later date when no bug reports exist related to 856 * grouping pages by mobility 857 */ 858 BUG_ON(page_zone(start_page) != page_zone(end_page)); 859#endif 860 861 for (page = start_page; page <= end_page;) { 862 /* Make sure we are not inadvertently changing nodes */ 863 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 864 865 if (!pfn_valid_within(page_to_pfn(page))) { 866 page++; 867 continue; 868 } 869 870 if (!PageBuddy(page)) { 871 page++; 872 continue; 873 } 874 875 order = page_order(page); 876 list_move(&page->lru, 877 &zone->free_area[order].free_list[migratetype]); 878 page += 1 << order; 879 pages_moved += 1 << order; 880 } 881 882 return pages_moved; 883} 884 885static int move_freepages_block(struct zone *zone, struct page *page, 886 int migratetype) 887{ 888 unsigned long start_pfn, end_pfn; 889 struct page *start_page, *end_page; 890 891 start_pfn = page_to_pfn(page); 892 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 893 start_page = pfn_to_page(start_pfn); 894 end_page = start_page + pageblock_nr_pages - 1; 895 end_pfn = start_pfn + pageblock_nr_pages - 1; 896 897 /* Do not cross zone boundaries */ 898 if (start_pfn < zone->zone_start_pfn) 899 start_page = page; 900 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 901 return 0; 902 903 return move_freepages(zone, start_page, end_page, migratetype); 904} 905 906static void change_pageblock_range(struct page *pageblock_page, 907 int start_order, int migratetype) 908{ 909 int nr_pageblocks = 1 << (start_order - pageblock_order); 910 911 while (nr_pageblocks--) { 912 set_pageblock_migratetype(pageblock_page, migratetype); 913 pageblock_page += pageblock_nr_pages; 914 } 915} 916 917/* Remove an element from the buddy allocator from the fallback list */ 918static inline struct page * 919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 920{ 921 struct free_area * area; 922 int current_order; 923 struct page *page; 924 int migratetype, i; 925 926 /* Find the largest possible block of pages in the other list */ 927 for (current_order = MAX_ORDER-1; current_order >= order; 928 --current_order) { 929 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 930 migratetype = fallbacks[start_migratetype][i]; 931 932 /* MIGRATE_RESERVE handled later if necessary */ 933 if (migratetype == MIGRATE_RESERVE) 934 continue; 935 936 area = &(zone->free_area[current_order]); 937 if (list_empty(&area->free_list[migratetype])) 938 continue; 939 940 page = list_entry(area->free_list[migratetype].next, 941 struct page, lru); 942 area->nr_free--; 943 944 /* 945 * If breaking a large block of pages, move all free 946 * pages to the preferred allocation list. If falling 947 * back for a reclaimable kernel allocation, be more 948 * aggressive about taking ownership of free pages 949 */ 950 if (unlikely(current_order >= (pageblock_order >> 1)) || 951 start_migratetype == MIGRATE_RECLAIMABLE || 952 page_group_by_mobility_disabled) { 953 unsigned long pages; 954 pages = move_freepages_block(zone, page, 955 start_migratetype); 956 957 /* Claim the whole block if over half of it is free */ 958 if (pages >= (1 << (pageblock_order-1)) || 959 page_group_by_mobility_disabled) 960 set_pageblock_migratetype(page, 961 start_migratetype); 962 963 migratetype = start_migratetype; 964 } 965 966 /* Remove the page from the freelists */ 967 list_del(&page->lru); 968 rmv_page_order(page); 969 970 /* Take ownership for orders >= pageblock_order */ 971 if (current_order >= pageblock_order) 972 change_pageblock_range(page, current_order, 973 start_migratetype); 974 975 expand(zone, page, order, current_order, area, migratetype); 976 977 trace_mm_page_alloc_extfrag(page, order, current_order, 978 start_migratetype, migratetype); 979 980 return page; 981 } 982 } 983 984 return NULL; 985} 986 987/* 988 * Do the hard work of removing an element from the buddy allocator. 989 * Call me with the zone->lock already held. 990 */ 991static struct page *__rmqueue(struct zone *zone, unsigned int order, 992 int migratetype) 993{ 994 struct page *page; 995 996retry_reserve: 997 page = __rmqueue_smallest(zone, order, migratetype); 998 999 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1000 page = __rmqueue_fallback(zone, order, migratetype); 1001 1002 /* 1003 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1004 * is used because __rmqueue_smallest is an inline function 1005 * and we want just one call site 1006 */ 1007 if (!page) { 1008 migratetype = MIGRATE_RESERVE; 1009 goto retry_reserve; 1010 } 1011 } 1012 1013 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1014 return page; 1015} 1016 1017/* 1018 * Obtain a specified number of elements from the buddy allocator, all under 1019 * a single hold of the lock, for efficiency. Add them to the supplied list. 1020 * Returns the number of new pages which were placed at *list. 1021 */ 1022static int rmqueue_bulk(struct zone *zone, unsigned int order, 1023 unsigned long count, struct list_head *list, 1024 int migratetype, int cold) 1025{ 1026 int i; 1027 1028 spin_lock(&zone->lock); 1029 for (i = 0; i < count; ++i) { 1030 struct page *page = __rmqueue(zone, order, migratetype); 1031 if (unlikely(page == NULL)) 1032 break; 1033 1034 /* 1035 * Split buddy pages returned by expand() are received here 1036 * in physical page order. The page is added to the callers and 1037 * list and the list head then moves forward. From the callers 1038 * perspective, the linked list is ordered by page number in 1039 * some conditions. This is useful for IO devices that can 1040 * merge IO requests if the physical pages are ordered 1041 * properly. 1042 */ 1043 if (likely(cold == 0)) 1044 list_add(&page->lru, list); 1045 else 1046 list_add_tail(&page->lru, list); 1047 set_page_private(page, migratetype); 1048 list = &page->lru; 1049 } 1050 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1051 spin_unlock(&zone->lock); 1052 return i; 1053} 1054 1055#ifdef CONFIG_NUMA 1056/* 1057 * Called from the vmstat counter updater to drain pagesets of this 1058 * currently executing processor on remote nodes after they have 1059 * expired. 1060 * 1061 * Note that this function must be called with the thread pinned to 1062 * a single processor. 1063 */ 1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1065{ 1066 unsigned long flags; 1067 int to_drain; 1068 1069 local_irq_save(flags); 1070 if (pcp->count >= pcp->batch) 1071 to_drain = pcp->batch; 1072 else 1073 to_drain = pcp->count; 1074 free_pcppages_bulk(zone, to_drain, pcp); 1075 pcp->count -= to_drain; 1076 local_irq_restore(flags); 1077} 1078#endif 1079 1080/* 1081 * Drain pages of the indicated processor. 1082 * 1083 * The processor must either be the current processor and the 1084 * thread pinned to the current processor or a processor that 1085 * is not online. 1086 */ 1087static void drain_pages(unsigned int cpu) 1088{ 1089 unsigned long flags; 1090 struct zone *zone; 1091 1092 for_each_populated_zone(zone) { 1093 struct per_cpu_pageset *pset; 1094 struct per_cpu_pages *pcp; 1095 1096 local_irq_save(flags); 1097 pset = per_cpu_ptr(zone->pageset, cpu); 1098 1099 pcp = &pset->pcp; 1100 if (pcp->count) { 1101 free_pcppages_bulk(zone, pcp->count, pcp); 1102 pcp->count = 0; 1103 } 1104 local_irq_restore(flags); 1105 } 1106} 1107 1108/* 1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1110 */ 1111void drain_local_pages(void *arg) 1112{ 1113 drain_pages(smp_processor_id()); 1114} 1115 1116/* 1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1118 */ 1119void drain_all_pages(void) 1120{ 1121 on_each_cpu(drain_local_pages, NULL, 1); 1122} 1123 1124#ifdef CONFIG_HIBERNATION 1125 1126void mark_free_pages(struct zone *zone) 1127{ 1128 unsigned long pfn, max_zone_pfn; 1129 unsigned long flags; 1130 int order, t; 1131 struct list_head *curr; 1132 1133 if (!zone->spanned_pages) 1134 return; 1135 1136 spin_lock_irqsave(&zone->lock, flags); 1137 1138 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1139 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1140 if (pfn_valid(pfn)) { 1141 struct page *page = pfn_to_page(pfn); 1142 1143 if (!swsusp_page_is_forbidden(page)) 1144 swsusp_unset_page_free(page); 1145 } 1146 1147 for_each_migratetype_order(order, t) { 1148 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1149 unsigned long i; 1150 1151 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1152 for (i = 0; i < (1UL << order); i++) 1153 swsusp_set_page_free(pfn_to_page(pfn + i)); 1154 } 1155 } 1156 spin_unlock_irqrestore(&zone->lock, flags); 1157} 1158#endif /* CONFIG_PM */ 1159 1160/* 1161 * Free a 0-order page 1162 * cold == 1 ? free a cold page : free a hot page 1163 */ 1164void free_hot_cold_page(struct page *page, int cold) 1165{ 1166 struct zone *zone = page_zone(page); 1167 struct per_cpu_pages *pcp; 1168 unsigned long flags; 1169 int migratetype; 1170 int wasMlocked = __TestClearPageMlocked(page); 1171 1172 if (!free_pages_prepare(page, 0)) 1173 return; 1174 1175 migratetype = get_pageblock_migratetype(page); 1176 set_page_private(page, migratetype); 1177 local_irq_save(flags); 1178 if (unlikely(wasMlocked)) 1179 free_page_mlock(page); 1180 __count_vm_event(PGFREE); 1181 1182 /* 1183 * We only track unmovable, reclaimable and movable on pcp lists. 1184 * Free ISOLATE pages back to the allocator because they are being 1185 * offlined but treat RESERVE as movable pages so we can get those 1186 * areas back if necessary. Otherwise, we may have to free 1187 * excessively into the page allocator 1188 */ 1189 if (migratetype >= MIGRATE_PCPTYPES) { 1190 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1191 free_one_page(zone, page, 0, migratetype); 1192 goto out; 1193 } 1194 migratetype = MIGRATE_MOVABLE; 1195 } 1196 1197 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1198 if (cold) 1199 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1200 else 1201 list_add(&page->lru, &pcp->lists[migratetype]); 1202 pcp->count++; 1203 if (pcp->count >= pcp->high) { 1204 free_pcppages_bulk(zone, pcp->batch, pcp); 1205 pcp->count -= pcp->batch; 1206 } 1207 1208out: 1209 local_irq_restore(flags); 1210} 1211 1212/* 1213 * split_page takes a non-compound higher-order page, and splits it into 1214 * n (1<<order) sub-pages: page[0..n] 1215 * Each sub-page must be freed individually. 1216 * 1217 * Note: this is probably too low level an operation for use in drivers. 1218 * Please consult with lkml before using this in your driver. 1219 */ 1220void split_page(struct page *page, unsigned int order) 1221{ 1222 int i; 1223 1224 VM_BUG_ON(PageCompound(page)); 1225 VM_BUG_ON(!page_count(page)); 1226 1227#ifdef CONFIG_KMEMCHECK 1228 /* 1229 * Split shadow pages too, because free(page[0]) would 1230 * otherwise free the whole shadow. 1231 */ 1232 if (kmemcheck_page_is_tracked(page)) 1233 split_page(virt_to_page(page[0].shadow), order); 1234#endif 1235 1236 for (i = 1; i < (1 << order); i++) 1237 set_page_refcounted(page + i); 1238} 1239 1240/* 1241 * Similar to split_page except the page is already free. As this is only 1242 * being used for migration, the migratetype of the block also changes. 1243 * As this is called with interrupts disabled, the caller is responsible 1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1245 * are enabled. 1246 * 1247 * Note: this is probably too low level an operation for use in drivers. 1248 * Please consult with lkml before using this in your driver. 1249 */ 1250int split_free_page(struct page *page) 1251{ 1252 unsigned int order; 1253 unsigned long watermark; 1254 struct zone *zone; 1255 1256 BUG_ON(!PageBuddy(page)); 1257 1258 zone = page_zone(page); 1259 order = page_order(page); 1260 1261 /* Obey watermarks as if the page was being allocated */ 1262 watermark = low_wmark_pages(zone) + (1 << order); 1263 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1264 return 0; 1265 1266 /* Remove page from free list */ 1267 list_del(&page->lru); 1268 zone->free_area[order].nr_free--; 1269 rmv_page_order(page); 1270 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1271 1272 /* Split into individual pages */ 1273 set_page_refcounted(page); 1274 split_page(page, order); 1275 1276 if (order >= pageblock_order - 1) { 1277 struct page *endpage = page + (1 << order) - 1; 1278 for (; page < endpage; page += pageblock_nr_pages) 1279 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1280 } 1281 1282 return 1 << order; 1283} 1284 1285/* 1286 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1287 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1288 * or two. 1289 */ 1290static inline 1291struct page *buffered_rmqueue(struct zone *preferred_zone, 1292 struct zone *zone, int order, gfp_t gfp_flags, 1293 int migratetype) 1294{ 1295 unsigned long flags; 1296 struct page *page; 1297 int cold = !!(gfp_flags & __GFP_COLD); 1298 1299again: 1300 if (likely(order == 0)) { 1301 struct per_cpu_pages *pcp; 1302 struct list_head *list; 1303 1304 local_irq_save(flags); 1305 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1306 list = &pcp->lists[migratetype]; 1307 if (list_empty(list)) { 1308 pcp->count += rmqueue_bulk(zone, 0, 1309 pcp->batch, list, 1310 migratetype, cold); 1311 if (unlikely(list_empty(list))) 1312 goto failed; 1313 } 1314 1315 if (cold) 1316 page = list_entry(list->prev, struct page, lru); 1317 else 1318 page = list_entry(list->next, struct page, lru); 1319 1320 list_del(&page->lru); 1321 pcp->count--; 1322 } else { 1323 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1324 /* 1325 * __GFP_NOFAIL is not to be used in new code. 1326 * 1327 * All __GFP_NOFAIL callers should be fixed so that they 1328 * properly detect and handle allocation failures. 1329 * 1330 * We most definitely don't want callers attempting to 1331 * allocate greater than order-1 page units with 1332 * __GFP_NOFAIL. 1333 */ 1334 WARN_ON_ONCE(order > 1); 1335 } 1336 spin_lock_irqsave(&zone->lock, flags); 1337 page = __rmqueue(zone, order, migratetype); 1338 spin_unlock(&zone->lock); 1339 if (!page) 1340 goto failed; 1341 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1342 } 1343 1344 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1345 zone_statistics(preferred_zone, zone, gfp_flags); 1346 local_irq_restore(flags); 1347 1348 VM_BUG_ON(bad_range(zone, page)); 1349 if (prep_new_page(page, order, gfp_flags)) 1350 goto again; 1351 return page; 1352 1353failed: 1354 local_irq_restore(flags); 1355 return NULL; 1356} 1357 1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1359#define ALLOC_WMARK_MIN WMARK_MIN 1360#define ALLOC_WMARK_LOW WMARK_LOW 1361#define ALLOC_WMARK_HIGH WMARK_HIGH 1362#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1363 1364/* Mask to get the watermark bits */ 1365#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1366 1367#define ALLOC_HARDER 0x10 /* try to alloc harder */ 1368#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1369#define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1370 1371#ifdef CONFIG_FAIL_PAGE_ALLOC 1372 1373static struct { 1374 struct fault_attr attr; 1375 1376 u32 ignore_gfp_highmem; 1377 u32 ignore_gfp_wait; 1378 u32 min_order; 1379} fail_page_alloc = { 1380 .attr = FAULT_ATTR_INITIALIZER, 1381 .ignore_gfp_wait = 1, 1382 .ignore_gfp_highmem = 1, 1383 .min_order = 1, 1384}; 1385 1386static int __init setup_fail_page_alloc(char *str) 1387{ 1388 return setup_fault_attr(&fail_page_alloc.attr, str); 1389} 1390__setup("fail_page_alloc=", setup_fail_page_alloc); 1391 1392static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1393{ 1394 if (order < fail_page_alloc.min_order) 1395 return 0; 1396 if (gfp_mask & __GFP_NOFAIL) 1397 return 0; 1398 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1399 return 0; 1400 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1401 return 0; 1402 1403 return should_fail(&fail_page_alloc.attr, 1 << order); 1404} 1405 1406#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1407 1408static int __init fail_page_alloc_debugfs(void) 1409{ 1410 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1411 struct dentry *dir; 1412 int err; 1413 1414 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1415 "fail_page_alloc"); 1416 if (err) 1417 return err; 1418 1419 dir = fail_page_alloc.attr.dir; 1420 1421 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 1422 &fail_page_alloc.ignore_gfp_wait)) 1423 goto fail; 1424 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1425 &fail_page_alloc.ignore_gfp_highmem)) 1426 goto fail; 1427 if (!debugfs_create_u32("min-order", mode, dir, 1428 &fail_page_alloc.min_order)) 1429 goto fail; 1430 1431 return 0; 1432fail: 1433 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1434 1435 return -ENOMEM; 1436} 1437 1438late_initcall(fail_page_alloc_debugfs); 1439 1440#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1441 1442#else /* CONFIG_FAIL_PAGE_ALLOC */ 1443 1444static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1445{ 1446 return 0; 1447} 1448 1449#endif /* CONFIG_FAIL_PAGE_ALLOC */ 1450 1451/* 1452 * Return true if free pages are above 'mark'. This takes into account the order 1453 * of the allocation. 1454 */ 1455static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1456 int classzone_idx, int alloc_flags, long free_pages) 1457{ 1458 /* free_pages my go negative - that's OK */ 1459 long min = mark; 1460 int o; 1461 1462 free_pages -= (1 << order) + 1; 1463 if (alloc_flags & ALLOC_HIGH) 1464 min -= min / 2; 1465 if (alloc_flags & ALLOC_HARDER) 1466 min -= min / 4; 1467 1468 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1469 return false; 1470 for (o = 0; o < order; o++) { 1471 /* At the next order, this order's pages become unavailable */ 1472 free_pages -= z->free_area[o].nr_free << o; 1473 1474 /* Require fewer higher order pages to be free */ 1475 min >>= 1; 1476 1477 if (free_pages <= min) 1478 return false; 1479 } 1480 return true; 1481} 1482 1483bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1484 int classzone_idx, int alloc_flags) 1485{ 1486 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1487 zone_page_state(z, NR_FREE_PAGES)); 1488} 1489 1490bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1491 int classzone_idx, int alloc_flags) 1492{ 1493 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1494 1495 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1496 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1497 1498 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1499 free_pages); 1500} 1501 1502#ifdef CONFIG_NUMA 1503/* 1504 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1505 * skip over zones that are not allowed by the cpuset, or that have 1506 * been recently (in last second) found to be nearly full. See further 1507 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1508 * that have to skip over a lot of full or unallowed zones. 