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/rwsem.h> 22 #include <linux/pagemap.h> 23 #include <linux/jiffies.h> 24 #include <linux/bootmem.h> 25 #include <linux/memblock.h> 26 #include <linux/compiler.h> 27 #include <linux/kernel.h> 28 #include <linux/kmemcheck.h> 29 #include <linux/kasan.h> 30 #include <linux/module.h> 31 #include <linux/suspend.h> 32 #include <linux/pagevec.h> 33 #include <linux/blkdev.h> 34 #include <linux/slab.h> 35 #include <linux/ratelimit.h> 36 #include <linux/oom.h> 37 #include <linux/notifier.h> 38 #include <linux/topology.h> 39 #include <linux/sysctl.h> 40 #include <linux/cpu.h> 41 #include <linux/cpuset.h> 42 #include <linux/memory_hotplug.h> 43 #include <linux/nodemask.h> 44 #include <linux/vmalloc.h> 45 #include <linux/vmstat.h> 46 #include <linux/mempolicy.h> 47 #include <linux/stop_machine.h> 48 #include <linux/sort.h> 49 #include <linux/pfn.h> 50 #include <linux/backing-dev.h> 51 #include <linux/fault-inject.h> 52 #include <linux/page-isolation.h> 53 #include <linux/page_ext.h> 54 #include <linux/debugobjects.h> 55 #include <linux/kmemleak.h> 56 #include <linux/compaction.h> 57 #include <trace/events/kmem.h> 58 #include <linux/prefetch.h> 59 #include <linux/mm_inline.h> 60 #include <linux/migrate.h> 61 #include <linux/page_ext.h> 62 #include <linux/hugetlb.h> 63 #include <linux/sched/rt.h> 64 #include <linux/page_owner.h> 65 #include <linux/kthread.h> 66 67 #include <asm/sections.h> 68 #include <asm/tlbflush.h> 69 #include <asm/div64.h> 70 #include "internal.h" 71 72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ 73 static DEFINE_MUTEX(pcp_batch_high_lock); 74 #define MIN_PERCPU_PAGELIST_FRACTION (8) 75 76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 77 DEFINE_PER_CPU(int, numa_node); 78 EXPORT_PER_CPU_SYMBOL(numa_node); 79 #endif 80 81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 82 /* 83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 86 * defined in <linux/topology.h>. 87 */ 88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 89 EXPORT_PER_CPU_SYMBOL(_numa_mem_); 90 int _node_numa_mem_[MAX_NUMNODES]; 91 #endif 92 93 /* 94 * Array of node states. 95 */ 96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 97 [N_POSSIBLE] = NODE_MASK_ALL, 98 [N_ONLINE] = { { [0] = 1UL } }, 99 #ifndef CONFIG_NUMA 100 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 101 #ifdef CONFIG_HIGHMEM 102 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 103 #endif 104 #ifdef CONFIG_MOVABLE_NODE 105 [N_MEMORY] = { { [0] = 1UL } }, 106 #endif 107 [N_CPU] = { { [0] = 1UL } }, 108 #endif /* NUMA */ 109 }; 110 EXPORT_SYMBOL(node_states); 111 112 /* Protect totalram_pages and zone->managed_pages */ 113 static DEFINE_SPINLOCK(managed_page_count_lock); 114 115 unsigned long totalram_pages __read_mostly; 116 unsigned long totalreserve_pages __read_mostly; 117 unsigned long totalcma_pages __read_mostly; 118 /* 119 * When calculating the number of globally allowed dirty pages, there 120 * is a certain number of per-zone reserves that should not be 121 * considered dirtyable memory. This is the sum of those reserves 122 * over all existing zones that contribute dirtyable memory. 123 */ 124 unsigned long dirty_balance_reserve __read_mostly; 125 126 int percpu_pagelist_fraction; 127 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 128 129 #ifdef CONFIG_PM_SLEEP 130 /* 131 * The following functions are used by the suspend/hibernate code to temporarily 132 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 133 * while devices are suspended. To avoid races with the suspend/hibernate code, 134 * they should always be called with pm_mutex held (gfp_allowed_mask also should 135 * only be modified with pm_mutex held, unless the suspend/hibernate code is 136 * guaranteed not to run in parallel with that modification). 137 */ 138 139 static gfp_t saved_gfp_mask; 140 141 void pm_restore_gfp_mask(void) 142 { 143 WARN_ON(!mutex_is_locked(&pm_mutex)); 144 if (saved_gfp_mask) { 145 gfp_allowed_mask = saved_gfp_mask; 146 saved_gfp_mask = 0; 147 } 148 } 149 150 void pm_restrict_gfp_mask(void) 151 { 152 WARN_ON(!mutex_is_locked(&pm_mutex)); 153 WARN_ON(saved_gfp_mask); 154 saved_gfp_mask = gfp_allowed_mask; 155 gfp_allowed_mask &= ~GFP_IOFS; 156 } 157 158 bool pm_suspended_storage(void) 159 { 160 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS) 161 return false; 162 return true; 163 } 164 #endif /* CONFIG_PM_SLEEP */ 165 166 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 167 int pageblock_order __read_mostly; 168 #endif 169 170 static void __free_pages_ok(struct page *page, unsigned int order); 171 172 /* 173 * results with 256, 32 in the lowmem_reserve sysctl: 174 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 175 * 1G machine -> (16M dma, 784M normal, 224M high) 176 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 177 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 178 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA 179 * 180 * TBD: should special case ZONE_DMA32 machines here - in those we normally 181 * don't need any ZONE_NORMAL reservation 182 */ 183 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 184 #ifdef CONFIG_ZONE_DMA 185 256, 186 #endif 187 #ifdef CONFIG_ZONE_DMA32 188 256, 189 #endif 190 #ifdef CONFIG_HIGHMEM 191 32, 192 #endif 193 32, 194 }; 195 196 EXPORT_SYMBOL(totalram_pages); 197 198 static char * const zone_names[MAX_NR_ZONES] = { 199 #ifdef CONFIG_ZONE_DMA 200 "DMA", 201 #endif 202 #ifdef CONFIG_ZONE_DMA32 203 "DMA32", 204 #endif 205 "Normal", 206 #ifdef CONFIG_HIGHMEM 207 "HighMem", 208 #endif 209 "Movable", 210 }; 211 212 int min_free_kbytes = 1024; 213 int user_min_free_kbytes = -1; 214 215 static unsigned long __meminitdata nr_kernel_pages; 216 static unsigned long __meminitdata nr_all_pages; 217 static unsigned long __meminitdata dma_reserve; 218 219 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 220 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 221 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 222 static unsigned long __initdata required_kernelcore; 223 static unsigned long __initdata required_movablecore; 224 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 225 226 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 227 int movable_zone; 228 EXPORT_SYMBOL(movable_zone); 229 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 230 231 #if MAX_NUMNODES > 1 232 int nr_node_ids __read_mostly = MAX_NUMNODES; 233 int nr_online_nodes __read_mostly = 1; 234 EXPORT_SYMBOL(nr_node_ids); 235 EXPORT_SYMBOL(nr_online_nodes); 236 #endif 237 238 int page_group_by_mobility_disabled __read_mostly; 239 240 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 241 static inline void reset_deferred_meminit(pg_data_t *pgdat) 242 { 243 pgdat->first_deferred_pfn = ULONG_MAX; 244 } 245 246 /* Returns true if the struct page for the pfn is uninitialised */ 247 static inline bool __meminit early_page_uninitialised(unsigned long pfn) 248 { 249 int nid = early_pfn_to_nid(pfn); 250 251 if (pfn >= NODE_DATA(nid)->first_deferred_pfn) 252 return true; 253 254 return false; 255 } 256 257 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid) 258 { 259 if (pfn >= NODE_DATA(nid)->first_deferred_pfn) 260 return true; 261 262 return false; 263 } 264 265 /* 266 * Returns false when the remaining initialisation should be deferred until 267 * later in the boot cycle when it can be parallelised. 268 */ 269 static inline bool update_defer_init(pg_data_t *pgdat, 270 unsigned long pfn, unsigned long zone_end, 271 unsigned long *nr_initialised) 272 { 273 /* Always populate low zones for address-contrained allocations */ 274 if (zone_end < pgdat_end_pfn(pgdat)) 275 return true; 276 277 /* Initialise at least 2G of the highest zone */ 278 (*nr_initialised)++; 279 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) && 280 (pfn & (PAGES_PER_SECTION - 1)) == 0) { 281 pgdat->first_deferred_pfn = pfn; 282 return false; 283 } 284 285 return true; 286 } 287 #else 288 static inline void reset_deferred_meminit(pg_data_t *pgdat) 289 { 290 } 291 292 static inline bool early_page_uninitialised(unsigned long pfn) 293 { 294 return false; 295 } 296 297 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid) 298 { 299 return false; 300 } 301 302 static inline bool update_defer_init(pg_data_t *pgdat, 303 unsigned long pfn, unsigned long zone_end, 304 unsigned long *nr_initialised) 305 { 306 return true; 307 } 308 #endif 309 310 311 void set_pageblock_migratetype(struct page *page, int migratetype) 312 { 313 if (unlikely(page_group_by_mobility_disabled && 314 migratetype < MIGRATE_PCPTYPES)) 315 migratetype = MIGRATE_UNMOVABLE; 316 317 set_pageblock_flags_group(page, (unsigned long)migratetype, 318 PB_migrate, PB_migrate_end); 319 } 320 321 #ifdef CONFIG_DEBUG_VM 322 static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 323 { 324 int ret = 0; 325 unsigned seq; 326 unsigned long pfn = page_to_pfn(page); 327 unsigned long sp, start_pfn; 328 329 do { 330 seq = zone_span_seqbegin(zone); 331 start_pfn = zone->zone_start_pfn; 332 sp = zone->spanned_pages; 333 if (!zone_spans_pfn(zone, pfn)) 334 ret = 1; 335 } while (zone_span_seqretry(zone, seq)); 336 337 if (ret) 338 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", 339 pfn, zone_to_nid(zone), zone->name, 340 start_pfn, start_pfn + sp); 341 342 return ret; 343 } 344 345 static int page_is_consistent(struct zone *zone, struct page *page) 346 { 347 if (!pfn_valid_within(page_to_pfn(page))) 348 return 0; 349 if (zone != page_zone(page)) 350 return 0; 351 352 return 1; 353 } 354 /* 355 * Temporary debugging check for pages not lying within a given zone. 356 */ 357 static int bad_range(struct zone *zone, struct page *page) 358 { 359 if (page_outside_zone_boundaries(zone, page)) 360 return 1; 361 if (!page_is_consistent(zone, page)) 362 return 1; 363 364 return 0; 365 } 366 #else 367 static inline int bad_range(struct zone *zone, struct page *page) 368 { 369 return 0; 370 } 371 #endif 372 373 static void bad_page(struct page *page, const char *reason, 374 unsigned long bad_flags) 375 { 376 static unsigned long resume; 377 static unsigned long nr_shown; 378 static unsigned long nr_unshown; 379 380 /* Don't complain about poisoned pages */ 381 if (PageHWPoison(page)) { 382 page_mapcount_reset(page); /* remove PageBuddy */ 383 return; 384 } 385 386 /* 387 * Allow a burst of 60 reports, then keep quiet for that minute; 388 * or allow a steady drip of one report per second. 389 */ 390 if (nr_shown == 60) { 391 if (time_before(jiffies, resume)) { 392 nr_unshown++; 393 goto out; 394 } 395 if (nr_unshown) { 396 printk(KERN_ALERT 397 "BUG: Bad page state: %lu messages suppressed\n", 398 nr_unshown); 399 nr_unshown = 0; 400 } 401 nr_shown = 0; 402 } 403 if (nr_shown++ == 0) 404 resume = jiffies + 60 * HZ; 405 406 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 407 current->comm, page_to_pfn(page)); 408 dump_page_badflags(page, reason, bad_flags); 409 410 print_modules(); 411 dump_stack(); 412 out: 413 /* Leave bad fields for debug, except PageBuddy could make trouble */ 414 page_mapcount_reset(page); /* remove PageBuddy */ 415 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 416 } 417 418 /* 419 * Higher-order pages are called "compound pages". They are structured thusly: 420 * 421 * The first PAGE_SIZE page is called the "head page". 422 * 423 * The remaining PAGE_SIZE pages are called "tail pages". 424 * 425 * All pages have PG_compound set. All tail pages have their ->first_page 426 * pointing at the head page. 427 * 428 * The first tail page's ->lru.next holds the address of the compound page's 429 * put_page() function. Its ->lru.prev holds the order of allocation. 430 * This usage means that zero-order pages may not be compound. 431 */ 432 433 static void free_compound_page(struct page *page) 434 { 435 __free_pages_ok(page, compound_order(page)); 436 } 437 438 void prep_compound_page(struct page *page, unsigned long order) 439 { 440 int i; 441 int nr_pages = 1 << order; 442 443 set_compound_page_dtor(page, free_compound_page); 444 set_compound_order(page, order); 445 __SetPageHead(page); 446 for (i = 1; i < nr_pages; i++) { 447 struct page *p = page + i; 448 set_page_count(p, 0); 449 p->first_page = page; 450 /* Make sure p->first_page is always valid for PageTail() */ 451 smp_wmb(); 452 __SetPageTail(p); 453 } 454 } 455 456 #ifdef CONFIG_DEBUG_PAGEALLOC 457 unsigned int _debug_guardpage_minorder; 458 bool _debug_pagealloc_enabled __read_mostly; 459 bool _debug_guardpage_enabled __read_mostly; 460 461 static int __init early_debug_pagealloc(char *buf) 462 { 463 if (!buf) 464 return -EINVAL; 465 466 if (strcmp(buf, "on") == 0) 467 _debug_pagealloc_enabled = true; 468 469 return 0; 470 } 471 early_param("debug_pagealloc", early_debug_pagealloc); 472 473 static bool need_debug_guardpage(void) 474 { 475 /* If we don't use debug_pagealloc, we don't need guard page */ 476 if (!debug_pagealloc_enabled()) 477 return false; 478 479 return true; 480 } 481 482 static void init_debug_guardpage(void) 483 { 484 if (!debug_pagealloc_enabled()) 485 return; 486 487 _debug_guardpage_enabled = true; 488 } 489 490 struct page_ext_operations debug_guardpage_ops = { 491 .need = need_debug_guardpage, 492 .init = init_debug_guardpage, 493 }; 494 495 static int __init debug_guardpage_minorder_setup(char *buf) 496 { 497 unsigned long res; 498 499 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 500 printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); 501 return 0; 502 } 503 _debug_guardpage_minorder = res; 504 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); 505 return 0; 506 } 507 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); 508 509 static inline void set_page_guard(struct zone *zone, struct page *page, 510 unsigned int order, int migratetype) 511 { 512 struct page_ext *page_ext; 513 514 if (!debug_guardpage_enabled()) 515 return; 516 517 page_ext = lookup_page_ext(page); 518 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 519 520 INIT_LIST_HEAD(&page->lru); 521 set_page_private(page, order); 522 /* Guard pages are not available for any usage */ 523 __mod_zone_freepage_state(zone, -(1 << order), migratetype); 524 } 525 526 static inline void clear_page_guard(struct zone *zone, struct page *page, 527 unsigned int order, int migratetype) 528 { 529 struct page_ext *page_ext; 530 531 if (!debug_guardpage_enabled()) 532 return; 533 534 page_ext = lookup_page_ext(page); 535 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 536 537 set_page_private(page, 0); 538 if (!is_migrate_isolate(migratetype)) 539 __mod_zone_freepage_state(zone, (1 << order), migratetype); 540 } 541 #else 542 struct page_ext_operations debug_guardpage_ops = { NULL, }; 543 static inline void set_page_guard(struct zone *zone, struct page *page, 544 unsigned int order, int migratetype) {} 545 static inline void clear_page_guard(struct zone *zone, struct page *page, 546 unsigned int order, int migratetype) {} 547 #endif 548 549 static inline void set_page_order(struct page *page, unsigned int order) 550 { 551 set_page_private(page, order); 552 __SetPageBuddy(page); 553 } 554 555 static inline void rmv_page_order(struct page *page) 556 { 557 __ClearPageBuddy(page); 558 set_page_private(page, 0); 559 } 560 561 /* 562 * This function checks whether a page is free && is the buddy 563 * we can do coalesce a page and its buddy if 564 * (a) the buddy is not in a hole && 565 * (b) the buddy is in the buddy system && 566 * (c) a page and its buddy have the same order && 567 * (d) a page and its buddy are in the same zone. 568 * 569 * For recording whether a page is in the buddy system, we set ->_mapcount 570 * PAGE_BUDDY_MAPCOUNT_VALUE. 571 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is 572 * serialized by zone->lock. 573 * 574 * For recording page's order, we use page_private(page). 575 */ 576 static inline int page_is_buddy(struct page *page, struct page *buddy, 577 unsigned int order) 578 { 579 if (!pfn_valid_within(page_to_pfn(buddy))) 580 return 0; 581 582 if (page_is_guard(buddy) && page_order(buddy) == order) { 583 if (page_zone_id(page) != page_zone_id(buddy)) 584 return 0; 585 586 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 587 588 return 1; 589 } 590 591 if (PageBuddy(buddy) && page_order(buddy) == order) { 592 /* 593 * zone check is done late to avoid uselessly 594 * calculating zone/node ids for pages that could 595 * never merge. 596 */ 597 if (page_zone_id(page) != page_zone_id(buddy)) 598 return 0; 599 600 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 601 602 return 1; 603 } 604 return 0; 605 } 606 607 /* 608 * Freeing function for a buddy system allocator. 609 * 610 * The concept of a buddy system is to maintain direct-mapped table 611 * (containing bit values) for memory blocks of various "orders". 612 * The bottom level table contains the map for the smallest allocatable 613 * units of memory (here, pages), and each level above it describes 614 * pairs of units from the levels below, hence, "buddies". 615 * At a high level, all that happens here is marking the table entry 616 * at the bottom level available, and propagating the changes upward 617 * as necessary, plus some accounting needed to play nicely with other 618 * parts of the VM system. 619 * At each level, we keep a list of pages, which are heads of continuous 620 * free pages of length of (1 << order) and marked with _mapcount 621 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) 622 * field. 623 * So when we are allocating or freeing one, we can derive the state of the 624 * other. That is, if we allocate a small block, and both were 625 * free, the remainder of the region must be split into blocks. 626 * If a block is freed, and its buddy is also free, then this 627 * triggers coalescing into a block of larger size. 628 * 629 * -- nyc 630 */ 631 632 static inline void __free_one_page(struct page *page, 633 unsigned long pfn, 634 struct zone *zone, unsigned int order, 635 int migratetype) 636 { 637 unsigned long page_idx; 638 unsigned long combined_idx; 639 unsigned long uninitialized_var(buddy_idx); 640 struct page *buddy; 641 int max_order = MAX_ORDER; 642 643 VM_BUG_ON(!zone_is_initialized(zone)); 644 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); 645 646 VM_BUG_ON(migratetype == -1); 647 if (is_migrate_isolate(migratetype)) { 648 /* 649 * We restrict max order of merging to prevent merge 650 * between freepages on isolate pageblock and normal 651 * pageblock. Without this, pageblock isolation 652 * could cause incorrect freepage accounting. 653 */ 654 max_order = min(MAX_ORDER, pageblock_order + 1); 655 } else { 656 __mod_zone_freepage_state(zone, 1 << order, migratetype); 657 } 658 659 page_idx = pfn & ((1 << max_order) - 1); 660 661 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page); 662 VM_BUG_ON_PAGE(bad_range(zone, page), page); 663 664 while (order < max_order - 1) { 665 buddy_idx = __find_buddy_index(page_idx, order); 666 buddy = page + (buddy_idx - page_idx); 667 if (!page_is_buddy(page, buddy, order)) 668 break; 669 /* 670 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 671 * merge with it and move up one order. 672 */ 673 if (page_is_guard(buddy)) { 674 clear_page_guard(zone, buddy, order, migratetype); 675 } else { 676 list_del(&buddy->lru); 677 zone->free_area[order].nr_free--; 678 rmv_page_order(buddy); 679 } 680 combined_idx = buddy_idx & page_idx; 681 page = page + (combined_idx - page_idx); 682 page_idx = combined_idx; 683 order++; 684 } 685 set_page_order(page, order); 686 687 /* 688 * If this is not the largest possible page, check if the buddy 689 * of the next-highest order is free. If it is, it's possible 690 * that pages are being freed that will coalesce soon. In case, 691 * that is happening, add the free page to the tail of the list 692 * so it's less likely to be used soon and more likely to be merged 693 * as a higher order page 694 */ 695 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 696 struct page *higher_page, *higher_buddy; 697 combined_idx = buddy_idx & page_idx; 698 higher_page = page + (combined_idx - page_idx); 699 buddy_idx = __find_buddy_index(combined_idx, order + 1); 700 higher_buddy = higher_page + (buddy_idx - combined_idx); 701 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 702 list_add_tail(&page->lru, 703 &zone->free_area[order].free_list[migratetype]); 704 goto out; 705 } 706 } 707 708 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 709 out: 710 zone->free_area[order].nr_free++; 711 } 712 713 static inline int free_pages_check(struct page *page) 714 { 715 const char *bad_reason = NULL; 716 unsigned long bad_flags = 0; 717 718 if (unlikely(page_mapcount(page))) 719 bad_reason = "nonzero mapcount"; 720 if (unlikely(page->mapping != NULL)) 721 bad_reason = "non-NULL mapping"; 722 if (unlikely(atomic_read(&page->_count) != 0)) 723 bad_reason = "nonzero _count"; 724 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { 725 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; 726 bad_flags = PAGE_FLAGS_CHECK_AT_FREE; 727 } 728 #ifdef CONFIG_MEMCG 729 if (unlikely(page->mem_cgroup)) 730 bad_reason = "page still charged to cgroup"; 731 #endif 732 if (unlikely(bad_reason)) { 733 bad_page(page, bad_reason, bad_flags); 734 return 1; 735 } 736 page_cpupid_reset_last(page); 737 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 738 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 739 return 0; 740 } 741 742 /* 743 * Frees a number of pages from the PCP lists 744 * Assumes all pages on list are in same zone, and of same order. 745 * count is the number of pages to free. 746 * 747 * If the zone was previously in an "all pages pinned" state then look to 748 * see if this freeing clears that state. 749 * 750 * And clear the zone's pages_scanned counter, to hold off the "all pages are 751 * pinned" detection logic. 752 */ 753 static void free_pcppages_bulk(struct zone *zone, int count, 754 struct per_cpu_pages *pcp) 755 { 756 int migratetype = 0; 757 int batch_free = 0; 758 int to_free = count; 759 unsigned long nr_scanned; 760 761 spin_lock(&zone->lock); 762 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 763 if (nr_scanned) 764 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 765 766 while (to_free) { 767 struct page *page; 768 struct list_head *list; 769 770 /* 771 * Remove pages from lists in a round-robin fashion. A 772 * batch_free count is maintained that is incremented when an 773 * empty list is encountered. This is so more pages are freed 774 * off fuller lists instead of spinning excessively around empty 775 * lists 776 */ 777 do { 778 batch_free++; 779 if (++migratetype == MIGRATE_PCPTYPES) 780 migratetype = 0; 781 list = &pcp->lists[migratetype]; 782 } while (list_empty(list)); 783 784 /* This is the only non-empty list. Free them all. */ 785 if (batch_free == MIGRATE_PCPTYPES) 786 batch_free = to_free; 787 788 do { 789 int mt; /* migratetype of the to-be-freed page */ 790 791 page = list_entry(list->prev, struct page, lru); 792 /* must delete as __free_one_page list manipulates */ 793 list_del(&page->lru); 794 mt = get_freepage_migratetype(page); 795 if (unlikely(has_isolate_pageblock(zone))) 796 mt = get_pageblock_migratetype(page); 797 798 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 799 __free_one_page(page, page_to_pfn(page), zone, 0, mt); 800 trace_mm_page_pcpu_drain(page, 0, mt); 801 } while (--to_free && --batch_free && !list_empty(list)); 802 } 803 spin_unlock(&zone->lock); 804 } 805 806 static void free_one_page(struct zone *zone, 807 struct page *page, unsigned long pfn, 808 unsigned int order, 809 int migratetype) 810 { 811 unsigned long nr_scanned; 812 spin_lock(&zone->lock); 813 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 814 if (nr_scanned) 815 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 816 817 if (unlikely(has_isolate_pageblock(zone) || 818 is_migrate_isolate(migratetype))) { 819 migratetype = get_pfnblock_migratetype(page, pfn); 820 } 821 __free_one_page(page, pfn, zone, order, migratetype); 822 spin_unlock(&zone->lock); 823 } 824 825 static int free_tail_pages_check(struct page *head_page, struct page *page) 826 { 827 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 828 return 0; 829 if (unlikely(!PageTail(page))) { 830 bad_page(page, "PageTail not set", 0); 831 return 1; 832 } 833 if (unlikely(page->first_page != head_page)) { 834 bad_page(page, "first_page not consistent", 0); 835 return 1; 836 } 837 return 0; 838 } 839 840 static void __meminit __init_single_page(struct page *page, unsigned long pfn, 841 unsigned long zone, int nid) 842 { 843 set_page_links(page, zone, nid, pfn); 844 init_page_count(page); 845 page_mapcount_reset(page); 846 page_cpupid_reset_last(page); 847 848 INIT_LIST_HEAD(&page->lru); 849 #ifdef WANT_PAGE_VIRTUAL 850 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 851 if (!is_highmem_idx(zone)) 852 set_page_address(page, __va(pfn << PAGE_SHIFT)); 853 #endif 854 } 855 856 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone, 857 int nid) 858 { 859 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid); 860 } 861 862 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 863 static void init_reserved_page(unsigned long pfn) 864 { 865 pg_data_t *pgdat; 866 int nid, zid; 867 868 if (!early_page_uninitialised(pfn)) 869 return; 870 871 nid = early_pfn_to_nid(pfn); 872 pgdat = NODE_DATA(nid); 873 874 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 875 struct zone *zone = &pgdat->node_zones[zid]; 876 877 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone)) 878 break; 879 } 880 __init_single_pfn(pfn, zid, nid); 881 } 882 #else 883 static inline void init_reserved_page(unsigned long pfn) 884 { 885 } 886 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 887 888 /* 889 * Initialised pages do not have PageReserved set. This function is 890 * called for each range allocated by the bootmem allocator and 891 * marks the pages PageReserved. The remaining valid pages are later 892 * sent to the buddy page allocator. 