1 /* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17 #include <linux/stddef.h> 18 #include <linux/mm.h> 19 #include <linux/swap.h> 20 #include <linux/interrupt.h> 21 #include <linux/pagemap.h> 22 #include <linux/jiffies.h> 23 #include <linux/bootmem.h> 24 #include <linux/compiler.h> 25 #include <linux/kernel.h> 26 #include <linux/module.h> 27 #include <linux/suspend.h> 28 #include <linux/pagevec.h> 29 #include <linux/blkdev.h> 30 #include <linux/slab.h> 31 #include <linux/oom.h> 32 #include <linux/notifier.h> 33 #include <linux/topology.h> 34 #include <linux/sysctl.h> 35 #include <linux/cpu.h> 36 #include <linux/cpuset.h> 37 #include <linux/memory_hotplug.h> 38 #include <linux/nodemask.h> 39 #include <linux/vmalloc.h> 40 #include <linux/mempolicy.h> 41 #include <linux/stop_machine.h> 42 #include <linux/sort.h> 43 #include <linux/pfn.h> 44 #include <linux/backing-dev.h> 45 #include <linux/fault-inject.h> 46 #include <linux/page-isolation.h> 47 #include <linux/memcontrol.h> 48 49 #include <asm/tlbflush.h> 50 #include <asm/div64.h> 51 #include "internal.h" 52 53 /* 54 * Array of node states. 55 */ 56 nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 57 [N_POSSIBLE] = NODE_MASK_ALL, 58 [N_ONLINE] = { { [0] = 1UL } }, 59 #ifndef CONFIG_NUMA 60 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 61 #ifdef CONFIG_HIGHMEM 62 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 63 #endif 64 [N_CPU] = { { [0] = 1UL } }, 65 #endif /* NUMA */ 66 }; 67 EXPORT_SYMBOL(node_states); 68 69 unsigned long totalram_pages __read_mostly; 70 unsigned long totalreserve_pages __read_mostly; 71 long nr_swap_pages; 72 int percpu_pagelist_fraction; 73 74 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 75 int pageblock_order __read_mostly; 76 #endif 77 78 static void __free_pages_ok(struct page *page, unsigned int order); 79 80 /* 81 * results with 256, 32 in the lowmem_reserve sysctl: 82 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 83 * 1G machine -> (16M dma, 784M normal, 224M high) 84 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 85 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 86 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 87 * 88 * TBD: should special case ZONE_DMA32 machines here - in those we normally 89 * don't need any ZONE_NORMAL reservation 90 */ 91 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 92 #ifdef CONFIG_ZONE_DMA 93 256, 94 #endif 95 #ifdef CONFIG_ZONE_DMA32 96 256, 97 #endif 98 #ifdef CONFIG_HIGHMEM 99 32, 100 #endif 101 32, 102 }; 103 104 EXPORT_SYMBOL(totalram_pages); 105 106 static char * const zone_names[MAX_NR_ZONES] = { 107 #ifdef CONFIG_ZONE_DMA 108 "DMA", 109 #endif 110 #ifdef CONFIG_ZONE_DMA32 111 "DMA32", 112 #endif 113 "Normal", 114 #ifdef CONFIG_HIGHMEM 115 "HighMem", 116 #endif 117 "Movable", 118 }; 119 120 int min_free_kbytes = 1024; 121 122 unsigned long __meminitdata nr_kernel_pages; 123 unsigned long __meminitdata nr_all_pages; 124 static unsigned long __meminitdata dma_reserve; 125 126 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP 127 /* 128 * MAX_ACTIVE_REGIONS determines the maximum number of distinct 129 * ranges of memory (RAM) that may be registered with add_active_range(). 130 * Ranges passed to add_active_range() will be merged if possible 131 * so the number of times add_active_range() can be called is 132 * related to the number of nodes and the number of holes 133 */ 134 #ifdef CONFIG_MAX_ACTIVE_REGIONS 135 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */ 136 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS 137 #else 138 #if MAX_NUMNODES >= 32 139 /* If there can be many nodes, allow up to 50 holes per node */ 140 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50) 141 #else 142 /* By default, allow up to 256 distinct regions */ 143 #define MAX_ACTIVE_REGIONS 256 144 #endif 145 #endif 146 147 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS]; 148 static int __meminitdata nr_nodemap_entries; 149 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 150 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 151 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE 152 static unsigned long __meminitdata node_boundary_start_pfn[MAX_NUMNODES]; 153 static unsigned long __meminitdata node_boundary_end_pfn[MAX_NUMNODES]; 154 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */ 155 unsigned long __initdata required_kernelcore; 156 static unsigned long __initdata required_movablecore; 157 unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 158 159 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 160 int movable_zone; 161 EXPORT_SYMBOL(movable_zone); 162 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 163 164 #if MAX_NUMNODES > 1 165 int nr_node_ids __read_mostly = MAX_NUMNODES; 166 EXPORT_SYMBOL(nr_node_ids); 167 #endif 168 169 int page_group_by_mobility_disabled __read_mostly; 170 171 static void set_pageblock_migratetype(struct page *page, int migratetype) 172 { 173 set_pageblock_flags_group(page, (unsigned long)migratetype, 174 PB_migrate, PB_migrate_end); 175 } 176 177 #ifdef CONFIG_DEBUG_VM 178 static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 179 { 180 int ret = 0; 181 unsigned seq; 182 unsigned long pfn = page_to_pfn(page); 183 184 do { 185 seq = zone_span_seqbegin(zone); 186 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 187 ret = 1; 188 else if (pfn < zone->zone_start_pfn) 189 ret = 1; 190 } while (zone_span_seqretry(zone, seq)); 191 192 return ret; 193 } 194 195 static int page_is_consistent(struct zone *zone, struct page *page) 196 { 197 if (!pfn_valid_within(page_to_pfn(page))) 198 return 0; 199 if (zone != page_zone(page)) 200 return 0; 201 202 return 1; 203 } 204 /* 205 * Temporary debugging check for pages not lying within a given zone. 206 */ 207 static int bad_range(struct zone *zone, struct page *page) 208 { 209 if (page_outside_zone_boundaries(zone, page)) 210 return 1; 211 if (!page_is_consistent(zone, page)) 212 return 1; 213 214 return 0; 215 } 216 #else 217 static inline int bad_range(struct zone *zone, struct page *page) 218 { 219 return 0; 220 } 221 #endif 222 223 static void bad_page(struct page *page) 224 { 225 void *pc = page_get_page_cgroup(page); 226 227 printk(KERN_EMERG "Bad page state in process '%s'\n" KERN_EMERG 228 "page:%p flags:0x%0*lx mapping:%p mapcount:%d count:%d\n", 229 current->comm, page, (int)(2*sizeof(unsigned long)), 230 (unsigned long)page->flags, page->mapping, 231 page_mapcount(page), page_count(page)); 232 if (pc) { 233 printk(KERN_EMERG "cgroup:%p\n", pc); 234 page_reset_bad_cgroup(page); 235 } 236 printk(KERN_EMERG "Trying to fix it up, but a reboot is needed\n" 237 KERN_EMERG "Backtrace:\n"); 238 dump_stack(); 239 page->flags &= ~(1 << PG_lru | 240 1 << PG_private | 241 1 << PG_locked | 242 1 << PG_active | 243 1 << PG_dirty | 244 1 << PG_reclaim | 245 1 << PG_slab | 246 1 << PG_swapcache | 247 1 << PG_writeback | 248 1 << PG_buddy ); 249 set_page_count(page, 0); 250 reset_page_mapcount(page); 251 page->mapping = NULL; 252 add_taint(TAINT_BAD_PAGE); 253 } 254 255 /* 256 * Higher-order pages are called "compound pages". They are structured thusly: 257 * 258 * The first PAGE_SIZE page is called the "head page". 259 * 260 * The remaining PAGE_SIZE pages are called "tail pages". 261 * 262 * All pages have PG_compound set. All pages have their ->private pointing at 263 * the head page (even the head page has this). 264 * 265 * The first tail page's ->lru.next holds the address of the compound page's 266 * put_page() function. Its ->lru.prev holds the order of allocation. 267 * This usage means that zero-order pages may not be compound. 268 */ 269 270 static void free_compound_page(struct page *page) 271 { 272 __free_pages_ok(page, compound_order(page)); 273 } 274 275 static void prep_compound_page(struct page *page, unsigned long order) 276 { 277 int i; 278 int nr_pages = 1 << order; 279 280 set_compound_page_dtor(page, free_compound_page); 281 set_compound_order(page, order); 282 __SetPageHead(page); 283 for (i = 1; i < nr_pages; i++) { 284 struct page *p = page + i; 285 286 __SetPageTail(p); 287 p->first_page = page; 288 } 289 } 290 291 static void destroy_compound_page(struct page *page, unsigned long order) 292 { 293 int i; 294 int nr_pages = 1 << order; 295 296 if (unlikely(compound_order(page) != order)) 297 bad_page(page); 298 299 if (unlikely(!PageHead(page))) 300 bad_page(page); 301 __ClearPageHead(page); 302 for (i = 1; i < nr_pages; i++) { 303 struct page *p = page + i; 304 305 if (unlikely(!PageTail(p) | 306 (p->first_page != page))) 307 bad_page(page); 308 __ClearPageTail(p); 309 } 310 } 311 312 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 313 { 314 int i; 315 316 /* 317 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 318 * and __GFP_HIGHMEM from hard or soft interrupt context. 319 */ 320 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 321 for (i = 0; i < (1 << order); i++) 322 clear_highpage(page + i); 323 } 324 325 static inline void set_page_order(struct page *page, int order) 326 { 327 set_page_private(page, order); 328 __SetPageBuddy(page); 329 } 330 331 static inline void rmv_page_order(struct page *page) 332 { 333 __ClearPageBuddy(page); 334 set_page_private(page, 0); 335 } 336 337 /* 338 * Locate the struct page for both the matching buddy in our 339 * pair (buddy1) and the combined O(n+1) page they form (page). 340 * 341 * 1) Any buddy B1 will have an order O twin B2 which satisfies 342 * the following equation: 343 * B2 = B1 ^ (1 << O) 344 * For example, if the starting buddy (buddy2) is #8 its order 345 * 1 buddy is #10: 346 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 347 * 348 * 2) Any buddy B will have an order O+1 parent P which 349 * satisfies the following equation: 350 * P = B & ~(1 << O) 351 * 352 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 353 */ 354 static inline struct page * 355 __page_find_buddy(struct page *page, unsigned long page_idx, unsigned int order) 356 { 357 unsigned long buddy_idx = page_idx ^ (1 << order); 358 359 return page + (buddy_idx - page_idx); 360 } 361 362 static inline unsigned long 363 __find_combined_index(unsigned long page_idx, unsigned int order) 364 { 365 return (page_idx & ~(1 << order)); 366 } 367 368 /* 369 * This function checks whether a page is free && is the buddy 370 * we can do coalesce a page and its buddy if 371 * (a) the buddy is not in a hole && 372 * (b) the buddy is in the buddy system && 373 * (c) a page and its buddy have the same order && 374 * (d) a page and its buddy are in the same zone. 375 * 376 * For recording whether a page is in the buddy system, we use PG_buddy. 377 * Setting, clearing, and testing PG_buddy is serialized by zone->lock. 378 * 379 * For recording page's order, we use page_private(page). 380 */ 381 static inline int page_is_buddy(struct page *page, struct page *buddy, 382 int order) 383 { 384 if (!pfn_valid_within(page_to_pfn(buddy))) 385 return 0; 386 387 if (page_zone_id(page) != page_zone_id(buddy)) 388 return 0; 389 390 if (PageBuddy(buddy) && page_order(buddy) == order) { 391 BUG_ON(page_count(buddy) != 0); 392 return 1; 393 } 394 return 0; 395 } 396 397 /* 398 * Freeing function for a buddy system allocator. 399 * 400 * The concept of a buddy system is to maintain direct-mapped table 401 * (containing bit values) for memory blocks of various "orders". 402 * The bottom level table contains the map for the smallest allocatable 403 * units of memory (here, pages), and each level above it describes 404 * pairs of units from the levels below, hence, "buddies". 405 * At a high level, all that happens here is marking the table entry 406 * at the bottom level available, and propagating the changes upward 407 * as necessary, plus some accounting needed to play nicely with other 408 * parts of the VM system. 409 * At each level, we keep a list of pages, which are heads of continuous 410 * free pages of length of (1 << order) and marked with PG_buddy. Page's 411 * order is recorded in page_private(page) field. 412 * So when we are allocating or freeing one, we can derive the state of the 413 * other. That is, if we allocate a small block, and both were 414 * free, the remainder of the region must be split into blocks. 415 * If a block is freed, and its buddy is also free, then this 416 * triggers coalescing into a block of larger size. 417 * 418 * -- wli 419 */ 420 421 static inline void __free_one_page(struct page *page, 422 struct zone *zone, unsigned int order) 423 { 424 unsigned long page_idx; 425 int order_size = 1 << order; 426 int migratetype = get_pageblock_migratetype(page); 427 428 if (unlikely(PageCompound(page))) 429 destroy_compound_page(page, order); 430 431 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 432 433 VM_BUG_ON(page_idx & (order_size - 1)); 434 VM_BUG_ON(bad_range(zone, page)); 435 436 __mod_zone_page_state(zone, NR_FREE_PAGES, order_size); 437 while (order < MAX_ORDER-1) { 438 unsigned long combined_idx; 439 struct page *buddy; 440 441 buddy = __page_find_buddy(page, page_idx, order); 442 if (!page_is_buddy(page, buddy, order)) 443 break; /* Move the buddy up one level. */ 444 445 list_del(&buddy->lru); 446 zone->free_area[order].nr_free--; 447 rmv_page_order(buddy); 448 combined_idx = __find_combined_index(page_idx, order); 449 page = page + (combined_idx - page_idx); 450 page_idx = combined_idx; 451 order++; 452 } 453 set_page_order(page, order); 454 list_add(&page->lru, 455 &zone->free_area[order].free_list[migratetype]); 456 zone->free_area[order].nr_free++; 457 } 458 459 static inline int free_pages_check(struct page *page) 460 { 461 if (unlikely(page_mapcount(page) | 462 (page->mapping != NULL) | 463 (page_get_page_cgroup(page) != NULL) | 464 (page_count(page) != 0) | 465 (page->flags & ( 466 1 << PG_lru | 467 1 << PG_private | 468 1 << PG_locked | 469 1 << PG_active | 470 1 << PG_slab | 471 1 << PG_swapcache | 472 1 << PG_writeback | 473 1 << PG_reserved | 474 1 << PG_buddy )))) 475 bad_page(page); 476 if (PageDirty(page)) 477 __ClearPageDirty(page); 478 /* 479 * For now, we report if PG_reserved was found set, but do not 480 * clear it, and do not free the page. But we shall soon need 481 * to do more, for when the ZERO_PAGE count wraps negative. 482 */ 483 return PageReserved(page); 484 } 485 486 /* 487 * Frees a list of pages. 488 * Assumes all pages on list are in same zone, and of same order. 489 * count is the number of pages to free. 490 * 491 * If the zone was previously in an "all pages pinned" state then look to 492 * see if this freeing clears that state. 493 * 494 * And clear the zone's pages_scanned counter, to hold off the "all pages are 495 * pinned" detection logic. 496 */ 497 static void free_pages_bulk(struct zone *zone, int count, 498 struct list_head *list, int order) 499 { 500 spin_lock(&zone->lock); 501 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 502 zone->pages_scanned = 0; 503 while (count--) { 504 struct page *page; 505 506 VM_BUG_ON(list_empty(list)); 507 page = list_entry(list->prev, struct page, lru); 508 /* have to delete it as __free_one_page list manipulates */ 509 list_del(&page->lru); 510 __free_one_page(page, zone, order); 511 } 512 spin_unlock(&zone->lock); 513 } 514 515 static void free_one_page(struct zone *zone, struct page *page, int order) 516 { 517 spin_lock(&zone->lock); 518 zone_clear_flag(zone, ZONE_ALL_UNRECLAIMABLE); 519 zone->pages_scanned = 0; 520 __free_one_page(page, zone, order); 521 spin_unlock(&zone->lock); 522 } 523 524 static void __free_pages_ok(struct page *page, unsigned int order) 525 { 526 unsigned long flags; 527 int i; 528 int reserved = 0; 529 530 for (i = 0 ; i < (1 << order) ; ++i) 531 reserved += free_pages_check(page + i); 532 if (reserved) 533 return; 534 535 if (!PageHighMem(page)) 536 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 537 arch_free_page(page, order); 538 kernel_map_pages(page, 1 << order, 0); 539 540 local_irq_save(flags); 541 __count_vm_events(PGFREE, 1 << order); 542 free_one_page(page_zone(page), page, order); 543 local_irq_restore(flags); 544 } 545 546 /* 547 * permit the bootmem allocator to evade page validation on high-order frees 548 */ 549 void __init __free_pages_bootmem(struct page *page, unsigned int order) 550 { 551 if (order == 0) { 552 __ClearPageReserved(page); 553 set_page_count(page, 0); 554 set_page_refcounted(page); 555 __free_page(page); 556 } else { 557 int loop; 558 559 prefetchw(page); 560 for (loop = 0; loop < BITS_PER_LONG; loop++) { 561 struct page *p = &page[loop]; 562 563 if (loop + 1 < BITS_PER_LONG) 564 prefetchw(p + 1); 565 __ClearPageReserved(p); 566 set_page_count(p, 0); 567 } 568 569 set_page_refcounted(page); 570 __free_pages(page, order); 571 } 572 } 573 574 575 /* 576 * The order of subdivision here is critical for the IO subsystem. 577 * Please do not alter this order without good reasons and regression 578 * testing. Specifically, as large blocks of memory are subdivided, 579 * the order in which smaller blocks are delivered depends on the order 580 * they're subdivided in this function. This is the primary factor 581 * influencing the order in which pages are delivered to the IO 582 * subsystem according to empirical testing, and this is also justified 583 * by considering the behavior of a buddy system containing a single 584 * large block of memory acted on by a series of small allocations. 