1 /* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17 #include <linux/stddef.h> 18 #include <linux/mm.h> 19 #include <linux/swap.h> 20 #include <linux/interrupt.h> 21 #include <linux/pagemap.h> 22 #include <linux/jiffies.h> 23 #include <linux/bootmem.h> 24 #include <linux/memblock.h> 25 #include <linux/compiler.h> 26 #include <linux/kernel.h> 27 #include <linux/kmemcheck.h> 28 #include <linux/module.h> 29 #include <linux/suspend.h> 30 #include <linux/pagevec.h> 31 #include <linux/blkdev.h> 32 #include <linux/slab.h> 33 #include <linux/ratelimit.h> 34 #include <linux/oom.h> 35 #include <linux/notifier.h> 36 #include <linux/topology.h> 37 #include <linux/sysctl.h> 38 #include <linux/cpu.h> 39 #include <linux/cpuset.h> 40 #include <linux/memory_hotplug.h> 41 #include <linux/nodemask.h> 42 #include <linux/vmalloc.h> 43 #include <linux/vmstat.h> 44 #include <linux/mempolicy.h> 45 #include <linux/stop_machine.h> 46 #include <linux/sort.h> 47 #include <linux/pfn.h> 48 #include <linux/backing-dev.h> 49 #include <linux/fault-inject.h> 50 #include <linux/page-isolation.h> 51 #include <linux/page_cgroup.h> 52 #include <linux/debugobjects.h> 53 #include <linux/kmemleak.h> 54 #include <linux/memory.h> 55 #include <linux/compaction.h> 56 #include <trace/events/kmem.h> 57 #include <linux/ftrace_event.h> 58 #include <linux/memcontrol.h> 59 #include <linux/prefetch.h> 60 #include <linux/page-debug-flags.h> 61 62 #include <asm/tlbflush.h> 63 #include <asm/div64.h> 64 #include "internal.h" 65 66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 67 DEFINE_PER_CPU(int, numa_node); 68 EXPORT_PER_CPU_SYMBOL(numa_node); 69 #endif 70 71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 72 /* 73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 76 * defined in <linux/topology.h>. 77 */ 78 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 79 EXPORT_PER_CPU_SYMBOL(_numa_mem_); 80 #endif 81 82 /* 83 * Array of node states. 84 */ 85 nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 86 [N_POSSIBLE] = NODE_MASK_ALL, 87 [N_ONLINE] = { { [0] = 1UL } }, 88 #ifndef CONFIG_NUMA 89 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 90 #ifdef CONFIG_HIGHMEM 91 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 92 #endif 93 [N_CPU] = { { [0] = 1UL } }, 94 #endif /* NUMA */ 95 }; 96 EXPORT_SYMBOL(node_states); 97 98 unsigned long totalram_pages __read_mostly; 99 unsigned long totalreserve_pages __read_mostly; 100 /* 101 * When calculating the number of globally allowed dirty pages, there 102 * is a certain number of per-zone reserves that should not be 103 * considered dirtyable memory. This is the sum of those reserves 104 * over all existing zones that contribute dirtyable memory. 105 */ 106 unsigned long dirty_balance_reserve __read_mostly; 107 108 int percpu_pagelist_fraction; 109 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 110 111 #ifdef CONFIG_PM_SLEEP 112 /* 113 * The following functions are used by the suspend/hibernate code to temporarily 114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 115 * while devices are suspended. To avoid races with the suspend/hibernate code, 116 * they should always be called with pm_mutex held (gfp_allowed_mask also should 117 * only be modified with pm_mutex held, unless the suspend/hibernate code is 118 * guaranteed not to run in parallel with that modification). 119 */ 120 121 static gfp_t saved_gfp_mask; 122 123 void pm_restore_gfp_mask(void) 124 { 125 WARN_ON(!mutex_is_locked(&pm_mutex)); 126 if (saved_gfp_mask) { 127 gfp_allowed_mask = saved_gfp_mask; 128 saved_gfp_mask = 0; 129 } 130 } 131 132 void pm_restrict_gfp_mask(void) 133 { 134 WARN_ON(!mutex_is_locked(&pm_mutex)); 135 WARN_ON(saved_gfp_mask); 136 saved_gfp_mask = gfp_allowed_mask; 137 gfp_allowed_mask &= ~GFP_IOFS; 138 } 139 140 bool pm_suspended_storage(void) 141 { 142 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) 143 return false; 144 return true; 145 } 146 #endif /* CONFIG_PM_SLEEP */ 147 148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 149 int pageblock_order __read_mostly; 150 #endif 151 152 static void __free_pages_ok(struct page *page, unsigned int order); 153 154 /* 155 * results with 256, 32 in the lowmem_reserve sysctl: 156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 157 * 1G machine -> (16M dma, 784M normal, 224M high) 158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA 161 * 162 * TBD: should special case ZONE_DMA32 machines here - in those we normally 163 * don't need any ZONE_NORMAL reservation 164 */ 165 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 166 #ifdef CONFIG_ZONE_DMA 167 256, 168 #endif 169 #ifdef CONFIG_ZONE_DMA32 170 256, 171 #endif 172 #ifdef CONFIG_HIGHMEM 173 32, 174 #endif 175 32, 176 }; 177 178 EXPORT_SYMBOL(totalram_pages); 179 180 static char * const zone_names[MAX_NR_ZONES] = { 181 #ifdef CONFIG_ZONE_DMA 182 "DMA", 183 #endif 184 #ifdef CONFIG_ZONE_DMA32 185 "DMA32", 186 #endif 187 "Normal", 188 #ifdef CONFIG_HIGHMEM 189 "HighMem", 190 #endif 191 "Movable", 192 }; 193 194 int min_free_kbytes = 1024; 195 196 static unsigned long __meminitdata nr_kernel_pages; 197 static unsigned long __meminitdata nr_all_pages; 198 static unsigned long __meminitdata dma_reserve; 199 200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 202 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 203 static unsigned long __initdata required_kernelcore; 204 static unsigned long __initdata required_movablecore; 205 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 206 207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 208 int movable_zone; 209 EXPORT_SYMBOL(movable_zone); 210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 211 212 #if MAX_NUMNODES > 1 213 int nr_node_ids __read_mostly = MAX_NUMNODES; 214 int nr_online_nodes __read_mostly = 1; 215 EXPORT_SYMBOL(nr_node_ids); 216 EXPORT_SYMBOL(nr_online_nodes); 217 #endif 218 219 int page_group_by_mobility_disabled __read_mostly; 220 221 static void set_pageblock_migratetype(struct page *page, int migratetype) 222 { 223 224 if (unlikely(page_group_by_mobility_disabled)) 225 migratetype = MIGRATE_UNMOVABLE; 226 227 set_pageblock_flags_group(page, (unsigned long)migratetype, 228 PB_migrate, PB_migrate_end); 229 } 230 231 bool oom_killer_disabled __read_mostly; 232 233 #ifdef CONFIG_DEBUG_VM 234 static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 235 { 236 int ret = 0; 237 unsigned seq; 238 unsigned long pfn = page_to_pfn(page); 239 240 do { 241 seq = zone_span_seqbegin(zone); 242 if (pfn >= zone->zone_start_pfn + zone->spanned_pages) 243 ret = 1; 244 else if (pfn < zone->zone_start_pfn) 245 ret = 1; 246 } while (zone_span_seqretry(zone, seq)); 247 248 return ret; 249 } 250 251 static int page_is_consistent(struct zone *zone, struct page *page) 252 { 253 if (!pfn_valid_within(page_to_pfn(page))) 254 return 0; 255 if (zone != page_zone(page)) 256 return 0; 257 258 return 1; 259 } 260 /* 261 * Temporary debugging check for pages not lying within a given zone. 262 */ 263 static int bad_range(struct zone *zone, struct page *page) 264 { 265 if (page_outside_zone_boundaries(zone, page)) 266 return 1; 267 if (!page_is_consistent(zone, page)) 268 return 1; 269 270 return 0; 271 } 272 #else 273 static inline int bad_range(struct zone *zone, struct page *page) 274 { 275 return 0; 276 } 277 #endif 278 279 static void bad_page(struct page *page) 280 { 281 static unsigned long resume; 282 static unsigned long nr_shown; 283 static unsigned long nr_unshown; 284 285 /* Don't complain about poisoned pages */ 286 if (PageHWPoison(page)) { 287 reset_page_mapcount(page); /* remove PageBuddy */ 288 return; 289 } 290 291 /* 292 * Allow a burst of 60 reports, then keep quiet for that minute; 293 * or allow a steady drip of one report per second. 294 */ 295 if (nr_shown == 60) { 296 if (time_before(jiffies, resume)) { 297 nr_unshown++; 298 goto out; 299 } 300 if (nr_unshown) { 301 printk(KERN_ALERT 302 "BUG: Bad page state: %lu messages suppressed\n", 303 nr_unshown); 304 nr_unshown = 0; 305 } 306 nr_shown = 0; 307 } 308 if (nr_shown++ == 0) 309 resume = jiffies + 60 * HZ; 310 311 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 312 current->comm, page_to_pfn(page)); 313 dump_page(page); 314 315 print_modules(); 316 dump_stack(); 317 out: 318 /* Leave bad fields for debug, except PageBuddy could make trouble */ 319 reset_page_mapcount(page); /* remove PageBuddy */ 320 add_taint(TAINT_BAD_PAGE); 321 } 322 323 /* 324 * Higher-order pages are called "compound pages". They are structured thusly: 325 * 326 * The first PAGE_SIZE page is called the "head page". 327 * 328 * The remaining PAGE_SIZE pages are called "tail pages". 329 * 330 * All pages have PG_compound set. All tail pages have their ->first_page 331 * pointing at the head page. 332 * 333 * The first tail page's ->lru.next holds the address of the compound page's 334 * put_page() function. Its ->lru.prev holds the order of allocation. 335 * This usage means that zero-order pages may not be compound. 336 */ 337 338 static void free_compound_page(struct page *page) 339 { 340 __free_pages_ok(page, compound_order(page)); 341 } 342 343 void prep_compound_page(struct page *page, unsigned long order) 344 { 345 int i; 346 int nr_pages = 1 << order; 347 348 set_compound_page_dtor(page, free_compound_page); 349 set_compound_order(page, order); 350 __SetPageHead(page); 351 for (i = 1; i < nr_pages; i++) { 352 struct page *p = page + i; 353 __SetPageTail(p); 354 set_page_count(p, 0); 355 p->first_page = page; 356 } 357 } 358 359 /* update __split_huge_page_refcount if you change this function */ 360 static int destroy_compound_page(struct page *page, unsigned long order) 361 { 362 int i; 363 int nr_pages = 1 << order; 364 int bad = 0; 365 366 if (unlikely(compound_order(page) != order) || 367 unlikely(!PageHead(page))) { 368 bad_page(page); 369 bad++; 370 } 371 372 __ClearPageHead(page); 373 374 for (i = 1; i < nr_pages; i++) { 375 struct page *p = page + i; 376 377 if (unlikely(!PageTail(p) || (p->first_page != page))) { 378 bad_page(page); 379 bad++; 380 } 381 __ClearPageTail(p); 382 } 383 384 return bad; 385 } 386 387 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags) 388 { 389 int i; 390 391 /* 392 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO 393 * and __GFP_HIGHMEM from hard or soft interrupt context. 394 */ 395 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt()); 396 for (i = 0; i < (1 << order); i++) 397 clear_highpage(page + i); 398 } 399 400 #ifdef CONFIG_DEBUG_PAGEALLOC 401 unsigned int _debug_guardpage_minorder; 402 403 static int __init debug_guardpage_minorder_setup(char *buf) 404 { 405 unsigned long res; 406 407 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 408 printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); 409 return 0; 410 } 411 _debug_guardpage_minorder = res; 412 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); 413 return 0; 414 } 415 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); 416 417 static inline void set_page_guard_flag(struct page *page) 418 { 419 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 420 } 421 422 static inline void clear_page_guard_flag(struct page *page) 423 { 424 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags); 425 } 426 #else 427 static inline void set_page_guard_flag(struct page *page) { } 428 static inline void clear_page_guard_flag(struct page *page) { } 429 #endif 430 431 static inline void set_page_order(struct page *page, int order) 432 { 433 set_page_private(page, order); 434 __SetPageBuddy(page); 435 } 436 437 static inline void rmv_page_order(struct page *page) 438 { 439 __ClearPageBuddy(page); 440 set_page_private(page, 0); 441 } 442 443 /* 444 * Locate the struct page for both the matching buddy in our 445 * pair (buddy1) and the combined O(n+1) page they form (page). 446 * 447 * 1) Any buddy B1 will have an order O twin B2 which satisfies 448 * the following equation: 449 * B2 = B1 ^ (1 << O) 450 * For example, if the starting buddy (buddy2) is #8 its order 451 * 1 buddy is #10: 452 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10 453 * 454 * 2) Any buddy B will have an order O+1 parent P which 455 * satisfies the following equation: 456 * P = B & ~(1 << O) 457 * 458 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER 459 */ 460 static inline unsigned long 461 __find_buddy_index(unsigned long page_idx, unsigned int order) 462 { 463 return page_idx ^ (1 << order); 464 } 465 466 /* 467 * This function checks whether a page is free && is the buddy 468 * we can do coalesce a page and its buddy if 469 * (a) the buddy is not in a hole && 470 * (b) the buddy is in the buddy system && 471 * (c) a page and its buddy have the same order && 472 * (d) a page and its buddy are in the same zone. 473 * 474 * For recording whether a page is in the buddy system, we set ->_mapcount -2. 475 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock. 476 * 477 * For recording page's order, we use page_private(page). 478 */ 479 static inline int page_is_buddy(struct page *page, struct page *buddy, 480 int order) 481 { 482 if (!pfn_valid_within(page_to_pfn(buddy))) 483 return 0; 484 485 if (page_zone_id(page) != page_zone_id(buddy)) 486 return 0; 487 488 if (page_is_guard(buddy) && page_order(buddy) == order) { 489 VM_BUG_ON(page_count(buddy) != 0); 490 return 1; 491 } 492 493 if (PageBuddy(buddy) && page_order(buddy) == order) { 494 VM_BUG_ON(page_count(buddy) != 0); 495 return 1; 496 } 497 return 0; 498 } 499 500 /* 501 * Freeing function for a buddy system allocator. 502 * 503 * The concept of a buddy system is to maintain direct-mapped table 504 * (containing bit values) for memory blocks of various "orders". 505 * The bottom level table contains the map for the smallest allocatable 506 * units of memory (here, pages), and each level above it describes 507 * pairs of units from the levels below, hence, "buddies". 508 * At a high level, all that happens here is marking the table entry 509 * at the bottom level available, and propagating the changes upward 510 * as necessary, plus some accounting needed to play nicely with other 511 * parts of the VM system. 512 * At each level, we keep a list of pages, which are heads of continuous 513 * free pages of length of (1 << order) and marked with _mapcount -2. Page's 514 * order is recorded in page_private(page) field. 515 * So when we are allocating or freeing one, we can derive the state of the 516 * other. That is, if we allocate a small block, and both were 517 * free, the remainder of the region must be split into blocks. 518 * If a block is freed, and its buddy is also free, then this 519 * triggers coalescing into a block of larger size. 520 * 521 * -- wli 522 */ 523 524 static inline void __free_one_page(struct page *page, 525 struct zone *zone, unsigned int order, 526 int migratetype) 527 { 528 unsigned long page_idx; 529 unsigned long combined_idx; 530 unsigned long uninitialized_var(buddy_idx); 531 struct page *buddy; 532 533 if (unlikely(PageCompound(page))) 534 if (unlikely(destroy_compound_page(page, order))) 535 return; 536 537 VM_BUG_ON(migratetype == -1); 538 539 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1); 540 541 VM_BUG_ON(page_idx & ((1 << order) - 1)); 542 VM_BUG_ON(bad_range(zone, page)); 543 544 while (order < MAX_ORDER-1) { 545 buddy_idx = __find_buddy_index(page_idx, order); 546 buddy = page + (buddy_idx - page_idx); 547 if (!page_is_buddy(page, buddy, order)) 548 break; 549 /* 550 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 551 * merge with it and move up one order. 552 */ 553 if (page_is_guard(buddy)) { 554 clear_page_guard_flag(buddy); 555 set_page_private(page, 0); 556 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 557 } else { 558 list_del(&buddy->lru); 559 zone->free_area[order].nr_free--; 560 rmv_page_order(buddy); 561 } 562 combined_idx = buddy_idx & page_idx; 563 page = page + (combined_idx - page_idx); 564 page_idx = combined_idx; 565 order++; 566 } 567 set_page_order(page, order); 568 569 /* 570 * If this is not the largest possible page, check if the buddy 571 * of the next-highest order is free. If it is, it's possible 572 * that pages are being freed that will coalesce soon. In case, 573 * that is happening, add the free page to the tail of the list 574 * so it's less likely to be used soon and more likely to be merged 575 * as a higher order page 576 */ 577 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 578 struct page *higher_page, *higher_buddy; 579 combined_idx = buddy_idx & page_idx; 580 higher_page = page + (combined_idx - page_idx); 581 buddy_idx = __find_buddy_index(combined_idx, order + 1); 582 higher_buddy = page + (buddy_idx - combined_idx); 583 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 584 list_add_tail(&page->lru, 585 &zone->free_area[order].free_list[migratetype]); 586 goto out; 587 } 588 } 589 590 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 591 out: 592 zone->free_area[order].nr_free++; 593 } 594 595 /* 596 * free_page_mlock() -- clean up attempts to free and mlocked() page. 597 * Page should not be on lru, so no need to fix that up. 598 * free_pages_check() will verify... 599 */ 600 static inline void free_page_mlock(struct page *page) 601 { 602 __dec_zone_page_state(page, NR_MLOCK); 603 __count_vm_event(UNEVICTABLE_MLOCKFREED); 604 } 605 606 static inline int free_pages_check(struct page *page) 607 { 608 if (unlikely(page_mapcount(page) | 609 (page->mapping != NULL) | 610 (atomic_read(&page->_count) != 0) | 611 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) | 612 (mem_cgroup_bad_page_check(page)))) { 613 bad_page(page); 614 return 1; 615 } 616 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 617 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 618 return 0; 619 } 620 621 /* 622 * Frees a number of pages from the PCP lists 623 * Assumes all pages on list are in same zone, and of same order. 624 * count is the number of pages to free. 625 * 626 * If the zone was previously in an "all pages pinned" state then look to 627 * see if this freeing clears that state. 628 * 629 * And clear the zone's pages_scanned counter, to hold off the "all pages are 630 * pinned" detection logic. 631 */ 632 static void free_pcppages_bulk(struct zone *zone, int count, 633 struct per_cpu_pages *pcp) 634 { 635 int migratetype = 0; 636 int batch_free = 0; 637 int to_free = count; 638 639 spin_lock(&zone->lock); 640 zone->all_unreclaimable = 0; 641 zone->pages_scanned = 0; 642 643 while (to_free) { 644 struct page *page; 645 struct list_head *list; 646 647 /* 648 * Remove pages from lists in a round-robin fashion. A 649 * batch_free count is maintained that is incremented when an 650 * empty list is encountered. This is so more pages are freed 651 * off fuller lists instead of spinning excessively around empty 652 * lists 653 */ 654 do { 655 batch_free++; 656 if (++migratetype == MIGRATE_PCPTYPES) 657 migratetype = 0; 658 list = &pcp->lists[migratetype]; 659 } while (list_empty(list)); 660 661 /* This is the only non-empty list. Free them all. */ 662 if (batch_free == MIGRATE_PCPTYPES) 663 batch_free = to_free; 664 665 do { 666 page = list_entry(list->prev, struct page, lru); 667 /* must delete as __free_one_page list manipulates */ 668 list_del(&page->lru); 669 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 670 __free_one_page(page, zone, 0, page_private(page)); 671 trace_mm_page_pcpu_drain(page, 0, page_private(page)); 672 } while (--to_free && --batch_free && !list_empty(list)); 673 } 674 __mod_zone_page_state(zone, NR_FREE_PAGES, count); 675 spin_unlock(&zone->lock); 676 } 677 678 static void free_one_page(struct zone *zone, struct page *page, int order, 679 int migratetype) 680 { 681 spin_lock(&zone->lock); 682 zone->all_unreclaimable = 0; 683 zone->pages_scanned = 0; 684 685 __free_one_page(page, zone, order, migratetype); 686 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order); 687 spin_unlock(&zone->lock); 688 } 689 690 static bool free_pages_prepare(struct page *page, unsigned int order) 691 { 692 int i; 693 int bad = 0; 694 695 trace_mm_page_free(page, order); 696 kmemcheck_free_shadow(page, order); 697 698 if (PageAnon(page)) 699 page->mapping = NULL; 700 for (i = 0; i < (1 << order); i++) 701 bad += free_pages_check(page + i); 702 if (bad) 703 return false; 704 705 if (!PageHighMem(page)) { 706 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order); 707 debug_check_no_obj_freed(page_address(page), 708 PAGE_SIZE << order); 709 } 710 arch_free_page(page, order); 711 kernel_map_pages(page, 1 << order, 0); 712 713 return true; 714 } 715 716 static void __free_pages_ok(struct page *page, unsigned int order) 717 { 718 unsigned long flags; 719 int wasMlocked = __TestClearPageMlocked(page); 720 721 if (!free_pages_prepare(page, order)) 722 return; 723 724 local_irq_save(flags); 725 if (unlikely(wasMlocked)) 726 free_page_mlock(page); 727 __count_vm_events(PGFREE, 1 << order); 728 free_one_page(page_zone(page), page, order, 729 get_pageblock_migratetype(page)); 730 local_irq_restore(flags); 731 } 732 733 void __meminit __free_pages_bootmem(struct page *page, unsigned int order) 734 { 735 unsigned int nr_pages = 1 << order; 736 unsigned int loop; 737 738 prefetchw(page); 739 for (loop = 0; loop < nr_pages; loop++) { 740 struct page *p = &page[loop]; 741 742 if (loop + 1 < nr_pages) 743 prefetchw(p + 1); 744 __ClearPageReserved(p); 745 set_page_count(p, 0); 746 } 747 748 set_page_refcounted(page); 749 __free_pages(page, order); 750 } 751 752 753 /* 754 * The order of subdivision here is critical for the IO subsystem. 