1509 * 1510 * If the zonelist cache is present in the passed in zonelist, then 1511 * returns a pointer to the allowed node mask (either the current 1512 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1513 * 1514 * If the zonelist cache is not available for this zonelist, does 1515 * nothing and returns NULL. 1516 * 1517 * If the fullzones BITMAP in the zonelist cache is stale (more than 1518 * a second since last zap'd) then we zap it out (clear its bits.) 1519 * 1520 * We hold off even calling zlc_setup, until after we've checked the 1521 * first zone in the zonelist, on the theory that most allocations will 1522 * be satisfied from that first zone, so best to examine that zone as 1523 * quickly as we can. 1524 */ 1525static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1526{ 1527 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1528 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1529 1530 zlc = zonelist->zlcache_ptr; 1531 if (!zlc) 1532 return NULL; 1533 1534 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1535 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1536 zlc->last_full_zap = jiffies; 1537 } 1538 1539 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1540 &cpuset_current_mems_allowed : 1541 &node_states[N_HIGH_MEMORY]; 1542 return allowednodes; 1543} 1544 1545/* 1546 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1547 * if it is worth looking at further for free memory: 1548 * 1) Check that the zone isn't thought to be full (doesn't have its 1549 * bit set in the zonelist_cache fullzones BITMAP). 1550 * 2) Check that the zones node (obtained from the zonelist_cache 1551 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1552 * Return true (non-zero) if zone is worth looking at further, or 1553 * else return false (zero) if it is not. 1554 * 1555 * This check -ignores- the distinction between various watermarks, 1556 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1557 * found to be full for any variation of these watermarks, it will 1558 * be considered full for up to one second by all requests, unless 1559 * we are so low on memory on all allowed nodes that we are forced 1560 * into the second scan of the zonelist. 1561 * 1562 * In the second scan we ignore this zonelist cache and exactly 1563 * apply the watermarks to all zones, even it is slower to do so. 1564 * We are low on memory in the second scan, and should leave no stone 1565 * unturned looking for a free page. 1566 */ 1567static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1568 nodemask_t *allowednodes) 1569{ 1570 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1571 int i; /* index of *z in zonelist zones */ 1572 int n; /* node that zone *z is on */ 1573 1574 zlc = zonelist->zlcache_ptr; 1575 if (!zlc) 1576 return 1; 1577 1578 i = z - zonelist->_zonerefs; 1579 n = zlc->z_to_n[i]; 1580 1581 /* This zone is worth trying if it is allowed but not full */ 1582 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1583} 1584 1585/* 1586 * Given 'z' scanning a zonelist, set the corresponding bit in 1587 * zlc->fullzones, so that subsequent attempts to allocate a page 1588 * from that zone don't waste time re-examining it. 1589 */ 1590static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1591{ 1592 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1593 int i; /* index of *z in zonelist zones */ 1594 1595 zlc = zonelist->zlcache_ptr; 1596 if (!zlc) 1597 return; 1598 1599 i = z - zonelist->_zonerefs; 1600 1601 set_bit(i, zlc->fullzones); 1602} 1603 1604/* 1605 * clear all zones full, called after direct reclaim makes progress so that 1606 * a zone that was recently full is not skipped over for up to a second 1607 */ 1608static void zlc_clear_zones_full(struct zonelist *zonelist) 1609{ 1610 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1611 1612 zlc = zonelist->zlcache_ptr; 1613 if (!zlc) 1614 return; 1615 1616 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1617} 1618 1619#else /* CONFIG_NUMA */ 1620 1621static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1622{ 1623 return NULL; 1624} 1625 1626static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1627 nodemask_t *allowednodes) 1628{ 1629 return 1; 1630} 1631 1632static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1633{ 1634} 1635 1636static void zlc_clear_zones_full(struct zonelist *zonelist) 1637{ 1638} 1639#endif /* CONFIG_NUMA */ 1640 1641/* 1642 * get_page_from_freelist goes through the zonelist trying to allocate 1643 * a page. 1644 */ 1645static struct page * 1646get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1647 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1648 struct zone *preferred_zone, int migratetype) 1649{ 1650 struct zoneref *z; 1651 struct page *page = NULL; 1652 int classzone_idx; 1653 struct zone *zone; 1654 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1655 int zlc_active = 0; /* set if using zonelist_cache */ 1656 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1657 1658 classzone_idx = zone_idx(preferred_zone); 1659zonelist_scan: 1660 /* 1661 * Scan zonelist, looking for a zone with enough free. 1662 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1663 */ 1664 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1665 high_zoneidx, nodemask) { 1666 if (NUMA_BUILD && zlc_active && 1667 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1668 continue; 1669 if ((alloc_flags & ALLOC_CPUSET) && 1670 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1671 continue; 1672 1673 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1674 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1675 unsigned long mark; 1676 int ret; 1677 1678 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1679 if (zone_watermark_ok(zone, order, mark, 1680 classzone_idx, alloc_flags)) 1681 goto try_this_zone; 1682 1683 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1684 /* 1685 * we do zlc_setup if there are multiple nodes 1686 * and before considering the first zone allowed 1687 * by the cpuset. 1688 */ 1689 allowednodes = zlc_setup(zonelist, alloc_flags); 1690 zlc_active = 1; 1691 did_zlc_setup = 1; 1692 } 1693 1694 if (zone_reclaim_mode == 0) 1695 goto this_zone_full; 1696 1697 /* 1698 * As we may have just activated ZLC, check if the first 1699 * eligible zone has failed zone_reclaim recently. 1700 */ 1701 if (NUMA_BUILD && zlc_active && 1702 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1703 continue; 1704 1705 ret = zone_reclaim(zone, gfp_mask, order); 1706 switch (ret) { 1707 case ZONE_RECLAIM_NOSCAN: 1708 /* did not scan */ 1709 continue; 1710 case ZONE_RECLAIM_FULL: 1711 /* scanned but unreclaimable */ 1712 continue; 1713 default: 1714 /* did we reclaim enough */ 1715 if (!zone_watermark_ok(zone, order, mark, 1716 classzone_idx, alloc_flags)) 1717 goto this_zone_full; 1718 } 1719 } 1720 1721try_this_zone: 1722 page = buffered_rmqueue(preferred_zone, zone, order, 1723 gfp_mask, migratetype); 1724 if (page) 1725 break; 1726this_zone_full: 1727 if (NUMA_BUILD) 1728 zlc_mark_zone_full(zonelist, z); 1729 } 1730 1731 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1732 /* Disable zlc cache for second zonelist scan */ 1733 zlc_active = 0; 1734 goto zonelist_scan; 1735 } 1736 return page; 1737} 1738 1739/* 1740 * Large machines with many possible nodes should not always dump per-node 1741 * meminfo in irq context. 1742 */ 1743static inline bool should_suppress_show_mem(void) 1744{ 1745 bool ret = false; 1746 1747#if NODES_SHIFT > 8 1748 ret = in_interrupt(); 1749#endif 1750 return ret; 1751} 1752 1753static DEFINE_RATELIMIT_STATE(nopage_rs, 1754 DEFAULT_RATELIMIT_INTERVAL, 1755 DEFAULT_RATELIMIT_BURST); 1756 1757void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 1758{ 1759 va_list args; 1760 unsigned int filter = SHOW_MEM_FILTER_NODES; 1761 1762 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs)) 1763 return; 1764 1765 /* 1766 * This documents exceptions given to allocations in certain 1767 * contexts that are allowed to allocate outside current's set 1768 * of allowed nodes. 1769 */ 1770 if (!(gfp_mask & __GFP_NOMEMALLOC)) 1771 if (test_thread_flag(TIF_MEMDIE) || 1772 (current->flags & (PF_MEMALLOC | PF_EXITING))) 1773 filter &= ~SHOW_MEM_FILTER_NODES; 1774 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 1775 filter &= ~SHOW_MEM_FILTER_NODES; 1776 1777 if (fmt) { 1778 printk(KERN_WARNING); 1779 va_start(args, fmt); 1780 vprintk(fmt, args); 1781 va_end(args); 1782 } 1783 1784 pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n", 1785 current->comm, order, gfp_mask); 1786 1787 dump_stack(); 1788 if (!should_suppress_show_mem()) 1789 show_mem(filter); 1790} 1791 1792static inline int 1793should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1794 unsigned long pages_reclaimed) 1795{ 1796 /* Do not loop if specifically requested */ 1797 if (gfp_mask & __GFP_NORETRY) 1798 return 0; 1799 1800 /* 1801 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1802 * means __GFP_NOFAIL, but that may not be true in other 1803 * implementations. 1804 */ 1805 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1806 return 1; 1807 1808 /* 1809 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1810 * specified, then we retry until we no longer reclaim any pages 1811 * (above), or we've reclaimed an order of pages at least as 1812 * large as the allocation's order. In both cases, if the 1813 * allocation still fails, we stop retrying. 1814 */ 1815 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1816 return 1; 1817 1818 /* 1819 * Don't let big-order allocations loop unless the caller 1820 * explicitly requests that. 1821 */ 1822 if (gfp_mask & __GFP_NOFAIL) 1823 return 1; 1824 1825 return 0; 1826} 1827 1828static inline struct page * 1829__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1830 struct zonelist *zonelist, enum zone_type high_zoneidx, 1831 nodemask_t *nodemask, struct zone *preferred_zone, 1832 int migratetype) 1833{ 1834 struct page *page; 1835 1836 /* Acquire the OOM killer lock for the zones in zonelist */ 1837 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1838 schedule_timeout_uninterruptible(1); 1839 return NULL; 1840 } 1841 1842 /* 1843 * Go through the zonelist yet one more time, keep very high watermark 1844 * here, this is only to catch a parallel oom killing, we must fail if 1845 * we're still under heavy pressure. 1846 */ 1847 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1848 order, zonelist, high_zoneidx, 1849 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1850 preferred_zone, migratetype); 1851 if (page) 1852 goto out; 1853 1854 if (!(gfp_mask & __GFP_NOFAIL)) { 1855 /* The OOM killer will not help higher order allocs */ 1856 if (order > PAGE_ALLOC_COSTLY_ORDER) 1857 goto out; 1858 /* The OOM killer does not needlessly kill tasks for lowmem */ 1859 if (high_zoneidx < ZONE_NORMAL) 1860 goto out; 1861 /* 1862 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1863 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1864 * The caller should handle page allocation failure by itself if 1865 * it specifies __GFP_THISNODE. 1866 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1867 */ 1868 if (gfp_mask & __GFP_THISNODE) 1869 goto out; 1870 } 1871 /* Exhausted what can be done so it's blamo time */ 1872 out_of_memory(zonelist, gfp_mask, order, nodemask); 1873 1874out: 1875 clear_zonelist_oom(zonelist, gfp_mask); 1876 return page; 1877} 1878 1879#ifdef CONFIG_COMPACTION 1880/* Try memory compaction for high-order allocations before reclaim */ 1881static struct page * 1882__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1883 struct zonelist *zonelist, enum zone_type high_zoneidx, 1884 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1885 int migratetype, unsigned long *did_some_progress, 1886 bool sync_migration) 1887{ 1888 struct page *page; 1889 1890 if (!order || compaction_deferred(preferred_zone)) 1891 return NULL; 1892 1893 current->flags |= PF_MEMALLOC; 1894 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 1895 nodemask, sync_migration); 1896 current->flags &= ~PF_MEMALLOC; 1897 if (*did_some_progress != COMPACT_SKIPPED) { 1898 1899 /* Page migration frees to the PCP lists but we want merging */ 1900 drain_pages(get_cpu()); 1901 put_cpu(); 1902 1903 page = get_page_from_freelist(gfp_mask, nodemask, 1904 order, zonelist, high_zoneidx, 1905 alloc_flags, preferred_zone, 1906 migratetype); 1907 if (page) { 1908 preferred_zone->compact_considered = 0; 1909 preferred_zone->compact_defer_shift = 0; 1910 count_vm_event(COMPACTSUCCESS); 1911 return page; 1912 } 1913 1914 /* 1915 * It's bad if compaction run occurs and fails. 1916 * The most likely reason is that pages exist, 1917 * but not enough to satisfy watermarks. 1918 */ 1919 count_vm_event(COMPACTFAIL); 1920 defer_compaction(preferred_zone); 1921 1922 cond_resched(); 1923 } 1924 1925 return NULL; 1926} 1927#else 1928static inline struct page * 1929__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1930 struct zonelist *zonelist, enum zone_type high_zoneidx, 1931 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1932 int migratetype, unsigned long *did_some_progress, 1933 bool sync_migration) 1934{ 1935 return NULL; 1936} 1937#endif /* CONFIG_COMPACTION */ 1938 1939/* The really slow allocator path where we enter direct reclaim */ 1940static inline struct page * 1941__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 1942 struct zonelist *zonelist, enum zone_type high_zoneidx, 1943 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1944 int migratetype, unsigned long *did_some_progress) 1945{ 1946 struct page *page = NULL; 1947 struct reclaim_state reclaim_state; 1948 bool drained = false; 1949 1950 cond_resched(); 1951 1952 /* We now go into synchronous reclaim */ 1953 cpuset_memory_pressure_bump(); 1954 current->flags |= PF_MEMALLOC; 1955 lockdep_set_current_reclaim_state(gfp_mask); 1956 reclaim_state.reclaimed_slab = 0; 1957 current->reclaim_state = &reclaim_state; 1958 1959 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 1960 1961 current->reclaim_state = NULL; 1962 lockdep_clear_current_reclaim_state(); 1963 current->flags &= ~PF_MEMALLOC; 1964 1965 cond_resched(); 1966 1967 if (unlikely(!(*did_some_progress))) 1968 return NULL; 1969 1970 /* After successful reclaim, reconsider all zones for allocation */ 1971 if (NUMA_BUILD) 1972 zlc_clear_zones_full(zonelist); 1973 1974retry: 1975 page = get_page_from_freelist(gfp_mask, nodemask, order, 1976 zonelist, high_zoneidx, 1977 alloc_flags, preferred_zone, 1978 migratetype); 1979 1980 /* 1981 * If an allocation failed after direct reclaim, it could be because 1982 * pages are pinned on the per-cpu lists. Drain them and try again 1983 */ 1984 if (!page && !drained) { 1985 drain_all_pages(); 1986 drained = true; 1987 goto retry; 1988 } 1989 1990 return page; 1991} 1992 1993/* 1994 * This is called in the allocator slow-path if the allocation request is of 1995 * sufficient urgency to ignore watermarks and take other desperate measures 1996 */ 1997static inline struct page * 1998__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 1999 struct zonelist *zonelist, enum zone_type high_zoneidx, 2000 nodemask_t *nodemask, struct zone *preferred_zone, 2001 int migratetype) 2002{ 2003 struct page *page; 2004 2005 do { 2006 page = get_page_from_freelist(gfp_mask, nodemask, order, 2007 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 2008 preferred_zone, migratetype); 2009 2010 if (!page && gfp_mask & __GFP_NOFAIL) 2011 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2012 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2013 2014 return page; 2015} 2016 2017static inline 2018void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 2019 enum zone_type high_zoneidx, 2020 enum zone_type classzone_idx) 2021{ 2022 struct zoneref *z; 2023 struct zone *zone; 2024 2025 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 2026 wakeup_kswapd(zone, order, classzone_idx); 2027} 2028 2029static inline int 2030gfp_to_alloc_flags(gfp_t gfp_mask) 2031{ 2032 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2033 const gfp_t wait = gfp_mask & __GFP_WAIT; 2034 2035 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2036 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2037 2038 /* 2039 * The caller may dip into page reserves a bit more if the caller 2040 * cannot run direct reclaim, or if the caller has realtime scheduling 2041 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2042 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 2043 */ 2044 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2045 2046 if (!wait) { 2047 /* 2048 * Not worth trying to allocate harder for 2049 * __GFP_NOMEMALLOC even if it can't schedule. 2050 */ 2051 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2052 alloc_flags |= ALLOC_HARDER; 2053 /* 2054 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 2055 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 2056 */ 2057 alloc_flags &= ~ALLOC_CPUSET; 2058 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2059 alloc_flags |= ALLOC_HARDER; 2060 2061 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2062 if (!in_interrupt() && 2063 ((current->flags & PF_MEMALLOC) || 2064 unlikely(test_thread_flag(TIF_MEMDIE)))) 2065 alloc_flags |= ALLOC_NO_WATERMARKS; 2066 } 2067 2068 return alloc_flags; 2069} 2070 2071static inline struct page * 2072__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2073 struct zonelist *zonelist, enum zone_type high_zoneidx, 2074 nodemask_t *nodemask, struct zone *preferred_zone, 2075 int migratetype) 2076{ 2077 const gfp_t wait = gfp_mask & __GFP_WAIT; 2078 struct page *page = NULL; 2079 int alloc_flags; 2080 unsigned long pages_reclaimed = 0; 2081 unsigned long did_some_progress; 2082 bool sync_migration = false; 2083 2084 /* 2085 * In the slowpath, we sanity check order to avoid ever trying to 2086 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2087 * be using allocators in order of preference for an area that is 2088 * too large. 2089 */ 2090 if (order >= MAX_ORDER) { 2091 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2092 return NULL; 2093 } 2094 2095 /* 2096 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2097 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2098 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2099 * using a larger set of nodes after it has established that the 2100 * allowed per node queues are empty and that nodes are 2101 * over allocated. 2102 */ 2103 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2104 goto nopage; 2105 2106restart: 2107 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2108 wake_all_kswapd(order, zonelist, high_zoneidx, 2109 zone_idx(preferred_zone)); 2110 2111 /* 2112 * OK, we're below the kswapd watermark and have kicked background 2113 * reclaim. Now things get more complex, so set up alloc_flags according 2114 * to how we want to proceed. 2115 */ 2116 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2117 2118 /* 2119 * Find the true preferred zone if the allocation is unconstrained by 2120 * cpusets. 