585 * This behavior is a critical factor in sglist merging's success. 586 * 587 * -- wli 588 */ 589 static inline void expand(struct zone *zone, struct page *page, 590 int low, int high, struct free_area *area, 591 int migratetype) 592 { 593 unsigned long size = 1 << high; 594 595 while (high > low) { 596 area--; 597 high--; 598 size >>= 1; 599 VM_BUG_ON(bad_range(zone, &page[size])); 600 list_add(&page[size].lru, &area->free_list[migratetype]); 601 area->nr_free++; 602 set_page_order(&page[size], high); 603 } 604 } 605 606 /* 607 * This page is about to be returned from the page allocator 608 */ 609 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 610 { 611 if (unlikely(page_mapcount(page) | 612 (page->mapping != NULL) | 613 (page_get_page_cgroup(page) != NULL) | 614 (page_count(page) != 0) | 615 (page->flags & ( 616 1 << PG_lru | 617 1 << PG_private | 618 1 << PG_locked | 619 1 << PG_active | 620 1 << PG_dirty | 621 1 << PG_slab | 622 1 << PG_swapcache | 623 1 << PG_writeback | 624 1 << PG_reserved | 625 1 << PG_buddy )))) 626 bad_page(page); 627 628 /* 629 * For now, we report if PG_reserved was found set, but do not 630 * clear it, and do not allocate the page: as a safety net. 631 */ 632 if (PageReserved(page)) 633 return 1; 634 635 page->flags &= ~(1 << PG_uptodate | 1 << PG_error | 1 << PG_readahead | 636 1 << PG_referenced | 1 << PG_arch_1 | 637 1 << PG_owner_priv_1 | 1 << PG_mappedtodisk); 638 set_page_private(page, 0); 639 set_page_refcounted(page); 640 641 arch_alloc_page(page, order); 642 kernel_map_pages(page, 1 << order, 1); 643 644 if (gfp_flags & __GFP_ZERO) 645 prep_zero_page(page, order, gfp_flags); 646 647 if (order && (gfp_flags & __GFP_COMP)) 648 prep_compound_page(page, order); 649 650 return 0; 651 } 652 653 /* 654 * Go through the free lists for the given migratetype and remove 655 * the smallest available page from the freelists 656 */ 657 static struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 658 int migratetype) 659 { 660 unsigned int current_order; 661 struct free_area * area; 662 struct page *page; 663 664 /* Find a page of the appropriate size in the preferred list */ 665 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 666 area = &(zone->free_area[current_order]); 667 if (list_empty(&area->free_list[migratetype])) 668 continue; 669 670 page = list_entry(area->free_list[migratetype].next, 671 struct page, lru); 672 list_del(&page->lru); 673 rmv_page_order(page); 674 area->nr_free--; 675 __mod_zone_page_state(zone, NR_FREE_PAGES, - (1UL << order)); 676 expand(zone, page, order, current_order, area, migratetype); 677 return page; 678 } 679 680 return NULL; 681 } 682 683 684 /* 685 * This array describes the order lists are fallen back to when 686 * the free lists for the desirable migrate type are depleted 687 */ 688 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 689 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 690 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 691 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 692 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 693 }; 694 695 /* 696 * Move the free pages in a range to the free lists of the requested type. 697 * Note that start_page and end_pages are not aligned on a pageblock 698 * boundary. If alignment is required, use move_freepages_block() 699 */ 700 int move_freepages(struct zone *zone, 701 struct page *start_page, struct page *end_page, 702 int migratetype) 703 { 704 struct page *page; 705 unsigned long order; 706 int pages_moved = 0; 707 708 #ifndef CONFIG_HOLES_IN_ZONE 709 /* 710 * page_zone is not safe to call in this context when 711 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 712 * anyway as we check zone boundaries in move_freepages_block(). 713 * Remove at a later date when no bug reports exist related to 714 * grouping pages by mobility 715 */ 716 BUG_ON(page_zone(start_page) != page_zone(end_page)); 717 #endif 718 719 for (page = start_page; page <= end_page;) { 720 if (!pfn_valid_within(page_to_pfn(page))) { 721 page++; 722 continue; 723 } 724 725 if (!PageBuddy(page)) { 726 page++; 727 continue; 728 } 729 730 order = page_order(page); 731 list_del(&page->lru); 732 list_add(&page->lru, 733 &zone->free_area[order].free_list[migratetype]); 734 page += 1 << order; 735 pages_moved += 1 << order; 736 } 737 738 return pages_moved; 739 } 740 741 int move_freepages_block(struct zone *zone, struct page *page, int migratetype) 742 { 743 unsigned long start_pfn, end_pfn; 744 struct page *start_page, *end_page; 745 746 start_pfn = page_to_pfn(page); 747 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 748 start_page = pfn_to_page(start_pfn); 749 end_page = start_page + pageblock_nr_pages - 1; 750 end_pfn = start_pfn + pageblock_nr_pages - 1; 751 752 /* Do not cross zone boundaries */ 753 if (start_pfn < zone->zone_start_pfn) 754 start_page = page; 755 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 756 return 0; 757 758 return move_freepages(zone, start_page, end_page, migratetype); 759 } 760 761 /* Remove an element from the buddy allocator from the fallback list */ 762 static struct page *__rmqueue_fallback(struct zone *zone, int order, 763 int start_migratetype) 764 { 765 struct free_area * area; 766 int current_order; 767 struct page *page; 768 int migratetype, i; 769 770 /* Find the largest possible block of pages in the other list */ 771 for (current_order = MAX_ORDER-1; current_order >= order; 772 --current_order) { 773 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 774 migratetype = fallbacks[start_migratetype][i]; 775 776 /* MIGRATE_RESERVE handled later if necessary */ 777 if (migratetype == MIGRATE_RESERVE) 778 continue; 779 780 area = &(zone->free_area[current_order]); 781 if (list_empty(&area->free_list[migratetype])) 782 continue; 783 784 page = list_entry(area->free_list[migratetype].next, 785 struct page, lru); 786 area->nr_free--; 787 788 /* 789 * If breaking a large block of pages, move all free 790 * pages to the preferred allocation list. If falling 791 * back for a reclaimable kernel allocation, be more 792 * agressive about taking ownership of free pages 793 */ 794 if (unlikely(current_order >= (pageblock_order >> 1)) || 795 start_migratetype == MIGRATE_RECLAIMABLE) { 796 unsigned long pages; 797 pages = move_freepages_block(zone, page, 798 start_migratetype); 799 800 /* Claim the whole block if over half of it is free */ 801 if (pages >= (1 << (pageblock_order-1))) 802 set_pageblock_migratetype(page, 803 start_migratetype); 804 805 migratetype = start_migratetype; 806 } 807 808 /* Remove the page from the freelists */ 809 list_del(&page->lru); 810 rmv_page_order(page); 811 __mod_zone_page_state(zone, NR_FREE_PAGES, 812 -(1UL << order)); 813 814 if (current_order == pageblock_order) 815 set_pageblock_migratetype(page, 816 start_migratetype); 817 818 expand(zone, page, order, current_order, area, migratetype); 819 return page; 820 } 821 } 822 823 /* Use MIGRATE_RESERVE rather than fail an allocation */ 824 return __rmqueue_smallest(zone, order, MIGRATE_RESERVE); 825 } 826 827 /* 828 * Do the hard work of removing an element from the buddy allocator. 829 * Call me with the zone->lock already held. 830 */ 831 static struct page *__rmqueue(struct zone *zone, unsigned int order, 832 int migratetype) 833 { 834 struct page *page; 835 836 page = __rmqueue_smallest(zone, order, migratetype); 837 838 if (unlikely(!page)) 839 page = __rmqueue_fallback(zone, order, migratetype); 840 841 return page; 842 } 843 844 /* 845 * Obtain a specified number of elements from the buddy allocator, all under 846 * a single hold of the lock, for efficiency. Add them to the supplied list. 847 * Returns the number of new pages which were placed at *list. 848 */ 849 static int rmqueue_bulk(struct zone *zone, unsigned int order, 850 unsigned long count, struct list_head *list, 851 int migratetype) 852 { 853 int i; 854 855 spin_lock(&zone->lock); 856 for (i = 0; i < count; ++i) { 857 struct page *page = __rmqueue(zone, order, migratetype); 858 if (unlikely(page == NULL)) 859 break; 860 861 /* 862 * Split buddy pages returned by expand() are received here 863 * in physical page order. The page is added to the callers and 864 * list and the list head then moves forward. From the callers 865 * perspective, the linked list is ordered by page number in 866 * some conditions. This is useful for IO devices that can 867 * merge IO requests if the physical pages are ordered 868 * properly. 869 */ 870 list_add(&page->lru, list); 871 set_page_private(page, migratetype); 872 list = &page->lru; 873 } 874 spin_unlock(&zone->lock); 875 return i; 876 } 877 878 #ifdef CONFIG_NUMA 879 /* 880 * Called from the vmstat counter updater to drain pagesets of this 881 * currently executing processor on remote nodes after they have 882 * expired. 883 * 884 * Note that this function must be called with the thread pinned to 885 * a single processor. 886 */ 887 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 888 { 889 unsigned long flags; 890 int to_drain; 891 892 local_irq_save(flags); 893 if (pcp->count >= pcp->batch) 894 to_drain = pcp->batch; 895 else 896 to_drain = pcp->count; 897 free_pages_bulk(zone, to_drain, &pcp->list, 0); 898 pcp->count -= to_drain; 899 local_irq_restore(flags); 900 } 901 #endif 902 903 /* 904 * Drain pages of the indicated processor. 905 * 906 * The processor must either be the current processor and the 907 * thread pinned to the current processor or a processor that 908 * is not online. 909 */ 910 static void drain_pages(unsigned int cpu) 911 { 912 unsigned long flags; 913 struct zone *zone; 914 915 for_each_zone(zone) { 916 struct per_cpu_pageset *pset; 917 struct per_cpu_pages *pcp; 918 919 if (!populated_zone(zone)) 920 continue; 921 922 pset = zone_pcp(zone, cpu); 923 924 pcp = &pset->pcp; 925 local_irq_save(flags); 926 free_pages_bulk(zone, pcp->count, &pcp->list, 0); 927 pcp->count = 0; 928 local_irq_restore(flags); 929 } 930 } 931 932 /* 933 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 934 */ 935 void drain_local_pages(void *arg) 936 { 937 drain_pages(smp_processor_id()); 938 } 939 940 /* 941 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 942 */ 943 void drain_all_pages(void) 944 { 945 on_each_cpu(drain_local_pages, NULL, 0, 1); 946 } 947 948 #ifdef CONFIG_HIBERNATION 949 950 void mark_free_pages(struct zone *zone) 951 { 952 unsigned long pfn, max_zone_pfn; 953 unsigned long flags; 954 int order, t; 955 struct list_head *curr; 956 957 if (!zone->spanned_pages) 958 return; 959 960 spin_lock_irqsave(&zone->lock, flags); 961 962 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 963 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 964 if (pfn_valid(pfn)) { 965 struct page *page = pfn_to_page(pfn); 966 967 if (!swsusp_page_is_forbidden(page)) 968 swsusp_unset_page_free(page); 969 } 970 971 for_each_migratetype_order(order, t) { 972 list_for_each(curr, &zone->free_area[order].free_list[t]) { 973 unsigned long i; 974 975 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 976 for (i = 0; i < (1UL << order); i++) 977 swsusp_set_page_free(pfn_to_page(pfn + i)); 978 } 979 } 980 spin_unlock_irqrestore(&zone->lock, flags); 981 } 982 #endif /* CONFIG_PM */ 983 984 /* 985 * Free a 0-order page 986 */ 987 static void free_hot_cold_page(struct page *page, int cold) 988 { 989 struct zone *zone = page_zone(page); 990 struct per_cpu_pages *pcp; 991 unsigned long flags; 992 993 if (PageAnon(page)) 994 page->mapping = NULL; 995 if (free_pages_check(page)) 996 return; 997 998 if (!PageHighMem(page)) 999 debug_check_no_locks_freed(page_address(page), PAGE_SIZE); 1000 arch_free_page(page, 0); 1001 kernel_map_pages(page, 1, 0); 1002 1003 pcp = &zone_pcp(zone, get_cpu())->pcp; 1004 local_irq_save(flags); 1005 __count_vm_event(PGFREE); 1006 if (cold) 1007 list_add_tail(&page->lru, &pcp->list); 1008 else 1009 list_add(&page->lru, &pcp->list); 1010 set_page_private(page, get_pageblock_migratetype(page)); 1011 pcp->count++; 1012 if (pcp->count >= pcp->high) { 1013 free_pages_bulk(zone, pcp->batch, &pcp->list, 0); 1014 pcp->count -= pcp->batch; 1015 } 1016 local_irq_restore(flags); 1017 put_cpu(); 1018 } 1019 1020 void free_hot_page(struct page *page) 1021 { 1022 free_hot_cold_page(page, 0); 1023 } 1024 1025 void free_cold_page(struct page *page) 1026 { 1027 free_hot_cold_page(page, 1); 1028 } 1029 1030 /* 1031 * split_page takes a non-compound higher-order page, and splits it into 1032 * n (1<<order) sub-pages: page[0..n] 1033 * Each sub-page must be freed individually. 1034 * 1035 * Note: this is probably too low level an operation for use in drivers. 1036 * Please consult with lkml before using this in your driver. 1037 */ 1038 void split_page(struct page *page, unsigned int order) 1039 { 1040 int i; 1041 1042 VM_BUG_ON(PageCompound(page)); 1043 VM_BUG_ON(!page_count(page)); 1044 for (i = 1; i < (1 << order); i++) 1045 set_page_refcounted(page + i); 1046 } 1047 1048 /* 1049 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1050 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1051 * or two. 1052 */ 1053 static struct page *buffered_rmqueue(struct zonelist *zonelist, 1054 struct zone *zone, int order, gfp_t gfp_flags) 1055 { 1056 unsigned long flags; 1057 struct page *page; 1058 int cold = !!(gfp_flags & __GFP_COLD); 1059 int cpu; 1060 int migratetype = allocflags_to_migratetype(gfp_flags); 1061 1062 again: 1063 cpu = get_cpu(); 1064 if (likely(order == 0)) { 1065 struct per_cpu_pages *pcp; 1066 1067 pcp = &zone_pcp(zone, cpu)->pcp; 1068 local_irq_save(flags); 1069 if (!pcp->count) { 1070 pcp->count = rmqueue_bulk(zone, 0, 1071 pcp->batch, &pcp->list, migratetype); 1072 if (unlikely(!pcp->count)) 1073 goto failed; 1074 } 1075 1076 /* Find a page of the appropriate migrate type */ 1077 if (cold) { 1078 list_for_each_entry_reverse(page, &pcp->list, lru) 1079 if (page_private(page) == migratetype) 1080 break; 1081 } else { 1082 list_for_each_entry(page, &pcp->list, lru) 1083 if (page_private(page) == migratetype) 1084 break; 1085 } 1086 1087 /* Allocate more to the pcp list if necessary */ 1088 if (unlikely(&page->lru == &pcp->list)) { 1089 pcp->count += rmqueue_bulk(zone, 0, 1090 pcp->batch, &pcp->list, migratetype); 1091 page = list_entry(pcp->list.next, struct page, lru); 1092 } 1093 1094 list_del(&page->lru); 1095 pcp->count--; 1096 } else { 1097 spin_lock_irqsave(&zone->lock, flags); 1098 page = __rmqueue(zone, order, migratetype); 1099 spin_unlock(&zone->lock); 1100 if (!page) 1101 goto failed; 1102 } 1103 1104 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1105 zone_statistics(zonelist, zone); 1106 local_irq_restore(flags); 1107 put_cpu(); 1108 1109 VM_BUG_ON(bad_range(zone, page)); 1110 if (prep_new_page(page, order, gfp_flags)) 1111 goto again; 1112 return page; 1113 1114 failed: 1115 local_irq_restore(flags); 1116 put_cpu(); 1117 return NULL; 1118 } 1119 1120 #define ALLOC_NO_WATERMARKS 0x01 /* don't check watermarks at all */ 1121 #define ALLOC_WMARK_MIN 0x02 /* use pages_min watermark */ 1122 #define ALLOC_WMARK_LOW 0x04 /* use pages_low watermark */ 1123 #define ALLOC_WMARK_HIGH 0x08 /* use pages_high watermark */ 1124 #define ALLOC_HARDER 0x10 /* try to alloc harder */ 1125 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1126 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1127 1128 #ifdef CONFIG_FAIL_PAGE_ALLOC 1129 1130 static struct fail_page_alloc_attr { 1131 struct fault_attr attr; 1132 1133 u32 ignore_gfp_highmem; 1134 u32 ignore_gfp_wait; 1135 u32 min_order; 1136 1137 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1138 1139 struct dentry *ignore_gfp_highmem_file; 1140 struct dentry *ignore_gfp_wait_file; 1141 struct dentry *min_order_file; 1142 1143 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1144 1145 } fail_page_alloc = { 1146 .attr = FAULT_ATTR_INITIALIZER, 1147 .ignore_gfp_wait = 1, 1148 .ignore_gfp_highmem = 1, 1149 .min_order = 1, 1150 }; 1151 1152 static int __init setup_fail_page_alloc(char *str) 1153 { 1154 return setup_fault_attr(&fail_page_alloc.attr, str); 1155 } 1156 __setup("fail_page_alloc=", setup_fail_page_alloc); 1157 1158 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1159 { 1160 if (order < fail_page_alloc.min_order) 1161 return 0; 1162 if (gfp_mask & __GFP_NOFAIL) 1163 return 0; 1164 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1165 return 0; 1166 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1167 return 0; 1168 1169 return should_fail(&fail_page_alloc.attr, 1 << order); 1170 } 1171 1172 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1173 1174 static int __init fail_page_alloc_debugfs(void) 1175 { 1176 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1177 struct dentry *dir; 1178 int err; 1179 1180 err = init_fault_attr_dentries(&fail_page_alloc.attr, 1181 "fail_page_alloc"); 1182 if (err) 1183 return err; 1184 dir = fail_page_alloc.attr.dentries.dir; 1185 1186 fail_page_alloc.ignore_gfp_wait_file = 1187 debugfs_create_bool("ignore-gfp-wait", mode, dir, 1188 &fail_page_alloc.ignore_gfp_wait); 1189 1190 fail_page_alloc.ignore_gfp_highmem_file = 1191 debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1192 &fail_page_alloc.ignore_gfp_highmem); 1193 fail_page_alloc.min_order_file = 1194 debugfs_create_u32("min-order", mode, dir, 1195 &fail_page_alloc.min_order); 1196 1197 if (!fail_page_alloc.ignore_gfp_wait_file || 1198 !fail_page_alloc.ignore_gfp_highmem_file || 1199 !fail_page_alloc.min_order_file) { 1200 err = -ENOMEM; 1201 debugfs_remove(fail_page_alloc.ignore_gfp_wait_file); 1202 debugfs_remove(fail_page_alloc.ignore_gfp_highmem_file); 1203 debugfs_remove(fail_page_alloc.min_order_file); 1204 cleanup_fault_attr_dentries(&fail_page_alloc.attr); 1205 } 1206 1207 return err; 1208 } 1209 1210 late_initcall(fail_page_alloc_debugfs); 1211 1212 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1213 1214 #else /* CONFIG_FAIL_PAGE_ALLOC */ 1215 1216 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1217 { 1218 return 0; 1219 } 1220 1221 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 1222 1223 /* 1224 * Return 1 if free pages are above 'mark'. This takes into account the order 1225 * of the allocation. 