755 * Please do not alter this order without good reasons and regression 756 * testing. Specifically, as large blocks of memory are subdivided, 757 * the order in which smaller blocks are delivered depends on the order 758 * they're subdivided in this function. This is the primary factor 759 * influencing the order in which pages are delivered to the IO 760 * subsystem according to empirical testing, and this is also justified 761 * by considering the behavior of a buddy system containing a single 762 * large block of memory acted on by a series of small allocations. 763 * This behavior is a critical factor in sglist merging's success. 764 * 765 * -- wli 766 */ 767 static inline void expand(struct zone *zone, struct page *page, 768 int low, int high, struct free_area *area, 769 int migratetype) 770 { 771 unsigned long size = 1 << high; 772 773 while (high > low) { 774 area--; 775 high--; 776 size >>= 1; 777 VM_BUG_ON(bad_range(zone, &page[size])); 778 779 #ifdef CONFIG_DEBUG_PAGEALLOC 780 if (high < debug_guardpage_minorder()) { 781 /* 782 * Mark as guard pages (or page), that will allow to 783 * merge back to allocator when buddy will be freed. 784 * Corresponding page table entries will not be touched, 785 * pages will stay not present in virtual address space 786 */ 787 INIT_LIST_HEAD(&page[size].lru); 788 set_page_guard_flag(&page[size]); 789 set_page_private(&page[size], high); 790 /* Guard pages are not available for any usage */ 791 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << high)); 792 continue; 793 } 794 #endif 795 list_add(&page[size].lru, &area->free_list[migratetype]); 796 area->nr_free++; 797 set_page_order(&page[size], high); 798 } 799 } 800 801 /* 802 * This page is about to be returned from the page allocator 803 */ 804 static inline int check_new_page(struct page *page) 805 { 806 if (unlikely(page_mapcount(page) | 807 (page->mapping != NULL) | 808 (atomic_read(&page->_count) != 0) | 809 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) | 810 (mem_cgroup_bad_page_check(page)))) { 811 bad_page(page); 812 return 1; 813 } 814 return 0; 815 } 816 817 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags) 818 { 819 int i; 820 821 for (i = 0; i < (1 << order); i++) { 822 struct page *p = page + i; 823 if (unlikely(check_new_page(p))) 824 return 1; 825 } 826 827 set_page_private(page, 0); 828 set_page_refcounted(page); 829 830 arch_alloc_page(page, order); 831 kernel_map_pages(page, 1 << order, 1); 832 833 if (gfp_flags & __GFP_ZERO) 834 prep_zero_page(page, order, gfp_flags); 835 836 if (order && (gfp_flags & __GFP_COMP)) 837 prep_compound_page(page, order); 838 839 return 0; 840 } 841 842 /* 843 * Go through the free lists for the given migratetype and remove 844 * the smallest available page from the freelists 845 */ 846 static inline 847 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 848 int migratetype) 849 { 850 unsigned int current_order; 851 struct free_area * area; 852 struct page *page; 853 854 /* Find a page of the appropriate size in the preferred list */ 855 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 856 area = &(zone->free_area[current_order]); 857 if (list_empty(&area->free_list[migratetype])) 858 continue; 859 860 page = list_entry(area->free_list[migratetype].next, 861 struct page, lru); 862 list_del(&page->lru); 863 rmv_page_order(page); 864 area->nr_free--; 865 expand(zone, page, order, current_order, area, migratetype); 866 return page; 867 } 868 869 return NULL; 870 } 871 872 873 /* 874 * This array describes the order lists are fallen back to when 875 * the free lists for the desirable migrate type are depleted 876 */ 877 static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = { 878 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 879 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 880 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 881 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */ 882 }; 883 884 /* 885 * Move the free pages in a range to the free lists of the requested type. 886 * Note that start_page and end_pages are not aligned on a pageblock 887 * boundary. If alignment is required, use move_freepages_block() 888 */ 889 static int move_freepages(struct zone *zone, 890 struct page *start_page, struct page *end_page, 891 int migratetype) 892 { 893 struct page *page; 894 unsigned long order; 895 int pages_moved = 0; 896 897 #ifndef CONFIG_HOLES_IN_ZONE 898 /* 899 * page_zone is not safe to call in this context when 900 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 901 * anyway as we check zone boundaries in move_freepages_block(). 902 * Remove at a later date when no bug reports exist related to 903 * grouping pages by mobility 904 */ 905 BUG_ON(page_zone(start_page) != page_zone(end_page)); 906 #endif 907 908 for (page = start_page; page <= end_page;) { 909 /* Make sure we are not inadvertently changing nodes */ 910 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone)); 911 912 if (!pfn_valid_within(page_to_pfn(page))) { 913 page++; 914 continue; 915 } 916 917 if (!PageBuddy(page)) { 918 page++; 919 continue; 920 } 921 922 order = page_order(page); 923 list_move(&page->lru, 924 &zone->free_area[order].free_list[migratetype]); 925 page += 1 << order; 926 pages_moved += 1 << order; 927 } 928 929 return pages_moved; 930 } 931 932 static int move_freepages_block(struct zone *zone, struct page *page, 933 int migratetype) 934 { 935 unsigned long start_pfn, end_pfn; 936 struct page *start_page, *end_page; 937 938 start_pfn = page_to_pfn(page); 939 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 940 start_page = pfn_to_page(start_pfn); 941 end_page = start_page + pageblock_nr_pages - 1; 942 end_pfn = start_pfn + pageblock_nr_pages - 1; 943 944 /* Do not cross zone boundaries */ 945 if (start_pfn < zone->zone_start_pfn) 946 start_page = page; 947 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages) 948 return 0; 949 950 return move_freepages(zone, start_page, end_page, migratetype); 951 } 952 953 static void change_pageblock_range(struct page *pageblock_page, 954 int start_order, int migratetype) 955 { 956 int nr_pageblocks = 1 << (start_order - pageblock_order); 957 958 while (nr_pageblocks--) { 959 set_pageblock_migratetype(pageblock_page, migratetype); 960 pageblock_page += pageblock_nr_pages; 961 } 962 } 963 964 /* Remove an element from the buddy allocator from the fallback list */ 965 static inline struct page * 966 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype) 967 { 968 struct free_area * area; 969 int current_order; 970 struct page *page; 971 int migratetype, i; 972 973 /* Find the largest possible block of pages in the other list */ 974 for (current_order = MAX_ORDER-1; current_order >= order; 975 --current_order) { 976 for (i = 0; i < MIGRATE_TYPES - 1; i++) { 977 migratetype = fallbacks[start_migratetype][i]; 978 979 /* MIGRATE_RESERVE handled later if necessary */ 980 if (migratetype == MIGRATE_RESERVE) 981 continue; 982 983 area = &(zone->free_area[current_order]); 984 if (list_empty(&area->free_list[migratetype])) 985 continue; 986 987 page = list_entry(area->free_list[migratetype].next, 988 struct page, lru); 989 area->nr_free--; 990 991 /* 992 * If breaking a large block of pages, move all free 993 * pages to the preferred allocation list. If falling 994 * back for a reclaimable kernel allocation, be more 995 * aggressive about taking ownership of free pages 996 */ 997 if (unlikely(current_order >= (pageblock_order >> 1)) || 998 start_migratetype == MIGRATE_RECLAIMABLE || 999 page_group_by_mobility_disabled) { 1000 unsigned long pages; 1001 pages = move_freepages_block(zone, page, 1002 start_migratetype); 1003 1004 /* Claim the whole block if over half of it is free */ 1005 if (pages >= (1 << (pageblock_order-1)) || 1006 page_group_by_mobility_disabled) 1007 set_pageblock_migratetype(page, 1008 start_migratetype); 1009 1010 migratetype = start_migratetype; 1011 } 1012 1013 /* Remove the page from the freelists */ 1014 list_del(&page->lru); 1015 rmv_page_order(page); 1016 1017 /* Take ownership for orders >= pageblock_order */ 1018 if (current_order >= pageblock_order) 1019 change_pageblock_range(page, current_order, 1020 start_migratetype); 1021 1022 expand(zone, page, order, current_order, area, migratetype); 1023 1024 trace_mm_page_alloc_extfrag(page, order, current_order, 1025 start_migratetype, migratetype); 1026 1027 return page; 1028 } 1029 } 1030 1031 return NULL; 1032 } 1033 1034 /* 1035 * Do the hard work of removing an element from the buddy allocator. 1036 * Call me with the zone->lock already held. 1037 */ 1038 static struct page *__rmqueue(struct zone *zone, unsigned int order, 1039 int migratetype) 1040 { 1041 struct page *page; 1042 1043 retry_reserve: 1044 page = __rmqueue_smallest(zone, order, migratetype); 1045 1046 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1047 page = __rmqueue_fallback(zone, order, migratetype); 1048 1049 /* 1050 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1051 * is used because __rmqueue_smallest is an inline function 1052 * and we want just one call site 1053 */ 1054 if (!page) { 1055 migratetype = MIGRATE_RESERVE; 1056 goto retry_reserve; 1057 } 1058 } 1059 1060 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1061 return page; 1062 } 1063 1064 /* 1065 * Obtain a specified number of elements from the buddy allocator, all under 1066 * a single hold of the lock, for efficiency. Add them to the supplied list. 1067 * Returns the number of new pages which were placed at *list. 1068 */ 1069 static int rmqueue_bulk(struct zone *zone, unsigned int order, 1070 unsigned long count, struct list_head *list, 1071 int migratetype, int cold) 1072 { 1073 int i; 1074 1075 spin_lock(&zone->lock); 1076 for (i = 0; i < count; ++i) { 1077 struct page *page = __rmqueue(zone, order, migratetype); 1078 if (unlikely(page == NULL)) 1079 break; 1080 1081 /* 1082 * Split buddy pages returned by expand() are received here 1083 * in physical page order. The page is added to the callers and 1084 * list and the list head then moves forward. From the callers 1085 * perspective, the linked list is ordered by page number in 1086 * some conditions. This is useful for IO devices that can 1087 * merge IO requests if the physical pages are ordered 1088 * properly. 1089 */ 1090 if (likely(cold == 0)) 1091 list_add(&page->lru, list); 1092 else 1093 list_add_tail(&page->lru, list); 1094 set_page_private(page, migratetype); 1095 list = &page->lru; 1096 } 1097 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1098 spin_unlock(&zone->lock); 1099 return i; 1100 } 1101 1102 #ifdef CONFIG_NUMA 1103 /* 1104 * Called from the vmstat counter updater to drain pagesets of this 1105 * currently executing processor on remote nodes after they have 1106 * expired. 1107 * 1108 * Note that this function must be called with the thread pinned to 1109 * a single processor. 1110 */ 1111 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1112 { 1113 unsigned long flags; 1114 int to_drain; 1115 1116 local_irq_save(flags); 1117 if (pcp->count >= pcp->batch) 1118 to_drain = pcp->batch; 1119 else 1120 to_drain = pcp->count; 1121 free_pcppages_bulk(zone, to_drain, pcp); 1122 pcp->count -= to_drain; 1123 local_irq_restore(flags); 1124 } 1125 #endif 1126 1127 /* 1128 * Drain pages of the indicated processor. 1129 * 1130 * The processor must either be the current processor and the 1131 * thread pinned to the current processor or a processor that 1132 * is not online. 1133 */ 1134 static void drain_pages(unsigned int cpu) 1135 { 1136 unsigned long flags; 1137 struct zone *zone; 1138 1139 for_each_populated_zone(zone) { 1140 struct per_cpu_pageset *pset; 1141 struct per_cpu_pages *pcp; 1142 1143 local_irq_save(flags); 1144 pset = per_cpu_ptr(zone->pageset, cpu); 1145 1146 pcp = &pset->pcp; 1147 if (pcp->count) { 1148 free_pcppages_bulk(zone, pcp->count, pcp); 1149 pcp->count = 0; 1150 } 1151 local_irq_restore(flags); 1152 } 1153 } 1154 1155 /* 1156 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1157 */ 1158 void drain_local_pages(void *arg) 1159 { 1160 drain_pages(smp_processor_id()); 1161 } 1162 1163 /* 1164 * Spill all the per-cpu pages from all CPUs back into the buddy allocator 1165 */ 1166 void drain_all_pages(void) 1167 { 1168 on_each_cpu(drain_local_pages, NULL, 1); 1169 } 1170 1171 #ifdef CONFIG_HIBERNATION 1172 1173 void mark_free_pages(struct zone *zone) 1174 { 1175 unsigned long pfn, max_zone_pfn; 1176 unsigned long flags; 1177 int order, t; 1178 struct list_head *curr; 1179 1180 if (!zone->spanned_pages) 1181 return; 1182 1183 spin_lock_irqsave(&zone->lock, flags); 1184 1185 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages; 1186 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1187 if (pfn_valid(pfn)) { 1188 struct page *page = pfn_to_page(pfn); 1189 1190 if (!swsusp_page_is_forbidden(page)) 1191 swsusp_unset_page_free(page); 1192 } 1193 1194 for_each_migratetype_order(order, t) { 1195 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1196 unsigned long i; 1197 1198 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1199 for (i = 0; i < (1UL << order); i++) 1200 swsusp_set_page_free(pfn_to_page(pfn + i)); 1201 } 1202 } 1203 spin_unlock_irqrestore(&zone->lock, flags); 1204 } 1205 #endif /* CONFIG_PM */ 1206 1207 /* 1208 * Free a 0-order page 1209 * cold == 1 ? free a cold page : free a hot page 1210 */ 1211 void free_hot_cold_page(struct page *page, int cold) 1212 { 1213 struct zone *zone = page_zone(page); 1214 struct per_cpu_pages *pcp; 1215 unsigned long flags; 1216 int migratetype; 1217 int wasMlocked = __TestClearPageMlocked(page); 1218 1219 if (!free_pages_prepare(page, 0)) 1220 return; 1221 1222 migratetype = get_pageblock_migratetype(page); 1223 set_page_private(page, migratetype); 1224 local_irq_save(flags); 1225 if (unlikely(wasMlocked)) 1226 free_page_mlock(page); 1227 __count_vm_event(PGFREE); 1228 1229 /* 1230 * We only track unmovable, reclaimable and movable on pcp lists. 1231 * Free ISOLATE pages back to the allocator because they are being 1232 * offlined but treat RESERVE as movable pages so we can get those 1233 * areas back if necessary. Otherwise, we may have to free 1234 * excessively into the page allocator 1235 */ 1236 if (migratetype >= MIGRATE_PCPTYPES) { 1237 if (unlikely(migratetype == MIGRATE_ISOLATE)) { 1238 free_one_page(zone, page, 0, migratetype); 1239 goto out; 1240 } 1241 migratetype = MIGRATE_MOVABLE; 1242 } 1243 1244 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1245 if (cold) 1246 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1247 else 1248 list_add(&page->lru, &pcp->lists[migratetype]); 1249 pcp->count++; 1250 if (pcp->count >= pcp->high) { 1251 free_pcppages_bulk(zone, pcp->batch, pcp); 1252 pcp->count -= pcp->batch; 1253 } 1254 1255 out: 1256 local_irq_restore(flags); 1257 } 1258 1259 /* 1260 * Free a list of 0-order pages 1261 */ 1262 void free_hot_cold_page_list(struct list_head *list, int cold) 1263 { 1264 struct page *page, *next; 1265 1266 list_for_each_entry_safe(page, next, list, lru) { 1267 trace_mm_page_free_batched(page, cold); 1268 free_hot_cold_page(page, cold); 1269 } 1270 } 1271 1272 /* 1273 * split_page takes a non-compound higher-order page, and splits it into 1274 * n (1<<order) sub-pages: page[0..n] 1275 * Each sub-page must be freed individually. 1276 * 1277 * Note: this is probably too low level an operation for use in drivers. 1278 * Please consult with lkml before using this in your driver. 1279 */ 1280 void split_page(struct page *page, unsigned int order) 1281 { 1282 int i; 1283 1284 VM_BUG_ON(PageCompound(page)); 1285 VM_BUG_ON(!page_count(page)); 1286 1287 #ifdef CONFIG_KMEMCHECK 1288 /* 1289 * Split shadow pages too, because free(page[0]) would 1290 * otherwise free the whole shadow. 1291 */ 1292 if (kmemcheck_page_is_tracked(page)) 1293 split_page(virt_to_page(page[0].shadow), order); 1294 #endif 1295 1296 for (i = 1; i < (1 << order); i++) 1297 set_page_refcounted(page + i); 1298 } 1299 1300 /* 1301 * Similar to split_page except the page is already free. As this is only 1302 * being used for migration, the migratetype of the block also changes. 1303 * As this is called with interrupts disabled, the caller is responsible 1304 * for calling arch_alloc_page() and kernel_map_page() after interrupts 1305 * are enabled. 1306 * 1307 * Note: this is probably too low level an operation for use in drivers. 1308 * Please consult with lkml before using this in your driver. 1309 */ 1310 int split_free_page(struct page *page) 1311 { 1312 unsigned int order; 1313 unsigned long watermark; 1314 struct zone *zone; 1315 1316 BUG_ON(!PageBuddy(page)); 1317 1318 zone = page_zone(page); 1319 order = page_order(page); 1320 1321 /* Obey watermarks as if the page was being allocated */ 1322 watermark = low_wmark_pages(zone) + (1 << order); 1323 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1324 return 0; 1325 1326 /* Remove page from free list */ 1327 list_del(&page->lru); 1328 zone->free_area[order].nr_free--; 1329 rmv_page_order(page); 1330 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order)); 1331 1332 /* Split into individual pages */ 1333 set_page_refcounted(page); 1334 split_page(page, order); 1335 1336 if (order >= pageblock_order - 1) { 1337 struct page *endpage = page + (1 << order) - 1; 1338 for (; page < endpage; page += pageblock_nr_pages) 1339 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1340 } 1341 1342 return 1 << order; 1343 } 1344 1345 /* 1346 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But 1347 * we cheat by calling it from here, in the order > 0 path. Saves a branch 1348 * or two. 1349 */ 1350 static inline 1351 struct page *buffered_rmqueue(struct zone *preferred_zone, 1352 struct zone *zone, int order, gfp_t gfp_flags, 1353 int migratetype) 1354 { 1355 unsigned long flags; 1356 struct page *page; 1357 int cold = !!(gfp_flags & __GFP_COLD); 1358 1359 again: 1360 if (likely(order == 0)) { 1361 struct per_cpu_pages *pcp; 1362 struct list_head *list; 1363 1364 local_irq_save(flags); 1365 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1366 list = &pcp->lists[migratetype]; 1367 if (list_empty(list)) { 1368 pcp->count += rmqueue_bulk(zone, 0, 1369 pcp->batch, list, 1370 migratetype, cold); 1371 if (unlikely(list_empty(list))) 1372 goto failed; 1373 } 1374 1375 if (cold) 1376 page = list_entry(list->prev, struct page, lru); 1377 else 1378 page = list_entry(list->next, struct page, lru); 1379 1380 list_del(&page->lru); 1381 pcp->count--; 1382 } else { 1383 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 1384 /* 1385 * __GFP_NOFAIL is not to be used in new code. 1386 * 1387 * All __GFP_NOFAIL callers should be fixed so that they 1388 * properly detect and handle allocation failures. 1389 * 1390 * We most definitely don't want callers attempting to 1391 * allocate greater than order-1 page units with 1392 * __GFP_NOFAIL. 1393 */ 1394 WARN_ON_ONCE(order > 1); 1395 } 1396 spin_lock_irqsave(&zone->lock, flags); 1397 page = __rmqueue(zone, order, migratetype); 1398 spin_unlock(&zone->lock); 1399 if (!page) 1400 goto failed; 1401 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order)); 1402 } 1403 1404 __count_zone_vm_events(PGALLOC, zone, 1 << order); 1405 zone_statistics(preferred_zone, zone, gfp_flags); 1406 local_irq_restore(flags); 1407 1408 VM_BUG_ON(bad_range(zone, page)); 1409 if (prep_new_page(page, order, gfp_flags)) 1410 goto again; 1411 return page; 1412 1413 failed: 1414 local_irq_restore(flags); 1415 return NULL; 1416 } 1417 1418 /* The ALLOC_WMARK bits are used as an index to zone->watermark */ 1419 #define ALLOC_WMARK_MIN WMARK_MIN 1420 #define ALLOC_WMARK_LOW WMARK_LOW 1421 #define ALLOC_WMARK_HIGH WMARK_HIGH 1422 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */ 1423 1424 /* Mask to get the watermark bits */ 1425 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1) 1426 1427 #define ALLOC_HARDER 0x10 /* try to alloc harder */ 1428 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */ 1429 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */ 1430 1431 #ifdef CONFIG_FAIL_PAGE_ALLOC 1432 1433 static struct { 1434 struct fault_attr attr; 1435 1436 u32 ignore_gfp_highmem; 1437 u32 ignore_gfp_wait; 1438 u32 min_order; 1439 } fail_page_alloc = { 1440 .attr = FAULT_ATTR_INITIALIZER, 1441 .ignore_gfp_wait = 1, 1442 .ignore_gfp_highmem = 1, 1443 .min_order = 1, 1444 }; 1445 1446 static int __init setup_fail_page_alloc(char *str) 1447 { 1448 return setup_fault_attr(&fail_page_alloc.attr, str); 1449 } 1450 __setup("fail_page_alloc=", setup_fail_page_alloc); 1451 1452 static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1453 { 1454 if (order < fail_page_alloc.min_order) 1455 return 0; 1456 if (gfp_mask & __GFP_NOFAIL) 1457 return 0; 1458 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 1459 return 0; 1460 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 1461 return 0; 1462 1463 return should_fail(&fail_page_alloc.attr, 1 << order); 1464 } 1465 1466 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 1467 1468 static int __init fail_page_alloc_debugfs(void) 1469 { 1470 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 1471 struct dentry *dir; 1472 1473 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 1474 &fail_page_alloc.attr); 1475 if (IS_ERR(dir)) 1476 return PTR_ERR(dir); 1477 1478 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 1479 &fail_page_alloc.ignore_gfp_wait)) 1480 goto fail; 1481 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 1482 &fail_page_alloc.ignore_gfp_highmem)) 1483 goto fail; 1484 if (!debugfs_create_u32("min-order", mode, dir, 1485 &fail_page_alloc.min_order)) 1486 goto fail; 1487 1488 return 0; 1489 fail: 1490 debugfs_remove_recursive(dir); 1491 1492 return -ENOMEM; 1493 } 1494 1495 late_initcall(fail_page_alloc_debugfs); 1496 1497 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 1498 1499 #else /* CONFIG_FAIL_PAGE_ALLOC */ 1500 1501 static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 1502 { 1503 return 0; 1504 } 1505 1506 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 1507 1508 /* 1509 * Return true if free pages are above 'mark'. This takes into account the order 1510 * of the allocation. 