2121 */ 2122 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) 2123 first_zones_zonelist(zonelist, high_zoneidx, NULL, 2124 &preferred_zone); 2125 2126rebalance: 2127 /* This is the last chance, in general, before the goto nopage. */ 2128 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2129 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2130 preferred_zone, migratetype); 2131 if (page) 2132 goto got_pg; 2133 2134 /* Allocate without watermarks if the context allows */ 2135 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2136 page = __alloc_pages_high_priority(gfp_mask, order, 2137 zonelist, high_zoneidx, nodemask, 2138 preferred_zone, migratetype); 2139 if (page) 2140 goto got_pg; 2141 } 2142 2143 /* Atomic allocations - we can't balance anything */ 2144 if (!wait) 2145 goto nopage; 2146 2147 /* Avoid recursion of direct reclaim */ 2148 if (current->flags & PF_MEMALLOC) 2149 goto nopage; 2150 2151 /* Avoid allocations with no watermarks from looping endlessly */ 2152 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2153 goto nopage; 2154 2155 /* 2156 * Try direct compaction. The first pass is asynchronous. Subsequent 2157 * attempts after direct reclaim are synchronous 2158 */ 2159 page = __alloc_pages_direct_compact(gfp_mask, order, 2160 zonelist, high_zoneidx, 2161 nodemask, 2162 alloc_flags, preferred_zone, 2163 migratetype, &did_some_progress, 2164 sync_migration); 2165 if (page) 2166 goto got_pg; 2167 sync_migration = true; 2168 2169 /* Try direct reclaim and then allocating */ 2170 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2171 zonelist, high_zoneidx, 2172 nodemask, 2173 alloc_flags, preferred_zone, 2174 migratetype, &did_some_progress); 2175 if (page) 2176 goto got_pg; 2177 2178 /* 2179 * If we failed to make any progress reclaiming, then we are 2180 * running out of options and have to consider going OOM 2181 */ 2182 if (!did_some_progress) { 2183 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2184 if (oom_killer_disabled) 2185 goto nopage; 2186 page = __alloc_pages_may_oom(gfp_mask, order, 2187 zonelist, high_zoneidx, 2188 nodemask, preferred_zone, 2189 migratetype); 2190 if (page) 2191 goto got_pg; 2192 2193 if (!(gfp_mask & __GFP_NOFAIL)) { 2194 /* 2195 * The oom killer is not called for high-order 2196 * allocations that may fail, so if no progress 2197 * is being made, there are no other options and 2198 * retrying is unlikely to help. 2199 */ 2200 if (order > PAGE_ALLOC_COSTLY_ORDER) 2201 goto nopage; 2202 /* 2203 * The oom killer is not called for lowmem 2204 * allocations to prevent needlessly killing 2205 * innocent tasks. 2206 */ 2207 if (high_zoneidx < ZONE_NORMAL) 2208 goto nopage; 2209 } 2210 2211 goto restart; 2212 } 2213 } 2214 2215 /* Check if we should retry the allocation */ 2216 pages_reclaimed += did_some_progress; 2217 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) { 2218 /* Wait for some write requests to complete then retry */ 2219 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2220 goto rebalance; 2221 } else { 2222 /* 2223 * High-order allocations do not necessarily loop after 2224 * direct reclaim and reclaim/compaction depends on compaction 2225 * being called after reclaim so call directly if necessary 2226 */ 2227 page = __alloc_pages_direct_compact(gfp_mask, order, 2228 zonelist, high_zoneidx, 2229 nodemask, 2230 alloc_flags, preferred_zone, 2231 migratetype, &did_some_progress, 2232 sync_migration); 2233 if (page) 2234 goto got_pg; 2235 } 2236 2237nopage: 2238 warn_alloc_failed(gfp_mask, order, NULL); 2239 return page; 2240got_pg: 2241 if (kmemcheck_enabled) 2242 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2243 return page; 2244 2245} 2246 2247/* 2248 * This is the 'heart' of the zoned buddy allocator. 2249 */ 2250struct page * 2251__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2252 struct zonelist *zonelist, nodemask_t *nodemask) 2253{ 2254 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2255 struct zone *preferred_zone; 2256 struct page *page; 2257 int migratetype = allocflags_to_migratetype(gfp_mask); 2258 2259 gfp_mask &= gfp_allowed_mask; 2260 2261 lockdep_trace_alloc(gfp_mask); 2262 2263 might_sleep_if(gfp_mask & __GFP_WAIT); 2264 2265 if (should_fail_alloc_page(gfp_mask, order)) 2266 return NULL; 2267 2268 /* 2269 * Check the zones suitable for the gfp_mask contain at least one 2270 * valid zone. It's possible to have an empty zonelist as a result 2271 * of GFP_THISNODE and a memoryless node 2272 */ 2273 if (unlikely(!zonelist->_zonerefs->zone)) 2274 return NULL; 2275 2276 get_mems_allowed(); 2277 /* The preferred zone is used for statistics later */ 2278 first_zones_zonelist(zonelist, high_zoneidx, 2279 nodemask ? : &cpuset_current_mems_allowed, 2280 &preferred_zone); 2281 if (!preferred_zone) { 2282 put_mems_allowed(); 2283 return NULL; 2284 } 2285 2286 /* First allocation attempt */ 2287 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2288 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2289 preferred_zone, migratetype); 2290 if (unlikely(!page)) 2291 page = __alloc_pages_slowpath(gfp_mask, order, 2292 zonelist, high_zoneidx, nodemask, 2293 preferred_zone, migratetype); 2294 put_mems_allowed(); 2295 2296 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2297 return page; 2298} 2299EXPORT_SYMBOL(__alloc_pages_nodemask); 2300 2301/* 2302 * Common helper functions. 2303 */ 2304unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2305{ 2306 struct page *page; 2307 2308 /* 2309 * __get_free_pages() returns a 32-bit address, which cannot represent 2310 * a highmem page 2311 */ 2312 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2313 2314 page = alloc_pages(gfp_mask, order); 2315 if (!page) 2316 return 0; 2317 return (unsigned long) page_address(page); 2318} 2319EXPORT_SYMBOL(__get_free_pages); 2320 2321unsigned long get_zeroed_page(gfp_t gfp_mask) 2322{ 2323 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2324} 2325EXPORT_SYMBOL(get_zeroed_page); 2326 2327void __pagevec_free(struct pagevec *pvec) 2328{ 2329 int i = pagevec_count(pvec); 2330 2331 while (--i >= 0) { 2332 trace_mm_pagevec_free(pvec->pages[i], pvec->cold); 2333 free_hot_cold_page(pvec->pages[i], pvec->cold); 2334 } 2335} 2336 2337void __free_pages(struct page *page, unsigned int order) 2338{ 2339 if (put_page_testzero(page)) { 2340 if (order == 0) 2341 free_hot_cold_page(page, 0); 2342 else 2343 __free_pages_ok(page, order); 2344 } 2345} 2346 2347EXPORT_SYMBOL(__free_pages); 2348 2349void free_pages(unsigned long addr, unsigned int order) 2350{ 2351 if (addr != 0) { 2352 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2353 __free_pages(virt_to_page((void *)addr), order); 2354 } 2355} 2356 2357EXPORT_SYMBOL(free_pages); 2358 2359static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 2360{ 2361 if (addr) { 2362 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2363 unsigned long used = addr + PAGE_ALIGN(size); 2364 2365 split_page(virt_to_page((void *)addr), order); 2366 while (used < alloc_end) { 2367 free_page(used); 2368 used += PAGE_SIZE; 2369 } 2370 } 2371 return (void *)addr; 2372} 2373 2374/** 2375 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2376 * @size: the number of bytes to allocate 2377 * @gfp_mask: GFP flags for the allocation 2378 * 2379 * This function is similar to alloc_pages(), except that it allocates the 2380 * minimum number of pages to satisfy the request. alloc_pages() can only 2381 * allocate memory in power-of-two pages. 2382 * 2383 * This function is also limited by MAX_ORDER. 2384 * 2385 * Memory allocated by this function must be released by free_pages_exact(). 2386 */ 2387void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2388{ 2389 unsigned int order = get_order(size); 2390 unsigned long addr; 2391 2392 addr = __get_free_pages(gfp_mask, order); 2393 return make_alloc_exact(addr, order, size); 2394} 2395EXPORT_SYMBOL(alloc_pages_exact); 2396 2397/** 2398 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 2399 * pages on a node. 2400 * @nid: the preferred node ID where memory should be allocated 2401 * @size: the number of bytes to allocate 2402 * @gfp_mask: GFP flags for the allocation 2403 * 2404 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 2405 * back. 2406 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 2407 * but is not exact. 2408 */ 2409void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 2410{ 2411 unsigned order = get_order(size); 2412 struct page *p = alloc_pages_node(nid, gfp_mask, order); 2413 if (!p) 2414 return NULL; 2415 return make_alloc_exact((unsigned long)page_address(p), order, size); 2416} 2417EXPORT_SYMBOL(alloc_pages_exact_nid); 2418 2419/** 2420 * free_pages_exact - release memory allocated via alloc_pages_exact() 2421 * @virt: the value returned by alloc_pages_exact. 2422 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2423 * 2424 * Release the memory allocated by a previous call to alloc_pages_exact. 2425 */ 2426void free_pages_exact(void *virt, size_t size) 2427{ 2428 unsigned long addr = (unsigned long)virt; 2429 unsigned long end = addr + PAGE_ALIGN(size); 2430 2431 while (addr < end) { 2432 free_page(addr); 2433 addr += PAGE_SIZE; 2434 } 2435} 2436EXPORT_SYMBOL(free_pages_exact); 2437 2438static unsigned int nr_free_zone_pages(int offset) 2439{ 2440 struct zoneref *z; 2441 struct zone *zone; 2442 2443 /* Just pick one node, since fallback list is circular */ 2444 unsigned int sum = 0; 2445 2446 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2447 2448 for_each_zone_zonelist(zone, z, zonelist, offset) { 2449 unsigned long size = zone->present_pages; 2450 unsigned long high = high_wmark_pages(zone); 2451 if (size > high) 2452 sum += size - high; 2453 } 2454 2455 return sum; 2456} 2457 2458/* 2459 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2460 */ 2461unsigned int nr_free_buffer_pages(void) 2462{ 2463 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2464} 2465EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2466 2467/* 2468 * Amount of free RAM allocatable within all zones 2469 */ 2470unsigned int nr_free_pagecache_pages(void) 2471{ 2472 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2473} 2474 2475static inline void show_node(struct zone *zone) 2476{ 2477 if (NUMA_BUILD) 2478 printk("Node %d ", zone_to_nid(zone)); 2479} 2480 2481void si_meminfo(struct sysinfo *val) 2482{ 2483 val->totalram = totalram_pages; 2484 val->sharedram = 0; 2485 val->freeram = global_page_state(NR_FREE_PAGES); 2486 val->bufferram = nr_blockdev_pages(); 2487 val->totalhigh = totalhigh_pages; 2488 val->freehigh = nr_free_highpages(); 2489 val->mem_unit = PAGE_SIZE; 2490} 2491 2492EXPORT_SYMBOL(si_meminfo); 2493 2494#ifdef CONFIG_NUMA 2495void si_meminfo_node(struct sysinfo *val, int nid) 2496{ 2497 pg_data_t *pgdat = NODE_DATA(nid); 2498 2499 val->totalram = pgdat->node_present_pages; 2500 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2501#ifdef CONFIG_HIGHMEM 2502 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2503 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2504 NR_FREE_PAGES); 2505#else 2506 val->totalhigh = 0; 2507 val->freehigh = 0; 2508#endif 2509 val->mem_unit = PAGE_SIZE; 2510} 2511#endif 2512 2513/* 2514 * Determine whether the node should be displayed or not, depending on whether 2515 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 2516 */ 2517bool skip_free_areas_node(unsigned int flags, int nid) 2518{ 2519 bool ret = false; 2520 2521 if (!(flags & SHOW_MEM_FILTER_NODES)) 2522 goto out; 2523 2524 get_mems_allowed(); 2525 ret = !node_isset(nid, cpuset_current_mems_allowed); 2526 put_mems_allowed(); 2527out: 2528 return ret; 2529} 2530 2531#define K(x) ((x) << (PAGE_SHIFT-10)) 2532 2533/* 2534 * Show free area list (used inside shift_scroll-lock stuff) 2535 * We also calculate the percentage fragmentation. We do this by counting the 2536 * memory on each free list with the exception of the first item on the list. 2537 * Suppresses nodes that are not allowed by current's cpuset if 2538 * SHOW_MEM_FILTER_NODES is passed. 2539 */ 2540void show_free_areas(unsigned int filter) 2541{ 2542 int cpu; 2543 struct zone *zone; 2544 2545 for_each_populated_zone(zone) { 2546 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2547 continue; 2548 show_node(zone); 2549 printk("%s per-cpu:\n", zone->name); 2550 2551 for_each_online_cpu(cpu) { 2552 struct per_cpu_pageset *pageset; 2553 2554 pageset = per_cpu_ptr(zone->pageset, cpu); 2555 2556 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2557 cpu, pageset->pcp.high, 2558 pageset->pcp.batch, pageset->pcp.count); 2559 } 2560 } 2561 2562 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2563 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2564 " unevictable:%lu" 2565 " dirty:%lu writeback:%lu unstable:%lu\n" 2566 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2567 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2568 global_page_state(NR_ACTIVE_ANON), 2569 global_page_state(NR_INACTIVE_ANON), 2570 global_page_state(NR_ISOLATED_ANON), 2571 global_page_state(NR_ACTIVE_FILE), 2572 global_page_state(NR_INACTIVE_FILE), 2573 global_page_state(NR_ISOLATED_FILE), 2574 global_page_state(NR_UNEVICTABLE), 2575 global_page_state(NR_FILE_DIRTY), 2576 global_page_state(NR_WRITEBACK), 2577 global_page_state(NR_UNSTABLE_NFS), 2578 global_page_state(NR_FREE_PAGES), 2579 global_page_state(NR_SLAB_RECLAIMABLE), 2580 global_page_state(NR_SLAB_UNRECLAIMABLE), 2581 global_page_state(NR_FILE_MAPPED), 2582 global_page_state(NR_SHMEM), 2583 global_page_state(NR_PAGETABLE), 2584 global_page_state(NR_BOUNCE)); 2585 2586 for_each_populated_zone(zone) { 2587 int i; 2588 2589 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2590 continue; 2591 show_node(zone); 2592 printk("%s" 2593 " free:%lukB" 2594 " min:%lukB" 2595 " low:%lukB" 2596 " high:%lukB" 2597 " active_anon:%lukB" 2598 " inactive_anon:%lukB" 2599 " active_file:%lukB" 2600 " inactive_file:%lukB" 2601 " unevictable:%lukB" 2602 " isolated(anon):%lukB" 2603 " isolated(file):%lukB" 2604 " present:%lukB" 2605 " mlocked:%lukB" 2606 " dirty:%lukB" 2607 " writeback:%lukB" 2608 " mapped:%lukB" 2609 " shmem:%lukB" 2610 " slab_reclaimable:%lukB" 2611 " slab_unreclaimable:%lukB" 2612 " kernel_stack:%lukB" 2613 " pagetables:%lukB" 2614 " unstable:%lukB" 2615 " bounce:%lukB" 2616 " writeback_tmp:%lukB" 2617 " pages_scanned:%lu" 2618 " all_unreclaimable? %s" 2619 "\n", 2620 zone->name, 2621 K(zone_page_state(zone, NR_FREE_PAGES)), 2622 K(min_wmark_pages(zone)), 2623 K(low_wmark_pages(zone)), 2624 K(high_wmark_pages(zone)), 2625 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2626 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2627 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2628 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2629 K(zone_page_state(zone, NR_UNEVICTABLE)), 2630 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2631 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2632 K(zone->present_pages), 2633 K(zone_page_state(zone, NR_MLOCK)), 2634 K(zone_page_state(zone, NR_FILE_DIRTY)), 2635 K(zone_page_state(zone, NR_WRITEBACK)), 2636 K(zone_page_state(zone, NR_FILE_MAPPED)), 2637 K(zone_page_state(zone, NR_SHMEM)), 2638 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2639 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2640 zone_page_state(zone, NR_KERNEL_STACK) * 2641 THREAD_SIZE / 1024, 2642 K(zone_page_state(zone, NR_PAGETABLE)), 2643 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2644 K(zone_page_state(zone, NR_BOUNCE)), 2645 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2646 zone->pages_scanned, 2647 (zone->all_unreclaimable ? "yes" : "no") 2648 ); 2649 printk("lowmem_reserve[]:"); 2650 for (i = 0; i < MAX_NR_ZONES; i++) 2651 printk(" %lu", zone->lowmem_reserve[i]); 2652 printk("\n"); 2653 } 2654 2655 for_each_populated_zone(zone) { 2656 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2657 2658 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2659 continue; 2660 show_node(zone); 2661 printk("%s: ", zone->name); 2662 2663 spin_lock_irqsave(&zone->lock, flags); 2664 for (order = 0; order < MAX_ORDER; order++) { 2665 nr[order] = zone->free_area[order].nr_free; 2666 total += nr[order] << order; 2667 } 2668 spin_unlock_irqrestore(&zone->lock, flags); 2669 for (order = 0; order < MAX_ORDER; order++) 2670 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2671 printk("= %lukB\n", K(total)); 2672 } 2673 2674 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2675 2676 show_swap_cache_info(); 2677} 2678 2679static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2680{ 2681 zoneref->zone = zone; 2682 zoneref->zone_idx = zone_idx(zone); 2683} 2684 2685/* 2686 * Builds allocation fallback zone lists. 2687 * 2688 * Add all populated zones of a node to the zonelist. 2689 */ 2690static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2691 int nr_zones, enum zone_type zone_type) 2692{ 2693 struct zone *zone; 2694 2695 BUG_ON(zone_type >= MAX_NR_ZONES); 2696 zone_type++; 2697 2698 do { 2699 zone_type--; 2700 zone = pgdat->node_zones + zone_type; 2701 if (populated_zone(zone)) { 2702 zoneref_set_zone(zone, 2703 &zonelist->_zonerefs[nr_zones++]); 2704 check_highest_zone(zone_type); 2705 } 2706 2707 } while (zone_type); 2708 return nr_zones; 2709} 2710 2711 2712/* 2713 * zonelist_order: 2714 * 0 = automatic detection of better ordering. 2715 * 1 = order by ([node] distance, -zonetype) 2716 * 2 = order by (-zonetype, [node] distance) 2717 * 2718 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2719 * the same zonelist. So only NUMA can configure this param. 2720 */ 2721#define ZONELIST_ORDER_DEFAULT 0 2722#define ZONELIST_ORDER_NODE 1 2723#define ZONELIST_ORDER_ZONE 2 2724 2725/* zonelist order in the kernel. 2726 * set_zonelist_order() will set this to NODE or ZONE. 2727 */ 2728static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2729static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2730 2731 2732#ifdef CONFIG_NUMA 2733/* The value user specified ....changed by config */ 2734static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2735/* string for sysctl */ 2736#define NUMA_ZONELIST_ORDER_LEN 16 2737char numa_zonelist_order[16] = "default"; 2738 2739/* 2740 * interface for configure zonelist ordering. 2741 * command line option "numa_zonelist_order" 2742 * = "[dD]efault - default, automatic configuration. 