1226 */ 1227 int zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1228 int classzone_idx, int alloc_flags) 1229 { 1230 /* free_pages my go negative - that's OK */ 1231 long min = mark; 1232 long free_pages = zone_page_state(z, NR_FREE_PAGES) - (1 << order) + 1; 1233 int o; 1234 1235 if (alloc_flags & ALLOC_HIGH) 1236 min -= min / 2; 1237 if (alloc_flags & ALLOC_HARDER) 1238 min -= min / 4; 1239 1240 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1241 return 0; 1242 for (o = 0; o < order; o++) { 1243 /* At the next order, this order's pages become unavailable */ 1244 free_pages -= z->free_area[o].nr_free << o; 1245 1246 /* Require fewer higher order pages to be free */ 1247 min >>= 1; 1248 1249 if (free_pages <= min) 1250 return 0; 1251 } 1252 return 1; 1253 } 1254 1255 #ifdef CONFIG_NUMA 1256 /* 1257 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1258 * skip over zones that are not allowed by the cpuset, or that have 1259 * been recently (in last second) found to be nearly full. See further 1260 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1261 * that have to skip over a lot of full or unallowed zones. 1262 * 1263 * If the zonelist cache is present in the passed in zonelist, then 1264 * returns a pointer to the allowed node mask (either the current 1265 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1266 * 1267 * If the zonelist cache is not available for this zonelist, does 1268 * nothing and returns NULL. 1269 * 1270 * If the fullzones BITMAP in the zonelist cache is stale (more than 1271 * a second since last zap'd) then we zap it out (clear its bits.) 1272 * 1273 * We hold off even calling zlc_setup, until after we've checked the 1274 * first zone in the zonelist, on the theory that most allocations will 1275 * be satisfied from that first zone, so best to examine that zone as 1276 * quickly as we can. 1277 */ 1278 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1279 { 1280 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1281 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1282 1283 zlc = zonelist->zlcache_ptr; 1284 if (!zlc) 1285 return NULL; 1286 1287 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1288 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1289 zlc->last_full_zap = jiffies; 1290 } 1291 1292 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1293 &cpuset_current_mems_allowed : 1294 &node_states[N_HIGH_MEMORY]; 1295 return allowednodes; 1296 } 1297 1298 /* 1299 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1300 * if it is worth looking at further for free memory: 1301 * 1) Check that the zone isn't thought to be full (doesn't have its 1302 * bit set in the zonelist_cache fullzones BITMAP). 1303 * 2) Check that the zones node (obtained from the zonelist_cache 1304 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1305 * Return true (non-zero) if zone is worth looking at further, or 1306 * else return false (zero) if it is not. 1307 * 1308 * This check -ignores- the distinction between various watermarks, 1309 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1310 * found to be full for any variation of these watermarks, it will 1311 * be considered full for up to one second by all requests, unless 1312 * we are so low on memory on all allowed nodes that we are forced 1313 * into the second scan of the zonelist. 1314 * 1315 * In the second scan we ignore this zonelist cache and exactly 1316 * apply the watermarks to all zones, even it is slower to do so. 1317 * We are low on memory in the second scan, and should leave no stone 1318 * unturned looking for a free page. 1319 */ 1320 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z, 1321 nodemask_t *allowednodes) 1322 { 1323 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1324 int i; /* index of *z in zonelist zones */ 1325 int n; /* node that zone *z is on */ 1326 1327 zlc = zonelist->zlcache_ptr; 1328 if (!zlc) 1329 return 1; 1330 1331 i = z - zonelist->zones; 1332 n = zlc->z_to_n[i]; 1333 1334 /* This zone is worth trying if it is allowed but not full */ 1335 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1336 } 1337 1338 /* 1339 * Given 'z' scanning a zonelist, set the corresponding bit in 1340 * zlc->fullzones, so that subsequent attempts to allocate a page 1341 * from that zone don't waste time re-examining it. 1342 */ 1343 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z) 1344 { 1345 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1346 int i; /* index of *z in zonelist zones */ 1347 1348 zlc = zonelist->zlcache_ptr; 1349 if (!zlc) 1350 return; 1351 1352 i = z - zonelist->zones; 1353 1354 set_bit(i, zlc->fullzones); 1355 } 1356 1357 #else /* CONFIG_NUMA */ 1358 1359 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1360 { 1361 return NULL; 1362 } 1363 1364 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zone **z, 1365 nodemask_t *allowednodes) 1366 { 1367 return 1; 1368 } 1369 1370 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zone **z) 1371 { 1372 } 1373 #endif /* CONFIG_NUMA */ 1374 1375 /* 1376 * get_page_from_freelist goes through the zonelist trying to allocate 1377 * a page. 1378 */ 1379 static struct page * 1380 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, 1381 struct zonelist *zonelist, int alloc_flags) 1382 { 1383 struct zone **z; 1384 struct page *page = NULL; 1385 int classzone_idx = zone_idx(zonelist->zones[0]); 1386 struct zone *zone; 1387 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1388 int zlc_active = 0; /* set if using zonelist_cache */ 1389 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1390 enum zone_type highest_zoneidx = -1; /* Gets set for policy zonelists */ 1391 1392 zonelist_scan: 1393 /* 1394 * Scan zonelist, looking for a zone with enough free. 1395 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1396 */ 1397 z = zonelist->zones; 1398 1399 do { 1400 /* 1401 * In NUMA, this could be a policy zonelist which contains 1402 * zones that may not be allowed by the current gfp_mask. 1403 * Check the zone is allowed by the current flags 1404 */ 1405 if (unlikely(alloc_should_filter_zonelist(zonelist))) { 1406 if (highest_zoneidx == -1) 1407 highest_zoneidx = gfp_zone(gfp_mask); 1408 if (zone_idx(*z) > highest_zoneidx) 1409 continue; 1410 } 1411 1412 if (NUMA_BUILD && zlc_active && 1413 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1414 continue; 1415 zone = *z; 1416 if ((alloc_flags & ALLOC_CPUSET) && 1417 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1418 goto try_next_zone; 1419 1420 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1421 unsigned long mark; 1422 if (alloc_flags & ALLOC_WMARK_MIN) 1423 mark = zone->pages_min; 1424 else if (alloc_flags & ALLOC_WMARK_LOW) 1425 mark = zone->pages_low; 1426 else 1427 mark = zone->pages_high; 1428 if (!zone_watermark_ok(zone, order, mark, 1429 classzone_idx, alloc_flags)) { 1430 if (!zone_reclaim_mode || 1431 !zone_reclaim(zone, gfp_mask, order)) 1432 goto this_zone_full; 1433 } 1434 } 1435 1436 page = buffered_rmqueue(zonelist, zone, order, gfp_mask); 1437 if (page) 1438 break; 1439 this_zone_full: 1440 if (NUMA_BUILD) 1441 zlc_mark_zone_full(zonelist, z); 1442 try_next_zone: 1443 if (NUMA_BUILD && !did_zlc_setup) { 1444 /* we do zlc_setup after the first zone is tried */ 1445 allowednodes = zlc_setup(zonelist, alloc_flags); 1446 zlc_active = 1; 1447 did_zlc_setup = 1; 1448 } 1449 } while (*(++z) != NULL); 1450 1451 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1452 /* Disable zlc cache for second zonelist scan */ 1453 zlc_active = 0; 1454 goto zonelist_scan; 1455 } 1456 return page; 1457 } 1458 1459 /* 1460 * This is the 'heart' of the zoned buddy allocator. 1461 */ 1462 struct page * 1463 __alloc_pages(gfp_t gfp_mask, unsigned int order, 1464 struct zonelist *zonelist) 1465 { 1466 const gfp_t wait = gfp_mask & __GFP_WAIT; 1467 struct zone **z; 1468 struct page *page; 1469 struct reclaim_state reclaim_state; 1470 struct task_struct *p = current; 1471 int do_retry; 1472 int alloc_flags; 1473 int did_some_progress; 1474 1475 might_sleep_if(wait); 1476 1477 if (should_fail_alloc_page(gfp_mask, order)) 1478 return NULL; 1479 1480 restart: 1481 z = zonelist->zones; /* the list of zones suitable for gfp_mask */ 1482 1483 if (unlikely(*z == NULL)) { 1484 /* 1485 * Happens if we have an empty zonelist as a result of 1486 * GFP_THISNODE being used on a memoryless node 1487 */ 1488 return NULL; 1489 } 1490 1491 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order, 1492 zonelist, ALLOC_WMARK_LOW|ALLOC_CPUSET); 1493 if (page) 1494 goto got_pg; 1495 1496 /* 1497 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 1498 * __GFP_NOWARN set) should not cause reclaim since the subsystem 1499 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 1500 * using a larger set of nodes after it has established that the 1501 * allowed per node queues are empty and that nodes are 1502 * over allocated. 1503 */ 1504 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 1505 goto nopage; 1506 1507 for (z = zonelist->zones; *z; z++) 1508 wakeup_kswapd(*z, order); 1509 1510 /* 1511 * OK, we're below the kswapd watermark and have kicked background 1512 * reclaim. Now things get more complex, so set up alloc_flags according 1513 * to how we want to proceed. 1514 * 1515 * The caller may dip into page reserves a bit more if the caller 1516 * cannot run direct reclaim, or if the caller has realtime scheduling 1517 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 1518 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 1519 */ 1520 alloc_flags = ALLOC_WMARK_MIN; 1521 if ((unlikely(rt_task(p)) && !in_interrupt()) || !wait) 1522 alloc_flags |= ALLOC_HARDER; 1523 if (gfp_mask & __GFP_HIGH) 1524 alloc_flags |= ALLOC_HIGH; 1525 if (wait) 1526 alloc_flags |= ALLOC_CPUSET; 1527 1528 /* 1529 * Go through the zonelist again. Let __GFP_HIGH and allocations 1530 * coming from realtime tasks go deeper into reserves. 1531 * 1532 * This is the last chance, in general, before the goto nopage. 1533 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 1534 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1535 */ 1536 page = get_page_from_freelist(gfp_mask, order, zonelist, alloc_flags); 1537 if (page) 1538 goto got_pg; 1539 1540 /* This allocation should allow future memory freeing. */ 1541 1542 rebalance: 1543 if (((p->flags & PF_MEMALLOC) || unlikely(test_thread_flag(TIF_MEMDIE))) 1544 && !in_interrupt()) { 1545 if (!(gfp_mask & __GFP_NOMEMALLOC)) { 1546 nofail_alloc: 1547 /* go through the zonelist yet again, ignoring mins */ 1548 page = get_page_from_freelist(gfp_mask, order, 1549 zonelist, ALLOC_NO_WATERMARKS); 1550 if (page) 1551 goto got_pg; 1552 if (gfp_mask & __GFP_NOFAIL) { 1553 congestion_wait(WRITE, HZ/50); 1554 goto nofail_alloc; 1555 } 1556 } 1557 goto nopage; 1558 } 1559 1560 /* Atomic allocations - we can't balance anything */ 1561 if (!wait) 1562 goto nopage; 1563 1564 cond_resched(); 1565 1566 /* We now go into synchronous reclaim */ 1567 cpuset_memory_pressure_bump(); 1568 p->flags |= PF_MEMALLOC; 1569 reclaim_state.reclaimed_slab = 0; 1570 p->reclaim_state = &reclaim_state; 1571 1572 did_some_progress = try_to_free_pages(zonelist->zones, order, gfp_mask); 1573 1574 p->reclaim_state = NULL; 1575 p->flags &= ~PF_MEMALLOC; 1576 1577 cond_resched(); 1578 1579 if (order != 0) 1580 drain_all_pages(); 1581 1582 if (likely(did_some_progress)) { 1583 page = get_page_from_freelist(gfp_mask, order, 1584 zonelist, alloc_flags); 1585 if (page) 1586 goto got_pg; 1587 } else if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 1588 if (!try_set_zone_oom(zonelist)) { 1589 schedule_timeout_uninterruptible(1); 1590 goto restart; 1591 } 1592 1593 /* 1594 * Go through the zonelist yet one more time, keep 1595 * very high watermark here, this is only to catch 1596 * a parallel oom killing, we must fail if we're still 1597 * under heavy pressure. 1598 */ 1599 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, order, 1600 zonelist, ALLOC_WMARK_HIGH|ALLOC_CPUSET); 1601 if (page) { 1602 clear_zonelist_oom(zonelist); 1603 goto got_pg; 1604 } 1605 1606 /* The OOM killer will not help higher order allocs so fail */ 1607 if (order > PAGE_ALLOC_COSTLY_ORDER) { 1608 clear_zonelist_oom(zonelist); 1609 goto nopage; 1610 } 1611 1612 out_of_memory(zonelist, gfp_mask, order); 1613 clear_zonelist_oom(zonelist); 1614 goto restart; 1615 } 1616 1617 /* 1618 * Don't let big-order allocations loop unless the caller explicitly 1619 * requests that. Wait for some write requests to complete then retry. 1620 * 1621 * In this implementation, __GFP_REPEAT means __GFP_NOFAIL for order 1622 * <= 3, but that may not be true in other implementations. 1623 */ 1624 do_retry = 0; 1625 if (!(gfp_mask & __GFP_NORETRY)) { 1626 if ((order <= PAGE_ALLOC_COSTLY_ORDER) || 1627 (gfp_mask & __GFP_REPEAT)) 1628 do_retry = 1; 1629 if (gfp_mask & __GFP_NOFAIL) 1630 do_retry = 1; 1631 } 1632 if (do_retry) { 1633 congestion_wait(WRITE, HZ/50); 1634 goto rebalance; 1635 } 1636 1637 nopage: 1638 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit()) { 1639 printk(KERN_WARNING "%s: page allocation failure." 1640 " order:%d, mode:0x%x\n", 1641 p->comm, order, gfp_mask); 1642 dump_stack(); 1643 show_mem(); 1644 } 1645 got_pg: 1646 return page; 1647 } 1648 1649 EXPORT_SYMBOL(__alloc_pages); 1650 1651 /* 1652 * Common helper functions. 1653 */ 1654 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 1655 { 1656 struct page * page; 1657 page = alloc_pages(gfp_mask, order); 1658 if (!page) 1659 return 0; 1660 return (unsigned long) page_address(page); 1661 } 1662 1663 EXPORT_SYMBOL(__get_free_pages); 1664 1665 unsigned long get_zeroed_page(gfp_t gfp_mask) 1666 { 1667 struct page * page; 1668 1669 /* 1670 * get_zeroed_page() returns a 32-bit address, which cannot represent 1671 * a highmem page 1672 */ 1673 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 1674 1675 page = alloc_pages(gfp_mask | __GFP_ZERO, 0); 1676 if (page) 1677 return (unsigned long) page_address(page); 1678 return 0; 1679 } 1680 1681 EXPORT_SYMBOL(get_zeroed_page); 1682 1683 void __pagevec_free(struct pagevec *pvec) 1684 { 1685 int i = pagevec_count(pvec); 1686 1687 while (--i >= 0) 1688 free_hot_cold_page(pvec->pages[i], pvec->cold); 1689 } 1690 1691 void __free_pages(struct page *page, unsigned int order) 1692 { 1693 if (put_page_testzero(page)) { 1694 if (order == 0) 1695 free_hot_page(page); 1696 else 1697 __free_pages_ok(page, order); 1698 } 1699 } 1700 1701 EXPORT_SYMBOL(__free_pages); 1702 1703 void free_pages(unsigned long addr, unsigned int order) 1704 { 1705 if (addr != 0) { 1706 VM_BUG_ON(!virt_addr_valid((void *)addr)); 1707 __free_pages(virt_to_page((void *)addr), order); 1708 } 1709 } 1710 1711 EXPORT_SYMBOL(free_pages); 1712 1713 static unsigned int nr_free_zone_pages(int offset) 1714 { 1715 /* Just pick one node, since fallback list is circular */ 1716 pg_data_t *pgdat = NODE_DATA(numa_node_id()); 1717 unsigned int sum = 0; 1718 1719 struct zonelist *zonelist = pgdat->node_zonelists + offset; 1720 struct zone **zonep = zonelist->zones; 1721 struct zone *zone; 1722 1723 for (zone = *zonep++; zone; zone = *zonep++) { 1724 unsigned long size = zone->present_pages; 1725 unsigned long high = zone->pages_high; 1726 if (size > high) 1727 sum += size - high; 1728 } 1729 1730 return sum; 1731 } 1732 1733 /* 1734 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 1735 */ 1736 unsigned int nr_free_buffer_pages(void) 1737 { 1738 return nr_free_zone_pages(gfp_zone(GFP_USER)); 1739 } 1740 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 1741 1742 /* 1743 * Amount of free RAM allocatable within all zones 1744 */ 1745 unsigned int nr_free_pagecache_pages(void) 1746 { 1747 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 1748 } 1749 1750 static inline void show_node(struct zone *zone) 1751 { 1752 if (NUMA_BUILD) 1753 printk("Node %d ", zone_to_nid(zone)); 1754 } 1755 1756 void si_meminfo(struct sysinfo *val) 1757 { 1758 val->totalram = totalram_pages; 1759 val->sharedram = 0; 1760 val->freeram = global_page_state(NR_FREE_PAGES); 1761 val->bufferram = nr_blockdev_pages(); 1762 val->totalhigh = totalhigh_pages; 1763 val->freehigh = nr_free_highpages(); 1764 val->mem_unit = PAGE_SIZE; 1765 } 1766 1767 EXPORT_SYMBOL(si_meminfo); 1768 1769 #ifdef CONFIG_NUMA 1770 void si_meminfo_node(struct sysinfo *val, int nid) 1771 { 1772 pg_data_t *pgdat = NODE_DATA(nid); 1773 1774 val->totalram = pgdat->node_present_pages; 