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 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn) 250 return true; 251 252 return false; 253 } 254 255 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid) 256 { 257 if (pfn >= NODE_DATA(nid)->first_deferred_pfn) 258 return true; 259 260 return false; 261 } 262 263 /* 264 * Returns false when the remaining initialisation should be deferred until 265 * later in the boot cycle when it can be parallelised. 266 */ 267 static inline bool update_defer_init(pg_data_t *pgdat, 268 unsigned long pfn, unsigned long zone_end, 269 unsigned long *nr_initialised) 270 { 271 /* Always populate low zones for address-contrained allocations */ 272 if (zone_end < pgdat_end_pfn(pgdat)) 273 return true; 274 275 /* Initialise at least 2G of the highest zone */ 276 (*nr_initialised)++; 277 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) && 278 (pfn & (PAGES_PER_SECTION - 1)) == 0) { 279 pgdat->first_deferred_pfn = pfn; 280 return false; 281 } 282 283 return true; 284 } 285 #else 286 static inline void reset_deferred_meminit(pg_data_t *pgdat) 287 { 288 } 289 290 static inline bool early_page_uninitialised(unsigned long pfn) 291 { 292 return false; 293 } 294 295 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid) 296 { 297 return false; 298 } 299 300 static inline bool update_defer_init(pg_data_t *pgdat, 301 unsigned long pfn, unsigned long zone_end, 302 unsigned long *nr_initialised) 303 { 304 return true; 305 } 306 #endif 307 308 309 void set_pageblock_migratetype(struct page *page, int migratetype) 310 { 311 if (unlikely(page_group_by_mobility_disabled && 312 migratetype < MIGRATE_PCPTYPES)) 313 migratetype = MIGRATE_UNMOVABLE; 314 315 set_pageblock_flags_group(page, (unsigned long)migratetype, 316 PB_migrate, PB_migrate_end); 317 } 318 319 #ifdef CONFIG_DEBUG_VM 320 static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 321 { 322 int ret = 0; 323 unsigned seq; 324 unsigned long pfn = page_to_pfn(page); 325 unsigned long sp, start_pfn; 326 327 do { 328 seq = zone_span_seqbegin(zone); 329 start_pfn = zone->zone_start_pfn; 330 sp = zone->spanned_pages; 331 if (!zone_spans_pfn(zone, pfn)) 332 ret = 1; 333 } while (zone_span_seqretry(zone, seq)); 334 335 if (ret) 336 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", 337 pfn, zone_to_nid(zone), zone->name, 338 start_pfn, start_pfn + sp); 339 340 return ret; 341 } 342 343 static int page_is_consistent(struct zone *zone, struct page *page) 344 { 345 if (!pfn_valid_within(page_to_pfn(page))) 346 return 0; 347 if (zone != page_zone(page)) 348 return 0; 349 350 return 1; 351 } 352 /* 353 * Temporary debugging check for pages not lying within a given zone. 354 */ 355 static int bad_range(struct zone *zone, struct page *page) 356 { 357 if (page_outside_zone_boundaries(zone, page)) 358 return 1; 359 if (!page_is_consistent(zone, page)) 360 return 1; 361 362 return 0; 363 } 364 #else 365 static inline int bad_range(struct zone *zone, struct page *page) 366 { 367 return 0; 368 } 369 #endif 370 371 static void bad_page(struct page *page, const char *reason, 372 unsigned long bad_flags) 373 { 374 static unsigned long resume; 375 static unsigned long nr_shown; 376 static unsigned long nr_unshown; 377 378 /* Don't complain about poisoned pages */ 379 if (PageHWPoison(page)) { 380 page_mapcount_reset(page); /* remove PageBuddy */ 381 return; 382 } 383 384 /* 385 * Allow a burst of 60 reports, then keep quiet for that minute; 386 * or allow a steady drip of one report per second. 387 */ 388 if (nr_shown == 60) { 389 if (time_before(jiffies, resume)) { 390 nr_unshown++; 391 goto out; 392 } 393 if (nr_unshown) { 394 printk(KERN_ALERT 395 "BUG: Bad page state: %lu messages suppressed\n", 396 nr_unshown); 397 nr_unshown = 0; 398 } 399 nr_shown = 0; 400 } 401 if (nr_shown++ == 0) 402 resume = jiffies + 60 * HZ; 403 404 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n", 405 current->comm, page_to_pfn(page)); 406 dump_page_badflags(page, reason, bad_flags); 407 408 print_modules(); 409 dump_stack(); 410 out: 411 /* Leave bad fields for debug, except PageBuddy could make trouble */ 412 page_mapcount_reset(page); /* remove PageBuddy */ 413 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 414 } 415 416 /* 417 * Higher-order pages are called "compound pages". They are structured thusly: 418 * 419 * The first PAGE_SIZE page is called the "head page". 420 * 421 * The remaining PAGE_SIZE pages are called "tail pages". 422 * 423 * All pages have PG_compound set. All tail pages have their ->first_page 424 * pointing at the head page. 425 * 426 * The first tail page's ->lru.next holds the address of the compound page's 427 * put_page() function. Its ->lru.prev holds the order of allocation. 428 * This usage means that zero-order pages may not be compound. 429 */ 430 431 static void free_compound_page(struct page *page) 432 { 433 __free_pages_ok(page, compound_order(page)); 434 } 435 436 void prep_compound_page(struct page *page, unsigned long order) 437 { 438 int i; 439 int nr_pages = 1 << order; 440 441 set_compound_page_dtor(page, free_compound_page); 442 set_compound_order(page, order); 443 __SetPageHead(page); 444 for (i = 1; i < nr_pages; i++) { 445 struct page *p = page + i; 446 set_page_count(p, 0); 447 p->first_page = page; 448 /* Make sure p->first_page is always valid for PageTail() */ 449 smp_wmb(); 450 __SetPageTail(p); 451 } 452 } 453 454 #ifdef CONFIG_DEBUG_PAGEALLOC 455 unsigned int _debug_guardpage_minorder; 456 bool _debug_pagealloc_enabled __read_mostly; 457 bool _debug_guardpage_enabled __read_mostly; 458 459 static int __init early_debug_pagealloc(char *buf) 460 { 461 if (!buf) 462 return -EINVAL; 463 464 if (strcmp(buf, "on") == 0) 465 _debug_pagealloc_enabled = true; 466 467 return 0; 468 } 469 early_param("debug_pagealloc", early_debug_pagealloc); 470 471 static bool need_debug_guardpage(void) 472 { 473 /* If we don't use debug_pagealloc, we don't need guard page */ 474 if (!debug_pagealloc_enabled()) 475 return false; 476 477 return true; 478 } 479 480 static void init_debug_guardpage(void) 481 { 482 if (!debug_pagealloc_enabled()) 483 return; 484 485 _debug_guardpage_enabled = true; 486 } 487 488 struct page_ext_operations debug_guardpage_ops = { 489 .need = need_debug_guardpage, 490 .init = init_debug_guardpage, 491 }; 492 493 static int __init debug_guardpage_minorder_setup(char *buf) 494 { 495 unsigned long res; 496 497 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 498 printk(KERN_ERR "Bad debug_guardpage_minorder value\n"); 499 return 0; 500 } 501 _debug_guardpage_minorder = res; 502 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res); 503 return 0; 504 } 505 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup); 506 507 static inline void set_page_guard(struct zone *zone, struct page *page, 508 unsigned int order, int migratetype) 509 { 510 struct page_ext *page_ext; 511 512 if (!debug_guardpage_enabled()) 513 return; 514 515 page_ext = lookup_page_ext(page); 516 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 517 518 INIT_LIST_HEAD(&page->lru); 519 set_page_private(page, order); 520 /* Guard pages are not available for any usage */ 521 __mod_zone_freepage_state(zone, -(1 << order), migratetype); 522 } 523 524 static inline void clear_page_guard(struct zone *zone, struct page *page, 525 unsigned int order, int migratetype) 526 { 527 struct page_ext *page_ext; 528 529 if (!debug_guardpage_enabled()) 530 return; 531 532 page_ext = lookup_page_ext(page); 533 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 534 535 set_page_private(page, 0); 536 if (!is_migrate_isolate(migratetype)) 537 __mod_zone_freepage_state(zone, (1 << order), migratetype); 538 } 539 #else 540 struct page_ext_operations debug_guardpage_ops = { NULL, }; 541 static inline void set_page_guard(struct zone *zone, struct page *page, 542 unsigned int order, int migratetype) {} 543 static inline void clear_page_guard(struct zone *zone, struct page *page, 544 unsigned int order, int migratetype) {} 545 #endif 546 547 static inline void set_page_order(struct page *page, unsigned int order) 548 { 549 set_page_private(page, order); 550 __SetPageBuddy(page); 551 } 552 553 static inline void rmv_page_order(struct page *page) 554 { 555 __ClearPageBuddy(page); 556 set_page_private(page, 0); 557 } 558 559 /* 560 * This function checks whether a page is free && is the buddy 561 * we can do coalesce a page and its buddy if 562 * (a) the buddy is not in a hole && 563 * (b) the buddy is in the buddy system && 564 * (c) a page and its buddy have the same order && 565 * (d) a page and its buddy are in the same zone. 566 * 567 * For recording whether a page is in the buddy system, we set ->_mapcount 568 * PAGE_BUDDY_MAPCOUNT_VALUE. 569 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is 570 * serialized by zone->lock. 571 * 572 * For recording page's order, we use page_private(page). 573 */ 574 static inline int page_is_buddy(struct page *page, struct page *buddy, 575 unsigned int order) 576 { 577 if (!pfn_valid_within(page_to_pfn(buddy))) 578 return 0; 579 580 if (page_is_guard(buddy) && page_order(buddy) == order) { 581 if (page_zone_id(page) != page_zone_id(buddy)) 582 return 0; 583 584 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 585 586 return 1; 587 } 588 589 if (PageBuddy(buddy) && page_order(buddy) == order) { 590 /* 591 * zone check is done late to avoid uselessly 592 * calculating zone/node ids for pages that could 593 * never merge. 594 */ 595 if (page_zone_id(page) != page_zone_id(buddy)) 596 return 0; 597 598 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 599 600 return 1; 601 } 602 return 0; 603 } 604 605 /* 606 * Freeing function for a buddy system allocator. 607 * 608 * The concept of a buddy system is to maintain direct-mapped table 609 * (containing bit values) for memory blocks of various "orders". 610 * The bottom level table contains the map for the smallest allocatable 611 * units of memory (here, pages), and each level above it describes 612 * pairs of units from the levels below, hence, "buddies". 613 * At a high level, all that happens here is marking the table entry 614 * at the bottom level available, and propagating the changes upward 615 * as necessary, plus some accounting needed to play nicely with other 616 * parts of the VM system. 617 * At each level, we keep a list of pages, which are heads of continuous 618 * free pages of length of (1 << order) and marked with _mapcount 619 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) 620 * field. 621 * So when we are allocating or freeing one, we can derive the state of the 622 * other. That is, if we allocate a small block, and both were 623 * free, the remainder of the region must be split into blocks. 624 * If a block is freed, and its buddy is also free, then this 625 * triggers coalescing into a block of larger size. 626 * 627 * -- nyc 628 */ 629 630 static inline void __free_one_page(struct page *page, 631 unsigned long pfn, 632 struct zone *zone, unsigned int order, 633 int migratetype) 634 { 635 unsigned long page_idx; 636 unsigned long combined_idx; 637 unsigned long uninitialized_var(buddy_idx); 638 struct page *buddy; 639 int max_order = MAX_ORDER; 640 641 VM_BUG_ON(!zone_is_initialized(zone)); 642 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); 643 644 VM_BUG_ON(migratetype == -1); 645 if (is_migrate_isolate(migratetype)) { 646 /* 647 * We restrict max order of merging to prevent merge 648 * between freepages on isolate pageblock and normal 649 * pageblock. Without this, pageblock isolation 650 * could cause incorrect freepage accounting. 651 */ 652 max_order = min(MAX_ORDER, pageblock_order + 1); 653 } else { 654 __mod_zone_freepage_state(zone, 1 << order, migratetype); 655 } 656 657 page_idx = pfn & ((1 << max_order) - 1); 658 659 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page); 660 VM_BUG_ON_PAGE(bad_range(zone, page), page); 661 662 while (order < max_order - 1) { 663 buddy_idx = __find_buddy_index(page_idx, order); 664 buddy = page + (buddy_idx - page_idx); 665 if (!page_is_buddy(page, buddy, order)) 666 break; 667 /* 668 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 669 * merge with it and move up one order. 670 */ 671 if (page_is_guard(buddy)) { 672 clear_page_guard(zone, buddy, order, migratetype); 673 } else { 674 list_del(&buddy->lru); 675 zone->free_area[order].nr_free--; 676 rmv_page_order(buddy); 677 } 678 combined_idx = buddy_idx & page_idx; 679 page = page + (combined_idx - page_idx); 680 page_idx = combined_idx; 681 order++; 682 } 683 set_page_order(page, order); 684 685 /* 686 * If this is not the largest possible page, check if the buddy 687 * of the next-highest order is free. If it is, it's possible 688 * that pages are being freed that will coalesce soon. In case, 689 * that is happening, add the free page to the tail of the list 690 * so it's less likely to be used soon and more likely to be merged 691 * as a higher order page 692 */ 693 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) { 694 struct page *higher_page, *higher_buddy; 695 combined_idx = buddy_idx & page_idx; 696 higher_page = page + (combined_idx - page_idx); 697 buddy_idx = __find_buddy_index(combined_idx, order + 1); 698 higher_buddy = higher_page + (buddy_idx - combined_idx); 699 if (page_is_buddy(higher_page, higher_buddy, order + 1)) { 700 list_add_tail(&page->lru, 701 &zone->free_area[order].free_list[migratetype]); 702 goto out; 703 } 704 } 705 706 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 707 out: 708 zone->free_area[order].nr_free++; 709 } 710 711 static inline int free_pages_check(struct page *page) 712 { 713 const char *bad_reason = NULL; 714 unsigned long bad_flags = 0; 715 716 if (unlikely(page_mapcount(page))) 717 bad_reason = "nonzero mapcount"; 718 if (unlikely(page->mapping != NULL)) 719 bad_reason = "non-NULL mapping"; 720 if (unlikely(atomic_read(&page->_count) != 0)) 721 bad_reason = "nonzero _count"; 722 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { 723 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; 724 bad_flags = PAGE_FLAGS_CHECK_AT_FREE; 725 } 726 #ifdef CONFIG_MEMCG 727 if (unlikely(page->mem_cgroup)) 728 bad_reason = "page still charged to cgroup"; 729 #endif 730 if (unlikely(bad_reason)) { 731 bad_page(page, bad_reason, bad_flags); 732 return 1; 733 } 734 page_cpupid_reset_last(page); 735 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP) 736 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 737 return 0; 738 } 739 740 /* 741 * Frees a number of pages from the PCP lists 742 * Assumes all pages on list are in same zone, and of same order. 