1511 */ 1512 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1513 int classzone_idx, int alloc_flags, long free_pages) 1514 { 1515 /* free_pages my go negative - that's OK */ 1516 long min = mark; 1517 int o; 1518 1519 free_pages -= (1 << order) - 1; 1520 if (alloc_flags & ALLOC_HIGH) 1521 min -= min / 2; 1522 if (alloc_flags & ALLOC_HARDER) 1523 min -= min / 4; 1524 1525 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 1526 return false; 1527 for (o = 0; o < order; o++) { 1528 /* At the next order, this order's pages become unavailable */ 1529 free_pages -= z->free_area[o].nr_free << o; 1530 1531 /* Require fewer higher order pages to be free */ 1532 min >>= 1; 1533 1534 if (free_pages <= min) 1535 return false; 1536 } 1537 return true; 1538 } 1539 1540 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark, 1541 int classzone_idx, int alloc_flags) 1542 { 1543 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1544 zone_page_state(z, NR_FREE_PAGES)); 1545 } 1546 1547 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark, 1548 int classzone_idx, int alloc_flags) 1549 { 1550 long free_pages = zone_page_state(z, NR_FREE_PAGES); 1551 1552 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 1553 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 1554 1555 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 1556 free_pages); 1557 } 1558 1559 #ifdef CONFIG_NUMA 1560 /* 1561 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 1562 * skip over zones that are not allowed by the cpuset, or that have 1563 * been recently (in last second) found to be nearly full. See further 1564 * comments in mmzone.h. Reduces cache footprint of zonelist scans 1565 * that have to skip over a lot of full or unallowed zones. 1566 * 1567 * If the zonelist cache is present in the passed in zonelist, then 1568 * returns a pointer to the allowed node mask (either the current 1569 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].) 1570 * 1571 * If the zonelist cache is not available for this zonelist, does 1572 * nothing and returns NULL. 1573 * 1574 * If the fullzones BITMAP in the zonelist cache is stale (more than 1575 * a second since last zap'd) then we zap it out (clear its bits.) 1576 * 1577 * We hold off even calling zlc_setup, until after we've checked the 1578 * first zone in the zonelist, on the theory that most allocations will 1579 * be satisfied from that first zone, so best to examine that zone as 1580 * quickly as we can. 1581 */ 1582 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1583 { 1584 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1585 nodemask_t *allowednodes; /* zonelist_cache approximation */ 1586 1587 zlc = zonelist->zlcache_ptr; 1588 if (!zlc) 1589 return NULL; 1590 1591 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 1592 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1593 zlc->last_full_zap = jiffies; 1594 } 1595 1596 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 1597 &cpuset_current_mems_allowed : 1598 &node_states[N_HIGH_MEMORY]; 1599 return allowednodes; 1600 } 1601 1602 /* 1603 * Given 'z' scanning a zonelist, run a couple of quick checks to see 1604 * if it is worth looking at further for free memory: 1605 * 1) Check that the zone isn't thought to be full (doesn't have its 1606 * bit set in the zonelist_cache fullzones BITMAP). 1607 * 2) Check that the zones node (obtained from the zonelist_cache 1608 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 1609 * Return true (non-zero) if zone is worth looking at further, or 1610 * else return false (zero) if it is not. 1611 * 1612 * This check -ignores- the distinction between various watermarks, 1613 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 1614 * found to be full for any variation of these watermarks, it will 1615 * be considered full for up to one second by all requests, unless 1616 * we are so low on memory on all allowed nodes that we are forced 1617 * into the second scan of the zonelist. 1618 * 1619 * In the second scan we ignore this zonelist cache and exactly 1620 * apply the watermarks to all zones, even it is slower to do so. 1621 * We are low on memory in the second scan, and should leave no stone 1622 * unturned looking for a free page. 1623 */ 1624 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1625 nodemask_t *allowednodes) 1626 { 1627 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1628 int i; /* index of *z in zonelist zones */ 1629 int n; /* node that zone *z is on */ 1630 1631 zlc = zonelist->zlcache_ptr; 1632 if (!zlc) 1633 return 1; 1634 1635 i = z - zonelist->_zonerefs; 1636 n = zlc->z_to_n[i]; 1637 1638 /* This zone is worth trying if it is allowed but not full */ 1639 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 1640 } 1641 1642 /* 1643 * Given 'z' scanning a zonelist, set the corresponding bit in 1644 * zlc->fullzones, so that subsequent attempts to allocate a page 1645 * from that zone don't waste time re-examining it. 1646 */ 1647 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1648 { 1649 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1650 int i; /* index of *z in zonelist zones */ 1651 1652 zlc = zonelist->zlcache_ptr; 1653 if (!zlc) 1654 return; 1655 1656 i = z - zonelist->_zonerefs; 1657 1658 set_bit(i, zlc->fullzones); 1659 } 1660 1661 /* 1662 * clear all zones full, called after direct reclaim makes progress so that 1663 * a zone that was recently full is not skipped over for up to a second 1664 */ 1665 static void zlc_clear_zones_full(struct zonelist *zonelist) 1666 { 1667 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 1668 1669 zlc = zonelist->zlcache_ptr; 1670 if (!zlc) 1671 return; 1672 1673 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 1674 } 1675 1676 #else /* CONFIG_NUMA */ 1677 1678 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 1679 { 1680 return NULL; 1681 } 1682 1683 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 1684 nodemask_t *allowednodes) 1685 { 1686 return 1; 1687 } 1688 1689 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 1690 { 1691 } 1692 1693 static void zlc_clear_zones_full(struct zonelist *zonelist) 1694 { 1695 } 1696 #endif /* CONFIG_NUMA */ 1697 1698 /* 1699 * get_page_from_freelist goes through the zonelist trying to allocate 1700 * a page. 1701 */ 1702 static struct page * 1703 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order, 1704 struct zonelist *zonelist, int high_zoneidx, int alloc_flags, 1705 struct zone *preferred_zone, int migratetype) 1706 { 1707 struct zoneref *z; 1708 struct page *page = NULL; 1709 int classzone_idx; 1710 struct zone *zone; 1711 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 1712 int zlc_active = 0; /* set if using zonelist_cache */ 1713 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 1714 1715 classzone_idx = zone_idx(preferred_zone); 1716 zonelist_scan: 1717 /* 1718 * Scan zonelist, looking for a zone with enough free. 1719 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 1720 */ 1721 for_each_zone_zonelist_nodemask(zone, z, zonelist, 1722 high_zoneidx, nodemask) { 1723 if (NUMA_BUILD && zlc_active && 1724 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1725 continue; 1726 if ((alloc_flags & ALLOC_CPUSET) && 1727 !cpuset_zone_allowed_softwall(zone, gfp_mask)) 1728 continue; 1729 /* 1730 * When allocating a page cache page for writing, we 1731 * want to get it from a zone that is within its dirty 1732 * limit, such that no single zone holds more than its 1733 * proportional share of globally allowed dirty pages. 1734 * The dirty limits take into account the zone's 1735 * lowmem reserves and high watermark so that kswapd 1736 * should be able to balance it without having to 1737 * write pages from its LRU list. 1738 * 1739 * This may look like it could increase pressure on 1740 * lower zones by failing allocations in higher zones 1741 * before they are full. But the pages that do spill 1742 * over are limited as the lower zones are protected 1743 * by this very same mechanism. It should not become 1744 * a practical burden to them. 1745 * 1746 * XXX: For now, allow allocations to potentially 1747 * exceed the per-zone dirty limit in the slowpath 1748 * (ALLOC_WMARK_LOW unset) before going into reclaim, 1749 * which is important when on a NUMA setup the allowed 1750 * zones are together not big enough to reach the 1751 * global limit. The proper fix for these situations 1752 * will require awareness of zones in the 1753 * dirty-throttling and the flusher threads. 1754 */ 1755 if ((alloc_flags & ALLOC_WMARK_LOW) && 1756 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone)) 1757 goto this_zone_full; 1758 1759 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 1760 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) { 1761 unsigned long mark; 1762 int ret; 1763 1764 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 1765 if (zone_watermark_ok(zone, order, mark, 1766 classzone_idx, alloc_flags)) 1767 goto try_this_zone; 1768 1769 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) { 1770 /* 1771 * we do zlc_setup if there are multiple nodes 1772 * and before considering the first zone allowed 1773 * by the cpuset. 1774 */ 1775 allowednodes = zlc_setup(zonelist, alloc_flags); 1776 zlc_active = 1; 1777 did_zlc_setup = 1; 1778 } 1779 1780 if (zone_reclaim_mode == 0) 1781 goto this_zone_full; 1782 1783 /* 1784 * As we may have just activated ZLC, check if the first 1785 * eligible zone has failed zone_reclaim recently. 1786 */ 1787 if (NUMA_BUILD && zlc_active && 1788 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 1789 continue; 1790 1791 ret = zone_reclaim(zone, gfp_mask, order); 1792 switch (ret) { 1793 case ZONE_RECLAIM_NOSCAN: 1794 /* did not scan */ 1795 continue; 1796 case ZONE_RECLAIM_FULL: 1797 /* scanned but unreclaimable */ 1798 continue; 1799 default: 1800 /* did we reclaim enough */ 1801 if (!zone_watermark_ok(zone, order, mark, 1802 classzone_idx, alloc_flags)) 1803 goto this_zone_full; 1804 } 1805 } 1806 1807 try_this_zone: 1808 page = buffered_rmqueue(preferred_zone, zone, order, 1809 gfp_mask, migratetype); 1810 if (page) 1811 break; 1812 this_zone_full: 1813 if (NUMA_BUILD) 1814 zlc_mark_zone_full(zonelist, z); 1815 } 1816 1817 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) { 1818 /* Disable zlc cache for second zonelist scan */ 1819 zlc_active = 0; 1820 goto zonelist_scan; 1821 } 1822 return page; 1823 } 1824 1825 /* 1826 * Large machines with many possible nodes should not always dump per-node 1827 * meminfo in irq context. 1828 */ 1829 static inline bool should_suppress_show_mem(void) 1830 { 1831 bool ret = false; 1832 1833 #if NODES_SHIFT > 8 1834 ret = in_interrupt(); 1835 #endif 1836 return ret; 1837 } 1838 1839 static DEFINE_RATELIMIT_STATE(nopage_rs, 1840 DEFAULT_RATELIMIT_INTERVAL, 1841 DEFAULT_RATELIMIT_BURST); 1842 1843 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 1844 { 1845 unsigned int filter = SHOW_MEM_FILTER_NODES; 1846 1847 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 1848 debug_guardpage_minorder() > 0) 1849 return; 1850 1851 /* 1852 * This documents exceptions given to allocations in certain 1853 * contexts that are allowed to allocate outside current's set 1854 * of allowed nodes. 1855 */ 1856 if (!(gfp_mask & __GFP_NOMEMALLOC)) 1857 if (test_thread_flag(TIF_MEMDIE) || 1858 (current->flags & (PF_MEMALLOC | PF_EXITING))) 1859 filter &= ~SHOW_MEM_FILTER_NODES; 1860 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 1861 filter &= ~SHOW_MEM_FILTER_NODES; 1862 1863 if (fmt) { 1864 struct va_format vaf; 1865 va_list args; 1866 1867 va_start(args, fmt); 1868 1869 vaf.fmt = fmt; 1870 vaf.va = &args; 1871 1872 pr_warn("%pV", &vaf); 1873 1874 va_end(args); 1875 } 1876 1877 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 1878 current->comm, order, gfp_mask); 1879 1880 dump_stack(); 1881 if (!should_suppress_show_mem()) 1882 show_mem(filter); 1883 } 1884 1885 static inline int 1886 should_alloc_retry(gfp_t gfp_mask, unsigned int order, 1887 unsigned long did_some_progress, 1888 unsigned long pages_reclaimed) 1889 { 1890 /* Do not loop if specifically requested */ 1891 if (gfp_mask & __GFP_NORETRY) 1892 return 0; 1893 1894 /* Always retry if specifically requested */ 1895 if (gfp_mask & __GFP_NOFAIL) 1896 return 1; 1897 1898 /* 1899 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim 1900 * making forward progress without invoking OOM. Suspend also disables 1901 * storage devices so kswapd will not help. Bail if we are suspending. 1902 */ 1903 if (!did_some_progress && pm_suspended_storage()) 1904 return 0; 1905 1906 /* 1907 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER 1908 * means __GFP_NOFAIL, but that may not be true in other 1909 * implementations. 1910 */ 1911 if (order <= PAGE_ALLOC_COSTLY_ORDER) 1912 return 1; 1913 1914 /* 1915 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is 1916 * specified, then we retry until we no longer reclaim any pages 1917 * (above), or we've reclaimed an order of pages at least as 1918 * large as the allocation's order. In both cases, if the 1919 * allocation still fails, we stop retrying. 1920 */ 1921 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order)) 1922 return 1; 1923 1924 return 0; 1925 } 1926 1927 static inline struct page * 1928 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 1929 struct zonelist *zonelist, enum zone_type high_zoneidx, 1930 nodemask_t *nodemask, struct zone *preferred_zone, 1931 int migratetype) 1932 { 1933 struct page *page; 1934 1935 /* Acquire the OOM killer lock for the zones in zonelist */ 1936 if (!try_set_zonelist_oom(zonelist, gfp_mask)) { 1937 schedule_timeout_uninterruptible(1); 1938 return NULL; 1939 } 1940 1941 /* 1942 * Go through the zonelist yet one more time, keep very high watermark 1943 * here, this is only to catch a parallel oom killing, we must fail if 1944 * we're still under heavy pressure. 1945 */ 1946 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, 1947 order, zonelist, high_zoneidx, 1948 ALLOC_WMARK_HIGH|ALLOC_CPUSET, 1949 preferred_zone, migratetype); 1950 if (page) 1951 goto out; 1952 1953 if (!(gfp_mask & __GFP_NOFAIL)) { 1954 /* The OOM killer will not help higher order allocs */ 1955 if (order > PAGE_ALLOC_COSTLY_ORDER) 1956 goto out; 1957 /* The OOM killer does not needlessly kill tasks for lowmem */ 1958 if (high_zoneidx < ZONE_NORMAL) 1959 goto out; 1960 /* 1961 * GFP_THISNODE contains __GFP_NORETRY and we never hit this. 1962 * Sanity check for bare calls of __GFP_THISNODE, not real OOM. 1963 * The caller should handle page allocation failure by itself if 1964 * it specifies __GFP_THISNODE. 1965 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER. 1966 */ 1967 if (gfp_mask & __GFP_THISNODE) 1968 goto out; 1969 } 1970 /* Exhausted what can be done so it's blamo time */ 1971 out_of_memory(zonelist, gfp_mask, order, nodemask); 1972 1973 out: 1974 clear_zonelist_oom(zonelist, gfp_mask); 1975 return page; 1976 } 1977 1978 #ifdef CONFIG_COMPACTION 1979 /* Try memory compaction for high-order allocations before reclaim */ 1980 static struct page * 1981 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 1982 struct zonelist *zonelist, enum zone_type high_zoneidx, 1983 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 1984 int migratetype, bool sync_migration, 1985 bool *deferred_compaction, 1986 unsigned long *did_some_progress) 1987 { 1988 struct page *page; 1989 1990 if (!order) 1991 return NULL; 1992 1993 if (compaction_deferred(preferred_zone)) { 1994 *deferred_compaction = true; 1995 return NULL; 1996 } 1997 1998 current->flags |= PF_MEMALLOC; 1999 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask, 2000 nodemask, sync_migration); 2001 current->flags &= ~PF_MEMALLOC; 2002 if (*did_some_progress != COMPACT_SKIPPED) { 2003 2004 /* Page migration frees to the PCP lists but we want merging */ 2005 drain_pages(get_cpu()); 2006 put_cpu(); 2007 2008 page = get_page_from_freelist(gfp_mask, nodemask, 2009 order, zonelist, high_zoneidx, 2010 alloc_flags, preferred_zone, 2011 migratetype); 2012 if (page) { 2013 preferred_zone->compact_considered = 0; 2014 preferred_zone->compact_defer_shift = 0; 2015 count_vm_event(COMPACTSUCCESS); 2016 return page; 2017 } 2018 2019 /* 2020 * It's bad if compaction run occurs and fails. 2021 * The most likely reason is that pages exist, 2022 * but not enough to satisfy watermarks. 2023 */ 2024 count_vm_event(COMPACTFAIL); 2025 2026 /* 2027 * As async compaction considers a subset of pageblocks, only 2028 * defer if the failure was a sync compaction failure. 2029 */ 2030 if (sync_migration) 2031 defer_compaction(preferred_zone); 2032 2033 cond_resched(); 2034 } 2035 2036 return NULL; 2037 } 2038 #else 2039 static inline struct page * 2040 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2041 struct zonelist *zonelist, enum zone_type high_zoneidx, 2042 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2043 int migratetype, bool sync_migration, 2044 bool *deferred_compaction, 2045 unsigned long *did_some_progress) 2046 { 2047 return NULL; 2048 } 2049 #endif /* CONFIG_COMPACTION */ 2050 2051 /* The really slow allocator path where we enter direct reclaim */ 2052 static inline struct page * 2053 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 2054 struct zonelist *zonelist, enum zone_type high_zoneidx, 2055 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone, 2056 int migratetype, unsigned long *did_some_progress) 2057 { 2058 struct page *page = NULL; 2059 struct reclaim_state reclaim_state; 2060 bool drained = false; 2061 2062 cond_resched(); 2063 2064 /* We now go into synchronous reclaim */ 2065 cpuset_memory_pressure_bump(); 2066 current->flags |= PF_MEMALLOC; 2067 lockdep_set_current_reclaim_state(gfp_mask); 2068 reclaim_state.reclaimed_slab = 0; 2069 current->reclaim_state = &reclaim_state; 2070 2071 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask); 2072 2073 current->reclaim_state = NULL; 2074 lockdep_clear_current_reclaim_state(); 2075 current->flags &= ~PF_MEMALLOC; 2076 2077 cond_resched(); 2078 2079 if (unlikely(!(*did_some_progress))) 2080 return NULL; 2081 2082 /* After successful reclaim, reconsider all zones for allocation */ 2083 if (NUMA_BUILD) 2084 zlc_clear_zones_full(zonelist); 2085 2086 retry: 2087 page = get_page_from_freelist(gfp_mask, nodemask, order, 2088 zonelist, high_zoneidx, 2089 alloc_flags, preferred_zone, 2090 migratetype); 2091 2092 /* 2093 * If an allocation failed after direct reclaim, it could be because 2094 * pages are pinned on the per-cpu lists. Drain them and try again 2095 */ 2096 if (!page && !drained) { 2097 drain_all_pages(); 2098 drained = true; 2099 goto retry; 2100 } 2101 2102 return page; 2103 } 2104 2105 /* 2106 * This is called in the allocator slow-path if the allocation request is of 2107 * sufficient urgency to ignore watermarks and take other desperate measures 2108 */ 2109 static inline struct page * 2110 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2111 struct zonelist *zonelist, enum zone_type high_zoneidx, 2112 nodemask_t *nodemask, struct zone *preferred_zone, 2113 int migratetype) 2114 { 2115 struct page *page; 2116 2117 do { 2118 page = get_page_from_freelist(gfp_mask, nodemask, order, 2119 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS, 2120 preferred_zone, migratetype); 2121 2122 if (!page && gfp_mask & __GFP_NOFAIL) 2123 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2124 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2125 2126 return page; 2127 } 2128 2129 static inline 2130 void wake_all_kswapd(unsigned int order, struct zonelist *zonelist, 2131 enum zone_type high_zoneidx, 2132 enum zone_type classzone_idx) 2133 { 2134 struct zoneref *z; 2135 struct zone *zone; 2136 2137 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) 2138 wakeup_kswapd(zone, order, classzone_idx); 2139 } 2140 2141 static inline int 2142 gfp_to_alloc_flags(gfp_t gfp_mask) 2143 { 2144 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2145 const gfp_t wait = gfp_mask & __GFP_WAIT; 2146 2147 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2148 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2149 2150 /* 2151 * The caller may dip into page reserves a bit more if the caller 2152 * cannot run direct reclaim, or if the caller has realtime scheduling 2153 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2154 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH). 2155 */ 2156 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2157 2158 if (!wait) { 2159 /* 2160 * Not worth trying to allocate harder for 2161 * __GFP_NOMEMALLOC even if it can't schedule. 