2743 * = "[nN]ode - order by node locality, then by zone within node 2744 * = "[zZ]one - order by zone, then by locality within zone 2745 */ 2746 2747static int __parse_numa_zonelist_order(char *s) 2748{ 2749 if (*s == 'd' || *s == 'D') { 2750 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2751 } else if (*s == 'n' || *s == 'N') { 2752 user_zonelist_order = ZONELIST_ORDER_NODE; 2753 } else if (*s == 'z' || *s == 'Z') { 2754 user_zonelist_order = ZONELIST_ORDER_ZONE; 2755 } else { 2756 printk(KERN_WARNING 2757 "Ignoring invalid numa_zonelist_order value: " 2758 "%s\n", s); 2759 return -EINVAL; 2760 } 2761 return 0; 2762} 2763 2764static __init int setup_numa_zonelist_order(char *s) 2765{ 2766 int ret; 2767 2768 if (!s) 2769 return 0; 2770 2771 ret = __parse_numa_zonelist_order(s); 2772 if (ret == 0) 2773 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 2774 2775 return ret; 2776} 2777early_param("numa_zonelist_order", setup_numa_zonelist_order); 2778 2779/* 2780 * sysctl handler for numa_zonelist_order 2781 */ 2782int numa_zonelist_order_handler(ctl_table *table, int write, 2783 void __user *buffer, size_t *length, 2784 loff_t *ppos) 2785{ 2786 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2787 int ret; 2788 static DEFINE_MUTEX(zl_order_mutex); 2789 2790 mutex_lock(&zl_order_mutex); 2791 if (write) 2792 strcpy(saved_string, (char*)table->data); 2793 ret = proc_dostring(table, write, buffer, length, ppos); 2794 if (ret) 2795 goto out; 2796 if (write) { 2797 int oldval = user_zonelist_order; 2798 if (__parse_numa_zonelist_order((char*)table->data)) { 2799 /* 2800 * bogus value. restore saved string 2801 */ 2802 strncpy((char*)table->data, saved_string, 2803 NUMA_ZONELIST_ORDER_LEN); 2804 user_zonelist_order = oldval; 2805 } else if (oldval != user_zonelist_order) { 2806 mutex_lock(&zonelists_mutex); 2807 build_all_zonelists(NULL); 2808 mutex_unlock(&zonelists_mutex); 2809 } 2810 } 2811out: 2812 mutex_unlock(&zl_order_mutex); 2813 return ret; 2814} 2815 2816 2817#define MAX_NODE_LOAD (nr_online_nodes) 2818static int node_load[MAX_NUMNODES]; 2819 2820/** 2821 * find_next_best_node - find the next node that should appear in a given node's fallback list 2822 * @node: node whose fallback list we're appending 2823 * @used_node_mask: nodemask_t of already used nodes 2824 * 2825 * We use a number of factors to determine which is the next node that should 2826 * appear on a given node's fallback list. The node should not have appeared 2827 * already in @node's fallback list, and it should be the next closest node 2828 * according to the distance array (which contains arbitrary distance values 2829 * from each node to each node in the system), and should also prefer nodes 2830 * with no CPUs, since presumably they'll have very little allocation pressure 2831 * on them otherwise. 2832 * It returns -1 if no node is found. 2833 */ 2834static int find_next_best_node(int node, nodemask_t *used_node_mask) 2835{ 2836 int n, val; 2837 int min_val = INT_MAX; 2838 int best_node = -1; 2839 const struct cpumask *tmp = cpumask_of_node(0); 2840 2841 /* Use the local node if we haven't already */ 2842 if (!node_isset(node, *used_node_mask)) { 2843 node_set(node, *used_node_mask); 2844 return node; 2845 } 2846 2847 for_each_node_state(n, N_HIGH_MEMORY) { 2848 2849 /* Don't want a node to appear more than once */ 2850 if (node_isset(n, *used_node_mask)) 2851 continue; 2852 2853 /* Use the distance array to find the distance */ 2854 val = node_distance(node, n); 2855 2856 /* Penalize nodes under us ("prefer the next node") */ 2857 val += (n < node); 2858 2859 /* Give preference to headless and unused nodes */ 2860 tmp = cpumask_of_node(n); 2861 if (!cpumask_empty(tmp)) 2862 val += PENALTY_FOR_NODE_WITH_CPUS; 2863 2864 /* Slight preference for less loaded node */ 2865 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2866 val += node_load[n]; 2867 2868 if (val < min_val) { 2869 min_val = val; 2870 best_node = n; 2871 } 2872 } 2873 2874 if (best_node >= 0) 2875 node_set(best_node, *used_node_mask); 2876 2877 return best_node; 2878} 2879 2880 2881/* 2882 * Build zonelists ordered by node and zones within node. 2883 * This results in maximum locality--normal zone overflows into local 2884 * DMA zone, if any--but risks exhausting DMA zone. 2885 */ 2886static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2887{ 2888 int j; 2889 struct zonelist *zonelist; 2890 2891 zonelist = &pgdat->node_zonelists[0]; 2892 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 2893 ; 2894 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 2895 MAX_NR_ZONES - 1); 2896 zonelist->_zonerefs[j].zone = NULL; 2897 zonelist->_zonerefs[j].zone_idx = 0; 2898} 2899 2900/* 2901 * Build gfp_thisnode zonelists 2902 */ 2903static void build_thisnode_zonelists(pg_data_t *pgdat) 2904{ 2905 int j; 2906 struct zonelist *zonelist; 2907 2908 zonelist = &pgdat->node_zonelists[1]; 2909 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 2910 zonelist->_zonerefs[j].zone = NULL; 2911 zonelist->_zonerefs[j].zone_idx = 0; 2912} 2913 2914/* 2915 * Build zonelists ordered by zone and nodes within zones. 2916 * This results in conserving DMA zone[s] until all Normal memory is 2917 * exhausted, but results in overflowing to remote node while memory 2918 * may still exist in local DMA zone. 2919 */ 2920static int node_order[MAX_NUMNODES]; 2921 2922static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2923{ 2924 int pos, j, node; 2925 int zone_type; /* needs to be signed */ 2926 struct zone *z; 2927 struct zonelist *zonelist; 2928 2929 zonelist = &pgdat->node_zonelists[0]; 2930 pos = 0; 2931 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 2932 for (j = 0; j < nr_nodes; j++) { 2933 node = node_order[j]; 2934 z = &NODE_DATA(node)->node_zones[zone_type]; 2935 if (populated_zone(z)) { 2936 zoneref_set_zone(z, 2937 &zonelist->_zonerefs[pos++]); 2938 check_highest_zone(zone_type); 2939 } 2940 } 2941 } 2942 zonelist->_zonerefs[pos].zone = NULL; 2943 zonelist->_zonerefs[pos].zone_idx = 0; 2944} 2945 2946static int default_zonelist_order(void) 2947{ 2948 int nid, zone_type; 2949 unsigned long low_kmem_size,total_size; 2950 struct zone *z; 2951 int average_size; 2952 /* 2953 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 2954 * If they are really small and used heavily, the system can fall 2955 * into OOM very easily. 2956 * This function detect ZONE_DMA/DMA32 size and configures zone order. 2957 */ 2958 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2959 low_kmem_size = 0; 2960 total_size = 0; 2961 for_each_online_node(nid) { 2962 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2963 z = &NODE_DATA(nid)->node_zones[zone_type]; 2964 if (populated_zone(z)) { 2965 if (zone_type < ZONE_NORMAL) 2966 low_kmem_size += z->present_pages; 2967 total_size += z->present_pages; 2968 } else if (zone_type == ZONE_NORMAL) { 2969 /* 2970 * If any node has only lowmem, then node order 2971 * is preferred to allow kernel allocations 2972 * locally; otherwise, they can easily infringe 2973 * on other nodes when there is an abundance of 2974 * lowmem available to allocate from. 2975 */ 2976 return ZONELIST_ORDER_NODE; 2977 } 2978 } 2979 } 2980 if (!low_kmem_size || /* there are no DMA area. */ 2981 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2982 return ZONELIST_ORDER_NODE; 2983 /* 2984 * look into each node's config. 2985 * If there is a node whose DMA/DMA32 memory is very big area on 2986 * local memory, NODE_ORDER may be suitable. 2987 */ 2988 average_size = total_size / 2989 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2990 for_each_online_node(nid) { 2991 low_kmem_size = 0; 2992 total_size = 0; 2993 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2994 z = &NODE_DATA(nid)->node_zones[zone_type]; 2995 if (populated_zone(z)) { 2996 if (zone_type < ZONE_NORMAL) 2997 low_kmem_size += z->present_pages; 2998 total_size += z->present_pages; 2999 } 3000 } 3001 if (low_kmem_size && 3002 total_size > average_size && /* ignore small node */ 3003 low_kmem_size > total_size * 70/100) 3004 return ZONELIST_ORDER_NODE; 3005 } 3006 return ZONELIST_ORDER_ZONE; 3007} 3008 3009static void set_zonelist_order(void) 3010{ 3011 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 3012 current_zonelist_order = default_zonelist_order(); 3013 else 3014 current_zonelist_order = user_zonelist_order; 3015} 3016 3017static void build_zonelists(pg_data_t *pgdat) 3018{ 3019 int j, node, load; 3020 enum zone_type i; 3021 nodemask_t used_mask; 3022 int local_node, prev_node; 3023 struct zonelist *zonelist; 3024 int order = current_zonelist_order; 3025 3026 /* initialize zonelists */ 3027 for (i = 0; i < MAX_ZONELISTS; i++) { 3028 zonelist = pgdat->node_zonelists + i; 3029 zonelist->_zonerefs[0].zone = NULL; 3030 zonelist->_zonerefs[0].zone_idx = 0; 3031 } 3032 3033 /* NUMA-aware ordering of nodes */ 3034 local_node = pgdat->node_id; 3035 load = nr_online_nodes; 3036 prev_node = local_node; 3037 nodes_clear(used_mask); 3038 3039 memset(node_order, 0, sizeof(node_order)); 3040 j = 0; 3041 3042 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 3043 int distance = node_distance(local_node, node); 3044 3045 /* 3046 * If another node is sufficiently far away then it is better 3047 * to reclaim pages in a zone before going off node. 3048 */ 3049 if (distance > RECLAIM_DISTANCE) 3050 zone_reclaim_mode = 1; 3051 3052 /* 3053 * We don't want to pressure a particular node. 3054 * So adding penalty to the first node in same 3055 * distance group to make it round-robin. 3056 */ 3057 if (distance != node_distance(local_node, prev_node)) 3058 node_load[node] = load; 3059 3060 prev_node = node; 3061 load--; 3062 if (order == ZONELIST_ORDER_NODE) 3063 build_zonelists_in_node_order(pgdat, node); 3064 else 3065 node_order[j++] = node; /* remember order */ 3066 } 3067 3068 if (order == ZONELIST_ORDER_ZONE) { 3069 /* calculate node order -- i.e., DMA last! */ 3070 build_zonelists_in_zone_order(pgdat, j); 3071 } 3072 3073 build_thisnode_zonelists(pgdat); 3074} 3075 3076/* Construct the zonelist performance cache - see further mmzone.h */ 3077static void build_zonelist_cache(pg_data_t *pgdat) 3078{ 3079 struct zonelist *zonelist; 3080 struct zonelist_cache *zlc; 3081 struct zoneref *z; 3082 3083 zonelist = &pgdat->node_zonelists[0]; 3084 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 3085 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 3086 for (z = zonelist->_zonerefs; z->zone; z++) 3087 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 3088} 3089 3090#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3091/* 3092 * Return node id of node used for "local" allocations. 3093 * I.e., first node id of first zone in arg node's generic zonelist. 3094 * Used for initializing percpu 'numa_mem', which is used primarily 3095 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 3096 */ 3097int local_memory_node(int node) 3098{ 3099 struct zone *zone; 3100 3101 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 3102 gfp_zone(GFP_KERNEL), 3103 NULL, 3104 &zone); 3105 return zone->node; 3106} 3107#endif 3108 3109#else /* CONFIG_NUMA */ 3110 3111static void set_zonelist_order(void) 3112{ 3113 current_zonelist_order = ZONELIST_ORDER_ZONE; 3114} 3115 3116static void build_zonelists(pg_data_t *pgdat) 3117{ 3118 int node, local_node; 3119 enum zone_type j; 3120 struct zonelist *zonelist; 3121 3122 local_node = pgdat->node_id; 3123 3124 zonelist = &pgdat->node_zonelists[0]; 3125 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3126 3127 /* 3128 * Now we build the zonelist so that it contains the zones 3129 * of all the other nodes. 3130 * We don't want to pressure a particular node, so when 3131 * building the zones for node N, we make sure that the 3132 * zones coming right after the local ones are those from 3133 * node N+1 (modulo N) 3134 */ 3135 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 3136 if (!node_online(node)) 3137 continue; 3138 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3139 MAX_NR_ZONES - 1); 3140 } 3141 for (node = 0; node < local_node; node++) { 3142 if (!node_online(node)) 3143 continue; 3144 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3145 MAX_NR_ZONES - 1); 3146 } 3147 3148 zonelist->_zonerefs[j].zone = NULL; 3149 zonelist->_zonerefs[j].zone_idx = 0; 3150} 3151 3152/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3153static void build_zonelist_cache(pg_data_t *pgdat) 3154{ 3155 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3156} 3157 3158#endif /* CONFIG_NUMA */ 3159 3160/* 3161 * Boot pageset table. One per cpu which is going to be used for all 3162 * zones and all nodes. The parameters will be set in such a way 3163 * that an item put on a list will immediately be handed over to 3164 * the buddy list. This is safe since pageset manipulation is done 3165 * with interrupts disabled. 3166 * 3167 * The boot_pagesets must be kept even after bootup is complete for 3168 * unused processors and/or zones. They do play a role for bootstrapping 3169 * hotplugged processors. 3170 * 3171 * zoneinfo_show() and maybe other functions do 3172 * not check if the processor is online before following the pageset pointer. 3173 * Other parts of the kernel may not check if the zone is available. 3174 */ 3175static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3176static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3177static void setup_zone_pageset(struct zone *zone); 3178 3179/* 3180 * Global mutex to protect against size modification of zonelists 3181 * as well as to serialize pageset setup for the new populated zone. 3182 */ 3183DEFINE_MUTEX(zonelists_mutex); 3184 3185/* return values int ....just for stop_machine() */ 3186static __init_refok int __build_all_zonelists(void *data) 3187{ 3188 int nid; 3189 int cpu; 3190 3191#ifdef CONFIG_NUMA 3192 memset(node_load, 0, sizeof(node_load)); 3193#endif 3194 for_each_online_node(nid) { 3195 pg_data_t *pgdat = NODE_DATA(nid); 3196 3197 build_zonelists(pgdat); 3198 build_zonelist_cache(pgdat); 3199 } 3200 3201 /* 3202 * Initialize the boot_pagesets that are going to be used 3203 * for bootstrapping processors. The real pagesets for 3204 * each zone will be allocated later when the per cpu 3205 * allocator is available. 3206 * 3207 * boot_pagesets are used also for bootstrapping offline 3208 * cpus if the system is already booted because the pagesets 3209 * are needed to initialize allocators on a specific cpu too. 3210 * F.e. the percpu allocator needs the page allocator which 3211 * needs the percpu allocator in order to allocate its pagesets 3212 * (a chicken-egg dilemma). 3213 */ 3214 for_each_possible_cpu(cpu) { 3215 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3216 3217#ifdef CONFIG_HAVE_MEMORYLESS_NODES 3218 /* 3219 * We now know the "local memory node" for each node-- 3220 * i.e., the node of the first zone in the generic zonelist. 3221 * Set up numa_mem percpu variable for on-line cpus. During 3222 * boot, only the boot cpu should be on-line; we'll init the 3223 * secondary cpus' numa_mem as they come on-line. During 3224 * node/memory hotplug, we'll fixup all on-line cpus. 3225 */ 3226 if (cpu_online(cpu)) 3227 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3228#endif 3229 } 3230 3231 return 0; 3232} 3233 3234/* 3235 * Called with zonelists_mutex held always 3236 * unless system_state == SYSTEM_BOOTING. 3237 */ 3238void __ref build_all_zonelists(void *data) 3239{ 3240 set_zonelist_order(); 3241 3242 if (system_state == SYSTEM_BOOTING) { 3243 __build_all_zonelists(NULL); 3244 mminit_verify_zonelist(); 3245 cpuset_init_current_mems_allowed(); 3246 } else { 3247 /* we have to stop all cpus to guarantee there is no user 3248 of zonelist */ 3249#ifdef CONFIG_MEMORY_HOTPLUG 3250 if (data) 3251 setup_zone_pageset((struct zone *)data); 3252#endif 3253 stop_machine(__build_all_zonelists, NULL, NULL); 3254 /* cpuset refresh routine should be here */ 3255 } 3256 vm_total_pages = nr_free_pagecache_pages(); 3257 /* 3258 * Disable grouping by mobility if the number of pages in the 3259 * system is too low to allow the mechanism to work. It would be 3260 * more accurate, but expensive to check per-zone. This check is 3261 * made on memory-hotadd so a system can start with mobility 3262 * disabled and enable it later 3263 */ 3264 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3265 page_group_by_mobility_disabled = 1; 3266 else 3267 page_group_by_mobility_disabled = 0; 3268 3269 printk("Built %i zonelists in %s order, mobility grouping %s. " 3270 "Total pages: %ld\n", 3271 nr_online_nodes, 3272 zonelist_order_name[current_zonelist_order], 3273 page_group_by_mobility_disabled ? "off" : "on", 3274 vm_total_pages); 3275#ifdef CONFIG_NUMA 3276 printk("Policy zone: %s\n", zone_names[policy_zone]); 3277#endif 3278} 3279 3280/* 3281 * Helper functions to size the waitqueue hash table. 3282 * Essentially these want to choose hash table sizes sufficiently 3283 * large so that collisions trying to wait on pages are rare. 3284 * But in fact, the number of active page waitqueues on typical 3285 * systems is ridiculously low, less than 200. So this is even 3286 * conservative, even though it seems large. 3287 * 3288 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3289 * waitqueues, i.e. the size of the waitq table given the number of pages. 3290 */ 3291#define PAGES_PER_WAITQUEUE 256 3292 3293#ifndef CONFIG_MEMORY_HOTPLUG 3294static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3295{ 3296 unsigned long size = 1; 3297 3298 pages /= PAGES_PER_WAITQUEUE; 3299 3300 while (size < pages) 3301 size <<= 1; 3302 3303 /* 3304 * Once we have dozens or even hundreds of threads sleeping 3305 * on IO we've got bigger problems than wait queue collision. 3306 * Limit the size of the wait table to a reasonable size. 3307 */ 3308 size = min(size, 4096UL); 3309 3310 return max(size, 4UL); 3311} 3312#else 3313/* 3314 * A zone's size might be changed by hot-add, so it is not possible to determine 3315 * a suitable size for its wait_table. So we use the maximum size now. 3316 * 3317 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3318 * 3319 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3320 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3321 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3322 * 3323 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3324 * or more by the traditional way. (See above). It equals: 3325 * 3326 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3327 * ia64(16K page size) : = ( 8G + 4M)byte. 3328 * powerpc (64K page size) : = (32G +16M)byte. 3329 */ 3330static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3331{ 3332 return 4096UL; 3333} 3334#endif 3335 3336/* 3337 * This is an integer logarithm so that shifts can be used later 3338 * to extract the more random high bits from the multiplicative 3339 * hash function before the remainder is taken. 