743 * count is the number of pages to free. 744 * 745 * If the zone was previously in an "all pages pinned" state then look to 746 * see if this freeing clears that state. 747 * 748 * And clear the zone's pages_scanned counter, to hold off the "all pages are 749 * pinned" detection logic. 750 */ 751 static void free_pcppages_bulk(struct zone *zone, int count, 752 struct per_cpu_pages *pcp) 753 { 754 int migratetype = 0; 755 int batch_free = 0; 756 int to_free = count; 757 unsigned long nr_scanned; 758 759 spin_lock(&zone->lock); 760 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 761 if (nr_scanned) 762 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 763 764 while (to_free) { 765 struct page *page; 766 struct list_head *list; 767 768 /* 769 * Remove pages from lists in a round-robin fashion. A 770 * batch_free count is maintained that is incremented when an 771 * empty list is encountered. This is so more pages are freed 772 * off fuller lists instead of spinning excessively around empty 773 * lists 774 */ 775 do { 776 batch_free++; 777 if (++migratetype == MIGRATE_PCPTYPES) 778 migratetype = 0; 779 list = &pcp->lists[migratetype]; 780 } while (list_empty(list)); 781 782 /* This is the only non-empty list. Free them all. */ 783 if (batch_free == MIGRATE_PCPTYPES) 784 batch_free = to_free; 785 786 do { 787 int mt; /* migratetype of the to-be-freed page */ 788 789 page = list_entry(list->prev, struct page, lru); 790 /* must delete as __free_one_page list manipulates */ 791 list_del(&page->lru); 792 mt = get_freepage_migratetype(page); 793 if (unlikely(has_isolate_pageblock(zone))) 794 mt = get_pageblock_migratetype(page); 795 796 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */ 797 __free_one_page(page, page_to_pfn(page), zone, 0, mt); 798 trace_mm_page_pcpu_drain(page, 0, mt); 799 } while (--to_free && --batch_free && !list_empty(list)); 800 } 801 spin_unlock(&zone->lock); 802 } 803 804 static void free_one_page(struct zone *zone, 805 struct page *page, unsigned long pfn, 806 unsigned int order, 807 int migratetype) 808 { 809 unsigned long nr_scanned; 810 spin_lock(&zone->lock); 811 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED); 812 if (nr_scanned) 813 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned); 814 815 if (unlikely(has_isolate_pageblock(zone) || 816 is_migrate_isolate(migratetype))) { 817 migratetype = get_pfnblock_migratetype(page, pfn); 818 } 819 __free_one_page(page, pfn, zone, order, migratetype); 820 spin_unlock(&zone->lock); 821 } 822 823 static int free_tail_pages_check(struct page *head_page, struct page *page) 824 { 825 if (!IS_ENABLED(CONFIG_DEBUG_VM)) 826 return 0; 827 if (unlikely(!PageTail(page))) { 828 bad_page(page, "PageTail not set", 0); 829 return 1; 830 } 831 if (unlikely(page->first_page != head_page)) { 832 bad_page(page, "first_page not consistent", 0); 833 return 1; 834 } 835 return 0; 836 } 837 838 static void __meminit __init_single_page(struct page *page, unsigned long pfn, 839 unsigned long zone, int nid) 840 { 841 set_page_links(page, zone, nid, pfn); 842 init_page_count(page); 843 page_mapcount_reset(page); 844 page_cpupid_reset_last(page); 845 846 INIT_LIST_HEAD(&page->lru); 847 #ifdef WANT_PAGE_VIRTUAL 848 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 849 if (!is_highmem_idx(zone)) 850 set_page_address(page, __va(pfn << PAGE_SHIFT)); 851 #endif 852 } 853 854 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone, 855 int nid) 856 { 857 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid); 858 } 859 860 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 861 static void init_reserved_page(unsigned long pfn) 862 { 863 pg_data_t *pgdat; 864 int nid, zid; 865 866 if (!early_page_uninitialised(pfn)) 867 return; 868 869 nid = early_pfn_to_nid(pfn); 870 pgdat = NODE_DATA(nid); 871 872 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 873 struct zone *zone = &pgdat->node_zones[zid]; 874 875 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone)) 876 break; 877 } 878 __init_single_pfn(pfn, zid, nid); 879 } 880 #else 881 static inline void init_reserved_page(unsigned long pfn) 882 { 883 } 884 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 885 886 /* 887 * Initialised pages do not have PageReserved set. This function is 888 * called for each range allocated by the bootmem allocator and 889 * marks the pages PageReserved. The remaining valid pages are later 890 * sent to the buddy page allocator. 891 */ 892 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end) 893 { 894 unsigned long start_pfn = PFN_DOWN(start); 895 unsigned long end_pfn = PFN_UP(end); 896 897 for (; start_pfn < end_pfn; start_pfn++) { 898 if (pfn_valid(start_pfn)) { 899 struct page *page = pfn_to_page(start_pfn); 900 901 init_reserved_page(start_pfn); 902 SetPageReserved(page); 903 } 904 } 905 } 906 907 static bool free_pages_prepare(struct page *page, unsigned int order) 908 { 909 bool compound = PageCompound(page); 910 int i, bad = 0; 911 912 VM_BUG_ON_PAGE(PageTail(page), page); 913 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); 914 915 trace_mm_page_free(page, order); 916 kmemcheck_free_shadow(page, order); 917 kasan_free_pages(page, order); 918 919 if (PageAnon(page)) 920 page->mapping = NULL; 921 bad += free_pages_check(page); 922 for (i = 1; i < (1 << order); i++) { 923 if (compound) 924 bad += free_tail_pages_check(page, page + i); 925 bad += free_pages_check(page + i); 926 } 927 if (bad) 928 return false; 929 930 reset_page_owner(page, order); 931 932 if (!PageHighMem(page)) { 933 debug_check_no_locks_freed(page_address(page), 934 PAGE_SIZE << order); 935 debug_check_no_obj_freed(page_address(page), 936 PAGE_SIZE << order); 937 } 938 arch_free_page(page, order); 939 kernel_map_pages(page, 1 << order, 0); 940 941 return true; 942 } 943 944 static void __free_pages_ok(struct page *page, unsigned int order) 945 { 946 unsigned long flags; 947 int migratetype; 948 unsigned long pfn = page_to_pfn(page); 949 950 if (!free_pages_prepare(page, order)) 951 return; 952 953 migratetype = get_pfnblock_migratetype(page, pfn); 954 local_irq_save(flags); 955 __count_vm_events(PGFREE, 1 << order); 956 set_freepage_migratetype(page, migratetype); 957 free_one_page(page_zone(page), page, pfn, order, migratetype); 958 local_irq_restore(flags); 959 } 960 961 static void __init __free_pages_boot_core(struct page *page, 962 unsigned long pfn, unsigned int order) 963 { 964 unsigned int nr_pages = 1 << order; 965 struct page *p = page; 966 unsigned int loop; 967 968 prefetchw(p); 969 for (loop = 0; loop < (nr_pages - 1); loop++, p++) { 970 prefetchw(p + 1); 971 __ClearPageReserved(p); 972 set_page_count(p, 0); 973 } 974 __ClearPageReserved(p); 975 set_page_count(p, 0); 976 977 page_zone(page)->managed_pages += nr_pages; 978 set_page_refcounted(page); 979 __free_pages(page, order); 980 } 981 982 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \ 983 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) 984 /* Only safe to use early in boot when initialisation is single-threaded */ 985 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata; 986 987 int __meminit early_pfn_to_nid(unsigned long pfn) 988 { 989 int nid; 990 991 /* The system will behave unpredictably otherwise */ 992 BUG_ON(system_state != SYSTEM_BOOTING); 993 994 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache); 995 if (nid >= 0) 996 return nid; 997 /* just returns 0 */ 998 return 0; 999 } 1000 #endif 1001 1002 #ifdef CONFIG_NODES_SPAN_OTHER_NODES 1003 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1004 struct mminit_pfnnid_cache *state) 1005 { 1006 int nid; 1007 1008 nid = __early_pfn_to_nid(pfn, state); 1009 if (nid >= 0 && nid != node) 1010 return false; 1011 return true; 1012 } 1013 1014 /* Only safe to use early in boot when initialisation is single-threaded */ 1015 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1016 { 1017 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache); 1018 } 1019 1020 #else 1021 1022 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1023 { 1024 return true; 1025 } 1026 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1027 struct mminit_pfnnid_cache *state) 1028 { 1029 return true; 1030 } 1031 #endif 1032 1033 1034 void __init __free_pages_bootmem(struct page *page, unsigned long pfn, 1035 unsigned int order) 1036 { 1037 if (early_page_uninitialised(pfn)) 1038 return; 1039 return __free_pages_boot_core(page, pfn, order); 1040 } 1041 1042 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 1043 static void __init deferred_free_range(struct page *page, 1044 unsigned long pfn, int nr_pages) 1045 { 1046 int i; 1047 1048 if (!page) 1049 return; 1050 1051 /* Free a large naturally-aligned chunk if possible */ 1052 if (nr_pages == MAX_ORDER_NR_PAGES && 1053 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) { 1054 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1055 __free_pages_boot_core(page, pfn, MAX_ORDER-1); 1056 return; 1057 } 1058 1059 for (i = 0; i < nr_pages; i++, page++, pfn++) 1060 __free_pages_boot_core(page, pfn, 0); 1061 } 1062 1063 static __initdata DECLARE_RWSEM(pgdat_init_rwsem); 1064 1065 /* Initialise remaining memory on a node */ 1066 static int __init deferred_init_memmap(void *data) 1067 { 1068 pg_data_t *pgdat = data; 1069 int nid = pgdat->node_id; 1070 struct mminit_pfnnid_cache nid_init_state = { }; 1071 unsigned long start = jiffies; 1072 unsigned long nr_pages = 0; 1073 unsigned long walk_start, walk_end; 1074 int i, zid; 1075 struct zone *zone; 1076 unsigned long first_init_pfn = pgdat->first_deferred_pfn; 1077 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 1078 1079 if (first_init_pfn == ULONG_MAX) { 1080 up_read(&pgdat_init_rwsem); 1081 return 0; 1082 } 1083 1084 /* Bind memory initialisation thread to a local node if possible */ 1085 if (!cpumask_empty(cpumask)) 1086 set_cpus_allowed_ptr(current, cpumask); 1087 1088 /* Sanity check boundaries */ 1089 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn); 1090 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat)); 1091 pgdat->first_deferred_pfn = ULONG_MAX; 1092 1093 /* Only the highest zone is deferred so find it */ 1094 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1095 zone = pgdat->node_zones + zid; 1096 if (first_init_pfn < zone_end_pfn(zone)) 1097 break; 1098 } 1099 1100 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) { 1101 unsigned long pfn, end_pfn; 1102 struct page *page = NULL; 1103 struct page *free_base_page = NULL; 1104 unsigned long free_base_pfn = 0; 1105 int nr_to_free = 0; 1106 1107 end_pfn = min(walk_end, zone_end_pfn(zone)); 1108 pfn = first_init_pfn; 1109 if (pfn < walk_start) 1110 pfn = walk_start; 1111 if (pfn < zone->zone_start_pfn) 1112 pfn = zone->zone_start_pfn; 1113 1114 for (; pfn < end_pfn; pfn++) { 1115 if (!pfn_valid_within(pfn)) 1116 goto free_range; 1117 1118 /* 1119 * Ensure pfn_valid is checked every 1120 * MAX_ORDER_NR_PAGES for memory holes 1121 */ 1122 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) { 1123 if (!pfn_valid(pfn)) { 1124 page = NULL; 1125 goto free_range; 1126 } 1127 } 1128 1129 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) { 1130 page = NULL; 1131 goto free_range; 1132 } 1133 1134 /* Minimise pfn page lookups and scheduler checks */ 1135 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) { 1136 page++; 1137 } else { 1138 nr_pages += nr_to_free; 1139 deferred_free_range(free_base_page, 1140 free_base_pfn, nr_to_free); 1141 free_base_page = NULL; 1142 free_base_pfn = nr_to_free = 0; 1143 1144 page = pfn_to_page(pfn); 1145 cond_resched(); 1146 } 1147 1148 if (page->flags) { 1149 VM_BUG_ON(page_zone(page) != zone); 1150 goto free_range; 1151 } 1152 1153 __init_single_page(page, pfn, zid, nid); 1154 if (!free_base_page) { 1155 free_base_page = page; 1156 free_base_pfn = pfn; 1157 nr_to_free = 0; 1158 } 1159 nr_to_free++; 1160 1161 /* Where possible, batch up pages for a single free */ 1162 continue; 1163 free_range: 1164 /* Free the current block of pages to allocator */ 1165 nr_pages += nr_to_free; 1166 deferred_free_range(free_base_page, free_base_pfn, 1167 nr_to_free); 1168 free_base_page = NULL; 1169 free_base_pfn = nr_to_free = 0; 1170 } 1171 1172 first_init_pfn = max(end_pfn, first_init_pfn); 1173 } 1174 1175 /* Sanity check that the next zone really is unpopulated */ 1176 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone)); 1177 1178 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages, 1179 jiffies_to_msecs(jiffies - start)); 1180 up_read(&pgdat_init_rwsem); 1181 return 0; 1182 } 1183 1184 void __init page_alloc_init_late(void) 1185 { 1186 int nid; 1187 1188 for_each_node_state(nid, N_MEMORY) { 1189 down_read(&pgdat_init_rwsem); 1190 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid); 1191 } 1192 1193 /* Block until all are initialised */ 1194 down_write(&pgdat_init_rwsem); 1195 up_write(&pgdat_init_rwsem); 1196 } 1197 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 1198 1199 #ifdef CONFIG_CMA 1200 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ 1201 void __init init_cma_reserved_pageblock(struct page *page) 1202 { 1203 unsigned i = pageblock_nr_pages; 1204 struct page *p = page; 1205 1206 do { 1207 __ClearPageReserved(p); 1208 set_page_count(p, 0); 1209 } while (++p, --i); 1210 1211 set_pageblock_migratetype(page, MIGRATE_CMA); 1212 1213 if (pageblock_order >= MAX_ORDER) { 1214 i = pageblock_nr_pages; 1215 p = page; 1216 do { 1217 set_page_refcounted(p); 1218 __free_pages(p, MAX_ORDER - 1); 1219 p += MAX_ORDER_NR_PAGES; 1220 } while (i -= MAX_ORDER_NR_PAGES); 1221 } else { 1222 set_page_refcounted(page); 1223 __free_pages(page, pageblock_order); 1224 } 1225 1226 adjust_managed_page_count(page, pageblock_nr_pages); 1227 } 1228 #endif 1229 1230 /* 1231 * The order of subdivision here is critical for the IO subsystem. 1232 * Please do not alter this order without good reasons and regression 1233 * testing. Specifically, as large blocks of memory are subdivided, 1234 * the order in which smaller blocks are delivered depends on the order 1235 * they're subdivided in this function. This is the primary factor 1236 * influencing the order in which pages are delivered to the IO 1237 * subsystem according to empirical testing, and this is also justified 1238 * by considering the behavior of a buddy system containing a single 1239 * large block of memory acted on by a series of small allocations. 1240 * This behavior is a critical factor in sglist merging's success. 1241 * 1242 * -- nyc 1243 */ 1244 static inline void expand(struct zone *zone, struct page *page, 1245 int low, int high, struct free_area *area, 1246 int migratetype) 1247 { 1248 unsigned long size = 1 << high; 1249 1250 while (high > low) { 1251 area--; 1252 high--; 1253 size >>= 1; 1254 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); 1255 1256 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) && 1257 debug_guardpage_enabled() && 1258 high < debug_guardpage_minorder()) { 1259 /* 1260 * Mark as guard pages (or page), that will allow to 1261 * merge back to allocator when buddy will be freed. 