2162 */ 2163 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2164 alloc_flags |= ALLOC_HARDER; 2165 /* 2166 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc. 2167 * See also cpuset_zone_allowed() comment in kernel/cpuset.c. 2168 */ 2169 alloc_flags &= ~ALLOC_CPUSET; 2170 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2171 alloc_flags |= ALLOC_HARDER; 2172 2173 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2174 if (!in_interrupt() && 2175 ((current->flags & PF_MEMALLOC) || 2176 unlikely(test_thread_flag(TIF_MEMDIE)))) 2177 alloc_flags |= ALLOC_NO_WATERMARKS; 2178 } 2179 2180 return alloc_flags; 2181 } 2182 2183 static inline struct page * 2184 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2185 struct zonelist *zonelist, enum zone_type high_zoneidx, 2186 nodemask_t *nodemask, struct zone *preferred_zone, 2187 int migratetype) 2188 { 2189 const gfp_t wait = gfp_mask & __GFP_WAIT; 2190 struct page *page = NULL; 2191 int alloc_flags; 2192 unsigned long pages_reclaimed = 0; 2193 unsigned long did_some_progress; 2194 bool sync_migration = false; 2195 bool deferred_compaction = false; 2196 2197 /* 2198 * In the slowpath, we sanity check order to avoid ever trying to 2199 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2200 * be using allocators in order of preference for an area that is 2201 * too large. 2202 */ 2203 if (order >= MAX_ORDER) { 2204 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2205 return NULL; 2206 } 2207 2208 /* 2209 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and 2210 * __GFP_NOWARN set) should not cause reclaim since the subsystem 2211 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim 2212 * using a larger set of nodes after it has established that the 2213 * allowed per node queues are empty and that nodes are 2214 * over allocated. 2215 */ 2216 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE) 2217 goto nopage; 2218 2219 restart: 2220 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2221 wake_all_kswapd(order, zonelist, high_zoneidx, 2222 zone_idx(preferred_zone)); 2223 2224 /* 2225 * OK, we're below the kswapd watermark and have kicked background 2226 * reclaim. Now things get more complex, so set up alloc_flags according 2227 * to how we want to proceed. 2228 */ 2229 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2230 2231 /* 2232 * Find the true preferred zone if the allocation is unconstrained by 2233 * cpusets. 2234 */ 2235 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask) 2236 first_zones_zonelist(zonelist, high_zoneidx, NULL, 2237 &preferred_zone); 2238 2239 rebalance: 2240 /* This is the last chance, in general, before the goto nopage. */ 2241 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist, 2242 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS, 2243 preferred_zone, migratetype); 2244 if (page) 2245 goto got_pg; 2246 2247 /* Allocate without watermarks if the context allows */ 2248 if (alloc_flags & ALLOC_NO_WATERMARKS) { 2249 page = __alloc_pages_high_priority(gfp_mask, order, 2250 zonelist, high_zoneidx, nodemask, 2251 preferred_zone, migratetype); 2252 if (page) 2253 goto got_pg; 2254 } 2255 2256 /* Atomic allocations - we can't balance anything */ 2257 if (!wait) 2258 goto nopage; 2259 2260 /* Avoid recursion of direct reclaim */ 2261 if (current->flags & PF_MEMALLOC) 2262 goto nopage; 2263 2264 /* Avoid allocations with no watermarks from looping endlessly */ 2265 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 2266 goto nopage; 2267 2268 /* 2269 * Try direct compaction. The first pass is asynchronous. Subsequent 2270 * attempts after direct reclaim are synchronous 2271 */ 2272 page = __alloc_pages_direct_compact(gfp_mask, order, 2273 zonelist, high_zoneidx, 2274 nodemask, 2275 alloc_flags, preferred_zone, 2276 migratetype, sync_migration, 2277 &deferred_compaction, 2278 &did_some_progress); 2279 if (page) 2280 goto got_pg; 2281 sync_migration = true; 2282 2283 /* 2284 * If compaction is deferred for high-order allocations, it is because 2285 * sync compaction recently failed. In this is the case and the caller 2286 * has requested the system not be heavily disrupted, fail the 2287 * allocation now instead of entering direct reclaim 2288 */ 2289 if (deferred_compaction && (gfp_mask & __GFP_NO_KSWAPD)) 2290 goto nopage; 2291 2292 /* Try direct reclaim and then allocating */ 2293 page = __alloc_pages_direct_reclaim(gfp_mask, order, 2294 zonelist, high_zoneidx, 2295 nodemask, 2296 alloc_flags, preferred_zone, 2297 migratetype, &did_some_progress); 2298 if (page) 2299 goto got_pg; 2300 2301 /* 2302 * If we failed to make any progress reclaiming, then we are 2303 * running out of options and have to consider going OOM 2304 */ 2305 if (!did_some_progress) { 2306 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) { 2307 if (oom_killer_disabled) 2308 goto nopage; 2309 page = __alloc_pages_may_oom(gfp_mask, order, 2310 zonelist, high_zoneidx, 2311 nodemask, preferred_zone, 2312 migratetype); 2313 if (page) 2314 goto got_pg; 2315 2316 if (!(gfp_mask & __GFP_NOFAIL)) { 2317 /* 2318 * The oom killer is not called for high-order 2319 * allocations that may fail, so if no progress 2320 * is being made, there are no other options and 2321 * retrying is unlikely to help. 2322 */ 2323 if (order > PAGE_ALLOC_COSTLY_ORDER) 2324 goto nopage; 2325 /* 2326 * The oom killer is not called for lowmem 2327 * allocations to prevent needlessly killing 2328 * innocent tasks. 2329 */ 2330 if (high_zoneidx < ZONE_NORMAL) 2331 goto nopage; 2332 } 2333 2334 goto restart; 2335 } 2336 } 2337 2338 /* Check if we should retry the allocation */ 2339 pages_reclaimed += did_some_progress; 2340 if (should_alloc_retry(gfp_mask, order, did_some_progress, 2341 pages_reclaimed)) { 2342 /* Wait for some write requests to complete then retry */ 2343 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50); 2344 goto rebalance; 2345 } else { 2346 /* 2347 * High-order allocations do not necessarily loop after 2348 * direct reclaim and reclaim/compaction depends on compaction 2349 * being called after reclaim so call directly if necessary 2350 */ 2351 page = __alloc_pages_direct_compact(gfp_mask, order, 2352 zonelist, high_zoneidx, 2353 nodemask, 2354 alloc_flags, preferred_zone, 2355 migratetype, sync_migration, 2356 &deferred_compaction, 2357 &did_some_progress); 2358 if (page) 2359 goto got_pg; 2360 } 2361 2362 nopage: 2363 warn_alloc_failed(gfp_mask, order, NULL); 2364 return page; 2365 got_pg: 2366 if (kmemcheck_enabled) 2367 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 2368 return page; 2369 2370 } 2371 2372 /* 2373 * This is the 'heart' of the zoned buddy allocator. 2374 */ 2375 struct page * 2376 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 2377 struct zonelist *zonelist, nodemask_t *nodemask) 2378 { 2379 enum zone_type high_zoneidx = gfp_zone(gfp_mask); 2380 struct zone *preferred_zone; 2381 struct page *page; 2382 int migratetype = allocflags_to_migratetype(gfp_mask); 2383 2384 gfp_mask &= gfp_allowed_mask; 2385 2386 lockdep_trace_alloc(gfp_mask); 2387 2388 might_sleep_if(gfp_mask & __GFP_WAIT); 2389 2390 if (should_fail_alloc_page(gfp_mask, order)) 2391 return NULL; 2392 2393 /* 2394 * Check the zones suitable for the gfp_mask contain at least one 2395 * valid zone. It's possible to have an empty zonelist as a result 2396 * of GFP_THISNODE and a memoryless node 2397 */ 2398 if (unlikely(!zonelist->_zonerefs->zone)) 2399 return NULL; 2400 2401 get_mems_allowed(); 2402 /* The preferred zone is used for statistics later */ 2403 first_zones_zonelist(zonelist, high_zoneidx, 2404 nodemask ? : &cpuset_current_mems_allowed, 2405 &preferred_zone); 2406 if (!preferred_zone) { 2407 put_mems_allowed(); 2408 return NULL; 2409 } 2410 2411 /* First allocation attempt */ 2412 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order, 2413 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET, 2414 preferred_zone, migratetype); 2415 if (unlikely(!page)) 2416 page = __alloc_pages_slowpath(gfp_mask, order, 2417 zonelist, high_zoneidx, nodemask, 2418 preferred_zone, migratetype); 2419 put_mems_allowed(); 2420 2421 trace_mm_page_alloc(page, order, gfp_mask, migratetype); 2422 return page; 2423 } 2424 EXPORT_SYMBOL(__alloc_pages_nodemask); 2425 2426 /* 2427 * Common helper functions. 2428 */ 2429 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 2430 { 2431 struct page *page; 2432 2433 /* 2434 * __get_free_pages() returns a 32-bit address, which cannot represent 2435 * a highmem page 2436 */ 2437 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 2438 2439 page = alloc_pages(gfp_mask, order); 2440 if (!page) 2441 return 0; 2442 return (unsigned long) page_address(page); 2443 } 2444 EXPORT_SYMBOL(__get_free_pages); 2445 2446 unsigned long get_zeroed_page(gfp_t gfp_mask) 2447 { 2448 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 2449 } 2450 EXPORT_SYMBOL(get_zeroed_page); 2451 2452 void __free_pages(struct page *page, unsigned int order) 2453 { 2454 if (put_page_testzero(page)) { 2455 if (order == 0) 2456 free_hot_cold_page(page, 0); 2457 else 2458 __free_pages_ok(page, order); 2459 } 2460 } 2461 2462 EXPORT_SYMBOL(__free_pages); 2463 2464 void free_pages(unsigned long addr, unsigned int order) 2465 { 2466 if (addr != 0) { 2467 VM_BUG_ON(!virt_addr_valid((void *)addr)); 2468 __free_pages(virt_to_page((void *)addr), order); 2469 } 2470 } 2471 2472 EXPORT_SYMBOL(free_pages); 2473 2474 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 2475 { 2476 if (addr) { 2477 unsigned long alloc_end = addr + (PAGE_SIZE << order); 2478 unsigned long used = addr + PAGE_ALIGN(size); 2479 2480 split_page(virt_to_page((void *)addr), order); 2481 while (used < alloc_end) { 2482 free_page(used); 2483 used += PAGE_SIZE; 2484 } 2485 } 2486 return (void *)addr; 2487 } 2488 2489 /** 2490 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 2491 * @size: the number of bytes to allocate 2492 * @gfp_mask: GFP flags for the allocation 2493 * 2494 * This function is similar to alloc_pages(), except that it allocates the 2495 * minimum number of pages to satisfy the request. alloc_pages() can only 2496 * allocate memory in power-of-two pages. 2497 * 2498 * This function is also limited by MAX_ORDER. 2499 * 2500 * Memory allocated by this function must be released by free_pages_exact(). 2501 */ 2502 void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 2503 { 2504 unsigned int order = get_order(size); 2505 unsigned long addr; 2506 2507 addr = __get_free_pages(gfp_mask, order); 2508 return make_alloc_exact(addr, order, size); 2509 } 2510 EXPORT_SYMBOL(alloc_pages_exact); 2511 2512 /** 2513 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 2514 * pages on a node. 2515 * @nid: the preferred node ID where memory should be allocated 2516 * @size: the number of bytes to allocate 2517 * @gfp_mask: GFP flags for the allocation 2518 * 2519 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 2520 * back. 2521 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 2522 * but is not exact. 2523 */ 2524 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 2525 { 2526 unsigned order = get_order(size); 2527 struct page *p = alloc_pages_node(nid, gfp_mask, order); 2528 if (!p) 2529 return NULL; 2530 return make_alloc_exact((unsigned long)page_address(p), order, size); 2531 } 2532 EXPORT_SYMBOL(alloc_pages_exact_nid); 2533 2534 /** 2535 * free_pages_exact - release memory allocated via alloc_pages_exact() 2536 * @virt: the value returned by alloc_pages_exact. 2537 * @size: size of allocation, same value as passed to alloc_pages_exact(). 2538 * 2539 * Release the memory allocated by a previous call to alloc_pages_exact. 2540 */ 2541 void free_pages_exact(void *virt, size_t size) 2542 { 2543 unsigned long addr = (unsigned long)virt; 2544 unsigned long end = addr + PAGE_ALIGN(size); 2545 2546 while (addr < end) { 2547 free_page(addr); 2548 addr += PAGE_SIZE; 2549 } 2550 } 2551 EXPORT_SYMBOL(free_pages_exact); 2552 2553 static unsigned int nr_free_zone_pages(int offset) 2554 { 2555 struct zoneref *z; 2556 struct zone *zone; 2557 2558 /* Just pick one node, since fallback list is circular */ 2559 unsigned int sum = 0; 2560 2561 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 2562 2563 for_each_zone_zonelist(zone, z, zonelist, offset) { 2564 unsigned long size = zone->present_pages; 2565 unsigned long high = high_wmark_pages(zone); 2566 if (size > high) 2567 sum += size - high; 2568 } 2569 2570 return sum; 2571 } 2572 2573 /* 2574 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL 2575 */ 2576 unsigned int nr_free_buffer_pages(void) 2577 { 2578 return nr_free_zone_pages(gfp_zone(GFP_USER)); 2579 } 2580 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 2581 2582 /* 2583 * Amount of free RAM allocatable within all zones 2584 */ 2585 unsigned int nr_free_pagecache_pages(void) 2586 { 2587 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 2588 } 2589 2590 static inline void show_node(struct zone *zone) 2591 { 2592 if (NUMA_BUILD) 2593 printk("Node %d ", zone_to_nid(zone)); 2594 } 2595 2596 void si_meminfo(struct sysinfo *val) 2597 { 2598 val->totalram = totalram_pages; 2599 val->sharedram = 0; 2600 val->freeram = global_page_state(NR_FREE_PAGES); 2601 val->bufferram = nr_blockdev_pages(); 2602 val->totalhigh = totalhigh_pages; 2603 val->freehigh = nr_free_highpages(); 2604 val->mem_unit = PAGE_SIZE; 2605 } 2606 2607 EXPORT_SYMBOL(si_meminfo); 2608 2609 #ifdef CONFIG_NUMA 2610 void si_meminfo_node(struct sysinfo *val, int nid) 2611 { 2612 pg_data_t *pgdat = NODE_DATA(nid); 2613 2614 val->totalram = pgdat->node_present_pages; 2615 val->freeram = node_page_state(nid, NR_FREE_PAGES); 2616 #ifdef CONFIG_HIGHMEM 2617 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages; 2618 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 2619 NR_FREE_PAGES); 2620 #else 2621 val->totalhigh = 0; 2622 val->freehigh = 0; 2623 #endif 2624 val->mem_unit = PAGE_SIZE; 2625 } 2626 #endif 2627 2628 /* 2629 * Determine whether the node should be displayed or not, depending on whether 2630 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 2631 */ 2632 bool skip_free_areas_node(unsigned int flags, int nid) 2633 { 2634 bool ret = false; 2635 2636 if (!(flags & SHOW_MEM_FILTER_NODES)) 2637 goto out; 2638 2639 get_mems_allowed(); 2640 ret = !node_isset(nid, cpuset_current_mems_allowed); 2641 put_mems_allowed(); 2642 out: 2643 return ret; 2644 } 2645 2646 #define K(x) ((x) << (PAGE_SHIFT-10)) 2647 2648 /* 2649 * Show free area list (used inside shift_scroll-lock stuff) 2650 * We also calculate the percentage fragmentation. We do this by counting the 2651 * memory on each free list with the exception of the first item on the list. 2652 * Suppresses nodes that are not allowed by current's cpuset if 2653 * SHOW_MEM_FILTER_NODES is passed. 2654 */ 2655 void show_free_areas(unsigned int filter) 2656 { 2657 int cpu; 2658 struct zone *zone; 2659 2660 for_each_populated_zone(zone) { 2661 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2662 continue; 2663 show_node(zone); 2664 printk("%s per-cpu:\n", zone->name); 2665 2666 for_each_online_cpu(cpu) { 2667 struct per_cpu_pageset *pageset; 2668 2669 pageset = per_cpu_ptr(zone->pageset, cpu); 2670 2671 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n", 2672 cpu, pageset->pcp.high, 2673 pageset->pcp.batch, pageset->pcp.count); 2674 } 2675 } 2676 2677 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 2678 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 2679 " unevictable:%lu" 2680 " dirty:%lu writeback:%lu unstable:%lu\n" 2681 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n" 2682 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n", 2683 global_page_state(NR_ACTIVE_ANON), 2684 global_page_state(NR_INACTIVE_ANON), 2685 global_page_state(NR_ISOLATED_ANON), 2686 global_page_state(NR_ACTIVE_FILE), 2687 global_page_state(NR_INACTIVE_FILE), 2688 global_page_state(NR_ISOLATED_FILE), 2689 global_page_state(NR_UNEVICTABLE), 2690 global_page_state(NR_FILE_DIRTY), 2691 global_page_state(NR_WRITEBACK), 2692 global_page_state(NR_UNSTABLE_NFS), 2693 global_page_state(NR_FREE_PAGES), 2694 global_page_state(NR_SLAB_RECLAIMABLE), 2695 global_page_state(NR_SLAB_UNRECLAIMABLE), 2696 global_page_state(NR_FILE_MAPPED), 2697 global_page_state(NR_SHMEM), 2698 global_page_state(NR_PAGETABLE), 2699 global_page_state(NR_BOUNCE)); 2700 2701 for_each_populated_zone(zone) { 2702 int i; 2703 2704 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2705 continue; 2706 show_node(zone); 2707 printk("%s" 2708 " free:%lukB" 2709 " min:%lukB" 2710 " low:%lukB" 2711 " high:%lukB" 2712 " active_anon:%lukB" 2713 " inactive_anon:%lukB" 2714 " active_file:%lukB" 2715 " inactive_file:%lukB" 2716 " unevictable:%lukB" 2717 " isolated(anon):%lukB" 2718 " isolated(file):%lukB" 2719 " present:%lukB" 2720 " mlocked:%lukB" 2721 " dirty:%lukB" 2722 " writeback:%lukB" 2723 " mapped:%lukB" 2724 " shmem:%lukB" 2725 " slab_reclaimable:%lukB" 2726 " slab_unreclaimable:%lukB" 2727 " kernel_stack:%lukB" 2728 " pagetables:%lukB" 2729 " unstable:%lukB" 2730 " bounce:%lukB" 2731 " writeback_tmp:%lukB" 2732 " pages_scanned:%lu" 2733 " all_unreclaimable? %s" 2734 "\n", 2735 zone->name, 2736 K(zone_page_state(zone, NR_FREE_PAGES)), 2737 K(min_wmark_pages(zone)), 2738 K(low_wmark_pages(zone)), 2739 K(high_wmark_pages(zone)), 2740 K(zone_page_state(zone, NR_ACTIVE_ANON)), 2741 K(zone_page_state(zone, NR_INACTIVE_ANON)), 2742 K(zone_page_state(zone, NR_ACTIVE_FILE)), 2743 K(zone_page_state(zone, NR_INACTIVE_FILE)), 2744 K(zone_page_state(zone, NR_UNEVICTABLE)), 2745 K(zone_page_state(zone, NR_ISOLATED_ANON)), 2746 K(zone_page_state(zone, NR_ISOLATED_FILE)), 2747 K(zone->present_pages), 2748 K(zone_page_state(zone, NR_MLOCK)), 2749 K(zone_page_state(zone, NR_FILE_DIRTY)), 2750 K(zone_page_state(zone, NR_WRITEBACK)), 2751 K(zone_page_state(zone, NR_FILE_MAPPED)), 2752 K(zone_page_state(zone, NR_SHMEM)), 2753 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 2754 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 2755 zone_page_state(zone, NR_KERNEL_STACK) * 2756 THREAD_SIZE / 1024, 2757 K(zone_page_state(zone, NR_PAGETABLE)), 2758 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 2759 K(zone_page_state(zone, NR_BOUNCE)), 2760 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 2761 zone->pages_scanned, 2762 (zone->all_unreclaimable ? "yes" : "no") 2763 ); 2764 printk("lowmem_reserve[]:"); 2765 for (i = 0; i < MAX_NR_ZONES; i++) 2766 printk(" %lu", zone->lowmem_reserve[i]); 2767 printk("\n"); 2768 } 2769 2770 for_each_populated_zone(zone) { 2771 unsigned long nr[MAX_ORDER], flags, order, total = 0; 2772 2773 if (skip_free_areas_node(filter, zone_to_nid(zone))) 2774 continue; 2775 show_node(zone); 2776 printk("%s: ", zone->name); 2777 2778 spin_lock_irqsave(&zone->lock, flags); 2779 for (order = 0; order < MAX_ORDER; order++) { 2780 nr[order] = zone->free_area[order].nr_free; 2781 total += nr[order] << order; 2782 } 2783 spin_unlock_irqrestore(&zone->lock, flags); 2784 for (order = 0; order < MAX_ORDER; order++) 2785 printk("%lu*%lukB ", nr[order], K(1UL) << order); 2786 printk("= %lukB\n", K(total)); 2787 } 2788 2789 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 2790 2791 show_swap_cache_info(); 2792 } 2793 2794 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 2795 { 2796 zoneref->zone = zone; 2797 zoneref->zone_idx = zone_idx(zone); 2798 } 2799 2800 /* 2801 * Builds allocation fallback zone lists. 2802 * 2803 * Add all populated zones of a node to the zonelist. 2804 */ 2805 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 2806 int nr_zones, enum zone_type zone_type) 2807 { 2808 struct zone *zone; 2809 2810 BUG_ON(zone_type >= MAX_NR_ZONES); 2811 zone_type++; 2812 2813 do { 2814 zone_type--; 2815 zone = pgdat->node_zones + zone_type; 2816 if (populated_zone(zone)) { 2817 zoneref_set_zone(zone, 2818 &zonelist->_zonerefs[nr_zones++]); 2819 check_highest_zone(zone_type); 2820 } 2821 2822 } while (zone_type); 2823 return nr_zones; 2824 } 2825 2826 2827 /* 2828 * zonelist_order: 2829 * 0 = automatic detection of better ordering. 2830 * 1 = order by ([node] distance, -zonetype) 2831 * 2 = order by (-zonetype, [node] distance) 2832 * 2833 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 2834 * the same zonelist. So only NUMA can configure this param. 2835 */ 2836 #define ZONELIST_ORDER_DEFAULT 0 2837 #define ZONELIST_ORDER_NODE 1 2838 #define ZONELIST_ORDER_ZONE 2 2839 2840 /* zonelist order in the kernel. 2841 * set_zonelist_order() will set this to NODE or ZONE. 2842 */ 2843 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 2844 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 2845 2846 2847 #ifdef CONFIG_NUMA 2848 /* The value user specified ....changed by config */ 2849 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2850 /* string for sysctl */ 2851 #define NUMA_ZONELIST_ORDER_LEN 16 2852 char numa_zonelist_order[16] = "default"; 2853 2854 /* 2855 * interface for configure zonelist ordering. 2856 * command line option "numa_zonelist_order" 2857 * = "[dD]efault - default, automatic configuration. 