3340 */ 3341static inline unsigned long wait_table_bits(unsigned long size) 3342{ 3343 return ffz(~size); 3344} 3345 3346#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3347 3348/* 3349 * Check if a pageblock contains reserved pages 3350 */ 3351static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 3352{ 3353 unsigned long pfn; 3354 3355 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3356 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 3357 return 1; 3358 } 3359 return 0; 3360} 3361 3362/* 3363 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3364 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3365 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3366 * higher will lead to a bigger reserve which will get freed as contiguous 3367 * blocks as reclaim kicks in 3368 */ 3369static void setup_zone_migrate_reserve(struct zone *zone) 3370{ 3371 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 3372 struct page *page; 3373 unsigned long block_migratetype; 3374 int reserve; 3375 3376 /* Get the start pfn, end pfn and the number of blocks to reserve */ 3377 start_pfn = zone->zone_start_pfn; 3378 end_pfn = start_pfn + zone->spanned_pages; 3379 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3380 pageblock_order; 3381 3382 /* 3383 * Reserve blocks are generally in place to help high-order atomic 3384 * allocations that are short-lived. A min_free_kbytes value that 3385 * would result in more than 2 reserve blocks for atomic allocations 3386 * is assumed to be in place to help anti-fragmentation for the 3387 * future allocation of hugepages at runtime. 3388 */ 3389 reserve = min(2, reserve); 3390 3391 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3392 if (!pfn_valid(pfn)) 3393 continue; 3394 page = pfn_to_page(pfn); 3395 3396 /* Watch out for overlapping nodes */ 3397 if (page_to_nid(page) != zone_to_nid(zone)) 3398 continue; 3399 3400 /* Blocks with reserved pages will never free, skip them. */ 3401 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 3402 if (pageblock_is_reserved(pfn, block_end_pfn)) 3403 continue; 3404 3405 block_migratetype = get_pageblock_migratetype(page); 3406 3407 /* If this block is reserved, account for it */ 3408 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 3409 reserve--; 3410 continue; 3411 } 3412 3413 /* Suitable for reserving if this block is movable */ 3414 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 3415 set_pageblock_migratetype(page, MIGRATE_RESERVE); 3416 move_freepages_block(zone, page, MIGRATE_RESERVE); 3417 reserve--; 3418 continue; 3419 } 3420 3421 /* 3422 * If the reserve is met and this is a previous reserved block, 3423 * take it back 3424 */ 3425 if (block_migratetype == MIGRATE_RESERVE) { 3426 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3427 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3428 } 3429 } 3430} 3431 3432/* 3433 * Initially all pages are reserved - free ones are freed 3434 * up by free_all_bootmem() once the early boot process is 3435 * done. Non-atomic initialization, single-pass. 3436 */ 3437void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3438 unsigned long start_pfn, enum memmap_context context) 3439{ 3440 struct page *page; 3441 unsigned long end_pfn = start_pfn + size; 3442 unsigned long pfn; 3443 struct zone *z; 3444 3445 if (highest_memmap_pfn < end_pfn - 1) 3446 highest_memmap_pfn = end_pfn - 1; 3447 3448 z = &NODE_DATA(nid)->node_zones[zone]; 3449 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3450 /* 3451 * There can be holes in boot-time mem_map[]s 3452 * handed to this function. They do not 3453 * exist on hotplugged memory. 3454 */ 3455 if (context == MEMMAP_EARLY) { 3456 if (!early_pfn_valid(pfn)) 3457 continue; 3458 if (!early_pfn_in_nid(pfn, nid)) 3459 continue; 3460 } 3461 page = pfn_to_page(pfn); 3462 set_page_links(page, zone, nid, pfn); 3463 mminit_verify_page_links(page, zone, nid, pfn); 3464 init_page_count(page); 3465 reset_page_mapcount(page); 3466 SetPageReserved(page); 3467 /* 3468 * Mark the block movable so that blocks are reserved for 3469 * movable at startup. This will force kernel allocations 3470 * to reserve their blocks rather than leaking throughout 3471 * the address space during boot when many long-lived 3472 * kernel allocations are made. Later some blocks near 3473 * the start are marked MIGRATE_RESERVE by 3474 * setup_zone_migrate_reserve() 3475 * 3476 * bitmap is created for zone's valid pfn range. but memmap 3477 * can be created for invalid pages (for alignment) 3478 * check here not to call set_pageblock_migratetype() against 3479 * pfn out of zone. 3480 */ 3481 if ((z->zone_start_pfn <= pfn) 3482 && (pfn < z->zone_start_pfn + z->spanned_pages) 3483 && !(pfn & (pageblock_nr_pages - 1))) 3484 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3485 3486 INIT_LIST_HEAD(&page->lru); 3487#ifdef WANT_PAGE_VIRTUAL 3488 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3489 if (!is_highmem_idx(zone)) 3490 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3491#endif 3492 } 3493} 3494 3495static void __meminit zone_init_free_lists(struct zone *zone) 3496{ 3497 int order, t; 3498 for_each_migratetype_order(order, t) { 3499 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3500 zone->free_area[order].nr_free = 0; 3501 } 3502} 3503 3504#ifndef __HAVE_ARCH_MEMMAP_INIT 3505#define memmap_init(size, nid, zone, start_pfn) \ 3506 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3507#endif 3508 3509static int zone_batchsize(struct zone *zone) 3510{ 3511#ifdef CONFIG_MMU 3512 int batch; 3513 3514 /* 3515 * The per-cpu-pages pools are set to around 1000th of the 3516 * size of the zone. But no more than 1/2 of a meg. 3517 * 3518 * OK, so we don't know how big the cache is. So guess. 3519 */ 3520 batch = zone->present_pages / 1024; 3521 if (batch * PAGE_SIZE > 512 * 1024) 3522 batch = (512 * 1024) / PAGE_SIZE; 3523 batch /= 4; /* We effectively *= 4 below */ 3524 if (batch < 1) 3525 batch = 1; 3526 3527 /* 3528 * Clamp the batch to a 2^n - 1 value. Having a power 3529 * of 2 value was found to be more likely to have 3530 * suboptimal cache aliasing properties in some cases. 3531 * 3532 * For example if 2 tasks are alternately allocating 3533 * batches of pages, one task can end up with a lot 3534 * of pages of one half of the possible page colors 3535 * and the other with pages of the other colors. 3536 */ 3537 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3538 3539 return batch; 3540 3541#else 3542 /* The deferral and batching of frees should be suppressed under NOMMU 3543 * conditions. 3544 * 3545 * The problem is that NOMMU needs to be able to allocate large chunks 3546 * of contiguous memory as there's no hardware page translation to 3547 * assemble apparent contiguous memory from discontiguous pages. 3548 * 3549 * Queueing large contiguous runs of pages for batching, however, 3550 * causes the pages to actually be freed in smaller chunks. As there 3551 * can be a significant delay between the individual batches being 3552 * recycled, this leads to the once large chunks of space being 3553 * fragmented and becoming unavailable for high-order allocations. 3554 */ 3555 return 0; 3556#endif 3557} 3558 3559static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3560{ 3561 struct per_cpu_pages *pcp; 3562 int migratetype; 3563 3564 memset(p, 0, sizeof(*p)); 3565 3566 pcp = &p->pcp; 3567 pcp->count = 0; 3568 pcp->high = 6 * batch; 3569 pcp->batch = max(1UL, 1 * batch); 3570 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3571 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3572} 3573 3574/* 3575 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3576 * to the value high for the pageset p. 3577 */ 3578 3579static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3580 unsigned long high) 3581{ 3582 struct per_cpu_pages *pcp; 3583 3584 pcp = &p->pcp; 3585 pcp->high = high; 3586 pcp->batch = max(1UL, high/4); 3587 if ((high/4) > (PAGE_SHIFT * 8)) 3588 pcp->batch = PAGE_SHIFT * 8; 3589} 3590 3591static void setup_zone_pageset(struct zone *zone) 3592{ 3593 int cpu; 3594 3595 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3596 3597 for_each_possible_cpu(cpu) { 3598 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3599 3600 setup_pageset(pcp, zone_batchsize(zone)); 3601 3602 if (percpu_pagelist_fraction) 3603 setup_pagelist_highmark(pcp, 3604 (zone->present_pages / 3605 percpu_pagelist_fraction)); 3606 } 3607} 3608 3609/* 3610 * Allocate per cpu pagesets and initialize them. 3611 * Before this call only boot pagesets were available. 3612 */ 3613void __init setup_per_cpu_pageset(void) 3614{ 3615 struct zone *zone; 3616 3617 for_each_populated_zone(zone) 3618 setup_zone_pageset(zone); 3619} 3620 3621static noinline __init_refok 3622int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3623{ 3624 int i; 3625 struct pglist_data *pgdat = zone->zone_pgdat; 3626 size_t alloc_size; 3627 3628 /* 3629 * The per-page waitqueue mechanism uses hashed waitqueues 3630 * per zone. 3631 */ 3632 zone->wait_table_hash_nr_entries = 3633 wait_table_hash_nr_entries(zone_size_pages); 3634 zone->wait_table_bits = 3635 wait_table_bits(zone->wait_table_hash_nr_entries); 3636 alloc_size = zone->wait_table_hash_nr_entries 3637 * sizeof(wait_queue_head_t); 3638 3639 if (!slab_is_available()) { 3640 zone->wait_table = (wait_queue_head_t *) 3641 alloc_bootmem_node_nopanic(pgdat, alloc_size); 3642 } else { 3643 /* 3644 * This case means that a zone whose size was 0 gets new memory 3645 * via memory hot-add. 3646 * But it may be the case that a new node was hot-added. In 3647 * this case vmalloc() will not be able to use this new node's 3648 * memory - this wait_table must be initialized to use this new 3649 * node itself as well. 3650 * To use this new node's memory, further consideration will be 3651 * necessary. 3652 */ 3653 zone->wait_table = vmalloc(alloc_size); 3654 } 3655 if (!zone->wait_table) 3656 return -ENOMEM; 3657 3658 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3659 init_waitqueue_head(zone->wait_table + i); 3660 3661 return 0; 3662} 3663 3664static int __zone_pcp_update(void *data) 3665{ 3666 struct zone *zone = data; 3667 int cpu; 3668 unsigned long batch = zone_batchsize(zone), flags; 3669 3670 for_each_possible_cpu(cpu) { 3671 struct per_cpu_pageset *pset; 3672 struct per_cpu_pages *pcp; 3673 3674 pset = per_cpu_ptr(zone->pageset, cpu); 3675 pcp = &pset->pcp; 3676 3677 local_irq_save(flags); 3678 free_pcppages_bulk(zone, pcp->count, pcp); 3679 setup_pageset(pset, batch); 3680 local_irq_restore(flags); 3681 } 3682 return 0; 3683} 3684 3685void zone_pcp_update(struct zone *zone) 3686{ 3687 stop_machine(__zone_pcp_update, zone, NULL); 3688} 3689 3690static __meminit void zone_pcp_init(struct zone *zone) 3691{ 3692 /* 3693 * per cpu subsystem is not up at this point. The following code 3694 * relies on the ability of the linker to provide the 3695 * offset of a (static) per cpu variable into the per cpu area. 3696 */ 3697 zone->pageset = &boot_pageset; 3698 3699 if (zone->present_pages) 3700 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3701 zone->name, zone->present_pages, 3702 zone_batchsize(zone)); 3703} 3704 3705__meminit int init_currently_empty_zone(struct zone *zone, 3706 unsigned long zone_start_pfn, 3707 unsigned long size, 3708 enum memmap_context context) 3709{ 3710 struct pglist_data *pgdat = zone->zone_pgdat; 3711 int ret; 3712 ret = zone_wait_table_init(zone, size); 3713 if (ret) 3714 return ret; 3715 pgdat->nr_zones = zone_idx(zone) + 1; 3716 3717 zone->zone_start_pfn = zone_start_pfn; 3718 3719 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3720 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3721 pgdat->node_id, 3722 (unsigned long)zone_idx(zone), 3723 zone_start_pfn, (zone_start_pfn + size)); 3724 3725 zone_init_free_lists(zone); 3726 3727 return 0; 3728} 3729 3730#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3731/* 3732 * Basic iterator support. Return the first range of PFNs for a node 3733 * Note: nid == MAX_NUMNODES returns first region regardless of node 3734 */ 3735static int __meminit first_active_region_index_in_nid(int nid) 3736{ 3737 int i; 3738 3739 for (i = 0; i < nr_nodemap_entries; i++) 3740 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3741 return i; 3742 3743 return -1; 3744} 3745 3746/* 3747 * Basic iterator support. Return the next active range of PFNs for a node 3748 * Note: nid == MAX_NUMNODES returns next region regardless of node 3749 */ 3750static int __meminit next_active_region_index_in_nid(int index, int nid) 3751{ 3752 for (index = index + 1; index < nr_nodemap_entries; index++) 3753 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3754 return index; 3755 3756 return -1; 3757} 3758 3759#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3760/* 3761 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3762 * Architectures may implement their own version but if add_active_range() 3763 * was used and there are no special requirements, this is a convenient 3764 * alternative 3765 */ 3766int __meminit __early_pfn_to_nid(unsigned long pfn) 3767{ 3768 int i; 3769 3770 for (i = 0; i < nr_nodemap_entries; i++) { 3771 unsigned long start_pfn = early_node_map[i].start_pfn; 3772 unsigned long end_pfn = early_node_map[i].end_pfn; 3773 3774 if (start_pfn <= pfn && pfn < end_pfn) 3775 return early_node_map[i].nid; 3776 } 3777 /* This is a memory hole */ 3778 return -1; 3779} 3780#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3781 3782int __meminit early_pfn_to_nid(unsigned long pfn) 3783{ 3784 int nid; 3785 3786 nid = __early_pfn_to_nid(pfn); 3787 if (nid >= 0) 3788 return nid; 3789 /* just returns 0 */ 3790 return 0; 3791} 3792 3793#ifdef CONFIG_NODES_SPAN_OTHER_NODES 3794bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3795{ 3796 int nid; 3797 3798 nid = __early_pfn_to_nid(pfn); 3799 if (nid >= 0 && nid != node) 3800 return false; 3801 return true; 3802} 3803#endif 3804 3805/* Basic iterator support to walk early_node_map[] */ 3806#define for_each_active_range_index_in_nid(i, nid) \ 3807 for (i = first_active_region_index_in_nid(nid); i != -1; \ 3808 i = next_active_region_index_in_nid(i, nid)) 3809 3810/** 3811 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3812 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3813 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3814 * 3815 * If an architecture guarantees that all ranges registered with 3816 * add_active_ranges() contain no holes and may be freed, this 3817 * this function may be used instead of calling free_bootmem() manually. 3818 */ 3819void __init free_bootmem_with_active_regions(int nid, 3820 unsigned long max_low_pfn) 3821{ 3822 int i; 3823 3824 for_each_active_range_index_in_nid(i, nid) { 3825 unsigned long size_pages = 0; 3826 unsigned long end_pfn = early_node_map[i].end_pfn; 3827 3828 if (early_node_map[i].start_pfn >= max_low_pfn) 3829 continue; 3830 3831 if (end_pfn > max_low_pfn) 3832 end_pfn = max_low_pfn; 3833 3834 size_pages = end_pfn - early_node_map[i].start_pfn; 3835 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 3836 PFN_PHYS(early_node_map[i].start_pfn), 3837 size_pages << PAGE_SHIFT); 3838 } 3839} 3840 3841#ifdef CONFIG_HAVE_MEMBLOCK 3842/* 3843 * Basic iterator support. Return the last range of PFNs for a node 3844 * Note: nid == MAX_NUMNODES returns last region regardless of node 3845 */ 3846static int __meminit last_active_region_index_in_nid(int nid) 3847{ 3848 int i; 3849 3850 for (i = nr_nodemap_entries - 1; i >= 0; i--) 3851 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 3852 return i; 3853 3854 return -1; 3855} 3856 3857/* 3858 * Basic iterator support. Return the previous active range of PFNs for a node 3859 * Note: nid == MAX_NUMNODES returns next region regardless of node 3860 */ 3861static int __meminit previous_active_region_index_in_nid(int index, int nid) 3862{ 3863 for (index = index - 1; index >= 0; index--) 3864 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 3865 return index; 3866 3867 return -1; 3868} 3869 3870#define for_each_active_range_index_in_nid_reverse(i, nid) \ 3871 for (i = last_active_region_index_in_nid(nid); i != -1; \ 3872 i = previous_active_region_index_in_nid(i, nid)) 3873 3874u64 __init find_memory_core_early(int nid, u64 size, u64 align, 3875 u64 goal, u64 limit) 3876{ 3877 int i; 3878 3879 /* Need to go over early_node_map to find out good range for node */ 3880 for_each_active_range_index_in_nid_reverse(i, nid) { 3881 u64 addr; 3882 u64 ei_start, ei_last; 3883 u64 final_start, final_end; 3884 3885 ei_last = early_node_map[i].end_pfn; 3886 ei_last <<= PAGE_SHIFT; 3887 ei_start = early_node_map[i].start_pfn; 3888 ei_start <<= PAGE_SHIFT; 3889 3890 final_start = max(ei_start, goal); 3891 final_end = min(ei_last, limit); 3892 3893 if (final_start >= final_end) 3894 continue; 3895 3896 addr = memblock_find_in_range(final_start, final_end, size, align); 3897 3898 if (addr == MEMBLOCK_ERROR) 3899 continue; 3900 3901 return addr; 3902 } 3903 3904 return MEMBLOCK_ERROR; 3905} 3906#endif 3907 3908int __init add_from_early_node_map(struct range *range, int az, 3909 int nr_range, int nid) 3910{ 3911 int i; 3912 u64 start, end; 3913 3914 /* need to go over early_node_map to find out good range for node */ 3915 for_each_active_range_index_in_nid(i, nid) { 3916 start = early_node_map[i].start_pfn; 3917 end = early_node_map[i].end_pfn; 3918 nr_range = add_range(range, az, nr_range, start, end); 3919 } 3920 return nr_range; 3921} 3922 3923void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data) 3924{ 3925 int i; 3926 int ret; 3927 3928 for_each_active_range_index_in_nid(i, nid) { 3929 ret = work_fn(early_node_map[i].start_pfn, 3930 early_node_map[i].end_pfn, data); 3931 if (ret) 3932 break; 3933 } 3934} 3935/** 3936 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3937 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3938 * 3939 * If an architecture guarantees that all ranges registered with 3940 * add_active_ranges() contain no holes and may be freed, this 3941 * function may be used instead of calling memory_present() manually. 3942 */ 3943void __init sparse_memory_present_with_active_regions(int nid) 3944{ 3945 int i; 3946 3947 for_each_active_range_index_in_nid(i, nid) 3948 memory_present(early_node_map[i].nid, 3949 early_node_map[i].start_pfn, 3950 early_node_map[i].end_pfn); 3951} 3952 3953/** 3954 * get_pfn_range_for_nid - Return the start and end page frames for a node 3955 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3956 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3957 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3958 * 3959 * It returns the start and end page frame of a node based on information 3960 * provided by an arch calling add_active_range(). If called for a node 3961 * with no available memory, a warning is printed and the start and end 3962 * PFNs will be 0. 