1262 * Corresponding page table entries will not be touched, 1263 * pages will stay not present in virtual address space 1264 */ 1265 set_page_guard(zone, &page[size], high, migratetype); 1266 continue; 1267 } 1268 list_add(&page[size].lru, &area->free_list[migratetype]); 1269 area->nr_free++; 1270 set_page_order(&page[size], high); 1271 } 1272 } 1273 1274 /* 1275 * This page is about to be returned from the page allocator 1276 */ 1277 static inline int check_new_page(struct page *page) 1278 { 1279 const char *bad_reason = NULL; 1280 unsigned long bad_flags = 0; 1281 1282 if (unlikely(page_mapcount(page))) 1283 bad_reason = "nonzero mapcount"; 1284 if (unlikely(page->mapping != NULL)) 1285 bad_reason = "non-NULL mapping"; 1286 if (unlikely(atomic_read(&page->_count) != 0)) 1287 bad_reason = "nonzero _count"; 1288 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { 1289 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; 1290 bad_flags = PAGE_FLAGS_CHECK_AT_PREP; 1291 } 1292 #ifdef CONFIG_MEMCG 1293 if (unlikely(page->mem_cgroup)) 1294 bad_reason = "page still charged to cgroup"; 1295 #endif 1296 if (unlikely(bad_reason)) { 1297 bad_page(page, bad_reason, bad_flags); 1298 return 1; 1299 } 1300 return 0; 1301 } 1302 1303 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, 1304 int alloc_flags) 1305 { 1306 int i; 1307 1308 for (i = 0; i < (1 << order); i++) { 1309 struct page *p = page + i; 1310 if (unlikely(check_new_page(p))) 1311 return 1; 1312 } 1313 1314 set_page_private(page, 0); 1315 set_page_refcounted(page); 1316 1317 arch_alloc_page(page, order); 1318 kernel_map_pages(page, 1 << order, 1); 1319 kasan_alloc_pages(page, order); 1320 1321 if (gfp_flags & __GFP_ZERO) 1322 for (i = 0; i < (1 << order); i++) 1323 clear_highpage(page + i); 1324 1325 if (order && (gfp_flags & __GFP_COMP)) 1326 prep_compound_page(page, order); 1327 1328 set_page_owner(page, order, gfp_flags); 1329 1330 /* 1331 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to 1332 * allocate the page. The expectation is that the caller is taking 1333 * steps that will free more memory. The caller should avoid the page 1334 * being used for !PFMEMALLOC purposes. 1335 */ 1336 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS); 1337 1338 return 0; 1339 } 1340 1341 /* 1342 * Go through the free lists for the given migratetype and remove 1343 * the smallest available page from the freelists 1344 */ 1345 static inline 1346 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 1347 int migratetype) 1348 { 1349 unsigned int current_order; 1350 struct free_area *area; 1351 struct page *page; 1352 1353 /* Find a page of the appropriate size in the preferred list */ 1354 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 1355 area = &(zone->free_area[current_order]); 1356 if (list_empty(&area->free_list[migratetype])) 1357 continue; 1358 1359 page = list_entry(area->free_list[migratetype].next, 1360 struct page, lru); 1361 list_del(&page->lru); 1362 rmv_page_order(page); 1363 area->nr_free--; 1364 expand(zone, page, order, current_order, area, migratetype); 1365 set_freepage_migratetype(page, migratetype); 1366 return page; 1367 } 1368 1369 return NULL; 1370 } 1371 1372 1373 /* 1374 * This array describes the order lists are fallen back to when 1375 * the free lists for the desirable migrate type are depleted 1376 */ 1377 static int fallbacks[MIGRATE_TYPES][4] = { 1378 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 1379 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE }, 1380 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE }, 1381 #ifdef CONFIG_CMA 1382 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */ 1383 #endif 1384 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */ 1385 #ifdef CONFIG_MEMORY_ISOLATION 1386 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */ 1387 #endif 1388 }; 1389 1390 #ifdef CONFIG_CMA 1391 static struct page *__rmqueue_cma_fallback(struct zone *zone, 1392 unsigned int order) 1393 { 1394 return __rmqueue_smallest(zone, order, MIGRATE_CMA); 1395 } 1396 #else 1397 static inline struct page *__rmqueue_cma_fallback(struct zone *zone, 1398 unsigned int order) { return NULL; } 1399 #endif 1400 1401 /* 1402 * Move the free pages in a range to the free lists of the requested type. 1403 * Note that start_page and end_pages are not aligned on a pageblock 1404 * boundary. If alignment is required, use move_freepages_block() 1405 */ 1406 int move_freepages(struct zone *zone, 1407 struct page *start_page, struct page *end_page, 1408 int migratetype) 1409 { 1410 struct page *page; 1411 unsigned long order; 1412 int pages_moved = 0; 1413 1414 #ifndef CONFIG_HOLES_IN_ZONE 1415 /* 1416 * page_zone is not safe to call in this context when 1417 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 1418 * anyway as we check zone boundaries in move_freepages_block(). 1419 * Remove at a later date when no bug reports exist related to 1420 * grouping pages by mobility 1421 */ 1422 VM_BUG_ON(page_zone(start_page) != page_zone(end_page)); 1423 #endif 1424 1425 for (page = start_page; page <= end_page;) { 1426 /* Make sure we are not inadvertently changing nodes */ 1427 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); 1428 1429 if (!pfn_valid_within(page_to_pfn(page))) { 1430 page++; 1431 continue; 1432 } 1433 1434 if (!PageBuddy(page)) { 1435 page++; 1436 continue; 1437 } 1438 1439 order = page_order(page); 1440 list_move(&page->lru, 1441 &zone->free_area[order].free_list[migratetype]); 1442 set_freepage_migratetype(page, migratetype); 1443 page += 1 << order; 1444 pages_moved += 1 << order; 1445 } 1446 1447 return pages_moved; 1448 } 1449 1450 int move_freepages_block(struct zone *zone, struct page *page, 1451 int migratetype) 1452 { 1453 unsigned long start_pfn, end_pfn; 1454 struct page *start_page, *end_page; 1455 1456 start_pfn = page_to_pfn(page); 1457 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 1458 start_page = pfn_to_page(start_pfn); 1459 end_page = start_page + pageblock_nr_pages - 1; 1460 end_pfn = start_pfn + pageblock_nr_pages - 1; 1461 1462 /* Do not cross zone boundaries */ 1463 if (!zone_spans_pfn(zone, start_pfn)) 1464 start_page = page; 1465 if (!zone_spans_pfn(zone, end_pfn)) 1466 return 0; 1467 1468 return move_freepages(zone, start_page, end_page, migratetype); 1469 } 1470 1471 static void change_pageblock_range(struct page *pageblock_page, 1472 int start_order, int migratetype) 1473 { 1474 int nr_pageblocks = 1 << (start_order - pageblock_order); 1475 1476 while (nr_pageblocks--) { 1477 set_pageblock_migratetype(pageblock_page, migratetype); 1478 pageblock_page += pageblock_nr_pages; 1479 } 1480 } 1481 1482 /* 1483 * When we are falling back to another migratetype during allocation, try to 1484 * steal extra free pages from the same pageblocks to satisfy further 1485 * allocations, instead of polluting multiple pageblocks. 1486 * 1487 * If we are stealing a relatively large buddy page, it is likely there will 1488 * be more free pages in the pageblock, so try to steal them all. For 1489 * reclaimable and unmovable allocations, we steal regardless of page size, 1490 * as fragmentation caused by those allocations polluting movable pageblocks 1491 * is worse than movable allocations stealing from unmovable and reclaimable 1492 * pageblocks. 1493 */ 1494 static bool can_steal_fallback(unsigned int order, int start_mt) 1495 { 1496 /* 1497 * Leaving this order check is intended, although there is 1498 * relaxed order check in next check. The reason is that 1499 * we can actually steal whole pageblock if this condition met, 1500 * but, below check doesn't guarantee it and that is just heuristic 1501 * so could be changed anytime. 1502 */ 1503 if (order >= pageblock_order) 1504 return true; 1505 1506 if (order >= pageblock_order / 2 || 1507 start_mt == MIGRATE_RECLAIMABLE || 1508 start_mt == MIGRATE_UNMOVABLE || 1509 page_group_by_mobility_disabled) 1510 return true; 1511 1512 return false; 1513 } 1514 1515 /* 1516 * This function implements actual steal behaviour. If order is large enough, 1517 * we can steal whole pageblock. If not, we first move freepages in this 1518 * pageblock and check whether half of pages are moved or not. If half of 1519 * pages are moved, we can change migratetype of pageblock and permanently 1520 * use it's pages as requested migratetype in the future. 1521 */ 1522 static void steal_suitable_fallback(struct zone *zone, struct page *page, 1523 int start_type) 1524 { 1525 int current_order = page_order(page); 1526 int pages; 1527 1528 /* Take ownership for orders >= pageblock_order */ 1529 if (current_order >= pageblock_order) { 1530 change_pageblock_range(page, current_order, start_type); 1531 return; 1532 } 1533 1534 pages = move_freepages_block(zone, page, start_type); 1535 1536 /* Claim the whole block if over half of it is free */ 1537 if (pages >= (1 << (pageblock_order-1)) || 1538 page_group_by_mobility_disabled) 1539 set_pageblock_migratetype(page, start_type); 1540 } 1541 1542 /* 1543 * Check whether there is a suitable fallback freepage with requested order. 1544 * If only_stealable is true, this function returns fallback_mt only if 1545 * we can steal other freepages all together. This would help to reduce 1546 * fragmentation due to mixed migratetype pages in one pageblock. 1547 */ 1548 int find_suitable_fallback(struct free_area *area, unsigned int order, 1549 int migratetype, bool only_stealable, bool *can_steal) 1550 { 1551 int i; 1552 int fallback_mt; 1553 1554 if (area->nr_free == 0) 1555 return -1; 1556 1557 *can_steal = false; 1558 for (i = 0;; i++) { 1559 fallback_mt = fallbacks[migratetype][i]; 1560 if (fallback_mt == MIGRATE_RESERVE) 1561 break; 1562 1563 if (list_empty(&area->free_list[fallback_mt])) 1564 continue; 1565 1566 if (can_steal_fallback(order, migratetype)) 1567 *can_steal = true; 1568 1569 if (!only_stealable) 1570 return fallback_mt; 1571 1572 if (*can_steal) 1573 return fallback_mt; 1574 } 1575 1576 return -1; 1577 } 1578 1579 /* Remove an element from the buddy allocator from the fallback list */ 1580 static inline struct page * 1581 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) 1582 { 1583 struct free_area *area; 1584 unsigned int current_order; 1585 struct page *page; 1586 int fallback_mt; 1587 bool can_steal; 1588 1589 /* Find the largest possible block of pages in the other list */ 1590 for (current_order = MAX_ORDER-1; 1591 current_order >= order && current_order <= MAX_ORDER-1; 1592 --current_order) { 1593 area = &(zone->free_area[current_order]); 1594 fallback_mt = find_suitable_fallback(area, current_order, 1595 start_migratetype, false, &can_steal); 1596 if (fallback_mt == -1) 1597 continue; 1598 1599 page = list_entry(area->free_list[fallback_mt].next, 1600 struct page, lru); 1601 if (can_steal) 1602 steal_suitable_fallback(zone, page, start_migratetype); 1603 1604 /* Remove the page from the freelists */ 1605 area->nr_free--; 1606 list_del(&page->lru); 1607 rmv_page_order(page); 1608 1609 expand(zone, page, order, current_order, area, 1610 start_migratetype); 1611 /* 1612 * The freepage_migratetype may differ from pageblock's 1613 * migratetype depending on the decisions in 1614 * try_to_steal_freepages(). This is OK as long as it 1615 * does not differ for MIGRATE_CMA pageblocks. For CMA 1616 * we need to make sure unallocated pages flushed from 1617 * pcp lists are returned to the correct freelist. 1618 */ 1619 set_freepage_migratetype(page, start_migratetype); 1620 1621 trace_mm_page_alloc_extfrag(page, order, current_order, 1622 start_migratetype, fallback_mt); 1623 1624 return page; 1625 } 1626 1627 return NULL; 1628 } 1629 1630 /* 1631 * Do the hard work of removing an element from the buddy allocator. 1632 * Call me with the zone->lock already held. 1633 */ 1634 static struct page *__rmqueue(struct zone *zone, unsigned int order, 1635 int migratetype) 1636 { 1637 struct page *page; 1638 1639 retry_reserve: 1640 page = __rmqueue_smallest(zone, order, migratetype); 1641 1642 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) { 1643 if (migratetype == MIGRATE_MOVABLE) 1644 page = __rmqueue_cma_fallback(zone, order); 1645 1646 if (!page) 1647 page = __rmqueue_fallback(zone, order, migratetype); 1648 1649 /* 1650 * Use MIGRATE_RESERVE rather than fail an allocation. goto 1651 * is used because __rmqueue_smallest is an inline function 1652 * and we want just one call site 1653 */ 1654 if (!page) { 1655 migratetype = MIGRATE_RESERVE; 1656 goto retry_reserve; 1657 } 1658 } 1659 1660 trace_mm_page_alloc_zone_locked(page, order, migratetype); 1661 return page; 1662 } 1663 1664 /* 1665 * Obtain a specified number of elements from the buddy allocator, all under 1666 * a single hold of the lock, for efficiency. Add them to the supplied list. 1667 * Returns the number of new pages which were placed at *list. 1668 */ 1669 static int rmqueue_bulk(struct zone *zone, unsigned int order, 1670 unsigned long count, struct list_head *list, 1671 int migratetype, bool cold) 1672 { 1673 int i; 1674 1675 spin_lock(&zone->lock); 1676 for (i = 0; i < count; ++i) { 1677 struct page *page = __rmqueue(zone, order, migratetype); 1678 if (unlikely(page == NULL)) 1679 break; 1680 1681 /* 1682 * Split buddy pages returned by expand() are received here 1683 * in physical page order. The page is added to the callers and 1684 * list and the list head then moves forward. From the callers 1685 * perspective, the linked list is ordered by page number in 1686 * some conditions. This is useful for IO devices that can 1687 * merge IO requests if the physical pages are ordered 1688 * properly. 1689 */ 1690 if (likely(!cold)) 1691 list_add(&page->lru, list); 1692 else 1693 list_add_tail(&page->lru, list); 1694 list = &page->lru; 1695 if (is_migrate_cma(get_freepage_migratetype(page))) 1696 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1697 -(1 << order)); 1698 } 1699 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 1700 spin_unlock(&zone->lock); 1701 return i; 1702 } 1703 1704 #ifdef CONFIG_NUMA 1705 /* 1706 * Called from the vmstat counter updater to drain pagesets of this 1707 * currently executing processor on remote nodes after they have 1708 * expired. 1709 * 1710 * Note that this function must be called with the thread pinned to 1711 * a single processor. 