2858 * = "[nN]ode - order by node locality, then by zone within node 2859 * = "[zZ]one - order by zone, then by locality within zone 2860 */ 2861 2862 static int __parse_numa_zonelist_order(char *s) 2863 { 2864 if (*s == 'd' || *s == 'D') { 2865 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 2866 } else if (*s == 'n' || *s == 'N') { 2867 user_zonelist_order = ZONELIST_ORDER_NODE; 2868 } else if (*s == 'z' || *s == 'Z') { 2869 user_zonelist_order = ZONELIST_ORDER_ZONE; 2870 } else { 2871 printk(KERN_WARNING 2872 "Ignoring invalid numa_zonelist_order value: " 2873 "%s\n", s); 2874 return -EINVAL; 2875 } 2876 return 0; 2877 } 2878 2879 static __init int setup_numa_zonelist_order(char *s) 2880 { 2881 int ret; 2882 2883 if (!s) 2884 return 0; 2885 2886 ret = __parse_numa_zonelist_order(s); 2887 if (ret == 0) 2888 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 2889 2890 return ret; 2891 } 2892 early_param("numa_zonelist_order", setup_numa_zonelist_order); 2893 2894 /* 2895 * sysctl handler for numa_zonelist_order 2896 */ 2897 int numa_zonelist_order_handler(ctl_table *table, int write, 2898 void __user *buffer, size_t *length, 2899 loff_t *ppos) 2900 { 2901 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 2902 int ret; 2903 static DEFINE_MUTEX(zl_order_mutex); 2904 2905 mutex_lock(&zl_order_mutex); 2906 if (write) 2907 strcpy(saved_string, (char*)table->data); 2908 ret = proc_dostring(table, write, buffer, length, ppos); 2909 if (ret) 2910 goto out; 2911 if (write) { 2912 int oldval = user_zonelist_order; 2913 if (__parse_numa_zonelist_order((char*)table->data)) { 2914 /* 2915 * bogus value. restore saved string 2916 */ 2917 strncpy((char*)table->data, saved_string, 2918 NUMA_ZONELIST_ORDER_LEN); 2919 user_zonelist_order = oldval; 2920 } else if (oldval != user_zonelist_order) { 2921 mutex_lock(&zonelists_mutex); 2922 build_all_zonelists(NULL); 2923 mutex_unlock(&zonelists_mutex); 2924 } 2925 } 2926 out: 2927 mutex_unlock(&zl_order_mutex); 2928 return ret; 2929 } 2930 2931 2932 #define MAX_NODE_LOAD (nr_online_nodes) 2933 static int node_load[MAX_NUMNODES]; 2934 2935 /** 2936 * find_next_best_node - find the next node that should appear in a given node's fallback list 2937 * @node: node whose fallback list we're appending 2938 * @used_node_mask: nodemask_t of already used nodes 2939 * 2940 * We use a number of factors to determine which is the next node that should 2941 * appear on a given node's fallback list. The node should not have appeared 2942 * already in @node's fallback list, and it should be the next closest node 2943 * according to the distance array (which contains arbitrary distance values 2944 * from each node to each node in the system), and should also prefer nodes 2945 * with no CPUs, since presumably they'll have very little allocation pressure 2946 * on them otherwise. 2947 * It returns -1 if no node is found. 2948 */ 2949 static int find_next_best_node(int node, nodemask_t *used_node_mask) 2950 { 2951 int n, val; 2952 int min_val = INT_MAX; 2953 int best_node = -1; 2954 const struct cpumask *tmp = cpumask_of_node(0); 2955 2956 /* Use the local node if we haven't already */ 2957 if (!node_isset(node, *used_node_mask)) { 2958 node_set(node, *used_node_mask); 2959 return node; 2960 } 2961 2962 for_each_node_state(n, N_HIGH_MEMORY) { 2963 2964 /* Don't want a node to appear more than once */ 2965 if (node_isset(n, *used_node_mask)) 2966 continue; 2967 2968 /* Use the distance array to find the distance */ 2969 val = node_distance(node, n); 2970 2971 /* Penalize nodes under us ("prefer the next node") */ 2972 val += (n < node); 2973 2974 /* Give preference to headless and unused nodes */ 2975 tmp = cpumask_of_node(n); 2976 if (!cpumask_empty(tmp)) 2977 val += PENALTY_FOR_NODE_WITH_CPUS; 2978 2979 /* Slight preference for less loaded node */ 2980 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 2981 val += node_load[n]; 2982 2983 if (val < min_val) { 2984 min_val = val; 2985 best_node = n; 2986 } 2987 } 2988 2989 if (best_node >= 0) 2990 node_set(best_node, *used_node_mask); 2991 2992 return best_node; 2993 } 2994 2995 2996 /* 2997 * Build zonelists ordered by node and zones within node. 2998 * This results in maximum locality--normal zone overflows into local 2999 * DMA zone, if any--but risks exhausting DMA zone. 3000 */ 3001 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 3002 { 3003 int j; 3004 struct zonelist *zonelist; 3005 3006 zonelist = &pgdat->node_zonelists[0]; 3007 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 3008 ; 3009 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3010 MAX_NR_ZONES - 1); 3011 zonelist->_zonerefs[j].zone = NULL; 3012 zonelist->_zonerefs[j].zone_idx = 0; 3013 } 3014 3015 /* 3016 * Build gfp_thisnode zonelists 3017 */ 3018 static void build_thisnode_zonelists(pg_data_t *pgdat) 3019 { 3020 int j; 3021 struct zonelist *zonelist; 3022 3023 zonelist = &pgdat->node_zonelists[1]; 3024 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3025 zonelist->_zonerefs[j].zone = NULL; 3026 zonelist->_zonerefs[j].zone_idx = 0; 3027 } 3028 3029 /* 3030 * Build zonelists ordered by zone and nodes within zones. 3031 * This results in conserving DMA zone[s] until all Normal memory is 3032 * exhausted, but results in overflowing to remote node while memory 3033 * may still exist in local DMA zone. 3034 */ 3035 static int node_order[MAX_NUMNODES]; 3036 3037 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 3038 { 3039 int pos, j, node; 3040 int zone_type; /* needs to be signed */ 3041 struct zone *z; 3042 struct zonelist *zonelist; 3043 3044 zonelist = &pgdat->node_zonelists[0]; 3045 pos = 0; 3046 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 3047 for (j = 0; j < nr_nodes; j++) { 3048 node = node_order[j]; 3049 z = &NODE_DATA(node)->node_zones[zone_type]; 3050 if (populated_zone(z)) { 3051 zoneref_set_zone(z, 3052 &zonelist->_zonerefs[pos++]); 3053 check_highest_zone(zone_type); 3054 } 3055 } 3056 } 3057 zonelist->_zonerefs[pos].zone = NULL; 3058 zonelist->_zonerefs[pos].zone_idx = 0; 3059 } 3060 3061 static int default_zonelist_order(void) 3062 { 3063 int nid, zone_type; 3064 unsigned long low_kmem_size,total_size; 3065 struct zone *z; 3066 int average_size; 3067 /* 3068 * ZONE_DMA and ZONE_DMA32 can be very small area in the system. 3069 * If they are really small and used heavily, the system can fall 3070 * into OOM very easily. 3071 * This function detect ZONE_DMA/DMA32 size and configures zone order. 3072 */ 3073 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */ 3074 low_kmem_size = 0; 3075 total_size = 0; 3076 for_each_online_node(nid) { 3077 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 3078 z = &NODE_DATA(nid)->node_zones[zone_type]; 3079 if (populated_zone(z)) { 3080 if (zone_type < ZONE_NORMAL) 3081 low_kmem_size += z->present_pages; 3082 total_size += z->present_pages; 3083 } else if (zone_type == ZONE_NORMAL) { 3084 /* 3085 * If any node has only lowmem, then node order 3086 * is preferred to allow kernel allocations 3087 * locally; otherwise, they can easily infringe 3088 * on other nodes when there is an abundance of 3089 * lowmem available to allocate from. 3090 */ 3091 return ZONELIST_ORDER_NODE; 3092 } 3093 } 3094 } 3095 if (!low_kmem_size || /* there are no DMA area. */ 3096 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */ 3097 return ZONELIST_ORDER_NODE; 3098 /* 3099 * look into each node's config. 3100 * If there is a node whose DMA/DMA32 memory is very big area on 3101 * local memory, NODE_ORDER may be suitable. 3102 */ 3103 average_size = total_size / 3104 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1); 3105 for_each_online_node(nid) { 3106 low_kmem_size = 0; 3107 total_size = 0; 3108 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 3109 z = &NODE_DATA(nid)->node_zones[zone_type]; 3110 if (populated_zone(z)) { 3111 if (zone_type < ZONE_NORMAL) 3112 low_kmem_size += z->present_pages; 3113 total_size += z->present_pages; 3114 } 3115 } 3116 if (low_kmem_size && 3117 total_size > average_size && /* ignore small node */ 3118 low_kmem_size > total_size * 70/100) 3119 return ZONELIST_ORDER_NODE; 3120 } 3121 return ZONELIST_ORDER_ZONE; 3122 } 3123 3124 static void set_zonelist_order(void) 3125 { 3126 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 3127 current_zonelist_order = default_zonelist_order(); 3128 else 3129 current_zonelist_order = user_zonelist_order; 3130 } 3131 3132 static void build_zonelists(pg_data_t *pgdat) 3133 { 3134 int j, node, load; 3135 enum zone_type i; 3136 nodemask_t used_mask; 3137 int local_node, prev_node; 3138 struct zonelist *zonelist; 3139 int order = current_zonelist_order; 3140 3141 /* initialize zonelists */ 3142 for (i = 0; i < MAX_ZONELISTS; i++) { 3143 zonelist = pgdat->node_zonelists + i; 3144 zonelist->_zonerefs[0].zone = NULL; 3145 zonelist->_zonerefs[0].zone_idx = 0; 3146 } 3147 3148 /* NUMA-aware ordering of nodes */ 3149 local_node = pgdat->node_id; 3150 load = nr_online_nodes; 3151 prev_node = local_node; 3152 nodes_clear(used_mask); 3153 3154 memset(node_order, 0, sizeof(node_order)); 3155 j = 0; 3156 3157 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 3158 int distance = node_distance(local_node, node); 3159 3160 /* 3161 * If another node is sufficiently far away then it is better 3162 * to reclaim pages in a zone before going off node. 3163 */ 3164 if (distance > RECLAIM_DISTANCE) 3165 zone_reclaim_mode = 1; 3166 3167 /* 3168 * We don't want to pressure a particular node. 3169 * So adding penalty to the first node in same 3170 * distance group to make it round-robin. 3171 */ 3172 if (distance != node_distance(local_node, prev_node)) 3173 node_load[node] = load; 3174 3175 prev_node = node; 3176 load--; 3177 if (order == ZONELIST_ORDER_NODE) 3178 build_zonelists_in_node_order(pgdat, node); 3179 else 3180 node_order[j++] = node; /* remember order */ 3181 } 3182 3183 if (order == ZONELIST_ORDER_ZONE) { 3184 /* calculate node order -- i.e., DMA last! */ 3185 build_zonelists_in_zone_order(pgdat, j); 3186 } 3187 3188 build_thisnode_zonelists(pgdat); 3189 } 3190 3191 /* Construct the zonelist performance cache - see further mmzone.h */ 3192 static void build_zonelist_cache(pg_data_t *pgdat) 3193 { 3194 struct zonelist *zonelist; 3195 struct zonelist_cache *zlc; 3196 struct zoneref *z; 3197 3198 zonelist = &pgdat->node_zonelists[0]; 3199 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 3200 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 3201 for (z = zonelist->_zonerefs; z->zone; z++) 3202 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 3203 } 3204 3205 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 3206 /* 3207 * Return node id of node used for "local" allocations. 3208 * I.e., first node id of first zone in arg node's generic zonelist. 3209 * Used for initializing percpu 'numa_mem', which is used primarily 3210 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 3211 */ 3212 int local_memory_node(int node) 3213 { 3214 struct zone *zone; 3215 3216 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 3217 gfp_zone(GFP_KERNEL), 3218 NULL, 3219 &zone); 3220 return zone->node; 3221 } 3222 #endif 3223 3224 #else /* CONFIG_NUMA */ 3225 3226 static void set_zonelist_order(void) 3227 { 3228 current_zonelist_order = ZONELIST_ORDER_ZONE; 3229 } 3230 3231 static void build_zonelists(pg_data_t *pgdat) 3232 { 3233 int node, local_node; 3234 enum zone_type j; 3235 struct zonelist *zonelist; 3236 3237 local_node = pgdat->node_id; 3238 3239 zonelist = &pgdat->node_zonelists[0]; 3240 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1); 3241 3242 /* 3243 * Now we build the zonelist so that it contains the zones 3244 * of all the other nodes. 3245 * We don't want to pressure a particular node, so when 3246 * building the zones for node N, we make sure that the 3247 * zones coming right after the local ones are those from 3248 * node N+1 (modulo N) 3249 */ 3250 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 3251 if (!node_online(node)) 3252 continue; 3253 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3254 MAX_NR_ZONES - 1); 3255 } 3256 for (node = 0; node < local_node; node++) { 3257 if (!node_online(node)) 3258 continue; 3259 j = build_zonelists_node(NODE_DATA(node), zonelist, j, 3260 MAX_NR_ZONES - 1); 3261 } 3262 3263 zonelist->_zonerefs[j].zone = NULL; 3264 zonelist->_zonerefs[j].zone_idx = 0; 3265 } 3266 3267 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 3268 static void build_zonelist_cache(pg_data_t *pgdat) 3269 { 3270 pgdat->node_zonelists[0].zlcache_ptr = NULL; 3271 } 3272 3273 #endif /* CONFIG_NUMA */ 3274 3275 /* 3276 * Boot pageset table. One per cpu which is going to be used for all 3277 * zones and all nodes. The parameters will be set in such a way 3278 * that an item put on a list will immediately be handed over to 3279 * the buddy list. This is safe since pageset manipulation is done 3280 * with interrupts disabled. 3281 * 3282 * The boot_pagesets must be kept even after bootup is complete for 3283 * unused processors and/or zones. They do play a role for bootstrapping 3284 * hotplugged processors. 3285 * 3286 * zoneinfo_show() and maybe other functions do 3287 * not check if the processor is online before following the pageset pointer. 3288 * Other parts of the kernel may not check if the zone is available. 3289 */ 3290 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 3291 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 3292 static void setup_zone_pageset(struct zone *zone); 3293 3294 /* 3295 * Global mutex to protect against size modification of zonelists 3296 * as well as to serialize pageset setup for the new populated zone. 3297 */ 3298 DEFINE_MUTEX(zonelists_mutex); 3299 3300 /* return values int ....just for stop_machine() */ 3301 static __init_refok int __build_all_zonelists(void *data) 3302 { 3303 int nid; 3304 int cpu; 3305 3306 #ifdef CONFIG_NUMA 3307 memset(node_load, 0, sizeof(node_load)); 3308 #endif 3309 for_each_online_node(nid) { 3310 pg_data_t *pgdat = NODE_DATA(nid); 3311 3312 build_zonelists(pgdat); 3313 build_zonelist_cache(pgdat); 3314 } 3315 3316 /* 3317 * Initialize the boot_pagesets that are going to be used 3318 * for bootstrapping processors. The real pagesets for 3319 * each zone will be allocated later when the per cpu 3320 * allocator is available. 3321 * 3322 * boot_pagesets are used also for bootstrapping offline 3323 * cpus if the system is already booted because the pagesets 3324 * are needed to initialize allocators on a specific cpu too. 3325 * F.e. the percpu allocator needs the page allocator which 3326 * needs the percpu allocator in order to allocate its pagesets 3327 * (a chicken-egg dilemma). 3328 */ 3329 for_each_possible_cpu(cpu) { 3330 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 3331 3332 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 3333 /* 3334 * We now know the "local memory node" for each node-- 3335 * i.e., the node of the first zone in the generic zonelist. 3336 * Set up numa_mem percpu variable for on-line cpus. During 3337 * boot, only the boot cpu should be on-line; we'll init the 3338 * secondary cpus' numa_mem as they come on-line. During 3339 * node/memory hotplug, we'll fixup all on-line cpus. 3340 */ 3341 if (cpu_online(cpu)) 3342 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 3343 #endif 3344 } 3345 3346 return 0; 3347 } 3348 3349 /* 3350 * Called with zonelists_mutex held always 3351 * unless system_state == SYSTEM_BOOTING. 3352 */ 3353 void __ref build_all_zonelists(void *data) 3354 { 3355 set_zonelist_order(); 3356 3357 if (system_state == SYSTEM_BOOTING) { 3358 __build_all_zonelists(NULL); 3359 mminit_verify_zonelist(); 3360 cpuset_init_current_mems_allowed(); 3361 } else { 3362 /* we have to stop all cpus to guarantee there is no user 3363 of zonelist */ 3364 #ifdef CONFIG_MEMORY_HOTPLUG 3365 if (data) 3366 setup_zone_pageset((struct zone *)data); 3367 #endif 3368 stop_machine(__build_all_zonelists, NULL, NULL); 3369 /* cpuset refresh routine should be here */ 3370 } 3371 vm_total_pages = nr_free_pagecache_pages(); 3372 /* 3373 * Disable grouping by mobility if the number of pages in the 3374 * system is too low to allow the mechanism to work. It would be 3375 * more accurate, but expensive to check per-zone. This check is 3376 * made on memory-hotadd so a system can start with mobility 3377 * disabled and enable it later 3378 */ 3379 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 3380 page_group_by_mobility_disabled = 1; 3381 else 3382 page_group_by_mobility_disabled = 0; 3383 3384 printk("Built %i zonelists in %s order, mobility grouping %s. " 3385 "Total pages: %ld\n", 3386 nr_online_nodes, 3387 zonelist_order_name[current_zonelist_order], 3388 page_group_by_mobility_disabled ? "off" : "on", 3389 vm_total_pages); 3390 #ifdef CONFIG_NUMA 3391 printk("Policy zone: %s\n", zone_names[policy_zone]); 3392 #endif 3393 } 3394 3395 /* 3396 * Helper functions to size the waitqueue hash table. 3397 * Essentially these want to choose hash table sizes sufficiently 3398 * large so that collisions trying to wait on pages are rare. 3399 * But in fact, the number of active page waitqueues on typical 3400 * systems is ridiculously low, less than 200. So this is even 3401 * conservative, even though it seems large. 3402 * 3403 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 3404 * waitqueues, i.e. the size of the waitq table given the number of pages. 3405 */ 3406 #define PAGES_PER_WAITQUEUE 256 3407 3408 #ifndef CONFIG_MEMORY_HOTPLUG 3409 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3410 { 3411 unsigned long size = 1; 3412 3413 pages /= PAGES_PER_WAITQUEUE; 3414 3415 while (size < pages) 3416 size <<= 1; 3417 3418 /* 3419 * Once we have dozens or even hundreds of threads sleeping 3420 * on IO we've got bigger problems than wait queue collision. 3421 * Limit the size of the wait table to a reasonable size. 3422 */ 3423 size = min(size, 4096UL); 3424 3425 return max(size, 4UL); 3426 } 3427 #else 3428 /* 3429 * A zone's size might be changed by hot-add, so it is not possible to determine 3430 * a suitable size for its wait_table. So we use the maximum size now. 3431 * 3432 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 3433 * 3434 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 3435 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 3436 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 3437 * 3438 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 3439 * or more by the traditional way. (See above). It equals: 3440 * 3441 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 3442 * ia64(16K page size) : = ( 8G + 4M)byte. 3443 * powerpc (64K page size) : = (32G +16M)byte. 3444 */ 3445 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 3446 { 3447 return 4096UL; 3448 } 3449 #endif 3450 3451 /* 3452 * This is an integer logarithm so that shifts can be used later 3453 * to extract the more random high bits from the multiplicative 3454 * hash function before the remainder is taken. 3455 */ 3456 static inline unsigned long wait_table_bits(unsigned long size) 3457 { 3458 return ffz(~size); 3459 } 3460 3461 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1)) 3462 3463 /* 3464 * Check if a pageblock contains reserved pages 3465 */ 3466 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 3467 { 3468 unsigned long pfn; 3469 3470 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3471 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 3472 return 1; 3473 } 3474 return 0; 3475 } 3476 3477 /* 3478 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 3479 * of blocks reserved is based on min_wmark_pages(zone). The memory within 3480 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 3481 * higher will lead to a bigger reserve which will get freed as contiguous 3482 * blocks as reclaim kicks in 3483 */ 3484 static void setup_zone_migrate_reserve(struct zone *zone) 3485 { 3486 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 3487 struct page *page; 3488 unsigned long block_migratetype; 3489 int reserve; 3490 3491 /* 3492 * Get the start pfn, end pfn and the number of blocks to reserve 3493 * We have to be careful to be aligned to pageblock_nr_pages to 3494 * make sure that we always check pfn_valid for the first page in 3495 * the block. 3496 */ 3497 start_pfn = zone->zone_start_pfn; 3498 end_pfn = start_pfn + zone->spanned_pages; 3499 start_pfn = roundup(start_pfn, pageblock_nr_pages); 3500 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 3501 pageblock_order; 3502 3503 /* 3504 * Reserve blocks are generally in place to help high-order atomic 3505 * allocations that are short-lived. A min_free_kbytes value that 3506 * would result in more than 2 reserve blocks for atomic allocations 3507 * is assumed to be in place to help anti-fragmentation for the 3508 * future allocation of hugepages at runtime. 3509 */ 3510 reserve = min(2, reserve); 3511 3512 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 3513 if (!pfn_valid(pfn)) 3514 continue; 3515 page = pfn_to_page(pfn); 3516 3517 /* Watch out for overlapping nodes */ 3518 if (page_to_nid(page) != zone_to_nid(zone)) 3519 continue; 3520 3521 block_migratetype = get_pageblock_migratetype(page); 3522 3523 /* Only test what is necessary when the reserves are not met */ 3524 if (reserve > 0) { 3525 /* 3526 * Blocks with reserved pages will never free, skip 3527 * them. 3528 */ 3529 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 3530 if (pageblock_is_reserved(pfn, block_end_pfn)) 3531 continue; 3532 3533 /* If this block is reserved, account for it */ 3534 if (block_migratetype == MIGRATE_RESERVE) { 3535 reserve--; 3536 continue; 3537 } 3538 3539 /* Suitable for reserving if this block is movable */ 3540 if (block_migratetype == MIGRATE_MOVABLE) { 3541 set_pageblock_migratetype(page, 3542 MIGRATE_RESERVE); 3543 move_freepages_block(zone, page, 3544 MIGRATE_RESERVE); 3545 reserve--; 3546 continue; 3547 } 3548 } 3549 3550 /* 3551 * If the reserve is met and this is a previous reserved block, 3552 * take it back 3553 */ 3554 if (block_migratetype == MIGRATE_RESERVE) { 3555 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3556 move_freepages_block(zone, page, MIGRATE_MOVABLE); 3557 } 3558 } 3559 } 3560 3561 /* 3562 * Initially all pages are reserved - free ones are freed 3563 * up by free_all_bootmem() once the early boot process is 3564 * done. Non-atomic initialization, single-pass. 3565 */ 3566 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 3567 unsigned long start_pfn, enum memmap_context context) 3568 { 3569 struct page *page; 3570 unsigned long end_pfn = start_pfn + size; 3571 unsigned long pfn; 3572 struct zone *z; 3573 3574 if (highest_memmap_pfn < end_pfn - 1) 3575 highest_memmap_pfn = end_pfn - 1; 3576 3577 z = &NODE_DATA(nid)->node_zones[zone]; 3578 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 3579 /* 3580 * There can be holes in boot-time mem_map[]s 3581 * handed to this function. They do not 3582 * exist on hotplugged memory. 