893 */ 894 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end) 895 { 896 unsigned long start_pfn = PFN_DOWN(start); 897 unsigned long end_pfn = PFN_UP(end); 898 899 for (; start_pfn < end_pfn; start_pfn++) { 900 if (pfn_valid(start_pfn)) { 901 struct page *page = pfn_to_page(start_pfn); 902 903 init_reserved_page(start_pfn); 904 SetPageReserved(page); 905 } 906 } 907 } 908 909 static bool free_pages_prepare(struct page *page, unsigned int order) 910 { 911 bool compound = PageCompound(page); 912 int i, bad = 0; 913 914 VM_BUG_ON_PAGE(PageTail(page), page); 915 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); 916 917 trace_mm_page_free(page, order); 918 kmemcheck_free_shadow(page, order); 919 kasan_free_pages(page, order); 920 921 if (PageAnon(page)) 922 page->mapping = NULL; 923 bad += free_pages_check(page); 924 for (i = 1; i < (1 << order); i++) { 925 if (compound) 926 bad += free_tail_pages_check(page, page + i); 927 bad += free_pages_check(page + i); 928 } 929 if (bad) 930 return false; 931 932 reset_page_owner(page, order); 933 934 if (!PageHighMem(page)) { 935 debug_check_no_locks_freed(page_address(page), 936 PAGE_SIZE << order); 937 debug_check_no_obj_freed(page_address(page), 938 PAGE_SIZE << order); 939 } 940 arch_free_page(page, order); 941 kernel_map_pages(page, 1 << order, 0); 942 943 return true; 944 } 945 946 static void __free_pages_ok(struct page *page, unsigned int order) 947 { 948 unsigned long flags; 949 int migratetype; 950 unsigned long pfn = page_to_pfn(page); 951 952 if (!free_pages_prepare(page, order)) 953 return; 954 955 migratetype = get_pfnblock_migratetype(page, pfn); 956 local_irq_save(flags); 957 __count_vm_events(PGFREE, 1 << order); 958 set_freepage_migratetype(page, migratetype); 959 free_one_page(page_zone(page), page, pfn, order, migratetype); 960 local_irq_restore(flags); 961 } 962 963 static void __init __free_pages_boot_core(struct page *page, 964 unsigned long pfn, unsigned int order) 965 { 966 unsigned int nr_pages = 1 << order; 967 struct page *p = page; 968 unsigned int loop; 969 970 prefetchw(p); 971 for (loop = 0; loop < (nr_pages - 1); loop++, p++) { 972 prefetchw(p + 1); 973 __ClearPageReserved(p); 974 set_page_count(p, 0); 975 } 976 __ClearPageReserved(p); 977 set_page_count(p, 0); 978 979 page_zone(page)->managed_pages += nr_pages; 980 set_page_refcounted(page); 981 __free_pages(page, order); 982 } 983 984 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \ 985 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) 986 /* Only safe to use early in boot when initialisation is single-threaded */ 987 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata; 988 989 int __meminit early_pfn_to_nid(unsigned long pfn) 990 { 991 int nid; 992 993 /* The system will behave unpredictably otherwise */ 994 BUG_ON(system_state != SYSTEM_BOOTING); 995 996 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache); 997 if (nid >= 0) 998 return nid; 999 /* just returns 0 */ 1000 return 0; 1001 } 1002 #endif 1003 1004 #ifdef CONFIG_NODES_SPAN_OTHER_NODES 1005 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1006 struct mminit_pfnnid_cache *state) 1007 { 1008 int nid; 1009 1010 nid = __early_pfn_to_nid(pfn, state); 1011 if (nid >= 0 && nid != node) 1012 return false; 1013 return true; 1014 } 1015 1016 /* Only safe to use early in boot when initialisation is single-threaded */ 1017 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1018 { 1019 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache); 1020 } 1021 1022 #else 1023 1024 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1025 { 1026 return true; 1027 } 1028 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1029 struct mminit_pfnnid_cache *state) 1030 { 1031 return true; 1032 } 1033 #endif 1034 1035 1036 void __init __free_pages_bootmem(struct page *page, unsigned long pfn, 1037 unsigned int order) 1038 { 1039 if (early_page_uninitialised(pfn)) 1040 return; 1041 return __free_pages_boot_core(page, pfn, order); 1042 } 1043 1044 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 1045 static void __init deferred_free_range(struct page *page, 1046 unsigned long pfn, int nr_pages) 1047 { 1048 int i; 1049 1050 if (!page) 1051 return; 1052 1053 /* Free a large naturally-aligned chunk if possible */ 1054 if (nr_pages == MAX_ORDER_NR_PAGES && 1055 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) { 1056 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1057 __free_pages_boot_core(page, pfn, MAX_ORDER-1); 1058 return; 1059 } 1060 1061 for (i = 0; i < nr_pages; i++, page++, pfn++) 1062 __free_pages_boot_core(page, pfn, 0); 1063 } 1064 1065 static __initdata DECLARE_RWSEM(pgdat_init_rwsem); 1066 1067 /* Initialise remaining memory on a node */ 1068 static int __init deferred_init_memmap(void *data) 1069 { 1070 pg_data_t *pgdat = data; 1071 int nid = pgdat->node_id; 1072 struct mminit_pfnnid_cache nid_init_state = { }; 1073 unsigned long start = jiffies; 1074 unsigned long nr_pages = 0; 1075 unsigned long walk_start, walk_end; 1076 int i, zid; 1077 struct zone *zone; 1078 unsigned long first_init_pfn = pgdat->first_deferred_pfn; 1079 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 1080 1081 if (first_init_pfn == ULONG_MAX) { 1082 up_read(&pgdat_init_rwsem); 1083 return 0; 1084 } 1085 1086 /* Bind memory initialisation thread to a local node if possible */ 1087 if (!cpumask_empty(cpumask)) 1088 set_cpus_allowed_ptr(current, cpumask); 1089 1090 /* Sanity check boundaries */ 1091 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn); 1092 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat)); 1093 pgdat->first_deferred_pfn = ULONG_MAX; 1094 1095 /* Only the highest zone is deferred so find it */ 1096 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1097 zone = pgdat->node_zones + zid; 1098 if (first_init_pfn < zone_end_pfn(zone)) 1099 break; 1100 } 1101 1102 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) { 1103 unsigned long pfn, end_pfn; 1104 struct page *page = NULL; 1105 struct page *free_base_page = NULL; 1106 unsigned long free_base_pfn = 0; 1107 int nr_to_free = 0; 1108 1109 end_pfn = min(walk_end, zone_end_pfn(zone)); 1110 pfn = first_init_pfn; 1111 if (pfn < walk_start) 1112 pfn = walk_start; 1113 if (pfn < zone->zone_start_pfn) 1114 pfn = zone->zone_start_pfn; 1115 1116 for (; pfn < end_pfn; pfn++) { 1117 if (!pfn_valid_within(pfn)) 1118 goto free_range; 1119 1120 /* 1121 * Ensure pfn_valid is checked every 1122 * MAX_ORDER_NR_PAGES for memory holes 1123 */ 1124 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { 1125 if (!pfn_valid(pfn)) { 1126 page = NULL; 1127 goto free_range; 1128 } 1129 } 1130 1131 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) { 1132 page = NULL; 1133 goto free_range; 1134 } 1135 1136 /* Minimise pfn page lookups and scheduler checks */ 1137 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) { 1138 page++; 1139 } else { 1140 nr_pages += nr_to_free; 1141 deferred_free_range(free_base_page, 1142 free_base_pfn, nr_to_free); 1143 free_base_page = NULL; 1144 free_base_pfn = nr_to_free = 0; 1145 1146 page = pfn_to_page(pfn); 1147 cond_resched(); 1148 } 1149 1150 if (page->flags) { 1151 VM_BUG_ON(page_zone(page) != zone); 1152 goto free_range; 1153 } 1154 1155 __init_single_page(page, pfn, zid, nid); 1156 if (!free_base_page) { 1157 free_base_page = page; 1158 free_base_pfn = pfn; 1159 nr_to_free = 0; 1160 } 1161 nr_to_free++; 1162 1163 /* Where possible, batch up pages for a single free */ 1164 continue; 1165 free_range: 1166 /* Free the current block of pages to allocator */ 1167 nr_pages += nr_to_free; 1168 deferred_free_range(free_base_page, free_base_pfn, 1169 nr_to_free); 1170 free_base_page = NULL; 1171 free_base_pfn = nr_to_free = 0; 1172 } 1173 1174 first_init_pfn = max(end_pfn, first_init_pfn); 1175 } 1176 1177 /* Sanity check that the next zone really is unpopulated */ 1178 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone)); 1179 1180 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages, 1181 jiffies_to_msecs(jiffies - start)); 1182 up_read(&pgdat_init_rwsem); 1183 return 0; 1184 } 1185 1186 void __init page_alloc_init_late(void) 1187 { 1188 int nid; 1189 1190 for_each_node_state(nid, N_MEMORY) { 1191 down_read(&pgdat_init_rwsem); 1192 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid); 1193 } 1194 1195 /* Block until all are initialised */ 1196 down_write(&pgdat_init_rwsem); 1197 up_write(&pgdat_init_rwsem); 1198 } 1199 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 1200 1201 #ifdef CONFIG_CMA 1202 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ 1203 void __init init_cma_reserved_pageblock(struct page *page) 1204 { 1205 unsigned i = pageblock_nr_pages; 1206 struct page *p = page; 1207 1208 do { 1209 __ClearPageReserved(p); 1210 set_page_count(p, 0); 1211 } while (++p, --i); 1212 1213 set_pageblock_migratetype(page, MIGRATE_CMA); 1214 1215 if (pageblock_order >= MAX_ORDER) { 1216 i = pageblock_nr_pages; 1217 p = page; 1218 do { 1219 set_page_refcounted(p); 1220 __free_pages(p, MAX_ORDER - 1); 1221 p += MAX_ORDER_NR_PAGES; 1222 } while (i -= MAX_ORDER_NR_PAGES); 1223 } else { 1224 set_page_refcounted(page); 1225 __free_pages(page, pageblock_order); 1226 } 1227 1228 adjust_managed_page_count(page, pageblock_nr_pages); 1229 } 1230 #endif 1231 1232 /* 1233 * The order of subdivision here is critical for the IO subsystem. 1234 * Please do not alter this order without good reasons and regression 1235 * testing. Specifically, as large blocks of memory are subdivided, 1236 * the order in which smaller blocks are delivered depends on the order 1237 * they're subdivided in this function. This is the primary factor 1238 * influencing the order in which pages are delivered to the IO 1239 * subsystem according to empirical testing, and this is also justified 1240 * by considering the behavior of a buddy system containing a single 1241 * large block of memory acted on by a series of small allocations. 1242 * This behavior is a critical factor in sglist merging's success. 1243 * 1244 * -- nyc 1245 */ 1246 static inline void expand(struct zone *zone, struct page *page, 1247 int low, int high, struct free_area *area, 1248 int migratetype) 1249 { 1250 unsigned long size = 1 << high; 1251 1252 while (high > low) { 1253 area--; 1254 high--; 1255 size >>= 1; 1256 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); 1257 1258 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) && 1259 debug_guardpage_enabled() && 1260 high < debug_guardpage_minorder()) { 1261 /* 1262 * Mark as guard pages (or page), that will allow to 1263 * merge back to allocator when buddy will be freed. 1264 * Corresponding page table entries will not be touched, 1265 * pages will stay not present in virtual address space 1266 */ 1267 set_page_guard(zone, &page[size], high, migratetype); 1268 continue; 1269 } 1270 list_add(&page[size].lru, &area->free_list[migratetype]); 1271 area->nr_free++; 1272 set_page_order(&page[size], high); 1273 } 1274 } 1275 1276 /* 1277 * This page is about to be returned from the page allocator 1278 */ 1279 static inline int check_new_page(struct page *page) 1280 { 1281 const char *bad_reason = NULL; 1282 unsigned long bad_flags = 0; 1283 1284 if (unlikely(page_mapcount(page))) 1285 bad_reason = "nonzero mapcount"; 1286 if (unlikely(page->mapping != NULL)) 1287 bad_reason = "non-NULL mapping"; 1288 if (unlikely(atomic_read(&page->_count) != 0)) 1289 bad_reason = "nonzero _count"; 1290 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { 1291 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; 1292 bad_flags = PAGE_FLAGS_CHECK_AT_PREP; 1293 } 1294 #ifdef CONFIG_MEMCG 1295 if (unlikely(page->mem_cgroup)) 1296 bad_reason = "page still charged to cgroup"; 1297 #endif 1298 if (unlikely(bad_reason)) { 1299 bad_page(page, bad_reason, bad_flags); 1300 return 1; 1301 } 1302 return 0; 1303 } 1304 1305 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, 1306 int alloc_flags) 1307 { 1308 int i; 1309 1310 for (i = 0; i < (1 << order); i++) { 1311 struct page *p = page + i; 1312 if (unlikely(check_new_page(p))) 1313 return 1; 1314 } 1315 1316 set_page_private(page, 0); 1317 set_page_refcounted(page); 1318 1319 arch_alloc_page(page, order); 1320 kernel_map_pages(page, 1 << order, 1); 1321 kasan_alloc_pages(page, order); 1322 1323 if (gfp_flags & __GFP_ZERO) 1324 for (i = 0; i < (1 << order); i++) 1325 clear_highpage(page + i); 1326 1327 if (order && (gfp_flags & __GFP_COMP)) 1328 prep_compound_page(page, order); 1329 1330 set_page_owner(page, order, gfp_flags); 1331 1332 /* 1333 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to 1334 * allocate the page. The expectation is that the caller is taking 1335 * steps that will free more memory. The caller should avoid the page 1336 * being used for !PFMEMALLOC purposes. 1337 */ 1338 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); 1339 1340 return 0; 1341 } 1342 1343 /* 1344 * Go through the free lists for the given migratetype and remove 1345 * the smallest available page from the freelists 1346 */ 1347 static inline 1348 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 1349 int migratetype) 1350 { 1351 unsigned int current_order; 1352 struct free_area *area; 1353 struct page *page; 1354 1355 /* Find a page of the appropriate size in the preferred list */ 1356 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 1357 area = &(zone->free_area[current_order]); 1358 if (list_empty(&area->free_list[migratetype])) 1359 continue; 1360 1361 page = list_entry(area->free_list[migratetype].next, 1362 struct page, lru); 1363 list_del(&page->lru); 1364 rmv_page_order(page); 1365 area->nr_free--; 1366 expand(zone, page, order, current_order, area, migratetype); 1367 set_freepage_migratetype(page, migratetype); 1368 return page; 1369 } 1370 1371 return NULL; 1372 } 1373 1374 1375 /* 1376 * This array describes the order lists are fallen back to when 1377 * the free lists for the desirable migrate type are depleted 1378 */ 1379 static int fallbacks[MIGRATE_TYPES][4] = { 1380 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 1381 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 1382 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 1383 #ifdef CONFIG_CMA 1384 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ 1385 #endif 1386 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ 1387 #ifdef CONFIG_MEMORY_ISOLATION 1388 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ 1389 #endif 1390 }; 1391 1392 #ifdef CONFIG_CMA 1393 static struct page *__rmqueue_cma_fallback(struct zone *zone, 1394 unsigned int order) 1395 { 1396 return __rmqueue_smallest(zone, order, MIGRATE_CMA); 1397 } 1398 #else 1399 static inline struct page *__rmqueue_cma_fallback(struct zone *zone, 1400 unsigned int order) { return NULL; } 1401 #endif 1402 1403 /* 1404 * Move the free pages in a range to the free lists of the requested type. 1405 * Note that start_page and end_pages are not aligned on a pageblock 1406 * boundary. If alignment is required, use move_freepages_block() 1407 */ 1408 int move_freepages(struct zone *zone, 1409 struct page *start_page, struct page *end_page, 1410 int migratetype) 1411 { 1412 struct page *page; 1413 unsigned long order; 1414 int pages_moved = 0; 1415 1416 #ifndef CONFIG_HOLES_IN_ZONE 1417 /* 1418 * page_zone is not safe to call in this context when 1419 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 1420 * anyway as we check zone boundaries in move_freepages_block(). 1421 * Remove at a later date when no bug reports exist related to 1422 * grouping pages by mobility 1423 */ 1424 VM_BUG_ON(page_zone(start_page) != page_zone(end_page)); 1425 #endif 1426 1427 for (page = start_page; page <= end_page;) { 1428 /* Make sure we are not inadvertently changing nodes */ 1429 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); 1430 1431 if (!pfn_valid_within(page_to_pfn(page))) { 1432 page++; 1433 continue; 1434 } 1435 1436 if (!PageBuddy(page)) { 1437 page++; 1438 continue; 1439 } 1440 1441 order = page_order(page); 1442 list_move(&page->lru, 1443 &zone->free_area[order].free_list[migratetype]); 1444 set_freepage_migratetype(page, migratetype); 1445 page += 1 << order; 1446 pages_moved += 1 << order; 1447 } 1448 1449 return pages_moved; 1450 } 1451 1452 int move_freepages_block(struct zone *zone, struct page *page, 1453 int migratetype) 1454 { 1455 unsigned long start_pfn, end_pfn; 1456 struct page *start_page, *end_page; 1457 1458 start_pfn = page_to_pfn(page); 1459 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 1460 start_page = pfn_to_page(start_pfn); 1461 end_page = start_page + pageblock_nr_pages - 1; 1462 end_pfn = start_pfn + pageblock_nr_pages - 1; 1463 1464 /* Do not cross zone boundaries */ 1465 if (!zone_spans_pfn(zone, start_pfn)) 1466 start_page = page; 1467 if (!zone_spans_pfn(zone, end_pfn)) 1468 return 0; 1469 1470 return move_freepages(zone, start_page, end_page, migratetype); 1471 } 1472 1473 static void change_pageblock_range(struct page *pageblock_page, 1474 int start_order, int migratetype) 1475 { 1476 int nr_pageblocks = 1 << (start_order - pageblock_order); 1477 1478 while (nr_pageblocks--) { 1479 set_pageblock_migratetype(pageblock_page, migratetype); 1480 pageblock_page += pageblock_nr_pages; 1481 } 1482 } 1483 1484 /* 1485 * When we are falling back to another migratetype during allocation, try to 1486 * steal extra free pages from the same pageblocks to satisfy further 1487 * allocations, instead of polluting multiple pageblocks. 1488 * 1489 * If we are stealing a relatively large buddy page, it is likely there will 1490 * be more free pages in the pageblock, so try to steal them all. For 1491 * reclaimable and unmovable allocations, we steal regardless of page size, 1492 * as fragmentation caused by those allocations polluting movable pageblocks 1493 * is worse than movable allocations stealing from unmovable and reclaimable 1494 * pageblocks. 1495 */ 1496 static bool can_steal_fallback(unsigned int order, int start_mt) 1497 { 1498 /* 1499 * Leaving this order check is intended, although there is 1500 * relaxed order check in next check. The reason is that 1501 * we can actually steal whole pageblock if this condition met, 1502 * but, below check doesn't guarantee it and that is just heuristic 1503 * so could be changed anytime. 1504 */ 1505 if (order >= pageblock_order) 1506 return true; 1507 1508 if (order >= pageblock_order / 2 || 1509 start_mt == MIGRATE_RECLAIMABLE || 1510 start_mt == MIGRATE_UNMOVABLE || 1511 page_group_by_mobility_disabled) 1512 return true; 1513 1514 return false; 1515 } 1516 1517 /* 1518 * This function implements actual steal behaviour. If order is large enough, 1519 * we can steal whole pageblock. If not, we first move freepages in this 1520 * pageblock and check whether half of pages are moved or not. If half of 1521 * pages are moved, we can change migratetype of pageblock and permanently 1522 * use it's pages as requested migratetype in the future. 1523 */ 1524 static void steal_suitable_fallback(struct zone *zone, struct page *page, 1525 int start_type) 1526 { 1527 int current_order = page_order(page); 1528 int pages; 1529 1530 /* Take ownership for orders >= pageblock_order */ 1531 if (current_order >= pageblock_order) { 1532 change_pageblock_range(page, current_order, start_type); 1533 return; 1534 } 1535 1536 pages = move_freepages_block(zone, page, start_type); 1537 1538 /* Claim the whole block if over half of it is free */ 1539 if (pages >= (1 << (pageblock_order-1)) || 1540 page_group_by_mobility_disabled) 1541 set_pageblock_migratetype(page, start_type); 1542 } 1543 1544 /* 1545 * Check whether there is a suitable fallback freepage with requested order. 1546 * If only_stealable is true, this function returns fallback_mt only if 1547 * we can steal other freepages all together. This would help to reduce 1548 * fragmentation due to mixed migratetype pages in one pageblock. 1549 */ 1550 int find_suitable_fallback(struct free_area *area, unsigned int order, 1551 int migratetype, bool only_stealable, bool *can_steal) 1552 { 1553 int i; 1554 int fallback_mt; 1555 1556 if (area->nr_free == 0) 1557 return -1; 1558 1559 *can_steal = false; 1560 for (i = 0;; i++) { 1561 fallback_mt = fallbacks[migratetype][i]; 1562 if (fallback_mt == MIGRATE_RESERVE) 1563 break; 1564 1565 if (list_empty(&area->free_list[fallback_mt])) 1566 continue; 1567 1568 if (can_steal_fallback(order, migratetype)) 1569 *can_steal = true; 1570 1571 if (!only_stealable) 1572 return fallback_mt; 1573 1574 if (*can_steal) 1575 return fallback_mt; 1576 } 1577 1578 return -1; 1579 } 1580 1581 /* Remove an element from the buddy allocator from the fallback list */ 1582 static inline struct page * 1583 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) 1584 { 1585 struct free_area *area; 1586 unsigned int current_order; 1587 struct page *page; 1588 int fallback_mt; 1589 bool can_steal; 1590 1591 /* Find the largest possible block of pages in the other list */ 1592 for (current_order = MAX_ORDER-1; 1593 current_order >= order && current_order <= MAX_ORDER-1; 1594 --current_order) { 1595 area = &(zone->free_area[current_order]); 1596 fallback_mt = find_suitable_fallback(area, current_order, 1597 start_migratetype, false, &can_steal); 1598 if (fallback_mt == -1) 1599 continue; 1600 1601 page = list_entry(area->free_list[fallback_mt].next, 1602 struct page, lru); 1603 if (can_steal) 1604 steal_suitable_fallback(zone, page, start_migratetype); 1605 1606 /* Remove the page from the freelists */ 1607 area->nr_free--; 1608 list_del(&page->lru); 1609 rmv_page_order(page); 1610 1611 expand(zone, page, order, current_order, area, 1612 start_migratetype); 1613 /* 1614 * The freepage_migratetype may differ from pageblock's 1615 * migratetype depending on the decisions in 1616 * try_to_steal_freepages(). This is OK as long as it 1617 * does not differ for MIGRATE_CMA pageblocks. For CMA 1618 * we need to make sure unallocated pages flushed from 1619 * pcp lists are returned to the correct freelist. 1620 */ 1621 set_freepage_migratetype(page, start_migratetype); 1622 1623 trace_mm_page_alloc_extfrag(page, order, current_order, 1624 start_migratetype, fallback_mt); 1625 1626 return page; 1627 } 1628 1629 return NULL; 1630 } 1631 1632 /* 1633 * Do the hard work of removing an element from the buddy allocator. 1634 * Call me with the zone->lock already held. 1635 */ 1636 static struct page *__rmqueue(struct zone *zone, unsigned int order, 1637 int migratetype) 1638 { 1639 struct page *page; 1640 1641 retry_reserve: 1642 page = __rmqueue_smallest(zone, order, migratetype); 1643 1644 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1645 if (migratetype == MIGRATE_MOVABLE) 1646 page = __rmqueue_cma_fallback(zone, order); 1647 1648 if (!page) 1649 page = __rmqueue_fallback(zone, order, migratetype); 1650 1651 /* 1652 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1653 * is used because __rmqueue_smallest is an inline function 1654 * and we want just one call site 1655 */ 1656 if (!page) { 1657 migratetype = MIGRATE_RESERVE; 1658 goto retry_reserve; 1659 } 1660 } 1661 1662 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1663 return page; 1664 } 1665 1666 /* 1667 * Obtain a specified number of elements from the buddy allocator, all under 1668 * a single hold of the lock, for efficiency. Add them to the supplied list. 1669 * Returns the number of new pages which were placed at *list. 1670 */ 1671 static int rmqueue_bulk(struct zone *zone, unsigned int order, 1672 unsigned long count, struct list_head *list, 1673 int migratetype, bool cold) 1674 { 1675 int i; 1676 1677 spin_lock(&zone->lock); 1678 for (i = 0; i < count; ++i) { 1679 struct page *page = __rmqueue(zone, order, migratetype); 1680 if (unlikely(page == NULL)) 1681 break; 1682 1683 /* 1684 * Split buddy pages returned by expand() are received here 1685 * in physical page order. The page is added to the callers and 1686 * list and the list head then moves forward. From the callers 1687 * perspective, the linked list is ordered by page number in 1688 * some conditions. This is useful for IO devices that can 1689 * merge IO requests if the physical pages are ordered 1690 * properly. 1691 */ 1692 if (likely(!cold)) 1693 list_add(&page->lru, list); 1694 else 1695 list_add_tail(&page->lru, list); 1696 list = &page->lru; 1697 if (is_migrate_cma(get_freepage_migratetype(page))) 1698 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1699 -(1 << order)); 1700 } 1701 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1702 spin_unlock(&zone->lock); 1703 return i; 1704 } 1705 1706 #ifdef CONFIG_NUMA 1707 /* 1708 * Called from the vmstat counter updater to drain pagesets of this 1709 * currently executing processor on remote nodes after they have 1710 * expired. 1711 * 1712 * Note that this function must be called with the thread pinned to 1713 * a single processor. 1714 */ 1715 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1716 { 1717 unsigned long flags; 1718 int to_drain, batch; 1719 1720 local_irq_save(flags); 1721 batch = READ_ONCE(pcp->batch); 1722 to_drain = min(pcp->count, batch); 1723 if (to_drain > 0) { 1724 free_pcppages_bulk(zone, to_drain, pcp); 1725 pcp->count -= to_drain; 1726 } 1727 local_irq_restore(flags); 1728 } 1729 #endif 1730 1731 /* 1732 * Drain pcplists of the indicated processor and zone. 1733 * 1734 * The processor must either be the current processor and the 1735 * thread pinned to the current processor or a processor that 1736 * is not online. 1737 */ 1738 static void drain_pages_zone(unsigned int cpu, struct zone *zone) 1739 { 1740 unsigned long flags; 1741 struct per_cpu_pageset *pset; 1742 struct per_cpu_pages *pcp; 1743 1744 local_irq_save(flags); 1745 pset = per_cpu_ptr(zone->pageset, cpu); 1746 1747 pcp = &pset->pcp; 1748 if (pcp->count) { 1749 free_pcppages_bulk(zone, pcp->count, pcp); 1750 pcp->count = 0; 1751 } 1752 local_irq_restore(flags); 1753 } 1754 1755 /* 1756 * Drain pcplists of all zones on the indicated processor. 1757 * 1758 * The processor must either be the current processor and the 1759 * thread pinned to the current processor or a processor that 1760 * is not online. 1761 */ 1762 static void drain_pages(unsigned int cpu) 1763 { 1764 struct zone *zone; 1765 1766 for_each_populated_zone(zone) { 1767 drain_pages_zone(cpu, zone); 1768 } 1769 } 1770 1771 /* 1772 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1773 * 1774 * The CPU has to be pinned. When zone parameter is non-NULL, spill just 1775 * the single zone's pages. 1776 */ 1777 void drain_local_pages(struct zone *zone) 1778 { 1779 int cpu = smp_processor_id(); 1780 1781 if (zone) 1782 drain_pages_zone(cpu, zone); 1783 else 1784 drain_pages(cpu); 1785 } 1786 1787 /* 1788 * Spill all the per-cpu pages from all CPUs back into the buddy allocator. 1789 * 1790 * When zone parameter is non-NULL, spill just the single zone's pages. 1791 * 1792 * Note that this code is protected against sending an IPI to an offline 1793 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: 1794 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but 1795 * nothing keeps CPUs from showing up after we populated the cpumask and 1796 * before the call to on_each_cpu_mask(). 