1712 */ 1713 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 1714 { 1715 unsigned long flags; 1716 int to_drain, batch; 1717 1718 local_irq_save(flags); 1719 batch = READ_ONCE(pcp->batch); 1720 to_drain = min(pcp->count, batch); 1721 if (to_drain > 0) { 1722 free_pcppages_bulk(zone, to_drain, pcp); 1723 pcp->count -= to_drain; 1724 } 1725 local_irq_restore(flags); 1726 } 1727 #endif 1728 1729 /* 1730 * Drain pcplists of the indicated processor and zone. 1731 * 1732 * The processor must either be the current processor and the 1733 * thread pinned to the current processor or a processor that 1734 * is not online. 1735 */ 1736 static void drain_pages_zone(unsigned int cpu, struct zone *zone) 1737 { 1738 unsigned long flags; 1739 struct per_cpu_pageset *pset; 1740 struct per_cpu_pages *pcp; 1741 1742 local_irq_save(flags); 1743 pset = per_cpu_ptr(zone->pageset, cpu); 1744 1745 pcp = &pset->pcp; 1746 if (pcp->count) { 1747 free_pcppages_bulk(zone, pcp->count, pcp); 1748 pcp->count = 0; 1749 } 1750 local_irq_restore(flags); 1751 } 1752 1753 /* 1754 * Drain pcplists of all zones on the indicated processor. 1755 * 1756 * The processor must either be the current processor and the 1757 * thread pinned to the current processor or a processor that 1758 * is not online. 1759 */ 1760 static void drain_pages(unsigned int cpu) 1761 { 1762 struct zone *zone; 1763 1764 for_each_populated_zone(zone) { 1765 drain_pages_zone(cpu, zone); 1766 } 1767 } 1768 1769 /* 1770 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 1771 * 1772 * The CPU has to be pinned. When zone parameter is non-NULL, spill just 1773 * the single zone's pages. 1774 */ 1775 void drain_local_pages(struct zone *zone) 1776 { 1777 int cpu = smp_processor_id(); 1778 1779 if (zone) 1780 drain_pages_zone(cpu, zone); 1781 else 1782 drain_pages(cpu); 1783 } 1784 1785 /* 1786 * Spill all the per-cpu pages from all CPUs back into the buddy allocator. 1787 * 1788 * When zone parameter is non-NULL, spill just the single zone's pages. 1789 * 1790 * Note that this code is protected against sending an IPI to an offline 1791 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs: 1792 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but 1793 * nothing keeps CPUs from showing up after we populated the cpumask and 1794 * before the call to on_each_cpu_mask(). 1795 */ 1796 void drain_all_pages(struct zone *zone) 1797 { 1798 int cpu; 1799 1800 /* 1801 * Allocate in the BSS so we wont require allocation in 1802 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y 1803 */ 1804 static cpumask_t cpus_with_pcps; 1805 1806 /* 1807 * We don't care about racing with CPU hotplug event 1808 * as offline notification will cause the notified 1809 * cpu to drain that CPU pcps and on_each_cpu_mask 1810 * disables preemption as part of its processing 1811 */ 1812 for_each_online_cpu(cpu) { 1813 struct per_cpu_pageset *pcp; 1814 struct zone *z; 1815 bool has_pcps = false; 1816 1817 if (zone) { 1818 pcp = per_cpu_ptr(zone->pageset, cpu); 1819 if (pcp->pcp.count) 1820 has_pcps = true; 1821 } else { 1822 for_each_populated_zone(z) { 1823 pcp = per_cpu_ptr(z->pageset, cpu); 1824 if (pcp->pcp.count) { 1825 has_pcps = true; 1826 break; 1827 } 1828 } 1829 } 1830 1831 if (has_pcps) 1832 cpumask_set_cpu(cpu, &cpus_with_pcps); 1833 else 1834 cpumask_clear_cpu(cpu, &cpus_with_pcps); 1835 } 1836 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages, 1837 zone, 1); 1838 } 1839 1840 #ifdef CONFIG_HIBERNATION 1841 1842 void mark_free_pages(struct zone *zone) 1843 { 1844 unsigned long pfn, max_zone_pfn; 1845 unsigned long flags; 1846 unsigned int order, t; 1847 struct list_head *curr; 1848 1849 if (zone_is_empty(zone)) 1850 return; 1851 1852 spin_lock_irqsave(&zone->lock, flags); 1853 1854 max_zone_pfn = zone_end_pfn(zone); 1855 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 1856 if (pfn_valid(pfn)) { 1857 struct page *page = pfn_to_page(pfn); 1858 1859 if (!swsusp_page_is_forbidden(page)) 1860 swsusp_unset_page_free(page); 1861 } 1862 1863 for_each_migratetype_order(order, t) { 1864 list_for_each(curr, &zone->free_area[order].free_list[t]) { 1865 unsigned long i; 1866 1867 pfn = page_to_pfn(list_entry(curr, struct page, lru)); 1868 for (i = 0; i < (1UL << order); i++) 1869 swsusp_set_page_free(pfn_to_page(pfn + i)); 1870 } 1871 } 1872 spin_unlock_irqrestore(&zone->lock, flags); 1873 } 1874 #endif /* CONFIG_PM */ 1875 1876 /* 1877 * Free a 0-order page 1878 * cold == true ? free a cold page : free a hot page 1879 */ 1880 void free_hot_cold_page(struct page *page, bool cold) 1881 { 1882 struct zone *zone = page_zone(page); 1883 struct per_cpu_pages *pcp; 1884 unsigned long flags; 1885 unsigned long pfn = page_to_pfn(page); 1886 int migratetype; 1887 1888 if (!free_pages_prepare(page, 0)) 1889 return; 1890 1891 migratetype = get_pfnblock_migratetype(page, pfn); 1892 set_freepage_migratetype(page, migratetype); 1893 local_irq_save(flags); 1894 __count_vm_event(PGFREE); 1895 1896 /* 1897 * We only track unmovable, reclaimable and movable on pcp lists. 1898 * Free ISOLATE pages back to the allocator because they are being 1899 * offlined but treat RESERVE as movable pages so we can get those 1900 * areas back if necessary. Otherwise, we may have to free 1901 * excessively into the page allocator 1902 */ 1903 if (migratetype >= MIGRATE_PCPTYPES) { 1904 if (unlikely(is_migrate_isolate(migratetype))) { 1905 free_one_page(zone, page, pfn, 0, migratetype); 1906 goto out; 1907 } 1908 migratetype = MIGRATE_MOVABLE; 1909 } 1910 1911 pcp = &this_cpu_ptr(zone->pageset)->pcp; 1912 if (!cold) 1913 list_add(&page->lru, &pcp->lists[migratetype]); 1914 else 1915 list_add_tail(&page->lru, &pcp->lists[migratetype]); 1916 pcp->count++; 1917 if (pcp->count >= pcp->high) { 1918 unsigned long batch = READ_ONCE(pcp->batch); 1919 free_pcppages_bulk(zone, batch, pcp); 1920 pcp->count -= batch; 1921 } 1922 1923 out: 1924 local_irq_restore(flags); 1925 } 1926 1927 /* 1928 * Free a list of 0-order pages 1929 */ 1930 void free_hot_cold_page_list(struct list_head *list, bool cold) 1931 { 1932 struct page *page, *next; 1933 1934 list_for_each_entry_safe(page, next, list, lru) { 1935 trace_mm_page_free_batched(page, cold); 1936 free_hot_cold_page(page, cold); 1937 } 1938 } 1939 1940 /* 1941 * split_page takes a non-compound higher-order page, and splits it into 1942 * n (1<<order) sub-pages: page[0..n] 1943 * Each sub-page must be freed individually. 1944 * 1945 * Note: this is probably too low level an operation for use in drivers. 1946 * Please consult with lkml before using this in your driver. 1947 */ 1948 void split_page(struct page *page, unsigned int order) 1949 { 1950 int i; 1951 gfp_t gfp_mask; 1952 1953 VM_BUG_ON_PAGE(PageCompound(page), page); 1954 VM_BUG_ON_PAGE(!page_count(page), page); 1955 1956 #ifdef CONFIG_KMEMCHECK 1957 /* 1958 * Split shadow pages too, because free(page[0]) would 1959 * otherwise free the whole shadow. 1960 */ 1961 if (kmemcheck_page_is_tracked(page)) 1962 split_page(virt_to_page(page[0].shadow), order); 1963 #endif 1964 1965 gfp_mask = get_page_owner_gfp(page); 1966 set_page_owner(page, 0, gfp_mask); 1967 for (i = 1; i < (1 << order); i++) { 1968 set_page_refcounted(page + i); 1969 set_page_owner(page + i, 0, gfp_mask); 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_page_owner(page, order, __GFP_MOVABLE); 2000 2001 /* Set the pageblock if the isolated page is at least a pageblock */ 2002 if (order >= pageblock_order - 1) { 2003 struct page *endpage = page + (1 << order) - 1; 2004 for (; page < endpage; page += pageblock_nr_pages) { 2005 int mt = get_pageblock_migratetype(page); 2006 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)) 2007 set_pageblock_migratetype(page, 2008 MIGRATE_MOVABLE); 2009 } 2010 } 2011 2012 2013 return 1UL << order; 2014 } 2015 2016 /* 2017 * Similar to split_page except the page is already free. As this is only 2018 * being used for migration, the migratetype of the block also changes. 2019 * As this is called with interrupts disabled, the caller is responsible 2020 * for calling arch_alloc_page() and kernel_map_page() after interrupts 2021 * are enabled. 2022 * 2023 * Note: this is probably too low level an operation for use in drivers. 2024 * Please consult with lkml before using this in your driver. 2025 */ 2026 int split_free_page(struct page *page) 2027 { 2028 unsigned int order; 2029 int nr_pages; 2030 2031 order = page_order(page); 2032 2033 nr_pages = __isolate_free_page(page, order); 2034 if (!nr_pages) 2035 return 0; 2036 2037 /* Split into individual pages */ 2038 set_page_refcounted(page); 2039 split_page(page, order); 2040 return nr_pages; 2041 } 2042 2043 /* 2044 * Allocate a page from the given zone. Use pcplists for order-0 allocations. 2045 */ 2046 static inline 2047 struct page *buffered_rmqueue(struct zone *preferred_zone, 2048 struct zone *zone, unsigned int order, 2049 gfp_t gfp_flags, int migratetype) 2050 { 2051 unsigned long flags; 2052 struct page *page; 2053 bool cold = ((gfp_flags & __GFP_COLD) != 0); 2054 2055 if (likely(order == 0)) { 2056 struct per_cpu_pages *pcp; 2057 struct list_head *list; 2058 2059 local_irq_save(flags); 2060 pcp = &this_cpu_ptr(zone->pageset)->pcp; 2061 list = &pcp->lists[migratetype]; 2062 if (list_empty(list)) { 2063 pcp->count += rmqueue_bulk(zone, 0, 2064 pcp->batch, list, 2065 migratetype, cold); 2066 if (unlikely(list_empty(list))) 2067 goto failed; 2068 } 2069 2070 if (cold) 2071 page = list_entry(list->prev, struct page, lru); 2072 else 2073 page = list_entry(list->next, struct page, lru); 2074 2075 list_del(&page->lru); 2076 pcp->count--; 2077 } else { 2078 if (unlikely(gfp_flags & __GFP_NOFAIL)) { 2079 /* 2080 * __GFP_NOFAIL is not to be used in new code. 2081 * 2082 * All __GFP_NOFAIL callers should be fixed so that they 2083 * properly detect and handle allocation failures. 2084 * 2085 * We most definitely don't want callers attempting to 2086 * allocate greater than order-1 page units with 2087 * __GFP_NOFAIL. 2088 */ 2089 WARN_ON_ONCE(order > 1); 2090 } 2091 spin_lock_irqsave(&zone->lock, flags); 2092 page = __rmqueue(zone, order, migratetype); 2093 spin_unlock(&zone->lock); 2094 if (!page) 2095 goto failed; 2096 __mod_zone_freepage_state(zone, -(1 << order), 2097 get_freepage_migratetype(page)); 2098 } 2099 2100 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order)); 2101 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 && 2102 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) 2103 set_bit(ZONE_FAIR_DEPLETED, &zone->flags); 2104 2105 __count_zone_vm_events(PGALLOC, zone, 1 << order); 2106 zone_statistics(preferred_zone, zone, gfp_flags); 2107 local_irq_restore(flags); 2108 2109 VM_BUG_ON_PAGE(bad_range(zone, page), page); 2110 return page; 2111 2112 failed: 2113 local_irq_restore(flags); 2114 return NULL; 2115 } 2116 2117 #ifdef CONFIG_FAIL_PAGE_ALLOC 2118 2119 static struct { 2120 struct fault_attr attr; 2121 2122 u32 ignore_gfp_highmem; 2123 u32 ignore_gfp_wait; 2124 u32 min_order; 2125 } fail_page_alloc = { 2126 .attr = FAULT_ATTR_INITIALIZER, 2127 .ignore_gfp_wait = 1, 2128 .ignore_gfp_highmem = 1, 2129 .min_order = 1, 2130 }; 2131 2132 static int __init setup_fail_page_alloc(char *str) 2133 { 2134 return setup_fault_attr(&fail_page_alloc.attr, str); 2135 } 2136 __setup("fail_page_alloc=", setup_fail_page_alloc); 2137 2138 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2139 { 2140 if (order < fail_page_alloc.min_order) 2141 return false; 2142 if (gfp_mask & __GFP_NOFAIL) 2143 return false; 2144 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 2145 return false; 2146 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT)) 2147 return false; 2148 2149 return should_fail(&fail_page_alloc.attr, 1 << order); 2150 } 2151 2152 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 2153 2154 static int __init fail_page_alloc_debugfs(void) 2155 { 2156 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 2157 struct dentry *dir; 2158 2159 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 2160 &fail_page_alloc.attr); 2161 if (IS_ERR(dir)) 2162 return PTR_ERR(dir); 2163 2164 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 2165 &fail_page_alloc.ignore_gfp_wait)) 2166 goto fail; 2167 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 2168 &fail_page_alloc.ignore_gfp_highmem)) 2169 goto fail; 2170 if (!debugfs_create_u32("min-order", mode, dir, 2171 &fail_page_alloc.min_order)) 2172 goto fail; 2173 2174 return 0; 2175 fail: 2176 debugfs_remove_recursive(dir); 2177 2178 return -ENOMEM; 2179 } 2180 2181 late_initcall(fail_page_alloc_debugfs); 2182 2183 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 2184 2185 #else /* CONFIG_FAIL_PAGE_ALLOC */ 2186 2187 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2188 { 2189 return false; 2190 } 2191 2192 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 2193 2194 /* 2195 * Return true if free pages are above 'mark'. This takes into account the order 2196 * of the allocation. 2197 */ 2198 static bool __zone_watermark_ok(struct zone *z, unsigned int order, 2199 unsigned long mark, int classzone_idx, int alloc_flags, 2200 long free_pages) 2201 { 2202 /* free_pages may go negative - that's OK */ 2203 long min = mark; 2204 int o; 2205 long free_cma = 0; 2206 2207 free_pages -= (1 << order) - 1; 2208 if (alloc_flags & ALLOC_HIGH) 2209 min -= min / 2; 2210 if (alloc_flags & ALLOC_HARDER) 2211 min -= min / 4; 2212 #ifdef CONFIG_CMA 2213 /* If allocation can't use CMA areas don't use free CMA pages */ 2214 if (!(alloc_flags & ALLOC_CMA)) 2215 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES); 2216 #endif 2217 2218 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx]) 2219 return false; 2220 for (o = 0; o < order; o++) { 2221 /* At the next order, this order's pages become unavailable */ 2222 free_pages -= z->free_area[o].nr_free << o; 2223 2224 /* Require fewer higher order pages to be free */ 2225 min >>= 1; 2226 2227 if (free_pages <= min) 2228 return false; 2229 } 2230 return true; 2231 } 2232 2233 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, 2234 int classzone_idx, int alloc_flags) 2235 { 2236 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2237 zone_page_state(z, NR_FREE_PAGES)); 2238 } 2239 2240 bool zone_watermark_ok_safe(struct zone *z, unsigned int order, 2241 unsigned long mark, int classzone_idx, int alloc_flags) 2242 { 2243 long free_pages = zone_page_state(z, NR_FREE_PAGES); 2244 2245 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 2246 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 2247 2248 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2249 free_pages); 2250 } 2251 2252 #ifdef CONFIG_NUMA 2253 /* 2254 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to 2255 * skip over zones that are not allowed by the cpuset, or that have 2256 * been recently (in last second) found to be nearly full. See further 2257 * comments in mmzone.h. Reduces cache footprint of zonelist scans 2258 * that have to skip over a lot of full or unallowed zones. 