3583 */ 3584 if (context == MEMMAP_EARLY) { 3585 if (!early_pfn_valid(pfn)) 3586 continue; 3587 if (!early_pfn_in_nid(pfn, nid)) 3588 continue; 3589 } 3590 page = pfn_to_page(pfn); 3591 set_page_links(page, zone, nid, pfn); 3592 mminit_verify_page_links(page, zone, nid, pfn); 3593 init_page_count(page); 3594 reset_page_mapcount(page); 3595 SetPageReserved(page); 3596 /* 3597 * Mark the block movable so that blocks are reserved for 3598 * movable at startup. This will force kernel allocations 3599 * to reserve their blocks rather than leaking throughout 3600 * the address space during boot when many long-lived 3601 * kernel allocations are made. Later some blocks near 3602 * the start are marked MIGRATE_RESERVE by 3603 * setup_zone_migrate_reserve() 3604 * 3605 * bitmap is created for zone's valid pfn range. but memmap 3606 * can be created for invalid pages (for alignment) 3607 * check here not to call set_pageblock_migratetype() against 3608 * pfn out of zone. 3609 */ 3610 if ((z->zone_start_pfn <= pfn) 3611 && (pfn < z->zone_start_pfn + z->spanned_pages) 3612 && !(pfn & (pageblock_nr_pages - 1))) 3613 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 3614 3615 INIT_LIST_HEAD(&page->lru); 3616 #ifdef WANT_PAGE_VIRTUAL 3617 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 3618 if (!is_highmem_idx(zone)) 3619 set_page_address(page, __va(pfn << PAGE_SHIFT)); 3620 #endif 3621 } 3622 } 3623 3624 static void __meminit zone_init_free_lists(struct zone *zone) 3625 { 3626 int order, t; 3627 for_each_migratetype_order(order, t) { 3628 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 3629 zone->free_area[order].nr_free = 0; 3630 } 3631 } 3632 3633 #ifndef __HAVE_ARCH_MEMMAP_INIT 3634 #define memmap_init(size, nid, zone, start_pfn) \ 3635 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 3636 #endif 3637 3638 static int zone_batchsize(struct zone *zone) 3639 { 3640 #ifdef CONFIG_MMU 3641 int batch; 3642 3643 /* 3644 * The per-cpu-pages pools are set to around 1000th of the 3645 * size of the zone. But no more than 1/2 of a meg. 3646 * 3647 * OK, so we don't know how big the cache is. So guess. 3648 */ 3649 batch = zone->present_pages / 1024; 3650 if (batch * PAGE_SIZE > 512 * 1024) 3651 batch = (512 * 1024) / PAGE_SIZE; 3652 batch /= 4; /* We effectively *= 4 below */ 3653 if (batch < 1) 3654 batch = 1; 3655 3656 /* 3657 * Clamp the batch to a 2^n - 1 value. Having a power 3658 * of 2 value was found to be more likely to have 3659 * suboptimal cache aliasing properties in some cases. 3660 * 3661 * For example if 2 tasks are alternately allocating 3662 * batches of pages, one task can end up with a lot 3663 * of pages of one half of the possible page colors 3664 * and the other with pages of the other colors. 3665 */ 3666 batch = rounddown_pow_of_two(batch + batch/2) - 1; 3667 3668 return batch; 3669 3670 #else 3671 /* The deferral and batching of frees should be suppressed under NOMMU 3672 * conditions. 3673 * 3674 * The problem is that NOMMU needs to be able to allocate large chunks 3675 * of contiguous memory as there's no hardware page translation to 3676 * assemble apparent contiguous memory from discontiguous pages. 3677 * 3678 * Queueing large contiguous runs of pages for batching, however, 3679 * causes the pages to actually be freed in smaller chunks. As there 3680 * can be a significant delay between the individual batches being 3681 * recycled, this leads to the once large chunks of space being 3682 * fragmented and becoming unavailable for high-order allocations. 3683 */ 3684 return 0; 3685 #endif 3686 } 3687 3688 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 3689 { 3690 struct per_cpu_pages *pcp; 3691 int migratetype; 3692 3693 memset(p, 0, sizeof(*p)); 3694 3695 pcp = &p->pcp; 3696 pcp->count = 0; 3697 pcp->high = 6 * batch; 3698 pcp->batch = max(1UL, 1 * batch); 3699 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 3700 INIT_LIST_HEAD(&pcp->lists[migratetype]); 3701 } 3702 3703 /* 3704 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist 3705 * to the value high for the pageset p. 3706 */ 3707 3708 static void setup_pagelist_highmark(struct per_cpu_pageset *p, 3709 unsigned long high) 3710 { 3711 struct per_cpu_pages *pcp; 3712 3713 pcp = &p->pcp; 3714 pcp->high = high; 3715 pcp->batch = max(1UL, high/4); 3716 if ((high/4) > (PAGE_SHIFT * 8)) 3717 pcp->batch = PAGE_SHIFT * 8; 3718 } 3719 3720 static void setup_zone_pageset(struct zone *zone) 3721 { 3722 int cpu; 3723 3724 zone->pageset = alloc_percpu(struct per_cpu_pageset); 3725 3726 for_each_possible_cpu(cpu) { 3727 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 3728 3729 setup_pageset(pcp, zone_batchsize(zone)); 3730 3731 if (percpu_pagelist_fraction) 3732 setup_pagelist_highmark(pcp, 3733 (zone->present_pages / 3734 percpu_pagelist_fraction)); 3735 } 3736 } 3737 3738 /* 3739 * Allocate per cpu pagesets and initialize them. 3740 * Before this call only boot pagesets were available. 3741 */ 3742 void __init setup_per_cpu_pageset(void) 3743 { 3744 struct zone *zone; 3745 3746 for_each_populated_zone(zone) 3747 setup_zone_pageset(zone); 3748 } 3749 3750 static noinline __init_refok 3751 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 3752 { 3753 int i; 3754 struct pglist_data *pgdat = zone->zone_pgdat; 3755 size_t alloc_size; 3756 3757 /* 3758 * The per-page waitqueue mechanism uses hashed waitqueues 3759 * per zone. 3760 */ 3761 zone->wait_table_hash_nr_entries = 3762 wait_table_hash_nr_entries(zone_size_pages); 3763 zone->wait_table_bits = 3764 wait_table_bits(zone->wait_table_hash_nr_entries); 3765 alloc_size = zone->wait_table_hash_nr_entries 3766 * sizeof(wait_queue_head_t); 3767 3768 if (!slab_is_available()) { 3769 zone->wait_table = (wait_queue_head_t *) 3770 alloc_bootmem_node_nopanic(pgdat, alloc_size); 3771 } else { 3772 /* 3773 * This case means that a zone whose size was 0 gets new memory 3774 * via memory hot-add. 3775 * But it may be the case that a new node was hot-added. In 3776 * this case vmalloc() will not be able to use this new node's 3777 * memory - this wait_table must be initialized to use this new 3778 * node itself as well. 3779 * To use this new node's memory, further consideration will be 3780 * necessary. 3781 */ 3782 zone->wait_table = vmalloc(alloc_size); 3783 } 3784 if (!zone->wait_table) 3785 return -ENOMEM; 3786 3787 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i) 3788 init_waitqueue_head(zone->wait_table + i); 3789 3790 return 0; 3791 } 3792 3793 static int __zone_pcp_update(void *data) 3794 { 3795 struct zone *zone = data; 3796 int cpu; 3797 unsigned long batch = zone_batchsize(zone), flags; 3798 3799 for_each_possible_cpu(cpu) { 3800 struct per_cpu_pageset *pset; 3801 struct per_cpu_pages *pcp; 3802 3803 pset = per_cpu_ptr(zone->pageset, cpu); 3804 pcp = &pset->pcp; 3805 3806 local_irq_save(flags); 3807 free_pcppages_bulk(zone, pcp->count, pcp); 3808 setup_pageset(pset, batch); 3809 local_irq_restore(flags); 3810 } 3811 return 0; 3812 } 3813 3814 void zone_pcp_update(struct zone *zone) 3815 { 3816 stop_machine(__zone_pcp_update, zone, NULL); 3817 } 3818 3819 static __meminit void zone_pcp_init(struct zone *zone) 3820 { 3821 /* 3822 * per cpu subsystem is not up at this point. The following code 3823 * relies on the ability of the linker to provide the 3824 * offset of a (static) per cpu variable into the per cpu area. 3825 */ 3826 zone->pageset = &boot_pageset; 3827 3828 if (zone->present_pages) 3829 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 3830 zone->name, zone->present_pages, 3831 zone_batchsize(zone)); 3832 } 3833 3834 __meminit int init_currently_empty_zone(struct zone *zone, 3835 unsigned long zone_start_pfn, 3836 unsigned long size, 3837 enum memmap_context context) 3838 { 3839 struct pglist_data *pgdat = zone->zone_pgdat; 3840 int ret; 3841 ret = zone_wait_table_init(zone, size); 3842 if (ret) 3843 return ret; 3844 pgdat->nr_zones = zone_idx(zone) + 1; 3845 3846 zone->zone_start_pfn = zone_start_pfn; 3847 3848 mminit_dprintk(MMINIT_TRACE, "memmap_init", 3849 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 3850 pgdat->node_id, 3851 (unsigned long)zone_idx(zone), 3852 zone_start_pfn, (zone_start_pfn + size)); 3853 3854 zone_init_free_lists(zone); 3855 3856 return 0; 3857 } 3858 3859 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 3860 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 3861 /* 3862 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 3863 * Architectures may implement their own version but if add_active_range() 3864 * was used and there are no special requirements, this is a convenient 3865 * alternative 3866 */ 3867 int __meminit __early_pfn_to_nid(unsigned long pfn) 3868 { 3869 unsigned long start_pfn, end_pfn; 3870 int i, nid; 3871 3872 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 3873 if (start_pfn <= pfn && pfn < end_pfn) 3874 return nid; 3875 /* This is a memory hole */ 3876 return -1; 3877 } 3878 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 3879 3880 int __meminit early_pfn_to_nid(unsigned long pfn) 3881 { 3882 int nid; 3883 3884 nid = __early_pfn_to_nid(pfn); 3885 if (nid >= 0) 3886 return nid; 3887 /* just returns 0 */ 3888 return 0; 3889 } 3890 3891 #ifdef CONFIG_NODES_SPAN_OTHER_NODES 3892 bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 3893 { 3894 int nid; 3895 3896 nid = __early_pfn_to_nid(pfn); 3897 if (nid >= 0 && nid != node) 3898 return false; 3899 return true; 3900 } 3901 #endif 3902 3903 /** 3904 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range 3905 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 3906 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node 3907 * 3908 * If an architecture guarantees that all ranges registered with 3909 * add_active_ranges() contain no holes and may be freed, this 3910 * this function may be used instead of calling free_bootmem() manually. 3911 */ 3912 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 3913 { 3914 unsigned long start_pfn, end_pfn; 3915 int i, this_nid; 3916 3917 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 3918 start_pfn = min(start_pfn, max_low_pfn); 3919 end_pfn = min(end_pfn, max_low_pfn); 3920 3921 if (start_pfn < end_pfn) 3922 free_bootmem_node(NODE_DATA(this_nid), 3923 PFN_PHYS(start_pfn), 3924 (end_pfn - start_pfn) << PAGE_SHIFT); 3925 } 3926 } 3927 3928 int __init add_from_early_node_map(struct range *range, int az, 3929 int nr_range, int nid) 3930 { 3931 unsigned long start_pfn, end_pfn; 3932 int i; 3933 3934 /* need to go over early_node_map to find out good range for node */ 3935 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) 3936 nr_range = add_range(range, az, nr_range, start_pfn, end_pfn); 3937 return nr_range; 3938 } 3939 3940 /** 3941 * sparse_memory_present_with_active_regions - Call memory_present for each active range 3942 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 3943 * 3944 * If an architecture guarantees that all ranges registered with 3945 * add_active_ranges() contain no holes and may be freed, this 3946 * function may be used instead of calling memory_present() manually. 3947 */ 3948 void __init sparse_memory_present_with_active_regions(int nid) 3949 { 3950 unsigned long start_pfn, end_pfn; 3951 int i, this_nid; 3952 3953 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 3954 memory_present(this_nid, start_pfn, end_pfn); 3955 } 3956 3957 /** 3958 * get_pfn_range_for_nid - Return the start and end page frames for a node 3959 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 3960 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 3961 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 3962 * 3963 * It returns the start and end page frame of a node based on information 3964 * provided by an arch calling add_active_range(). If called for a node 3965 * with no available memory, a warning is printed and the start and end 3966 * PFNs will be 0. 3967 */ 3968 void __meminit get_pfn_range_for_nid(unsigned int nid, 3969 unsigned long *start_pfn, unsigned long *end_pfn) 3970 { 3971 unsigned long this_start_pfn, this_end_pfn; 3972 int i; 3973 3974 *start_pfn = -1UL; 3975 *end_pfn = 0; 3976 3977 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 3978 *start_pfn = min(*start_pfn, this_start_pfn); 3979 *end_pfn = max(*end_pfn, this_end_pfn); 3980 } 3981 3982 if (*start_pfn == -1UL) 3983 *start_pfn = 0; 3984 } 3985 3986 /* 3987 * This finds a zone that can be used for ZONE_MOVABLE pages. The 3988 * assumption is made that zones within a node are ordered in monotonic 3989 * increasing memory addresses so that the "highest" populated zone is used 3990 */ 3991 static void __init find_usable_zone_for_movable(void) 3992 { 3993 int zone_index; 3994 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 3995 if (zone_index == ZONE_MOVABLE) 3996 continue; 3997 3998 if (arch_zone_highest_possible_pfn[zone_index] > 3999 arch_zone_lowest_possible_pfn[zone_index]) 4000 break; 4001 } 4002 4003 VM_BUG_ON(zone_index == -1); 4004 movable_zone = zone_index; 4005 } 4006 4007 /* 4008 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 4009 * because it is sized independent of architecture. Unlike the other zones, 4010 * the starting point for ZONE_MOVABLE is not fixed. It may be different 4011 * in each node depending on the size of each node and how evenly kernelcore 4012 * is distributed. This helper function adjusts the zone ranges 4013 * provided by the architecture for a given node by using the end of the 4014 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 4015 * zones within a node are in order of monotonic increases memory addresses 4016 */ 4017 static void __meminit adjust_zone_range_for_zone_movable(int nid, 4018 unsigned long zone_type, 4019 unsigned long node_start_pfn, 4020 unsigned long node_end_pfn, 4021 unsigned long *zone_start_pfn, 4022 unsigned long *zone_end_pfn) 4023 { 4024 /* Only adjust if ZONE_MOVABLE is on this node */ 4025 if (zone_movable_pfn[nid]) { 4026 /* Size ZONE_MOVABLE */ 4027 if (zone_type == ZONE_MOVABLE) { 4028 *zone_start_pfn = zone_movable_pfn[nid]; 4029 *zone_end_pfn = min(node_end_pfn, 4030 arch_zone_highest_possible_pfn[movable_zone]); 4031 4032 /* Adjust for ZONE_MOVABLE starting within this range */ 4033 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 4034 *zone_end_pfn > zone_movable_pfn[nid]) { 4035 *zone_end_pfn = zone_movable_pfn[nid]; 4036 4037 /* Check if this whole range is within ZONE_MOVABLE */ 4038 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 4039 *zone_start_pfn = *zone_end_pfn; 4040 } 4041 } 4042 4043 /* 4044 * Return the number of pages a zone spans in a node, including holes 4045 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 4046 */ 4047 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 4048 unsigned long zone_type, 4049 unsigned long *ignored) 4050 { 4051 unsigned long node_start_pfn, node_end_pfn; 4052 unsigned long zone_start_pfn, zone_end_pfn; 4053 4054 /* Get the start and end of the node and zone */ 4055 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4056 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 4057 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 4058 adjust_zone_range_for_zone_movable(nid, zone_type, 4059 node_start_pfn, node_end_pfn, 4060 &zone_start_pfn, &zone_end_pfn); 4061 4062 /* Check that this node has pages within the zone's required range */ 4063 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 4064 return 0; 4065 4066 /* Move the zone boundaries inside the node if necessary */ 4067 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 4068 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 4069 4070 /* Return the spanned pages */ 4071 return zone_end_pfn - zone_start_pfn; 4072 } 4073 4074 /* 4075 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 4076 * then all holes in the requested range will be accounted for. 4077 */ 4078 unsigned long __meminit __absent_pages_in_range(int nid, 4079 unsigned long range_start_pfn, 4080 unsigned long range_end_pfn) 4081 { 4082 unsigned long nr_absent = range_end_pfn - range_start_pfn; 4083 unsigned long start_pfn, end_pfn; 4084 int i; 4085 4086 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 4087 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 4088 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 4089 nr_absent -= end_pfn - start_pfn; 4090 } 4091 return nr_absent; 4092 } 4093 4094 /** 4095 * absent_pages_in_range - Return number of page frames in holes within a range 4096 * @start_pfn: The start PFN to start searching for holes 4097 * @end_pfn: The end PFN to stop searching for holes 4098 * 4099 * It returns the number of pages frames in memory holes within a range. 4100 */ 4101 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 4102 unsigned long end_pfn) 4103 { 4104 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 4105 } 4106 4107 /* Return the number of page frames in holes in a zone on a node */ 4108 static unsigned long __meminit zone_absent_pages_in_node(int nid, 4109 unsigned long zone_type, 4110 unsigned long *ignored) 4111 { 4112 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 4113 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 4114 unsigned long node_start_pfn, node_end_pfn; 4115 unsigned long zone_start_pfn, zone_end_pfn; 4116 4117 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn); 4118 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 4119 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 4120 4121 adjust_zone_range_for_zone_movable(nid, zone_type, 4122 node_start_pfn, node_end_pfn, 4123 &zone_start_pfn, &zone_end_pfn); 4124 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 4125 } 4126 4127 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4128 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 4129 unsigned long zone_type, 4130 unsigned long *zones_size) 4131 { 4132 return zones_size[zone_type]; 4133 } 4134 4135 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 4136 unsigned long zone_type, 4137 unsigned long *zholes_size) 4138 { 4139 if (!zholes_size) 4140 return 0; 4141 4142 return zholes_size[zone_type]; 4143 } 4144 4145 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4146 4147 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 4148 unsigned long *zones_size, unsigned long *zholes_size) 4149 { 4150 unsigned long realtotalpages, totalpages = 0; 4151 enum zone_type i; 4152 4153 for (i = 0; i < MAX_NR_ZONES; i++) 4154 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i, 4155 zones_size); 4156 pgdat->node_spanned_pages = totalpages; 4157 4158 realtotalpages = totalpages; 4159 for (i = 0; i < MAX_NR_ZONES; i++) 4160 realtotalpages -= 4161 zone_absent_pages_in_node(pgdat->node_id, i, 4162 zholes_size); 4163 pgdat->node_present_pages = realtotalpages; 4164 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 4165 realtotalpages); 4166 } 4167 4168 #ifndef CONFIG_SPARSEMEM 4169 /* 4170 * Calculate the size of the zone->blockflags rounded to an unsigned long 4171 * Start by making sure zonesize is a multiple of pageblock_order by rounding 4172 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 4173 * round what is now in bits to nearest long in bits, then return it in 4174 * bytes. 