1797 */ 1798 void drain_all_pages(struct zone *zone) 1799 { 1800 int cpu; 1801 1802 /* 1803 * Allocate in the BSS so we wont require allocation in 1804 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y 1805 */ 1806 static cpumask_t cpus_with_pcps; 1807 1808 /* 1809 * We don't care about racing with CPU hotplug event 1810 * as offline notification will cause the notified 1811 * cpu to drain that CPU pcps and on_each_cpu_mask 1812 * disables preemption as part of its processing 1813 */ 1814 for_each_online_cpu(cpu) { 1815 struct per_cpu_pageset *pcp; 1816 struct zone *z; 1817 bool has_pcps = false; 1818 1819 if (zone) { 1820 pcp = per_cpu_ptr(zone->pageset, cpu); 1821 if (pcp->pcp.count) 1822 has_pcps = true; 1823 } else { 1824 for_each_populated_zone(z) { 1825 pcp = per_cpu_ptr(z->pageset, cpu); 1826 if (pcp->pcp.count) { 1827 has_pcps = true; 1828 break; 1829 } 1830 } 1831 } 1832 1833 if (has_pcps) 1834 cpumask_set_cpu(cpu, &cpus_with_pcps); 1835 else 1836 cpumask_clear_cpu(cpu, &cpus_with_pcps); 1837 } 1838 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages, 1839 zone, 1); 1840 } 1841 1842 #ifdef CONFIG_HIBERNATION 1843 1844 void mark_free_pages(struct zone *zone) 1845 { 1846 unsigned long pfn, max_zone_pfn; 1847 unsigned long flags; 1848 unsigned int order, t; 1849 struct list_head *curr; 1850 1851 if (zone_is_empty(zone)) 1852 return; 1853 1854 spin_lock_irqsave(&zone->lock, flags); 1855 1856 max_zone_pfn = zone_end_pfn(zone); 1857 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1858 if (pfn_valid(pfn)) { 1859 struct page *page = pfn_to_page(pfn); 1860 1861 if (!swsusp_page_is_forbidden(page)) 1862 swsusp_unset_page_free(page); 1863 } 1864 1865 for_each_migratetype_order(order, t) { 1866 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1867 unsigned long i; 1868 1869 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1870 for (i = 0; i < (1UL << order); i++) 1871 swsusp_set_page_free(pfn_to_page(pfn + i)); 1872 } 1873 } 1874 spin_unlock_irqrestore(&zone->lock, flags); 1875 } 1876 #endif /* CONFIG_PM */ 1877 1878 /* 1879 * Free a 0-order page 1880 * cold == true ? free a cold page : free a hot page 1881 */ 1882 void free_hot_cold_page(struct page *page, bool cold) 1883 { 1884 struct zone *zone = page_zone(page); 1885 struct per_cpu_pages *pcp; 1886 unsigned long flags; 1887 unsigned long pfn = page_to_pfn(page); 1888 int migratetype; 1889 1890 if (!free_pages_prepare(page, 0)) 1891 return; 1892 1893 migratetype = get_pfnblock_migratetype(page, pfn); 1894 set_freepage_migratetype(page, migratetype); 1895 local_irq_save(flags); 1896 __count_vm_event(PGFREE); 1897 1898 /* 1899 * We only track unmovable, reclaimable and movable on pcp lists. 1900 * Free ISOLATE pages back to the allocator because they are being 1901 * offlined but treat RESERVE as movable pages so we can get those 1902 * areas back if necessary. Otherwise, we may have to free 1903 * excessively into the page allocator 1904 */ 1905 if (migratetype >= MIGRATE_PCPTYPES) { 1906 if (unlikely(is_migrate_isolate(migratetype))) { 1907 free_one_page(zone, page, pfn, 0, migratetype); 1908 goto out; 1909 } 1910 migratetype = MIGRATE_MOVABLE; 1911 } 1912 1913 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1914 if (!cold) 1915 list_add(&page->lru, &pcp->lists[migratetype]); 1916 else 1917 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1918 pcp->count++; 1919 if (pcp->count >= pcp->high) { 1920 unsigned long batch = READ_ONCE(pcp->batch); 1921 free_pcppages_bulk(zone, batch, pcp); 1922 pcp->count -= batch; 1923 } 1924 1925 out: 1926 local_irq_restore(flags); 1927 } 1928 1929 /* 1930 * Free a list of 0-order pages 1931 */ 1932 void free_hot_cold_page_list(struct list_head *list, bool cold) 1933 { 1934 struct page *page, *next; 1935 1936 list_for_each_entry_safe(page, next, list, lru) { 1937 trace_mm_page_free_batched(page, cold); 1938 free_hot_cold_page(page, cold); 1939 } 1940 } 1941 1942 /* 1943 * split_page takes a non-compound higher-order page, and splits it into 1944 * n (1<<order) sub-pages: page[0..n] 1945 * Each sub-page must be freed individually. 1946 * 1947 * Note: this is probably too low level an operation for use in drivers. 1948 * Please consult with lkml before using this in your driver. 1949 */ 1950 void split_page(struct page *page, unsigned int order) 1951 { 1952 int i; 1953 1954 VM_BUG_ON_PAGE(PageCompound(page), page); 1955 VM_BUG_ON_PAGE(!page_count(page), page); 1956 1957 #ifdef CONFIG_KMEMCHECK 1958 /* 1959 * Split shadow pages too, because free(page[0]) would 1960 * otherwise free the whole shadow. 1961 */ 1962 if (kmemcheck_page_is_tracked(page)) 1963 split_page(virt_to_page(page[0].shadow), order); 1964 #endif 1965 1966 set_page_owner(page, 0, 0); 1967 for (i = 1; i < (1 << order); i++) { 1968 set_page_refcounted(page + i); 1969 set_page_owner(page + i, 0, 0); 1970 } 1971 } 1972 EXPORT_SYMBOL_GPL(split_page); 1973 1974 int __isolate_free_page(struct page *page, unsigned int order) 1975 { 1976 unsigned long watermark; 1977 struct zone *zone; 1978 int mt; 1979 1980 BUG_ON(!PageBuddy(page)); 1981 1982 zone = page_zone(page); 1983 mt = get_pageblock_migratetype(page); 1984 1985 if (!is_migrate_isolate(mt)) { 1986 /* Obey watermarks as if the page was being allocated */ 1987 watermark = low_wmark_pages(zone) + (1 << order); 1988 if (!zone_watermark_ok(zone, 0, watermark, 0, 0)) 1989 return 0; 1990 1991 __mod_zone_freepage_state(zone, -(1UL << order), mt); 1992 } 1993 1994 /* Remove page from free list */ 1995 list_del(&page->lru); 1996 zone->free_area[order].nr_free--; 1997 rmv_page_order(page); 1998 1999 /* Set the pageblock if the isolated page is at least a pageblock */ 2000 if (order >= pageblock_order - 1) { 2001 struct page *endpage = page + (1 << order) - 1; 2002 for (; page < endpage; page += pageblock_nr_pages) { 2003 int mt = get_pageblock_migratetype(page); 2004 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) 2005 set_pageblock_migratetype(page, 2006 MIGRATE_MOVABLE); 2007 } 2008 } 2009 2010 set_page_owner(page, order, 0); 2011 return 1UL << order; 2012 } 2013 2014 /* 2015 * Similar to split_page except the page is already free. As this is only 2016 * being used for migration, the migratetype of the block also changes. 2017 * As this is called with interrupts disabled, the caller is responsible 2018 * for calling arch_alloc_page() and kernel_map_page() after interrupts 2019 * are enabled. 2020 * 2021 * Note: this is probably too low level an operation for use in drivers. 2022 * Please consult with lkml before using this in your driver. 2023 */ 2024 int split_free_page(struct page *page) 2025 { 2026 unsigned int order; 2027 int nr_pages; 2028 2029 order = page_order(page); 2030 2031 nr_pages = __isolate_free_page(page, order); 2032 if (!nr_pages) 2033 return 0; 2034 2035 /* Split into individual pages */ 2036 set_page_refcounted(page); 2037 split_page(page, order); 2038 return nr_pages; 2039 } 2040 2041 /* 2042 * Allocate a page from the given zone. Use pcplists for order-0 allocations. 2043 */ 2044 static inline 2045 struct page *buffered_rmqueue(struct zone *preferred_zone, 2046 struct zone *zone, unsigned int order, 2047 gfp_t gfp_flags, int migratetype) 2048 { 2049 unsigned long flags; 2050 struct page *page; 2051 bool cold = ((gfp_flags & __GFP_COLD) != 0); 2052 2053 if (likely(order == 0)) { 2054 struct per_cpu_pages *pcp; 2055 struct list_head *list; 2056 2057 local_irq_save(flags); 2058 pcp = &this_cpu_ptr(zone->pageset)->pcp; 2059 list = &pcp->lists[migratetype]; 2060 if (list_empty(list)) { 2061 pcp->count += rmqueue_bulk(zone, 0, 2062 pcp->batch, list, 2063 migratetype, cold); 2064 if (unlikely(list_empty(list))) 2065 goto failed; 2066 } 2067 2068 if (cold) 2069 page = list_entry(list->prev, struct page, lru); 2070 else 2071 page = list_entry(list->next, struct page, lru); 2072 2073 list_del(&page->lru); 2074 pcp->count--; 2075 } else { 2076 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 2077 /* 2078 * __GFP_NOFAIL is not to be used in new code. 2079 * 2080 * All __GFP_NOFAIL callers should be fixed so that they 2081 * properly detect and handle allocation failures. 2082 * 2083 * We most definitely don't want callers attempting to 2084 * allocate greater than order-1 page units with 2085 * __GFP_NOFAIL. 2086 */ 2087 WARN_ON_ONCE(order > 1); 2088 } 2089 spin_lock_irqsave(&zone->lock, flags); 2090 page = __rmqueue(zone, order, migratetype); 2091 spin_unlock(&zone->lock); 2092 if (!page) 2093 goto failed; 2094 __mod_zone_freepage_state(zone, -(1 << order), 2095 get_freepage_migratetype(page)); 2096 } 2097 2098 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order)); 2099 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 && 2100 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) 2101 set_bit(ZONE_FAIR_DEPLETED, &zone->flags); 2102 2103 __count_zone_vm_events(PGALLOC, zone, 1 << order); 2104 zone_statistics(preferred_zone, zone, gfp_flags); 2105 local_irq_restore(flags); 2106 2107 VM_BUG_ON_PAGE(bad_range(zone, page), page); 2108 return page; 2109 2110 failed: 2111 local_irq_restore(flags); 2112 return NULL; 2113 } 2114 2115 #ifdef CONFIG_FAIL_PAGE_ALLOC 2116 2117 static struct { 2118 struct fault_attr attr; 2119 2120 u32 ignore_gfp_highmem; 2121 u32 ignore_gfp_wait; 2122 u32 min_order; 2123 } fail_page_alloc = { 2124 .attr = FAULT_ATTR_INITIALIZER, 2125 .ignore_gfp_wait = 1, 2126 .ignore_gfp_highmem = 1, 2127 .min_order = 1, 2128 }; 2129 2130 static int __init setup_fail_page_alloc(char *str) 2131 { 2132 return setup_fault_attr(&fail_page_alloc.attr, str); 2133 } 2134 __setup("fail_page_alloc=", setup_fail_page_alloc); 2135 2136 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2137 { 2138 if (order < fail_page_alloc.min_order) 2139 return false; 2140 if (gfp_mask & __GFP_NOFAIL) 2141 return false; 2142 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 2143 return false; 2144 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 2145 return false; 2146 2147 return should_fail(&fail_page_alloc.attr, 1 << order); 2148 } 2149 2150 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 2151 2152 static int __init fail_page_alloc_debugfs(void) 2153 { 2154 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 2155 struct dentry *dir; 2156 2157 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 2158 &fail_page_alloc.attr); 2159 if (IS_ERR(dir)) 2160 return PTR_ERR(dir); 2161 2162 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 2163 &fail_page_alloc.ignore_gfp_wait)) 2164 goto fail; 2165 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 2166 &fail_page_alloc.ignore_gfp_highmem)) 2167 goto fail; 2168 if (!debugfs_create_u32("min-order", mode, dir, 2169 &fail_page_alloc.min_order)) 2170 goto fail; 2171 2172 return 0; 2173 fail: 2174 debugfs_remove_recursive(dir); 2175 2176 return -ENOMEM; 2177 } 2178 2179 late_initcall(fail_page_alloc_debugfs); 2180 2181 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 2182 2183 #else /* CONFIG_FAIL_PAGE_ALLOC */ 2184 2185 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2186 { 2187 return false; 2188 } 2189 2190 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 2191 2192 /* 2193 * Return true if free pages are above 'mark'. This takes into account the order 2194 * of the allocation. 2195 */ 2196 static bool __zone_watermark_ok(struct zone *z, unsigned int order, 2197 unsigned long mark, int classzone_idx, int alloc_flags, 2198 long free_pages) 2199 { 2200 /* free_pages may go negative - that's OK */ 2201 long min = mark; 2202 int o; 2203 long free_cma = 0; 2204 2205 free_pages -= (1 << order) - 1; 2206 if (alloc_flags & ALLOC_HIGH) 2207 min -= min / 2; 2208 if (alloc_flags & ALLOC_HARDER) 2209 min -= min / 4; 2210 #ifdef CONFIG_CMA 2211 /* If allocation can't use CMA areas don't use free CMA pages */ 2212 if (!(alloc_flags & ALLOC_CMA)) 2213 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES); 2214 #endif 2215 2216 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx]) 2217 return false; 2218 for (o = 0; o < order; o++) { 2219 /* At the next order, this order's pages become unavailable */ 2220 free_pages -= z->free_area[o].nr_free << o; 2221 2222 /* Require fewer higher order pages to be free */ 2223 min >>= 1; 2224 2225 if (free_pages <= min) 2226 return false; 2227 } 2228 return true; 2229 } 2230 2231 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, 2232 int classzone_idx, int alloc_flags) 2233 { 2234 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2235 zone_page_state(z, NR_FREE_PAGES)); 2236 } 2237 2238 bool zone_watermark_ok_safe(struct zone *z, unsigned int order, 2239 unsigned long mark, int classzone_idx, int alloc_flags) 2240 { 2241 long free_pages = zone_page_state(z, NR_FREE_PAGES); 2242 2243 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 2244 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 2245 2246 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2247 free_pages); 2248 } 2249 2250 #ifdef CONFIG_NUMA 2251 /* 2252 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 2253 * skip over zones that are not allowed by the cpuset, or that have 2254 * been recently (in last second) found to be nearly full. See further 2255 * comments in mmzone.h. Reduces cache footprint of zonelist scans 2256 * that have to skip over a lot of full or unallowed zones. 2257 * 2258 * If the zonelist cache is present in the passed zonelist, then 2259 * returns a pointer to the allowed node mask (either the current 2260 * tasks mems_allowed, or node_states[N_MEMORY].) 2261 * 2262 * If the zonelist cache is not available for this zonelist, does 2263 * nothing and returns NULL. 2264 * 2265 * If the fullzones BITMAP in the zonelist cache is stale (more than 2266 * a second since last zap'd) then we zap it out (clear its bits.) 2267 * 2268 * We hold off even calling zlc_setup, until after we've checked the 2269 * first zone in the zonelist, on the theory that most allocations will 2270 * be satisfied from that first zone, so best to examine that zone as 2271 * quickly as we can. 2272 */ 2273 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 2274 { 2275 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2276 nodemask_t *allowednodes; /* zonelist_cache approximation */ 2277 2278 zlc = zonelist->zlcache_ptr; 2279 if (!zlc) 2280 return NULL; 2281 2282 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 2283 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2284 zlc->last_full_zap = jiffies; 2285 } 2286 2287 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 2288 &cpuset_current_mems_allowed : 2289 &node_states[N_MEMORY]; 2290 return allowednodes; 2291 } 2292 2293 /* 2294 * Given 'z' scanning a zonelist, run a couple of quick checks to see 2295 * if it is worth looking at further for free memory: 2296 * 1) Check that the zone isn't thought to be full (doesn't have its 2297 * bit set in the zonelist_cache fullzones BITMAP). 2298 * 2) Check that the zones node (obtained from the zonelist_cache 2299 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 2300 * Return true (non-zero) if zone is worth looking at further, or 2301 * else return false (zero) if it is not. 2302 * 2303 * This check -ignores- the distinction between various watermarks, 2304 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 2305 * found to be full for any variation of these watermarks, it will 2306 * be considered full for up to one second by all requests, unless 2307 * we are so low on memory on all allowed nodes that we are forced 2308 * into the second scan of the zonelist. 2309 * 2310 * In the second scan we ignore this zonelist cache and exactly 2311 * apply the watermarks to all zones, even it is slower to do so. 2312 * We are low on memory in the second scan, and should leave no stone 2313 * unturned looking for a free page. 2314 */ 2315 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 2316 nodemask_t *allowednodes) 2317 { 2318 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2319 int i; /* index of *z in zonelist zones */ 2320 int n; /* node that zone *z is on */ 2321 2322 zlc = zonelist->zlcache_ptr; 2323 if (!zlc) 2324 return 1; 2325 2326 i = z - zonelist->_zonerefs; 2327 n = zlc->z_to_n[i]; 2328 2329 /* This zone is worth trying if it is allowed but not full */ 2330 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 2331 } 2332 2333 /* 2334 * Given 'z' scanning a zonelist, set the corresponding bit in 2335 * zlc->fullzones, so that subsequent attempts to allocate a page 2336 * from that zone don't waste time re-examining it. 2337 */ 2338 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 2339 { 2340 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2341 int i; /* index of *z in zonelist zones */ 2342 2343 zlc = zonelist->zlcache_ptr; 2344 if (!zlc) 2345 return; 2346 2347 i = z - zonelist->_zonerefs; 2348 2349 set_bit(i, zlc->fullzones); 2350 } 2351 2352 /* 2353 * clear all zones full, called after direct reclaim makes progress so that 2354 * a zone that was recently full is not skipped over for up to a second 2355 */ 2356 static void zlc_clear_zones_full(struct zonelist *zonelist) 2357 { 2358 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2359 2360 zlc = zonelist->zlcache_ptr; 2361 if (!zlc) 2362 return; 2363 2364 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2365 } 2366 2367 static bool zone_local(struct zone *local_zone, struct zone *zone) 2368 { 2369 return local_zone->node == zone->node; 2370 } 2371 2372 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2373 { 2374 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) < 2375 RECLAIM_DISTANCE; 2376 } 2377 2378 #else /* CONFIG_NUMA */ 2379 2380 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 2381 { 2382 return NULL; 2383 } 2384 2385 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 2386 nodemask_t *allowednodes) 2387 { 2388 return 1; 2389 } 2390 2391 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 2392 { 2393 } 2394 2395 static void zlc_clear_zones_full(struct zonelist *zonelist) 2396 { 2397 } 2398 2399 static bool zone_local(struct zone *local_zone, struct zone *zone) 2400 { 2401 return true; 2402 } 2403 2404 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2405 { 2406 return true; 2407 } 2408 2409 #endif /* CONFIG_NUMA */ 2410 2411 static void reset_alloc_batches(struct zone *preferred_zone) 2412 { 2413 struct zone *zone = preferred_zone->zone_pgdat->node_zones; 2414 2415 do { 2416 mod_zone_page_state(zone, NR_ALLOC_BATCH, 2417 high_wmark_pages(zone) - low_wmark_pages(zone) - 2418 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 2419 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags); 2420 } while (zone++ != preferred_zone); 2421 } 2422 2423 /* 2424 * get_page_from_freelist goes through the zonelist trying to allocate 2425 * a page. 2426 */ 2427 static struct page * 2428 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, 2429 const struct alloc_context *ac) 2430 { 2431 struct zonelist *zonelist = ac->zonelist; 2432 struct zoneref *z; 2433 struct page *page = NULL; 2434 struct zone *zone; 2435 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 2436 int zlc_active = 0; /* set if using zonelist_cache */ 2437 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 2438 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) && 2439 (gfp_mask & __GFP_WRITE); 2440 int nr_fair_skipped = 0; 2441 bool zonelist_rescan; 2442 2443 zonelist_scan: 2444 zonelist_rescan = false; 2445 2446 /* 2447 * Scan zonelist, looking for a zone with enough free. 2448 * See also __cpuset_node_allowed() comment in kernel/cpuset.c. 2449 */ 2450 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx, 2451 ac->nodemask) { 2452 unsigned long mark; 2453 2454 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 2455 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 2456 continue; 2457 if (cpusets_enabled() && 2458 (alloc_flags & ALLOC_CPUSET) && 2459 !cpuset_zone_allowed(zone, gfp_mask)) 2460 continue; 2461 /* 2462 * Distribute pages in proportion to the individual 2463 * zone size to ensure fair page aging. The zone a 2464 * page was allocated in should have no effect on the 2465 * time the page has in memory before being reclaimed. 2466 */ 2467 if (alloc_flags & ALLOC_FAIR) { 2468 if (!zone_local(ac->preferred_zone, zone)) 2469 break; 2470 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) { 2471 nr_fair_skipped++; 2472 continue; 2473 } 2474 } 2475 /* 2476 * When allocating a page cache page for writing, we 2477 * want to get it from a zone that is within its dirty 2478 * limit, such that no single zone holds more than its 2479 * proportional share of globally allowed dirty pages. 2480 * The dirty limits take into account the zone's 2481 * lowmem reserves and high watermark so that kswapd 2482 * should be able to balance it without having to 2483 * write pages from its LRU list. 2484 * 2485 * This may look like it could increase pressure on 2486 * lower zones by failing allocations in higher zones 2487 * before they are full. But the pages that do spill 2488 * over are limited as the lower zones are protected 2489 * by this very same mechanism. It should not become 2490 * a practical burden to them. 2491 * 2492 * XXX: For now, allow allocations to potentially 2493 * exceed the per-zone dirty limit in the slowpath 2494 * (ALLOC_WMARK_LOW unset) before going into reclaim, 2495 * which is important when on a NUMA setup the allowed 2496 * zones are together not big enough to reach the 2497 * global limit. The proper fix for these situations 2498 * will require awareness of zones in the 2499 * dirty-throttling and the flusher threads. 2500 */ 2501 if (consider_zone_dirty && !zone_dirty_ok(zone)) 2502 continue; 2503 2504 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 2505 if (!zone_watermark_ok(zone, order, mark, 2506 ac->classzone_idx, alloc_flags)) { 2507 int ret; 2508 2509 /* Checked here to keep the fast path fast */ 2510 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 2511 if (alloc_flags & ALLOC_NO_WATERMARKS) 2512 goto try_this_zone; 2513 2514 if (IS_ENABLED(CONFIG_NUMA) && 2515 !did_zlc_setup && nr_online_nodes > 1) { 2516 /* 2517 * we do zlc_setup if there are multiple nodes 2518 * and before considering the first zone allowed 2519 * by the cpuset. 2520 */ 2521 allowednodes = zlc_setup(zonelist, alloc_flags); 2522 zlc_active = 1; 2523 did_zlc_setup = 1; 2524 } 2525 2526 if (zone_reclaim_mode == 0 || 2527 !zone_allows_reclaim(ac->preferred_zone, zone)) 2528 goto this_zone_full; 2529 2530 /* 2531 * As we may have just activated ZLC, check if the first 2532 * eligible zone has failed zone_reclaim recently. 2533 */ 2534 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 2535 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 2536 continue; 2537 2538 ret = zone_reclaim(zone, gfp_mask, order); 2539 switch (ret) { 2540 case ZONE_RECLAIM_NOSCAN: 2541 /* did not scan */ 2542 continue; 2543 case ZONE_RECLAIM_FULL: 2544 /* scanned but unreclaimable */ 2545 continue; 2546 default: 2547 /* did we reclaim enough */ 2548 if (zone_watermark_ok(zone, order, mark, 2549 ac->classzone_idx, alloc_flags)) 2550 goto try_this_zone; 2551 2552 /* 2553 * Failed to reclaim enough to meet watermark. 2554 * Only mark the zone full if checking the min 2555 * watermark or if we failed to reclaim just 2556 * 1<<order pages or else the page allocator 2557 * fastpath will prematurely mark zones full 2558 * when the watermark is between the low and 2559 * min watermarks. 2560 */ 2561 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) || 2562 ret == ZONE_RECLAIM_SOME) 2563 goto this_zone_full; 2564 2565 continue; 2566 } 2567 } 2568 2569 try_this_zone: 2570 page = buffered_rmqueue(ac->preferred_zone, zone, order, 2571 gfp_mask, ac->migratetype); 2572 if (page) { 2573 if (prep_new_page(page, order, gfp_mask, alloc_flags)) 2574 goto try_this_zone; 2575 return page; 2576 } 2577 this_zone_full: 2578 if (IS_ENABLED(CONFIG_NUMA) && zlc_active) 2579 zlc_mark_zone_full(zonelist, z); 2580 } 2581 2582 /* 2583 * The first pass makes sure allocations are spread fairly within the 2584 * local node. However, the local node might have free pages left 2585 * after the fairness batches are exhausted, and remote zones haven't 2586 * even been considered yet. Try once more without fairness, and 2587 * include remote zones now, before entering the slowpath and waking 2588 * kswapd: prefer spilling to a remote zone over swapping locally. 2589 */ 2590 if (alloc_flags & ALLOC_FAIR) { 2591 alloc_flags &= ~ALLOC_FAIR; 2592 if (nr_fair_skipped) { 2593 zonelist_rescan = true; 2594 reset_alloc_batches(ac->preferred_zone); 2595 } 2596 if (nr_online_nodes > 1) 2597 zonelist_rescan = true; 2598 } 2599 2600 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) { 2601 /* Disable zlc cache for second zonelist scan */ 2602 zlc_active = 0; 2603 zonelist_rescan = true; 2604 } 2605 2606 if (zonelist_rescan) 2607 goto zonelist_scan; 2608 2609 return NULL; 2610 } 2611 2612 /* 2613 * Large machines with many possible nodes should not always dump per-node 2614 * meminfo in irq context. 2615 */ 2616 static inline bool should_suppress_show_mem(void) 2617 { 2618 bool ret = false; 2619 2620 #if NODES_SHIFT > 8 2621 ret = in_interrupt(); 2622 #endif 2623 return ret; 2624 } 2625 2626 static DEFINE_RATELIMIT_STATE(nopage_rs, 2627 DEFAULT_RATELIMIT_INTERVAL, 2628 DEFAULT_RATELIMIT_BURST); 2629 2630 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 2631 { 2632 unsigned int filter = SHOW_MEM_FILTER_NODES; 2633 2634 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 2635 debug_guardpage_minorder() > 0) 2636 return; 2637 2638 /* 2639 * This documents exceptions given to allocations in certain 2640 * contexts that are allowed to allocate outside current's set 2641 * of allowed nodes. 2642 */ 2643 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2644 if (test_thread_flag(TIF_MEMDIE) || 2645 (current->flags & (PF_MEMALLOC | PF_EXITING))) 2646 filter &= ~SHOW_MEM_FILTER_NODES; 2647 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 2648 filter &= ~SHOW_MEM_FILTER_NODES; 2649 2650 if (fmt) { 2651 struct va_format vaf; 2652 va_list args; 2653 2654 va_start(args, fmt); 2655 2656 vaf.fmt = fmt; 2657 vaf.va = &args; 2658 2659 pr_warn("%pV", &vaf); 2660 2661 va_end(args); 2662 } 2663 2664 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 2665 current->comm, order, gfp_mask); 2666 2667 dump_stack(); 2668 if (!should_suppress_show_mem()) 2669 show_mem(filter); 2670 } 2671 2672 static inline struct page * 2673 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 2674 const struct alloc_context *ac, unsigned long *did_some_progress) 2675 { 2676 struct page *page; 2677 2678 *did_some_progress = 0; 2679 2680 /* 2681 * Acquire the oom lock. If that fails, somebody else is 2682 * making progress for us. 2683 */ 2684 if (!mutex_trylock(&oom_lock)) { 2685 *did_some_progress = 1; 2686 schedule_timeout_uninterruptible(1); 2687 return NULL; 2688 } 2689 2690 /* 2691 * Go through the zonelist yet one more time, keep very high watermark 2692 * here, this is only to catch a parallel oom killing, we must fail if 2693 * we're still under heavy pressure. 2694 */ 2695 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order, 2696 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); 2697 if (page) 2698 goto out; 2699 2700 if (!