1775 val->freeram = node_page_state(nid, NR_FREE_PAGES); 1776 #ifdef CONFIG_HIGHMEM 1777 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 1778 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 1779 NR_FREE_PAGES); 1780 #else 1781 val->totalhigh = 0; 1782 val->freehigh = 0; 1783 #endif 1784 val->mem_unit = PAGE_SIZE; 1785 } 1786 #endif 1787 1788 #define K(x) ((x) << (PAGE_SHIFT-10)) 1789 1790 /* 1791 * Show free area list (used inside shift_scroll-lock stuff) 1792 * We also calculate the percentage fragmentation. We do this by counting the 1793 * memory on each free list with the exception of the first item on the list. 1794 */ 1795 void show_free_areas(void) 1796 { 1797 int cpu; 1798 struct zone *zone; 1799 1800 for_each_zone(zone) { 1801 if (!populated_zone(zone)) 1802 continue; 1803 1804 show_node(zone); 1805 printk("%s per-cpu:\n", zone->name); 1806 1807 for_each_online_cpu(cpu) { 1808 struct per_cpu_pageset *pageset; 1809 1810 pageset = zone_pcp(zone, cpu); 1811 1812 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 1813 cpu, pageset->pcp.high, 1814 pageset->pcp.batch, pageset->pcp.count); 1815 } 1816 } 1817 1818 printk("Active:%lu inactive:%lu dirty:%lu writeback:%lu unstable:%lu\n" 1819 " free:%lu slab:%lu mapped:%lu pagetables:%lu bounce:%lu\n", 1820 global_page_state(NR_ACTIVE), 1821 global_page_state(NR_INACTIVE), 1822 global_page_state(NR_FILE_DIRTY), 1823 global_page_state(NR_WRITEBACK), 1824 global_page_state(NR_UNSTABLE_NFS), 1825 global_page_state(NR_FREE_PAGES), 1826 global_page_state(NR_SLAB_RECLAIMABLE) + 1827 global_page_state(NR_SLAB_UNRECLAIMABLE), 1828 global_page_state(NR_FILE_MAPPED), 1829 global_page_state(NR_PAGETABLE), 1830 global_page_state(NR_BOUNCE)); 1831 1832 for_each_zone(zone) { 1833 int i; 1834 1835 if (!populated_zone(zone)) 1836 continue; 1837 1838 show_node(zone); 1839 printk("%s" 1840 " free:%lukB" 1841 " min:%lukB" 1842 " low:%lukB" 1843 " high:%lukB" 1844 " active:%lukB" 1845 " inactive:%lukB" 1846 " present:%lukB" 1847 " pages_scanned:%lu" 1848 " all_unreclaimable? %s" 1849 "\n", 1850 zone->name, 1851 K(zone_page_state(zone, NR_FREE_PAGES)), 1852 K(zone->pages_min), 1853 K(zone->pages_low), 1854 K(zone->pages_high), 1855 K(zone_page_state(zone, NR_ACTIVE)), 1856 K(zone_page_state(zone, NR_INACTIVE)), 1857 K(zone->present_pages), 1858 zone->pages_scanned, 1859 (zone_is_all_unreclaimable(zone) ? "yes" : "no") 1860 ); 1861 printk("lowmem_reserve[]:"); 1862 for (i = 0; i < MAX_NR_ZONES; i++) 1863 printk(" %lu", zone->lowmem_reserve[i]); 1864 printk("\n"); 1865 } 1866 1867 for_each_zone(zone) { 1868 unsigned long nr[MAX_ORDER], flags, order, total = 0; 1869 1870 if (!populated_zone(zone)) 1871 continue; 1872 1873 show_node(zone); 1874 printk("%s: ", zone->name); 1875 1876 spin_lock_irqsave(&zone->lock, flags); 1877 for (order = 0; order < MAX_ORDER; order++) { 1878 nr[order] = zone->free_area[order].nr_free; 1879 total += nr[order] << order; 1880 } 1881 spin_unlock_irqrestore(&zone->lock, flags); 1882 for (order = 0; order < MAX_ORDER; order++) 1883 printk("%lu*%lukB ", nr[order], K(1UL) << order); 1884 printk("= %lukB\n", K(total)); 1885 } 1886 1887 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 1888 1889 show_swap_cache_info(); 1890 } 1891 1892 /* 1893 * Builds allocation fallback zone lists. 1894 * 1895 * Add all populated zones of a node to the zonelist. 1896 */ 1897 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 1898 int nr_zones, enum zone_type zone_type) 1899 { 1900 struct zone *zone; 1901 1902 BUG_ON(zone_type >= MAX_NR_ZONES); 1903 zone_type++; 1904 1905 do { 1906 zone_type--; 1907 zone = pgdat->node_zones + zone_type; 1908 if (populated_zone(zone)) { 1909 zonelist->zones[nr_zones++] = zone; 1910 check_highest_zone(zone_type); 1911 } 1912 1913 } while (zone_type); 1914 return nr_zones; 1915 } 1916 1917 1918 /* 1919 * zonelist_order: 1920 * 0 = automatic detection of better ordering. 1921 * 1 = order by ([node] distance, -zonetype) 1922 * 2 = order by (-zonetype, [node] distance) 1923 * 1924 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 1925 * the same zonelist. So only NUMA can configure this param. 1926 */ 1927 #define ZONELIST_ORDER_DEFAULT 0 1928 #define ZONELIST_ORDER_NODE 1 1929 #define ZONELIST_ORDER_ZONE 2 1930 1931 /* zonelist order in the kernel. 1932 * set_zonelist_order() will set this to NODE or ZONE. 1933 */ 1934 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 1935 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 1936 1937 1938 #ifdef CONFIG_NUMA 1939 /* The value user specified ....changed by config */ 1940 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 1941 /* string for sysctl */ 1942 #define NUMA_ZONELIST_ORDER_LEN 16 1943 char numa_zonelist_order[16] = "default"; 1944 1945 /* 1946 * interface for configure zonelist ordering. 1947 * command line option "numa_zonelist_order" 1948 * = "[dD]efault - default, automatic configuration. 1949 * = "[nN]ode - order by node locality, then by zone within node 1950 * = "[zZ]one - order by zone, then by locality within zone 1951 */ 1952 1953 static int __parse_numa_zonelist_order(char *s) 1954 { 1955 if (*s == 'd' || *s == 'D') { 1956 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 1957 } else if (*s == 'n' || *s == 'N') { 1958 user_zonelist_order = ZONELIST_ORDER_NODE; 1959 } else if (*s == 'z' || *s == 'Z') { 1960 user_zonelist_order = ZONELIST_ORDER_ZONE; 1961 } else { 1962 printk(KERN_WARNING 1963 "Ignoring invalid numa_zonelist_order value: " 1964 "%s\n", s); 1965 return -EINVAL; 1966 } 1967 return 0; 1968 } 1969 1970 static __init int setup_numa_zonelist_order(char *s) 1971 { 1972 if (s) 1973 return __parse_numa_zonelist_order(s); 1974 return 0; 1975 } 1976 early_param("numa_zonelist_order", setup_numa_zonelist_order); 1977 1978 /* 1979 * sysctl handler for numa_zonelist_order 1980 */ 1981 int numa_zonelist_order_handler(ctl_table *table, int write, 1982 struct file *file, void __user *buffer, size_t *length, 1983 loff_t *ppos) 1984 { 1985 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 1986 int ret; 1987 1988 if (write) 1989 strncpy(saved_string, (char*)table->data, 1990 NUMA_ZONELIST_ORDER_LEN); 1991 ret = proc_dostring(table, write, file, buffer, length, ppos); 1992 if (ret) 1993 return ret; 1994 if (write) { 1995 int oldval = user_zonelist_order; 1996 if (__parse_numa_zonelist_order((char*)table->data)) { 1997 /* 1998 * bogus value. restore saved string 1999 */ 2000 strncpy((char*)table->data, saved_string, 2001 NUMA_ZONELIST_ORDER_LEN); 2002 user_zonelist_order = oldval; 2003 } else if (oldval != user_zonelist_order) 2004 build_all_zonelists(); 2005 } 2006 return 0; 2007 } 2008 2009 2010 #define MAX_NODE_LOAD (num_online_nodes()) 2011 static int node_load[MAX_NUMNODES]; 2012 2013 /** 2014 * find_next_best_node - find the next node that should appear in a given node's fallback list 2015 * @node: node whose fallback list we're appending 2016 * @used_node_mask: nodemask_t of already used nodes 2017 * 2018 * We use a number of factors to determine which is the next node that should 2019 * appear on a given node's fallback list. The node should not have appeared 2020 * already in @node's fallback list, and it should be the next closest node 2021 * according to the distance array (which contains arbitrary distance values 2022 * from each node to each node in the system), and should also prefer nodes 2023 * with no CPUs, since presumably they'll have very little allocation pressure 2024 * on them otherwise. 2025 * It returns -1 if no node is found. 2026 */ 2027 static int find_next_best_node(int node, nodemask_t *used_node_mask) 2028 { 2029 int n, val; 2030 int min_val = INT_MAX; 2031 int best_node = -1; 2032 2033 /* Use the local node if we haven't already */ 2034 if (!node_isset(node, *used_node_mask)) { 2035 node_set(node, *used_node_mask); 2036 return node; 2037 } 2038 2039 for_each_node_state(n, N_HIGH_MEMORY) { 2040 cpumask_t tmp; 2041 2042 /* Don't want a node to appear more than once */ 2043 if (node_isset(n, *used_node_mask)) 2044 continue; 2045 2046 /* Use the distance array to find the distance */ 2047 val = node_distance(node, n); 2048 2049 /* Penalize nodes under us ("prefer the next node") */ 2050 val += (n < node); 2051 2052 /* Give preference to headless and unused nodes */ 2053 tmp = node_to_cpumask(n); 2054 if (!cpus_empty(tmp)) 2055 val += PENALTY_FOR_NODE_WITH_CPUS; 2056 2057 /* Slight preference for less loaded node */ 2058 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2059 val += node_load[n]; 2060 2061 if (val < min_val) { 2062 min_val = val; 2063 best_node = n; 2064 } 2065 } 2066 2067 if (best_node >= 0) 2068 node_set(best_node, *used_node_mask); 2069 2070 return best_node; 2071 } 2072 2073 2074 /* 2075 * Build zonelists ordered by node and zones within node. 2076 * This results in maximum locality--normal zone overflows into local 2077 * DMA zone, if any--but risks exhausting DMA zone. 2078 */ 2079 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 2080 { 2081 enum zone_type i; 2082 int j; 2083 struct zonelist *zonelist; 2084 2085 for (i = 0; i < MAX_NR_ZONES; i++) { 2086 zonelist = pgdat->node_zonelists + i; 2087 for (j = 0; zonelist->zones[j] != NULL; j++) 2088 ; 2089 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i); 2090 zonelist->zones[j] = NULL; 2091 } 2092 } 2093 2094 /* 2095 * Build gfp_thisnode zonelists 2096 */ 2097 static void build_thisnode_zonelists(pg_data_t *pgdat) 2098 { 2099 enum zone_type i; 2100 int j; 2101 struct zonelist *zonelist; 2102 2103 for (i = 0; i < MAX_NR_ZONES; i++) { 2104 zonelist = pgdat->node_zonelists + MAX_NR_ZONES + i; 2105 j = build_zonelists_node(pgdat, zonelist, 0, i); 2106 zonelist->zones[j] = NULL; 2107 } 2108 } 2109 2110 /* 2111 * Build zonelists ordered by zone and nodes within zones. 2112 * This results in conserving DMA zone[s] until all Normal memory is 2113 * exhausted, but results in overflowing to remote node while memory 2114 * may still exist in local DMA zone. 2115 */ 2116 static int node_order[MAX_NUMNODES]; 2117 2118 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 2119 { 2120 enum zone_type i; 2121 int pos, j, node; 2122 int zone_type; /* needs to be signed */ 2123 struct zone *z; 2124 struct zonelist *zonelist; 2125 2126 for (i = 0; i < MAX_NR_ZONES; i++) { 2127 zonelist = pgdat->node_zonelists + i; 2128 pos = 0; 2129 for (zone_type = i; zone_type >= 0; zone_type--) { 2130 for (j = 0; j < nr_nodes; j++) { 2131 node = node_order[j]; 2132 z = &NODE_DATA(node)->node_zones[zone_type]; 2133 if (populated_zone(z)) { 2134 zonelist->zones[pos++] = z; 2135 check_highest_zone(zone_type); 2136 } 2137 } 2138 } 2139 zonelist->zones[pos] = NULL; 2140 } 2141 } 2142 2143 static int default_zonelist_order(void) 2144 { 2145 int nid, zone_type; 2146 unsigned long low_kmem_size,total_size; 2147 struct zone *z; 2148 int average_size; 2149 /* 2150 * ZONE_DMA and ZONE_DMA32 can be very small area in the sytem. 2151 * If they are really small and used heavily, the system can fall 2152 * into OOM very easily. 2153 * This function detect ZONE_DMA/DMA32 size and confgigures zone order. 2154 */ 2155 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 2156 low_kmem_size = 0; 2157 total_size = 0; 2158 for_each_online_node(nid) { 2159 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2160 z = &NODE_DATA(nid)->node_zones[zone_type]; 2161 if (populated_zone(z)) { 2162 if (zone_type < ZONE_NORMAL) 2163 low_kmem_size += z->present_pages; 2164 total_size += z->present_pages; 2165 } 2166 } 2167 } 2168 if (!low_kmem_size || /* there are no DMA area. */ 2169 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 2170 return ZONELIST_ORDER_NODE; 2171 /* 2172 * look into each node's config. 2173 * If there is a node whose DMA/DMA32 memory is very big area on 2174 * local memory, NODE_ORDER may be suitable. 2175 */ 2176 average_size = total_size / 2177 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 2178 for_each_online_node(nid) { 2179 low_kmem_size = 0; 2180 total_size = 0; 2181 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 2182 z = &NODE_DATA(nid)->node_zones[zone_type]; 2183 if (populated_zone(z)) { 2184 if (zone_type < ZONE_NORMAL) 2185 low_kmem_size += z->present_pages; 2186 total_size += z->present_pages; 2187 } 2188 } 2189 if (low_kmem_size && 2190 total_size > average_size && /* ignore small node */ 2191 low_kmem_size > total_size * 70/100) 2192 return ZONELIST_ORDER_NODE; 2193 } 2194 return ZONELIST_ORDER_ZONE; 2195 } 2196 2197 static void set_zonelist_order(void) 2198 { 2199 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 2200 current_zonelist_order = default_zonelist_order(); 2201 else 2202 current_zonelist_order = user_zonelist_order; 2203 } 2204 2205 static void build_zonelists(pg_data_t *pgdat) 2206 { 2207 int j, node, load; 2208 enum zone_type i; 2209 nodemask_t used_mask; 2210 int local_node, prev_node; 2211 struct zonelist *zonelist; 2212 int order = current_zonelist_order; 2213 2214 /* initialize zonelists */ 2215 for (i = 0; i < MAX_ZONELISTS; i++) { 2216 zonelist = pgdat->node_zonelists + i; 2217 zonelist->zones[0] = NULL; 2218 } 2219 2220 /* NUMA-aware ordering of nodes */ 2221 local_node = pgdat->node_id; 2222 load = num_online_nodes(); 2223 prev_node = local_node; 2224 nodes_clear(used_mask); 2225 2226 memset(node_load, 0, sizeof(node_load)); 2227 memset(node_order, 0, sizeof(node_order)); 2228 j = 0; 2229 2230 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 2231 int distance = node_distance(local_node, node); 2232 2233 /* 2234 * If another node is sufficiently far away then it is better 2235 * to reclaim pages in a zone before going off node. 2236 */ 2237 if (distance > RECLAIM_DISTANCE) 2238 zone_reclaim_mode = 1; 2239 2240 /* 2241 * We don't want to pressure a particular node. 2242 * So adding penalty to the first node in same 2243 * distance group to make it round-robin. 2244 */ 2245 if (distance != node_distance(local_node, prev_node)) 2246 node_load[node] = load; 2247 2248 prev_node = node; 2249 load--; 2250 if (order == ZONELIST_ORDER_NODE) 2251 build_zonelists_in_node_order(pgdat, node); 2252 else 2253 node_order[j++] = node; /* remember order */ 2254 } 2255 2256 if (order == ZONELIST_ORDER_ZONE) { 2257 /* calculate node order -- i.e., DMA last! */ 2258 build_zonelists_in_zone_order(pgdat, j); 2259 } 2260 2261 build_thisnode_zonelists(pgdat); 2262 } 2263 2264 /* Construct the zonelist performance cache - see further mmzone.h */ 2265 static void build_zonelist_cache(pg_data_t *pgdat) 2266 { 2267 int i; 2268 2269 for (i = 0; i < MAX_NR_ZONES; i++) { 2270 struct zonelist *zonelist; 2271 struct zonelist_cache *zlc; 2272 struct zone **z; 2273 2274 zonelist = pgdat->node_zonelists + i; 2275 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 2276 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2277 for (z = zonelist->zones; *z; z++) 2278 zlc->z_to_n[z - zonelist->zones] = zone_to_nid(*z); 2279 } 2280 } 2281 2282 2283 #else /* CONFIG_NUMA */ 2284 2285 static void set_zonelist_order(void) 2286 { 2287 current_zonelist_order = ZONELIST_ORDER_ZONE; 2288 } 2289 2290 static void build_zonelists(pg_data_t *pgdat) 2291 { 2292 int node, local_node; 2293 enum zone_type i,j; 2294 2295 local_node = pgdat->node_id; 2296 for (i = 0; i < MAX_NR_ZONES; i++) { 2297 struct zonelist *zonelist; 2298 2299 zonelist = pgdat->node_zonelists + i; 2300 2301 j = build_zonelists_node(pgdat, zonelist, 0, i); 2302 /* 2303 * Now we build the zonelist so that it contains the zones 2304 * of all the other nodes. 2305 * We don't want to pressure a particular node, so when 2306 * building the zones for node N, we make sure that the 2307 * zones coming right after the local ones are those from 2308 * node N+1 (modulo N) 2309 */ 2310 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 2311 if (!node_online(node)) 2312 continue; 2313 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i); 2314 } 2315 for (node = 0; node < local_node; node++) { 2316 if (!node_online(node)) 2317 continue; 2318 j = build_zonelists_node(NODE_DATA(node), zonelist, j, i); 2319 } 2320 2321 zonelist->zones[j] = NULL; 2322 } 2323 } 2324 2325 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 2326 static void build_zonelist_cache(pg_data_t *pgdat) 2327 { 2328 int i; 2329 2330 for (i = 0; i < MAX_NR_ZONES; i++) 2331 pgdat->node_zonelists[i].zlcache_ptr = NULL; 2332 } 2333 2334 #endif /* CONFIG_NUMA */ 2335 2336 /* return values int ....just for stop_machine_run() */ 2337 static int __build_all_zonelists(void *dummy) 2338 { 2339 int nid; 2340 2341 for_each_online_node(nid) { 2342 pg_data_t *pgdat = NODE_DATA(nid); 2343 2344 build_zonelists(pgdat); 2345 build_zonelist_cache(pgdat); 2346 } 2347 return 0; 2348 } 2349 2350 void build_all_zonelists(void) 2351 { 2352 set_zonelist_order(); 2353 2354 if (system_state == SYSTEM_BOOTING) { 2355 __build_all_zonelists(NULL); 2356 cpuset_init_current_mems_allowed(); 2357 } else { 2358 /* we have to stop all cpus to guarantee there is no user 2359 of zonelist */ 2360 stop_machine_run(__build_all_zonelists, NULL, NR_CPUS); 2361 /* cpuset refresh routine should be here */ 2362 } 2363 vm_total_pages = nr_free_pagecache_pages(); 2364 /* 2365 * Disable grouping by mobility if the number of pages in the 2366 * system is too low to allow the mechanism to work. It would be 2367 * more accurate, but expensive to check per-zone. This check is 2368 * made on memory-hotadd so a system can start with mobility 2369 * disabled and enable it later 2370 */ 2371 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 2372 page_group_by_mobility_disabled = 1; 2373 else 2374 page_group_by_mobility_disabled = 0; 2375 2376 printk("Built %i zonelists in %s order, mobility grouping %s. " 2377 "Total pages: %ld\n", 2378 num_online_nodes(), 2379 zonelist_order_name[current_zonelist_order], 2380 page_group_by_mobility_disabled ? "off" : "on", 2381 vm_total_pages); 2382 #ifdef CONFIG_NUMA 2383 printk("Policy zone: %s\n", zone_names[policy_zone]); 2384 #endif 2385 } 2386 2387 /* 2388 * Helper functions to size the waitqueue hash table. 