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