(gfp_mask & __GFP_NOFAIL)) { 2701 /* Coredumps can quickly deplete all memory reserves */ 2702 if (current->flags & PF_DUMPCORE) 2703 goto out; 2704 /* The OOM killer will not help higher order allocs */ 2705 if (order > PAGE_ALLOC_COSTLY_ORDER) 2706 goto out; 2707 /* The OOM killer does not needlessly kill tasks for lowmem */ 2708 if (ac->high_zoneidx < ZONE_NORMAL) 2709 goto out; 2710 /* The OOM killer does not compensate for IO-less reclaim */ 2711 if (!(gfp_mask & __GFP_FS)) { 2712 /* 2713 * XXX: Page reclaim didn't yield anything, 2714 * and the OOM killer can't be invoked, but 2715 * keep looping as per tradition. 2716 */ 2717 *did_some_progress = 1; 2718 goto out; 2719 } 2720 if (pm_suspended_storage()) 2721 goto out; 2722 /* The OOM killer may not free memory on a specific node */ 2723 if (gfp_mask & __GFP_THISNODE) 2724 goto out; 2725 } 2726 /* Exhausted what can be done so it's blamo time */ 2727 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false) 2728 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) 2729 *did_some_progress = 1; 2730 out: 2731 mutex_unlock(&oom_lock); 2732 return page; 2733 } 2734 2735 #ifdef CONFIG_COMPACTION 2736 /* Try memory compaction for high-order allocations before reclaim */ 2737 static struct page * 2738 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2739 int alloc_flags, const struct alloc_context *ac, 2740 enum migrate_mode mode, int *contended_compaction, 2741 bool *deferred_compaction) 2742 { 2743 unsigned long compact_result; 2744 struct page *page; 2745 2746 if (!order) 2747 return NULL; 2748 2749 current->flags |= PF_MEMALLOC; 2750 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, 2751 mode, contended_compaction); 2752 current->flags &= ~PF_MEMALLOC; 2753 2754 switch (compact_result) { 2755 case COMPACT_DEFERRED: 2756 *deferred_compaction = true; 2757 /* fall-through */ 2758 case COMPACT_SKIPPED: 2759 return NULL; 2760 default: 2761 break; 2762 } 2763 2764 /* 2765 * At least in one zone compaction wasn't deferred or skipped, so let's 2766 * count a compaction stall 2767 */ 2768 count_vm_event(COMPACTSTALL); 2769 2770 page = get_page_from_freelist(gfp_mask, order, 2771 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 2772 2773 if (page) { 2774 struct zone *zone = page_zone(page); 2775 2776 zone->compact_blockskip_flush = false; 2777 compaction_defer_reset(zone, order, true); 2778 count_vm_event(COMPACTSUCCESS); 2779 return page; 2780 } 2781 2782 /* 2783 * It's bad if compaction run occurs and fails. The most likely reason 2784 * is that pages exist, but not enough to satisfy watermarks. 2785 */ 2786 count_vm_event(COMPACTFAIL); 2787 2788 cond_resched(); 2789 2790 return NULL; 2791 } 2792 #else 2793 static inline struct page * 2794 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2795 int alloc_flags, const struct alloc_context *ac, 2796 enum migrate_mode mode, int *contended_compaction, 2797 bool *deferred_compaction) 2798 { 2799 return NULL; 2800 } 2801 #endif /* CONFIG_COMPACTION */ 2802 2803 /* Perform direct synchronous page reclaim */ 2804 static int 2805 __perform_reclaim(gfp_t gfp_mask, unsigned int order, 2806 const struct alloc_context *ac) 2807 { 2808 struct reclaim_state reclaim_state; 2809 int progress; 2810 2811 cond_resched(); 2812 2813 /* We now go into synchronous reclaim */ 2814 cpuset_memory_pressure_bump(); 2815 current->flags |= PF_MEMALLOC; 2816 lockdep_set_current_reclaim_state(gfp_mask); 2817 reclaim_state.reclaimed_slab = 0; 2818 current->reclaim_state = &reclaim_state; 2819 2820 progress = try_to_free_pages(ac->zonelist, order, gfp_mask, 2821 ac->nodemask); 2822 2823 current->reclaim_state = NULL; 2824 lockdep_clear_current_reclaim_state(); 2825 current->flags &= ~PF_MEMALLOC; 2826 2827 cond_resched(); 2828 2829 return progress; 2830 } 2831 2832 /* The really slow allocator path where we enter direct reclaim */ 2833 static inline struct page * 2834 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 2835 int alloc_flags, const struct alloc_context *ac, 2836 unsigned long *did_some_progress) 2837 { 2838 struct page *page = NULL; 2839 bool drained = false; 2840 2841 *did_some_progress = __perform_reclaim(gfp_mask, order, ac); 2842 if (unlikely(!(*did_some_progress))) 2843 return NULL; 2844 2845 /* After successful reclaim, reconsider all zones for allocation */ 2846 if (IS_ENABLED(CONFIG_NUMA)) 2847 zlc_clear_zones_full(ac->zonelist); 2848 2849 retry: 2850 page = get_page_from_freelist(gfp_mask, order, 2851 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 2852 2853 /* 2854 * If an allocation failed after direct reclaim, it could be because 2855 * pages are pinned on the per-cpu lists. Drain them and try again 2856 */ 2857 if (!page && !drained) { 2858 drain_all_pages(NULL); 2859 drained = true; 2860 goto retry; 2861 } 2862 2863 return page; 2864 } 2865 2866 /* 2867 * This is called in the allocator slow-path if the allocation request is of 2868 * sufficient urgency to ignore watermarks and take other desperate measures 2869 */ 2870 static inline struct page * 2871 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2872 const struct alloc_context *ac) 2873 { 2874 struct page *page; 2875 2876 do { 2877 page = get_page_from_freelist(gfp_mask, order, 2878 ALLOC_NO_WATERMARKS, ac); 2879 2880 if (!page && gfp_mask & __GFP_NOFAIL) 2881 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, 2882 HZ/50); 2883 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2884 2885 return page; 2886 } 2887 2888 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac) 2889 { 2890 struct zoneref *z; 2891 struct zone *zone; 2892 2893 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 2894 ac->high_zoneidx, ac->nodemask) 2895 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone)); 2896 } 2897 2898 static inline int 2899 gfp_to_alloc_flags(gfp_t gfp_mask) 2900 { 2901 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2902 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD)); 2903 2904 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2905 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2906 2907 /* 2908 * The caller may dip into page reserves a bit more if the caller 2909 * cannot run direct reclaim, or if the caller has realtime scheduling 2910 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2911 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH). 2912 */ 2913 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2914 2915 if (atomic) { 2916 /* 2917 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even 2918 * if it can't schedule. 2919 */ 2920 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2921 alloc_flags |= ALLOC_HARDER; 2922 /* 2923 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the 2924 * comment for __cpuset_node_allowed(). 2925 */ 2926 alloc_flags &= ~ALLOC_CPUSET; 2927 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2928 alloc_flags |= ALLOC_HARDER; 2929 2930 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2931 if (gfp_mask & __GFP_MEMALLOC) 2932 alloc_flags |= ALLOC_NO_WATERMARKS; 2933 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) 2934 alloc_flags |= ALLOC_NO_WATERMARKS; 2935 else if (!in_interrupt() && 2936 ((current->flags & PF_MEMALLOC) || 2937 unlikely(test_thread_flag(TIF_MEMDIE)))) 2938 alloc_flags |= ALLOC_NO_WATERMARKS; 2939 } 2940 #ifdef CONFIG_CMA 2941 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) 2942 alloc_flags |= ALLOC_CMA; 2943 #endif 2944 return alloc_flags; 2945 } 2946 2947 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) 2948 { 2949 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); 2950 } 2951 2952 static inline struct page * 2953 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2954 struct alloc_context *ac) 2955 { 2956 const gfp_t wait = gfp_mask & __GFP_WAIT; 2957 struct page *page = NULL; 2958 int alloc_flags; 2959 unsigned long pages_reclaimed = 0; 2960 unsigned long did_some_progress; 2961 enum migrate_mode migration_mode = MIGRATE_ASYNC; 2962 bool deferred_compaction = false; 2963 int contended_compaction = COMPACT_CONTENDED_NONE; 2964 2965 /* 2966 * In the slowpath, we sanity check order to avoid ever trying to 2967 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2968 * be using allocators in order of preference for an area that is 2969 * too large. 2970 */ 2971 if (order >= MAX_ORDER) { 2972 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2973 return NULL; 2974 } 2975 2976 /* 2977 * If this allocation cannot block and it is for a specific node, then 2978 * fail early. There's no need to wakeup kswapd or retry for a 2979 * speculative node-specific allocation. 2980 */ 2981 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait) 2982 goto nopage; 2983 2984 retry: 2985 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2986 wake_all_kswapds(order, ac); 2987 2988 /* 2989 * OK, we're below the kswapd watermark and have kicked background 2990 * reclaim. Now things get more complex, so set up alloc_flags according 2991 * to how we want to proceed. 2992 */ 2993 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2994 2995 /* 2996 * Find the true preferred zone if the allocation is unconstrained by 2997 * cpusets. 2998 */ 2999 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) { 3000 struct zoneref *preferred_zoneref; 3001 preferred_zoneref = first_zones_zonelist(ac->zonelist, 3002 ac->high_zoneidx, NULL, &ac->preferred_zone); 3003 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref); 3004 } 3005 3006 /* This is the last chance, in general, before the goto nopage. */ 3007 page = get_page_from_freelist(gfp_mask, order, 3008 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 3009 if (page) 3010 goto got_pg; 3011 3012 /* Allocate without watermarks if the context allows */ 3013 if (alloc_flags & ALLOC_NO_WATERMARKS) { 3014 /* 3015 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds 3016 * the allocation is high priority and these type of 3017 * allocations are system rather than user orientated 3018 */ 3019 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask); 3020 3021 page = __alloc_pages_high_priority(gfp_mask, order, ac); 3022 3023 if (page) { 3024 goto got_pg; 3025 } 3026 } 3027 3028 /* Atomic allocations - we can't balance anything */ 3029 if (!wait) { 3030 /* 3031 * All existing users of the deprecated __GFP_NOFAIL are 3032 * blockable, so warn of any new users that actually allow this 3033 * type of allocation to fail. 3034 */ 3035 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL); 3036 goto nopage; 3037 } 3038 3039 /* Avoid recursion of direct reclaim */ 3040 if (current->flags & PF_MEMALLOC) 3041 goto nopage; 3042 3043 /* Avoid allocations with no watermarks from looping endlessly */ 3044 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 3045 goto nopage; 3046 3047 /* 3048 * Try direct compaction. The first pass is asynchronous. Subsequent 3049 * attempts after direct reclaim are synchronous 3050 */ 3051 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, 3052 migration_mode, 3053 &contended_compaction, 3054 &deferred_compaction); 3055 if (page) 3056 goto got_pg; 3057 3058 /* Checks for THP-specific high-order allocations */ 3059 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) { 3060 /* 3061 * If compaction is deferred for high-order allocations, it is 3062 * because sync compaction recently failed. If this is the case 3063 * and the caller requested a THP allocation, we do not want 3064 * to heavily disrupt the system, so we fail the allocation 3065 * instead of entering direct reclaim. 3066 */ 3067 if (deferred_compaction) 3068 goto nopage; 3069 3070 /* 3071 * In all zones where compaction was attempted (and not 3072 * deferred or skipped), lock contention has been detected. 3073 * For THP allocation we do not want to disrupt the others 3074 * so we fallback to base pages instead. 3075 */ 3076 if (contended_compaction == COMPACT_CONTENDED_LOCK) 3077 goto nopage; 3078 3079 /* 3080 * If compaction was aborted due to need_resched(), we do not 3081 * want to further increase allocation latency, unless it is 3082 * khugepaged trying to collapse. 3083 */ 3084 if (contended_compaction == COMPACT_CONTENDED_SCHED 3085 && !(current->flags & PF_KTHREAD)) 3086 goto nopage; 3087 } 3088 3089 /* 3090 * It can become very expensive to allocate transparent hugepages at 3091 * fault, so use asynchronous memory compaction for THP unless it is 3092 * khugepaged trying to collapse. 3093 */ 3094 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE || 3095 (current->flags & PF_KTHREAD)) 3096 migration_mode = MIGRATE_SYNC_LIGHT; 3097 3098 /* Try direct reclaim and then allocating */ 3099 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, 3100 &did_some_progress); 3101 if (page) 3102 goto got_pg; 3103 3104 /* Do not loop if specifically requested */ 3105 if (gfp_mask & __GFP_NORETRY) 3106 goto noretry; 3107 3108 /* Keep reclaiming pages as long as there is reasonable progress */ 3109 pages_reclaimed += did_some_progress; 3110 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) || 3111 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) { 3112 /* Wait for some write requests to complete then retry */ 3113 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50); 3114 goto retry; 3115 } 3116 3117 /* Reclaim has failed us, start killing things */ 3118 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress); 3119 if (page) 3120 goto got_pg; 3121 3122 /* Retry as long as the OOM killer is making progress */ 3123 if (did_some_progress) 3124 goto retry; 3125 3126 noretry: 3127 /* 3128 * High-order allocations do not necessarily loop after 3129 * direct reclaim and reclaim/compaction depends on compaction 3130 * being called after reclaim so call directly if necessary 3131 */ 3132 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, 3133 ac, migration_mode, 3134 &contended_compaction, 3135 &deferred_compaction); 3136 if (page) 3137 goto got_pg; 3138 nopage: 3139 warn_alloc_failed(gfp_mask, order, NULL); 3140 got_pg: 3141 return page; 3142 } 3143 3144 /* 3145 * This is the 'heart' of the zoned buddy allocator. 3146 */ 3147 struct page * 3148 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 3149 struct zonelist *zonelist, nodemask_t *nodemask) 3150 { 3151 struct zoneref *preferred_zoneref; 3152 struct page *page = NULL; 3153 unsigned int cpuset_mems_cookie; 3154 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR; 3155 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */ 3156 struct alloc_context ac = { 3157 .high_zoneidx = gfp_zone(gfp_mask), 3158 .nodemask = nodemask, 3159 .migratetype = gfpflags_to_migratetype(gfp_mask), 3160 }; 3161 3162 gfp_mask &= gfp_allowed_mask; 3163 3164 lockdep_trace_alloc(gfp_mask); 3165 3166 might_sleep_if(gfp_mask & __GFP_WAIT); 3167 3168 if (should_fail_alloc_page(gfp_mask, order)) 3169 return NULL; 3170 3171 /* 3172 * Check the zones suitable for the gfp_mask contain at least one 3173 * valid zone. It's possible to have an empty zonelist as a result 3174 * of __GFP_THISNODE and a memoryless node 3175 */ 3176 if (unlikely(!zonelist->_zonerefs->zone)) 3177 return NULL; 3178 3179 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE) 3180 alloc_flags |= ALLOC_CMA; 3181 3182 retry_cpuset: 3183 cpuset_mems_cookie = read_mems_allowed_begin(); 3184 3185 /* We set it here, as __alloc_pages_slowpath might have changed it */ 3186 ac.zonelist = zonelist; 3187 /* The preferred zone is used for statistics later */ 3188 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx, 3189 ac.nodemask ? : &cpuset_current_mems_allowed, 3190 &ac.preferred_zone); 3191 if (!ac.preferred_zone) 3192 goto out; 3193 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref); 3194 3195 /* First allocation attempt */ 3196 alloc_mask = gfp_mask|__GFP_HARDWALL; 3197 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac); 3198 if (unlikely(!page)) { 3199 /* 3200 * Runtime PM, block IO and its error handling path 3201 * can deadlock because I/O on the device might not 3202 * complete. 3203 */ 3204 alloc_mask = memalloc_noio_flags(gfp_mask); 3205 3206 page = __alloc_pages_slowpath(alloc_mask, order, &ac); 3207 } 3208 3209 if (kmemcheck_enabled && page) 3210 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 3211 3212 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype); 3213 3214 out: 3215 /* 3216 * When updating a task's mems_allowed, it is possible to race with 3217 * parallel threads in such a way that an allocation can fail while 3218 * the mask is being updated. If a page allocation is about to fail, 3219 * check if the cpuset changed during allocation and if so, retry. 3220 */ 3221 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 3222 goto retry_cpuset; 3223 3224 return page; 3225 } 3226 EXPORT_SYMBOL(__alloc_pages_nodemask); 3227 3228 /* 3229 * Common helper functions. 3230 */ 3231 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 3232 { 3233 struct page *page; 3234 3235 /* 3236 * __get_free_pages() returns a 32-bit address, which cannot represent 3237 * a highmem page 3238 */ 3239 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 3240 3241 page = alloc_pages(gfp_mask, order); 3242 if (!page) 3243 return 0; 3244 return (unsigned long) page_address(page); 3245 } 3246 EXPORT_SYMBOL(__get_free_pages); 3247 3248 unsigned long get_zeroed_page(gfp_t gfp_mask) 3249 { 3250 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 3251 } 3252 EXPORT_SYMBOL(get_zeroed_page); 3253 3254 void __free_pages(struct page *page, unsigned int order) 3255 { 3256 if (put_page_testzero(page)) { 3257 if (order == 0) 3258 free_hot_cold_page(page, false); 3259 else 3260 __free_pages_ok(page, order); 3261 } 3262 } 3263 3264 EXPORT_SYMBOL(__free_pages); 3265 3266 void free_pages(unsigned long addr, unsigned int order) 3267 { 3268 if (addr != 0) { 3269 VM_BUG_ON(!virt_addr_valid((void *)addr)); 3270 __free_pages(virt_to_page((void *)addr), order); 3271 } 3272 } 3273 3274 EXPORT_SYMBOL(free_pages); 3275 3276 /* 3277 * Page Fragment: 3278 * An arbitrary-length arbitrary-offset area of memory which resides 3279 * within a 0 or higher order page. Multiple fragments within that page 3280 * are individually refcounted, in the page's reference counter. 3281 * 3282 * The page_frag functions below provide a simple allocation framework for 3283 * page fragments. This is used by the network stack and network device 3284 * drivers to provide a backing region of memory for use as either an 3285 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info. 3286 */ 3287 static struct page *__page_frag_refill(struct page_frag_cache *nc, 3288 gfp_t gfp_mask) 3289 { 3290 struct page *page = NULL; 3291 gfp_t gfp = gfp_mask; 3292 3293 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3294 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY | 3295 __GFP_NOMEMALLOC; 3296 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, 3297 PAGE_FRAG_CACHE_MAX_ORDER); 3298 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE; 3299 #endif 3300 if (unlikely(!page)) 3301 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0); 3302 3303 nc->va = page ? page_address(page) : NULL; 3304 3305 return page; 3306 } 3307 3308 void *__alloc_page_frag(struct page_frag_cache *nc, 3309 unsigned int fragsz, gfp_t gfp_mask) 3310 { 3311 unsigned int size = PAGE_SIZE; 3312 struct page *page; 3313 int offset; 3314 3315 if (unlikely(!nc->va)) { 3316 refill: 3317 page = __page_frag_refill(nc, gfp_mask); 3318 if (!page) 3319 return NULL; 3320 3321 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3322 /* if size can vary use size else just use PAGE_SIZE */ 3323 size = nc->size; 3324 #endif 3325 /* Even if we own the page, we do not use atomic_set(). 3326 * This would break get_page_unless_zero() users. 3327 */ 3328 atomic_add(size - 1, &page->_count); 3329 3330 /* reset page count bias and offset to start of new frag */ 3331 nc->pfmemalloc = page->pfmemalloc; 3332 nc->pagecnt_bias = size; 3333 nc->offset = size; 3334 } 3335 3336 offset = nc->offset - fragsz; 3337 if (unlikely(offset < 0)) { 3338 page = virt_to_page(nc->va); 3339 3340 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count)) 3341 goto refill; 3342 3343 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3344 /* if size can vary use size else just use PAGE_SIZE */ 3345 size = nc->size; 3346 #endif 3347 /* OK, page count is 0, we can safely set it */ 3348 atomic_set(&page->_count, size); 3349 3350 /* reset page count bias and offset to start of new frag */ 3351 nc->pagecnt_bias = size; 3352 offset = size - fragsz; 3353 } 3354 3355 nc->pagecnt_bias--; 3356 nc->offset = offset; 3357 3358 return nc->va + offset; 3359 } 3360 EXPORT_SYMBOL(__alloc_page_frag); 3361 3362 /* 3363 * Frees a page fragment allocated out of either a compound or order 0 page. 3364 */ 3365 void __free_page_frag(void *addr) 3366 { 3367 struct page *page = virt_to_head_page(addr); 3368 3369 if (unlikely(put_page_testzero(page))) 3370 __free_pages_ok(page, compound_order(page)); 3371 } 3372 EXPORT_SYMBOL(__free_page_frag); 3373 3374 /* 3375 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter 3376 * of the current memory cgroup. 3377 * 3378 * It should be used when the caller would like to use kmalloc, but since the 3379 * allocation is large, it has to fall back to the page allocator. 3380 */ 3381 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order) 3382 { 3383 struct page *page; 3384 struct mem_cgroup *memcg = NULL; 3385 3386 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 3387 return NULL; 3388 page = alloc_pages(gfp_mask, order); 3389 memcg_kmem_commit_charge(page, memcg, order); 3390 return page; 3391 } 3392 3393 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order) 3394 { 3395 struct page *page; 3396 struct mem_cgroup *memcg = NULL; 3397 3398 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 3399 return NULL; 3400 page = alloc_pages_node(nid, gfp_mask, order); 3401 memcg_kmem_commit_charge(page, memcg, order); 3402 return page; 3403 } 3404 3405 /* 3406 * __free_kmem_pages and free_kmem_pages will free pages allocated with 3407 * alloc_kmem_pages. 3408 */ 3409 void __free_kmem_pages(struct page *page, unsigned int order) 3410 { 3411 memcg_kmem_uncharge_pages(page, order); 3412 __free_pages(page, order); 3413 } 3414 3415 void free_kmem_pages(unsigned long addr, unsigned int order) 3416 { 3417 if (addr != 0) { 3418 VM_BUG_ON(!virt_addr_valid((void *)addr)); 3419 __free_kmem_pages(virt_to_page((void *)addr), order); 3420 } 3421 } 3422 3423 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 3424 { 3425 if (addr) { 3426 unsigned long alloc_end = addr + (PAGE_SIZE << order); 3427 unsigned long used = addr + PAGE_ALIGN(size); 3428 3429 split_page(virt_to_page((void *)addr), order); 3430 while (used < alloc_end) { 3431 free_page(used); 3432 used += PAGE_SIZE; 3433 } 3434 } 3435 return (void *)addr; 3436 } 3437 3438 /** 3439 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 3440 * @size: the number of bytes to allocate 3441 * @gfp_mask: GFP flags for the allocation 3442 * 3443 * This function is similar to alloc_pages(), except that it allocates the 3444 * minimum number of pages to satisfy the request. alloc_pages() can only 3445 * allocate memory in power-of-two pages. 3446 * 3447 * This function is also limited by MAX_ORDER. 3448 * 3449 * Memory allocated by this function must be released by free_pages_exact(). 3450 */ 3451 void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 3452 { 3453 unsigned int order = get_order(size); 3454 unsigned long addr; 3455 3456 addr = __get_free_pages(gfp_mask, order); 3457 return make_alloc_exact(addr, order, size); 3458 } 3459 EXPORT_SYMBOL(alloc_pages_exact); 3460 3461 /** 3462 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 3463 * pages on a node. 3464 * @nid: the preferred node ID where memory should be allocated 3465 * @size: the number of bytes to allocate 3466 * @gfp_mask: GFP flags for the allocation 3467 * 3468 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 3469 * back. 3470 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 3471 * but is not exact. 3472 */ 3473 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 3474 { 3475 unsigned order = get_order(size); 3476 struct page *p = alloc_pages_node(nid, gfp_mask, order); 3477 if (!p) 3478 return NULL; 3479 return make_alloc_exact((unsigned long)page_address(p), order, size); 3480 } 3481 3482 /** 3483 * free_pages_exact - release memory allocated via alloc_pages_exact() 3484 * @virt: the value returned by alloc_pages_exact. 3485 * @size: size of allocation, same value as passed to alloc_pages_exact(). 3486 * 3487 * Release the memory allocated by a previous call to alloc_pages_exact. 3488 */ 3489 void free_pages_exact(void *virt, size_t size) 3490 { 3491 unsigned long addr = (unsigned long)virt; 3492 unsigned long end = addr + PAGE_ALIGN(size); 3493 3494 while (addr < end) { 3495 free_page(addr); 3496 addr += PAGE_SIZE; 3497 } 3498 } 3499 EXPORT_SYMBOL(free_pages_exact); 3500 3501 /** 3502 * nr_free_zone_pages - count number of pages beyond high watermark 3503 * @offset: The zone index of the highest zone 3504 * 3505 * nr_free_zone_pages() counts the number of counts pages which are beyond the 3506 * high watermark within all zones at or below a given zone index. For each 3507 * zone, the number of pages is calculated as: 3508 * managed_pages - high_pages 3509 */ 3510 static unsigned long nr_free_zone_pages(int offset) 3511 { 3512 struct zoneref *z; 3513 struct zone *zone; 3514 3515 /* Just pick one node, since fallback list is circular */ 3516 unsigned long sum = 0; 3517 3518 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 3519 3520 for_each_zone_zonelist(zone, z, zonelist, offset) { 3521 unsigned long size = zone->managed_pages; 3522 unsigned long high = high_wmark_pages(zone); 3523 if (size > high) 3524 sum += size - high; 3525 } 3526 3527 return sum; 3528 } 3529 3530 /** 3531 * nr_free_buffer_pages - count number of pages beyond high watermark 3532 * 3533 * nr_free_buffer_pages() counts the number of pages which are beyond the high 3534 * watermark within ZONE_DMA and ZONE_NORMAL. 3535 */ 3536 unsigned long nr_free_buffer_pages(void) 3537 { 3538 return nr_free_zone_pages(gfp_zone(GFP_USER)); 3539 } 3540 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 3541 3542 /** 3543 * nr_free_pagecache_pages - count number of pages beyond high watermark 3544 * 3545 * nr_free_pagecache_pages() counts the number of pages which are beyond the 3546 * high watermark within all zones. 