2389 * Essentially these want to choose hash table sizes sufficiently 2390 * large so that collisions trying to wait on pages are rare. 2391 * But in fact, the number of active page waitqueues on typical 2392 * systems is ridiculously low, less than 200. So this is even 2393 * conservative, even though it seems large. 2394 * 2395 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 2396 * waitqueues, i.e. the size of the waitq table given the number of pages. 2397 */ 2398 #define PAGES_PER_WAITQUEUE 256 2399 2400 #ifndef CONFIG_MEMORY_HOTPLUG 2401 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2402 { 2403 unsigned long size = 1; 2404 2405 pages /= PAGES_PER_WAITQUEUE; 2406 2407 while (size < pages) 2408 size <<= 1; 2409 2410 /* 2411 * Once we have dozens or even hundreds of threads sleeping 2412 * on IO we've got bigger problems than wait queue collision. 2413 * Limit the size of the wait table to a reasonable size. 2414 */ 2415 size = min(size, 4096UL); 2416 2417 return max(size, 4UL); 2418 } 2419 #else 2420 /* 2421 * A zone's size might be changed by hot-add, so it is not possible to determine 2422 * a suitable size for its wait_table. So we use the maximum size now. 2423 * 2424 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 2425 * 2426 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 2427 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 2428 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 2429 * 2430 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 2431 * or more by the traditional way. (See above). It equals: 2432 * 2433 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 2434 * ia64(16K page size) : = ( 8G + 4M)byte. 2435 * powerpc (64K page size) : = (32G +16M)byte. 2436 */ 2437 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 2438 { 2439 return 4096UL; 2440 } 2441 #endif 2442 2443 /* 2444 * This is an integer logarithm so that shifts can be used later 2445 * to extract the more random high bits from the multiplicative 2446 * hash function before the remainder is taken. 2447 */ 2448 static inline unsigned long wait_table_bits(unsigned long size) 2449 { 2450 return ffz(~size); 2451 } 2452 2453 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 2454 2455 /* 2456 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 2457 * of blocks reserved is based on zone->pages_min. The memory within the 2458 * reserve will tend to store contiguous free pages. Setting min_free_kbytes 2459 * higher will lead to a bigger reserve which will get freed as contiguous 2460 * blocks as reclaim kicks in 2461 */ 2462 static void setup_zone_migrate_reserve(struct zone *zone) 2463 { 2464 unsigned long start_pfn, pfn, end_pfn; 2465 struct page *page; 2466 unsigned long reserve, block_migratetype; 2467 2468 /* Get the start pfn, end pfn and the number of blocks to reserve */ 2469 start_pfn = zone->zone_start_pfn; 2470 end_pfn = start_pfn + zone->spanned_pages; 2471 reserve = roundup(zone->pages_min, pageblock_nr_pages) >> 2472 pageblock_order; 2473 2474 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 2475 if (!pfn_valid(pfn)) 2476 continue; 2477 page = pfn_to_page(pfn); 2478 2479 /* Blocks with reserved pages will never free, skip them. */ 2480 if (PageReserved(page)) 2481 continue; 2482 2483 block_migratetype = get_pageblock_migratetype(page); 2484 2485 /* If this block is reserved, account for it */ 2486 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) { 2487 reserve--; 2488 continue; 2489 } 2490 2491 /* Suitable for reserving if this block is movable */ 2492 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) { 2493 set_pageblock_migratetype(page, MIGRATE_RESERVE); 2494 move_freepages_block(zone, page, MIGRATE_RESERVE); 2495 reserve--; 2496 continue; 2497 } 2498 2499 /* 2500 * If the reserve is met and this is a previous reserved block, 2501 * take it back 2502 */ 2503 if (block_migratetype == MIGRATE_RESERVE) { 2504 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2505 move_freepages_block(zone, page, MIGRATE_MOVABLE); 2506 } 2507 } 2508 } 2509 2510 /* 2511 * Initially all pages are reserved - free ones are freed 2512 * up by free_all_bootmem() once the early boot process is 2513 * done. Non-atomic initialization, single-pass. 2514 */ 2515 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 2516 unsigned long start_pfn, enum memmap_context context) 2517 { 2518 struct page *page; 2519 unsigned long end_pfn = start_pfn + size; 2520 unsigned long pfn; 2521 2522 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 2523 /* 2524 * There can be holes in boot-time mem_map[]s 2525 * handed to this function. They do not 2526 * exist on hotplugged memory. 2527 */ 2528 if (context == MEMMAP_EARLY) { 2529 if (!early_pfn_valid(pfn)) 2530 continue; 2531 if (!early_pfn_in_nid(pfn, nid)) 2532 continue; 2533 } 2534 page = pfn_to_page(pfn); 2535 set_page_links(page, zone, nid, pfn); 2536 init_page_count(page); 2537 reset_page_mapcount(page); 2538 SetPageReserved(page); 2539 2540 /* 2541 * Mark the block movable so that blocks are reserved for 2542 * movable at startup. This will force kernel allocations 2543 * to reserve their blocks rather than leaking throughout 2544 * the address space during boot when many long-lived 2545 * kernel allocations are made. Later some blocks near 2546 * the start are marked MIGRATE_RESERVE by 2547 * setup_zone_migrate_reserve() 2548 */ 2549 if ((pfn & (pageblock_nr_pages-1))) 2550 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 2551 2552 INIT_LIST_HEAD(&page->lru); 2553 #ifdef WANT_PAGE_VIRTUAL 2554 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 2555 if (!is_highmem_idx(zone)) 2556 set_page_address(page, __va(pfn << PAGE_SHIFT)); 2557 #endif 2558 } 2559 } 2560 2561 static void __meminit zone_init_free_lists(struct zone *zone) 2562 { 2563 int order, t; 2564 for_each_migratetype_order(order, t) { 2565 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 2566 zone->free_area[order].nr_free = 0; 2567 } 2568 } 2569 2570 #ifndef __HAVE_ARCH_MEMMAP_INIT 2571 #define memmap_init(size, nid, zone, start_pfn) \ 2572 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 2573 #endif 2574 2575 static int zone_batchsize(struct zone *zone) 2576 { 2577 int batch; 2578 2579 /* 2580 * The per-cpu-pages pools are set to around 1000th of the 2581 * size of the zone. But no more than 1/2 of a meg. 2582 * 2583 * OK, so we don't know how big the cache is. So guess. 2584 */ 2585 batch = zone->present_pages / 1024; 2586 if (batch * PAGE_SIZE > 512 * 1024) 2587 batch = (512 * 1024) / PAGE_SIZE; 2588 batch /= 4; /* We effectively *= 4 below */ 2589 if (batch < 1) 2590 batch = 1; 2591 2592 /* 2593 * Clamp the batch to a 2^n - 1 value. Having a power 2594 * of 2 value was found to be more likely to have 2595 * suboptimal cache aliasing properties in some cases. 2596 * 2597 * For example if 2 tasks are alternately allocating 2598 * batches of pages, one task can end up with a lot 2599 * of pages of one half of the possible page colors 2600 * and the other with pages of the other colors. 2601 */ 2602 batch = (1 << (fls(batch + batch/2)-1)) - 1; 2603 2604 return batch; 2605 } 2606 2607 inline void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 2608 { 2609 struct per_cpu_pages *pcp; 2610 2611 memset(p, 0, sizeof(*p)); 2612 2613 pcp = &p->pcp; 2614 pcp->count = 0; 2615 pcp->high = 6 * batch; 2616 pcp->batch = max(1UL, 1 * batch); 2617 INIT_LIST_HEAD(&pcp->list); 2618 } 2619 2620 /* 2621 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 2622 * to the value high for the pageset p. 2623 */ 2624 2625 static void setup_pagelist_highmark(struct per_cpu_pageset *p, 2626 unsigned long high) 2627 { 2628 struct per_cpu_pages *pcp; 2629 2630 pcp = &p->pcp; 2631 pcp->high = high; 2632 pcp->batch = max(1UL, high/4); 2633 if ((high/4) > (PAGE_SHIFT * 8)) 2634 pcp->batch = PAGE_SHIFT * 8; 2635 } 2636 2637 2638 #ifdef CONFIG_NUMA 2639 /* 2640 * Boot pageset table. One per cpu which is going to be used for all 2641 * zones and all nodes. The parameters will be set in such a way 2642 * that an item put on a list will immediately be handed over to 2643 * the buddy list. This is safe since pageset manipulation is done 2644 * with interrupts disabled. 2645 * 2646 * Some NUMA counter updates may also be caught by the boot pagesets. 2647 * 2648 * The boot_pagesets must be kept even after bootup is complete for 2649 * unused processors and/or zones. They do play a role for bootstrapping 2650 * hotplugged processors. 2651 * 2652 * zoneinfo_show() and maybe other functions do 2653 * not check if the processor is online before following the pageset pointer. 2654 * Other parts of the kernel may not check if the zone is available. 2655 */ 2656 static struct per_cpu_pageset boot_pageset[NR_CPUS]; 2657 2658 /* 2659 * Dynamically allocate memory for the 2660 * per cpu pageset array in struct zone. 2661 */ 2662 static int __cpuinit process_zones(int cpu) 2663 { 2664 struct zone *zone, *dzone; 2665 int node = cpu_to_node(cpu); 2666 2667 node_set_state(node, N_CPU); /* this node has a cpu */ 2668 2669 for_each_zone(zone) { 2670 2671 if (!populated_zone(zone)) 2672 continue; 2673 2674 zone_pcp(zone, cpu) = kmalloc_node(sizeof(struct per_cpu_pageset), 2675 GFP_KERNEL, node); 2676 if (!zone_pcp(zone, cpu)) 2677 goto bad; 2678 2679 setup_pageset(zone_pcp(zone, cpu), zone_batchsize(zone)); 2680 2681 if (percpu_pagelist_fraction) 2682 setup_pagelist_highmark(zone_pcp(zone, cpu), 2683 (zone->present_pages / percpu_pagelist_fraction)); 2684 } 2685 2686 return 0; 2687 bad: 2688 for_each_zone(dzone) { 2689 if (!populated_zone(dzone)) 2690 continue; 2691 if (dzone == zone) 2692 break; 2693 kfree(zone_pcp(dzone, cpu)); 2694 zone_pcp(dzone, cpu) = NULL; 2695 } 2696 return -ENOMEM; 2697 } 2698 2699 static inline void free_zone_pagesets(int cpu) 2700 { 2701 struct zone *zone; 2702 2703 for_each_zone(zone) { 2704 struct per_cpu_pageset *pset = zone_pcp(zone, cpu); 2705 2706 /* Free per_cpu_pageset if it is slab allocated */ 2707 if (pset != &boot_pageset[cpu]) 2708 kfree(pset); 2709 zone_pcp(zone, cpu) = NULL; 2710 } 2711 } 2712 2713 static int __cpuinit pageset_cpuup_callback(struct notifier_block *nfb, 2714 unsigned long action, 2715 void *hcpu) 2716 { 2717 int cpu = (long)hcpu; 2718 int ret = NOTIFY_OK; 2719 2720 switch (action) { 2721 case CPU_UP_PREPARE: 2722 case CPU_UP_PREPARE_FROZEN: 2723 if (process_zones(cpu)) 2724 ret = NOTIFY_BAD; 2725 break; 2726 case CPU_UP_CANCELED: 2727 case CPU_UP_CANCELED_FROZEN: 2728 case CPU_DEAD: 2729 case CPU_DEAD_FROZEN: 2730 free_zone_pagesets(cpu); 2731 break; 2732 default: 2733 break; 2734 } 2735 return ret; 2736 } 2737 2738 static struct notifier_block __cpuinitdata pageset_notifier = 2739 { &pageset_cpuup_callback, NULL, 0 }; 2740 2741 void __init setup_per_cpu_pageset(void) 2742 { 2743 int err; 2744 2745 /* Initialize per_cpu_pageset for cpu 0. 2746 * A cpuup callback will do this for every cpu 2747 * as it comes online 2748 */ 2749 err = process_zones(smp_processor_id()); 2750 BUG_ON(err); 2751 register_cpu_notifier(&pageset_notifier); 2752 } 2753 2754 #endif 2755 2756 static noinline __init_refok 2757 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 2758 { 2759 int i; 2760 struct pglist_data *pgdat = zone->zone_pgdat; 2761 size_t alloc_size; 2762 2763 /* 2764 * The per-page waitqueue mechanism uses hashed waitqueues 2765 * per zone. 2766 */ 2767 zone->wait_table_hash_nr_entries = 2768 wait_table_hash_nr_entries(zone_size_pages); 2769 zone->wait_table_bits = 2770 wait_table_bits(zone->wait_table_hash_nr_entries); 2771 alloc_size = zone->wait_table_hash_nr_entries 2772 * sizeof(wait_queue_head_t); 2773 2774 if (system_state == SYSTEM_BOOTING) { 2775 zone->wait_table = (wait_queue_head_t *) 2776 alloc_bootmem_node(pgdat, alloc_size); 2777 } else { 2778 /* 2779 * This case means that a zone whose size was 0 gets new memory 2780 * via memory hot-add. 2781 * But it may be the case that a new node was hot-added. In 2782 * this case vmalloc() will not be able to use this new node's 2783 * memory - this wait_table must be initialized to use this new 2784 * node itself as well. 2785 * To use this new node's memory, further consideration will be 2786 * necessary. 2787 */ 2788 zone->wait_table = vmalloc(alloc_size); 2789 } 2790 if (!zone->wait_table) 2791 return -ENOMEM; 2792 2793 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 2794 init_waitqueue_head(zone->wait_table + i); 2795 2796 return 0; 2797 } 2798 2799 static __meminit void zone_pcp_init(struct zone *zone) 2800 { 2801 int cpu; 2802 unsigned long batch = zone_batchsize(zone); 2803 2804 for (cpu = 0; cpu < NR_CPUS; cpu++) { 2805 #ifdef CONFIG_NUMA 2806 /* Early boot. Slab allocator not functional yet */ 2807 zone_pcp(zone, cpu) = &boot_pageset[cpu]; 2808 setup_pageset(&boot_pageset[cpu],0); 2809 #else 2810 setup_pageset(zone_pcp(zone,cpu), batch); 2811 #endif 2812 } 2813 if (zone->present_pages) 2814 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%lu\n", 2815 zone->name, zone->present_pages, batch); 2816 } 2817 2818 __meminit int init_currently_empty_zone(struct zone *zone, 2819 unsigned long zone_start_pfn, 2820 unsigned long size, 2821 enum memmap_context context) 2822 { 2823 struct pglist_data *pgdat = zone->zone_pgdat; 2824 int ret; 2825 ret = zone_wait_table_init(zone, size); 2826 if (ret) 2827 return ret; 2828 pgdat->nr_zones = zone_idx(zone) + 1; 2829 2830 zone->zone_start_pfn = zone_start_pfn; 2831 2832 memmap_init(size, pgdat->node_id, zone_idx(zone), zone_start_pfn); 2833 2834 zone_init_free_lists(zone); 2835 2836 return 0; 2837 } 2838 2839 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP 2840 /* 2841 * Basic iterator support. Return the first range of PFNs for a node 2842 * Note: nid == MAX_NUMNODES returns first region regardless of node 2843 */ 2844 static int __meminit first_active_region_index_in_nid(int nid) 2845 { 2846 int i; 2847 2848 for (i = 0; i < nr_nodemap_entries; i++) 2849 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid) 2850 return i; 2851 2852 return -1; 2853 } 2854 2855 /* 2856 * Basic iterator support. Return the next active range of PFNs for a node 2857 * Note: nid == MAX_NUMNODES returns next region regardless of node 2858 */ 2859 static int __meminit next_active_region_index_in_nid(int index, int nid) 2860 { 2861 for (index = index + 1; index < nr_nodemap_entries; index++) 2862 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid) 2863 return index; 2864 2865 return -1; 2866 } 2867 2868 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 2869 /* 2870 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 2871 * Architectures may implement their own version but if add_active_range() 2872 * was used and there are no special requirements, this is a convenient 2873 * alternative 2874 */ 2875 int __meminit early_pfn_to_nid(unsigned long pfn) 2876 { 2877 int i; 2878 2879 for (i = 0; i < nr_nodemap_entries; i++) { 2880 unsigned long start_pfn = early_node_map[i].start_pfn; 2881 unsigned long end_pfn = early_node_map[i].end_pfn; 2882 2883 if (start_pfn <= pfn && pfn < end_pfn) 2884 return early_node_map[i].nid; 2885 } 2886 2887 return 0; 2888 } 2889 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 2890 2891 /* Basic iterator support to walk early_node_map[] */ 2892 #define for_each_active_range_index_in_nid(i, nid) \ 2893 for (i = first_active_region_index_in_nid(nid); i != -1; \ 2894 i = next_active_region_index_in_nid(i, nid)) 2895 2896 /** 2897 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 2898 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 2899 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 2900 * 2901 * If an architecture guarantees that all ranges registered with 2902 * add_active_ranges() contain no holes and may be freed, this 2903 * this function may be used instead of calling free_bootmem() manually. 2904 */ 2905 void __init free_bootmem_with_active_regions(int nid, 2906 unsigned long max_low_pfn) 2907 { 2908 int i; 2909 2910 for_each_active_range_index_in_nid(i, nid) { 2911 unsigned long size_pages = 0; 2912 unsigned long end_pfn = early_node_map[i].end_pfn; 2913 2914 if (early_node_map[i].start_pfn >= max_low_pfn) 2915 continue; 2916 2917 if (end_pfn > max_low_pfn) 2918 end_pfn = max_low_pfn; 2919 2920 size_pages = end_pfn - early_node_map[i].start_pfn; 2921 free_bootmem_node(NODE_DATA(early_node_map[i].nid), 2922 PFN_PHYS(early_node_map[i].start_pfn), 2923 size_pages << PAGE_SHIFT); 2924 } 2925 } 2926 2927 /** 2928 * sparse_memory_present_with_active_regions - Call memory_present for each active range 2929 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 2930 * 2931 * If an architecture guarantees that all ranges registered with 2932 * add_active_ranges() contain no holes and may be freed, this 2933 * function may be used instead of calling memory_present() manually. 