3963 */ 3964void __meminit get_pfn_range_for_nid(unsigned int nid, 3965 unsigned long *start_pfn, unsigned long *end_pfn) 3966{ 3967 int i; 3968 *start_pfn = -1UL; 3969 *end_pfn = 0; 3970 3971 for_each_active_range_index_in_nid(i, nid) { 3972 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3973 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3974 } 3975 3976 if (*start_pfn == -1UL) 3977 *start_pfn = 0; 3978} 3979 3980/* 3981 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3982 * assumption is made that zones within a node are ordered in monotonic 3983 * increasing memory addresses so that the "highest" populated zone is used 3984 */ 3985static void __init find_usable_zone_for_movable(void) 3986{ 3987 int zone_index; 3988 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3989 if (zone_index == ZONE_MOVABLE) 3990 continue; 3991 3992 if (arch_zone_highest_possible_pfn[zone_index] > 3993 arch_zone_lowest_possible_pfn[zone_index]) 3994 break; 3995 } 3996 3997 VM_BUG_ON(zone_index == -1); 3998 movable_zone = zone_index; 3999} 4000 4001/* 4002 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 4003 * because it is sized independent of architecture. Unlike the other zones, 4004 * the starting point for ZONE_MOVABLE is not fixed. It may be different 4005 * in each node depending on the size of each node and how evenly kernelcore 4006 * is distributed. This helper function adjusts the zone ranges 4007 * provided by the architecture for a given node by using the end of the 4008 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 4009 * zones within a node are in order of monotonic increases memory addresses 4010 */ 4011static void __meminit adjust_zone_range_for_zone_movable(int nid, 4012 unsigned long zone_type, 4013 unsigned long node_start_pfn, 4014 unsigned long node_end_pfn, 4015 unsigned long *zone_start_pfn, 4016 unsigned long *zone_end_pfn) 4017{ 4018 /* Only adjust if ZONE_MOVABLE is on this node */ 4019 if (zone_movable_pfn[nid]) { 4020 /* Size ZONE_MOVABLE */ 4021 if (zone_type == ZONE_MOVABLE) { 4022 *zone_start_pfn = zone_movable_pfn[nid]; 4023 *zone_end_pfn = min(node_end_pfn, 4024 arch_zone_highest_possible_pfn[movable_zone]); 4025 4026 /* Adjust for ZONE_MOVABLE starting within this range */ 4027 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 4028 *zone_end_pfn > zone_movable_pfn[nid]) { 4029 *zone_end_pfn = zone_movable_pfn[nid]; 4030 4031 /* Check if this whole range is within ZONE_MOVABLE */ 4032 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 4033 *zone_start_pfn = *zone_end_pfn; 4034 } 4035} 4036 4037/* 4038 * Return the number of pages a zone spans in a node, including holes 4039 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 4040 */ 4041static unsigned long __meminit zone_spanned_pages_in_node(int nid, 4042 unsigned long zone_type, 4043 unsigned long *ignored) 4044{ 4045 unsigned long node_start_pfn, node_end_pfn; 4046 unsigned long zone_start_pfn, zone_end_pfn; 4047 4048 /* Get the start and end of the node and zone */ 4049 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4050 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 4051 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 4052 adjust_zone_range_for_zone_movable(nid, zone_type, 4053 node_start_pfn, node_end_pfn, 4054 &zone_start_pfn, &zone_end_pfn); 4055 4056 /* Check that this node has pages within the zone's required range */ 4057 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 4058 return 0; 4059 4060 /* Move the zone boundaries inside the node if necessary */ 4061 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 4062 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 4063 4064 /* Return the spanned pages */ 4065 return zone_end_pfn - zone_start_pfn; 4066} 4067 4068/* 4069 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 4070 * then all holes in the requested range will be accounted for. 4071 */ 4072unsigned long __meminit __absent_pages_in_range(int nid, 4073 unsigned long range_start_pfn, 4074 unsigned long range_end_pfn) 4075{ 4076 int i = 0; 4077 unsigned long prev_end_pfn = 0, hole_pages = 0; 4078 unsigned long start_pfn; 4079 4080 /* Find the end_pfn of the first active range of pfns in the node */ 4081 i = first_active_region_index_in_nid(nid); 4082 if (i == -1) 4083 return 0; 4084 4085 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 4086 4087 /* Account for ranges before physical memory on this node */ 4088 if (early_node_map[i].start_pfn > range_start_pfn) 4089 hole_pages = prev_end_pfn - range_start_pfn; 4090 4091 /* Find all holes for the zone within the node */ 4092 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 4093 4094 /* No need to continue if prev_end_pfn is outside the zone */ 4095 if (prev_end_pfn >= range_end_pfn) 4096 break; 4097 4098 /* Make sure the end of the zone is not within the hole */ 4099 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 4100 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 4101 4102 /* Update the hole size cound and move on */ 4103 if (start_pfn > range_start_pfn) { 4104 BUG_ON(prev_end_pfn > start_pfn); 4105 hole_pages += start_pfn - prev_end_pfn; 4106 } 4107 prev_end_pfn = early_node_map[i].end_pfn; 4108 } 4109 4110 /* Account for ranges past physical memory on this node */ 4111 if (range_end_pfn > prev_end_pfn) 4112 hole_pages += range_end_pfn - 4113 max(range_start_pfn, prev_end_pfn); 4114 4115 return hole_pages; 4116} 4117 4118/** 4119 * absent_pages_in_range - Return number of page frames in holes within a range 4120 * @start_pfn: The start PFN to start searching for holes 4121 * @end_pfn: The end PFN to stop searching for holes 4122 * 4123 * It returns the number of pages frames in memory holes within a range. 4124 */ 4125unsigned long __init absent_pages_in_range(unsigned long start_pfn, 4126 unsigned long end_pfn) 4127{ 4128 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 4129} 4130 4131/* Return the number of page frames in holes in a zone on a node */ 4132static unsigned long __meminit zone_absent_pages_in_node(int nid, 4133 unsigned long zone_type, 4134 unsigned long *ignored) 4135{ 4136 unsigned long node_start_pfn, node_end_pfn; 4137 unsigned long zone_start_pfn, zone_end_pfn; 4138 4139 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4140 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 4141 node_start_pfn); 4142 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 4143 node_end_pfn); 4144 4145 adjust_zone_range_for_zone_movable(nid, zone_type, 4146 node_start_pfn, node_end_pfn, 4147 &zone_start_pfn, &zone_end_pfn); 4148 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 4149} 4150 4151#else 4152static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 4153 unsigned long zone_type, 4154 unsigned long *zones_size) 4155{ 4156 return zones_size[zone_type]; 4157} 4158 4159static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4160 unsigned long zone_type, 4161 unsigned long *zholes_size) 4162{ 4163 if (!zholes_size) 4164 return 0; 4165 4166 return zholes_size[zone_type]; 4167} 4168 4169#endif 4170 4171static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4172 unsigned long *zones_size, unsigned long *zholes_size) 4173{ 4174 unsigned long realtotalpages, totalpages = 0; 4175 enum zone_type i; 4176 4177 for (i = 0; i < MAX_NR_ZONES; i++) 4178 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4179 zones_size); 4180 pgdat->node_spanned_pages = totalpages; 4181 4182 realtotalpages = totalpages; 4183 for (i = 0; i < MAX_NR_ZONES; i++) 4184 realtotalpages -= 4185 zone_absent_pages_in_node(pgdat->node_id, i, 4186 zholes_size); 4187 pgdat->node_present_pages = realtotalpages; 4188 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4189 realtotalpages); 4190} 4191 4192#ifndef CONFIG_SPARSEMEM 4193/* 4194 * Calculate the size of the zone->blockflags rounded to an unsigned long 4195 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4196 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4197 * round what is now in bits to nearest long in bits, then return it in 4198 * bytes. 4199 */ 4200static unsigned long __init usemap_size(unsigned long zonesize) 4201{ 4202 unsigned long usemapsize; 4203 4204 usemapsize = roundup(zonesize, pageblock_nr_pages); 4205 usemapsize = usemapsize >> pageblock_order; 4206 usemapsize *= NR_PAGEBLOCK_BITS; 4207 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4208 4209 return usemapsize / 8; 4210} 4211 4212static void __init setup_usemap(struct pglist_data *pgdat, 4213 struct zone *zone, unsigned long zonesize) 4214{ 4215 unsigned long usemapsize = usemap_size(zonesize); 4216 zone->pageblock_flags = NULL; 4217 if (usemapsize) 4218 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, 4219 usemapsize); 4220} 4221#else 4222static inline void setup_usemap(struct pglist_data *pgdat, 4223 struct zone *zone, unsigned long zonesize) {} 4224#endif /* CONFIG_SPARSEMEM */ 4225 4226#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4227 4228/* Return a sensible default order for the pageblock size. */ 4229static inline int pageblock_default_order(void) 4230{ 4231 if (HPAGE_SHIFT > PAGE_SHIFT) 4232 return HUGETLB_PAGE_ORDER; 4233 4234 return MAX_ORDER-1; 4235} 4236 4237/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4238static inline void __init set_pageblock_order(unsigned int order) 4239{ 4240 /* Check that pageblock_nr_pages has not already been setup */ 4241 if (pageblock_order) 4242 return; 4243 4244 /* 4245 * Assume the largest contiguous order of interest is a huge page. 4246 * This value may be variable depending on boot parameters on IA64 4247 */ 4248 pageblock_order = order; 4249} 4250#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4251 4252/* 4253 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4254 * and pageblock_default_order() are unused as pageblock_order is set 4255 * at compile-time. See include/linux/pageblock-flags.h for the values of 4256 * pageblock_order based on the kernel config 4257 */ 4258static inline int pageblock_default_order(unsigned int order) 4259{ 4260 return MAX_ORDER-1; 4261} 4262#define set_pageblock_order(x) do {} while (0) 4263 4264#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4265 4266/* 4267 * Set up the zone data structures: 4268 * - mark all pages reserved 4269 * - mark all memory queues empty 4270 * - clear the memory bitmaps 4271 */ 4272static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4273 unsigned long *zones_size, unsigned long *zholes_size) 4274{ 4275 enum zone_type j; 4276 int nid = pgdat->node_id; 4277 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4278 int ret; 4279 4280 pgdat_resize_init(pgdat); 4281 pgdat->nr_zones = 0; 4282 init_waitqueue_head(&pgdat->kswapd_wait); 4283 pgdat->kswapd_max_order = 0; 4284 pgdat_page_cgroup_init(pgdat); 4285 4286 for (j = 0; j < MAX_NR_ZONES; j++) { 4287 struct zone *zone = pgdat->node_zones + j; 4288 unsigned long size, realsize, memmap_pages; 4289 enum lru_list l; 4290 4291 size = zone_spanned_pages_in_node(nid, j, zones_size); 4292 realsize = size - zone_absent_pages_in_node(nid, j, 4293 zholes_size); 4294 4295 /* 4296 * Adjust realsize so that it accounts for how much memory 4297 * is used by this zone for memmap. This affects the watermark 4298 * and per-cpu initialisations 4299 */ 4300 memmap_pages = 4301 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4302 if (realsize >= memmap_pages) { 4303 realsize -= memmap_pages; 4304 if (memmap_pages) 4305 printk(KERN_DEBUG 4306 " %s zone: %lu pages used for memmap\n", 4307 zone_names[j], memmap_pages); 4308 } else 4309 printk(KERN_WARNING 4310 " %s zone: %lu pages exceeds realsize %lu\n", 4311 zone_names[j], memmap_pages, realsize); 4312 4313 /* Account for reserved pages */ 4314 if (j == 0 && realsize > dma_reserve) { 4315 realsize -= dma_reserve; 4316 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4317 zone_names[0], dma_reserve); 4318 } 4319 4320 if (!is_highmem_idx(j)) 4321 nr_kernel_pages += realsize; 4322 nr_all_pages += realsize; 4323 4324 zone->spanned_pages = size; 4325 zone->present_pages = realsize; 4326#ifdef CONFIG_NUMA 4327 zone->node = nid; 4328 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4329 / 100; 4330 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4331#endif 4332 zone->name = zone_names[j]; 4333 spin_lock_init(&zone->lock); 4334 spin_lock_init(&zone->lru_lock); 4335 zone_seqlock_init(zone); 4336 zone->zone_pgdat = pgdat; 4337 4338 zone_pcp_init(zone); 4339 for_each_lru(l) 4340 INIT_LIST_HEAD(&zone->lru[l].list); 4341 zone->reclaim_stat.recent_rotated[0] = 0; 4342 zone->reclaim_stat.recent_rotated[1] = 0; 4343 zone->reclaim_stat.recent_scanned[0] = 0; 4344 zone->reclaim_stat.recent_scanned[1] = 0; 4345 zap_zone_vm_stats(zone); 4346 zone->flags = 0; 4347 if (!size) 4348 continue; 4349 4350 set_pageblock_order(pageblock_default_order()); 4351 setup_usemap(pgdat, zone, size); 4352 ret = init_currently_empty_zone(zone, zone_start_pfn, 4353 size, MEMMAP_EARLY); 4354 BUG_ON(ret); 4355 memmap_init(size, nid, j, zone_start_pfn); 4356 zone_start_pfn += size; 4357 } 4358} 4359 4360static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4361{ 4362 /* Skip empty nodes */ 4363 if (!pgdat->node_spanned_pages) 4364 return; 4365 4366#ifdef CONFIG_FLAT_NODE_MEM_MAP 4367 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4368 if (!pgdat->node_mem_map) { 4369 unsigned long size, start, end; 4370 struct page *map; 4371 4372 /* 4373 * The zone's endpoints aren't required to be MAX_ORDER 4374 * aligned but the node_mem_map endpoints must be in order 4375 * for the buddy allocator to function correctly. 4376 */ 4377 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4378 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4379 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4380 size = (end - start) * sizeof(struct page); 4381 map = alloc_remap(pgdat->node_id, size); 4382 if (!map) 4383 map = alloc_bootmem_node_nopanic(pgdat, size); 4384 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4385 } 4386#ifndef CONFIG_NEED_MULTIPLE_NODES 4387 /* 4388 * With no DISCONTIG, the global mem_map is just set as node 0's 4389 */ 4390 if (pgdat == NODE_DATA(0)) { 4391 mem_map = NODE_DATA(0)->node_mem_map; 4392#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4393 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4394 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4395#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4396 } 4397#endif 4398#endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4399} 4400 4401void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4402 unsigned long node_start_pfn, unsigned long *zholes_size) 4403{ 4404 pg_data_t *pgdat = NODE_DATA(nid); 4405 4406 pgdat->node_id = nid; 4407 pgdat->node_start_pfn = node_start_pfn; 4408 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4409 4410 alloc_node_mem_map(pgdat); 4411#ifdef CONFIG_FLAT_NODE_MEM_MAP 4412 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4413 nid, (unsigned long)pgdat, 4414 (unsigned long)pgdat->node_mem_map); 4415#endif 4416 4417 free_area_init_core(pgdat, zones_size, zholes_size); 4418} 4419 4420#ifdef CONFIG_ARCH_POPULATES_NODE_MAP 4421 4422#if MAX_NUMNODES > 1 4423/* 4424 * Figure out the number of possible node ids. 4425 */ 4426static void __init setup_nr_node_ids(void) 4427{ 4428 unsigned int node; 4429 unsigned int highest = 0; 4430 4431 for_each_node_mask(node, node_possible_map) 4432 highest = node; 4433 nr_node_ids = highest + 1; 4434} 4435#else 4436static inline void setup_nr_node_ids(void) 4437{ 4438} 4439#endif 4440 4441/** 4442 * add_active_range - Register a range of PFNs backed by physical memory 4443 * @nid: The node ID the range resides on 4444 * @start_pfn: The start PFN of the available physical memory 4445 * @end_pfn: The end PFN of the available physical memory 4446 * 4447 * These ranges are stored in an early_node_map[] and later used by 4448 * free_area_init_nodes() to calculate zone sizes and holes. If the 4449 * range spans a memory hole, it is up to the architecture to ensure 4450 * the memory is not freed by the bootmem allocator. If possible 4451 * the range being registered will be merged with existing ranges. 4452 */ 4453void __init add_active_range(unsigned int nid, unsigned long start_pfn, 4454 unsigned long end_pfn) 4455{ 4456 int i; 4457 4458 mminit_dprintk(MMINIT_TRACE, "memory_register", 4459 "Entering add_active_range(%d, %#lx, %#lx) " 4460 "%d entries of %d used\n", 4461 nid, start_pfn, end_pfn, 4462 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 4463 4464 mminit_validate_memmodel_limits(&start_pfn, &end_pfn); 4465 4466 /* Merge with existing active regions if possible */ 4467 for (i = 0; i < nr_nodemap_entries; i++) { 4468 if (early_node_map[i].nid != nid) 4469 continue; 4470 4471 /* Skip if an existing region covers this new one */ 4472 if (start_pfn >= early_node_map[i].start_pfn && 4473 end_pfn <= early_node_map[i].end_pfn) 4474 return; 4475 4476 /* Merge forward if suitable */ 4477 if (start_pfn <= early_node_map[i].end_pfn && 4478 end_pfn > early_node_map[i].end_pfn) { 4479 early_node_map[i].end_pfn = end_pfn; 4480 return; 4481 } 4482 4483 /* Merge backward if suitable */ 4484 if (start_pfn < early_node_map[i].start_pfn && 4485 end_pfn >= early_node_map[i].start_pfn) { 4486 early_node_map[i].start_pfn = start_pfn; 4487 return; 4488 } 4489 } 4490 4491 /* Check that early_node_map is large enough */ 4492 if (i >= MAX_ACTIVE_REGIONS) { 4493 printk(KERN_CRIT "More than %d memory regions, truncating\n", 4494 MAX_ACTIVE_REGIONS); 4495 return; 4496 } 4497 4498 early_node_map[i].nid = nid; 4499 early_node_map[i].start_pfn = start_pfn; 4500 early_node_map[i].end_pfn = end_pfn; 4501 nr_nodemap_entries = i + 1; 4502} 4503 4504/** 4505 * remove_active_range - Shrink an existing registered range of PFNs 4506 * @nid: The node id the range is on that should be shrunk 4507 * @start_pfn: The new PFN of the range 4508 * @end_pfn: The new PFN of the range 4509 * 4510 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 4511 * The map is kept near the end physical page range that has already been 4512 * registered. This function allows an arch to shrink an existing registered 4513 * range. 4514 */ 4515void __init remove_active_range(unsigned int nid, unsigned long start_pfn, 4516 unsigned long end_pfn) 4517{ 4518 int i, j; 4519 int removed = 0; 4520 4521 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n", 4522 nid, start_pfn, end_pfn); 4523 4524 /* Find the old active region end and shrink */ 4525 for_each_active_range_index_in_nid(i, nid) { 4526 if (early_node_map[i].start_pfn >= start_pfn && 4527 early_node_map[i].end_pfn <= end_pfn) { 4528 /* clear it */ 4529 early_node_map[i].start_pfn = 0; 4530 early_node_map[i].end_pfn = 0; 4531 removed = 1; 4532 continue; 4533 } 4534 if (early_node_map[i].start_pfn < start_pfn && 4535 early_node_map[i].end_pfn > start_pfn) { 4536 unsigned long temp_end_pfn = early_node_map[i].end_pfn; 4537 early_node_map[i].end_pfn = start_pfn; 4538 if (temp_end_pfn > end_pfn) 4539 add_active_range(nid, end_pfn, temp_end_pfn); 4540 continue; 4541 } 4542 if (early_node_map[i].start_pfn >= start_pfn && 4543 early_node_map[i].end_pfn > end_pfn && 4544 early_node_map[i].start_pfn < end_pfn) { 4545 early_node_map[i].start_pfn = end_pfn; 4546 continue; 4547 } 4548 } 4549 4550 if (!removed) 4551 return; 4552 4553 /* remove the blank ones */ 4554 for (i = nr_nodemap_entries - 1; i > 0; i--) { 4555 if (early_node_map[i].nid != nid) 4556 continue; 4557 if (early_node_map[i].end_pfn) 4558 continue; 4559 /* we found it, get rid of it */ 4560 for (j = i; j < nr_nodemap_entries - 1; j++) 4561 memcpy(&early_node_map[j], &early_node_map[j+1], 4562 sizeof(early_node_map[j])); 4563 j = nr_nodemap_entries - 1; 4564 memset(&early_node_map[j], 0, sizeof(early_node_map[j])); 4565 nr_nodemap_entries--; 4566 } 4567} 4568 4569/** 4570 * remove_all_active_ranges - Remove all currently registered regions 4571 * 4572 * During discovery, it may be found that a table like SRAT is invalid 4573 * and an alternative discovery method must be used. This function removes 4574 * all currently registered regions. 