2259 * 2260 * If the zonelist cache is present in the passed zonelist, then 2261 * returns a pointer to the allowed node mask (either the current 2262 * tasks mems_allowed, or node_states[N_MEMORY].) 2263 * 2264 * If the zonelist cache is not available for this zonelist, does 2265 * nothing and returns NULL. 2266 * 2267 * If the fullzones BITMAP in the zonelist cache is stale (more than 2268 * a second since last zap'd) then we zap it out (clear its bits.) 2269 * 2270 * We hold off even calling zlc_setup, until after we've checked the 2271 * first zone in the zonelist, on the theory that most allocations will 2272 * be satisfied from that first zone, so best to examine that zone as 2273 * quickly as we can. 2274 */ 2275 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 2276 { 2277 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2278 nodemask_t *allowednodes; /* zonelist_cache approximation */ 2279 2280 zlc = zonelist->zlcache_ptr; 2281 if (!zlc) 2282 return NULL; 2283 2284 if (time_after(jiffies, zlc->last_full_zap + HZ)) { 2285 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2286 zlc->last_full_zap = jiffies; 2287 } 2288 2289 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ? 2290 &cpuset_current_mems_allowed : 2291 &node_states[N_MEMORY]; 2292 return allowednodes; 2293 } 2294 2295 /* 2296 * Given 'z' scanning a zonelist, run a couple of quick checks to see 2297 * if it is worth looking at further for free memory: 2298 * 1) Check that the zone isn't thought to be full (doesn't have its 2299 * bit set in the zonelist_cache fullzones BITMAP). 2300 * 2) Check that the zones node (obtained from the zonelist_cache 2301 * z_to_n[] mapping) is allowed in the passed in allowednodes mask. 2302 * Return true (non-zero) if zone is worth looking at further, or 2303 * else return false (zero) if it is not. 2304 * 2305 * This check -ignores- the distinction between various watermarks, 2306 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is 2307 * found to be full for any variation of these watermarks, it will 2308 * be considered full for up to one second by all requests, unless 2309 * we are so low on memory on all allowed nodes that we are forced 2310 * into the second scan of the zonelist. 2311 * 2312 * In the second scan we ignore this zonelist cache and exactly 2313 * apply the watermarks to all zones, even it is slower to do so. 2314 * We are low on memory in the second scan, and should leave no stone 2315 * unturned looking for a free page. 2316 */ 2317 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 2318 nodemask_t *allowednodes) 2319 { 2320 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2321 int i; /* index of *z in zonelist zones */ 2322 int n; /* node that zone *z is on */ 2323 2324 zlc = zonelist->zlcache_ptr; 2325 if (!zlc) 2326 return 1; 2327 2328 i = z - zonelist->_zonerefs; 2329 n = zlc->z_to_n[i]; 2330 2331 /* This zone is worth trying if it is allowed but not full */ 2332 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones); 2333 } 2334 2335 /* 2336 * Given 'z' scanning a zonelist, set the corresponding bit in 2337 * zlc->fullzones, so that subsequent attempts to allocate a page 2338 * from that zone don't waste time re-examining it. 2339 */ 2340 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 2341 { 2342 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2343 int i; /* index of *z in zonelist zones */ 2344 2345 zlc = zonelist->zlcache_ptr; 2346 if (!zlc) 2347 return; 2348 2349 i = z - zonelist->_zonerefs; 2350 2351 set_bit(i, zlc->fullzones); 2352 } 2353 2354 /* 2355 * clear all zones full, called after direct reclaim makes progress so that 2356 * a zone that was recently full is not skipped over for up to a second 2357 */ 2358 static void zlc_clear_zones_full(struct zonelist *zonelist) 2359 { 2360 struct zonelist_cache *zlc; /* cached zonelist speedup info */ 2361 2362 zlc = zonelist->zlcache_ptr; 2363 if (!zlc) 2364 return; 2365 2366 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 2367 } 2368 2369 static bool zone_local(struct zone *local_zone, struct zone *zone) 2370 { 2371 return local_zone->node == zone->node; 2372 } 2373 2374 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2375 { 2376 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) < 2377 RECLAIM_DISTANCE; 2378 } 2379 2380 #else /* CONFIG_NUMA */ 2381 2382 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags) 2383 { 2384 return NULL; 2385 } 2386 2387 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z, 2388 nodemask_t *allowednodes) 2389 { 2390 return 1; 2391 } 2392 2393 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z) 2394 { 2395 } 2396 2397 static void zlc_clear_zones_full(struct zonelist *zonelist) 2398 { 2399 } 2400 2401 static bool zone_local(struct zone *local_zone, struct zone *zone) 2402 { 2403 return true; 2404 } 2405 2406 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2407 { 2408 return true; 2409 } 2410 2411 #endif /* CONFIG_NUMA */ 2412 2413 static void reset_alloc_batches(struct zone *preferred_zone) 2414 { 2415 struct zone *zone = preferred_zone->zone_pgdat->node_zones; 2416 2417 do { 2418 mod_zone_page_state(zone, NR_ALLOC_BATCH, 2419 high_wmark_pages(zone) - low_wmark_pages(zone) - 2420 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 2421 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags); 2422 } while (zone++ != preferred_zone); 2423 } 2424 2425 /* 2426 * get_page_from_freelist goes through the zonelist trying to allocate 2427 * a page. 2428 */ 2429 static struct page * 2430 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, 2431 const struct alloc_context *ac) 2432 { 2433 struct zonelist *zonelist = ac->zonelist; 2434 struct zoneref *z; 2435 struct page *page = NULL; 2436 struct zone *zone; 2437 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */ 2438 int zlc_active = 0; /* set if using zonelist_cache */ 2439 int did_zlc_setup = 0; /* just call zlc_setup() one time */ 2440 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) && 2441 (gfp_mask & __GFP_WRITE); 2442 int nr_fair_skipped = 0; 2443 bool zonelist_rescan; 2444 2445 zonelist_scan: 2446 zonelist_rescan = false; 2447 2448 /* 2449 * Scan zonelist, looking for a zone with enough free. 2450 * See also __cpuset_node_allowed() comment in kernel/cpuset.c. 2451 */ 2452 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx, 2453 ac->nodemask) { 2454 unsigned long mark; 2455 2456 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 2457 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 2458 continue; 2459 if (cpusets_enabled() && 2460 (alloc_flags & ALLOC_CPUSET) && 2461 !cpuset_zone_allowed(zone, gfp_mask)) 2462 continue; 2463 /* 2464 * Distribute pages in proportion to the individual 2465 * zone size to ensure fair page aging. The zone a 2466 * page was allocated in should have no effect on the 2467 * time the page has in memory before being reclaimed. 2468 */ 2469 if (alloc_flags & ALLOC_FAIR) { 2470 if (!zone_local(ac->preferred_zone, zone)) 2471 break; 2472 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) { 2473 nr_fair_skipped++; 2474 continue; 2475 } 2476 } 2477 /* 2478 * When allocating a page cache page for writing, we 2479 * want to get it from a zone that is within its dirty 2480 * limit, such that no single zone holds more than its 2481 * proportional share of globally allowed dirty pages. 2482 * The dirty limits take into account the zone's 2483 * lowmem reserves and high watermark so that kswapd 2484 * should be able to balance it without having to 2485 * write pages from its LRU list. 2486 * 2487 * This may look like it could increase pressure on 2488 * lower zones by failing allocations in higher zones 2489 * before they are full. But the pages that do spill 2490 * over are limited as the lower zones are protected 2491 * by this very same mechanism. It should not become 2492 * a practical burden to them. 2493 * 2494 * XXX: For now, allow allocations to potentially 2495 * exceed the per-zone dirty limit in the slowpath 2496 * (ALLOC_WMARK_LOW unset) before going into reclaim, 2497 * which is important when on a NUMA setup the allowed 2498 * zones are together not big enough to reach the 2499 * global limit. The proper fix for these situations 2500 * will require awareness of zones in the 2501 * dirty-throttling and the flusher threads. 2502 */ 2503 if (consider_zone_dirty && !zone_dirty_ok(zone)) 2504 continue; 2505 2506 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 2507 if (!zone_watermark_ok(zone, order, mark, 2508 ac->classzone_idx, alloc_flags)) { 2509 int ret; 2510 2511 /* Checked here to keep the fast path fast */ 2512 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 2513 if (alloc_flags & ALLOC_NO_WATERMARKS) 2514 goto try_this_zone; 2515 2516 if (IS_ENABLED(CONFIG_NUMA) && 2517 !did_zlc_setup && nr_online_nodes > 1) { 2518 /* 2519 * we do zlc_setup if there are multiple nodes 2520 * and before considering the first zone allowed 2521 * by the cpuset. 2522 */ 2523 allowednodes = zlc_setup(zonelist, alloc_flags); 2524 zlc_active = 1; 2525 did_zlc_setup = 1; 2526 } 2527 2528 if (zone_reclaim_mode == 0 || 2529 !zone_allows_reclaim(ac->preferred_zone, zone)) 2530 goto this_zone_full; 2531 2532 /* 2533 * As we may have just activated ZLC, check if the first 2534 * eligible zone has failed zone_reclaim recently. 2535 */ 2536 if (IS_ENABLED(CONFIG_NUMA) && zlc_active && 2537 !zlc_zone_worth_trying(zonelist, z, allowednodes)) 2538 continue; 2539 2540 ret = zone_reclaim(zone, gfp_mask, order); 2541 switch (ret) { 2542 case ZONE_RECLAIM_NOSCAN: 2543 /* did not scan */ 2544 continue; 2545 case ZONE_RECLAIM_FULL: 2546 /* scanned but unreclaimable */ 2547 continue; 2548 default: 2549 /* did we reclaim enough */ 2550 if (zone_watermark_ok(zone, order, mark, 2551 ac->classzone_idx, alloc_flags)) 2552 goto try_this_zone; 2553 2554 /* 2555 * Failed to reclaim enough to meet watermark. 2556 * Only mark the zone full if checking the min 2557 * watermark or if we failed to reclaim just 2558 * 1<<order pages or else the page allocator 2559 * fastpath will prematurely mark zones full 2560 * when the watermark is between the low and 2561 * min watermarks. 2562 */ 2563 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) || 2564 ret == ZONE_RECLAIM_SOME) 2565 goto this_zone_full; 2566 2567 continue; 2568 } 2569 } 2570 2571 try_this_zone: 2572 page = buffered_rmqueue(ac->preferred_zone, zone, order, 2573 gfp_mask, ac->migratetype); 2574 if (page) { 2575 if (prep_new_page(page, order, gfp_mask, alloc_flags)) 2576 goto try_this_zone; 2577 return page; 2578 } 2579 this_zone_full: 2580 if (IS_ENABLED(CONFIG_NUMA) && zlc_active) 2581 zlc_mark_zone_full(zonelist, z); 2582 } 2583 2584 /* 2585 * The first pass makes sure allocations are spread fairly within the 2586 * local node. However, the local node might have free pages left 2587 * after the fairness batches are exhausted, and remote zones haven't 2588 * even been considered yet. Try once more without fairness, and 2589 * include remote zones now, before entering the slowpath and waking 2590 * kswapd: prefer spilling to a remote zone over swapping locally. 2591 */ 2592 if (alloc_flags & ALLOC_FAIR) { 2593 alloc_flags &= ~ALLOC_FAIR; 2594 if (nr_fair_skipped) { 2595 zonelist_rescan = true; 2596 reset_alloc_batches(ac->preferred_zone); 2597 } 2598 if (nr_online_nodes > 1) 2599 zonelist_rescan = true; 2600 } 2601 2602 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) { 2603 /* Disable zlc cache for second zonelist scan */ 2604 zlc_active = 0; 2605 zonelist_rescan = true; 2606 } 2607 2608 if (zonelist_rescan) 2609 goto zonelist_scan; 2610 2611 return NULL; 2612 } 2613 2614 /* 2615 * Large machines with many possible nodes should not always dump per-node 2616 * meminfo in irq context. 2617 */ 2618 static inline bool should_suppress_show_mem(void) 2619 { 2620 bool ret = false; 2621 2622 #if NODES_SHIFT > 8 2623 ret = in_interrupt(); 2624 #endif 2625 return ret; 2626 } 2627 2628 static DEFINE_RATELIMIT_STATE(nopage_rs, 2629 DEFAULT_RATELIMIT_INTERVAL, 2630 DEFAULT_RATELIMIT_BURST); 2631 2632 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...) 2633 { 2634 unsigned int filter = SHOW_MEM_FILTER_NODES; 2635 2636 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 2637 debug_guardpage_minorder() > 0) 2638 return; 2639 2640 /* 2641 * This documents exceptions given to allocations in certain 2642 * contexts that are allowed to allocate outside current's set 2643 * of allowed nodes. 2644 */ 2645 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2646 if (test_thread_flag(TIF_MEMDIE) || 2647 (current->flags & (PF_MEMALLOC | PF_EXITING))) 2648 filter &= ~SHOW_MEM_FILTER_NODES; 2649 if (in_interrupt() || !(gfp_mask & __GFP_WAIT)) 2650 filter &= ~SHOW_MEM_FILTER_NODES; 2651 2652 if (fmt) { 2653 struct va_format vaf; 2654 va_list args; 2655 2656 va_start(args, fmt); 2657 2658 vaf.fmt = fmt; 2659 vaf.va = &args; 2660 2661 pr_warn("%pV", &vaf); 2662 2663 va_end(args); 2664 } 2665 2666 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n", 2667 current->comm, order, gfp_mask); 2668 2669 dump_stack(); 2670 if (!should_suppress_show_mem()) 2671 show_mem(filter); 2672 } 2673 2674 static inline struct page * 2675 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 2676 const struct alloc_context *ac, unsigned long *did_some_progress) 2677 { 2678 struct page *page; 2679 2680 *did_some_progress = 0; 2681 2682 /* 2683 * Acquire the oom lock. If that fails, somebody else is 2684 * making progress for us. 2685 */ 2686 if (!mutex_trylock(&oom_lock)) { 2687 *did_some_progress = 1; 2688 schedule_timeout_uninterruptible(1); 2689 return NULL; 2690 } 2691 2692 /* 2693 * Go through the zonelist yet one more time, keep very high watermark 2694 * here, this is only to catch a parallel oom killing, we must fail if 2695 * we're still under heavy pressure. 2696 */ 2697 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order, 2698 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); 2699 if (page) 2700 goto out; 2701 2702 if (!(gfp_mask & __GFP_NOFAIL)) { 2703 /* Coredumps can quickly deplete all memory reserves */ 2704 if (current->flags & PF_DUMPCORE) 2705 goto out; 2706 /* The OOM killer will not help higher order allocs */ 2707 if (order > PAGE_ALLOC_COSTLY_ORDER) 2708 goto out; 2709 /* The OOM killer does not needlessly kill tasks for lowmem */ 2710 if (ac->high_zoneidx < ZONE_NORMAL) 2711 goto out; 2712 /* The OOM killer does not compensate for IO-less reclaim */ 2713 if (!