2934 */ 2935 void __init sparse_memory_present_with_active_regions(int nid) 2936 { 2937 int i; 2938 2939 for_each_active_range_index_in_nid(i, nid) 2940 memory_present(early_node_map[i].nid, 2941 early_node_map[i].start_pfn, 2942 early_node_map[i].end_pfn); 2943 } 2944 2945 /** 2946 * push_node_boundaries - Push node boundaries to at least the requested boundary 2947 * @nid: The nid of the node to push the boundary for 2948 * @start_pfn: The start pfn of the node 2949 * @end_pfn: The end pfn of the node 2950 * 2951 * In reserve-based hot-add, mem_map is allocated that is unused until hotadd 2952 * time. Specifically, on x86_64, SRAT will report ranges that can potentially 2953 * be hotplugged even though no physical memory exists. This function allows 2954 * an arch to push out the node boundaries so mem_map is allocated that can 2955 * be used later. 2956 */ 2957 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE 2958 void __init push_node_boundaries(unsigned int nid, 2959 unsigned long start_pfn, unsigned long end_pfn) 2960 { 2961 printk(KERN_DEBUG "Entering push_node_boundaries(%u, %lu, %lu)\n", 2962 nid, start_pfn, end_pfn); 2963 2964 /* Initialise the boundary for this node if necessary */ 2965 if (node_boundary_end_pfn[nid] == 0) 2966 node_boundary_start_pfn[nid] = -1UL; 2967 2968 /* Update the boundaries */ 2969 if (node_boundary_start_pfn[nid] > start_pfn) 2970 node_boundary_start_pfn[nid] = start_pfn; 2971 if (node_boundary_end_pfn[nid] < end_pfn) 2972 node_boundary_end_pfn[nid] = end_pfn; 2973 } 2974 2975 /* If necessary, push the node boundary out for reserve hotadd */ 2976 static void __meminit account_node_boundary(unsigned int nid, 2977 unsigned long *start_pfn, unsigned long *end_pfn) 2978 { 2979 printk(KERN_DEBUG "Entering account_node_boundary(%u, %lu, %lu)\n", 2980 nid, *start_pfn, *end_pfn); 2981 2982 /* Return if boundary information has not been provided */ 2983 if (node_boundary_end_pfn[nid] == 0) 2984 return; 2985 2986 /* Check the boundaries and update if necessary */ 2987 if (node_boundary_start_pfn[nid] < *start_pfn) 2988 *start_pfn = node_boundary_start_pfn[nid]; 2989 if (node_boundary_end_pfn[nid] > *end_pfn) 2990 *end_pfn = node_boundary_end_pfn[nid]; 2991 } 2992 #else 2993 void __init push_node_boundaries(unsigned int nid, 2994 unsigned long start_pfn, unsigned long end_pfn) {} 2995 2996 static void __meminit account_node_boundary(unsigned int nid, 2997 unsigned long *start_pfn, unsigned long *end_pfn) {} 2998 #endif 2999 3000 3001 /** 3002 * get_pfn_range_for_nid - Return the start and end page frames for a node 3003 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3004 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3005 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3006 * 3007 * It returns the start and end page frame of a node based on information 3008 * provided by an arch calling add_active_range(). If called for a node 3009 * with no available memory, a warning is printed and the start and end 3010 * PFNs will be 0. 3011 */ 3012 void __meminit get_pfn_range_for_nid(unsigned int nid, 3013 unsigned long *start_pfn, unsigned long *end_pfn) 3014 { 3015 int i; 3016 *start_pfn = -1UL; 3017 *end_pfn = 0; 3018 3019 for_each_active_range_index_in_nid(i, nid) { 3020 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn); 3021 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn); 3022 } 3023 3024 if (*start_pfn == -1UL) 3025 *start_pfn = 0; 3026 3027 /* Push the node boundaries out if requested */ 3028 account_node_boundary(nid, start_pfn, end_pfn); 3029 } 3030 3031 /* 3032 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3033 * assumption is made that zones within a node are ordered in monotonic 3034 * increasing memory addresses so that the "highest" populated zone is used 3035 */ 3036 void __init find_usable_zone_for_movable(void) 3037 { 3038 int zone_index; 3039 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3040 if (zone_index == ZONE_MOVABLE) 3041 continue; 3042 3043 if (arch_zone_highest_possible_pfn[zone_index] > 3044 arch_zone_lowest_possible_pfn[zone_index]) 3045 break; 3046 } 3047 3048 VM_BUG_ON(zone_index == -1); 3049 movable_zone = zone_index; 3050 } 3051 3052 /* 3053 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 3054 * because it is sized independant of architecture. Unlike the other zones, 3055 * the starting point for ZONE_MOVABLE is not fixed. It may be different 3056 * in each node depending on the size of each node and how evenly kernelcore 3057 * is distributed. This helper function adjusts the zone ranges 3058 * provided by the architecture for a given node by using the end of the 3059 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 3060 * zones within a node are in order of monotonic increases memory addresses 3061 */ 3062 void __meminit adjust_zone_range_for_zone_movable(int nid, 3063 unsigned long zone_type, 3064 unsigned long node_start_pfn, 3065 unsigned long node_end_pfn, 3066 unsigned long *zone_start_pfn, 3067 unsigned long *zone_end_pfn) 3068 { 3069 /* Only adjust if ZONE_MOVABLE is on this node */ 3070 if (zone_movable_pfn[nid]) { 3071 /* Size ZONE_MOVABLE */ 3072 if (zone_type == ZONE_MOVABLE) { 3073 *zone_start_pfn = zone_movable_pfn[nid]; 3074 *zone_end_pfn = min(node_end_pfn, 3075 arch_zone_highest_possible_pfn[movable_zone]); 3076 3077 /* Adjust for ZONE_MOVABLE starting within this range */ 3078 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 3079 *zone_end_pfn > zone_movable_pfn[nid]) { 3080 *zone_end_pfn = zone_movable_pfn[nid]; 3081 3082 /* Check if this whole range is within ZONE_MOVABLE */ 3083 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 3084 *zone_start_pfn = *zone_end_pfn; 3085 } 3086 } 3087 3088 /* 3089 * Return the number of pages a zone spans in a node, including holes 3090 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 3091 */ 3092 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 3093 unsigned long zone_type, 3094 unsigned long *ignored) 3095 { 3096 unsigned long node_start_pfn, node_end_pfn; 3097 unsigned long zone_start_pfn, zone_end_pfn; 3098 3099 /* Get the start and end of the node and zone */ 3100 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3101 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 3102 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 3103 adjust_zone_range_for_zone_movable(nid, zone_type, 3104 node_start_pfn, node_end_pfn, 3105 &zone_start_pfn, &zone_end_pfn); 3106 3107 /* Check that this node has pages within the zone's required range */ 3108 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 3109 return 0; 3110 3111 /* Move the zone boundaries inside the node if necessary */ 3112 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 3113 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 3114 3115 /* Return the spanned pages */ 3116 return zone_end_pfn - zone_start_pfn; 3117 } 3118 3119 /* 3120 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 3121 * then all holes in the requested range will be accounted for. 3122 */ 3123 unsigned long __meminit __absent_pages_in_range(int nid, 3124 unsigned long range_start_pfn, 3125 unsigned long range_end_pfn) 3126 { 3127 int i = 0; 3128 unsigned long prev_end_pfn = 0, hole_pages = 0; 3129 unsigned long start_pfn; 3130 3131 /* Find the end_pfn of the first active range of pfns in the node */ 3132 i = first_active_region_index_in_nid(nid); 3133 if (i == -1) 3134 return 0; 3135 3136 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3137 3138 /* Account for ranges before physical memory on this node */ 3139 if (early_node_map[i].start_pfn > range_start_pfn) 3140 hole_pages = prev_end_pfn - range_start_pfn; 3141 3142 /* Find all holes for the zone within the node */ 3143 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) { 3144 3145 /* No need to continue if prev_end_pfn is outside the zone */ 3146 if (prev_end_pfn >= range_end_pfn) 3147 break; 3148 3149 /* Make sure the end of the zone is not within the hole */ 3150 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn); 3151 prev_end_pfn = max(prev_end_pfn, range_start_pfn); 3152 3153 /* Update the hole size cound and move on */ 3154 if (start_pfn > range_start_pfn) { 3155 BUG_ON(prev_end_pfn > start_pfn); 3156 hole_pages += start_pfn - prev_end_pfn; 3157 } 3158 prev_end_pfn = early_node_map[i].end_pfn; 3159 } 3160 3161 /* Account for ranges past physical memory on this node */ 3162 if (range_end_pfn > prev_end_pfn) 3163 hole_pages += range_end_pfn - 3164 max(range_start_pfn, prev_end_pfn); 3165 3166 return hole_pages; 3167 } 3168 3169 /** 3170 * absent_pages_in_range - Return number of page frames in holes within a range 3171 * @start_pfn: The start PFN to start searching for holes 3172 * @end_pfn: The end PFN to stop searching for holes 3173 * 3174 * It returns the number of pages frames in memory holes within a range. 3175 */ 3176 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 3177 unsigned long end_pfn) 3178 { 3179 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 3180 } 3181 3182 /* Return the number of page frames in holes in a zone on a node */ 3183 static unsigned long __meminit zone_absent_pages_in_node(int nid, 3184 unsigned long zone_type, 3185 unsigned long *ignored) 3186 { 3187 unsigned long node_start_pfn, node_end_pfn; 3188 unsigned long zone_start_pfn, zone_end_pfn; 3189 3190 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 3191 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type], 3192 node_start_pfn); 3193 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type], 3194 node_end_pfn); 3195 3196 adjust_zone_range_for_zone_movable(nid, zone_type, 3197 node_start_pfn, node_end_pfn, 3198 &zone_start_pfn, &zone_end_pfn); 3199 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 3200 } 3201 3202 #else 3203 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 3204 unsigned long zone_type, 3205 unsigned long *zones_size) 3206 { 3207 return zones_size[zone_type]; 3208 } 3209 3210 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 3211 unsigned long zone_type, 3212 unsigned long *zholes_size) 3213 { 3214 if (!zholes_size) 3215 return 0; 3216 3217 return zholes_size[zone_type]; 3218 } 3219 3220 #endif 3221 3222 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 3223 unsigned long *zones_size, unsigned long *zholes_size) 3224 { 3225 unsigned long realtotalpages, totalpages = 0; 3226 enum zone_type i; 3227 3228 for (i = 0; i < MAX_NR_ZONES; i++) 3229 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 3230 zones_size); 3231 pgdat->node_spanned_pages = totalpages; 3232 3233 realtotalpages = totalpages; 3234 for (i = 0; i < MAX_NR_ZONES; i++) 3235 realtotalpages -= 3236 zone_absent_pages_in_node(pgdat->node_id, i, 3237 zholes_size); 3238 pgdat->node_present_pages = realtotalpages; 3239 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 3240 realtotalpages); 3241 } 3242 3243 #ifndef CONFIG_SPARSEMEM 3244 /* 3245 * Calculate the size of the zone->blockflags rounded to an unsigned long 3246 * Start by making sure zonesize is a multiple of pageblock_order by rounding 3247 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 3248 * round what is now in bits to nearest long in bits, then return it in 3249 * bytes. 3250 */ 3251 static unsigned long __init usemap_size(unsigned long zonesize) 3252 { 3253 unsigned long usemapsize; 3254 3255 usemapsize = roundup(zonesize, pageblock_nr_pages); 3256 usemapsize = usemapsize >> pageblock_order; 3257 usemapsize *= NR_PAGEBLOCK_BITS; 3258 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 3259 3260 return usemapsize / 8; 3261 } 3262 3263 static void __init setup_usemap(struct pglist_data *pgdat, 3264 struct zone *zone, unsigned long zonesize) 3265 { 3266 unsigned long usemapsize = usemap_size(zonesize); 3267 zone->pageblock_flags = NULL; 3268 if (usemapsize) { 3269 zone->pageblock_flags = alloc_bootmem_node(pgdat, usemapsize); 3270 memset(zone->pageblock_flags, 0, usemapsize); 3271 } 3272 } 3273 #else 3274 static void inline setup_usemap(struct pglist_data *pgdat, 3275 struct zone *zone, unsigned long zonesize) {} 3276 #endif /* CONFIG_SPARSEMEM */ 3277 3278 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 3279 3280 /* Return a sensible default order for the pageblock size. */ 3281 static inline int pageblock_default_order(void) 3282 { 3283 if (HPAGE_SHIFT > PAGE_SHIFT) 3284 return HUGETLB_PAGE_ORDER; 3285 3286 return MAX_ORDER-1; 3287 } 3288 3289 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 3290 static inline void __init set_pageblock_order(unsigned int order) 3291 { 3292 /* Check that pageblock_nr_pages has not already been setup */ 3293 if (pageblock_order) 3294 return; 3295 3296 /* 3297 * Assume the largest contiguous order of interest is a huge page. 3298 * This value may be variable depending on boot parameters on IA64 3299 */ 3300 pageblock_order = order; 3301 } 3302 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3303 3304 /* 3305 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 3306 * and pageblock_default_order() are unused as pageblock_order is set 3307 * at compile-time. See include/linux/pageblock-flags.h for the values of 3308 * pageblock_order based on the kernel config 3309 */ 3310 static inline int pageblock_default_order(unsigned int order) 3311 { 3312 return MAX_ORDER-1; 3313 } 3314 #define set_pageblock_order(x) do {} while (0) 3315 3316 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 3317 3318 /* 3319 * Set up the zone data structures: 3320 * - mark all pages reserved 3321 * - mark all memory queues empty 3322 * - clear the memory bitmaps 3323 */ 3324 static void __paginginit free_area_init_core(struct pglist_data *pgdat, 3325 unsigned long *zones_size, unsigned long *zholes_size) 3326 { 3327 enum zone_type j; 3328 int nid = pgdat->node_id; 3329 unsigned long zone_start_pfn = pgdat->node_start_pfn; 3330 int ret; 3331 3332 pgdat_resize_init(pgdat); 3333 pgdat->nr_zones = 0; 3334 init_waitqueue_head(&pgdat->kswapd_wait); 3335 pgdat->kswapd_max_order = 0; 3336 3337 for (j = 0; j < MAX_NR_ZONES; j++) { 3338 struct zone *zone = pgdat->node_zones + j; 3339 unsigned long size, realsize, memmap_pages; 3340 3341 size = zone_spanned_pages_in_node(nid, j, zones_size); 3342 realsize = size - zone_absent_pages_in_node(nid, j, 3343 zholes_size); 3344 3345 /* 3346 * Adjust realsize so that it accounts for how much memory 3347 * is used by this zone for memmap. This affects the watermark 3348 * and per-cpu initialisations 3349 */ 3350 memmap_pages = (size * sizeof(struct page)) >> PAGE_SHIFT; 3351 if (realsize >= memmap_pages) { 3352 realsize -= memmap_pages; 3353 printk(KERN_DEBUG 3354 " %s zone: %lu pages used for memmap\n", 3355 zone_names[j], memmap_pages); 3356 } else 3357 printk(KERN_WARNING 3358 " %s zone: %lu pages exceeds realsize %lu\n", 3359 zone_names[j], memmap_pages, realsize); 3360 3361 /* Account for reserved pages */ 3362 if (j == 0 && realsize > dma_reserve) { 3363 realsize -= dma_reserve; 3364 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 3365 zone_names[0], dma_reserve); 3366 } 3367 3368 if (!is_highmem_idx(j)) 3369 nr_kernel_pages += realsize; 3370 nr_all_pages += realsize; 3371 3372 zone->spanned_pages = size; 3373 zone->present_pages = realsize; 3374 #ifdef CONFIG_NUMA 3375 zone->node = nid; 3376 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 3377 / 100; 3378 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 3379 #endif 3380 zone->name = zone_names[j]; 3381 spin_lock_init(&zone->lock); 3382 spin_lock_init(&zone->lru_lock); 3383 zone_seqlock_init(zone); 3384 zone->zone_pgdat = pgdat; 3385 3386 zone->prev_priority = DEF_PRIORITY; 3387 3388 zone_pcp_init(zone); 3389 INIT_LIST_HEAD(&zone->active_list); 3390 INIT_LIST_HEAD(&zone->inactive_list); 3391 zone->nr_scan_active = 0; 3392 zone->nr_scan_inactive = 0; 3393 zap_zone_vm_stats(zone); 3394 zone->flags = 0; 3395 if (!size) 3396 continue; 3397 3398 set_pageblock_order(pageblock_default_order()); 3399 setup_usemap(pgdat, zone, size); 3400 ret = init_currently_empty_zone(zone, zone_start_pfn, 3401 size, MEMMAP_EARLY); 3402 BUG_ON(ret); 3403 zone_start_pfn += size; 3404 } 3405 } 3406 3407 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 3408 { 3409 /* Skip empty nodes */ 3410 if (!pgdat->node_spanned_pages) 3411 return; 3412 3413 #ifdef CONFIG_FLAT_NODE_MEM_MAP 3414 /* ia64 gets its own node_mem_map, before this, without bootmem */ 3415 if (!pgdat->node_mem_map) { 3416 unsigned long size, start, end; 3417 struct page *map; 3418 3419 /* 3420 * The zone's endpoints aren't required to be MAX_ORDER 3421 * aligned but the node_mem_map endpoints must be in order 3422 * for the buddy allocator to function correctly. 