3547 */ 3548 unsigned long nr_free_pagecache_pages(void) 3549 { 3550 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 3551 } 3552 3553 static inline void show_node(struct zone *zone) 3554 { 3555 if (IS_ENABLED(CONFIG_NUMA)) 3556 printk("Node %d ", zone_to_nid(zone)); 3557 } 3558 3559 void si_meminfo(struct sysinfo *val) 3560 { 3561 val->totalram = totalram_pages; 3562 val->sharedram = global_page_state(NR_SHMEM); 3563 val->freeram = global_page_state(NR_FREE_PAGES); 3564 val->bufferram = nr_blockdev_pages(); 3565 val->totalhigh = totalhigh_pages; 3566 val->freehigh = nr_free_highpages(); 3567 val->mem_unit = PAGE_SIZE; 3568 } 3569 3570 EXPORT_SYMBOL(si_meminfo); 3571 3572 #ifdef CONFIG_NUMA 3573 void si_meminfo_node(struct sysinfo *val, int nid) 3574 { 3575 int zone_type; /* needs to be signed */ 3576 unsigned long managed_pages = 0; 3577 pg_data_t *pgdat = NODE_DATA(nid); 3578 3579 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 3580 managed_pages += pgdat->node_zones[zone_type].managed_pages; 3581 val->totalram = managed_pages; 3582 val->sharedram = node_page_state(nid, NR_SHMEM); 3583 val->freeram = node_page_state(nid, NR_FREE_PAGES); 3584 #ifdef CONFIG_HIGHMEM 3585 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; 3586 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 3587 NR_FREE_PAGES); 3588 #else 3589 val->totalhigh = 0; 3590 val->freehigh = 0; 3591 #endif 3592 val->mem_unit = PAGE_SIZE; 3593 } 3594 #endif 3595 3596 /* 3597 * Determine whether the node should be displayed or not, depending on whether 3598 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 3599 */ 3600 bool skip_free_areas_node(unsigned int flags, int nid) 3601 { 3602 bool ret = false; 3603 unsigned int cpuset_mems_cookie; 3604 3605 if (!(flags & SHOW_MEM_FILTER_NODES)) 3606 goto out; 3607 3608 do { 3609 cpuset_mems_cookie = read_mems_allowed_begin(); 3610 ret = !node_isset(nid, cpuset_current_mems_allowed); 3611 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 3612 out: 3613 return ret; 3614 } 3615 3616 #define K(x) ((x) << (PAGE_SHIFT-10)) 3617 3618 static void show_migration_types(unsigned char type) 3619 { 3620 static const char types[MIGRATE_TYPES] = { 3621 [MIGRATE_UNMOVABLE] = 'U', 3622 [MIGRATE_RECLAIMABLE] = 'E', 3623 [MIGRATE_MOVABLE] = 'M', 3624 [MIGRATE_RESERVE] = 'R', 3625 #ifdef CONFIG_CMA 3626 [MIGRATE_CMA] = 'C', 3627 #endif 3628 #ifdef CONFIG_MEMORY_ISOLATION 3629 [MIGRATE_ISOLATE] = 'I', 3630 #endif 3631 }; 3632 char tmp[MIGRATE_TYPES + 1]; 3633 char *p = tmp; 3634 int i; 3635 3636 for (i = 0; i < MIGRATE_TYPES; i++) { 3637 if (type & (1 << i)) 3638 *p++ = types[i]; 3639 } 3640 3641 *p = '\0'; 3642 printk("(%s) ", tmp); 3643 } 3644 3645 /* 3646 * Show free area list (used inside shift_scroll-lock stuff) 3647 * We also calculate the percentage fragmentation. We do this by counting the 3648 * memory on each free list with the exception of the first item on the list. 3649 * 3650 * Bits in @filter: 3651 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's 3652 * cpuset. 3653 */ 3654 void show_free_areas(unsigned int filter) 3655 { 3656 unsigned long free_pcp = 0; 3657 int cpu; 3658 struct zone *zone; 3659 3660 for_each_populated_zone(zone) { 3661 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3662 continue; 3663 3664 for_each_online_cpu(cpu) 3665 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 3666 } 3667 3668 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 3669 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 3670 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n" 3671 " slab_reclaimable:%lu slab_unreclaimable:%lu\n" 3672 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" 3673 " free:%lu free_pcp:%lu free_cma:%lu\n", 3674 global_page_state(NR_ACTIVE_ANON), 3675 global_page_state(NR_INACTIVE_ANON), 3676 global_page_state(NR_ISOLATED_ANON), 3677 global_page_state(NR_ACTIVE_FILE), 3678 global_page_state(NR_INACTIVE_FILE), 3679 global_page_state(NR_ISOLATED_FILE), 3680 global_page_state(NR_UNEVICTABLE), 3681 global_page_state(NR_FILE_DIRTY), 3682 global_page_state(NR_WRITEBACK), 3683 global_page_state(NR_UNSTABLE_NFS), 3684 global_page_state(NR_SLAB_RECLAIMABLE), 3685 global_page_state(NR_SLAB_UNRECLAIMABLE), 3686 global_page_state(NR_FILE_MAPPED), 3687 global_page_state(NR_SHMEM), 3688 global_page_state(NR_PAGETABLE), 3689 global_page_state(NR_BOUNCE), 3690 global_page_state(NR_FREE_PAGES), 3691 free_pcp, 3692 global_page_state(NR_FREE_CMA_PAGES)); 3693 3694 for_each_populated_zone(zone) { 3695 int i; 3696 3697 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3698 continue; 3699 3700 free_pcp = 0; 3701 for_each_online_cpu(cpu) 3702 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 3703 3704 show_node(zone); 3705 printk("%s" 3706 " free:%lukB" 3707 " min:%lukB" 3708 " low:%lukB" 3709 " high:%lukB" 3710 " active_anon:%lukB" 3711 " inactive_anon:%lukB" 3712 " active_file:%lukB" 3713 " inactive_file:%lukB" 3714 " unevictable:%lukB" 3715 " isolated(anon):%lukB" 3716 " isolated(file):%lukB" 3717 " present:%lukB" 3718 " managed:%lukB" 3719 " mlocked:%lukB" 3720 " dirty:%lukB" 3721 " writeback:%lukB" 3722 " mapped:%lukB" 3723 " shmem:%lukB" 3724 " slab_reclaimable:%lukB" 3725 " slab_unreclaimable:%lukB" 3726 " kernel_stack:%lukB" 3727 " pagetables:%lukB" 3728 " unstable:%lukB" 3729 " bounce:%lukB" 3730 " free_pcp:%lukB" 3731 " local_pcp:%ukB" 3732 " free_cma:%lukB" 3733 " writeback_tmp:%lukB" 3734 " pages_scanned:%lu" 3735 " all_unreclaimable? %s" 3736 "\n", 3737 zone->name, 3738 K(zone_page_state(zone, NR_FREE_PAGES)), 3739 K(min_wmark_pages(zone)), 3740 K(low_wmark_pages(zone)), 3741 K(high_wmark_pages(zone)), 3742 K(zone_page_state(zone, NR_ACTIVE_ANON)), 3743 K(zone_page_state(zone, NR_INACTIVE_ANON)), 3744 K(zone_page_state(zone, NR_ACTIVE_FILE)), 3745 K(zone_page_state(zone, NR_INACTIVE_FILE)), 3746 K(zone_page_state(zone, NR_UNEVICTABLE)), 3747 K(zone_page_state(zone, NR_ISOLATED_ANON)), 3748 K(zone_page_state(zone, NR_ISOLATED_FILE)), 3749 K(zone->present_pages), 3750 K(zone->managed_pages), 3751 K(zone_page_state(zone, NR_MLOCK)), 3752 K(zone_page_state(zone, NR_FILE_DIRTY)), 3753 K(zone_page_state(zone, NR_WRITEBACK)), 3754 K(zone_page_state(zone, NR_FILE_MAPPED)), 3755 K(zone_page_state(zone, NR_SHMEM)), 3756 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 3757 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 3758 zone_page_state(zone, NR_KERNEL_STACK) * 3759 THREAD_SIZE / 1024, 3760 K(zone_page_state(zone, NR_PAGETABLE)), 3761 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 3762 K(zone_page_state(zone, NR_BOUNCE)), 3763 K(free_pcp), 3764 K(this_cpu_read(zone->pageset->pcp.count)), 3765 K(zone_page_state(zone, NR_FREE_CMA_PAGES)), 3766 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 3767 K(zone_page_state(zone, NR_PAGES_SCANNED)), 3768 (!zone_reclaimable(zone) ? "yes" : "no") 3769 ); 3770 printk("lowmem_reserve[]:"); 3771 for (i = 0; i < MAX_NR_ZONES; i++) 3772 printk(" %ld", zone->lowmem_reserve[i]); 3773 printk("\n"); 3774 } 3775 3776 for_each_populated_zone(zone) { 3777 unsigned long nr[MAX_ORDER], flags, order, total = 0; 3778 unsigned char types[MAX_ORDER]; 3779 3780 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3781 continue; 3782 show_node(zone); 3783 printk("%s: ", zone->name); 3784 3785 spin_lock_irqsave(&zone->lock, flags); 3786 for (order = 0; order < MAX_ORDER; order++) { 3787 struct free_area *area = &zone->free_area[order]; 3788 int type; 3789 3790 nr[order] = area->nr_free; 3791 total += nr[order] << order; 3792 3793 types[order] = 0; 3794 for (type = 0; type < MIGRATE_TYPES; type++) { 3795 if (!list_empty(&area->free_list[type])) 3796 types[order] |= 1 << type; 3797 } 3798 } 3799 spin_unlock_irqrestore(&zone->lock, flags); 3800 for (order = 0; order < MAX_ORDER; order++) { 3801 printk("%lu*%lukB ", nr[order], K(1UL) << order); 3802 if (nr[order]) 3803 show_migration_types(types[order]); 3804 } 3805 printk("= %lukB\n", K(total)); 3806 } 3807 3808 hugetlb_show_meminfo(); 3809 3810 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 3811 3812 show_swap_cache_info(); 3813 } 3814 3815 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 3816 { 3817 zoneref->zone = zone; 3818 zoneref->zone_idx = zone_idx(zone); 3819 } 3820 3821 /* 3822 * Builds allocation fallback zone lists. 3823 * 3824 * Add all populated zones of a node to the zonelist. 3825 */ 3826 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 3827 int nr_zones) 3828 { 3829 struct zone *zone; 3830 enum zone_type zone_type = MAX_NR_ZONES; 3831 3832 do { 3833 zone_type--; 3834 zone = pgdat->node_zones + zone_type; 3835 if (populated_zone(zone)) { 3836 zoneref_set_zone(zone, 3837 &zonelist->_zonerefs[nr_zones++]); 3838 check_highest_zone(zone_type); 3839 } 3840 } while (zone_type); 3841 3842 return nr_zones; 3843 } 3844 3845 3846 /* 3847 * zonelist_order: 3848 * 0 = automatic detection of better ordering. 3849 * 1 = order by ([node] distance, -zonetype) 3850 * 2 = order by (-zonetype, [node] distance) 3851 * 3852 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 3853 * the same zonelist. So only NUMA can configure this param. 3854 */ 3855 #define ZONELIST_ORDER_DEFAULT 0 3856 #define ZONELIST_ORDER_NODE 1 3857 #define ZONELIST_ORDER_ZONE 2 3858 3859 /* zonelist order in the kernel. 3860 * set_zonelist_order() will set this to NODE or ZONE. 3861 */ 3862 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 3863 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 3864 3865 3866 #ifdef CONFIG_NUMA 3867 /* The value user specified ....changed by config */ 3868 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3869 /* string for sysctl */ 3870 #define NUMA_ZONELIST_ORDER_LEN 16 3871 char numa_zonelist_order[16] = "default"; 3872 3873 /* 3874 * interface for configure zonelist ordering. 3875 * command line option "numa_zonelist_order" 3876 * = "[dD]efault - default, automatic configuration. 3877 * = "[nN]ode - order by node locality, then by zone within node 3878 * = "[zZ]one - order by zone, then by locality within zone 3879 */ 3880 3881 static int __parse_numa_zonelist_order(char *s) 3882 { 3883 if (*s == 'd' || *s == 'D') { 3884 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3885 } else if (*s == 'n' || *s == 'N') { 3886 user_zonelist_order = ZONELIST_ORDER_NODE; 3887 } else if (*s == 'z' || *s == 'Z') { 3888 user_zonelist_order = ZONELIST_ORDER_ZONE; 3889 } else { 3890 printk(KERN_WARNING 3891 "Ignoring invalid numa_zonelist_order value: " 3892 "%s\n", s); 3893 return -EINVAL; 3894 } 3895 return 0; 3896 } 3897 3898 static __init int setup_numa_zonelist_order(char *s) 3899 { 3900 int ret; 3901 3902 if (!s) 3903 return 0; 3904 3905 ret = __parse_numa_zonelist_order(s); 3906 if (ret == 0) 3907 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 3908 3909 return ret; 3910 } 3911 early_param("numa_zonelist_order", setup_numa_zonelist_order); 3912 3913 /* 3914 * sysctl handler for numa_zonelist_order 3915 */ 3916 int numa_zonelist_order_handler(struct ctl_table *table, int write, 3917 void __user *buffer, size_t *length, 3918 loff_t *ppos) 3919 { 3920 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 3921 int ret; 3922 static DEFINE_MUTEX(zl_order_mutex); 3923 3924 mutex_lock(&zl_order_mutex); 3925 if (write) { 3926 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { 3927 ret = -EINVAL; 3928 goto out; 3929 } 3930 strcpy(saved_string, (char *)table->data); 3931 } 3932 ret = proc_dostring(table, write, buffer, length, ppos); 3933 if (ret) 3934 goto out; 3935 if (write) { 3936 int oldval = user_zonelist_order; 3937 3938 ret = __parse_numa_zonelist_order((char *)table->data); 3939 if (ret) { 3940 /* 3941 * bogus value. restore saved string 3942 */ 3943 strncpy((char *)table->data, saved_string, 3944 NUMA_ZONELIST_ORDER_LEN); 3945 user_zonelist_order = oldval; 3946 } else if (oldval != user_zonelist_order) { 3947 mutex_lock(&zonelists_mutex); 3948 build_all_zonelists(NULL, NULL); 3949 mutex_unlock(&zonelists_mutex); 3950 } 3951 } 3952 out: 3953 mutex_unlock(&zl_order_mutex); 3954 return ret; 3955 } 3956 3957 3958 #define MAX_NODE_LOAD (nr_online_nodes) 3959 static int node_load[MAX_NUMNODES]; 3960 3961 /** 3962 * find_next_best_node - find the next node that should appear in a given node's fallback list 3963 * @node: node whose fallback list we're appending 3964 * @used_node_mask: nodemask_t of already used nodes 3965 * 3966 * We use a number of factors to determine which is the next node that should 3967 * appear on a given node's fallback list. The node should not have appeared 3968 * already in @node's fallback list, and it should be the next closest node 3969 * according to the distance array (which contains arbitrary distance values 3970 * from each node to each node in the system), and should also prefer nodes 3971 * with no CPUs, since presumably they'll have very little allocation pressure 3972 * on them otherwise. 3973 * It returns -1 if no node is found. 3974 */ 3975 static int find_next_best_node(int node, nodemask_t *used_node_mask) 3976 { 3977 int n, val; 3978 int min_val = INT_MAX; 3979 int best_node = NUMA_NO_NODE; 3980 const struct cpumask *tmp = cpumask_of_node(0); 3981 3982 /* Use the local node if we haven't already */ 3983 if (!node_isset(node, *used_node_mask)) { 3984 node_set(node, *used_node_mask); 3985 return node; 3986 } 3987 3988 for_each_node_state(n, N_MEMORY) { 3989 3990 /* Don't want a node to appear more than once */ 3991 if (node_isset(n, *used_node_mask)) 3992 continue; 3993 3994 /* Use the distance array to find the distance */ 3995 val = node_distance(node, n); 3996 3997 /* Penalize nodes under us ("prefer the next node") */ 3998 val += (n < node); 3999 4000 /* Give preference to headless and unused nodes */ 4001 tmp = cpumask_of_node(n); 4002 if (!cpumask_empty(tmp)) 4003 val += PENALTY_FOR_NODE_WITH_CPUS; 4004 4005 /* Slight preference for less loaded node */ 4006 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 4007 val += node_load[n]; 4008 4009 if (val < min_val) { 4010 min_val = val; 4011 best_node = n; 4012 } 4013 } 4014 4015 if (best_node >= 0) 4016 node_set(best_node, *used_node_mask); 4017 4018 return best_node; 4019 } 4020 4021 4022 /* 4023 * Build zonelists ordered by node and zones within node. 4024 * This results in maximum locality--normal zone overflows into local 4025 * DMA zone, if any--but risks exhausting DMA zone. 4026 */ 4027 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 4028 { 4029 int j; 4030 struct zonelist *zonelist; 4031 4032 zonelist = &pgdat->node_zonelists[0]; 4033 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 4034 ; 4035 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4036 zonelist->_zonerefs[j].zone = NULL; 4037 zonelist->_zonerefs[j].zone_idx = 0; 4038 } 4039 4040 /* 4041 * Build gfp_thisnode zonelists 4042 */ 4043 static void build_thisnode_zonelists(pg_data_t *pgdat) 4044 { 4045 int j; 4046 struct zonelist *zonelist; 4047 4048 zonelist = &pgdat->node_zonelists[1]; 4049 j = build_zonelists_node(pgdat, zonelist, 0); 4050 zonelist->_zonerefs[j].zone = NULL; 4051 zonelist->_zonerefs[j].zone_idx = 0; 4052 } 4053 4054 /* 4055 * Build zonelists ordered by zone and nodes within zones. 4056 * This results in conserving DMA zone[s] until all Normal memory is 4057 * exhausted, but results in overflowing to remote node while memory 4058 * may still exist in local DMA zone. 4059 */ 4060 static int node_order[MAX_NUMNODES]; 4061 4062 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 4063 { 4064 int pos, j, node; 4065 int zone_type; /* needs to be signed */ 4066 struct zone *z; 4067 struct zonelist *zonelist; 4068 4069 zonelist = &pgdat->node_zonelists[0]; 4070 pos = 0; 4071 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 4072 for (j = 0; j < nr_nodes; j++) { 4073 node = node_order[j]; 4074 z = &NODE_DATA(node)->node_zones[zone_type]; 4075 if (populated_zone(z)) { 4076 zoneref_set_zone(z, 4077 &zonelist->_zonerefs[pos++]); 4078 check_highest_zone(zone_type); 4079 } 4080 } 4081 } 4082 zonelist->_zonerefs[pos].zone = NULL; 4083 zonelist->_zonerefs[pos].zone_idx = 0; 4084 } 4085 4086 #if defined(CONFIG_64BIT) 4087 /* 4088 * Devices that require DMA32/DMA are relatively rare and do not justify a 4089 * penalty to every machine in case the specialised case applies. Default 4090 * to Node-ordering on 64-bit NUMA machines 4091 */ 4092 static int default_zonelist_order(void) 4093 { 4094 return ZONELIST_ORDER_NODE; 4095 } 4096 #else 4097 /* 4098 * On 32-bit, the Normal zone needs to be preserved for allocations accessible 4099 * by the kernel. If processes running on node 0 deplete the low memory zone 4100 * then reclaim will occur more frequency increasing stalls and potentially 4101 * be easier to OOM if a large percentage of the zone is under writeback or 4102 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set. 4103 * Hence, default to zone ordering on 32-bit. 4104 */ 4105 static int default_zonelist_order(void) 4106 { 4107 return ZONELIST_ORDER_ZONE; 4108 } 4109 #endif /* CONFIG_64BIT */ 4110 4111 static void set_zonelist_order(void) 4112 { 4113 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 4114 current_zonelist_order = default_zonelist_order(); 4115 else 4116 current_zonelist_order = user_zonelist_order; 4117 } 4118 4119 static void build_zonelists(pg_data_t *pgdat) 4120 { 4121 int j, node, load; 4122 enum zone_type i; 4123 nodemask_t used_mask; 4124 int local_node, prev_node; 4125 struct zonelist *zonelist; 4126 int order = current_zonelist_order; 4127 4128 /* initialize zonelists */ 4129 for (i = 0; i < MAX_ZONELISTS; i++) { 4130 zonelist = pgdat->node_zonelists + i; 4131 zonelist->_zonerefs[0].zone = NULL; 4132 zonelist->_zonerefs[0].zone_idx = 0; 4133 } 4134 4135 /* NUMA-aware ordering of nodes */ 4136 local_node = pgdat->node_id; 4137 load = nr_online_nodes; 4138 prev_node = local_node; 4139 nodes_clear(used_mask); 4140 4141 memset(node_order, 0, sizeof(node_order)); 4142 j = 0; 4143 4144 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 4145 /* 4146 * We don't want to pressure a particular node. 4147 * So adding penalty to the first node in same 4148 * distance group to make it round-robin. 4149 */ 4150 if (node_distance(local_node, node) != 4151 node_distance(local_node, prev_node)) 4152 node_load[node] = load; 4153 4154 prev_node = node; 4155 load--; 4156 if (order == ZONELIST_ORDER_NODE) 4157 build_zonelists_in_node_order(pgdat, node); 4158 else 4159 node_order[j++] = node; /* remember order */ 4160 } 4161 4162 if (order == ZONELIST_ORDER_ZONE) { 4163 /* calculate node order -- i.e., DMA last! */ 4164 build_zonelists_in_zone_order(pgdat, j); 4165 } 4166 4167 build_thisnode_zonelists(pgdat); 4168 } 4169 4170 /* Construct the zonelist performance cache - see further mmzone.h */ 4171 static void build_zonelist_cache(pg_data_t *pgdat) 4172 { 4173 struct zonelist *zonelist; 4174 struct zonelist_cache *zlc; 4175 struct zoneref *z; 4176 4177 zonelist = &pgdat->node_zonelists[0]; 4178 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 4179 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 4180 for (z = zonelist->_zonerefs; z->zone; z++) 4181 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 4182 } 4183 4184 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 4185 /* 4186 * Return node id of node used for "local" allocations. 4187 * I.e., first node id of first zone in arg node's generic zonelist. 4188 * Used for initializing percpu 'numa_mem', which is used primarily 4189 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 4190 */ 4191 int local_memory_node(int node) 4192 { 4193 struct zone *zone; 4194 4195 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 4196 gfp_zone(GFP_KERNEL), 4197 NULL, 4198 &zone); 4199 return zone->node; 4200 } 4201 #endif 4202 4203 #else /* CONFIG_NUMA */ 4204 4205 static void set_zonelist_order(void) 4206 { 4207 current_zonelist_order = ZONELIST_ORDER_ZONE; 4208 } 4209 4210 static void build_zonelists(pg_data_t *pgdat) 4211 { 4212 int node, local_node; 4213 enum zone_type j; 4214 struct zonelist *zonelist; 4215 4216 local_node = pgdat->node_id; 4217 4218 zonelist = &pgdat->node_zonelists[0]; 4219 j = build_zonelists_node(pgdat, zonelist, 0); 4220 4221 /* 4222 * Now we build the zonelist so that it contains the zones 4223 * of all the other nodes. 4224 * We don't want to pressure a particular node, so when 4225 * building the zones for node N, we make sure that the 4226 * zones coming right after the local ones are those from 4227 * node N+1 (modulo N) 4228 */ 4229 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 4230 if (!node_online(node)) 4231 continue; 4232 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4233 } 4234 for (node = 0; node < local_node; node++) { 4235 if (!node_online(node)) 4236 continue; 4237 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4238 } 4239 4240 zonelist->_zonerefs[j].zone = NULL; 4241 zonelist->_zonerefs[j].zone_idx = 0; 4242 } 4243 4244 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 4245 static void build_zonelist_cache(pg_data_t *pgdat) 4246 { 4247 pgdat->node_zonelists[0].zlcache_ptr = NULL; 4248 } 4249 4250 #endif /* CONFIG_NUMA */ 4251 4252 /* 4253 * Boot pageset table. One per cpu which is going to be used for all 4254 * zones and all nodes. The parameters will be set in such a way 4255 * that an item put on a list will immediately be handed over to 4256 * the buddy list. This is safe since pageset manipulation is done 4257 * with interrupts disabled. 4258 * 4259 * The boot_pagesets must be kept even after bootup is complete for 4260 * unused processors and/or zones. They do play a role for bootstrapping 4261 * hotplugged processors. 4262 * 4263 * zoneinfo_show() and maybe other functions do 4264 * not check if the processor is online before following the pageset pointer. 4265 * Other parts of the kernel may not check if the zone is available. 4266 */ 4267 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 4268 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 4269 static void setup_zone_pageset(struct zone *zone); 4270 4271 /* 4272 * Global mutex to protect against size modification of zonelists 4273 * as well as to serialize pageset setup for the new populated zone. 4274 */ 4275 DEFINE_MUTEX(zonelists_mutex); 4276 4277 /* return values int ....just for stop_machine() */ 4278 static int __build_all_zonelists(void *data) 4279 { 4280 int nid; 4281 int cpu; 4282 pg_data_t *self = data; 4283 4284 #ifdef CONFIG_NUMA 4285 memset(node_load, 0, sizeof(node_load)); 4286 #endif 4287 4288 if (self && !node_online(self->node_id)) { 4289 build_zonelists(self); 4290 build_zonelist_cache(self); 4291 } 4292 4293 for_each_online_node(nid) { 4294 pg_data_t *pgdat = NODE_DATA(nid); 4295 4296 build_zonelists(pgdat); 4297 build_zonelist_cache(pgdat); 4298 } 4299 4300 /* 4301 * Initialize the boot_pagesets that are going to be used 4302 * for bootstrapping processors. The real pagesets for 4303 * each zone will be allocated later when the per cpu 4304 * allocator is available. 4305 * 4306 * boot_pagesets are used also for bootstrapping offline 4307 * cpus if the system is already booted because the pagesets 4308 * are needed to initialize allocators on a specific cpu too. 4309 * F.e. the percpu allocator needs the page allocator which 4310 * needs the percpu allocator in order to allocate its pagesets 4311 * (a chicken-egg dilemma). 4312 */ 4313 for_each_possible_cpu(cpu) { 4314 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 4315 4316 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 4317 /* 4318 * We now know the "local memory node" for each node-- 4319 * i.e., the node of the first zone in the generic zonelist. 4320 * Set up numa_mem percpu variable for on-line cpus. During 4321 * boot, only the boot cpu should be on-line; we'll init the 4322 * secondary cpus' numa_mem as they come on-line. During 4323 * node/memory hotplug, we'll fixup all on-line cpus. 4324 */ 4325 if (cpu_online(cpu)) 4326 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 4327 #endif 4328 } 4329 4330 return 0; 4331 } 4332 4333 static noinline void __init 4334 build_all_zonelists_init(void) 4335 { 4336 __build_all_zonelists(NULL); 4337 mminit_verify_zonelist(); 4338 cpuset_init_current_mems_allowed(); 4339 } 4340 4341 /* 4342 * Called with zonelists_mutex held always 4343 * unless system_state == SYSTEM_BOOTING. 4344 * 4345 * __ref due to (1) call of __meminit annotated setup_zone_pageset 4346 * [we're only called with non-NULL zone through __meminit paths] and 4347 * (2) call of __init annotated helper build_all_zonelists_init 4348 * [protected by SYSTEM_BOOTING]. 4349 */ 4350 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) 4351 { 4352 set_zonelist_order(); 4353 4354 if (system_state == SYSTEM_BOOTING) { 4355 build_all_zonelists_init(); 4356 } else { 4357 #ifdef CONFIG_MEMORY_HOTPLUG 4358 if (zone) 4359 setup_zone_pageset(zone); 4360 #endif 4361 /* we have to stop all cpus to guarantee there is no user 4362 of zonelist */ 4363 stop_machine(__build_all_zonelists, pgdat, NULL); 4364 /* cpuset refresh routine should be here */ 4365 } 4366 vm_total_pages = nr_free_pagecache_pages(); 4367 /* 4368 * Disable grouping by mobility if the number of pages in the 4369 * system is too low to allow the mechanism to work. It would be 4370 * more accurate, but expensive to check per-zone. This check is 4371 * made on memory-hotadd so a system can start with mobility 4372 * disabled and enable it later 4373 */ 4374 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 4375 page_group_by_mobility_disabled = 1; 4376 else 4377 page_group_by_mobility_disabled = 0; 4378 4379 pr_info("Built %i zonelists in %s order, mobility grouping %s. " 4380 "Total pages: %ld\n", 4381 nr_online_nodes, 4382 zonelist_order_name[current_zonelist_order], 4383 page_group_by_mobility_disabled ? "off" : "on", 4384 vm_total_pages); 4385 #ifdef CONFIG_NUMA 4386 pr_info("Policy zone: %s\n", zone_names[policy_zone]); 4387 #endif 4388 } 4389 4390 /* 4391 * Helper functions to size the waitqueue hash table. 4392 * Essentially these want to choose hash table sizes sufficiently 4393 * large so that collisions trying to wait on pages are rare. 4394 * But in fact, the number of active page waitqueues on typical 4395 * systems is ridiculously low, less than 200. So this is even 4396 * conservative, even though it seems large. 4397 * 4398 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 4399 * waitqueues, i.e. the size of the waitq table given the number of pages. 4400 */ 4401 #define PAGES_PER_WAITQUEUE 256 4402 4403 #ifndef CONFIG_MEMORY_HOTPLUG 4404 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 4405 { 4406 unsigned long size = 1; 4407 4408 pages /= PAGES_PER_WAITQUEUE; 4409 4410 while (size < pages) 4411 size <<= 1; 4412 4413 /* 4414 * Once we have dozens or even hundreds of threads sleeping 4415 * on IO we've got bigger problems than wait queue collision. 