(gfp_mask & __GFP_FS)) { 2714 /* 2715 * XXX: Page reclaim didn't yield anything, 2716 * and the OOM killer can't be invoked, but 2717 * keep looping as per tradition. 2718 */ 2719 *did_some_progress = 1; 2720 goto out; 2721 } 2722 if (pm_suspended_storage()) 2723 goto out; 2724 /* The OOM killer may not free memory on a specific node */ 2725 if (gfp_mask & __GFP_THISNODE) 2726 goto out; 2727 } 2728 /* Exhausted what can be done so it's blamo time */ 2729 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false) 2730 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) 2731 *did_some_progress = 1; 2732 out: 2733 mutex_unlock(&oom_lock); 2734 return page; 2735 } 2736 2737 #ifdef CONFIG_COMPACTION 2738 /* Try memory compaction for high-order allocations before reclaim */ 2739 static struct page * 2740 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2741 int alloc_flags, const struct alloc_context *ac, 2742 enum migrate_mode mode, int *contended_compaction, 2743 bool *deferred_compaction) 2744 { 2745 unsigned long compact_result; 2746 struct page *page; 2747 2748 if (!order) 2749 return NULL; 2750 2751 current->flags |= PF_MEMALLOC; 2752 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, 2753 mode, contended_compaction); 2754 current->flags &= ~PF_MEMALLOC; 2755 2756 switch (compact_result) { 2757 case COMPACT_DEFERRED: 2758 *deferred_compaction = true; 2759 /* fall-through */ 2760 case COMPACT_SKIPPED: 2761 return NULL; 2762 default: 2763 break; 2764 } 2765 2766 /* 2767 * At least in one zone compaction wasn't deferred or skipped, so let's 2768 * count a compaction stall 2769 */ 2770 count_vm_event(COMPACTSTALL); 2771 2772 page = get_page_from_freelist(gfp_mask, order, 2773 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 2774 2775 if (page) { 2776 struct zone *zone = page_zone(page); 2777 2778 zone->compact_blockskip_flush = false; 2779 compaction_defer_reset(zone, order, true); 2780 count_vm_event(COMPACTSUCCESS); 2781 return page; 2782 } 2783 2784 /* 2785 * It's bad if compaction run occurs and fails. The most likely reason 2786 * is that pages exist, but not enough to satisfy watermarks. 2787 */ 2788 count_vm_event(COMPACTFAIL); 2789 2790 cond_resched(); 2791 2792 return NULL; 2793 } 2794 #else 2795 static inline struct page * 2796 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 2797 int alloc_flags, const struct alloc_context *ac, 2798 enum migrate_mode mode, int *contended_compaction, 2799 bool *deferred_compaction) 2800 { 2801 return NULL; 2802 } 2803 #endif /* CONFIG_COMPACTION */ 2804 2805 /* Perform direct synchronous page reclaim */ 2806 static int 2807 __perform_reclaim(gfp_t gfp_mask, unsigned int order, 2808 const struct alloc_context *ac) 2809 { 2810 struct reclaim_state reclaim_state; 2811 int progress; 2812 2813 cond_resched(); 2814 2815 /* We now go into synchronous reclaim */ 2816 cpuset_memory_pressure_bump(); 2817 current->flags |= PF_MEMALLOC; 2818 lockdep_set_current_reclaim_state(gfp_mask); 2819 reclaim_state.reclaimed_slab = 0; 2820 current->reclaim_state = &reclaim_state; 2821 2822 progress = try_to_free_pages(ac->zonelist, order, gfp_mask, 2823 ac->nodemask); 2824 2825 current->reclaim_state = NULL; 2826 lockdep_clear_current_reclaim_state(); 2827 current->flags &= ~PF_MEMALLOC; 2828 2829 cond_resched(); 2830 2831 return progress; 2832 } 2833 2834 /* The really slow allocator path where we enter direct reclaim */ 2835 static inline struct page * 2836 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 2837 int alloc_flags, const struct alloc_context *ac, 2838 unsigned long *did_some_progress) 2839 { 2840 struct page *page = NULL; 2841 bool drained = false; 2842 2843 *did_some_progress = __perform_reclaim(gfp_mask, order, ac); 2844 if (unlikely(!(*did_some_progress))) 2845 return NULL; 2846 2847 /* After successful reclaim, reconsider all zones for allocation */ 2848 if (IS_ENABLED(CONFIG_NUMA)) 2849 zlc_clear_zones_full(ac->zonelist); 2850 2851 retry: 2852 page = get_page_from_freelist(gfp_mask, order, 2853 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 2854 2855 /* 2856 * If an allocation failed after direct reclaim, it could be because 2857 * pages are pinned on the per-cpu lists. Drain them and try again 2858 */ 2859 if (!page && !drained) { 2860 drain_all_pages(NULL); 2861 drained = true; 2862 goto retry; 2863 } 2864 2865 return page; 2866 } 2867 2868 /* 2869 * This is called in the allocator slow-path if the allocation request is of 2870 * sufficient urgency to ignore watermarks and take other desperate measures 2871 */ 2872 static inline struct page * 2873 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order, 2874 const struct alloc_context *ac) 2875 { 2876 struct page *page; 2877 2878 do { 2879 page = get_page_from_freelist(gfp_mask, order, 2880 ALLOC_NO_WATERMARKS, ac); 2881 2882 if (!page && gfp_mask & __GFP_NOFAIL) 2883 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, 2884 HZ/50); 2885 } while (!page && (gfp_mask & __GFP_NOFAIL)); 2886 2887 return page; 2888 } 2889 2890 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac) 2891 { 2892 struct zoneref *z; 2893 struct zone *zone; 2894 2895 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 2896 ac->high_zoneidx, ac->nodemask) 2897 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone)); 2898 } 2899 2900 static inline int 2901 gfp_to_alloc_flags(gfp_t gfp_mask) 2902 { 2903 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 2904 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD)); 2905 2906 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 2907 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 2908 2909 /* 2910 * The caller may dip into page reserves a bit more if the caller 2911 * cannot run direct reclaim, or if the caller has realtime scheduling 2912 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 2913 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH). 2914 */ 2915 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 2916 2917 if (atomic) { 2918 /* 2919 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even 2920 * if it can't schedule. 2921 */ 2922 if (!(gfp_mask & __GFP_NOMEMALLOC)) 2923 alloc_flags |= ALLOC_HARDER; 2924 /* 2925 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the 2926 * comment for __cpuset_node_allowed(). 2927 */ 2928 alloc_flags &= ~ALLOC_CPUSET; 2929 } else if (unlikely(rt_task(current)) && !in_interrupt()) 2930 alloc_flags |= ALLOC_HARDER; 2931 2932 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) { 2933 if (gfp_mask & __GFP_MEMALLOC) 2934 alloc_flags |= ALLOC_NO_WATERMARKS; 2935 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) 2936 alloc_flags |= ALLOC_NO_WATERMARKS; 2937 else if (!in_interrupt() && 2938 ((current->flags & PF_MEMALLOC) || 2939 unlikely(test_thread_flag(TIF_MEMDIE)))) 2940 alloc_flags |= ALLOC_NO_WATERMARKS; 2941 } 2942 #ifdef CONFIG_CMA 2943 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) 2944 alloc_flags |= ALLOC_CMA; 2945 #endif 2946 return alloc_flags; 2947 } 2948 2949 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) 2950 { 2951 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS); 2952 } 2953 2954 static inline struct page * 2955 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 2956 struct alloc_context *ac) 2957 { 2958 const gfp_t wait = gfp_mask & __GFP_WAIT; 2959 struct page *page = NULL; 2960 int alloc_flags; 2961 unsigned long pages_reclaimed = 0; 2962 unsigned long did_some_progress; 2963 enum migrate_mode migration_mode = MIGRATE_ASYNC; 2964 bool deferred_compaction = false; 2965 int contended_compaction = COMPACT_CONTENDED_NONE; 2966 2967 /* 2968 * In the slowpath, we sanity check order to avoid ever trying to 2969 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 2970 * be using allocators in order of preference for an area that is 2971 * too large. 2972 */ 2973 if (order >= MAX_ORDER) { 2974 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 2975 return NULL; 2976 } 2977 2978 /* 2979 * If this allocation cannot block and it is for a specific node, then 2980 * fail early. There's no need to wakeup kswapd or retry for a 2981 * speculative node-specific allocation. 2982 */ 2983 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait) 2984 goto nopage; 2985 2986 retry: 2987 if (!(gfp_mask & __GFP_NO_KSWAPD)) 2988 wake_all_kswapds(order, ac); 2989 2990 /* 2991 * OK, we're below the kswapd watermark and have kicked background 2992 * reclaim. Now things get more complex, so set up alloc_flags according 2993 * to how we want to proceed. 2994 */ 2995 alloc_flags = gfp_to_alloc_flags(gfp_mask); 2996 2997 /* 2998 * Find the true preferred zone if the allocation is unconstrained by 2999 * cpusets. 3000 */ 3001 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) { 3002 struct zoneref *preferred_zoneref; 3003 preferred_zoneref = first_zones_zonelist(ac->zonelist, 3004 ac->high_zoneidx, NULL, &ac->preferred_zone); 3005 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref); 3006 } 3007 3008 /* This is the last chance, in general, before the goto nopage. */ 3009 page = get_page_from_freelist(gfp_mask, order, 3010 alloc_flags & ~ALLOC_NO_WATERMARKS, ac); 3011 if (page) 3012 goto got_pg; 3013 3014 /* Allocate without watermarks if the context allows */ 3015 if (alloc_flags & ALLOC_NO_WATERMARKS) { 3016 /* 3017 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds 3018 * the allocation is high priority and these type of 3019 * allocations are system rather than user orientated 3020 */ 3021 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask); 3022 3023 page = __alloc_pages_high_priority(gfp_mask, order, ac); 3024 3025 if (page) { 3026 goto got_pg; 3027 } 3028 } 3029 3030 /* Atomic allocations - we can't balance anything */ 3031 if (!wait) { 3032 /* 3033 * All existing users of the deprecated __GFP_NOFAIL are 3034 * blockable, so warn of any new users that actually allow this 3035 * type of allocation to fail. 3036 */ 3037 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL); 3038 goto nopage; 3039 } 3040 3041 /* Avoid recursion of direct reclaim */ 3042 if (current->flags & PF_MEMALLOC) 3043 goto nopage; 3044 3045 /* Avoid allocations with no watermarks from looping endlessly */ 3046 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL)) 3047 goto nopage; 3048 3049 /* 3050 * Try direct compaction. The first pass is asynchronous. Subsequent 3051 * attempts after direct reclaim are synchronous 3052 */ 3053 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, 3054 migration_mode, 3055 &contended_compaction, 3056 &deferred_compaction); 3057 if (page) 3058 goto got_pg; 3059 3060 /* Checks for THP-specific high-order allocations */ 3061 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) { 3062 /* 3063 * If compaction is deferred for high-order allocations, it is 3064 * because sync compaction recently failed. If this is the case 3065 * and the caller requested a THP allocation, we do not want 3066 * to heavily disrupt the system, so we fail the allocation 3067 * instead of entering direct reclaim. 3068 */ 3069 if (deferred_compaction) 3070 goto nopage; 3071 3072 /* 3073 * In all zones where compaction was attempted (and not 3074 * deferred or skipped), lock contention has been detected. 3075 * For THP allocation we do not want to disrupt the others 3076 * so we fallback to base pages instead. 3077 */ 3078 if (contended_compaction == COMPACT_CONTENDED_LOCK) 3079 goto nopage; 3080 3081 /* 3082 * If compaction was aborted due to need_resched(), we do not 3083 * want to further increase allocation latency, unless it is 3084 * khugepaged trying to collapse. 3085 */ 3086 if (contended_compaction == COMPACT_CONTENDED_SCHED 3087 && !(current->flags & PF_KTHREAD)) 3088 goto nopage; 3089 } 3090 3091 /* 3092 * It can become very expensive to allocate transparent hugepages at 3093 * fault, so use asynchronous memory compaction for THP unless it is 3094 * khugepaged trying to collapse. 3095 */ 3096 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE || 3097 (current->flags & PF_KTHREAD)) 3098 migration_mode = MIGRATE_SYNC_LIGHT; 3099 3100 /* Try direct reclaim and then allocating */ 3101 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, 3102 &did_some_progress); 3103 if (page) 3104 goto got_pg; 3105 3106 /* Do not loop if specifically requested */ 3107 if (gfp_mask & __GFP_NORETRY) 3108 goto noretry; 3109 3110 /* Keep reclaiming pages as long as there is reasonable progress */ 3111 pages_reclaimed += did_some_progress; 3112 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) || 3113 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) { 3114 /* Wait for some write requests to complete then retry */ 3115 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50); 3116 goto retry; 3117 } 3118 3119 /* Reclaim has failed us, start killing things */ 3120 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress); 3121 if (page) 3122 goto got_pg; 3123 3124 /* Retry as long as the OOM killer is making progress */ 3125 if (did_some_progress) 3126 goto retry; 3127 3128 noretry: 3129 /* 3130 * High-order allocations do not necessarily loop after 3131 * direct reclaim and reclaim/compaction depends on compaction 3132 * being called after reclaim so call directly if necessary 3133 */ 3134 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, 3135 ac, migration_mode, 3136 &contended_compaction, 3137 &deferred_compaction); 3138 if (page) 3139 goto got_pg; 3140 nopage: 3141 warn_alloc_failed(gfp_mask, order, NULL); 3142 got_pg: 3143 return page; 3144 } 3145 3146 /* 3147 * This is the 'heart' of the zoned buddy allocator. 3148 */ 3149 struct page * 3150 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 3151 struct zonelist *zonelist, nodemask_t *nodemask) 3152 { 3153 struct zoneref *preferred_zoneref; 3154 struct page *page = NULL; 3155 unsigned int cpuset_mems_cookie; 3156 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR; 3157 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */ 3158 struct alloc_context ac = { 3159 .high_zoneidx = gfp_zone(gfp_mask), 3160 .nodemask = nodemask, 3161 .migratetype = gfpflags_to_migratetype(gfp_mask), 3162 }; 3163 3164 gfp_mask &= gfp_allowed_mask; 3165 3166 lockdep_trace_alloc(gfp_mask); 3167 3168 might_sleep_if(gfp_mask & __GFP_WAIT); 3169 3170 if (should_fail_alloc_page(gfp_mask, order)) 3171 return NULL; 3172 3173 /* 3174 * Check the zones suitable for the gfp_mask contain at least one 3175 * valid zone. It's possible to have an empty zonelist as a result 3176 * of __GFP_THISNODE and a memoryless node 3177 */ 3178 if (unlikely(!zonelist->_zonerefs->zone)) 3179 return NULL; 3180 3181 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE) 3182 alloc_flags |= ALLOC_CMA; 3183 3184 retry_cpuset: 3185 cpuset_mems_cookie = read_mems_allowed_begin(); 3186 3187 /* We set it here, as __alloc_pages_slowpath might have changed it */ 3188 ac.zonelist = zonelist; 3189 /* The preferred zone is used for statistics later */ 3190 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx, 3191 ac.nodemask ? : &cpuset_current_mems_allowed, 3192 &ac.preferred_zone); 3193 if (!ac.preferred_zone) 3194 goto out; 3195 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref); 3196 3197 /* First allocation attempt */ 3198 alloc_mask = gfp_mask|__GFP_HARDWALL; 3199 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac); 3200 if (unlikely(!page)) { 3201 /* 3202 * Runtime PM, block IO and its error handling path 3203 * can deadlock because I/O on the device might not 3204 * complete. 