3423 */ 3424 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 3425 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 3426 end = ALIGN(end, MAX_ORDER_NR_PAGES); 3427 size = (end - start) * sizeof(struct page); 3428 map = alloc_remap(pgdat->node_id, size); 3429 if (!map) 3430 map = alloc_bootmem_node(pgdat, size); 3431 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 3432 } 3433 #ifndef CONFIG_NEED_MULTIPLE_NODES 3434 /* 3435 * With no DISCONTIG, the global mem_map is just set as node 0's 3436 */ 3437 if (pgdat == NODE_DATA(0)) { 3438 mem_map = NODE_DATA(0)->node_mem_map; 3439 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3440 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 3441 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 3442 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3443 } 3444 #endif 3445 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 3446 } 3447 3448 void __paginginit free_area_init_node(int nid, struct pglist_data *pgdat, 3449 unsigned long *zones_size, unsigned long node_start_pfn, 3450 unsigned long *zholes_size) 3451 { 3452 pgdat->node_id = nid; 3453 pgdat->node_start_pfn = node_start_pfn; 3454 calculate_node_totalpages(pgdat, zones_size, zholes_size); 3455 3456 alloc_node_mem_map(pgdat); 3457 3458 free_area_init_core(pgdat, zones_size, zholes_size); 3459 } 3460 3461 #ifdef CONFIG_ARCH_POPULATES_NODE_MAP 3462 3463 #if MAX_NUMNODES > 1 3464 /* 3465 * Figure out the number of possible node ids. 3466 */ 3467 static void __init setup_nr_node_ids(void) 3468 { 3469 unsigned int node; 3470 unsigned int highest = 0; 3471 3472 for_each_node_mask(node, node_possible_map) 3473 highest = node; 3474 nr_node_ids = highest + 1; 3475 } 3476 #else 3477 static inline void setup_nr_node_ids(void) 3478 { 3479 } 3480 #endif 3481 3482 /** 3483 * add_active_range - Register a range of PFNs backed by physical memory 3484 * @nid: The node ID the range resides on 3485 * @start_pfn: The start PFN of the available physical memory 3486 * @end_pfn: The end PFN of the available physical memory 3487 * 3488 * These ranges are stored in an early_node_map[] and later used by 3489 * free_area_init_nodes() to calculate zone sizes and holes. If the 3490 * range spans a memory hole, it is up to the architecture to ensure 3491 * the memory is not freed by the bootmem allocator. If possible 3492 * the range being registered will be merged with existing ranges. 3493 */ 3494 void __init add_active_range(unsigned int nid, unsigned long start_pfn, 3495 unsigned long end_pfn) 3496 { 3497 int i; 3498 3499 printk(KERN_DEBUG "Entering add_active_range(%d, %lu, %lu) " 3500 "%d entries of %d used\n", 3501 nid, start_pfn, end_pfn, 3502 nr_nodemap_entries, MAX_ACTIVE_REGIONS); 3503 3504 /* Merge with existing active regions if possible */ 3505 for (i = 0; i < nr_nodemap_entries; i++) { 3506 if (early_node_map[i].nid != nid) 3507 continue; 3508 3509 /* Skip if an existing region covers this new one */ 3510 if (start_pfn >= early_node_map[i].start_pfn && 3511 end_pfn <= early_node_map[i].end_pfn) 3512 return; 3513 3514 /* Merge forward if suitable */ 3515 if (start_pfn <= early_node_map[i].end_pfn && 3516 end_pfn > early_node_map[i].end_pfn) { 3517 early_node_map[i].end_pfn = end_pfn; 3518 return; 3519 } 3520 3521 /* Merge backward if suitable */ 3522 if (start_pfn < early_node_map[i].end_pfn && 3523 end_pfn >= early_node_map[i].start_pfn) { 3524 early_node_map[i].start_pfn = start_pfn; 3525 return; 3526 } 3527 } 3528 3529 /* Check that early_node_map is large enough */ 3530 if (i >= MAX_ACTIVE_REGIONS) { 3531 printk(KERN_CRIT "More than %d memory regions, truncating\n", 3532 MAX_ACTIVE_REGIONS); 3533 return; 3534 } 3535 3536 early_node_map[i].nid = nid; 3537 early_node_map[i].start_pfn = start_pfn; 3538 early_node_map[i].end_pfn = end_pfn; 3539 nr_nodemap_entries = i + 1; 3540 } 3541 3542 /** 3543 * shrink_active_range - Shrink an existing registered range of PFNs 3544 * @nid: The node id the range is on that should be shrunk 3545 * @old_end_pfn: The old end PFN of the range 3546 * @new_end_pfn: The new PFN of the range 3547 * 3548 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node. 3549 * The map is kept at the end physical page range that has already been 3550 * registered with add_active_range(). This function allows an arch to shrink 3551 * an existing registered range. 3552 */ 3553 void __init shrink_active_range(unsigned int nid, unsigned long old_end_pfn, 3554 unsigned long new_end_pfn) 3555 { 3556 int i; 3557 3558 /* Find the old active region end and shrink */ 3559 for_each_active_range_index_in_nid(i, nid) 3560 if (early_node_map[i].end_pfn == old_end_pfn) { 3561 early_node_map[i].end_pfn = new_end_pfn; 3562 break; 3563 } 3564 } 3565 3566 /** 3567 * remove_all_active_ranges - Remove all currently registered regions 3568 * 3569 * During discovery, it may be found that a table like SRAT is invalid 3570 * and an alternative discovery method must be used. This function removes 3571 * all currently registered regions. 3572 */ 3573 void __init remove_all_active_ranges(void) 3574 { 3575 memset(early_node_map, 0, sizeof(early_node_map)); 3576 nr_nodemap_entries = 0; 3577 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE 3578 memset(node_boundary_start_pfn, 0, sizeof(node_boundary_start_pfn)); 3579 memset(node_boundary_end_pfn, 0, sizeof(node_boundary_end_pfn)); 3580 #endif /* CONFIG_MEMORY_HOTPLUG_RESERVE */ 3581 } 3582 3583 /* Compare two active node_active_regions */ 3584 static int __init cmp_node_active_region(const void *a, const void *b) 3585 { 3586 struct node_active_region *arange = (struct node_active_region *)a; 3587 struct node_active_region *brange = (struct node_active_region *)b; 3588 3589 /* Done this way to avoid overflows */ 3590 if (arange->start_pfn > brange->start_pfn) 3591 return 1; 3592 if (arange->start_pfn < brange->start_pfn) 3593 return -1; 3594 3595 return 0; 3596 } 3597 3598 /* sort the node_map by start_pfn */ 3599 static void __init sort_node_map(void) 3600 { 3601 sort(early_node_map, (size_t)nr_nodemap_entries, 3602 sizeof(struct node_active_region), 3603 cmp_node_active_region, NULL); 3604 } 3605 3606 /* Find the lowest pfn for a node */ 3607 unsigned long __init find_min_pfn_for_node(unsigned long nid) 3608 { 3609 int i; 3610 unsigned long min_pfn = ULONG_MAX; 3611 3612 /* Assuming a sorted map, the first range found has the starting pfn */ 3613 for_each_active_range_index_in_nid(i, nid) 3614 min_pfn = min(min_pfn, early_node_map[i].start_pfn); 3615 3616 if (min_pfn == ULONG_MAX) { 3617 printk(KERN_WARNING 3618 "Could not find start_pfn for node %lu\n", nid); 3619 return 0; 3620 } 3621 3622 return min_pfn; 3623 } 3624 3625 /** 3626 * find_min_pfn_with_active_regions - Find the minimum PFN registered 3627 * 3628 * It returns the minimum PFN based on information provided via 3629 * add_active_range(). 3630 */ 3631 unsigned long __init find_min_pfn_with_active_regions(void) 3632 { 3633 return find_min_pfn_for_node(MAX_NUMNODES); 3634 } 3635 3636 /** 3637 * find_max_pfn_with_active_regions - Find the maximum PFN registered 3638 * 3639 * It returns the maximum PFN based on information provided via 3640 * add_active_range(). 3641 */ 3642 unsigned long __init find_max_pfn_with_active_regions(void) 3643 { 3644 int i; 3645 unsigned long max_pfn = 0; 3646 3647 for (i = 0; i < nr_nodemap_entries; i++) 3648 max_pfn = max(max_pfn, early_node_map[i].end_pfn); 3649 3650 return max_pfn; 3651 } 3652 3653 /* 3654 * early_calculate_totalpages() 3655 * Sum pages in active regions for movable zone. 3656 * Populate N_HIGH_MEMORY for calculating usable_nodes. 3657 */ 3658 static unsigned long __init early_calculate_totalpages(void) 3659 { 3660 int i; 3661 unsigned long totalpages = 0; 3662 3663 for (i = 0; i < nr_nodemap_entries; i++) { 3664 unsigned long pages = early_node_map[i].end_pfn - 3665 early_node_map[i].start_pfn; 3666 totalpages += pages; 3667 if (pages) 3668 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY); 3669 } 3670 return totalpages; 3671 } 3672 3673 /* 3674 * Find the PFN the Movable zone begins in each node. Kernel memory 3675 * is spread evenly between nodes as long as the nodes have enough 3676 * memory. When they don't, some nodes will have more kernelcore than 3677 * others 3678 */ 3679 void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 3680 { 3681 int i, nid; 3682 unsigned long usable_startpfn; 3683 unsigned long kernelcore_node, kernelcore_remaining; 3684 unsigned long totalpages = early_calculate_totalpages(); 3685 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 3686 3687 /* 3688 * If movablecore was specified, calculate what size of 3689 * kernelcore that corresponds so that memory usable for 3690 * any allocation type is evenly spread. If both kernelcore 3691 * and movablecore are specified, then the value of kernelcore 3692 * will be used for required_kernelcore if it's greater than 3693 * what movablecore would have allowed. 3694 */ 3695 if (required_movablecore) { 3696 unsigned long corepages; 3697 3698 /* 3699 * Round-up so that ZONE_MOVABLE is at least as large as what 3700 * was requested by the user 3701 */ 3702 required_movablecore = 3703 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 3704 corepages = totalpages - required_movablecore; 3705 3706 required_kernelcore = max(required_kernelcore, corepages); 3707 } 3708 3709 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 3710 if (!required_kernelcore) 3711 return; 3712 3713 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 3714 find_usable_zone_for_movable(); 3715 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 3716 3717 restart: 3718 /* Spread kernelcore memory as evenly as possible throughout nodes */ 3719 kernelcore_node = required_kernelcore / usable_nodes; 3720 for_each_node_state(nid, N_HIGH_MEMORY) { 3721 /* 3722 * Recalculate kernelcore_node if the division per node 3723 * now exceeds what is necessary to satisfy the requested 3724 * amount of memory for the kernel 3725 */ 3726 if (required_kernelcore < kernelcore_node) 3727 kernelcore_node = required_kernelcore / usable_nodes; 3728 3729 /* 3730 * As the map is walked, we track how much memory is usable 3731 * by the kernel using kernelcore_remaining. When it is 3732 * 0, the rest of the node is usable by ZONE_MOVABLE 3733 */ 3734 kernelcore_remaining = kernelcore_node; 3735 3736 /* Go through each range of PFNs within this node */ 3737 for_each_active_range_index_in_nid(i, nid) { 3738 unsigned long start_pfn, end_pfn; 3739 unsigned long size_pages; 3740 3741 start_pfn = max(early_node_map[i].start_pfn, 3742 zone_movable_pfn[nid]); 3743 end_pfn = early_node_map[i].end_pfn; 3744 if (start_pfn >= end_pfn) 3745 continue; 3746 3747 /* Account for what is only usable for kernelcore */ 3748 if (start_pfn < usable_startpfn) { 3749 unsigned long kernel_pages; 3750 kernel_pages = min(end_pfn, usable_startpfn) 3751 - start_pfn; 3752 3753 kernelcore_remaining -= min(kernel_pages, 3754 kernelcore_remaining); 3755 required_kernelcore -= min(kernel_pages, 3756 required_kernelcore); 3757 3758 /* Continue if range is now fully accounted */ 3759 if (end_pfn <= usable_startpfn) { 3760 3761 /* 3762 * Push zone_movable_pfn to the end so 3763 * that if we have to rebalance 3764 * kernelcore across nodes, we will 3765 * not double account here 3766 */ 3767 zone_movable_pfn[nid] = end_pfn; 3768 continue; 3769 } 3770 start_pfn = usable_startpfn; 3771 } 3772 3773 /* 3774 * The usable PFN range for ZONE_MOVABLE is from 3775 * start_pfn->end_pfn. Calculate size_pages as the 3776 * number of pages used as kernelcore 3777 */ 3778 size_pages = end_pfn - start_pfn; 3779 if (size_pages > kernelcore_remaining) 3780 size_pages = kernelcore_remaining; 3781 zone_movable_pfn[nid] = start_pfn + size_pages; 3782 3783 /* 3784 * Some kernelcore has been met, update counts and 3785 * break if the kernelcore for this node has been 3786 * satisified 3787 */ 3788 required_kernelcore -= min(required_kernelcore, 3789 size_pages); 3790 kernelcore_remaining -= size_pages; 3791 if (!kernelcore_remaining) 3792 break; 3793 } 3794 } 3795 3796 /* 3797 * If there is still required_kernelcore, we do another pass with one 3798 * less node in the count. This will push zone_movable_pfn[nid] further 3799 * along on the nodes that still have memory until kernelcore is 3800 * satisified 3801 */ 3802 usable_nodes--; 3803 if (usable_nodes && required_kernelcore > usable_nodes) 3804 goto restart; 3805 3806 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 3807 for (nid = 0; nid < MAX_NUMNODES; nid++) 3808 zone_movable_pfn[nid] = 3809 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 3810 } 3811 3812 /* Any regular memory on that node ? */ 3813 static void check_for_regular_memory(pg_data_t *pgdat) 3814 { 3815 #ifdef CONFIG_HIGHMEM 3816 enum zone_type zone_type; 3817 3818 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 3819 struct zone *zone = &pgdat->node_zones[zone_type]; 3820 if (zone->present_pages) 3821 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 3822 } 3823 #endif 3824 } 3825 3826 /** 3827 * free_area_init_nodes - Initialise all pg_data_t and zone data 3828 * @max_zone_pfn: an array of max PFNs for each zone 3829 * 3830 * This will call free_area_init_node() for each active node in the system. 3831 * Using the page ranges provided by add_active_range(), the size of each 3832 * zone in each node and their holes is calculated. If the maximum PFN 3833 * between two adjacent zones match, it is assumed that the zone is empty. 3834 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 3835 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 3836 * starts where the previous one ended. For example, ZONE_DMA32 starts 3837 * at arch_max_dma_pfn. 3838 */ 3839 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 3840 { 3841 unsigned long nid; 3842 enum zone_type i; 3843 3844 /* Sort early_node_map as initialisation assumes it is sorted */ 3845 sort_node_map(); 3846 3847 /* Record where the zone boundaries are */ 3848 memset(arch_zone_lowest_possible_pfn, 0, 3849 sizeof(arch_zone_lowest_possible_pfn)); 3850 memset(arch_zone_highest_possible_pfn, 0, 3851 sizeof(arch_zone_highest_possible_pfn)); 3852 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 3853 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 3854 for (i = 1; i < MAX_NR_ZONES; i++) { 3855 if (i == ZONE_MOVABLE) 3856 continue; 3857 arch_zone_lowest_possible_pfn[i] = 3858 arch_zone_highest_possible_pfn[i-1]; 3859 arch_zone_highest_possible_pfn[i] = 3860 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 3861 } 3862 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 3863 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 3864 3865 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 3866 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 3867 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 3868 3869 /* Print out the zone ranges */ 3870 printk("Zone PFN ranges:\n"); 3871 for (i = 0; i < MAX_NR_ZONES; i++) { 3872 if (i == ZONE_MOVABLE) 3873 continue; 3874 printk(" %-8s %8lu -> %8lu\n", 3875 zone_names[i], 3876 arch_zone_lowest_possible_pfn[i], 3877 arch_zone_highest_possible_pfn[i]); 3878 } 3879 3880 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 3881 printk("Movable zone start PFN for each node\n"); 3882 for (i = 0; i < MAX_NUMNODES; i++) { 3883 if (zone_movable_pfn[i]) 3884 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 3885 } 3886 3887 /* Print out the early_node_map[] */ 3888 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries); 3889 for (i = 0; i < nr_nodemap_entries; i++) 3890 printk(" %3d: %8lu -> %8lu\n", early_node_map[i].nid, 3891 early_node_map[i].start_pfn, 3892 early_node_map[i].end_pfn); 3893 3894 /* Initialise every node */ 3895 setup_nr_node_ids(); 3896 for_each_online_node(nid) { 3897 pg_data_t *pgdat = NODE_DATA(nid); 3898 free_area_init_node(nid, pgdat, NULL, 3899 find_min_pfn_for_node(nid), NULL); 3900 3901 /* Any memory on that node */ 3902 if (pgdat->node_present_pages) 3903 node_set_state(nid, N_HIGH_MEMORY); 3904 check_for_regular_memory(pgdat); 3905 } 3906 } 3907 3908 static int __init cmdline_parse_core(char *p, unsigned long *core) 3909 { 3910 unsigned long long coremem; 3911 if (!p) 3912 return -EINVAL; 3913 3914 coremem = memparse(p, &p); 3915 *core = coremem >> PAGE_SHIFT; 3916 3917 /* Paranoid check that UL is enough for the coremem value */ 3918 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 3919 3920 return 0; 3921 } 3922 3923 /* 3924 * kernelcore=size sets the amount of memory for use for allocations that 3925 * cannot be reclaimed or migrated. 3926 */ 3927 static int __init cmdline_parse_kernelcore(char *p) 3928 { 3929 return cmdline_parse_core(p, &required_kernelcore); 3930 } 3931 3932 /* 3933 * movablecore=size sets the amount of memory for use for allocations that 3934 * can be reclaimed or migrated. 3935 */ 3936 static int __init cmdline_parse_movablecore(char *p) 3937 { 3938 return cmdline_parse_core(p, &required_movablecore); 3939 } 3940 3941 early_param("kernelcore", cmdline_parse_kernelcore); 3942 early_param("movablecore", cmdline_parse_movablecore); 3943 3944 #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */ 3945 3946 /** 3947 * set_dma_reserve - set the specified number of pages reserved in the first zone 3948 * @new_dma_reserve: The number of pages to mark reserved 3949 * 3950 * The per-cpu batchsize and zone watermarks are determined by present_pages. 3951 * In the DMA zone, a significant percentage may be consumed by kernel image 3952 * and other unfreeable allocations which can skew the watermarks badly. This 3953 * function may optionally be used to account for unfreeable pages in the 3954 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 3955 * smaller per-cpu batchsize. 