4416 * Limit the size of the wait table to a reasonable size. 4417 */ 4418 size = min(size, 4096UL); 4419 4420 return max(size, 4UL); 4421 } 4422 #else 4423 /* 4424 * A zone's size might be changed by hot-add, so it is not possible to determine 4425 * a suitable size for its wait_table. So we use the maximum size now. 4426 * 4427 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 4428 * 4429 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 4430 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 4431 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 4432 * 4433 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 4434 * or more by the traditional way. (See above). It equals: 4435 * 4436 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 4437 * ia64(16K page size) : = ( 8G + 4M)byte. 4438 * powerpc (64K page size) : = (32G +16M)byte. 4439 */ 4440 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 4441 { 4442 return 4096UL; 4443 } 4444 #endif 4445 4446 /* 4447 * This is an integer logarithm so that shifts can be used later 4448 * to extract the more random high bits from the multiplicative 4449 * hash function before the remainder is taken. 4450 */ 4451 static inline unsigned long wait_table_bits(unsigned long size) 4452 { 4453 return ffz(~size); 4454 } 4455 4456 /* 4457 * Check if a pageblock contains reserved pages 4458 */ 4459 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 4460 { 4461 unsigned long pfn; 4462 4463 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 4464 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 4465 return 1; 4466 } 4467 return 0; 4468 } 4469 4470 /* 4471 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 4472 * of blocks reserved is based on min_wmark_pages(zone). The memory within 4473 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 4474 * higher will lead to a bigger reserve which will get freed as contiguous 4475 * blocks as reclaim kicks in 4476 */ 4477 static void setup_zone_migrate_reserve(struct zone *zone) 4478 { 4479 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 4480 struct page *page; 4481 unsigned long block_migratetype; 4482 int reserve; 4483 int old_reserve; 4484 4485 /* 4486 * Get the start pfn, end pfn and the number of blocks to reserve 4487 * We have to be careful to be aligned to pageblock_nr_pages to 4488 * make sure that we always check pfn_valid for the first page in 4489 * the block. 4490 */ 4491 start_pfn = zone->zone_start_pfn; 4492 end_pfn = zone_end_pfn(zone); 4493 start_pfn = roundup(start_pfn, pageblock_nr_pages); 4494 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 4495 pageblock_order; 4496 4497 /* 4498 * Reserve blocks are generally in place to help high-order atomic 4499 * allocations that are short-lived. A min_free_kbytes value that 4500 * would result in more than 2 reserve blocks for atomic allocations 4501 * is assumed to be in place to help anti-fragmentation for the 4502 * future allocation of hugepages at runtime. 4503 */ 4504 reserve = min(2, reserve); 4505 old_reserve = zone->nr_migrate_reserve_block; 4506 4507 /* When memory hot-add, we almost always need to do nothing */ 4508 if (reserve == old_reserve) 4509 return; 4510 zone->nr_migrate_reserve_block = reserve; 4511 4512 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 4513 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone))) 4514 return; 4515 4516 if (!pfn_valid(pfn)) 4517 continue; 4518 page = pfn_to_page(pfn); 4519 4520 /* Watch out for overlapping nodes */ 4521 if (page_to_nid(page) != zone_to_nid(zone)) 4522 continue; 4523 4524 block_migratetype = get_pageblock_migratetype(page); 4525 4526 /* Only test what is necessary when the reserves are not met */ 4527 if (reserve > 0) { 4528 /* 4529 * Blocks with reserved pages will never free, skip 4530 * them. 4531 */ 4532 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 4533 if (pageblock_is_reserved(pfn, block_end_pfn)) 4534 continue; 4535 4536 /* If this block is reserved, account for it */ 4537 if (block_migratetype == MIGRATE_RESERVE) { 4538 reserve--; 4539 continue; 4540 } 4541 4542 /* Suitable for reserving if this block is movable */ 4543 if (block_migratetype == MIGRATE_MOVABLE) { 4544 set_pageblock_migratetype(page, 4545 MIGRATE_RESERVE); 4546 move_freepages_block(zone, page, 4547 MIGRATE_RESERVE); 4548 reserve--; 4549 continue; 4550 } 4551 } else if (!old_reserve) { 4552 /* 4553 * At boot time we don't need to scan the whole zone 4554 * for turning off MIGRATE_RESERVE. 4555 */ 4556 break; 4557 } 4558 4559 /* 4560 * If the reserve is met and this is a previous reserved block, 4561 * take it back 4562 */ 4563 if (block_migratetype == MIGRATE_RESERVE) { 4564 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4565 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4566 } 4567 } 4568 } 4569 4570 /* 4571 * Initially all pages are reserved - free ones are freed 4572 * up by free_all_bootmem() once the early boot process is 4573 * done. Non-atomic initialization, single-pass. 4574 */ 4575 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 4576 unsigned long start_pfn, enum memmap_context context) 4577 { 4578 pg_data_t *pgdat = NODE_DATA(nid); 4579 unsigned long end_pfn = start_pfn + size; 4580 unsigned long pfn; 4581 struct zone *z; 4582 unsigned long nr_initialised = 0; 4583 4584 if (highest_memmap_pfn < end_pfn - 1) 4585 highest_memmap_pfn = end_pfn - 1; 4586 4587 z = &pgdat->node_zones[zone]; 4588 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 4589 /* 4590 * There can be holes in boot-time mem_map[]s 4591 * handed to this function. They do not 4592 * exist on hotplugged memory. 4593 */ 4594 if (context == MEMMAP_EARLY) { 4595 if (!early_pfn_valid(pfn)) 4596 continue; 4597 if (!early_pfn_in_nid(pfn, nid)) 4598 continue; 4599 if (!update_defer_init(pgdat, pfn, end_pfn, 4600 &nr_initialised)) 4601 break; 4602 } 4603 4604 /* 4605 * Mark the block movable so that blocks are reserved for 4606 * movable at startup. This will force kernel allocations 4607 * to reserve their blocks rather than leaking throughout 4608 * the address space during boot when many long-lived 4609 * kernel allocations are made. Later some blocks near 4610 * the start are marked MIGRATE_RESERVE by 4611 * setup_zone_migrate_reserve() 4612 * 4613 * bitmap is created for zone's valid pfn range. but memmap 4614 * can be created for invalid pages (for alignment) 4615 * check here not to call set_pageblock_migratetype() against 4616 * pfn out of zone. 4617 */ 4618 if (!(pfn & (pageblock_nr_pages - 1))) { 4619 struct page *page = pfn_to_page(pfn); 4620 4621 __init_single_page(page, pfn, zone, nid); 4622 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4623 } else { 4624 __init_single_pfn(pfn, zone, nid); 4625 } 4626 } 4627 } 4628 4629 static void __meminit zone_init_free_lists(struct zone *zone) 4630 { 4631 unsigned int order, t; 4632 for_each_migratetype_order(order, t) { 4633 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 4634 zone->free_area[order].nr_free = 0; 4635 } 4636 } 4637 4638 #ifndef __HAVE_ARCH_MEMMAP_INIT 4639 #define memmap_init(size, nid, zone, start_pfn) \ 4640 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 4641 #endif 4642 4643 static int zone_batchsize(struct zone *zone) 4644 { 4645 #ifdef CONFIG_MMU 4646 int batch; 4647 4648 /* 4649 * The per-cpu-pages pools are set to around 1000th of the 4650 * size of the zone. But no more than 1/2 of a meg. 4651 * 4652 * OK, so we don't know how big the cache is. So guess. 4653 */ 4654 batch = zone->managed_pages / 1024; 4655 if (batch * PAGE_SIZE > 512 * 1024) 4656 batch = (512 * 1024) / PAGE_SIZE; 4657 batch /= 4; /* We effectively *= 4 below */ 4658 if (batch < 1) 4659 batch = 1; 4660 4661 /* 4662 * Clamp the batch to a 2^n - 1 value. Having a power 4663 * of 2 value was found to be more likely to have 4664 * suboptimal cache aliasing properties in some cases. 4665 * 4666 * For example if 2 tasks are alternately allocating 4667 * batches of pages, one task can end up with a lot 4668 * of pages of one half of the possible page colors 4669 * and the other with pages of the other colors. 4670 */ 4671 batch = rounddown_pow_of_two(batch + batch/2) - 1; 4672 4673 return batch; 4674 4675 #else 4676 /* The deferral and batching of frees should be suppressed under NOMMU 4677 * conditions. 4678 * 4679 * The problem is that NOMMU needs to be able to allocate large chunks 4680 * of contiguous memory as there's no hardware page translation to 4681 * assemble apparent contiguous memory from discontiguous pages. 4682 * 4683 * Queueing large contiguous runs of pages for batching, however, 4684 * causes the pages to actually be freed in smaller chunks. As there 4685 * can be a significant delay between the individual batches being 4686 * recycled, this leads to the once large chunks of space being 4687 * fragmented and becoming unavailable for high-order allocations. 4688 */ 4689 return 0; 4690 #endif 4691 } 4692 4693 /* 4694 * pcp->high and pcp->batch values are related and dependent on one another: 4695 * ->batch must never be higher then ->high. 4696 * The following function updates them in a safe manner without read side 4697 * locking. 4698 * 4699 * Any new users of pcp->batch and pcp->high should ensure they can cope with 4700 * those fields changing asynchronously (acording the the above rule). 4701 * 4702 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function 4703 * outside of boot time (or some other assurance that no concurrent updaters 4704 * exist). 4705 */ 4706 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, 4707 unsigned long batch) 4708 { 4709 /* start with a fail safe value for batch */ 4710 pcp->batch = 1; 4711 smp_wmb(); 4712 4713 /* Update high, then batch, in order */ 4714 pcp->high = high; 4715 smp_wmb(); 4716 4717 pcp->batch = batch; 4718 } 4719 4720 /* a companion to pageset_set_high() */ 4721 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) 4722 { 4723 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); 4724 } 4725 4726 static void pageset_init(struct per_cpu_pageset *p) 4727 { 4728 struct per_cpu_pages *pcp; 4729 int migratetype; 4730 4731 memset(p, 0, sizeof(*p)); 4732 4733 pcp = &p->pcp; 4734 pcp->count = 0; 4735 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 4736 INIT_LIST_HEAD(&pcp->lists[migratetype]); 4737 } 4738 4739 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 4740 { 4741 pageset_init(p); 4742 pageset_set_batch(p, batch); 4743 } 4744 4745 /* 4746 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist 4747 * to the value high for the pageset p. 4748 */ 4749 static void pageset_set_high(struct per_cpu_pageset *p, 4750 unsigned long high) 4751 { 4752 unsigned long batch = max(1UL, high / 4); 4753 if ((high / 4) > (PAGE_SHIFT * 8)) 4754 batch = PAGE_SHIFT * 8; 4755 4756 pageset_update(&p->pcp, high, batch); 4757 } 4758 4759 static void pageset_set_high_and_batch(struct zone *zone, 4760 struct per_cpu_pageset *pcp) 4761 { 4762 if (percpu_pagelist_fraction) 4763 pageset_set_high(pcp, 4764 (zone->managed_pages / 4765 percpu_pagelist_fraction)); 4766 else 4767 pageset_set_batch(pcp, zone_batchsize(zone)); 4768 } 4769 4770 static void __meminit zone_pageset_init(struct zone *zone, int cpu) 4771 { 4772 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 4773 4774 pageset_init(pcp); 4775 pageset_set_high_and_batch(zone, pcp); 4776 } 4777 4778 static void __meminit setup_zone_pageset(struct zone *zone) 4779 { 4780 int cpu; 4781 zone->pageset = alloc_percpu(struct per_cpu_pageset); 4782 for_each_possible_cpu(cpu) 4783 zone_pageset_init(zone, cpu); 4784 } 4785 4786 /* 4787 * Allocate per cpu pagesets and initialize them. 4788 * Before this call only boot pagesets were available. 4789 */ 4790 void __init setup_per_cpu_pageset(void) 4791 { 4792 struct zone *zone; 4793 4794 for_each_populated_zone(zone) 4795 setup_zone_pageset(zone); 4796 } 4797 4798 static noinline __init_refok 4799 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 4800 { 4801 int i; 4802 size_t alloc_size; 4803 4804 /* 4805 * The per-page waitqueue mechanism uses hashed waitqueues 4806 * per zone. 4807 */ 4808 zone->wait_table_hash_nr_entries = 4809 wait_table_hash_nr_entries(zone_size_pages); 4810 zone->wait_table_bits = 4811 wait_table_bits(zone->wait_table_hash_nr_entries); 4812 alloc_size = zone->wait_table_hash_nr_entries 4813 * sizeof(wait_queue_head_t); 4814 4815 if (!slab_is_available()) { 4816 zone->wait_table = (wait_queue_head_t *) 4817 memblock_virt_alloc_node_nopanic( 4818 alloc_size, zone->zone_pgdat->node_id); 4819 } else { 4820 /* 4821 * This case means that a zone whose size was 0 gets new memory 4822 * via memory hot-add. 4823 * But it may be the case that a new node was hot-added. In 4824 * this case vmalloc() will not be able to use this new node's 4825 * memory - this wait_table must be initialized to use this new 4826 * node itself as well. 4827 * To use this new node's memory, further consideration will be 4828 * necessary. 4829 */ 4830 zone->wait_table = vmalloc(alloc_size); 4831 } 4832 if (!zone->wait_table) 4833 return -ENOMEM; 4834 4835 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i) 4836 init_waitqueue_head(zone->wait_table + i); 4837 4838 return 0; 4839 } 4840 4841 static __meminit void zone_pcp_init(struct zone *zone) 4842 { 4843 /* 4844 * per cpu subsystem is not up at this point. The following code 4845 * relies on the ability of the linker to provide the 4846 * offset of a (static) per cpu variable into the per cpu area. 4847 */ 4848 zone->pageset = &boot_pageset; 4849 4850 if (populated_zone(zone)) 4851 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 4852 zone->name, zone->present_pages, 4853 zone_batchsize(zone)); 4854 } 4855 4856 int __meminit init_currently_empty_zone(struct zone *zone, 4857 unsigned long zone_start_pfn, 4858 unsigned long size, 4859 enum memmap_context context) 4860 { 4861 struct pglist_data *pgdat = zone->zone_pgdat; 4862 int ret; 4863 ret = zone_wait_table_init(zone, size); 4864 if (ret) 4865 return ret; 4866 pgdat->nr_zones = zone_idx(zone) + 1; 4867 4868 zone->zone_start_pfn = zone_start_pfn; 4869 4870 mminit_dprintk(MMINIT_TRACE, "memmap_init", 4871 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 4872 pgdat->node_id, 4873 (unsigned long)zone_idx(zone), 4874 zone_start_pfn, (zone_start_pfn + size)); 4875 4876 zone_init_free_lists(zone); 4877 4878 return 0; 4879 } 4880 4881 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4882 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 4883 4884 /* 4885 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 4886 */ 4887 int __meminit __early_pfn_to_nid(unsigned long pfn, 4888 struct mminit_pfnnid_cache *state) 4889 { 4890 unsigned long start_pfn, end_pfn; 4891 int nid; 4892 4893 if (state->last_start <= pfn && pfn < state->last_end) 4894 return state->last_nid; 4895 4896 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); 4897 if (nid != -1) { 4898 state->last_start = start_pfn; 4899 state->last_end = end_pfn; 4900 state->last_nid = nid; 4901 } 4902 4903 return nid; 4904 } 4905 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 4906 4907 /** 4908 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range 4909 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 4910 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid 4911 * 4912 * If an architecture guarantees that all ranges registered contain no holes 4913 * and may be freed, this this function may be used instead of calling 4914 * memblock_free_early_nid() manually. 4915 */ 4916 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 4917 { 4918 unsigned long start_pfn, end_pfn; 4919 int i, this_nid; 4920 4921 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 4922 start_pfn = min(start_pfn, max_low_pfn); 4923 end_pfn = min(end_pfn, max_low_pfn); 4924 4925 if (start_pfn < end_pfn) 4926 memblock_free_early_nid(PFN_PHYS(start_pfn), 4927 (end_pfn - start_pfn) << PAGE_SHIFT, 4928 this_nid); 4929 } 4930 } 4931 4932 /** 4933 * sparse_memory_present_with_active_regions - Call memory_present for each active range 4934 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 4935 * 4936 * If an architecture guarantees that all ranges registered contain no holes and may 4937 * be freed, this function may be used instead of calling memory_present() manually. 4938 */ 4939 void __init sparse_memory_present_with_active_regions(int nid) 4940 { 4941 unsigned long start_pfn, end_pfn; 4942 int i, this_nid; 4943 4944 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 4945 memory_present(this_nid, start_pfn, end_pfn); 4946 } 4947 4948 /** 4949 * get_pfn_range_for_nid - Return the start and end page frames for a node 4950 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 4951 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 4952 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 4953 * 4954 * It returns the start and end page frame of a node based on information 4955 * provided by memblock_set_node(). If called for a node 4956 * with no available memory, a warning is printed and the start and end 4957 * PFNs will be 0. 4958 */ 4959 void __meminit get_pfn_range_for_nid(unsigned int nid, 4960 unsigned long *start_pfn, unsigned long *end_pfn) 4961 { 4962 unsigned long this_start_pfn, this_end_pfn; 4963 int i; 4964 4965 *start_pfn = -1UL; 4966 *end_pfn = 0; 4967 4968 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 4969 *start_pfn = min(*start_pfn, this_start_pfn); 4970 *end_pfn = max(*end_pfn, this_end_pfn); 4971 } 4972 4973 if (*start_pfn == -1UL) 4974 *start_pfn = 0; 4975 } 4976 4977 /* 4978 * This finds a zone that can be used for ZONE_MOVABLE pages. The 4979 * assumption is made that zones within a node are ordered in monotonic 4980 * increasing memory addresses so that the "highest" populated zone is used 4981 */ 4982 static void __init find_usable_zone_for_movable(void) 4983 { 4984 int zone_index; 4985 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 4986 if (zone_index == ZONE_MOVABLE) 4987 continue; 4988 4989 if (arch_zone_highest_possible_pfn[zone_index] > 4990 arch_zone_lowest_possible_pfn[zone_index]) 4991 break; 4992 } 4993 4994 VM_BUG_ON(zone_index == -1); 4995 movable_zone = zone_index; 4996 } 4997 4998 /* 4999 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 5000 * because it is sized independent of architecture. Unlike the other zones, 5001 * the starting point for ZONE_MOVABLE is not fixed. It may be different 5002 * in each node depending on the size of each node and how evenly kernelcore 5003 * is distributed. This helper function adjusts the zone ranges 5004 * provided by the architecture for a given node by using the end of the 5005 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 5006 * zones within a node are in order of monotonic increases memory addresses 5007 */ 5008 static void __meminit adjust_zone_range_for_zone_movable(int nid, 5009 unsigned long zone_type, 5010 unsigned long node_start_pfn, 5011 unsigned long node_end_pfn, 5012 unsigned long *zone_start_pfn, 5013 unsigned long *zone_end_pfn) 5014 { 5015 /* Only adjust if ZONE_MOVABLE is on this node */ 5016 if (zone_movable_pfn[nid]) { 5017 /* Size ZONE_MOVABLE */ 5018 if (zone_type == ZONE_MOVABLE) { 5019 *zone_start_pfn = zone_movable_pfn[nid]; 5020 *zone_end_pfn = min(node_end_pfn, 5021 arch_zone_highest_possible_pfn[movable_zone]); 5022 5023 /* Adjust for ZONE_MOVABLE starting within this range */ 5024 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 5025 *zone_end_pfn > zone_movable_pfn[nid]) { 5026 *zone_end_pfn = zone_movable_pfn[nid]; 5027 5028 /* Check if this whole range is within ZONE_MOVABLE */ 5029 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 5030 *zone_start_pfn = *zone_end_pfn; 5031 } 5032 } 5033 5034 /* 5035 * Return the number of pages a zone spans in a node, including holes 5036 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 5037 */ 5038 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 5039 unsigned long zone_type, 5040 unsigned long node_start_pfn, 5041 unsigned long node_end_pfn, 5042 unsigned long *ignored) 5043 { 5044 unsigned long zone_start_pfn, zone_end_pfn; 5045 5046 /* Get the start and end of the zone */ 5047 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 5048 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 5049 adjust_zone_range_for_zone_movable(nid, zone_type, 5050 node_start_pfn, node_end_pfn, 5051 &zone_start_pfn, &zone_end_pfn); 5052 5053 /* Check that this node has pages within the zone's required range */ 5054 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 5055 return 0; 5056 5057 /* Move the zone boundaries inside the node if necessary */ 5058 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 5059 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 5060 5061 /* Return the spanned pages */ 5062 return zone_end_pfn - zone_start_pfn; 5063 } 5064 5065 /* 5066 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 5067 * then all holes in the requested range will be accounted for. 5068 */ 5069 unsigned long __meminit __absent_pages_in_range(int nid, 5070 unsigned long range_start_pfn, 5071 unsigned long range_end_pfn) 5072 { 5073 unsigned long nr_absent = range_end_pfn - range_start_pfn; 5074 unsigned long start_pfn, end_pfn; 5075 int i; 5076 5077 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5078 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 5079 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 5080 nr_absent -= end_pfn - start_pfn; 5081 } 5082 return nr_absent; 5083 } 5084 5085 /** 5086 * absent_pages_in_range - Return number of page frames in holes within a range 5087 * @start_pfn: The start PFN to start searching for holes 5088 * @end_pfn: The end PFN to stop searching for holes 5089 * 5090 * It returns the number of pages frames in memory holes within a range. 5091 */ 5092 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 5093 unsigned long end_pfn) 5094 { 5095 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 5096 } 5097 5098 /* Return the number of page frames in holes in a zone on a node */ 5099 static unsigned long __meminit zone_absent_pages_in_node(int nid, 5100 unsigned long zone_type, 5101 unsigned long node_start_pfn, 5102 unsigned long node_end_pfn, 5103 unsigned long *ignored) 5104 { 5105 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 5106 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 5107 unsigned long zone_start_pfn, zone_end_pfn; 5108 5109 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 5110 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 5111 5112 adjust_zone_range_for_zone_movable(nid, zone_type, 5113 node_start_pfn, node_end_pfn, 5114 &zone_start_pfn, &zone_end_pfn); 5115 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 5116 } 5117 5118 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5119 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 5120 unsigned long zone_type, 5121 unsigned long node_start_pfn, 5122 unsigned long node_end_pfn, 5123 unsigned long *zones_size) 5124 { 5125 return zones_size[zone_type]; 5126 } 5127 5128 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 5129 unsigned long zone_type, 5130 unsigned long node_start_pfn, 5131 unsigned long node_end_pfn, 5132 unsigned long *zholes_size) 5133 { 5134 if (!zholes_size) 5135 return 0; 5136 5137 return zholes_size[zone_type]; 5138 } 5139 5140 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5141 5142 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 5143 unsigned long node_start_pfn, 5144 unsigned long node_end_pfn, 5145 unsigned long *zones_size, 5146 unsigned long *zholes_size) 5147 { 5148 unsigned long realtotalpages = 0, totalpages = 0; 5149 enum zone_type i; 5150 5151 for (i = 0; i < MAX_NR_ZONES; i++) { 5152 struct zone *zone = pgdat->node_zones + i; 5153 unsigned long size, real_size; 5154 5155 size = zone_spanned_pages_in_node(pgdat->node_id, i, 5156 node_start_pfn, 5157 node_end_pfn, 5158 zones_size); 5159 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i, 5160 node_start_pfn, node_end_pfn, 5161 zholes_size); 5162 zone->spanned_pages = size; 5163 zone->present_pages = real_size; 5164 5165 totalpages += size; 5166 realtotalpages += real_size; 5167 } 5168 5169 pgdat->node_spanned_pages = totalpages; 5170 pgdat->node_present_pages = realtotalpages; 5171 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 5172 realtotalpages); 5173 } 5174 5175 #ifndef CONFIG_SPARSEMEM 5176 /* 5177 * Calculate the size of the zone->blockflags rounded to an unsigned long 5178 * Start by making sure zonesize is a multiple of pageblock_order by rounding 5179 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 5180 * round what is now in bits to nearest long in bits, then return it in 5181 * bytes. 5182 */ 5183 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) 5184 { 5185 unsigned long usemapsize; 5186 5187 zonesize += zone_start_pfn & (pageblock_nr_pages-1); 5188 usemapsize = roundup(zonesize, pageblock_nr_pages); 5189 usemapsize = usemapsize >> pageblock_order; 5190 usemapsize *= NR_PAGEBLOCK_BITS; 5191 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 5192 5193 return usemapsize / 8; 5194 } 5195 5196 static void __init setup_usemap(struct pglist_data *pgdat, 5197 struct zone *zone, 5198 unsigned long zone_start_pfn, 5199 unsigned long zonesize) 5200 { 5201 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); 5202 zone->pageblock_flags = NULL; 5203 if (usemapsize) 5204 zone->pageblock_flags = 5205 memblock_virt_alloc_node_nopanic(usemapsize, 5206 pgdat->node_id); 5207 } 5208 #else 5209 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, 5210 unsigned long zone_start_pfn, unsigned long zonesize) {} 5211 #endif /* CONFIG_SPARSEMEM */ 5212 5213 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 5214 5215 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 5216 void __paginginit set_pageblock_order(void) 5217 { 5218 unsigned int order; 5219 5220 /* Check that pageblock_nr_pages has not already been setup */ 5221 if (pageblock_order) 5222 return; 5223 5224 if (HPAGE_SHIFT > PAGE_SHIFT) 5225 order = HUGETLB_PAGE_ORDER; 5226 else 5227 order = MAX_ORDER - 1; 5228 5229 /* 5230 * Assume the largest contiguous order of interest is a huge page. 5231 * This value may be variable depending on boot parameters on IA64 and 5232 * powerpc. 5233 */ 5234 pageblock_order = order; 5235 } 5236 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5237 5238 /* 5239 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 5240 * is unused as pageblock_order is set at compile-time. See 5241 * include/linux/pageblock-flags.h for the values of pageblock_order based on 5242 * the kernel config 5243 */ 5244 void __paginginit set_pageblock_order(void) 5245 { 5246 } 5247 5248 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5249 5250 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, 5251 unsigned long present_pages) 5252 { 5253 unsigned long pages = spanned_pages; 5254 5255 /* 5256 * Provide a more accurate estimation if there are holes within 5257 * the zone and SPARSEMEM is in use. If there are holes within the 5258 * zone, each populated memory region may cost us one or two extra 5259 * memmap pages due to alignment because memmap pages for each 5260 * populated regions may not naturally algined on page boundary. 