4575 */ 4576void __init remove_all_active_ranges(void) 4577{ 4578 memset(early_node_map, 0, sizeof(early_node_map)); 4579 nr_nodemap_entries = 0; 4580} 4581 4582/* Compare two active node_active_regions */ 4583static int __init cmp_node_active_region(const void *a, const void *b) 4584{ 4585 struct node_active_region *arange = (struct node_active_region *)a; 4586 struct node_active_region *brange = (struct node_active_region *)b; 4587 4588 /* Done this way to avoid overflows */ 4589 if (arange->start_pfn > brange->start_pfn) 4590 return 1; 4591 if (arange->start_pfn < brange->start_pfn) 4592 return -1; 4593 4594 return 0; 4595} 4596 4597/* sort the node_map by start_pfn */ 4598void __init sort_node_map(void) 4599{ 4600 sort(early_node_map, (size_t)nr_nodemap_entries, 4601 sizeof(struct node_active_region), 4602 cmp_node_active_region, NULL); 4603} 4604 4605/** 4606 * node_map_pfn_alignment - determine the maximum internode alignment 4607 * 4608 * This function should be called after node map is populated and sorted. 4609 * It calculates the maximum power of two alignment which can distinguish 4610 * all the nodes. 4611 * 4612 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 4613 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 4614 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 4615 * shifted, 1GiB is enough and this function will indicate so. 4616 * 4617 * This is used to test whether pfn -> nid mapping of the chosen memory 4618 * model has fine enough granularity to avoid incorrect mapping for the 4619 * populated node map. 4620 * 4621 * Returns the determined alignment in pfn's. 0 if there is no alignment 4622 * requirement (single node). 4623 */ 4624unsigned long __init node_map_pfn_alignment(void) 4625{ 4626 unsigned long accl_mask = 0, last_end = 0; 4627 int last_nid = -1; 4628 int i; 4629 4630 for_each_active_range_index_in_nid(i, MAX_NUMNODES) { 4631 int nid = early_node_map[i].nid; 4632 unsigned long start = early_node_map[i].start_pfn; 4633 unsigned long end = early_node_map[i].end_pfn; 4634 unsigned long mask; 4635 4636 if (!start || last_nid < 0 || last_nid == nid) { 4637 last_nid = nid; 4638 last_end = end; 4639 continue; 4640 } 4641 4642 /* 4643 * Start with a mask granular enough to pin-point to the 4644 * start pfn and tick off bits one-by-one until it becomes 4645 * too coarse to separate the current node from the last. 4646 */ 4647 mask = ~((1 << __ffs(start)) - 1); 4648 while (mask && last_end <= (start & (mask << 1))) 4649 mask <<= 1; 4650 4651 /* accumulate all internode masks */ 4652 accl_mask |= mask; 4653 } 4654 4655 /* convert mask to number of pages */ 4656 return ~accl_mask + 1; 4657} 4658 4659/* Find the lowest pfn for a node */ 4660static unsigned long __init find_min_pfn_for_node(int nid) 4661{ 4662 int i; 4663 unsigned long min_pfn = ULONG_MAX; 4664 4665 /* Assuming a sorted map, the first range found has the starting pfn */ 4666 for_each_active_range_index_in_nid(i, nid) 4667 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 4668 4669 if (min_pfn == ULONG_MAX) { 4670 printk(KERN_WARNING 4671 "Could not find start_pfn for node %d\n", nid); 4672 return 0; 4673 } 4674 4675 return min_pfn; 4676} 4677 4678/** 4679 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4680 * 4681 * It returns the minimum PFN based on information provided via 4682 * add_active_range(). 4683 */ 4684unsigned long __init find_min_pfn_with_active_regions(void) 4685{ 4686 return find_min_pfn_for_node(MAX_NUMNODES); 4687} 4688 4689/* 4690 * early_calculate_totalpages() 4691 * Sum pages in active regions for movable zone. 4692 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4693 */ 4694static unsigned long __init early_calculate_totalpages(void) 4695{ 4696 int i; 4697 unsigned long totalpages = 0; 4698 4699 for (i = 0; i < nr_nodemap_entries; i++) { 4700 unsigned long pages = early_node_map[i].end_pfn - 4701 early_node_map[i].start_pfn; 4702 totalpages += pages; 4703 if (pages) 4704 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 4705 } 4706 return totalpages; 4707} 4708 4709/* 4710 * Find the PFN the Movable zone begins in each node. Kernel memory 4711 * is spread evenly between nodes as long as the nodes have enough 4712 * memory. When they don't, some nodes will have more kernelcore than 4713 * others 4714 */ 4715static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4716{ 4717 int i, nid; 4718 unsigned long usable_startpfn; 4719 unsigned long kernelcore_node, kernelcore_remaining; 4720 /* save the state before borrow the nodemask */ 4721 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4722 unsigned long totalpages = early_calculate_totalpages(); 4723 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4724 4725 /* 4726 * If movablecore was specified, calculate what size of 4727 * kernelcore that corresponds so that memory usable for 4728 * any allocation type is evenly spread. If both kernelcore 4729 * and movablecore are specified, then the value of kernelcore 4730 * will be used for required_kernelcore if it's greater than 4731 * what movablecore would have allowed. 4732 */ 4733 if (required_movablecore) { 4734 unsigned long corepages; 4735 4736 /* 4737 * Round-up so that ZONE_MOVABLE is at least as large as what 4738 * was requested by the user 4739 */ 4740 required_movablecore = 4741 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4742 corepages = totalpages - required_movablecore; 4743 4744 required_kernelcore = max(required_kernelcore, corepages); 4745 } 4746 4747 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4748 if (!required_kernelcore) 4749 goto out; 4750 4751 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4752 find_usable_zone_for_movable(); 4753 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4754 4755restart: 4756 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4757 kernelcore_node = required_kernelcore / usable_nodes; 4758 for_each_node_state(nid, N_HIGH_MEMORY) { 4759 /* 4760 * Recalculate kernelcore_node if the division per node 4761 * now exceeds what is necessary to satisfy the requested 4762 * amount of memory for the kernel 4763 */ 4764 if (required_kernelcore < kernelcore_node) 4765 kernelcore_node = required_kernelcore / usable_nodes; 4766 4767 /* 4768 * As the map is walked, we track how much memory is usable 4769 * by the kernel using kernelcore_remaining. When it is 4770 * 0, the rest of the node is usable by ZONE_MOVABLE 4771 */ 4772 kernelcore_remaining = kernelcore_node; 4773 4774 /* Go through each range of PFNs within this node */ 4775 for_each_active_range_index_in_nid(i, nid) { 4776 unsigned long start_pfn, end_pfn; 4777 unsigned long size_pages; 4778 4779 start_pfn = max(early_node_map[i].start_pfn, 4780 zone_movable_pfn[nid]); 4781 end_pfn = early_node_map[i].end_pfn; 4782 if (start_pfn >= end_pfn) 4783 continue; 4784 4785 /* Account for what is only usable for kernelcore */ 4786 if (start_pfn < usable_startpfn) { 4787 unsigned long kernel_pages; 4788 kernel_pages = min(end_pfn, usable_startpfn) 4789 - start_pfn; 4790 4791 kernelcore_remaining -= min(kernel_pages, 4792 kernelcore_remaining); 4793 required_kernelcore -= min(kernel_pages, 4794 required_kernelcore); 4795 4796 /* Continue if range is now fully accounted */ 4797 if (end_pfn <= usable_startpfn) { 4798 4799 /* 4800 * Push zone_movable_pfn to the end so 4801 * that if we have to rebalance 4802 * kernelcore across nodes, we will 4803 * not double account here 4804 */ 4805 zone_movable_pfn[nid] = end_pfn; 4806 continue; 4807 } 4808 start_pfn = usable_startpfn; 4809 } 4810 4811 /* 4812 * The usable PFN range for ZONE_MOVABLE is from 4813 * start_pfn->end_pfn. Calculate size_pages as the 4814 * number of pages used as kernelcore 4815 */ 4816 size_pages = end_pfn - start_pfn; 4817 if (size_pages > kernelcore_remaining) 4818 size_pages = kernelcore_remaining; 4819 zone_movable_pfn[nid] = start_pfn + size_pages; 4820 4821 /* 4822 * Some kernelcore has been met, update counts and 4823 * break if the kernelcore for this node has been 4824 * satisified 4825 */ 4826 required_kernelcore -= min(required_kernelcore, 4827 size_pages); 4828 kernelcore_remaining -= size_pages; 4829 if (!kernelcore_remaining) 4830 break; 4831 } 4832 } 4833 4834 /* 4835 * If there is still required_kernelcore, we do another pass with one 4836 * less node in the count. This will push zone_movable_pfn[nid] further 4837 * along on the nodes that still have memory until kernelcore is 4838 * satisified 4839 */ 4840 usable_nodes--; 4841 if (usable_nodes && required_kernelcore > usable_nodes) 4842 goto restart; 4843 4844 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4845 for (nid = 0; nid < MAX_NUMNODES; nid++) 4846 zone_movable_pfn[nid] = 4847 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4848 4849out: 4850 /* restore the node_state */ 4851 node_states[N_HIGH_MEMORY] = saved_node_state; 4852} 4853 4854/* Any regular memory on that node ? */ 4855static void check_for_regular_memory(pg_data_t *pgdat) 4856{ 4857#ifdef CONFIG_HIGHMEM 4858 enum zone_type zone_type; 4859 4860 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4861 struct zone *zone = &pgdat->node_zones[zone_type]; 4862 if (zone->present_pages) 4863 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4864 } 4865#endif 4866} 4867 4868/** 4869 * free_area_init_nodes - Initialise all pg_data_t and zone data 4870 * @max_zone_pfn: an array of max PFNs for each zone 4871 * 4872 * This will call free_area_init_node() for each active node in the system. 4873 * Using the page ranges provided by add_active_range(), the size of each 4874 * zone in each node and their holes is calculated. If the maximum PFN 4875 * between two adjacent zones match, it is assumed that the zone is empty. 4876 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4877 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4878 * starts where the previous one ended. For example, ZONE_DMA32 starts 4879 * at arch_max_dma_pfn. 4880 */ 4881void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4882{ 4883 unsigned long nid; 4884 int i; 4885 4886 /* Sort early_node_map as initialisation assumes it is sorted */ 4887 sort_node_map(); 4888 4889 /* Record where the zone boundaries are */ 4890 memset(arch_zone_lowest_possible_pfn, 0, 4891 sizeof(arch_zone_lowest_possible_pfn)); 4892 memset(arch_zone_highest_possible_pfn, 0, 4893 sizeof(arch_zone_highest_possible_pfn)); 4894 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4895 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4896 for (i = 1; i < MAX_NR_ZONES; i++) { 4897 if (i == ZONE_MOVABLE) 4898 continue; 4899 arch_zone_lowest_possible_pfn[i] = 4900 arch_zone_highest_possible_pfn[i-1]; 4901 arch_zone_highest_possible_pfn[i] = 4902 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4903 } 4904 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4905 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4906 4907 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4908 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4909 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4910 4911 /* Print out the zone ranges */ 4912 printk("Zone PFN ranges:\n"); 4913 for (i = 0; i < MAX_NR_ZONES; i++) { 4914 if (i == ZONE_MOVABLE) 4915 continue; 4916 printk(" %-8s ", zone_names[i]); 4917 if (arch_zone_lowest_possible_pfn[i] == 4918 arch_zone_highest_possible_pfn[i]) 4919 printk("empty\n"); 4920 else 4921 printk("%0#10lx -> %0#10lx\n", 4922 arch_zone_lowest_possible_pfn[i], 4923 arch_zone_highest_possible_pfn[i]); 4924 } 4925 4926 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4927 printk("Movable zone start PFN for each node\n"); 4928 for (i = 0; i < MAX_NUMNODES; i++) { 4929 if (zone_movable_pfn[i]) 4930 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4931 } 4932 4933 /* Print out the early_node_map[] */ 4934 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 4935 for (i = 0; i < nr_nodemap_entries; i++) 4936 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid, 4937 early_node_map[i].start_pfn, 4938 early_node_map[i].end_pfn); 4939 4940 /* Initialise every node */ 4941 mminit_verify_pageflags_layout(); 4942 setup_nr_node_ids(); 4943 for_each_online_node(nid) { 4944 pg_data_t *pgdat = NODE_DATA(nid); 4945 free_area_init_node(nid, NULL, 4946 find_min_pfn_for_node(nid), NULL); 4947 4948 /* Any memory on that node */ 4949 if (pgdat->node_present_pages) 4950 node_set_state(nid, N_HIGH_MEMORY); 4951 check_for_regular_memory(pgdat); 4952 } 4953} 4954 4955static int __init cmdline_parse_core(char *p, unsigned long *core) 4956{ 4957 unsigned long long coremem; 4958 if (!p) 4959 return -EINVAL; 4960 4961 coremem = memparse(p, &p); 4962 *core = coremem >> PAGE_SHIFT; 4963 4964 /* Paranoid check that UL is enough for the coremem value */ 4965 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4966 4967 return 0; 4968} 4969 4970/* 4971 * kernelcore=size sets the amount of memory for use for allocations that 4972 * cannot be reclaimed or migrated. 4973 */ 4974static int __init cmdline_parse_kernelcore(char *p) 4975{ 4976 return cmdline_parse_core(p, &required_kernelcore); 4977} 4978 4979/* 4980 * movablecore=size sets the amount of memory for use for allocations that 4981 * can be reclaimed or migrated. 4982 */ 4983static int __init cmdline_parse_movablecore(char *p) 4984{ 4985 return cmdline_parse_core(p, &required_movablecore); 4986} 4987 4988early_param("kernelcore", cmdline_parse_kernelcore); 4989early_param("movablecore", cmdline_parse_movablecore); 4990 4991#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 4992 4993/** 4994 * set_dma_reserve - set the specified number of pages reserved in the first zone 4995 * @new_dma_reserve: The number of pages to mark reserved 4996 * 4997 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4998 * In the DMA zone, a significant percentage may be consumed by kernel image 4999 * and other unfreeable allocations which can skew the watermarks badly. This 5000 * function may optionally be used to account for unfreeable pages in the 5001 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 5002 * smaller per-cpu batchsize. 5003 */ 5004void __init set_dma_reserve(unsigned long new_dma_reserve) 5005{ 5006 dma_reserve = new_dma_reserve; 5007} 5008 5009void __init free_area_init(unsigned long *zones_size) 5010{ 5011 free_area_init_node(0, zones_size, 5012 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 5013} 5014 5015static int page_alloc_cpu_notify(struct notifier_block *self, 5016 unsigned long action, void *hcpu) 5017{ 5018 int cpu = (unsigned long)hcpu; 5019 5020 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 5021 drain_pages(cpu); 5022 5023 /* 5024 * Spill the event counters of the dead processor 5025 * into the current processors event counters. 5026 * This artificially elevates the count of the current 5027 * processor. 5028 */ 5029 vm_events_fold_cpu(cpu); 5030 5031 /* 5032 * Zero the differential counters of the dead processor 5033 * so that the vm statistics are consistent. 5034 * 5035 * This is only okay since the processor is dead and cannot 5036 * race with what we are doing. 5037 */ 5038 refresh_cpu_vm_stats(cpu); 5039 } 5040 return NOTIFY_OK; 5041} 5042 5043void __init page_alloc_init(void) 5044{ 5045 hotcpu_notifier(page_alloc_cpu_notify, 0); 5046} 5047 5048/* 5049 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 5050 * or min_free_kbytes changes. 5051 */ 5052static void calculate_totalreserve_pages(void) 5053{ 5054 struct pglist_data *pgdat; 5055 unsigned long reserve_pages = 0; 5056 enum zone_type i, j; 5057 5058 for_each_online_pgdat(pgdat) { 5059 for (i = 0; i < MAX_NR_ZONES; i++) { 5060 struct zone *zone = pgdat->node_zones + i; 5061 unsigned long max = 0; 5062 5063 /* Find valid and maximum lowmem_reserve in the zone */ 5064 for (j = i; j < MAX_NR_ZONES; j++) { 5065 if (zone->lowmem_reserve[j] > max) 5066 max = zone->lowmem_reserve[j]; 5067 } 5068 5069 /* we treat the high watermark as reserved pages. */ 5070 max += high_wmark_pages(zone); 5071 5072 if (max > zone->present_pages) 5073 max = zone->present_pages; 5074 reserve_pages += max; 5075 } 5076 } 5077 totalreserve_pages = reserve_pages; 5078} 5079 5080/* 5081 * setup_per_zone_lowmem_reserve - called whenever 5082 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 5083 * has a correct pages reserved value, so an adequate number of 5084 * pages are left in the zone after a successful __alloc_pages(). 5085 */ 5086static void setup_per_zone_lowmem_reserve(void) 5087{ 5088 struct pglist_data *pgdat; 5089 enum zone_type j, idx; 5090 5091 for_each_online_pgdat(pgdat) { 5092 for (j = 0; j < MAX_NR_ZONES; j++) { 5093 struct zone *zone = pgdat->node_zones + j; 5094 unsigned long present_pages = zone->present_pages; 5095 5096 zone->lowmem_reserve[j] = 0; 5097 5098 idx = j; 5099 while (idx) { 5100 struct zone *lower_zone; 5101 5102 idx--; 5103 5104 if (sysctl_lowmem_reserve_ratio[idx] < 1) 5105 sysctl_lowmem_reserve_ratio[idx] = 1; 5106 5107 lower_zone = pgdat->node_zones + idx; 5108 lower_zone->lowmem_reserve[j] = present_pages / 5109 sysctl_lowmem_reserve_ratio[idx]; 5110 present_pages += lower_zone->present_pages; 5111 } 5112 } 5113 } 5114 5115 /* update totalreserve_pages */ 5116 calculate_totalreserve_pages(); 5117} 5118 5119/** 5120 * setup_per_zone_wmarks - called when min_free_kbytes changes 5121 * or when memory is hot-{added|removed} 5122 * 5123 * Ensures that the watermark[min,low,high] values for each zone are set 5124 * correctly with respect to min_free_kbytes. 5125 */ 5126void setup_per_zone_wmarks(void) 5127{ 5128 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 5129 unsigned long lowmem_pages = 0; 5130 struct zone *zone; 5131 unsigned long flags; 5132 5133 /* Calculate total number of !ZONE_HIGHMEM pages */ 5134 for_each_zone(zone) { 5135 if (!is_highmem(zone)) 5136 lowmem_pages += zone->present_pages; 5137 } 5138 5139 for_each_zone(zone) { 5140 u64 tmp; 5141 5142 spin_lock_irqsave(&zone->lock, flags); 5143 tmp = (u64)pages_min * zone->present_pages; 5144 do_div(tmp, lowmem_pages); 5145 if (is_highmem(zone)) { 5146 /* 5147 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 5148 * need highmem pages, so cap pages_min to a small 5149 * value here. 5150 * 5151 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 5152 * deltas controls asynch page reclaim, and so should 5153 * not be capped for highmem. 