3205 */ 3206 alloc_mask = memalloc_noio_flags(gfp_mask); 3207 3208 page = __alloc_pages_slowpath(alloc_mask, order, &ac); 3209 } 3210 3211 if (kmemcheck_enabled && page) 3212 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 3213 3214 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype); 3215 3216 out: 3217 /* 3218 * When updating a task's mems_allowed, it is possible to race with 3219 * parallel threads in such a way that an allocation can fail while 3220 * the mask is being updated. If a page allocation is about to fail, 3221 * check if the cpuset changed during allocation and if so, retry. 3222 */ 3223 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) 3224 goto retry_cpuset; 3225 3226 return page; 3227 } 3228 EXPORT_SYMBOL(__alloc_pages_nodemask); 3229 3230 /* 3231 * Common helper functions. 3232 */ 3233 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 3234 { 3235 struct page *page; 3236 3237 /* 3238 * __get_free_pages() returns a 32-bit address, which cannot represent 3239 * a highmem page 3240 */ 3241 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 3242 3243 page = alloc_pages(gfp_mask, order); 3244 if (!page) 3245 return 0; 3246 return (unsigned long) page_address(page); 3247 } 3248 EXPORT_SYMBOL(__get_free_pages); 3249 3250 unsigned long get_zeroed_page(gfp_t gfp_mask) 3251 { 3252 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 3253 } 3254 EXPORT_SYMBOL(get_zeroed_page); 3255 3256 void __free_pages(struct page *page, unsigned int order) 3257 { 3258 if (put_page_testzero(page)) { 3259 if (order == 0) 3260 free_hot_cold_page(page, false); 3261 else 3262 __free_pages_ok(page, order); 3263 } 3264 } 3265 3266 EXPORT_SYMBOL(__free_pages); 3267 3268 void free_pages(unsigned long addr, unsigned int order) 3269 { 3270 if (addr != 0) { 3271 VM_BUG_ON(!virt_addr_valid((void *)addr)); 3272 __free_pages(virt_to_page((void *)addr), order); 3273 } 3274 } 3275 3276 EXPORT_SYMBOL(free_pages); 3277 3278 /* 3279 * Page Fragment: 3280 * An arbitrary-length arbitrary-offset area of memory which resides 3281 * within a 0 or higher order page. Multiple fragments within that page 3282 * are individually refcounted, in the page's reference counter. 3283 * 3284 * The page_frag functions below provide a simple allocation framework for 3285 * page fragments. This is used by the network stack and network device 3286 * drivers to provide a backing region of memory for use as either an 3287 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info. 3288 */ 3289 static struct page *__page_frag_refill(struct page_frag_cache *nc, 3290 gfp_t gfp_mask) 3291 { 3292 struct page *page = NULL; 3293 gfp_t gfp = gfp_mask; 3294 3295 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3296 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY | 3297 __GFP_NOMEMALLOC; 3298 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, 3299 PAGE_FRAG_CACHE_MAX_ORDER); 3300 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE; 3301 #endif 3302 if (unlikely(!page)) 3303 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0); 3304 3305 nc->va = page ? page_address(page) : NULL; 3306 3307 return page; 3308 } 3309 3310 void *__alloc_page_frag(struct page_frag_cache *nc, 3311 unsigned int fragsz, gfp_t gfp_mask) 3312 { 3313 unsigned int size = PAGE_SIZE; 3314 struct page *page; 3315 int offset; 3316 3317 if (unlikely(!nc->va)) { 3318 refill: 3319 page = __page_frag_refill(nc, gfp_mask); 3320 if (!page) 3321 return NULL; 3322 3323 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3324 /* if size can vary use size else just use PAGE_SIZE */ 3325 size = nc->size; 3326 #endif 3327 /* Even if we own the page, we do not use atomic_set(). 3328 * This would break get_page_unless_zero() users. 3329 */ 3330 atomic_add(size - 1, &page->_count); 3331 3332 /* reset page count bias and offset to start of new frag */ 3333 nc->pfmemalloc = page->pfmemalloc; 3334 nc->pagecnt_bias = size; 3335 nc->offset = size; 3336 } 3337 3338 offset = nc->offset - fragsz; 3339 if (unlikely(offset < 0)) { 3340 page = virt_to_page(nc->va); 3341 3342 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count)) 3343 goto refill; 3344 3345 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 3346 /* if size can vary use size else just use PAGE_SIZE */ 3347 size = nc->size; 3348 #endif 3349 /* OK, page count is 0, we can safely set it */ 3350 atomic_set(&page->_count, size); 3351 3352 /* reset page count bias and offset to start of new frag */ 3353 nc->pagecnt_bias = size; 3354 offset = size - fragsz; 3355 } 3356 3357 nc->pagecnt_bias--; 3358 nc->offset = offset; 3359 3360 return nc->va + offset; 3361 } 3362 EXPORT_SYMBOL(__alloc_page_frag); 3363 3364 /* 3365 * Frees a page fragment allocated out of either a compound or order 0 page. 3366 */ 3367 void __free_page_frag(void *addr) 3368 { 3369 struct page *page = virt_to_head_page(addr); 3370 3371 if (unlikely(put_page_testzero(page))) 3372 __free_pages_ok(page, compound_order(page)); 3373 } 3374 EXPORT_SYMBOL(__free_page_frag); 3375 3376 /* 3377 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter 3378 * of the current memory cgroup. 3379 * 3380 * It should be used when the caller would like to use kmalloc, but since the 3381 * allocation is large, it has to fall back to the page allocator. 3382 */ 3383 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order) 3384 { 3385 struct page *page; 3386 struct mem_cgroup *memcg = NULL; 3387 3388 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 3389 return NULL; 3390 page = alloc_pages(gfp_mask, order); 3391 memcg_kmem_commit_charge(page, memcg, order); 3392 return page; 3393 } 3394 3395 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order) 3396 { 3397 struct page *page; 3398 struct mem_cgroup *memcg = NULL; 3399 3400 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order)) 3401 return NULL; 3402 page = alloc_pages_node(nid, gfp_mask, order); 3403 memcg_kmem_commit_charge(page, memcg, order); 3404 return page; 3405 } 3406 3407 /* 3408 * __free_kmem_pages and free_kmem_pages will free pages allocated with 3409 * alloc_kmem_pages. 3410 */ 3411 void __free_kmem_pages(struct page *page, unsigned int order) 3412 { 3413 memcg_kmem_uncharge_pages(page, order); 3414 __free_pages(page, order); 3415 } 3416 3417 void free_kmem_pages(unsigned long addr, unsigned int order) 3418 { 3419 if (addr != 0) { 3420 VM_BUG_ON(!virt_addr_valid((void *)addr)); 3421 __free_kmem_pages(virt_to_page((void *)addr), order); 3422 } 3423 } 3424 3425 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size) 3426 { 3427 if (addr) { 3428 unsigned long alloc_end = addr + (PAGE_SIZE << order); 3429 unsigned long used = addr + PAGE_ALIGN(size); 3430 3431 split_page(virt_to_page((void *)addr), order); 3432 while (used < alloc_end) { 3433 free_page(used); 3434 used += PAGE_SIZE; 3435 } 3436 } 3437 return (void *)addr; 3438 } 3439 3440 /** 3441 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 3442 * @size: the number of bytes to allocate 3443 * @gfp_mask: GFP flags for the allocation 3444 * 3445 * This function is similar to alloc_pages(), except that it allocates the 3446 * minimum number of pages to satisfy the request. alloc_pages() can only 3447 * allocate memory in power-of-two pages. 3448 * 3449 * This function is also limited by MAX_ORDER. 3450 * 3451 * Memory allocated by this function must be released by free_pages_exact(). 3452 */ 3453 void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 3454 { 3455 unsigned int order = get_order(size); 3456 unsigned long addr; 3457 3458 addr = __get_free_pages(gfp_mask, order); 3459 return make_alloc_exact(addr, order, size); 3460 } 3461 EXPORT_SYMBOL(alloc_pages_exact); 3462 3463 /** 3464 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 3465 * pages on a node. 3466 * @nid: the preferred node ID where memory should be allocated 3467 * @size: the number of bytes to allocate 3468 * @gfp_mask: GFP flags for the allocation 3469 * 3470 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 3471 * back. 3472 * Note this is not alloc_pages_exact_node() which allocates on a specific node, 3473 * but is not exact. 3474 */ 3475 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 3476 { 3477 unsigned order = get_order(size); 3478 struct page *p = alloc_pages_node(nid, gfp_mask, order); 3479 if (!p) 3480 return NULL; 3481 return make_alloc_exact((unsigned long)page_address(p), order, size); 3482 } 3483 3484 /** 3485 * free_pages_exact - release memory allocated via alloc_pages_exact() 3486 * @virt: the value returned by alloc_pages_exact. 3487 * @size: size of allocation, same value as passed to alloc_pages_exact(). 3488 * 3489 * Release the memory allocated by a previous call to alloc_pages_exact. 3490 */ 3491 void free_pages_exact(void *virt, size_t size) 3492 { 3493 unsigned long addr = (unsigned long)virt; 3494 unsigned long end = addr + PAGE_ALIGN(size); 3495 3496 while (addr < end) { 3497 free_page(addr); 3498 addr += PAGE_SIZE; 3499 } 3500 } 3501 EXPORT_SYMBOL(free_pages_exact); 3502 3503 /** 3504 * nr_free_zone_pages - count number of pages beyond high watermark 3505 * @offset: The zone index of the highest zone 3506 * 3507 * nr_free_zone_pages() counts the number of counts pages which are beyond the 3508 * high watermark within all zones at or below a given zone index. For each 3509 * zone, the number of pages is calculated as: 3510 * managed_pages - high_pages 3511 */ 3512 static unsigned long nr_free_zone_pages(int offset) 3513 { 3514 struct zoneref *z; 3515 struct zone *zone; 3516 3517 /* Just pick one node, since fallback list is circular */ 3518 unsigned long sum = 0; 3519 3520 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 3521 3522 for_each_zone_zonelist(zone, z, zonelist, offset) { 3523 unsigned long size = zone->managed_pages; 3524 unsigned long high = high_wmark_pages(zone); 3525 if (size > high) 3526 sum += size - high; 3527 } 3528 3529 return sum; 3530 } 3531 3532 /** 3533 * nr_free_buffer_pages - count number of pages beyond high watermark 3534 * 3535 * nr_free_buffer_pages() counts the number of pages which are beyond the high 3536 * watermark within ZONE_DMA and ZONE_NORMAL. 3537 */ 3538 unsigned long nr_free_buffer_pages(void) 3539 { 3540 return nr_free_zone_pages(gfp_zone(GFP_USER)); 3541 } 3542 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 3543 3544 /** 3545 * nr_free_pagecache_pages - count number of pages beyond high watermark 3546 * 3547 * nr_free_pagecache_pages() counts the number of pages which are beyond the 3548 * high watermark within all zones. 3549 */ 3550 unsigned long nr_free_pagecache_pages(void) 3551 { 3552 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 3553 } 3554 3555 static inline void show_node(struct zone *zone) 3556 { 3557 if (IS_ENABLED(CONFIG_NUMA)) 3558 printk("Node %d ", zone_to_nid(zone)); 3559 } 3560 3561 void si_meminfo(struct sysinfo *val) 3562 { 3563 val->totalram = totalram_pages; 3564 val->sharedram = global_page_state(NR_SHMEM); 3565 val->freeram = global_page_state(NR_FREE_PAGES); 3566 val->bufferram = nr_blockdev_pages(); 3567 val->totalhigh = totalhigh_pages; 3568 val->freehigh = nr_free_highpages(); 3569 val->mem_unit = PAGE_SIZE; 3570 } 3571 3572 EXPORT_SYMBOL(si_meminfo); 3573 3574 #ifdef CONFIG_NUMA 3575 void si_meminfo_node(struct sysinfo *val, int nid) 3576 { 3577 int zone_type; /* needs to be signed */ 3578 unsigned long managed_pages = 0; 3579 pg_data_t *pgdat = NODE_DATA(nid); 3580 3581 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 3582 managed_pages += pgdat->node_zones[zone_type].managed_pages; 3583 val->totalram = managed_pages; 3584 val->sharedram = node_page_state(nid, NR_SHMEM); 3585 val->freeram = node_page_state(nid, NR_FREE_PAGES); 3586 #ifdef CONFIG_HIGHMEM 3587 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages; 3588 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM], 3589 NR_FREE_PAGES); 3590 #else 3591 val->totalhigh = 0; 3592 val->freehigh = 0; 3593 #endif 3594 val->mem_unit = PAGE_SIZE; 3595 } 3596 #endif 3597 3598 /* 3599 * Determine whether the node should be displayed or not, depending on whether 3600 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 3601 */ 3602 bool skip_free_areas_node(unsigned int flags, int nid) 3603 { 3604 bool ret = false; 3605 unsigned int cpuset_mems_cookie; 3606 3607 if (!(flags & SHOW_MEM_FILTER_NODES)) 3608 goto out; 3609 3610 do { 3611 cpuset_mems_cookie = read_mems_allowed_begin(); 3612 ret = !node_isset(nid, cpuset_current_mems_allowed); 3613 } while (read_mems_allowed_retry(cpuset_mems_cookie)); 3614 out: 3615 return ret; 3616 } 3617 3618 #define K(x) ((x) << (PAGE_SHIFT-10)) 3619 3620 static void show_migration_types(unsigned char type) 3621 { 3622 static const char types[MIGRATE_TYPES] = { 3623 [MIGRATE_UNMOVABLE] = 'U', 3624 [MIGRATE_RECLAIMABLE] = 'E', 3625 [MIGRATE_MOVABLE] = 'M', 3626 [MIGRATE_RESERVE] = 'R', 3627 #ifdef CONFIG_CMA 3628 [MIGRATE_CMA] = 'C', 3629 #endif 3630 #ifdef CONFIG_MEMORY_ISOLATION 3631 [MIGRATE_ISOLATE] = 'I', 3632 #endif 3633 }; 3634 char tmp[MIGRATE_TYPES + 1]; 3635 char *p = tmp; 3636 int i; 3637 3638 for (i = 0; i < MIGRATE_TYPES; i++) { 3639 if (type & (1 << i)) 3640 *p++ = types[i]; 3641 } 3642 3643 *p = '\0'; 3644 printk("(%s) ", tmp); 3645 } 3646 3647 /* 3648 * Show free area list (used inside shift_scroll-lock stuff) 3649 * We also calculate the percentage fragmentation. We do this by counting the 3650 * memory on each free list with the exception of the first item on the list. 3651 * 3652 * Bits in @filter: 3653 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's 3654 * cpuset. 