4175 */ 4176 static unsigned long __init usemap_size(unsigned long zonesize) 4177 { 4178 unsigned long usemapsize; 4179 4180 usemapsize = roundup(zonesize, pageblock_nr_pages); 4181 usemapsize = usemapsize >> pageblock_order; 4182 usemapsize *= NR_PAGEBLOCK_BITS; 4183 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 4184 4185 return usemapsize / 8; 4186 } 4187 4188 static void __init setup_usemap(struct pglist_data *pgdat, 4189 struct zone *zone, unsigned long zonesize) 4190 { 4191 unsigned long usemapsize = usemap_size(zonesize); 4192 zone->pageblock_flags = NULL; 4193 if (usemapsize) 4194 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat, 4195 usemapsize); 4196 } 4197 #else 4198 static inline void setup_usemap(struct pglist_data *pgdat, 4199 struct zone *zone, unsigned long zonesize) {} 4200 #endif /* CONFIG_SPARSEMEM */ 4201 4202 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 4203 4204 /* Return a sensible default order for the pageblock size. */ 4205 static inline int pageblock_default_order(void) 4206 { 4207 if (HPAGE_SHIFT > PAGE_SHIFT) 4208 return HUGETLB_PAGE_ORDER; 4209 4210 return MAX_ORDER-1; 4211 } 4212 4213 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 4214 static inline void __init set_pageblock_order(unsigned int order) 4215 { 4216 /* Check that pageblock_nr_pages has not already been setup */ 4217 if (pageblock_order) 4218 return; 4219 4220 /* 4221 * Assume the largest contiguous order of interest is a huge page. 4222 * This value may be variable depending on boot parameters on IA64 4223 */ 4224 pageblock_order = order; 4225 } 4226 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4227 4228 /* 4229 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 4230 * and pageblock_default_order() are unused as pageblock_order is set 4231 * at compile-time. See include/linux/pageblock-flags.h for the values of 4232 * pageblock_order based on the kernel config 4233 */ 4234 static inline int pageblock_default_order(unsigned int order) 4235 { 4236 return MAX_ORDER-1; 4237 } 4238 #define set_pageblock_order(x) do {} while (0) 4239 4240 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 4241 4242 /* 4243 * Set up the zone data structures: 4244 * - mark all pages reserved 4245 * - mark all memory queues empty 4246 * - clear the memory bitmaps 4247 */ 4248 static void __paginginit free_area_init_core(struct pglist_data *pgdat, 4249 unsigned long *zones_size, unsigned long *zholes_size) 4250 { 4251 enum zone_type j; 4252 int nid = pgdat->node_id; 4253 unsigned long zone_start_pfn = pgdat->node_start_pfn; 4254 int ret; 4255 4256 pgdat_resize_init(pgdat); 4257 pgdat->nr_zones = 0; 4258 init_waitqueue_head(&pgdat->kswapd_wait); 4259 pgdat->kswapd_max_order = 0; 4260 pgdat_page_cgroup_init(pgdat); 4261 4262 for (j = 0; j < MAX_NR_ZONES; j++) { 4263 struct zone *zone = pgdat->node_zones + j; 4264 unsigned long size, realsize, memmap_pages; 4265 enum lru_list lru; 4266 4267 size = zone_spanned_pages_in_node(nid, j, zones_size); 4268 realsize = size - zone_absent_pages_in_node(nid, j, 4269 zholes_size); 4270 4271 /* 4272 * Adjust realsize so that it accounts for how much memory 4273 * is used by this zone for memmap. This affects the watermark 4274 * and per-cpu initialisations 4275 */ 4276 memmap_pages = 4277 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT; 4278 if (realsize >= memmap_pages) { 4279 realsize -= memmap_pages; 4280 if (memmap_pages) 4281 printk(KERN_DEBUG 4282 " %s zone: %lu pages used for memmap\n", 4283 zone_names[j], memmap_pages); 4284 } else 4285 printk(KERN_WARNING 4286 " %s zone: %lu pages exceeds realsize %lu\n", 4287 zone_names[j], memmap_pages, realsize); 4288 4289 /* Account for reserved pages */ 4290 if (j == 0 && realsize > dma_reserve) { 4291 realsize -= dma_reserve; 4292 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 4293 zone_names[0], dma_reserve); 4294 } 4295 4296 if (!is_highmem_idx(j)) 4297 nr_kernel_pages += realsize; 4298 nr_all_pages += realsize; 4299 4300 zone->spanned_pages = size; 4301 zone->present_pages = realsize; 4302 #ifdef CONFIG_NUMA 4303 zone->node = nid; 4304 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio) 4305 / 100; 4306 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100; 4307 #endif 4308 zone->name = zone_names[j]; 4309 spin_lock_init(&zone->lock); 4310 spin_lock_init(&zone->lru_lock); 4311 zone_seqlock_init(zone); 4312 zone->zone_pgdat = pgdat; 4313 4314 zone_pcp_init(zone); 4315 for_each_lru(lru) 4316 INIT_LIST_HEAD(&zone->lruvec.lists[lru]); 4317 zone->reclaim_stat.recent_rotated[0] = 0; 4318 zone->reclaim_stat.recent_rotated[1] = 0; 4319 zone->reclaim_stat.recent_scanned[0] = 0; 4320 zone->reclaim_stat.recent_scanned[1] = 0; 4321 zap_zone_vm_stats(zone); 4322 zone->flags = 0; 4323 if (!size) 4324 continue; 4325 4326 set_pageblock_order(pageblock_default_order()); 4327 setup_usemap(pgdat, zone, size); 4328 ret = init_currently_empty_zone(zone, zone_start_pfn, 4329 size, MEMMAP_EARLY); 4330 BUG_ON(ret); 4331 memmap_init(size, nid, j, zone_start_pfn); 4332 zone_start_pfn += size; 4333 } 4334 } 4335 4336 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 4337 { 4338 /* Skip empty nodes */ 4339 if (!pgdat->node_spanned_pages) 4340 return; 4341 4342 #ifdef CONFIG_FLAT_NODE_MEM_MAP 4343 /* ia64 gets its own node_mem_map, before this, without bootmem */ 4344 if (!pgdat->node_mem_map) { 4345 unsigned long size, start, end; 4346 struct page *map; 4347 4348 /* 4349 * The zone's endpoints aren't required to be MAX_ORDER 4350 * aligned but the node_mem_map endpoints must be in order 4351 * for the buddy allocator to function correctly. 4352 */ 4353 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 4354 end = pgdat->node_start_pfn + pgdat->node_spanned_pages; 4355 end = ALIGN(end, MAX_ORDER_NR_PAGES); 4356 size = (end - start) * sizeof(struct page); 4357 map = alloc_remap(pgdat->node_id, size); 4358 if (!map) 4359 map = alloc_bootmem_node_nopanic(pgdat, size); 4360 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 4361 } 4362 #ifndef CONFIG_NEED_MULTIPLE_NODES 4363 /* 4364 * With no DISCONTIG, the global mem_map is just set as node 0's 4365 */ 4366 if (pgdat == NODE_DATA(0)) { 4367 mem_map = NODE_DATA(0)->node_mem_map; 4368 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4369 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 4370 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 4371 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4372 } 4373 #endif 4374 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 4375 } 4376 4377 void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 4378 unsigned long node_start_pfn, unsigned long *zholes_size) 4379 { 4380 pg_data_t *pgdat = NODE_DATA(nid); 4381 4382 pgdat->node_id = nid; 4383 pgdat->node_start_pfn = node_start_pfn; 4384 calculate_node_totalpages(pgdat, zones_size, zholes_size); 4385 4386 alloc_node_mem_map(pgdat); 4387 #ifdef CONFIG_FLAT_NODE_MEM_MAP 4388 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 4389 nid, (unsigned long)pgdat, 4390 (unsigned long)pgdat->node_mem_map); 4391 #endif 4392 4393 free_area_init_core(pgdat, zones_size, zholes_size); 4394 } 4395 4396 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4397 4398 #if MAX_NUMNODES > 1 4399 /* 4400 * Figure out the number of possible node ids. 4401 */ 4402 static void __init setup_nr_node_ids(void) 4403 { 4404 unsigned int node; 4405 unsigned int highest = 0; 4406 4407 for_each_node_mask(node, node_possible_map) 4408 highest = node; 4409 nr_node_ids = highest + 1; 4410 } 4411 #else 4412 static inline void setup_nr_node_ids(void) 4413 { 4414 } 4415 #endif 4416 4417 /** 4418 * node_map_pfn_alignment - determine the maximum internode alignment 4419 * 4420 * This function should be called after node map is populated and sorted. 4421 * It calculates the maximum power of two alignment which can distinguish 4422 * all the nodes. 4423 * 4424 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 4425 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 4426 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 4427 * shifted, 1GiB is enough and this function will indicate so. 4428 * 4429 * This is used to test whether pfn -> nid mapping of the chosen memory 4430 * model has fine enough granularity to avoid incorrect mapping for the 4431 * populated node map. 4432 * 4433 * Returns the determined alignment in pfn's. 0 if there is no alignment 4434 * requirement (single node). 4435 */ 4436 unsigned long __init node_map_pfn_alignment(void) 4437 { 4438 unsigned long accl_mask = 0, last_end = 0; 4439 unsigned long start, end, mask; 4440 int last_nid = -1; 4441 int i, nid; 4442 4443 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 4444 if (!start || last_nid < 0 || last_nid == nid) { 4445 last_nid = nid; 4446 last_end = end; 4447 continue; 4448 } 4449 4450 /* 4451 * Start with a mask granular enough to pin-point to the 4452 * start pfn and tick off bits one-by-one until it becomes 4453 * too coarse to separate the current node from the last. 4454 */ 4455 mask = ~((1 << __ffs(start)) - 1); 4456 while (mask && last_end <= (start & (mask << 1))) 4457 mask <<= 1; 4458 4459 /* accumulate all internode masks */ 4460 accl_mask |= mask; 4461 } 4462 4463 /* convert mask to number of pages */ 4464 return ~accl_mask + 1; 4465 } 4466 4467 /* Find the lowest pfn for a node */ 4468 static unsigned long __init find_min_pfn_for_node(int nid) 4469 { 4470 unsigned long min_pfn = ULONG_MAX; 4471 unsigned long start_pfn; 4472 int i; 4473 4474 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 4475 min_pfn = min(min_pfn, start_pfn); 4476 4477 if (min_pfn == ULONG_MAX) { 4478 printk(KERN_WARNING 4479 "Could not find start_pfn for node %d\n", nid); 4480 return 0; 4481 } 4482 4483 return min_pfn; 4484 } 4485 4486 /** 4487 * find_min_pfn_with_active_regions - Find the minimum PFN registered 4488 * 4489 * It returns the minimum PFN based on information provided via 4490 * add_active_range(). 4491 */ 4492 unsigned long __init find_min_pfn_with_active_regions(void) 4493 { 4494 return find_min_pfn_for_node(MAX_NUMNODES); 4495 } 4496 4497 /* 4498 * early_calculate_totalpages() 4499 * Sum pages in active regions for movable zone. 4500 * Populate N_HIGH_MEMORY for calculating usable_nodes. 4501 */ 4502 static unsigned long __init early_calculate_totalpages(void) 4503 { 4504 unsigned long totalpages = 0; 4505 unsigned long start_pfn, end_pfn; 4506 int i, nid; 4507 4508 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 4509 unsigned long pages = end_pfn - start_pfn; 4510 4511 totalpages += pages; 4512 if (pages) 4513 node_set_state(nid, N_HIGH_MEMORY); 4514 } 4515 return totalpages; 4516 } 4517 4518 /* 4519 * Find the PFN the Movable zone begins in each node. Kernel memory 4520 * is spread evenly between nodes as long as the nodes have enough 4521 * memory. When they don't, some nodes will have more kernelcore than 4522 * others 4523 */ 4524 static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn) 4525 { 4526 int i, nid; 4527 unsigned long usable_startpfn; 4528 unsigned long kernelcore_node, kernelcore_remaining; 4529 /* save the state before borrow the nodemask */ 4530 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY]; 4531 unsigned long totalpages = early_calculate_totalpages(); 4532 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]); 4533 4534 /* 4535 * If movablecore was specified, calculate what size of 4536 * kernelcore that corresponds so that memory usable for 4537 * any allocation type is evenly spread. If both kernelcore 4538 * and movablecore are specified, then the value of kernelcore 4539 * will be used for required_kernelcore if it's greater than 4540 * what movablecore would have allowed. 4541 */ 4542 if (required_movablecore) { 4543 unsigned long corepages; 4544 4545 /* 4546 * Round-up so that ZONE_MOVABLE is at least as large as what 4547 * was requested by the user 4548 */ 4549 required_movablecore = 4550 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 4551 corepages = totalpages - required_movablecore; 4552 4553 required_kernelcore = max(required_kernelcore, corepages); 4554 } 4555 4556 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 4557 if (!required_kernelcore) 4558 goto out; 4559 4560 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 4561 find_usable_zone_for_movable(); 4562 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 4563 4564 restart: 4565 /* Spread kernelcore memory as evenly as possible throughout nodes */ 4566 kernelcore_node = required_kernelcore / usable_nodes; 4567 for_each_node_state(nid, N_HIGH_MEMORY) { 4568 unsigned long start_pfn, end_pfn; 4569 4570 /* 4571 * Recalculate kernelcore_node if the division per node 4572 * now exceeds what is necessary to satisfy the requested 4573 * amount of memory for the kernel 4574 */ 4575 if (required_kernelcore < kernelcore_node) 4576 kernelcore_node = required_kernelcore / usable_nodes; 4577 4578 /* 4579 * As the map is walked, we track how much memory is usable 4580 * by the kernel using kernelcore_remaining. When it is 4581 * 0, the rest of the node is usable by ZONE_MOVABLE 4582 */ 4583 kernelcore_remaining = kernelcore_node; 4584 4585 /* Go through each range of PFNs within this node */ 4586 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 4587 unsigned long size_pages; 4588 4589 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 4590 if (start_pfn >= end_pfn) 4591 continue; 4592 4593 /* Account for what is only usable for kernelcore */ 4594 if (start_pfn < usable_startpfn) { 4595 unsigned long kernel_pages; 4596 kernel_pages = min(end_pfn, usable_startpfn) 4597 - start_pfn; 4598 4599 kernelcore_remaining -= min(kernel_pages, 4600 kernelcore_remaining); 4601 required_kernelcore -= min(kernel_pages, 4602 required_kernelcore); 4603 4604 /* Continue if range is now fully accounted */ 4605 if (end_pfn <= usable_startpfn) { 4606 4607 /* 4608 * Push zone_movable_pfn to the end so 4609 * that if we have to rebalance 4610 * kernelcore across nodes, we will 4611 * not double account here 4612 */ 4613 zone_movable_pfn[nid] = end_pfn; 4614 continue; 4615 } 4616 start_pfn = usable_startpfn; 4617 } 4618 4619 /* 4620 * The usable PFN range for ZONE_MOVABLE is from 4621 * start_pfn->end_pfn. Calculate size_pages as the 4622 * number of pages used as kernelcore 4623 */ 4624 size_pages = end_pfn - start_pfn; 4625 if (size_pages > kernelcore_remaining) 4626 size_pages = kernelcore_remaining; 4627 zone_movable_pfn[nid] = start_pfn + size_pages; 4628 4629 /* 4630 * Some kernelcore has been met, update counts and 4631 * break if the kernelcore for this node has been 4632 * satisified 4633 */ 4634 required_kernelcore -= min(required_kernelcore, 4635 size_pages); 4636 kernelcore_remaining -= size_pages; 4637 if (!kernelcore_remaining) 4638 break; 4639 } 4640 } 4641 4642 /* 4643 * If there is still required_kernelcore, we do another pass with one 4644 * less node in the count. This will push zone_movable_pfn[nid] further 4645 * along on the nodes that still have memory until kernelcore is 4646 * satisified 4647 */ 4648 usable_nodes--; 4649 if (usable_nodes && required_kernelcore > usable_nodes) 4650 goto restart; 4651 4652 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 4653 for (nid = 0; nid < MAX_NUMNODES; nid++) 4654 zone_movable_pfn[nid] = 4655 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 4656 4657 out: 4658 /* restore the node_state */ 4659 node_states[N_HIGH_MEMORY] = saved_node_state; 4660 } 4661 4662 /* Any regular memory on that node ? */ 4663 static void check_for_regular_memory(pg_data_t *pgdat) 4664 { 4665 #ifdef CONFIG_HIGHMEM 4666 enum zone_type zone_type; 4667 4668 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) { 4669 struct zone *zone = &pgdat->node_zones[zone_type]; 4670 if (zone->present_pages) { 4671 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY); 4672 break; 4673 } 4674 } 4675 #endif 4676 } 4677 4678 /** 4679 * free_area_init_nodes - Initialise all pg_data_t and zone data 4680 * @max_zone_pfn: an array of max PFNs for each zone 4681 * 4682 * This will call free_area_init_node() for each active node in the system. 4683 * Using the page ranges provided by add_active_range(), the size of each 4684 * zone in each node and their holes is calculated. If the maximum PFN 4685 * between two adjacent zones match, it is assumed that the zone is empty. 4686 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 4687 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 4688 * starts where the previous one ended. For example, ZONE_DMA32 starts 4689 * at arch_max_dma_pfn. 4690 */ 4691 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 4692 { 4693 unsigned long start_pfn, end_pfn; 4694 int i, nid; 4695 4696 /* Record where the zone boundaries are */ 4697 memset(arch_zone_lowest_possible_pfn, 0, 4698 sizeof(arch_zone_lowest_possible_pfn)); 4699 memset(arch_zone_highest_possible_pfn, 0, 4700 sizeof(arch_zone_highest_possible_pfn)); 4701 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 4702 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 4703 for (i = 1; i < MAX_NR_ZONES; i++) { 4704 if (i == ZONE_MOVABLE) 4705 continue; 4706 arch_zone_lowest_possible_pfn[i] = 4707 arch_zone_highest_possible_pfn[i-1]; 4708 arch_zone_highest_possible_pfn[i] = 4709 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 4710 } 4711 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 4712 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 4713 4714 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 4715 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 4716 find_zone_movable_pfns_for_nodes(zone_movable_pfn); 4717 4718 /* Print out the zone ranges */ 4719 printk("Zone PFN ranges:\n"); 4720 for (i = 0; i < MAX_NR_ZONES; i++) { 4721 if (i == ZONE_MOVABLE) 4722 continue; 4723 printk(" %-8s ", zone_names[i]); 4724 if (arch_zone_lowest_possible_pfn[i] == 4725 arch_zone_highest_possible_pfn[i]) 4726 printk("empty\n"); 4727 else 4728 printk("%0#10lx -> %0#10lx\n", 4729 arch_zone_lowest_possible_pfn[i], 4730 arch_zone_highest_possible_pfn[i]); 4731 } 4732 4733 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 4734 printk("Movable zone start PFN for each node\n"); 4735 for (i = 0; i < MAX_NUMNODES; i++) { 4736 if (zone_movable_pfn[i]) 4737 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]); 4738 } 4739 4740 /* Print out the early_node_map[] */ 4741 printk("Early memory PFN ranges\n"); 4742 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 4743 printk(" %3d: %0#10lx -> %0#10lx\n", nid, start_pfn, end_pfn); 4744 4745 /* Initialise every node */ 4746 mminit_verify_pageflags_layout(); 4747 setup_nr_node_ids(); 4748 for_each_online_node(nid) { 4749 pg_data_t *pgdat = NODE_DATA(nid); 4750 free_area_init_node(nid, NULL, 4751 find_min_pfn_for_node(nid), NULL); 4752 4753 /* Any memory on that node */ 4754 if (pgdat->node_present_pages) 4755 node_set_state(nid, N_HIGH_MEMORY); 4756 check_for_regular_memory(pgdat); 4757 } 4758 } 4759 4760 static int __init cmdline_parse_core(char *p, unsigned long *core) 4761 { 4762 unsigned long long coremem; 4763 if (!p) 4764 return -EINVAL; 4765 4766 coremem = memparse(p, &p); 4767 *core = coremem >> PAGE_SHIFT; 4768 4769 /* Paranoid check that UL is enough for the coremem value */ 4770 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 4771 4772 return 0; 4773 } 4774 4775 /* 4776 * kernelcore=size sets the amount of memory for use for allocations that 4777 * cannot be reclaimed or migrated. 4778 */ 4779 static int __init cmdline_parse_kernelcore(char *p) 4780 { 4781 return cmdline_parse_core(p, &required_kernelcore); 4782 } 4783 4784 /* 4785 * movablecore=size sets the amount of memory for use for allocations that 4786 * can be reclaimed or migrated. 4787 */ 4788 static int __init cmdline_parse_movablecore(char *p) 4789 { 4790 return cmdline_parse_core(p, &required_movablecore); 4791 } 4792 4793 early_param("kernelcore", cmdline_parse_kernelcore); 4794 early_param("movablecore", cmdline_parse_movablecore); 4795 4796 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 4797 4798 /** 4799 * set_dma_reserve - set the specified number of pages reserved in the first zone 4800 * @new_dma_reserve: The number of pages to mark reserved 4801 * 4802 * The per-cpu batchsize and zone watermarks are determined by present_pages. 4803 * In the DMA zone, a significant percentage may be consumed by kernel image 4804 * and other unfreeable allocations which can skew the watermarks badly. This 4805 * function may optionally be used to account for unfreeable pages in the 4806 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 4807 * smaller per-cpu batchsize. 4808 */ 4809 void __init set_dma_reserve(unsigned long new_dma_reserve) 4810 { 4811 dma_reserve = new_dma_reserve; 4812 } 4813 4814 void __init free_area_init(unsigned long *zones_size) 4815 { 4816 free_area_init_node(0, zones_size, 4817 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 4818 } 4819 4820 static int page_alloc_cpu_notify(struct notifier_block *self, 4821 unsigned long action, void *hcpu) 4822 { 4823 int cpu = (unsigned long)hcpu; 4824 4825 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 4826 drain_pages(cpu); 4827 4828 /* 4829 * Spill the event counters of the dead processor 4830 * into the current processors event counters. 4831 * This artificially elevates the count of the current 4832 * processor. 4833 */ 4834 vm_events_fold_cpu(cpu); 4835 4836 /* 4837 * Zero the differential counters of the dead processor 4838 * so that the vm statistics are consistent. 4839 * 4840 * This is only okay since the processor is dead and cannot 4841 * race with what we are doing. 4842 */ 4843 refresh_cpu_vm_stats(cpu); 4844 } 4845 return NOTIFY_OK; 4846 } 4847 4848 void __init page_alloc_init(void) 4849 { 4850 hotcpu_notifier(page_alloc_cpu_notify, 0); 4851 } 4852 4853 /* 4854 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 4855 * or min_free_kbytes changes. 4856 */ 4857 static void calculate_totalreserve_pages(void) 4858 { 4859 struct pglist_data *pgdat; 4860 unsigned long reserve_pages = 0; 4861 enum zone_type i, j; 4862 4863 for_each_online_pgdat(pgdat) { 4864 for (i = 0; i < MAX_NR_ZONES; i++) { 4865 struct zone *zone = pgdat->node_zones + i; 4866 unsigned long max = 0; 4867 4868 /* Find valid and maximum lowmem_reserve in the zone */ 4869 for (j = i; j < MAX_NR_ZONES; j++) { 4870 if (zone->lowmem_reserve[j] > max) 4871 max = zone->lowmem_reserve[j]; 4872 } 4873 4874 /* we treat the high watermark as reserved pages. */ 4875 max += high_wmark_pages(zone); 4876 4877 if (max > zone->present_pages) 4878 max = zone->present_pages; 4879 reserve_pages += max; 4880 /* 4881 * Lowmem reserves are not available to 4882 * GFP_HIGHUSER page cache allocations and 4883 * kswapd tries to balance zones to their high 4884 * watermark. As a result, neither should be 4885 * regarded as dirtyable memory, to prevent a 4886 * situation where reclaim has to clean pages 4887 * in order to balance the zones. 4888 */ 4889 zone->dirty_balance_reserve = max; 4890 } 4891 } 4892 dirty_balance_reserve = reserve_pages; 4893 totalreserve_pages = reserve_pages; 4894 } 4895 4896 /* 4897 * setup_per_zone_lowmem_reserve - called whenever 4898 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 4899 * has a correct pages reserved value, so an adequate number of 4900 * pages are left in the zone after a successful __alloc_pages(). 