5154 */ 5155 int min_pages; 5156 5157 min_pages = zone->present_pages / 1024; 5158 if (min_pages < SWAP_CLUSTER_MAX) 5159 min_pages = SWAP_CLUSTER_MAX; 5160 if (min_pages > 128) 5161 min_pages = 128; 5162 zone->watermark[WMARK_MIN] = min_pages; 5163 } else { 5164 /* 5165 * If it's a lowmem zone, reserve a number of pages 5166 * proportionate to the zone's size. 5167 */ 5168 zone->watermark[WMARK_MIN] = tmp; 5169 } 5170 5171 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 5172 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 5173 setup_zone_migrate_reserve(zone); 5174 spin_unlock_irqrestore(&zone->lock, flags); 5175 } 5176 5177 /* update totalreserve_pages */ 5178 calculate_totalreserve_pages(); 5179} 5180 5181/* 5182 * The inactive anon list should be small enough that the VM never has to 5183 * do too much work, but large enough that each inactive page has a chance 5184 * to be referenced again before it is swapped out. 5185 * 5186 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 5187 * INACTIVE_ANON pages on this zone's LRU, maintained by the 5188 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 5189 * the anonymous pages are kept on the inactive list. 5190 * 5191 * total target max 5192 * memory ratio inactive anon 5193 * ------------------------------------- 5194 * 10MB 1 5MB 5195 * 100MB 1 50MB 5196 * 1GB 3 250MB 5197 * 10GB 10 0.9GB 5198 * 100GB 31 3GB 5199 * 1TB 101 10GB 5200 * 10TB 320 32GB 5201 */ 5202static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 5203{ 5204 unsigned int gb, ratio; 5205 5206 /* Zone size in gigabytes */ 5207 gb = zone->present_pages >> (30 - PAGE_SHIFT); 5208 if (gb) 5209 ratio = int_sqrt(10 * gb); 5210 else 5211 ratio = 1; 5212 5213 zone->inactive_ratio = ratio; 5214} 5215 5216static void __meminit setup_per_zone_inactive_ratio(void) 5217{ 5218 struct zone *zone; 5219 5220 for_each_zone(zone) 5221 calculate_zone_inactive_ratio(zone); 5222} 5223 5224/* 5225 * Initialise min_free_kbytes. 5226 * 5227 * For small machines we want it small (128k min). For large machines 5228 * we want it large (64MB max). But it is not linear, because network 5229 * bandwidth does not increase linearly with machine size. We use 5230 * 5231 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5232 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5233 * 5234 * which yields 5235 * 5236 * 16MB: 512k 5237 * 32MB: 724k 5238 * 64MB: 1024k 5239 * 128MB: 1448k 5240 * 256MB: 2048k 5241 * 512MB: 2896k 5242 * 1024MB: 4096k 5243 * 2048MB: 5792k 5244 * 4096MB: 8192k 5245 * 8192MB: 11584k 5246 * 16384MB: 16384k 5247 */ 5248int __meminit init_per_zone_wmark_min(void) 5249{ 5250 unsigned long lowmem_kbytes; 5251 5252 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5253 5254 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5255 if (min_free_kbytes < 128) 5256 min_free_kbytes = 128; 5257 if (min_free_kbytes > 65536) 5258 min_free_kbytes = 65536; 5259 setup_per_zone_wmarks(); 5260 refresh_zone_stat_thresholds(); 5261 setup_per_zone_lowmem_reserve(); 5262 setup_per_zone_inactive_ratio(); 5263 return 0; 5264} 5265module_init(init_per_zone_wmark_min) 5266 5267/* 5268 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5269 * that we can call two helper functions whenever min_free_kbytes 5270 * changes. 5271 */ 5272int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5273 void __user *buffer, size_t *length, loff_t *ppos) 5274{ 5275 proc_dointvec(table, write, buffer, length, ppos); 5276 if (write) 5277 setup_per_zone_wmarks(); 5278 return 0; 5279} 5280 5281#ifdef CONFIG_NUMA 5282int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5283 void __user *buffer, size_t *length, loff_t *ppos) 5284{ 5285 struct zone *zone; 5286 int rc; 5287 5288 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5289 if (rc) 5290 return rc; 5291 5292 for_each_zone(zone) 5293 zone->min_unmapped_pages = (zone->present_pages * 5294 sysctl_min_unmapped_ratio) / 100; 5295 return 0; 5296} 5297 5298int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5299 void __user *buffer, size_t *length, loff_t *ppos) 5300{ 5301 struct zone *zone; 5302 int rc; 5303 5304 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5305 if (rc) 5306 return rc; 5307 5308 for_each_zone(zone) 5309 zone->min_slab_pages = (zone->present_pages * 5310 sysctl_min_slab_ratio) / 100; 5311 return 0; 5312} 5313#endif 5314 5315/* 5316 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5317 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5318 * whenever sysctl_lowmem_reserve_ratio changes. 5319 * 5320 * The reserve ratio obviously has absolutely no relation with the 5321 * minimum watermarks. The lowmem reserve ratio can only make sense 5322 * if in function of the boot time zone sizes. 5323 */ 5324int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5325 void __user *buffer, size_t *length, loff_t *ppos) 5326{ 5327 proc_dointvec_minmax(table, write, buffer, length, ppos); 5328 setup_per_zone_lowmem_reserve(); 5329 return 0; 5330} 5331 5332/* 5333 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5334 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5335 * can have before it gets flushed back to buddy allocator. 5336 */ 5337 5338int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5339 void __user *buffer, size_t *length, loff_t *ppos) 5340{ 5341 struct zone *zone; 5342 unsigned int cpu; 5343 int ret; 5344 5345 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5346 if (!write || (ret == -EINVAL)) 5347 return ret; 5348 for_each_populated_zone(zone) { 5349 for_each_possible_cpu(cpu) { 5350 unsigned long high; 5351 high = zone->present_pages / percpu_pagelist_fraction; 5352 setup_pagelist_highmark( 5353 per_cpu_ptr(zone->pageset, cpu), high); 5354 } 5355 } 5356 return 0; 5357} 5358 5359int hashdist = HASHDIST_DEFAULT; 5360 5361#ifdef CONFIG_NUMA 5362static int __init set_hashdist(char *str) 5363{ 5364 if (!str) 5365 return 0; 5366 hashdist = simple_strtoul(str, &str, 0); 5367 return 1; 5368} 5369__setup("hashdist=", set_hashdist); 5370#endif 5371 5372/* 5373 * allocate a large system hash table from bootmem 5374 * - it is assumed that the hash table must contain an exact power-of-2 5375 * quantity of entries 5376 * - limit is the number of hash buckets, not the total allocation size 5377 */ 5378void *__init alloc_large_system_hash(const char *tablename, 5379 unsigned long bucketsize, 5380 unsigned long numentries, 5381 int scale, 5382 int flags, 5383 unsigned int *_hash_shift, 5384 unsigned int *_hash_mask, 5385 unsigned long limit) 5386{ 5387 unsigned long long max = limit; 5388 unsigned long log2qty, size; 5389 void *table = NULL; 5390 5391 /* allow the kernel cmdline to have a say */ 5392 if (!numentries) { 5393 /* round applicable memory size up to nearest megabyte */ 5394 numentries = nr_kernel_pages; 5395 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5396 numentries >>= 20 - PAGE_SHIFT; 5397 numentries <<= 20 - PAGE_SHIFT; 5398 5399 /* limit to 1 bucket per 2^scale bytes of low memory */ 5400 if (scale > PAGE_SHIFT) 5401 numentries >>= (scale - PAGE_SHIFT); 5402 else 5403 numentries <<= (PAGE_SHIFT - scale); 5404 5405 /* Make sure we've got at least a 0-order allocation.. */ 5406 if (unlikely(flags & HASH_SMALL)) { 5407 /* Makes no sense without HASH_EARLY */ 5408 WARN_ON(!(flags & HASH_EARLY)); 5409 if (!(numentries >> *_hash_shift)) { 5410 numentries = 1UL << *_hash_shift; 5411 BUG_ON(!numentries); 5412 } 5413 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5414 numentries = PAGE_SIZE / bucketsize; 5415 } 5416 numentries = roundup_pow_of_two(numentries); 5417 5418 /* limit allocation size to 1/16 total memory by default */ 5419 if (max == 0) { 5420 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5421 do_div(max, bucketsize); 5422 } 5423 5424 if (numentries > max) 5425 numentries = max; 5426 5427 log2qty = ilog2(numentries); 5428 5429 do { 5430 size = bucketsize << log2qty; 5431 if (flags & HASH_EARLY) 5432 table = alloc_bootmem_nopanic(size); 5433 else if (hashdist) 5434 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5435 else { 5436 /* 5437 * If bucketsize is not a power-of-two, we may free 5438 * some pages at the end of hash table which 5439 * alloc_pages_exact() automatically does 5440 */ 5441 if (get_order(size) < MAX_ORDER) { 5442 table = alloc_pages_exact(size, GFP_ATOMIC); 5443 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5444 } 5445 } 5446 } while (!table && size > PAGE_SIZE && --log2qty); 5447 5448 if (!table) 5449 panic("Failed to allocate %s hash table\n", tablename); 5450 5451 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5452 tablename, 5453 (1UL << log2qty), 5454 ilog2(size) - PAGE_SHIFT, 5455 size); 5456 5457 if (_hash_shift) 5458 *_hash_shift = log2qty; 5459 if (_hash_mask) 5460 *_hash_mask = (1 << log2qty) - 1; 5461 5462 return table; 5463} 5464 5465/* Return a pointer to the bitmap storing bits affecting a block of pages */ 5466static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5467 unsigned long pfn) 5468{ 5469#ifdef CONFIG_SPARSEMEM 5470 return __pfn_to_section(pfn)->pageblock_flags; 5471#else 5472 return zone->pageblock_flags; 5473#endif /* CONFIG_SPARSEMEM */ 5474} 5475 5476static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5477{ 5478#ifdef CONFIG_SPARSEMEM 5479 pfn &= (PAGES_PER_SECTION-1); 5480 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5481#else 5482 pfn = pfn - zone->zone_start_pfn; 5483 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5484#endif /* CONFIG_SPARSEMEM */ 5485} 5486 5487/** 5488 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5489 * @page: The page within the block of interest 5490 * @start_bitidx: The first bit of interest to retrieve 5491 * @end_bitidx: The last bit of interest 5492 * returns pageblock_bits flags 5493 */ 5494unsigned long get_pageblock_flags_group(struct page *page, 5495 int start_bitidx, int end_bitidx) 5496{ 5497 struct zone *zone; 5498 unsigned long *bitmap; 5499 unsigned long pfn, bitidx; 5500 unsigned long flags = 0; 5501 unsigned long value = 1; 5502 5503 zone = page_zone(page); 5504 pfn = page_to_pfn(page); 5505 bitmap = get_pageblock_bitmap(zone, pfn); 5506 bitidx = pfn_to_bitidx(zone, pfn); 5507 5508 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5509 if (test_bit(bitidx + start_bitidx, bitmap)) 5510 flags |= value; 5511 5512 return flags; 5513} 5514 5515/** 5516 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5517 * @page: The page within the block of interest 5518 * @start_bitidx: The first bit of interest 5519 * @end_bitidx: The last bit of interest 5520 * @flags: The flags to set 5521 */ 5522void set_pageblock_flags_group(struct page *page, unsigned long flags, 5523 int start_bitidx, int end_bitidx) 5524{ 5525 struct zone *zone; 5526 unsigned long *bitmap; 5527 unsigned long pfn, bitidx; 5528 unsigned long value = 1; 5529 5530 zone = page_zone(page); 5531 pfn = page_to_pfn(page); 5532 bitmap = get_pageblock_bitmap(zone, pfn); 5533 bitidx = pfn_to_bitidx(zone, pfn); 5534 VM_BUG_ON(pfn < zone->zone_start_pfn); 5535 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5536 5537 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5538 if (flags & value) 5539 __set_bit(bitidx + start_bitidx, bitmap); 5540 else 5541 __clear_bit(bitidx + start_bitidx, bitmap); 5542} 5543 5544/* 5545 * This is designed as sub function...plz see page_isolation.c also. 5546 * set/clear page block's type to be ISOLATE. 5547 * page allocater never alloc memory from ISOLATE block. 5548 */ 5549 5550static int 5551__count_immobile_pages(struct zone *zone, struct page *page, int count) 5552{ 5553 unsigned long pfn, iter, found; 5554 /* 5555 * For avoiding noise data, lru_add_drain_all() should be called 5556 * If ZONE_MOVABLE, the zone never contains immobile pages 5557 */ 5558 if (zone_idx(zone) == ZONE_MOVABLE) 5559 return true; 5560 5561 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE) 5562 return true; 5563 5564 pfn = page_to_pfn(page); 5565 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 5566 unsigned long check = pfn + iter; 5567 5568 if (!pfn_valid_within(check)) 5569 continue; 5570 5571 page = pfn_to_page(check); 5572 if (!page_count(page)) { 5573 if (PageBuddy(page)) 5574 iter += (1 << page_order(page)) - 1; 5575 continue; 5576 } 5577 if (!PageLRU(page)) 5578 found++; 5579 /* 5580 * If there are RECLAIMABLE pages, we need to check it. 5581 * But now, memory offline itself doesn't call shrink_slab() 5582 * and it still to be fixed. 5583 */ 5584 /* 5585 * If the page is not RAM, page_count()should be 0. 5586 * we don't need more check. This is an _used_ not-movable page. 5587 * 5588 * The problematic thing here is PG_reserved pages. PG_reserved 5589 * is set to both of a memory hole page and a _used_ kernel 5590 * page at boot. 5591 */ 5592 if (found > count) 5593 return false; 5594 } 5595 return true; 5596} 5597 5598bool is_pageblock_removable_nolock(struct page *page) 5599{ 5600 struct zone *zone = page_zone(page); 5601 return __count_immobile_pages(zone, page, 0); 5602} 5603 5604int set_migratetype_isolate(struct page *page) 5605{ 5606 struct zone *zone; 5607 unsigned long flags, pfn; 5608 struct memory_isolate_notify arg; 5609 int notifier_ret; 5610 int ret = -EBUSY; 5611 5612 zone = page_zone(page); 5613 5614 spin_lock_irqsave(&zone->lock, flags); 5615 5616 pfn = page_to_pfn(page); 5617 arg.start_pfn = pfn; 5618 arg.nr_pages = pageblock_nr_pages; 5619 arg.pages_found = 0; 5620 5621 /* 5622 * It may be possible to isolate a pageblock even if the 5623 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5624 * notifier chain is used by balloon drivers to return the 5625 * number of pages in a range that are held by the balloon 5626 * driver to shrink memory. If all the pages are accounted for 5627 * by balloons, are free, or on the LRU, isolation can continue. 5628 * Later, for example, when memory hotplug notifier runs, these 5629 * pages reported as "can be isolated" should be isolated(freed) 5630 * by the balloon driver through the memory notifier chain. 5631 */ 5632 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5633 notifier_ret = notifier_to_errno(notifier_ret); 5634 if (notifier_ret) 5635 goto out; 5636 /* 5637 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. 5638 * We just check MOVABLE pages. 5639 */ 5640 if (__count_immobile_pages(zone, page, arg.pages_found)) 5641 ret = 0; 5642 5643 /* 5644 * immobile means "not-on-lru" paes. If immobile is larger than 5645 * removable-by-driver pages reported by notifier, we'll fail. 5646 */ 5647 5648out: 5649 if (!ret) { 5650 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5651 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5652 } 5653 5654 spin_unlock_irqrestore(&zone->lock, flags); 5655 if (!ret) 5656 drain_all_pages(); 5657 return ret; 5658} 5659 5660void unset_migratetype_isolate(struct page *page) 5661{ 5662 struct zone *zone; 5663 unsigned long flags; 5664 zone = page_zone(page); 5665 spin_lock_irqsave(&zone->lock, flags); 5666 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5667 goto out; 5668 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5669 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5670out: 5671 spin_unlock_irqrestore(&zone->lock, flags); 5672} 5673 5674#ifdef CONFIG_MEMORY_HOTREMOVE 5675/* 5676 * All pages in the range must be isolated before calling this. 5677 */ 5678void 5679__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5680{ 5681 struct page *page; 5682 struct zone *zone; 5683 int order, i; 5684 unsigned long pfn; 5685 unsigned long flags; 5686 /* find the first valid pfn */ 5687 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5688 if (pfn_valid(pfn)) 5689 break; 5690 if (pfn == end_pfn) 5691 return; 5692 zone = page_zone(pfn_to_page(pfn)); 5693 spin_lock_irqsave(&zone->lock, flags); 5694 pfn = start_pfn; 5695 while (pfn < end_pfn) { 5696 if (!pfn_valid(pfn)) { 5697 pfn++; 5698 continue; 5699 } 5700 page = pfn_to_page(pfn); 5701 BUG_ON(page_count(page)); 5702 BUG_ON(!PageBuddy(page)); 5703 order = page_order(page); 5704#ifdef CONFIG_DEBUG_VM 5705 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5706 pfn, 1 << order, end_pfn); 5707#endif 5708 list_del(&page->lru); 5709 rmv_page_order(page); 5710 zone->free_area[order].nr_free--; 5711 __mod_zone_page_state(zone, NR_FREE_PAGES, 5712 - (1UL << order)); 5713 for (i = 0; i < (1 << order); i++) 5714 SetPageReserved((page+i)); 5715 pfn += (1 << order); 5716 } 5717 spin_unlock_irqrestore(&zone->lock, flags); 5718} 5719#endif 5720 5721#ifdef CONFIG_MEMORY_FAILURE 5722bool is_free_buddy_page(struct page *page) 5723{ 5724 struct zone *zone = page_zone(page); 5725 unsigned long pfn = page_to_pfn(page); 5726 unsigned long flags; 5727 int order; 5728 5729 spin_lock_irqsave(&zone->lock, flags); 5730 for (order = 0; order < MAX_ORDER; order++) { 5731 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5732 5733 if (PageBuddy(page_head) && page_order(page_head) >= order) 5734 break; 5735 } 5736 spin_unlock_irqrestore(&zone->lock, flags); 5737 5738 return order < MAX_ORDER; 5739} 5740#endif 5741 5742static struct trace_print_flags pageflag_names[] = { 5743 {1UL << PG_locked, "locked" }, 5744 {1UL << PG_error, "error" }, 5745 {1UL << PG_referenced, "referenced" }, 5746 {1UL << PG_uptodate, "uptodate" }, 5747 {1UL << PG_dirty, "dirty" }, 5748 {1UL << PG_lru, "lru" }, 5749 {1UL << PG_active, "active" }, 5750 {1UL << PG_slab, "slab" }, 5751 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5752 {1UL << PG_arch_1, "arch_1" }, 5753 {1UL << PG_reserved, "reserved" }, 5754 {1UL << PG_private, "private" }, 5755 {1UL << PG_private_2, "private_2" }, 5756 {1UL << PG_writeback, "writeback" }, 5757#ifdef CONFIG_PAGEFLAGS_EXTENDED 5758 {1UL << PG_head, "head" }, 5759 {1UL << PG_tail, "tail" }, 5760#else 5761 {1UL << PG_compound, "compound" }, 5762#endif 5763 {1UL << PG_swapcache, "swapcache" }, 5764 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5765 {1UL << PG_reclaim, "reclaim" }, 5766 {1UL << PG_swapbacked, "swapbacked" }, 5767 {1UL << PG_unevictable, "unevictable" }, 5768#ifdef CONFIG_MMU 5769 {1UL << PG_mlocked, "mlocked" }, 5770#endif 5771#ifdef CONFIG_ARCH_USES_PG_UNCACHED 5772 {1UL << PG_uncached, "uncached" }, 5773#endif 5774#ifdef CONFIG_MEMORY_FAILURE 5775 {1UL << PG_hwpoison, "hwpoison" }, 5776#endif 5777 {-1UL, NULL }, 5778}; 5779 5780static void dump_page_flags(unsigned long flags) 5781{ 5782 const char *delim = ""; 5783 unsigned long mask; 5784 int i; 5785 5786 printk(KERN_ALERT "page flags: %#lx(", flags); 5787 5788 /* remove zone id */ 5789 flags &= (1UL << NR_PAGEFLAGS) - 1; 5790 5791 for (i = 0; pageflag_names[i].name && flags; i++) { 5792 5793 mask = pageflag_names[i].mask; 5794 if ((flags & mask) != mask) 5795 continue; 5796 5797 flags &= ~mask; 5798 printk("%s%s", delim, pageflag_names[i].name); 5799 delim = "|"; 5800 } 5801 5802 /* check for left over flags */ 5803 if (flags) 5804 printk("%s%#lx", delim, flags); 5805 5806 printk(")\n"); 5807} 5808 5809void dump_page(struct page *page) 5810{ 5811 printk(KERN_ALERT 5812 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5813 page, atomic_read(&page->_count), page_mapcount(page), 5814 page->mapping, page->index); 5815 dump_page_flags(page->flags); 5816 mem_cgroup_print_bad_page(page); 5817} 5818