3956 */ 3957 void __init set_dma_reserve(unsigned long new_dma_reserve) 3958 { 3959 dma_reserve = new_dma_reserve; 3960 } 3961 3962 #ifndef CONFIG_NEED_MULTIPLE_NODES 3963 static bootmem_data_t contig_bootmem_data; 3964 struct pglist_data contig_page_data = { .bdata = &contig_bootmem_data }; 3965 3966 EXPORT_SYMBOL(contig_page_data); 3967 #endif 3968 3969 void __init free_area_init(unsigned long *zones_size) 3970 { 3971 free_area_init_node(0, NODE_DATA(0), zones_size, 3972 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 3973 } 3974 3975 static int page_alloc_cpu_notify(struct notifier_block *self, 3976 unsigned long action, void *hcpu) 3977 { 3978 int cpu = (unsigned long)hcpu; 3979 3980 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 3981 drain_pages(cpu); 3982 3983 /* 3984 * Spill the event counters of the dead processor 3985 * into the current processors event counters. 3986 * This artificially elevates the count of the current 3987 * processor. 3988 */ 3989 vm_events_fold_cpu(cpu); 3990 3991 /* 3992 * Zero the differential counters of the dead processor 3993 * so that the vm statistics are consistent. 3994 * 3995 * This is only okay since the processor is dead and cannot 3996 * race with what we are doing. 3997 */ 3998 refresh_cpu_vm_stats(cpu); 3999 } 4000 return NOTIFY_OK; 4001 } 4002 4003 void __init page_alloc_init(void) 4004 { 4005 hotcpu_notifier(page_alloc_cpu_notify, 0); 4006 } 4007 4008 /* 4009 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4010 * or min_free_kbytes changes. 4011 */ 4012 static void calculate_totalreserve_pages(void) 4013 { 4014 struct pglist_data *pgdat; 4015 unsigned long reserve_pages = 0; 4016 enum zone_type i, j; 4017 4018 for_each_online_pgdat(pgdat) { 4019 for (i = 0; i < MAX_NR_ZONES; i++) { 4020 struct zone *zone = pgdat->node_zones + i; 4021 unsigned long max = 0; 4022 4023 /* Find valid and maximum lowmem_reserve in the zone */ 4024 for (j = i; j < MAX_NR_ZONES; j++) { 4025 if (zone->lowmem_reserve[j] > max) 4026 max = zone->lowmem_reserve[j]; 4027 } 4028 4029 /* we treat pages_high as reserved pages. */ 4030 max += zone->pages_high; 4031 4032 if (max > zone->present_pages) 4033 max = zone->present_pages; 4034 reserve_pages += max; 4035 } 4036 } 4037 totalreserve_pages = reserve_pages; 4038 } 4039 4040 /* 4041 * setup_per_zone_lowmem_reserve - called whenever 4042 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4043 * has a correct pages reserved value, so an adequate number of 4044 * pages are left in the zone after a successful __alloc_pages(). 4045 */ 4046 static void setup_per_zone_lowmem_reserve(void) 4047 { 4048 struct pglist_data *pgdat; 4049 enum zone_type j, idx; 4050 4051 for_each_online_pgdat(pgdat) { 4052 for (j = 0; j < MAX_NR_ZONES; j++) { 4053 struct zone *zone = pgdat->node_zones + j; 4054 unsigned long present_pages = zone->present_pages; 4055 4056 zone->lowmem_reserve[j] = 0; 4057 4058 idx = j; 4059 while (idx) { 4060 struct zone *lower_zone; 4061 4062 idx--; 4063 4064 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4065 sysctl_lowmem_reserve_ratio[idx] = 1; 4066 4067 lower_zone = pgdat->node_zones + idx; 4068 lower_zone->lowmem_reserve[j] = present_pages / 4069 sysctl_lowmem_reserve_ratio[idx]; 4070 present_pages += lower_zone->present_pages; 4071 } 4072 } 4073 } 4074 4075 /* update totalreserve_pages */ 4076 calculate_totalreserve_pages(); 4077 } 4078 4079 /** 4080 * setup_per_zone_pages_min - called when min_free_kbytes changes. 4081 * 4082 * Ensures that the pages_{min,low,high} values for each zone are set correctly 4083 * with respect to min_free_kbytes. 4084 */ 4085 void setup_per_zone_pages_min(void) 4086 { 4087 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4088 unsigned long lowmem_pages = 0; 4089 struct zone *zone; 4090 unsigned long flags; 4091 4092 /* Calculate total number of !ZONE_HIGHMEM pages */ 4093 for_each_zone(zone) { 4094 if (!is_highmem(zone)) 4095 lowmem_pages += zone->present_pages; 4096 } 4097 4098 for_each_zone(zone) { 4099 u64 tmp; 4100 4101 spin_lock_irqsave(&zone->lru_lock, flags); 4102 tmp = (u64)pages_min * zone->present_pages; 4103 do_div(tmp, lowmem_pages); 4104 if (is_highmem(zone)) { 4105 /* 4106 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4107 * need highmem pages, so cap pages_min to a small 4108 * value here. 4109 * 4110 * The (pages_high-pages_low) and (pages_low-pages_min) 4111 * deltas controls asynch page reclaim, and so should 4112 * not be capped for highmem. 4113 */ 4114 int min_pages; 4115 4116 min_pages = zone->present_pages / 1024; 4117 if (min_pages < SWAP_CLUSTER_MAX) 4118 min_pages = SWAP_CLUSTER_MAX; 4119 if (min_pages > 128) 4120 min_pages = 128; 4121 zone->pages_min = min_pages; 4122 } else { 4123 /* 4124 * If it's a lowmem zone, reserve a number of pages 4125 * proportionate to the zone's size. 4126 */ 4127 zone->pages_min = tmp; 4128 } 4129 4130 zone->pages_low = zone->pages_min + (tmp >> 2); 4131 zone->pages_high = zone->pages_min + (tmp >> 1); 4132 setup_zone_migrate_reserve(zone); 4133 spin_unlock_irqrestore(&zone->lru_lock, flags); 4134 } 4135 4136 /* update totalreserve_pages */ 4137 calculate_totalreserve_pages(); 4138 } 4139 4140 /* 4141 * Initialise min_free_kbytes. 4142 * 4143 * For small machines we want it small (128k min). For large machines 4144 * we want it large (64MB max). But it is not linear, because network 4145 * bandwidth does not increase linearly with machine size. We use 4146 * 4147 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 4148 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 4149 * 4150 * which yields 4151 * 4152 * 16MB: 512k 4153 * 32MB: 724k 4154 * 64MB: 1024k 4155 * 128MB: 1448k 4156 * 256MB: 2048k 4157 * 512MB: 2896k 4158 * 1024MB: 4096k 4159 * 2048MB: 5792k 4160 * 4096MB: 8192k 4161 * 8192MB: 11584k 4162 * 16384MB: 16384k 4163 */ 4164 static int __init init_per_zone_pages_min(void) 4165 { 4166 unsigned long lowmem_kbytes; 4167 4168 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 4169 4170 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 4171 if (min_free_kbytes < 128) 4172 min_free_kbytes = 128; 4173 if (min_free_kbytes > 65536) 4174 min_free_kbytes = 65536; 4175 setup_per_zone_pages_min(); 4176 setup_per_zone_lowmem_reserve(); 4177 return 0; 4178 } 4179 module_init(init_per_zone_pages_min) 4180 4181 /* 4182 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 4183 * that we can call two helper functions whenever min_free_kbytes 4184 * changes. 4185 */ 4186 int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 4187 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4188 { 4189 proc_dointvec(table, write, file, buffer, length, ppos); 4190 if (write) 4191 setup_per_zone_pages_min(); 4192 return 0; 4193 } 4194 4195 #ifdef CONFIG_NUMA 4196 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 4197 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4198 { 4199 struct zone *zone; 4200 int rc; 4201 4202 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4203 if (rc) 4204 return rc; 4205 4206 for_each_zone(zone) 4207 zone->min_unmapped_pages = (zone->present_pages * 4208 sysctl_min_unmapped_ratio) / 100; 4209 return 0; 4210 } 4211 4212 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 4213 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4214 { 4215 struct zone *zone; 4216 int rc; 4217 4218 rc = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4219 if (rc) 4220 return rc; 4221 4222 for_each_zone(zone) 4223 zone->min_slab_pages = (zone->present_pages * 4224 sysctl_min_slab_ratio) / 100; 4225 return 0; 4226 } 4227 #endif 4228 4229 /* 4230 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 4231 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 4232 * whenever sysctl_lowmem_reserve_ratio changes. 4233 * 4234 * The reserve ratio obviously has absolutely no relation with the 4235 * pages_min watermarks. The lowmem reserve ratio can only make sense 4236 * if in function of the boot time zone sizes. 4237 */ 4238 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 4239 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4240 { 4241 proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4242 setup_per_zone_lowmem_reserve(); 4243 return 0; 4244 } 4245 4246 /* 4247 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 4248 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 4249 * can have before it gets flushed back to buddy allocator. 4250 */ 4251 4252 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 4253 struct file *file, void __user *buffer, size_t *length, loff_t *ppos) 4254 { 4255 struct zone *zone; 4256 unsigned int cpu; 4257 int ret; 4258 4259 ret = proc_dointvec_minmax(table, write, file, buffer, length, ppos); 4260 if (!write || (ret == -EINVAL)) 4261 return ret; 4262 for_each_zone(zone) { 4263 for_each_online_cpu(cpu) { 4264 unsigned long high; 4265 high = zone->present_pages / percpu_pagelist_fraction; 4266 setup_pagelist_highmark(zone_pcp(zone, cpu), high); 4267 } 4268 } 4269 return 0; 4270 } 4271 4272 int hashdist = HASHDIST_DEFAULT; 4273 4274 #ifdef CONFIG_NUMA 4275 static int __init set_hashdist(char *str) 4276 { 4277 if (!str) 4278 return 0; 4279 hashdist = simple_strtoul(str, &str, 0); 4280 return 1; 4281 } 4282 __setup("hashdist=", set_hashdist); 4283 #endif 4284 4285 /* 4286 * allocate a large system hash table from bootmem 4287 * - it is assumed that the hash table must contain an exact power-of-2 4288 * quantity of entries 4289 * - limit is the number of hash buckets, not the total allocation size 4290 */ 4291 void *__init alloc_large_system_hash(const char *tablename, 4292 unsigned long bucketsize, 4293 unsigned long numentries, 4294 int scale, 4295 int flags, 4296 unsigned int *_hash_shift, 4297 unsigned int *_hash_mask, 4298 unsigned long limit) 4299 { 4300 unsigned long long max = limit; 4301 unsigned long log2qty, size; 4302 void *table = NULL; 4303 4304 /* allow the kernel cmdline to have a say */ 4305 if (!numentries) { 4306 /* round applicable memory size up to nearest megabyte */ 4307 numentries = nr_kernel_pages; 4308 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 4309 numentries >>= 20 - PAGE_SHIFT; 4310 numentries <<= 20 - PAGE_SHIFT; 4311 4312 /* limit to 1 bucket per 2^scale bytes of low memory */ 4313 if (scale > PAGE_SHIFT) 4314 numentries >>= (scale - PAGE_SHIFT); 4315 else 4316 numentries <<= (PAGE_SHIFT - scale); 4317 4318 /* Make sure we've got at least a 0-order allocation.. */ 4319 if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 4320 numentries = PAGE_SIZE / bucketsize; 4321 } 4322 numentries = roundup_pow_of_two(numentries); 4323 4324 /* limit allocation size to 1/16 total memory by default */ 4325 if (max == 0) { 4326 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 4327 do_div(max, bucketsize); 4328 } 4329 4330 if (numentries > max) 4331 numentries = max; 4332 4333 log2qty = ilog2(numentries); 4334 4335 do { 4336 size = bucketsize << log2qty; 4337 if (flags & HASH_EARLY) 4338 table = alloc_bootmem(size); 4339 else if (hashdist) 4340 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 4341 else { 4342 unsigned long order; 4343 for (order = 0; ((1UL << order) << PAGE_SHIFT) < size; order++) 4344 ; 4345 table = (void*) __get_free_pages(GFP_ATOMIC, order); 4346 /* 4347 * If bucketsize is not a power-of-two, we may free 4348 * some pages at the end of hash table. 4349 */ 4350 if (table) { 4351 unsigned long alloc_end = (unsigned long)table + 4352 (PAGE_SIZE << order); 4353 unsigned long used = (unsigned long)table + 4354 PAGE_ALIGN(size); 4355 split_page(virt_to_page(table), order); 4356 while (used < alloc_end) { 4357 free_page(used); 4358 used += PAGE_SIZE; 4359 } 4360 } 4361 } 4362 } while (!table && size > PAGE_SIZE && --log2qty); 4363 4364 if (!table) 4365 panic("Failed to allocate %s hash table\n", tablename); 4366 4367 printk(KERN_INFO "%s hash table entries: %d (order: %d, %lu bytes)\n", 4368 tablename, 4369 (1U << log2qty), 4370 ilog2(size) - PAGE_SHIFT, 4371 size); 4372 4373 if (_hash_shift) 4374 *_hash_shift = log2qty; 4375 if (_hash_mask) 4376 *_hash_mask = (1 << log2qty) - 1; 4377 4378 return table; 4379 } 4380 4381 #ifdef CONFIG_OUT_OF_LINE_PFN_TO_PAGE 4382 struct page *pfn_to_page(unsigned long pfn) 4383 { 4384 return __pfn_to_page(pfn); 4385 } 4386 unsigned long page_to_pfn(struct page *page) 4387 { 4388 return __page_to_pfn(page); 4389 } 4390 EXPORT_SYMBOL(pfn_to_page); 4391 EXPORT_SYMBOL(page_to_pfn); 4392 #endif /* CONFIG_OUT_OF_LINE_PFN_TO_PAGE */ 4393 4394 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 4395 static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 4396 unsigned long pfn) 4397 { 4398 #ifdef CONFIG_SPARSEMEM 4399 return __pfn_to_section(pfn)->pageblock_flags; 4400 #else 4401 return zone->pageblock_flags; 4402 #endif /* CONFIG_SPARSEMEM */ 4403 } 4404 4405 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 4406 { 4407 #ifdef CONFIG_SPARSEMEM 4408 pfn &= (PAGES_PER_SECTION-1); 4409 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4410 #else 4411 pfn = pfn - zone->zone_start_pfn; 4412 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 4413 #endif /* CONFIG_SPARSEMEM */ 4414 } 4415 4416 /** 4417 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 4418 * @page: The page within the block of interest 4419 * @start_bitidx: The first bit of interest to retrieve 4420 * @end_bitidx: The last bit of interest 4421 * returns pageblock_bits flags 4422 */ 4423 unsigned long get_pageblock_flags_group(struct page *page, 4424 int start_bitidx, int end_bitidx) 4425 { 4426 struct zone *zone; 4427 unsigned long *bitmap; 4428 unsigned long pfn, bitidx; 4429 unsigned long flags = 0; 4430 unsigned long value = 1; 4431 4432 zone = page_zone(page); 4433 pfn = page_to_pfn(page); 4434 bitmap = get_pageblock_bitmap(zone, pfn); 4435 bitidx = pfn_to_bitidx(zone, pfn); 4436 4437 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4438 if (test_bit(bitidx + start_bitidx, bitmap)) 4439 flags |= value; 4440 4441 return flags; 4442 } 4443 4444 /** 4445 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 4446 * @page: The page within the block of interest 4447 * @start_bitidx: The first bit of interest 4448 * @end_bitidx: The last bit of interest 4449 * @flags: The flags to set 4450 */ 4451 void set_pageblock_flags_group(struct page *page, unsigned long flags, 4452 int start_bitidx, int end_bitidx) 4453 { 4454 struct zone *zone; 4455 unsigned long *bitmap; 4456 unsigned long pfn, bitidx; 4457 unsigned long value = 1; 4458 4459 zone = page_zone(page); 4460 pfn = page_to_pfn(page); 4461 bitmap = get_pageblock_bitmap(zone, pfn); 4462 bitidx = pfn_to_bitidx(zone, pfn); 4463 4464 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 4465 if (flags & value) 4466 __set_bit(bitidx + start_bitidx, bitmap); 4467 else 4468 __clear_bit(bitidx + start_bitidx, bitmap); 4469 } 4470 4471 /* 4472 * This is designed as sub function...plz see page_isolation.c also. 4473 * set/clear page block's type to be ISOLATE. 4474 * page allocater never alloc memory from ISOLATE block. 4475 */ 4476 4477 int set_migratetype_isolate(struct page *page) 4478 { 4479 struct zone *zone; 4480 unsigned long flags; 4481 int ret = -EBUSY; 4482 4483 zone = page_zone(page); 4484 spin_lock_irqsave(&zone->lock, flags); 4485 /* 4486 * In future, more migrate types will be able to be isolation target. 4487 */ 4488 if (get_pageblock_migratetype(page) != MIGRATE_MOVABLE) 4489 goto out; 4490 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 4491 move_freepages_block(zone, page, MIGRATE_ISOLATE); 4492 ret = 0; 4493 out: 4494 spin_unlock_irqrestore(&zone->lock, flags); 4495 if (!ret) 4496 drain_all_pages(); 4497 return ret; 4498 } 4499 4500 void unset_migratetype_isolate(struct page *page) 4501 { 4502 struct zone *zone; 4503 unsigned long flags; 4504 zone = page_zone(page); 4505 spin_lock_irqsave(&zone->lock, flags); 4506 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 4507 goto out; 4508 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4509 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4510 out: 4511 spin_unlock_irqrestore(&zone->lock, flags); 4512 } 4513 4514 #ifdef CONFIG_MEMORY_HOTREMOVE 4515 /* 4516 * All pages in the range must be isolated before calling this. 4517 */ 4518 void 4519 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 4520 { 4521 struct page *page; 4522 struct zone *zone; 4523 int order, i; 4524 unsigned long pfn; 4525 unsigned long flags; 4526 /* find the first valid pfn */ 4527 for (pfn = start_pfn; pfn < end_pfn; pfn++) 4528 if (pfn_valid(pfn)) 4529 break; 4530 if (pfn == end_pfn) 4531 return; 4532 zone = page_zone(pfn_to_page(pfn)); 4533 spin_lock_irqsave(&zone->lock, flags); 4534 pfn = start_pfn; 4535 while (pfn < end_pfn) { 4536 if (!pfn_valid(pfn)) { 4537 pfn++; 4538 continue; 4539 } 4540 page = pfn_to_page(pfn); 4541 BUG_ON(page_count(page)); 4542 BUG_ON(!PageBuddy(page)); 4543 order = page_order(page); 4544 #ifdef CONFIG_DEBUG_VM 4545 printk(KERN_INFO "remove from free list %lx %d %lx\n", 4546 pfn, 1 << order, end_pfn); 4547 #endif 4548 list_del(&page->lru); 4549 rmv_page_order(page); 4550 zone->free_area[order].nr_free--; 4551 __mod_zone_page_state(zone, NR_FREE_PAGES, 4552 - (1UL << order)); 4553 for (i = 0; i < (1 << order); i++) 4554 SetPageReserved((page+i)); 4555 pfn += (1 << order); 4556 } 4557 spin_unlock_irqrestore(&zone->lock, flags); 4558 } 4559 #endif 4560