4901 */ 4902 static void setup_per_zone_lowmem_reserve(void) 4903 { 4904 struct pglist_data *pgdat; 4905 enum zone_type j, idx; 4906 4907 for_each_online_pgdat(pgdat) { 4908 for (j = 0; j < MAX_NR_ZONES; j++) { 4909 struct zone *zone = pgdat->node_zones + j; 4910 unsigned long present_pages = zone->present_pages; 4911 4912 zone->lowmem_reserve[j] = 0; 4913 4914 idx = j; 4915 while (idx) { 4916 struct zone *lower_zone; 4917 4918 idx--; 4919 4920 if (sysctl_lowmem_reserve_ratio[idx] < 1) 4921 sysctl_lowmem_reserve_ratio[idx] = 1; 4922 4923 lower_zone = pgdat->node_zones + idx; 4924 lower_zone->lowmem_reserve[j] = present_pages / 4925 sysctl_lowmem_reserve_ratio[idx]; 4926 present_pages += lower_zone->present_pages; 4927 } 4928 } 4929 } 4930 4931 /* update totalreserve_pages */ 4932 calculate_totalreserve_pages(); 4933 } 4934 4935 /** 4936 * setup_per_zone_wmarks - called when min_free_kbytes changes 4937 * or when memory is hot-{added|removed} 4938 * 4939 * Ensures that the watermark[min,low,high] values for each zone are set 4940 * correctly with respect to min_free_kbytes. 4941 */ 4942 void setup_per_zone_wmarks(void) 4943 { 4944 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 4945 unsigned long lowmem_pages = 0; 4946 struct zone *zone; 4947 unsigned long flags; 4948 4949 /* Calculate total number of !ZONE_HIGHMEM pages */ 4950 for_each_zone(zone) { 4951 if (!is_highmem(zone)) 4952 lowmem_pages += zone->present_pages; 4953 } 4954 4955 for_each_zone(zone) { 4956 u64 tmp; 4957 4958 spin_lock_irqsave(&zone->lock, flags); 4959 tmp = (u64)pages_min * zone->present_pages; 4960 do_div(tmp, lowmem_pages); 4961 if (is_highmem(zone)) { 4962 /* 4963 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 4964 * need highmem pages, so cap pages_min to a small 4965 * value here. 4966 * 4967 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 4968 * deltas controls asynch page reclaim, and so should 4969 * not be capped for highmem. 4970 */ 4971 int min_pages; 4972 4973 min_pages = zone->present_pages / 1024; 4974 if (min_pages < SWAP_CLUSTER_MAX) 4975 min_pages = SWAP_CLUSTER_MAX; 4976 if (min_pages > 128) 4977 min_pages = 128; 4978 zone->watermark[WMARK_MIN] = min_pages; 4979 } else { 4980 /* 4981 * If it's a lowmem zone, reserve a number of pages 4982 * proportionate to the zone's size. 4983 */ 4984 zone->watermark[WMARK_MIN] = tmp; 4985 } 4986 4987 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 4988 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 4989 setup_zone_migrate_reserve(zone); 4990 spin_unlock_irqrestore(&zone->lock, flags); 4991 } 4992 4993 /* update totalreserve_pages */ 4994 calculate_totalreserve_pages(); 4995 } 4996 4997 /* 4998 * The inactive anon list should be small enough that the VM never has to 4999 * do too much work, but large enough that each inactive page has a chance 5000 * to be referenced again before it is swapped out. 5001 * 5002 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 5003 * INACTIVE_ANON pages on this zone's LRU, maintained by the 5004 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 5005 * the anonymous pages are kept on the inactive list. 5006 * 5007 * total target max 5008 * memory ratio inactive anon 5009 * ------------------------------------- 5010 * 10MB 1 5MB 5011 * 100MB 1 50MB 5012 * 1GB 3 250MB 5013 * 10GB 10 0.9GB 5014 * 100GB 31 3GB 5015 * 1TB 101 10GB 5016 * 10TB 320 32GB 5017 */ 5018 static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 5019 { 5020 unsigned int gb, ratio; 5021 5022 /* Zone size in gigabytes */ 5023 gb = zone->present_pages >> (30 - PAGE_SHIFT); 5024 if (gb) 5025 ratio = int_sqrt(10 * gb); 5026 else 5027 ratio = 1; 5028 5029 zone->inactive_ratio = ratio; 5030 } 5031 5032 static void __meminit setup_per_zone_inactive_ratio(void) 5033 { 5034 struct zone *zone; 5035 5036 for_each_zone(zone) 5037 calculate_zone_inactive_ratio(zone); 5038 } 5039 5040 /* 5041 * Initialise min_free_kbytes. 5042 * 5043 * For small machines we want it small (128k min). For large machines 5044 * we want it large (64MB max). But it is not linear, because network 5045 * bandwidth does not increase linearly with machine size. We use 5046 * 5047 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 5048 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 5049 * 5050 * which yields 5051 * 5052 * 16MB: 512k 5053 * 32MB: 724k 5054 * 64MB: 1024k 5055 * 128MB: 1448k 5056 * 256MB: 2048k 5057 * 512MB: 2896k 5058 * 1024MB: 4096k 5059 * 2048MB: 5792k 5060 * 4096MB: 8192k 5061 * 8192MB: 11584k 5062 * 16384MB: 16384k 5063 */ 5064 int __meminit init_per_zone_wmark_min(void) 5065 { 5066 unsigned long lowmem_kbytes; 5067 5068 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 5069 5070 min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 5071 if (min_free_kbytes < 128) 5072 min_free_kbytes = 128; 5073 if (min_free_kbytes > 65536) 5074 min_free_kbytes = 65536; 5075 setup_per_zone_wmarks(); 5076 refresh_zone_stat_thresholds(); 5077 setup_per_zone_lowmem_reserve(); 5078 setup_per_zone_inactive_ratio(); 5079 return 0; 5080 } 5081 module_init(init_per_zone_wmark_min) 5082 5083 /* 5084 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 5085 * that we can call two helper functions whenever min_free_kbytes 5086 * changes. 5087 */ 5088 int min_free_kbytes_sysctl_handler(ctl_table *table, int write, 5089 void __user *buffer, size_t *length, loff_t *ppos) 5090 { 5091 proc_dointvec(table, write, buffer, length, ppos); 5092 if (write) 5093 setup_per_zone_wmarks(); 5094 return 0; 5095 } 5096 5097 #ifdef CONFIG_NUMA 5098 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write, 5099 void __user *buffer, size_t *length, loff_t *ppos) 5100 { 5101 struct zone *zone; 5102 int rc; 5103 5104 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5105 if (rc) 5106 return rc; 5107 5108 for_each_zone(zone) 5109 zone->min_unmapped_pages = (zone->present_pages * 5110 sysctl_min_unmapped_ratio) / 100; 5111 return 0; 5112 } 5113 5114 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write, 5115 void __user *buffer, size_t *length, loff_t *ppos) 5116 { 5117 struct zone *zone; 5118 int rc; 5119 5120 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 5121 if (rc) 5122 return rc; 5123 5124 for_each_zone(zone) 5125 zone->min_slab_pages = (zone->present_pages * 5126 sysctl_min_slab_ratio) / 100; 5127 return 0; 5128 } 5129 #endif 5130 5131 /* 5132 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 5133 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 5134 * whenever sysctl_lowmem_reserve_ratio changes. 5135 * 5136 * The reserve ratio obviously has absolutely no relation with the 5137 * minimum watermarks. The lowmem reserve ratio can only make sense 5138 * if in function of the boot time zone sizes. 5139 */ 5140 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write, 5141 void __user *buffer, size_t *length, loff_t *ppos) 5142 { 5143 proc_dointvec_minmax(table, write, buffer, length, ppos); 5144 setup_per_zone_lowmem_reserve(); 5145 return 0; 5146 } 5147 5148 /* 5149 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 5150 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist 5151 * can have before it gets flushed back to buddy allocator. 5152 */ 5153 5154 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write, 5155 void __user *buffer, size_t *length, loff_t *ppos) 5156 { 5157 struct zone *zone; 5158 unsigned int cpu; 5159 int ret; 5160 5161 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 5162 if (!write || (ret == -EINVAL)) 5163 return ret; 5164 for_each_populated_zone(zone) { 5165 for_each_possible_cpu(cpu) { 5166 unsigned long high; 5167 high = zone->present_pages / percpu_pagelist_fraction; 5168 setup_pagelist_highmark( 5169 per_cpu_ptr(zone->pageset, cpu), high); 5170 } 5171 } 5172 return 0; 5173 } 5174 5175 int hashdist = HASHDIST_DEFAULT; 5176 5177 #ifdef CONFIG_NUMA 5178 static int __init set_hashdist(char *str) 5179 { 5180 if (!str) 5181 return 0; 5182 hashdist = simple_strtoul(str, &str, 0); 5183 return 1; 5184 } 5185 __setup("hashdist=", set_hashdist); 5186 #endif 5187 5188 /* 5189 * allocate a large system hash table from bootmem 5190 * - it is assumed that the hash table must contain an exact power-of-2 5191 * quantity of entries 5192 * - limit is the number of hash buckets, not the total allocation size 5193 */ 5194 void *__init alloc_large_system_hash(const char *tablename, 5195 unsigned long bucketsize, 5196 unsigned long numentries, 5197 int scale, 5198 int flags, 5199 unsigned int *_hash_shift, 5200 unsigned int *_hash_mask, 5201 unsigned long limit) 5202 { 5203 unsigned long long max = limit; 5204 unsigned long log2qty, size; 5205 void *table = NULL; 5206 5207 /* allow the kernel cmdline to have a say */ 5208 if (!numentries) { 5209 /* round applicable memory size up to nearest megabyte */ 5210 numentries = nr_kernel_pages; 5211 numentries += (1UL << (20 - PAGE_SHIFT)) - 1; 5212 numentries >>= 20 - PAGE_SHIFT; 5213 numentries <<= 20 - PAGE_SHIFT; 5214 5215 /* limit to 1 bucket per 2^scale bytes of low memory */ 5216 if (scale > PAGE_SHIFT) 5217 numentries >>= (scale - PAGE_SHIFT); 5218 else 5219 numentries <<= (PAGE_SHIFT - scale); 5220 5221 /* Make sure we've got at least a 0-order allocation.. */ 5222 if (unlikely(flags & HASH_SMALL)) { 5223 /* Makes no sense without HASH_EARLY */ 5224 WARN_ON(!(flags & HASH_EARLY)); 5225 if (!(numentries >> *_hash_shift)) { 5226 numentries = 1UL << *_hash_shift; 5227 BUG_ON(!numentries); 5228 } 5229 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 5230 numentries = PAGE_SIZE / bucketsize; 5231 } 5232 numentries = roundup_pow_of_two(numentries); 5233 5234 /* limit allocation size to 1/16 total memory by default */ 5235 if (max == 0) { 5236 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 5237 do_div(max, bucketsize); 5238 } 5239 max = min(max, 0x80000000ULL); 5240 5241 if (numentries > max) 5242 numentries = max; 5243 5244 log2qty = ilog2(numentries); 5245 5246 do { 5247 size = bucketsize << log2qty; 5248 if (flags & HASH_EARLY) 5249 table = alloc_bootmem_nopanic(size); 5250 else if (hashdist) 5251 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 5252 else { 5253 /* 5254 * If bucketsize is not a power-of-two, we may free 5255 * some pages at the end of hash table which 5256 * alloc_pages_exact() automatically does 5257 */ 5258 if (get_order(size) < MAX_ORDER) { 5259 table = alloc_pages_exact(size, GFP_ATOMIC); 5260 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 5261 } 5262 } 5263 } while (!table && size > PAGE_SIZE && --log2qty); 5264 5265 if (!table) 5266 panic("Failed to allocate %s hash table\n", tablename); 5267 5268 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 5269 tablename, 5270 (1UL << log2qty), 5271 ilog2(size) - PAGE_SHIFT, 5272 size); 5273 5274 if (_hash_shift) 5275 *_hash_shift = log2qty; 5276 if (_hash_mask) 5277 *_hash_mask = (1 << log2qty) - 1; 5278 5279 return table; 5280 } 5281 5282 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 5283 static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 5284 unsigned long pfn) 5285 { 5286 #ifdef CONFIG_SPARSEMEM 5287 return __pfn_to_section(pfn)->pageblock_flags; 5288 #else 5289 return zone->pageblock_flags; 5290 #endif /* CONFIG_SPARSEMEM */ 5291 } 5292 5293 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 5294 { 5295 #ifdef CONFIG_SPARSEMEM 5296 pfn &= (PAGES_PER_SECTION-1); 5297 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5298 #else 5299 pfn = pfn - zone->zone_start_pfn; 5300 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 5301 #endif /* CONFIG_SPARSEMEM */ 5302 } 5303 5304 /** 5305 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages 5306 * @page: The page within the block of interest 5307 * @start_bitidx: The first bit of interest to retrieve 5308 * @end_bitidx: The last bit of interest 5309 * returns pageblock_bits flags 5310 */ 5311 unsigned long get_pageblock_flags_group(struct page *page, 5312 int start_bitidx, int end_bitidx) 5313 { 5314 struct zone *zone; 5315 unsigned long *bitmap; 5316 unsigned long pfn, bitidx; 5317 unsigned long flags = 0; 5318 unsigned long value = 1; 5319 5320 zone = page_zone(page); 5321 pfn = page_to_pfn(page); 5322 bitmap = get_pageblock_bitmap(zone, pfn); 5323 bitidx = pfn_to_bitidx(zone, pfn); 5324 5325 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5326 if (test_bit(bitidx + start_bitidx, bitmap)) 5327 flags |= value; 5328 5329 return flags; 5330 } 5331 5332 /** 5333 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages 5334 * @page: The page within the block of interest 5335 * @start_bitidx: The first bit of interest 5336 * @end_bitidx: The last bit of interest 5337 * @flags: The flags to set 5338 */ 5339 void set_pageblock_flags_group(struct page *page, unsigned long flags, 5340 int start_bitidx, int end_bitidx) 5341 { 5342 struct zone *zone; 5343 unsigned long *bitmap; 5344 unsigned long pfn, bitidx; 5345 unsigned long value = 1; 5346 5347 zone = page_zone(page); 5348 pfn = page_to_pfn(page); 5349 bitmap = get_pageblock_bitmap(zone, pfn); 5350 bitidx = pfn_to_bitidx(zone, pfn); 5351 VM_BUG_ON(pfn < zone->zone_start_pfn); 5352 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages); 5353 5354 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1) 5355 if (flags & value) 5356 __set_bit(bitidx + start_bitidx, bitmap); 5357 else 5358 __clear_bit(bitidx + start_bitidx, bitmap); 5359 } 5360 5361 /* 5362 * This is designed as sub function...plz see page_isolation.c also. 5363 * set/clear page block's type to be ISOLATE. 5364 * page allocater never alloc memory from ISOLATE block. 5365 */ 5366 5367 static int 5368 __count_immobile_pages(struct zone *zone, struct page *page, int count) 5369 { 5370 unsigned long pfn, iter, found; 5371 /* 5372 * For avoiding noise data, lru_add_drain_all() should be called 5373 * If ZONE_MOVABLE, the zone never contains immobile pages 5374 */ 5375 if (zone_idx(zone) == ZONE_MOVABLE) 5376 return true; 5377 5378 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE) 5379 return true; 5380 5381 pfn = page_to_pfn(page); 5382 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 5383 unsigned long check = pfn + iter; 5384 5385 if (!pfn_valid_within(check)) 5386 continue; 5387 5388 page = pfn_to_page(check); 5389 if (!page_count(page)) { 5390 if (PageBuddy(page)) 5391 iter += (1 << page_order(page)) - 1; 5392 continue; 5393 } 5394 if (!PageLRU(page)) 5395 found++; 5396 /* 5397 * If there are RECLAIMABLE pages, we need to check it. 5398 * But now, memory offline itself doesn't call shrink_slab() 5399 * and it still to be fixed. 5400 */ 5401 /* 5402 * If the page is not RAM, page_count()should be 0. 5403 * we don't need more check. This is an _used_ not-movable page. 5404 * 5405 * The problematic thing here is PG_reserved pages. PG_reserved 5406 * is set to both of a memory hole page and a _used_ kernel 5407 * page at boot. 5408 */ 5409 if (found > count) 5410 return false; 5411 } 5412 return true; 5413 } 5414 5415 bool is_pageblock_removable_nolock(struct page *page) 5416 { 5417 struct zone *zone; 5418 unsigned long pfn; 5419 5420 /* 5421 * We have to be careful here because we are iterating over memory 5422 * sections which are not zone aware so we might end up outside of 5423 * the zone but still within the section. 5424 * We have to take care about the node as well. If the node is offline 5425 * its NODE_DATA will be NULL - see page_zone. 5426 */ 5427 if (!node_online(page_to_nid(page))) 5428 return false; 5429 5430 zone = page_zone(page); 5431 pfn = page_to_pfn(page); 5432 if (zone->zone_start_pfn > pfn || 5433 zone->zone_start_pfn + zone->spanned_pages <= pfn) 5434 return false; 5435 5436 return __count_immobile_pages(zone, page, 0); 5437 } 5438 5439 int set_migratetype_isolate(struct page *page) 5440 { 5441 struct zone *zone; 5442 unsigned long flags, pfn; 5443 struct memory_isolate_notify arg; 5444 int notifier_ret; 5445 int ret = -EBUSY; 5446 5447 zone = page_zone(page); 5448 5449 spin_lock_irqsave(&zone->lock, flags); 5450 5451 pfn = page_to_pfn(page); 5452 arg.start_pfn = pfn; 5453 arg.nr_pages = pageblock_nr_pages; 5454 arg.pages_found = 0; 5455 5456 /* 5457 * It may be possible to isolate a pageblock even if the 5458 * migratetype is not MIGRATE_MOVABLE. The memory isolation 5459 * notifier chain is used by balloon drivers to return the 5460 * number of pages in a range that are held by the balloon 5461 * driver to shrink memory. If all the pages are accounted for 5462 * by balloons, are free, or on the LRU, isolation can continue. 5463 * Later, for example, when memory hotplug notifier runs, these 5464 * pages reported as "can be isolated" should be isolated(freed) 5465 * by the balloon driver through the memory notifier chain. 5466 */ 5467 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg); 5468 notifier_ret = notifier_to_errno(notifier_ret); 5469 if (notifier_ret) 5470 goto out; 5471 /* 5472 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. 5473 * We just check MOVABLE pages. 5474 */ 5475 if (__count_immobile_pages(zone, page, arg.pages_found)) 5476 ret = 0; 5477 5478 /* 5479 * immobile means "not-on-lru" paes. If immobile is larger than 5480 * removable-by-driver pages reported by notifier, we'll fail. 5481 */ 5482 5483 out: 5484 if (!ret) { 5485 set_pageblock_migratetype(page, MIGRATE_ISOLATE); 5486 move_freepages_block(zone, page, MIGRATE_ISOLATE); 5487 } 5488 5489 spin_unlock_irqrestore(&zone->lock, flags); 5490 if (!ret) 5491 drain_all_pages(); 5492 return ret; 5493 } 5494 5495 void unset_migratetype_isolate(struct page *page) 5496 { 5497 struct zone *zone; 5498 unsigned long flags; 5499 zone = page_zone(page); 5500 spin_lock_irqsave(&zone->lock, flags); 5501 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE) 5502 goto out; 5503 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5504 move_freepages_block(zone, page, MIGRATE_MOVABLE); 5505 out: 5506 spin_unlock_irqrestore(&zone->lock, flags); 5507 } 5508 5509 #ifdef CONFIG_MEMORY_HOTREMOVE 5510 /* 5511 * All pages in the range must be isolated before calling this. 5512 */ 5513 void 5514 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 5515 { 5516 struct page *page; 5517 struct zone *zone; 5518 int order, i; 5519 unsigned long pfn; 5520 unsigned long flags; 5521 /* find the first valid pfn */ 5522 for (pfn = start_pfn; pfn < end_pfn; pfn++) 5523 if (pfn_valid(pfn)) 5524 break; 5525 if (pfn == end_pfn) 5526 return; 5527 zone = page_zone(pfn_to_page(pfn)); 5528 spin_lock_irqsave(&zone->lock, flags); 5529 pfn = start_pfn; 5530 while (pfn < end_pfn) { 5531 if (!pfn_valid(pfn)) { 5532 pfn++; 5533 continue; 5534 } 5535 page = pfn_to_page(pfn); 5536 BUG_ON(page_count(page)); 5537 BUG_ON(!PageBuddy(page)); 5538 order = page_order(page); 5539 #ifdef CONFIG_DEBUG_VM 5540 printk(KERN_INFO "remove from free list %lx %d %lx\n", 5541 pfn, 1 << order, end_pfn); 5542 #endif 5543 list_del(&page->lru); 5544 rmv_page_order(page); 5545 zone->free_area[order].nr_free--; 5546 __mod_zone_page_state(zone, NR_FREE_PAGES, 5547 - (1UL << order)); 5548 for (i = 0; i < (1 << order); i++) 5549 SetPageReserved((page+i)); 5550 pfn += (1 << order); 5551 } 5552 spin_unlock_irqrestore(&zone->lock, flags); 5553 } 5554 #endif 5555 5556 #ifdef CONFIG_MEMORY_FAILURE 5557 bool is_free_buddy_page(struct page *page) 5558 { 5559 struct zone *zone = page_zone(page); 5560 unsigned long pfn = page_to_pfn(page); 5561 unsigned long flags; 5562 int order; 5563 5564 spin_lock_irqsave(&zone->lock, flags); 5565 for (order = 0; order < MAX_ORDER; order++) { 5566 struct page *page_head = page - (pfn & ((1 << order) - 1)); 5567 5568 if (PageBuddy(page_head) && page_order(page_head) >= order) 5569 break; 5570 } 5571 spin_unlock_irqrestore(&zone->lock, flags); 5572 5573 return order < MAX_ORDER; 5574 } 5575 #endif 5576 5577 static struct trace_print_flags pageflag_names[] = { 5578 {1UL << PG_locked, "locked" }, 5579 {1UL << PG_error, "error" }, 5580 {1UL << PG_referenced, "referenced" }, 5581 {1UL << PG_uptodate, "uptodate" }, 5582 {1UL << PG_dirty, "dirty" }, 5583 {1UL << PG_lru, "lru" }, 5584 {1UL << PG_active, "active" }, 5585 {1UL << PG_slab, "slab" }, 5586 {1UL << PG_owner_priv_1, "owner_priv_1" }, 5587 {1UL << PG_arch_1, "arch_1" }, 5588 {1UL << PG_reserved, "reserved" }, 5589 {1UL << PG_private, "private" }, 5590 {1UL << PG_private_2, "private_2" }, 5591 {1UL << PG_writeback, "writeback" }, 5592 #ifdef CONFIG_PAGEFLAGS_EXTENDED 5593 {1UL << PG_head, "head" }, 5594 {1UL << PG_tail, "tail" }, 5595 #else 5596 {1UL << PG_compound, "compound" }, 5597 #endif 5598 {1UL << PG_swapcache, "swapcache" }, 5599 {1UL << PG_mappedtodisk, "mappedtodisk" }, 5600 {1UL << PG_reclaim, "reclaim" }, 5601 {1UL << PG_swapbacked, "swapbacked" }, 5602 {1UL << PG_unevictable, "unevictable" }, 5603 #ifdef CONFIG_MMU 5604 {1UL << PG_mlocked, "mlocked" }, 5605 #endif 5606 #ifdef CONFIG_ARCH_USES_PG_UNCACHED 5607 {1UL << PG_uncached, "uncached" }, 5608 #endif 5609 #ifdef CONFIG_MEMORY_FAILURE 5610 {1UL << PG_hwpoison, "hwpoison" }, 5611 #endif 5612 {-1UL, NULL }, 5613 }; 5614 5615 static void dump_page_flags(unsigned long flags) 5616 { 5617 const char *delim = ""; 5618 unsigned long mask; 5619 int i; 5620 5621 printk(KERN_ALERT "page flags: %#lx(", flags); 5622 5623 /* remove zone id */ 5624 flags &= (1UL << NR_PAGEFLAGS) - 1; 5625 5626 for (i = 0; pageflag_names[i].name && flags; i++) { 5627 5628 mask = pageflag_names[i].mask; 5629 if ((flags & mask) != mask) 5630 continue; 5631 5632 flags &= ~mask; 5633 printk("%s%s", delim, pageflag_names[i].name); 5634 delim = "|"; 5635 } 5636 5637 /* check for left over flags */ 5638 if (flags) 5639 printk("%s%#lx", delim, flags); 5640 5641 printk(")\n"); 5642 } 5643 5644 void dump_page(struct page *page) 5645 { 5646 printk(KERN_ALERT 5647 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n", 5648 page, atomic_read(&page->_count), page_mapcount(page), 5649 page->mapping, page->index); 5650 dump_page_flags(page->flags); 5651 mem_cgroup_print_bad_page(page); 5652 } 5653