5261 * So the (present_pages >> 4) heuristic is a tradeoff for that. 5262 */ 5263 if (spanned_pages > present_pages + (present_pages >> 4) && 5264 IS_ENABLED(CONFIG_SPARSEMEM)) 5265 pages = present_pages; 5266 5267 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; 5268 } 5269 5270 /* 5271 * Set up the zone data structures: 5272 * - mark all pages reserved 5273 * - mark all memory queues empty 5274 * - clear the memory bitmaps 5275 * 5276 * NOTE: pgdat should get zeroed by caller. 5277 */ 5278 static void __paginginit free_area_init_core(struct pglist_data *pgdat, 5279 unsigned long node_start_pfn, unsigned long node_end_pfn) 5280 { 5281 enum zone_type j; 5282 int nid = pgdat->node_id; 5283 unsigned long zone_start_pfn = pgdat->node_start_pfn; 5284 int ret; 5285 5286 pgdat_resize_init(pgdat); 5287 #ifdef CONFIG_NUMA_BALANCING 5288 spin_lock_init(&pgdat->numabalancing_migrate_lock); 5289 pgdat->numabalancing_migrate_nr_pages = 0; 5290 pgdat->numabalancing_migrate_next_window = jiffies; 5291 #endif 5292 init_waitqueue_head(&pgdat->kswapd_wait); 5293 init_waitqueue_head(&pgdat->pfmemalloc_wait); 5294 pgdat_page_ext_init(pgdat); 5295 5296 for (j = 0; j < MAX_NR_ZONES; j++) { 5297 struct zone *zone = pgdat->node_zones + j; 5298 unsigned long size, realsize, freesize, memmap_pages; 5299 5300 size = zone->spanned_pages; 5301 realsize = freesize = zone->present_pages; 5302 5303 /* 5304 * Adjust freesize so that it accounts for how much memory 5305 * is used by this zone for memmap. This affects the watermark 5306 * and per-cpu initialisations 5307 */ 5308 memmap_pages = calc_memmap_size(size, realsize); 5309 if (!is_highmem_idx(j)) { 5310 if (freesize >= memmap_pages) { 5311 freesize -= memmap_pages; 5312 if (memmap_pages) 5313 printk(KERN_DEBUG 5314 " %s zone: %lu pages used for memmap\n", 5315 zone_names[j], memmap_pages); 5316 } else 5317 printk(KERN_WARNING 5318 " %s zone: %lu pages exceeds freesize %lu\n", 5319 zone_names[j], memmap_pages, freesize); 5320 } 5321 5322 /* Account for reserved pages */ 5323 if (j == 0 && freesize > dma_reserve) { 5324 freesize -= dma_reserve; 5325 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 5326 zone_names[0], dma_reserve); 5327 } 5328 5329 if (!is_highmem_idx(j)) 5330 nr_kernel_pages += freesize; 5331 /* Charge for highmem memmap if there are enough kernel pages */ 5332 else if (nr_kernel_pages > memmap_pages * 2) 5333 nr_kernel_pages -= memmap_pages; 5334 nr_all_pages += freesize; 5335 5336 /* 5337 * Set an approximate value for lowmem here, it will be adjusted 5338 * when the bootmem allocator frees pages into the buddy system. 5339 * And all highmem pages will be managed by the buddy system. 5340 */ 5341 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; 5342 #ifdef CONFIG_NUMA 5343 zone->node = nid; 5344 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) 5345 / 100; 5346 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; 5347 #endif 5348 zone->name = zone_names[j]; 5349 spin_lock_init(&zone->lock); 5350 spin_lock_init(&zone->lru_lock); 5351 zone_seqlock_init(zone); 5352 zone->zone_pgdat = pgdat; 5353 zone_pcp_init(zone); 5354 5355 /* For bootup, initialized properly in watermark setup */ 5356 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages); 5357 5358 lruvec_init(&zone->lruvec); 5359 if (!size) 5360 continue; 5361 5362 set_pageblock_order(); 5363 setup_usemap(pgdat, zone, zone_start_pfn, size); 5364 ret = init_currently_empty_zone(zone, zone_start_pfn, 5365 size, MEMMAP_EARLY); 5366 BUG_ON(ret); 5367 memmap_init(size, nid, j, zone_start_pfn); 5368 zone_start_pfn += size; 5369 } 5370 } 5371 5372 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 5373 { 5374 /* Skip empty nodes */ 5375 if (!pgdat->node_spanned_pages) 5376 return; 5377 5378 #ifdef CONFIG_FLAT_NODE_MEM_MAP 5379 /* ia64 gets its own node_mem_map, before this, without bootmem */ 5380 if (!pgdat->node_mem_map) { 5381 unsigned long size, start, end; 5382 struct page *map; 5383 5384 /* 5385 * The zone's endpoints aren't required to be MAX_ORDER 5386 * aligned but the node_mem_map endpoints must be in order 5387 * for the buddy allocator to function correctly. 5388 */ 5389 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 5390 end = pgdat_end_pfn(pgdat); 5391 end = ALIGN(end, MAX_ORDER_NR_PAGES); 5392 size = (end - start) * sizeof(struct page); 5393 map = alloc_remap(pgdat->node_id, size); 5394 if (!map) 5395 map = memblock_virt_alloc_node_nopanic(size, 5396 pgdat->node_id); 5397 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 5398 } 5399 #ifndef CONFIG_NEED_MULTIPLE_NODES 5400 /* 5401 * With no DISCONTIG, the global mem_map is just set as node 0's 5402 */ 5403 if (pgdat == NODE_DATA(0)) { 5404 mem_map = NODE_DATA(0)->node_mem_map; 5405 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5406 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 5407 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 5408 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5409 } 5410 #endif 5411 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 5412 } 5413 5414 void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 5415 unsigned long node_start_pfn, unsigned long *zholes_size) 5416 { 5417 pg_data_t *pgdat = NODE_DATA(nid); 5418 unsigned long start_pfn = 0; 5419 unsigned long end_pfn = 0; 5420 5421 /* pg_data_t should be reset to zero when it's allocated */ 5422 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); 5423 5424 reset_deferred_meminit(pgdat); 5425 pgdat->node_id = nid; 5426 pgdat->node_start_pfn = node_start_pfn; 5427 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5428 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 5429 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid, 5430 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1); 5431 #endif 5432 calculate_node_totalpages(pgdat, start_pfn, end_pfn, 5433 zones_size, zholes_size); 5434 5435 alloc_node_mem_map(pgdat); 5436 #ifdef CONFIG_FLAT_NODE_MEM_MAP 5437 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 5438 nid, (unsigned long)pgdat, 5439 (unsigned long)pgdat->node_mem_map); 5440 #endif 5441 5442 free_area_init_core(pgdat, start_pfn, end_pfn); 5443 } 5444 5445 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5446 5447 #if MAX_NUMNODES > 1 5448 /* 5449 * Figure out the number of possible node ids. 5450 */ 5451 void __init setup_nr_node_ids(void) 5452 { 5453 unsigned int node; 5454 unsigned int highest = 0; 5455 5456 for_each_node_mask(node, node_possible_map) 5457 highest = node; 5458 nr_node_ids = highest + 1; 5459 } 5460 #endif 5461 5462 /** 5463 * node_map_pfn_alignment - determine the maximum internode alignment 5464 * 5465 * This function should be called after node map is populated and sorted. 5466 * It calculates the maximum power of two alignment which can distinguish 5467 * all the nodes. 5468 * 5469 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 5470 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 5471 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 5472 * shifted, 1GiB is enough and this function will indicate so. 5473 * 5474 * This is used to test whether pfn -> nid mapping of the chosen memory 5475 * model has fine enough granularity to avoid incorrect mapping for the 5476 * populated node map. 5477 * 5478 * Returns the determined alignment in pfn's. 0 if there is no alignment 5479 * requirement (single node). 5480 */ 5481 unsigned long __init node_map_pfn_alignment(void) 5482 { 5483 unsigned long accl_mask = 0, last_end = 0; 5484 unsigned long start, end, mask; 5485 int last_nid = -1; 5486 int i, nid; 5487 5488 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 5489 if (!start || last_nid < 0 || last_nid == nid) { 5490 last_nid = nid; 5491 last_end = end; 5492 continue; 5493 } 5494 5495 /* 5496 * Start with a mask granular enough to pin-point to the 5497 * start pfn and tick off bits one-by-one until it becomes 5498 * too coarse to separate the current node from the last. 5499 */ 5500 mask = ~((1 << __ffs(start)) - 1); 5501 while (mask && last_end <= (start & (mask << 1))) 5502 mask <<= 1; 5503 5504 /* accumulate all internode masks */ 5505 accl_mask |= mask; 5506 } 5507 5508 /* convert mask to number of pages */ 5509 return ~accl_mask + 1; 5510 } 5511 5512 /* Find the lowest pfn for a node */ 5513 static unsigned long __init find_min_pfn_for_node(int nid) 5514 { 5515 unsigned long min_pfn = ULONG_MAX; 5516 unsigned long start_pfn; 5517 int i; 5518 5519 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 5520 min_pfn = min(min_pfn, start_pfn); 5521 5522 if (min_pfn == ULONG_MAX) { 5523 printk(KERN_WARNING 5524 "Could not find start_pfn for node %d\n", nid); 5525 return 0; 5526 } 5527 5528 return min_pfn; 5529 } 5530 5531 /** 5532 * find_min_pfn_with_active_regions - Find the minimum PFN registered 5533 * 5534 * It returns the minimum PFN based on information provided via 5535 * memblock_set_node(). 5536 */ 5537 unsigned long __init find_min_pfn_with_active_regions(void) 5538 { 5539 return find_min_pfn_for_node(MAX_NUMNODES); 5540 } 5541 5542 /* 5543 * early_calculate_totalpages() 5544 * Sum pages in active regions for movable zone. 5545 * Populate N_MEMORY for calculating usable_nodes. 5546 */ 5547 static unsigned long __init early_calculate_totalpages(void) 5548 { 5549 unsigned long totalpages = 0; 5550 unsigned long start_pfn, end_pfn; 5551 int i, nid; 5552 5553 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 5554 unsigned long pages = end_pfn - start_pfn; 5555 5556 totalpages += pages; 5557 if (pages) 5558 node_set_state(nid, N_MEMORY); 5559 } 5560 return totalpages; 5561 } 5562 5563 /* 5564 * Find the PFN the Movable zone begins in each node. Kernel memory 5565 * is spread evenly between nodes as long as the nodes have enough 5566 * memory. When they don't, some nodes will have more kernelcore than 5567 * others 5568 */ 5569 static void __init find_zone_movable_pfns_for_nodes(void) 5570 { 5571 int i, nid; 5572 unsigned long usable_startpfn; 5573 unsigned long kernelcore_node, kernelcore_remaining; 5574 /* save the state before borrow the nodemask */ 5575 nodemask_t saved_node_state = node_states[N_MEMORY]; 5576 unsigned long totalpages = early_calculate_totalpages(); 5577 int usable_nodes = nodes_weight(node_states[N_MEMORY]); 5578 struct memblock_region *r; 5579 5580 /* Need to find movable_zone earlier when movable_node is specified. */ 5581 find_usable_zone_for_movable(); 5582 5583 /* 5584 * If movable_node is specified, ignore kernelcore and movablecore 5585 * options. 5586 */ 5587 if (movable_node_is_enabled()) { 5588 for_each_memblock(memory, r) { 5589 if (!memblock_is_hotpluggable(r)) 5590 continue; 5591 5592 nid = r->nid; 5593 5594 usable_startpfn = PFN_DOWN(r->base); 5595 zone_movable_pfn[nid] = zone_movable_pfn[nid] ? 5596 min(usable_startpfn, zone_movable_pfn[nid]) : 5597 usable_startpfn; 5598 } 5599 5600 goto out2; 5601 } 5602 5603 /* 5604 * If movablecore=nn[KMG] was specified, calculate what size of 5605 * kernelcore that corresponds so that memory usable for 5606 * any allocation type is evenly spread. If both kernelcore 5607 * and movablecore are specified, then the value of kernelcore 5608 * will be used for required_kernelcore if it's greater than 5609 * what movablecore would have allowed. 5610 */ 5611 if (required_movablecore) { 5612 unsigned long corepages; 5613 5614 /* 5615 * Round-up so that ZONE_MOVABLE is at least as large as what 5616 * was requested by the user 5617 */ 5618 required_movablecore = 5619 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 5620 corepages = totalpages - required_movablecore; 5621 5622 required_kernelcore = max(required_kernelcore, corepages); 5623 } 5624 5625 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 5626 if (!required_kernelcore) 5627 goto out; 5628 5629 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 5630 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 5631 5632 restart: 5633 /* Spread kernelcore memory as evenly as possible throughout nodes */ 5634 kernelcore_node = required_kernelcore / usable_nodes; 5635 for_each_node_state(nid, N_MEMORY) { 5636 unsigned long start_pfn, end_pfn; 5637 5638 /* 5639 * Recalculate kernelcore_node if the division per node 5640 * now exceeds what is necessary to satisfy the requested 5641 * amount of memory for the kernel 5642 */ 5643 if (required_kernelcore < kernelcore_node) 5644 kernelcore_node = required_kernelcore / usable_nodes; 5645 5646 /* 5647 * As the map is walked, we track how much memory is usable 5648 * by the kernel using kernelcore_remaining. When it is 5649 * 0, the rest of the node is usable by ZONE_MOVABLE 5650 */ 5651 kernelcore_remaining = kernelcore_node; 5652 5653 /* Go through each range of PFNs within this node */ 5654 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5655 unsigned long size_pages; 5656 5657 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 5658 if (start_pfn >= end_pfn) 5659 continue; 5660 5661 /* Account for what is only usable for kernelcore */ 5662 if (start_pfn < usable_startpfn) { 5663 unsigned long kernel_pages; 5664 kernel_pages = min(end_pfn, usable_startpfn) 5665 - start_pfn; 5666 5667 kernelcore_remaining -= min(kernel_pages, 5668 kernelcore_remaining); 5669 required_kernelcore -= min(kernel_pages, 5670 required_kernelcore); 5671 5672 /* Continue if range is now fully accounted */ 5673 if (end_pfn <= usable_startpfn) { 5674 5675 /* 5676 * Push zone_movable_pfn to the end so 5677 * that if we have to rebalance 5678 * kernelcore across nodes, we will 5679 * not double account here 5680 */ 5681 zone_movable_pfn[nid] = end_pfn; 5682 continue; 5683 } 5684 start_pfn = usable_startpfn; 5685 } 5686 5687 /* 5688 * The usable PFN range for ZONE_MOVABLE is from 5689 * start_pfn->end_pfn. Calculate size_pages as the 5690 * number of pages used as kernelcore 5691 */ 5692 size_pages = end_pfn - start_pfn; 5693 if (size_pages > kernelcore_remaining) 5694 size_pages = kernelcore_remaining; 5695 zone_movable_pfn[nid] = start_pfn + size_pages; 5696 5697 /* 5698 * Some kernelcore has been met, update counts and 5699 * break if the kernelcore for this node has been 5700 * satisfied 5701 */ 5702 required_kernelcore -= min(required_kernelcore, 5703 size_pages); 5704 kernelcore_remaining -= size_pages; 5705 if (!kernelcore_remaining) 5706 break; 5707 } 5708 } 5709 5710 /* 5711 * If there is still required_kernelcore, we do another pass with one 5712 * less node in the count. This will push zone_movable_pfn[nid] further 5713 * along on the nodes that still have memory until kernelcore is 5714 * satisfied 5715 */ 5716 usable_nodes--; 5717 if (usable_nodes && required_kernelcore > usable_nodes) 5718 goto restart; 5719 5720 out2: 5721 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 5722 for (nid = 0; nid < MAX_NUMNODES; nid++) 5723 zone_movable_pfn[nid] = 5724 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 5725 5726 out: 5727 /* restore the node_state */ 5728 node_states[N_MEMORY] = saved_node_state; 5729 } 5730 5731 /* Any regular or high memory on that node ? */ 5732 static void check_for_memory(pg_data_t *pgdat, int nid) 5733 { 5734 enum zone_type zone_type; 5735 5736 if (N_MEMORY == N_NORMAL_MEMORY) 5737 return; 5738 5739 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { 5740 struct zone *zone = &pgdat->node_zones[zone_type]; 5741 if (populated_zone(zone)) { 5742 node_set_state(nid, N_HIGH_MEMORY); 5743 if (N_NORMAL_MEMORY != N_HIGH_MEMORY && 5744 zone_type <= ZONE_NORMAL) 5745 node_set_state(nid, N_NORMAL_MEMORY); 5746 break; 5747 } 5748 } 5749 } 5750 5751 /** 5752 * free_area_init_nodes - Initialise all pg_data_t and zone data 5753 * @max_zone_pfn: an array of max PFNs for each zone 5754 * 5755 * This will call free_area_init_node() for each active node in the system. 5756 * Using the page ranges provided by memblock_set_node(), the size of each 5757 * zone in each node and their holes is calculated. If the maximum PFN 5758 * between two adjacent zones match, it is assumed that the zone is empty. 5759 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 5760 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 5761 * starts where the previous one ended. For example, ZONE_DMA32 starts 5762 * at arch_max_dma_pfn. 5763 */ 5764 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 5765 { 5766 unsigned long start_pfn, end_pfn; 5767 int i, nid; 5768 5769 /* Record where the zone boundaries are */ 5770 memset(arch_zone_lowest_possible_pfn, 0, 5771 sizeof(arch_zone_lowest_possible_pfn)); 5772 memset(arch_zone_highest_possible_pfn, 0, 5773 sizeof(arch_zone_highest_possible_pfn)); 5774 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 5775 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 5776 for (i = 1; i < MAX_NR_ZONES; i++) { 5777 if (i == ZONE_MOVABLE) 5778 continue; 5779 arch_zone_lowest_possible_pfn[i] = 5780 arch_zone_highest_possible_pfn[i-1]; 5781 arch_zone_highest_possible_pfn[i] = 5782 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 5783 } 5784 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 5785 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 5786 5787 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 5788 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 5789 find_zone_movable_pfns_for_nodes(); 5790 5791 /* Print out the zone ranges */ 5792 pr_info("Zone ranges:\n"); 5793 for (i = 0; i < MAX_NR_ZONES; i++) { 5794 if (i == ZONE_MOVABLE) 5795 continue; 5796 pr_info(" %-8s ", zone_names[i]); 5797 if (arch_zone_lowest_possible_pfn[i] == 5798 arch_zone_highest_possible_pfn[i]) 5799 pr_cont("empty\n"); 5800 else 5801 pr_cont("[mem %#018Lx-%#018Lx]\n", 5802 (u64)arch_zone_lowest_possible_pfn[i] 5803 << PAGE_SHIFT, 5804 ((u64)arch_zone_highest_possible_pfn[i] 5805 << PAGE_SHIFT) - 1); 5806 } 5807 5808 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 5809 pr_info("Movable zone start for each node\n"); 5810 for (i = 0; i < MAX_NUMNODES; i++) { 5811 if (zone_movable_pfn[i]) 5812 pr_info(" Node %d: %#018Lx\n", i, 5813 (u64)zone_movable_pfn[i] << PAGE_SHIFT); 5814 } 5815 5816 /* Print out the early node map */ 5817 pr_info("Early memory node ranges\n"); 5818 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 5819 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid, 5820 (u64)start_pfn << PAGE_SHIFT, 5821 ((u64)end_pfn << PAGE_SHIFT) - 1); 5822 5823 /* Initialise every node */ 5824 mminit_verify_pageflags_layout(); 5825 setup_nr_node_ids(); 5826 for_each_online_node(nid) { 5827 pg_data_t *pgdat = NODE_DATA(nid); 5828 free_area_init_node(nid, NULL, 5829 find_min_pfn_for_node(nid), NULL); 5830 5831 /* Any memory on that node */ 5832 if (pgdat->node_present_pages) 5833 node_set_state(nid, N_MEMORY); 5834 check_for_memory(pgdat, nid); 5835 } 5836 } 5837 5838 static int __init cmdline_parse_core(char *p, unsigned long *core) 5839 { 5840 unsigned long long coremem; 5841 if (!p) 5842 return -EINVAL; 5843 5844 coremem = memparse(p, &p); 5845 *core = coremem >> PAGE_SHIFT; 5846 5847 /* Paranoid check that UL is enough for the coremem value */ 5848 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 5849 5850 return 0; 5851 } 5852 5853 /* 5854 * kernelcore=size sets the amount of memory for use for allocations that 5855 * cannot be reclaimed or migrated. 5856 */ 5857 static int __init cmdline_parse_kernelcore(char *p) 5858 { 5859 return cmdline_parse_core(p, &required_kernelcore); 5860 } 5861 5862 /* 5863 * movablecore=size sets the amount of memory for use for allocations that 5864 * can be reclaimed or migrated. 5865 */ 5866 static int __init cmdline_parse_movablecore(char *p) 5867 { 5868 return cmdline_parse_core(p, &required_movablecore); 5869 } 5870 5871 early_param("kernelcore", cmdline_parse_kernelcore); 5872 early_param("movablecore", cmdline_parse_movablecore); 5873 5874 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5875 5876 void adjust_managed_page_count(struct page *page, long count) 5877 { 5878 spin_lock(&managed_page_count_lock); 5879 page_zone(page)->managed_pages += count; 5880 totalram_pages += count; 5881 #ifdef CONFIG_HIGHMEM 5882 if (PageHighMem(page)) 5883 totalhigh_pages += count; 5884 #endif 5885 spin_unlock(&managed_page_count_lock); 5886 } 5887 EXPORT_SYMBOL(adjust_managed_page_count); 5888 5889 unsigned long free_reserved_area(void *start, void *end, int poison, char *s) 5890 { 5891 void *pos; 5892 unsigned long pages = 0; 5893 5894 start = (void *)PAGE_ALIGN((unsigned long)start); 5895 end = (void *)((unsigned long)end & PAGE_MASK); 5896 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { 5897 if ((unsigned int)poison <= 0xFF) 5898 memset(pos, poison, PAGE_SIZE); 5899 free_reserved_page(virt_to_page(pos)); 5900 } 5901 5902 if (pages && s) 5903 pr_info("Freeing %s memory: %ldK (%p - %p)\n", 5904 s, pages << (PAGE_SHIFT - 10), start, end); 5905 5906 return pages; 5907 } 5908 EXPORT_SYMBOL(free_reserved_area); 5909 5910 #ifdef CONFIG_HIGHMEM 5911 void free_highmem_page(struct page *page) 5912 { 5913 __free_reserved_page(page); 5914 totalram_pages++; 5915 page_zone(page)->managed_pages++; 5916 totalhigh_pages++; 5917 } 5918 #endif 5919 5920 5921 void __init mem_init_print_info(const char *str) 5922 { 5923 unsigned long physpages, codesize, datasize, rosize, bss_size; 5924 unsigned long init_code_size, init_data_size; 5925 5926 physpages = get_num_physpages(); 5927 codesize = _etext - _stext; 5928 datasize = _edata - _sdata; 5929 rosize = __end_rodata - __start_rodata; 5930 bss_size = __bss_stop - __bss_start; 5931 init_data_size = __init_end - __init_begin; 5932 init_code_size = _einittext - _sinittext; 5933 5934 /* 5935 * Detect special cases and adjust section sizes accordingly: 5936 * 1) .init.* may be embedded into .data sections 5937 * 2) .init.text.* may be out of [__init_begin, __init_end], 5938 * please refer to arch/tile/kernel/vmlinux.lds.S. 5939 * 3) .rodata.* may be embedded into .text or .data sections. 5940 */ 5941 #define adj_init_size(start, end, size, pos, adj) \ 5942 do { \ 5943 if (start <= pos && pos < end && size > adj) \ 5944 size -= adj; \ 5945 } while (0) 5946 5947 adj_init_size(__init_begin, __init_end, init_data_size, 5948 _sinittext, init_code_size); 5949 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); 5950 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); 5951 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); 5952 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); 5953 5954 #undef adj_init_size 5955 5956 pr_info("Memory: %luK/%luK available " 5957 "(%luK kernel code, %luK rwdata, %luK rodata, " 5958 "%luK init, %luK bss, %luK reserved, %luK cma-reserved" 5959 #ifdef CONFIG_HIGHMEM 5960 ", %luK highmem" 5961 #endif 5962 "%s%s)\n", 5963 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10), 5964 codesize >> 10, datasize >> 10, rosize >> 10, 5965 (init_data_size + init_code_size) >> 10, bss_size >> 10, 5966 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10), 5967 totalcma_pages << (PAGE_SHIFT-10), 5968 #ifdef CONFIG_HIGHMEM 5969 totalhigh_pages << (PAGE_SHIFT-10), 5970 #endif 5971 str ? ", " : "", str ? str : ""); 5972 } 5973 5974 /** 5975 * set_dma_reserve - set the specified number of pages reserved in the first zone 5976 * @new_dma_reserve: The number of pages to mark reserved 5977 * 5978 * The per-cpu batchsize and zone watermarks are determined by present_pages. 5979 * In the DMA zone, a significant percentage may be consumed by kernel image 5980 * and other unfreeable allocations which can skew the watermarks badly. This 5981 * function may optionally be used to account for unfreeable pages in the 5982 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 5983 * smaller per-cpu batchsize. 5984 */ 5985 void __init set_dma_reserve(unsigned long new_dma_reserve) 5986 { 5987 dma_reserve = new_dma_reserve; 5988 } 5989 5990 void __init free_area_init(unsigned long *zones_size) 5991 { 5992 free_area_init_node(0, zones_size, 5993 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 5994 } 5995 5996 static int page_alloc_cpu_notify(struct notifier_block *self, 5997 unsigned long action, void *hcpu) 5998 { 5999 int cpu = (unsigned long)hcpu; 6000 6001 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 6002 lru_add_drain_cpu(cpu); 6003 drain_pages(cpu); 6004 6005 /* 6006 * Spill the event counters of the dead processor 6007 * into the current processors event counters. 6008 * This artificially elevates the count of the current 6009 * processor. 6010 */ 6011 vm_events_fold_cpu(cpu); 6012 6013 /* 6014 * Zero the differential counters of the dead processor 6015 * so that the vm statistics are consistent. 6016 * 6017 * This is only okay since the processor is dead and cannot 6018 * race with what we are doing. 6019 */ 6020 cpu_vm_stats_fold(cpu); 6021 } 6022 return NOTIFY_OK; 6023 } 6024 6025 void __init page_alloc_init(void) 6026 { 6027 hotcpu_notifier(page_alloc_cpu_notify, 0); 6028 } 6029 6030 /* 6031 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 6032 * or min_free_kbytes changes. 6033 */ 6034 static void calculate_totalreserve_pages(void) 6035 { 6036 struct pglist_data *pgdat; 6037 unsigned long reserve_pages = 0; 6038 enum zone_type i, j; 6039 6040 for_each_online_pgdat(pgdat) { 6041 for (i = 0; i < MAX_NR_ZONES; i++) { 6042 struct zone *zone = pgdat->node_zones + i; 6043 long max = 0; 6044 6045 /* Find valid and maximum lowmem_reserve in the zone */ 6046 for (j = i; j < MAX_NR_ZONES; j++) { 6047 if (zone->lowmem_reserve[j] > max) 6048 max = zone->lowmem_reserve[j]; 6049 } 6050 6051 /* we treat the high watermark as reserved pages. */ 6052 max += high_wmark_pages(zone); 6053 6054 if (max > zone->managed_pages) 6055 max = zone->managed_pages; 6056 reserve_pages += max; 6057 /* 6058 * Lowmem reserves are not available to 6059 * GFP_HIGHUSER page cache allocations and 6060 * kswapd tries to balance zones to their high 6061 * watermark. As a result, neither should be 6062 * regarded as dirtyable memory, to prevent a 6063 * situation where reclaim has to clean pages 6064 * in order to balance the zones. 6065 */ 6066 zone->dirty_balance_reserve = max; 6067 } 6068 } 6069 dirty_balance_reserve = reserve_pages; 6070 totalreserve_pages = reserve_pages; 6071 } 6072 6073 /* 6074 * setup_per_zone_lowmem_reserve - called whenever 6075 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 6076 * has a correct pages reserved value, so an adequate number of 6077 * pages are left in the zone after a successful __alloc_pages(). 6078 */ 6079 static void setup_per_zone_lowmem_reserve(void) 6080 { 6081 struct pglist_data *pgdat; 6082 enum zone_type j, idx; 6083 6084 for_each_online_pgdat(pgdat) { 6085 for (j = 0; j < MAX_NR_ZONES; j++) { 6086 struct zone *zone = pgdat->node_zones + j; 6087 unsigned long managed_pages = zone->managed_pages; 6088 6089 zone->lowmem_reserve[j] = 0; 6090 6091 idx = j; 6092 while (idx) { 6093 struct zone *lower_zone; 6094 6095 idx--; 6096 6097 if (sysctl_lowmem_reserve_ratio[idx] < 1) 6098 sysctl_lowmem_reserve_ratio[idx] = 1; 6099 6100 lower_zone = pgdat->node_zones + idx; 6101 lower_zone->lowmem_reserve[j] = managed_pages / 6102 sysctl_lowmem_reserve_ratio[idx]; 6103 managed_pages += lower_zone->managed_pages; 6104 } 6105 } 6106 } 6107 6108 /* update totalreserve_pages */ 6109 calculate_totalreserve_pages(); 6110 } 6111 6112 static void __setup_per_zone_wmarks(void) 6113 { 6114 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 6115 unsigned long lowmem_pages = 0; 6116 struct zone *zone; 6117 unsigned long flags; 6118 6119 /* Calculate total number of !ZONE_HIGHMEM pages */ 6120 for_each_zone(zone) { 6121 if (!is_highmem(zone)) 6122 lowmem_pages += zone->managed_pages; 6123 } 6124 6125 for_each_zone(zone) { 6126 u64 tmp; 6127 6128 spin_lock_irqsave(&zone->lock, flags); 6129 tmp = (u64)pages_min * zone->managed_pages; 6130 do_div(tmp, lowmem_pages); 6131 if (is_highmem(zone)) { 6132 /* 6133 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 6134 * need highmem pages, so cap pages_min to a small 6135 * value here. 6136 * 6137 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 6138 * deltas control asynch page reclaim, and so should 6139 * not be capped for highmem. 6140 */ 6141 unsigned long min_pages; 6142 6143 min_pages = zone->managed_pages / 1024; 6144 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); 6145 zone->watermark[WMARK_MIN] = min_pages; 6146 } else { 6147 /* 6148 * If it's a lowmem zone, reserve a number of pages 6149 * proportionate to the zone's size. 6150 */ 6151 zone->watermark[WMARK_MIN] = tmp; 6152 } 6153 6154 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 6155 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 6156 6157 __mod_zone_page_state(zone, NR_ALLOC_BATCH, 6158 high_wmark_pages(zone) - low_wmark_pages(zone) - 6159 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 6160 6161 setup_zone_migrate_reserve(zone); 6162 spin_unlock_irqrestore(&zone->lock, flags); 6163 } 6164 6165 /* update totalreserve_pages */ 6166 calculate_totalreserve_pages(); 6167 } 6168 6169 /** 6170 * setup_per_zone_wmarks - called when min_free_kbytes changes 6171 * or when memory is hot-{added|removed} 6172 * 6173 * Ensures that the watermark[min,low,high] values for each zone are set 6174 * correctly with respect to min_free_kbytes. 6175 */ 6176 void setup_per_zone_wmarks(void) 6177 { 6178 mutex_lock(&zonelists_mutex); 6179 __setup_per_zone_wmarks(); 6180 mutex_unlock(&zonelists_mutex); 6181 } 6182 6183 /* 6184 * The inactive anon list should be small enough that the VM never has to 6185 * do too much work, but large enough that each inactive page has a chance 6186 * to be referenced again before it is swapped out. 6187 * 6188 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 6189 * INACTIVE_ANON pages on this zone's LRU, maintained by the 6190 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 6191 * the anonymous pages are kept on the inactive list. 6192 * 6193 * total target max 6194 * memory ratio inactive anon 6195 * ------------------------------------- 6196 * 10MB 1 5MB 6197 * 100MB 1 50MB 6198 * 1GB 3 250MB 6199 * 10GB 10 0.9GB 6200 * 100GB 31 3GB 6201 * 1TB 101 10GB 6202 * 10TB 320 32GB 6203 */ 6204 static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 6205 { 6206 unsigned int gb, ratio; 6207 6208 /* Zone size in gigabytes */ 6209 gb = zone->managed_pages >> (30 - PAGE_SHIFT); 6210 if (gb) 6211 ratio = int_sqrt(10 * gb); 6212 else 6213 ratio = 1; 6214 6215 zone->inactive_ratio = ratio; 6216 } 6217 6218 static void __meminit setup_per_zone_inactive_ratio(void) 6219 { 6220 struct zone *zone; 6221 6222 for_each_zone(zone) 6223 calculate_zone_inactive_ratio(zone); 6224 } 6225 6226 /* 6227 * Initialise min_free_kbytes. 6228 * 6229 * For small machines we want it small (128k min). For large machines 6230 * we want it large (64MB max). But it is not linear, because network 6231 * bandwidth does not increase linearly with machine size. We use 6232 * 6233 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 6234 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 6235 * 6236 * which yields 6237 * 6238 * 16MB: 512k 6239 * 32MB: 724k 6240 * 64MB: 1024k 6241 * 128MB: 1448k 6242 * 256MB: 2048k 6243 * 512MB: 2896k 6244 * 1024MB: 4096k 6245 * 2048MB: 5792k 6246 * 4096MB: 8192k 6247 * 8192MB: 11584k 6248 * 16384MB: 16384k 6249 */ 6250 int __meminit init_per_zone_wmark_min(void) 6251 { 6252 unsigned long lowmem_kbytes; 6253 int new_min_free_kbytes; 6254 6255 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 6256 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 6257 6258 if (new_min_free_kbytes > user_min_free_kbytes) { 6259 min_free_kbytes = new_min_free_kbytes; 6260 if (min_free_kbytes < 128) 6261 min_free_kbytes = 128; 6262 if (min_free_kbytes > 65536) 6263 min_free_kbytes = 65536; 6264 } else { 6265 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", 6266 new_min_free_kbytes, user_min_free_kbytes); 6267 } 6268 setup_per_zone_wmarks(); 6269 refresh_zone_stat_thresholds(); 6270 setup_per_zone_lowmem_reserve(); 6271 setup_per_zone_inactive_ratio(); 6272 return 0; 6273 } 6274 module_init(init_per_zone_wmark_min) 6275 6276 /* 6277 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 6278 * that we can call two helper functions whenever min_free_kbytes 6279 * changes. 6280 */ 6281 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, 6282 void __user *buffer, size_t *length, loff_t *ppos) 6283 { 6284 int rc; 6285 6286 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6287 if (rc) 6288 return rc; 6289 6290 if (write) { 6291 user_min_free_kbytes = min_free_kbytes; 6292 setup_per_zone_wmarks(); 6293 } 6294 return 0; 6295 } 6296 6297 #ifdef CONFIG_NUMA 6298 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, 6299 void __user *buffer, size_t *length, loff_t *ppos) 6300 { 6301 struct zone *zone; 6302 int rc; 6303 6304 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6305 if (rc) 6306 return rc; 6307 6308 for_each_zone(zone) 6309 zone->min_unmapped_pages = (zone->managed_pages * 6310 sysctl_min_unmapped_ratio) / 100; 6311 return 0; 6312 } 6313 6314 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, 6315 void __user *buffer, size_t *length, loff_t *ppos) 6316 { 6317 struct zone *zone; 6318 int rc; 6319 6320 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6321 if (rc) 6322 return rc; 6323 6324 for_each_zone(zone) 6325 zone->min_slab_pages = (zone->managed_pages * 6326 sysctl_min_slab_ratio) / 100; 6327 return 0; 6328 } 6329 #endif 6330 6331 /* 6332 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 6333 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 6334 * whenever sysctl_lowmem_reserve_ratio changes. 6335 * 6336 * The reserve ratio obviously has absolutely no relation with the 6337 * minimum watermarks. The lowmem reserve ratio can only make sense 6338 * if in function of the boot time zone sizes. 6339 */ 6340 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write, 6341 void __user *buffer, size_t *length, loff_t *ppos) 6342 { 6343 proc_dointvec_minmax(table, write, buffer, length, ppos); 6344 setup_per_zone_lowmem_reserve(); 6345 return 0; 6346 } 6347 6348 /* 6349 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 6350 * cpu. It is the fraction of total pages in each zone that a hot per cpu 6351 * pagelist can have before it gets flushed back to buddy allocator. 6352 */ 6353 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, 6354 void __user *buffer, size_t *length, loff_t *ppos) 6355 { 6356 struct zone *zone; 6357 int old_percpu_pagelist_fraction; 6358 int ret; 6359 6360 mutex_lock(&pcp_batch_high_lock); 6361 old_percpu_pagelist_fraction = percpu_pagelist_fraction; 6362 6363 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 6364 if (!write || ret < 0) 6365 goto out; 6366 6367 /* Sanity checking to avoid pcp imbalance */ 6368 if (percpu_pagelist_fraction && 6369 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) { 6370 percpu_pagelist_fraction = old_percpu_pagelist_fraction; 6371 ret = -EINVAL; 6372 goto out; 6373 } 6374 6375 /* No change? */ 6376 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction) 6377 goto out; 6378 6379 for_each_populated_zone(zone) { 6380 unsigned int cpu; 6381 6382 for_each_possible_cpu(cpu) 6383 pageset_set_high_and_batch(zone, 6384 per_cpu_ptr(zone->pageset, cpu)); 6385 } 6386 out: 6387 mutex_unlock(&pcp_batch_high_lock); 6388 return ret; 6389 } 6390 6391 #ifdef CONFIG_NUMA 6392 int hashdist = HASHDIST_DEFAULT; 6393 6394 static int __init set_hashdist(char *str) 6395 { 6396 if (!str) 6397 return 0; 6398 hashdist = simple_strtoul(str, &str, 0); 6399 return 1; 6400 } 6401 __setup("hashdist=", set_hashdist); 6402 #endif 6403 6404 /* 6405 * allocate a large system hash table from bootmem 6406 * - it is assumed that the hash table must contain an exact power-of-2 6407 * quantity of entries 6408 * - limit is the number of hash buckets, not the total allocation size 6409 */ 6410 void *__init alloc_large_system_hash(const char *tablename, 6411 unsigned long bucketsize, 6412 unsigned long numentries, 6413 int scale, 6414 int flags, 6415 unsigned int *_hash_shift, 6416 unsigned int *_hash_mask, 6417 unsigned long low_limit, 6418 unsigned long high_limit) 6419 { 6420 unsigned long long max = high_limit; 6421 unsigned long log2qty, size; 6422 void *table = NULL; 6423 6424 /* allow the kernel cmdline to have a say */ 6425 if (!numentries) { 6426 /* round applicable memory size up to nearest megabyte */ 6427 numentries = nr_kernel_pages; 6428 6429 /* It isn't necessary when PAGE_SIZE >= 1MB */ 6430 if (PAGE_SHIFT < 20) 6431 numentries = round_up(numentries, (1<<20)/PAGE_SIZE); 6432 6433 /* limit to 1 bucket per 2^scale bytes of low memory */ 6434 if (scale > PAGE_SHIFT) 6435 numentries >>= (scale - PAGE_SHIFT); 6436 else 6437 numentries <<= (PAGE_SHIFT - scale); 6438 6439 /* Make sure we've got at least a 0-order allocation.. */ 6440 if (unlikely(flags & HASH_SMALL)) { 6441 /* Makes no sense without HASH_EARLY */ 6442 WARN_ON(!(flags & HASH_EARLY)); 6443 if (!(numentries >> *_hash_shift)) { 6444 numentries = 1UL << *_hash_shift; 6445 BUG_ON(!numentries); 6446 } 6447 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 6448 numentries = PAGE_SIZE / bucketsize; 6449 } 6450 numentries = roundup_pow_of_two(numentries); 6451 6452 /* limit allocation size to 1/16 total memory by default */ 6453 if (max == 0) { 6454 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 6455 do_div(max, bucketsize); 6456 } 6457 max = min(max, 0x80000000ULL); 6458 6459 if (numentries < low_limit) 6460 numentries = low_limit; 6461 if (numentries > max) 6462 numentries = max; 6463 6464 log2qty = ilog2(numentries); 6465 6466 do { 6467 size = bucketsize << log2qty; 6468 if (flags & HASH_EARLY) 6469 table = memblock_virt_alloc_nopanic(size, 0); 6470 else if (hashdist) 6471 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 6472 else { 6473 /* 6474 * If bucketsize is not a power-of-two, we may free 6475 * some pages at the end of hash table which 6476 * alloc_pages_exact() automatically does 6477 */ 6478 if (get_order(size) < MAX_ORDER) { 6479 table = alloc_pages_exact(size, GFP_ATOMIC); 6480 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 6481 } 6482 } 6483 } while (!table && size > PAGE_SIZE && --log2qty); 6484 6485 if (!table) 6486 panic("Failed to allocate %s hash table\n", tablename); 6487 6488 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 6489 tablename, 6490 (1UL << log2qty), 6491 ilog2(size) - PAGE_SHIFT, 6492 size); 6493 6494 if (_hash_shift) 6495 *_hash_shift = log2qty; 6496 if (_hash_mask) 6497 *_hash_mask = (1 << log2qty) - 1; 6498 6499 return table; 6500 } 6501 6502 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 6503 static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 6504 unsigned long pfn) 6505 { 6506 #ifdef CONFIG_SPARSEMEM 6507 return __pfn_to_section(pfn)->pageblock_flags; 6508 #else 6509 return zone->pageblock_flags; 6510 #endif /* CONFIG_SPARSEMEM */ 6511 } 6512 6513 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 6514 { 6515 #ifdef CONFIG_SPARSEMEM 6516 pfn &= (PAGES_PER_SECTION-1); 6517 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6518 #else 6519 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); 6520 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6521 #endif /* CONFIG_SPARSEMEM */ 6522 } 6523 6524 /** 6525 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages 6526 * @page: The page within the block of interest 6527 * @pfn: The target page frame number 6528 * @end_bitidx: The last bit of interest to retrieve 6529 * @mask: mask of bits that the caller is interested in 6530 * 6531 * Return: pageblock_bits flags 6532 */ 6533 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, 6534 unsigned long end_bitidx, 6535 unsigned long mask) 6536 { 6537 struct zone *zone; 6538 unsigned long *bitmap; 6539 unsigned long bitidx, word_bitidx; 6540 unsigned long word; 6541 6542 zone = page_zone(page); 6543 bitmap = get_pageblock_bitmap(zone, pfn); 6544 bitidx = pfn_to_bitidx(zone, pfn); 6545 word_bitidx = bitidx / BITS_PER_LONG; 6546 bitidx &= (BITS_PER_LONG-1); 6547 6548 word = bitmap[word_bitidx]; 6549 bitidx += end_bitidx; 6550 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; 6551 } 6552 6553 /** 6554 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages 6555 * @page: The page within the block of interest 6556 * @flags: The flags to set 6557 * @pfn: The target page frame number 6558 * @end_bitidx: The last bit of interest 6559 * @mask: mask of bits that the caller is interested in 6560 */ 6561 void set_pfnblock_flags_mask(struct page *page, unsigned long flags, 6562 unsigned long pfn, 6563 unsigned long end_bitidx, 6564 unsigned long mask) 6565 { 6566 struct zone *zone; 6567 unsigned long *bitmap; 6568 unsigned long bitidx, word_bitidx; 6569 unsigned long old_word, word; 6570 6571 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); 6572 6573 zone = page_zone(page); 6574 bitmap = get_pageblock_bitmap(zone, pfn); 6575 bitidx = pfn_to_bitidx(zone, pfn); 6576 word_bitidx = bitidx / BITS_PER_LONG; 6577 bitidx &= (BITS_PER_LONG-1); 6578 6579 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page); 6580 6581 bitidx += end_bitidx; 6582 mask <<= (BITS_PER_LONG - bitidx - 1); 6583 flags <<= (BITS_PER_LONG - bitidx - 1); 6584 6585 word = READ_ONCE(bitmap[word_bitidx]); 6586 for (;;) { 6587 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); 6588 if (word == old_word) 6589 break; 6590 word = old_word; 6591 } 6592 } 6593 6594 /* 6595 * This function checks whether pageblock includes unmovable pages or not. 6596 * If @count is not zero, it is okay to include less @count unmovable pages 6597 * 6598 * PageLRU check without isolation or lru_lock could race so that 6599 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't 6600 * expect this function should be exact. 6601 */ 6602 bool has_unmovable_pages(struct zone *zone, struct page *page, int count, 6603 bool skip_hwpoisoned_pages) 6604 { 6605 unsigned long pfn, iter, found; 6606 int mt; 6607 6608 /* 6609 * For avoiding noise data, lru_add_drain_all() should be called 6610 * If ZONE_MOVABLE, the zone never contains unmovable pages 6611 */ 6612 if (zone_idx(zone) == ZONE_MOVABLE) 6613 return false; 6614 mt = get_pageblock_migratetype(page); 6615 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) 6616 return false; 6617 6618 pfn = page_to_pfn(page); 6619 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 6620 unsigned long check = pfn + iter; 6621 6622 if (!pfn_valid_within(check)) 6623 continue; 6624 6625 page = pfn_to_page(check); 6626 6627 /* 6628 * Hugepages are not in LRU lists, but they're movable. 6629 * We need not scan over tail pages bacause we don't 6630 * handle each tail page individually in migration. 6631 */ 6632 if (PageHuge(page)) { 6633 iter = round_up(iter + 1, 1<<compound_order(page)) - 1; 6634 continue; 6635 } 6636 6637 /* 6638 * We can't use page_count without pin a page 6639 * because another CPU can free compound page. 6640 * This check already skips compound tails of THP 6641 * because their page->_count is zero at all time. 6642 */ 6643 if (!atomic_read(&page->_count)) { 6644 if (PageBuddy(page)) 6645 iter += (1 << page_order(page)) - 1; 6646 continue; 6647 } 6648 6649 /* 6650 * The HWPoisoned page may be not in buddy system, and 6651 * page_count() is not 0. 6652 */ 6653 if (skip_hwpoisoned_pages && PageHWPoison(page)) 6654 continue; 6655 6656 if (!PageLRU(page)) 6657 found++; 6658 /* 6659 * If there are RECLAIMABLE pages, we need to check 6660 * it. But now, memory offline itself doesn't call 6661 * shrink_node_slabs() and it still to be fixed. 6662 */ 6663 /* 6664 * If the page is not RAM, page_count()should be 0. 6665 * we don't need more check. This is an _used_ not-movable page. 6666 * 6667 * The problematic thing here is PG_reserved pages. PG_reserved 6668 * is set to both of a memory hole page and a _used_ kernel 6669 * page at boot. 6670 */ 6671 if (found > count) 6672 return true; 6673 } 6674 return false; 6675 } 6676 6677 bool is_pageblock_removable_nolock(struct page *page) 6678 { 6679 struct zone *zone; 6680 unsigned long pfn; 6681 6682 /* 6683 * We have to be careful here because we are iterating over memory 6684 * sections which are not zone aware so we might end up outside of 6685 * the zone but still within the section. 6686 * We have to take care about the node as well. If the node is offline 6687 * its NODE_DATA will be NULL - see page_zone. 6688 */ 6689 if (!node_online(page_to_nid(page))) 6690 return false; 6691 6692 zone = page_zone(page); 6693 pfn = page_to_pfn(page); 6694 if (!zone_spans_pfn(zone, pfn)) 6695 return false; 6696 6697 return !has_unmovable_pages(zone, page, 0, true); 6698 } 6699 6700 #ifdef CONFIG_CMA 6701 6702 static unsigned long pfn_max_align_down(unsigned long pfn) 6703 { 6704 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, 6705 pageblock_nr_pages) - 1); 6706 } 6707 6708 static unsigned long pfn_max_align_up(unsigned long pfn) 6709 { 6710 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, 6711 pageblock_nr_pages)); 6712 } 6713 6714 /* [start, end) must belong to a single zone. */ 6715 static int __alloc_contig_migrate_range(struct compact_control *cc, 6716 unsigned long start, unsigned long end) 6717 { 6718 /* This function is based on compact_zone() from compaction.c. */ 6719 unsigned long nr_reclaimed; 6720 unsigned long pfn = start; 6721 unsigned int tries = 0; 6722 int ret = 0; 6723 6724 migrate_prep(); 6725 6726 while (pfn < end || !list_empty(&cc->migratepages)) { 6727 if (fatal_signal_pending(current)) { 6728 ret = -EINTR; 6729 break; 6730 } 6731 6732 if (list_empty(&cc->migratepages)) { 6733 cc->nr_migratepages = 0; 6734 pfn = isolate_migratepages_range(cc, pfn, end); 6735 if (!pfn) { 6736 ret = -EINTR; 6737 break; 6738 } 6739 tries = 0; 6740 } else if (++tries == 5) { 6741 ret = ret < 0 ? ret : -EBUSY; 6742 break; 6743 } 6744 6745 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, 6746 &cc->migratepages); 6747 cc->nr_migratepages -= nr_reclaimed; 6748 6749 ret = migrate_pages(&cc->migratepages, alloc_migrate_target, 6750 NULL, 0, cc->mode, MR_CMA); 6751 } 6752 if (ret < 0) { 6753 putback_movable_pages(&cc->migratepages); 6754 return ret; 6755 } 6756 return 0; 6757 } 6758 6759 /** 6760 * alloc_contig_range() -- tries to allocate given range of pages 6761 * @start: start PFN to allocate 6762 * @end: one-past-the-last PFN to allocate 6763 * @migratetype: migratetype of the underlaying pageblocks (either 6764 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks 6765 * in range must have the same migratetype and it must 6766 * be either of the two. 6767 * 6768 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES 6769 * aligned, however it's the caller's responsibility to guarantee that 6770 * we are the only thread that changes migrate type of pageblocks the 6771 * pages fall in. 6772 * 6773 * The PFN range must belong to a single zone. 6774 * 6775 * Returns zero on success or negative error code. On success all 6776 * pages which PFN is in [start, end) are allocated for the caller and 6777 * need to be freed with free_contig_range(). 6778 */ 6779 int alloc_contig_range(unsigned long start, unsigned long end, 6780 unsigned migratetype) 6781 { 6782 unsigned long outer_start, outer_end; 6783 int ret = 0, order; 6784 6785 struct compact_control cc = { 6786 .nr_migratepages = 0, 6787 .order = -1, 6788 .zone = page_zone(pfn_to_page(start)), 6789 .mode = MIGRATE_SYNC, 6790 .ignore_skip_hint = true, 6791 }; 6792 INIT_LIST_HEAD(&cc.migratepages); 6793 6794 /* 6795 * What we do here is we mark all pageblocks in range as 6796 * MIGRATE_ISOLATE. Because pageblock and max order pages may 6797 * have different sizes, and due to the way page allocator 6798 * work, we align the range to biggest of the two pages so 6799 * that page allocator won't try to merge buddies from 6800 * different pageblocks and change MIGRATE_ISOLATE to some 6801 * other migration type. 6802 * 6803 * Once the pageblocks are marked as MIGRATE_ISOLATE, we 6804 * migrate the pages from an unaligned range (ie. pages that 6805 * we are interested in). This will put all the pages in 6806 * range back to page allocator as MIGRATE_ISOLATE. 6807 * 6808 * When this is done, we take the pages in range from page 6809 * allocator removing them from the buddy system. This way 6810 * page allocator will never consider using them. 6811 * 6812 * This lets us mark the pageblocks back as 6813 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the 6814 * aligned range but not in the unaligned, original range are 6815 * put back to page allocator so that buddy can use them. 6816 */ 6817 6818 ret = start_isolate_page_range(pfn_max_align_down(start), 6819 pfn_max_align_up(end), migratetype, 6820 false); 6821 if (ret) 6822 return ret; 6823 6824 ret = __alloc_contig_migrate_range(&cc, start, end); 6825 if (ret) 6826 goto done; 6827 6828 /* 6829 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES 6830 * aligned blocks that are marked as MIGRATE_ISOLATE. What's 6831 * more, all pages in [start, end) are free in page allocator. 6832 * What we are going to do is to allocate all pages from 6833 * [start, end) (that is remove them from page allocator). 6834 * 6835 * The only problem is that pages at the beginning and at the 6836 * end of interesting range may be not aligned with pages that 6837 * page allocator holds, ie. they can be part of higher order 6838 * pages. Because of this, we reserve the bigger range and 6839 * once this is done free the pages we are not interested in. 6840 * 6841 * We don't have to hold zone->lock here because the pages are 6842 * isolated thus they won't get removed from buddy. 6843 */ 6844 6845 lru_add_drain_all(); 6846 drain_all_pages(cc.zone); 6847 6848 order = 0; 6849 outer_start = start; 6850 while (!PageBuddy(pfn_to_page(outer_start))) { 6851 if (++order >= MAX_ORDER) { 6852 ret = -EBUSY; 6853 goto done; 6854 } 6855 outer_start &= ~0UL << order; 6856 } 6857 6858 /* Make sure the range is really isolated. */ 6859 if (test_pages_isolated(outer_start, end, false)) { 6860 pr_info("%s: [%lx, %lx) PFNs busy\n", 6861 __func__, outer_start, end); 6862 ret = -EBUSY; 6863 goto done; 6864 } 6865 6866 /* Grab isolated pages from freelists. */ 6867 outer_end = isolate_freepages_range(&cc, outer_start, end); 6868 if (!outer_end) { 6869 ret = -EBUSY; 6870 goto done; 6871 } 6872 6873 /* Free head and tail (if any) */ 6874 if (start != outer_start) 6875 free_contig_range(outer_start, start - outer_start); 6876 if (end != outer_end) 6877 free_contig_range(end, outer_end - end); 6878 6879 done: 6880 undo_isolate_page_range(pfn_max_align_down(start), 6881 pfn_max_align_up(end), migratetype); 6882 return ret; 6883 } 6884 6885 void free_contig_range(unsigned long pfn, unsigned nr_pages) 6886 { 6887 unsigned int count = 0; 6888 6889 for (; nr_pages--; pfn++) { 6890 struct page *page = pfn_to_page(pfn); 6891 6892 count += page_count(page) != 1; 6893 __free_page(page); 6894 } 6895 WARN(count != 0, "%d pages are still in use!\n", count); 6896 } 6897 #endif 6898 6899 #ifdef CONFIG_MEMORY_HOTPLUG 6900 /* 6901 * The zone indicated has a new number of managed_pages; batch sizes and percpu 6902 * page high values need to be recalulated. 6903 */ 6904 void __meminit zone_pcp_update(struct zone *zone) 6905 { 6906 unsigned cpu; 6907 mutex_lock(&pcp_batch_high_lock); 6908 for_each_possible_cpu(cpu) 6909 pageset_set_high_and_batch(zone, 6910 per_cpu_ptr(zone->pageset, cpu)); 6911 mutex_unlock(&pcp_batch_high_lock); 6912 } 6913 #endif 6914 6915 void zone_pcp_reset(struct zone *zone) 6916 { 6917 unsigned long flags; 6918 int cpu; 6919 struct per_cpu_pageset *pset; 6920 6921 /* avoid races with drain_pages() */ 6922 local_irq_save(flags); 6923 if (zone->pageset != &boot_pageset) { 6924 for_each_online_cpu(cpu) { 6925 pset = per_cpu_ptr(zone->pageset, cpu); 6926 drain_zonestat(zone, pset); 6927 } 6928 free_percpu(zone->pageset); 6929 zone->pageset = &boot_pageset; 6930 } 6931 local_irq_restore(flags); 6932 } 6933 6934 #ifdef CONFIG_MEMORY_HOTREMOVE 6935 /* 6936 * All pages in the range must be isolated before calling this. 6937 */ 6938 void 6939 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 6940 { 6941 struct page *page; 6942 struct zone *zone; 6943 unsigned int order, i; 6944 unsigned long pfn; 6945 unsigned long flags; 6946 /* find the first valid pfn */ 6947 for (pfn = start_pfn; pfn < end_pfn; pfn++) 6948 if (pfn_valid(pfn)) 6949 break; 6950 if (pfn == end_pfn) 6951 return; 6952 zone = page_zone(pfn_to_page(pfn)); 6953 spin_lock_irqsave(&zone->lock, flags); 6954 pfn = start_pfn; 6955 while (pfn < end_pfn) { 6956 if (!pfn_valid(pfn)) { 6957 pfn++; 6958 continue; 6959 } 6960 page = pfn_to_page(pfn); 6961 /* 6962 * The HWPoisoned page may be not in buddy system, and 6963 * page_count() is not 0. 6964 */ 6965 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { 6966 pfn++; 6967 SetPageReserved(page); 6968 continue; 6969 } 6970 6971 BUG_ON(page_count(page)); 6972 BUG_ON(!PageBuddy(page)); 6973 order = page_order(page); 6974 #ifdef CONFIG_DEBUG_VM 6975 printk(KERN_INFO "remove from free list %lx %d %lx\n", 6976 pfn, 1 << order, end_pfn); 6977 #endif 6978 list_del(&page->lru); 6979 rmv_page_order(page); 6980 zone->free_area[order].nr_free--; 6981 for (i = 0; i < (1 << order); i++) 6982 SetPageReserved((page+i)); 6983 pfn += (1 << order); 6984 } 6985 spin_unlock_irqrestore(&zone->lock, flags); 6986 } 6987 #endif 6988 6989 #ifdef CONFIG_MEMORY_FAILURE 6990 bool is_free_buddy_page(struct page *page) 6991 { 6992 struct zone *zone = page_zone(page); 6993 unsigned long pfn = page_to_pfn(page); 6994 unsigned long flags; 6995 unsigned int order; 6996 6997 spin_lock_irqsave(&zone->lock, flags); 6998 for (order = 0; order < MAX_ORDER; order++) { 6999 struct page *page_head = page - (pfn & ((1 << order) - 1)); 7000 7001 if (PageBuddy(page_head) && page_order(page_head) >= order) 7002 break; 7003 } 7004 spin_unlock_irqrestore(&zone->lock, flags); 7005 7006 return order < MAX_ORDER; 7007 } 7008 #endif 7009