3655 */ 3656 void show_free_areas(unsigned int filter) 3657 { 3658 unsigned long free_pcp = 0; 3659 int cpu; 3660 struct zone *zone; 3661 3662 for_each_populated_zone(zone) { 3663 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3664 continue; 3665 3666 for_each_online_cpu(cpu) 3667 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 3668 } 3669 3670 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 3671 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 3672 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n" 3673 " slab_reclaimable:%lu slab_unreclaimable:%lu\n" 3674 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" 3675 " free:%lu free_pcp:%lu free_cma:%lu\n", 3676 global_page_state(NR_ACTIVE_ANON), 3677 global_page_state(NR_INACTIVE_ANON), 3678 global_page_state(NR_ISOLATED_ANON), 3679 global_page_state(NR_ACTIVE_FILE), 3680 global_page_state(NR_INACTIVE_FILE), 3681 global_page_state(NR_ISOLATED_FILE), 3682 global_page_state(NR_UNEVICTABLE), 3683 global_page_state(NR_FILE_DIRTY), 3684 global_page_state(NR_WRITEBACK), 3685 global_page_state(NR_UNSTABLE_NFS), 3686 global_page_state(NR_SLAB_RECLAIMABLE), 3687 global_page_state(NR_SLAB_UNRECLAIMABLE), 3688 global_page_state(NR_FILE_MAPPED), 3689 global_page_state(NR_SHMEM), 3690 global_page_state(NR_PAGETABLE), 3691 global_page_state(NR_BOUNCE), 3692 global_page_state(NR_FREE_PAGES), 3693 free_pcp, 3694 global_page_state(NR_FREE_CMA_PAGES)); 3695 3696 for_each_populated_zone(zone) { 3697 int i; 3698 3699 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3700 continue; 3701 3702 free_pcp = 0; 3703 for_each_online_cpu(cpu) 3704 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 3705 3706 show_node(zone); 3707 printk("%s" 3708 " free:%lukB" 3709 " min:%lukB" 3710 " low:%lukB" 3711 " high:%lukB" 3712 " active_anon:%lukB" 3713 " inactive_anon:%lukB" 3714 " active_file:%lukB" 3715 " inactive_file:%lukB" 3716 " unevictable:%lukB" 3717 " isolated(anon):%lukB" 3718 " isolated(file):%lukB" 3719 " present:%lukB" 3720 " managed:%lukB" 3721 " mlocked:%lukB" 3722 " dirty:%lukB" 3723 " writeback:%lukB" 3724 " mapped:%lukB" 3725 " shmem:%lukB" 3726 " slab_reclaimable:%lukB" 3727 " slab_unreclaimable:%lukB" 3728 " kernel_stack:%lukB" 3729 " pagetables:%lukB" 3730 " unstable:%lukB" 3731 " bounce:%lukB" 3732 " free_pcp:%lukB" 3733 " local_pcp:%ukB" 3734 " free_cma:%lukB" 3735 " writeback_tmp:%lukB" 3736 " pages_scanned:%lu" 3737 " all_unreclaimable? %s" 3738 "\n", 3739 zone->name, 3740 K(zone_page_state(zone, NR_FREE_PAGES)), 3741 K(min_wmark_pages(zone)), 3742 K(low_wmark_pages(zone)), 3743 K(high_wmark_pages(zone)), 3744 K(zone_page_state(zone, NR_ACTIVE_ANON)), 3745 K(zone_page_state(zone, NR_INACTIVE_ANON)), 3746 K(zone_page_state(zone, NR_ACTIVE_FILE)), 3747 K(zone_page_state(zone, NR_INACTIVE_FILE)), 3748 K(zone_page_state(zone, NR_UNEVICTABLE)), 3749 K(zone_page_state(zone, NR_ISOLATED_ANON)), 3750 K(zone_page_state(zone, NR_ISOLATED_FILE)), 3751 K(zone->present_pages), 3752 K(zone->managed_pages), 3753 K(zone_page_state(zone, NR_MLOCK)), 3754 K(zone_page_state(zone, NR_FILE_DIRTY)), 3755 K(zone_page_state(zone, NR_WRITEBACK)), 3756 K(zone_page_state(zone, NR_FILE_MAPPED)), 3757 K(zone_page_state(zone, NR_SHMEM)), 3758 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 3759 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 3760 zone_page_state(zone, NR_KERNEL_STACK) * 3761 THREAD_SIZE / 1024, 3762 K(zone_page_state(zone, NR_PAGETABLE)), 3763 K(zone_page_state(zone, NR_UNSTABLE_NFS)), 3764 K(zone_page_state(zone, NR_BOUNCE)), 3765 K(free_pcp), 3766 K(this_cpu_read(zone->pageset->pcp.count)), 3767 K(zone_page_state(zone, NR_FREE_CMA_PAGES)), 3768 K(zone_page_state(zone, NR_WRITEBACK_TEMP)), 3769 K(zone_page_state(zone, NR_PAGES_SCANNED)), 3770 (!zone_reclaimable(zone) ? "yes" : "no") 3771 ); 3772 printk("lowmem_reserve[]:"); 3773 for (i = 0; i < MAX_NR_ZONES; i++) 3774 printk(" %ld", zone->lowmem_reserve[i]); 3775 printk("\n"); 3776 } 3777 3778 for_each_populated_zone(zone) { 3779 unsigned long nr[MAX_ORDER], flags, order, total = 0; 3780 unsigned char types[MAX_ORDER]; 3781 3782 if (skip_free_areas_node(filter, zone_to_nid(zone))) 3783 continue; 3784 show_node(zone); 3785 printk("%s: ", zone->name); 3786 3787 spin_lock_irqsave(&zone->lock, flags); 3788 for (order = 0; order < MAX_ORDER; order++) { 3789 struct free_area *area = &zone->free_area[order]; 3790 int type; 3791 3792 nr[order] = area->nr_free; 3793 total += nr[order] << order; 3794 3795 types[order] = 0; 3796 for (type = 0; type < MIGRATE_TYPES; type++) { 3797 if (!list_empty(&area->free_list[type])) 3798 types[order] |= 1 << type; 3799 } 3800 } 3801 spin_unlock_irqrestore(&zone->lock, flags); 3802 for (order = 0; order < MAX_ORDER; order++) { 3803 printk("%lu*%lukB ", nr[order], K(1UL) << order); 3804 if (nr[order]) 3805 show_migration_types(types[order]); 3806 } 3807 printk("= %lukB\n", K(total)); 3808 } 3809 3810 hugetlb_show_meminfo(); 3811 3812 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES)); 3813 3814 show_swap_cache_info(); 3815 } 3816 3817 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 3818 { 3819 zoneref->zone = zone; 3820 zoneref->zone_idx = zone_idx(zone); 3821 } 3822 3823 /* 3824 * Builds allocation fallback zone lists. 3825 * 3826 * Add all populated zones of a node to the zonelist. 3827 */ 3828 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 3829 int nr_zones) 3830 { 3831 struct zone *zone; 3832 enum zone_type zone_type = MAX_NR_ZONES; 3833 3834 do { 3835 zone_type--; 3836 zone = pgdat->node_zones + zone_type; 3837 if (populated_zone(zone)) { 3838 zoneref_set_zone(zone, 3839 &zonelist->_zonerefs[nr_zones++]); 3840 check_highest_zone(zone_type); 3841 } 3842 } while (zone_type); 3843 3844 return nr_zones; 3845 } 3846 3847 3848 /* 3849 * zonelist_order: 3850 * 0 = automatic detection of better ordering. 3851 * 1 = order by ([node] distance, -zonetype) 3852 * 2 = order by (-zonetype, [node] distance) 3853 * 3854 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 3855 * the same zonelist. So only NUMA can configure this param. 3856 */ 3857 #define ZONELIST_ORDER_DEFAULT 0 3858 #define ZONELIST_ORDER_NODE 1 3859 #define ZONELIST_ORDER_ZONE 2 3860 3861 /* zonelist order in the kernel. 3862 * set_zonelist_order() will set this to NODE or ZONE. 3863 */ 3864 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 3865 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 3866 3867 3868 #ifdef CONFIG_NUMA 3869 /* The value user specified ....changed by config */ 3870 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3871 /* string for sysctl */ 3872 #define NUMA_ZONELIST_ORDER_LEN 16 3873 char numa_zonelist_order[16] = "default"; 3874 3875 /* 3876 * interface for configure zonelist ordering. 3877 * command line option "numa_zonelist_order" 3878 * = "[dD]efault - default, automatic configuration. 3879 * = "[nN]ode - order by node locality, then by zone within node 3880 * = "[zZ]one - order by zone, then by locality within zone 3881 */ 3882 3883 static int __parse_numa_zonelist_order(char *s) 3884 { 3885 if (*s == 'd' || *s == 'D') { 3886 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 3887 } else if (*s == 'n' || *s == 'N') { 3888 user_zonelist_order = ZONELIST_ORDER_NODE; 3889 } else if (*s == 'z' || *s == 'Z') { 3890 user_zonelist_order = ZONELIST_ORDER_ZONE; 3891 } else { 3892 printk(KERN_WARNING 3893 "Ignoring invalid numa_zonelist_order value: " 3894 "%s\n", s); 3895 return -EINVAL; 3896 } 3897 return 0; 3898 } 3899 3900 static __init int setup_numa_zonelist_order(char *s) 3901 { 3902 int ret; 3903 3904 if (!s) 3905 return 0; 3906 3907 ret = __parse_numa_zonelist_order(s); 3908 if (ret == 0) 3909 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 3910 3911 return ret; 3912 } 3913 early_param("numa_zonelist_order", setup_numa_zonelist_order); 3914 3915 /* 3916 * sysctl handler for numa_zonelist_order 3917 */ 3918 int numa_zonelist_order_handler(struct ctl_table *table, int write, 3919 void __user *buffer, size_t *length, 3920 loff_t *ppos) 3921 { 3922 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 3923 int ret; 3924 static DEFINE_MUTEX(zl_order_mutex); 3925 3926 mutex_lock(&zl_order_mutex); 3927 if (write) { 3928 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { 3929 ret = -EINVAL; 3930 goto out; 3931 } 3932 strcpy(saved_string, (char *)table->data); 3933 } 3934 ret = proc_dostring(table, write, buffer, length, ppos); 3935 if (ret) 3936 goto out; 3937 if (write) { 3938 int oldval = user_zonelist_order; 3939 3940 ret = __parse_numa_zonelist_order((char *)table->data); 3941 if (ret) { 3942 /* 3943 * bogus value. restore saved string 3944 */ 3945 strncpy((char *)table->data, saved_string, 3946 NUMA_ZONELIST_ORDER_LEN); 3947 user_zonelist_order = oldval; 3948 } else if (oldval != user_zonelist_order) { 3949 mutex_lock(&zonelists_mutex); 3950 build_all_zonelists(NULL, NULL); 3951 mutex_unlock(&zonelists_mutex); 3952 } 3953 } 3954 out: 3955 mutex_unlock(&zl_order_mutex); 3956 return ret; 3957 } 3958 3959 3960 #define MAX_NODE_LOAD (nr_online_nodes) 3961 static int node_load[MAX_NUMNODES]; 3962 3963 /** 3964 * find_next_best_node - find the next node that should appear in a given node's fallback list 3965 * @node: node whose fallback list we're appending 3966 * @used_node_mask: nodemask_t of already used nodes 3967 * 3968 * We use a number of factors to determine which is the next node that should 3969 * appear on a given node's fallback list. The node should not have appeared 3970 * already in @node's fallback list, and it should be the next closest node 3971 * according to the distance array (which contains arbitrary distance values 3972 * from each node to each node in the system), and should also prefer nodes 3973 * with no CPUs, since presumably they'll have very little allocation pressure 3974 * on them otherwise. 3975 * It returns -1 if no node is found. 3976 */ 3977 static int find_next_best_node(int node, nodemask_t *used_node_mask) 3978 { 3979 int n, val; 3980 int min_val = INT_MAX; 3981 int best_node = NUMA_NO_NODE; 3982 const struct cpumask *tmp = cpumask_of_node(0); 3983 3984 /* Use the local node if we haven't already */ 3985 if (!node_isset(node, *used_node_mask)) { 3986 node_set(node, *used_node_mask); 3987 return node; 3988 } 3989 3990 for_each_node_state(n, N_MEMORY) { 3991 3992 /* Don't want a node to appear more than once */ 3993 if (node_isset(n, *used_node_mask)) 3994 continue; 3995 3996 /* Use the distance array to find the distance */ 3997 val = node_distance(node, n); 3998 3999 /* Penalize nodes under us ("prefer the next node") */ 4000 val += (n < node); 4001 4002 /* Give preference to headless and unused nodes */ 4003 tmp = cpumask_of_node(n); 4004 if (!cpumask_empty(tmp)) 4005 val += PENALTY_FOR_NODE_WITH_CPUS; 4006 4007 /* Slight preference for less loaded node */ 4008 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 4009 val += node_load[n]; 4010 4011 if (val < min_val) { 4012 min_val = val; 4013 best_node = n; 4014 } 4015 } 4016 4017 if (best_node >= 0) 4018 node_set(best_node, *used_node_mask); 4019 4020 return best_node; 4021 } 4022 4023 4024 /* 4025 * Build zonelists ordered by node and zones within node. 4026 * This results in maximum locality--normal zone overflows into local 4027 * DMA zone, if any--but risks exhausting DMA zone. 4028 */ 4029 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 4030 { 4031 int j; 4032 struct zonelist *zonelist; 4033 4034 zonelist = &pgdat->node_zonelists[0]; 4035 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 4036 ; 4037 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4038 zonelist->_zonerefs[j].zone = NULL; 4039 zonelist->_zonerefs[j].zone_idx = 0; 4040 } 4041 4042 /* 4043 * Build gfp_thisnode zonelists 4044 */ 4045 static void build_thisnode_zonelists(pg_data_t *pgdat) 4046 { 4047 int j; 4048 struct zonelist *zonelist; 4049 4050 zonelist = &pgdat->node_zonelists[1]; 4051 j = build_zonelists_node(pgdat, zonelist, 0); 4052 zonelist->_zonerefs[j].zone = NULL; 4053 zonelist->_zonerefs[j].zone_idx = 0; 4054 } 4055 4056 /* 4057 * Build zonelists ordered by zone and nodes within zones. 4058 * This results in conserving DMA zone[s] until all Normal memory is 4059 * exhausted, but results in overflowing to remote node while memory 4060 * may still exist in local DMA zone. 4061 */ 4062 static int node_order[MAX_NUMNODES]; 4063 4064 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 4065 { 4066 int pos, j, node; 4067 int zone_type; /* needs to be signed */ 4068 struct zone *z; 4069 struct zonelist *zonelist; 4070 4071 zonelist = &pgdat->node_zonelists[0]; 4072 pos = 0; 4073 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 4074 for (j = 0; j < nr_nodes; j++) { 4075 node = node_order[j]; 4076 z = &NODE_DATA(node)->node_zones[zone_type]; 4077 if (populated_zone(z)) { 4078 zoneref_set_zone(z, 4079 &zonelist->_zonerefs[pos++]); 4080 check_highest_zone(zone_type); 4081 } 4082 } 4083 } 4084 zonelist->_zonerefs[pos].zone = NULL; 4085 zonelist->_zonerefs[pos].zone_idx = 0; 4086 } 4087 4088 #if defined(CONFIG_64BIT) 4089 /* 4090 * Devices that require DMA32/DMA are relatively rare and do not justify a 4091 * penalty to every machine in case the specialised case applies. Default 4092 * to Node-ordering on 64-bit NUMA machines 4093 */ 4094 static int default_zonelist_order(void) 4095 { 4096 return ZONELIST_ORDER_NODE; 4097 } 4098 #else 4099 /* 4100 * On 32-bit, the Normal zone needs to be preserved for allocations accessible 4101 * by the kernel. If processes running on node 0 deplete the low memory zone 4102 * then reclaim will occur more frequency increasing stalls and potentially 4103 * be easier to OOM if a large percentage of the zone is under writeback or 4104 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set. 4105 * Hence, default to zone ordering on 32-bit. 4106 */ 4107 static int default_zonelist_order(void) 4108 { 4109 return ZONELIST_ORDER_ZONE; 4110 } 4111 #endif /* CONFIG_64BIT */ 4112 4113 static void set_zonelist_order(void) 4114 { 4115 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 4116 current_zonelist_order = default_zonelist_order(); 4117 else 4118 current_zonelist_order = user_zonelist_order; 4119 } 4120 4121 static void build_zonelists(pg_data_t *pgdat) 4122 { 4123 int j, node, load; 4124 enum zone_type i; 4125 nodemask_t used_mask; 4126 int local_node, prev_node; 4127 struct zonelist *zonelist; 4128 int order = current_zonelist_order; 4129 4130 /* initialize zonelists */ 4131 for (i = 0; i < MAX_ZONELISTS; i++) { 4132 zonelist = pgdat->node_zonelists + i; 4133 zonelist->_zonerefs[0].zone = NULL; 4134 zonelist->_zonerefs[0].zone_idx = 0; 4135 } 4136 4137 /* NUMA-aware ordering of nodes */ 4138 local_node = pgdat->node_id; 4139 load = nr_online_nodes; 4140 prev_node = local_node; 4141 nodes_clear(used_mask); 4142 4143 memset(node_order, 0, sizeof(node_order)); 4144 j = 0; 4145 4146 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 4147 /* 4148 * We don't want to pressure a particular node. 4149 * So adding penalty to the first node in same 4150 * distance group to make it round-robin. 4151 */ 4152 if (node_distance(local_node, node) != 4153 node_distance(local_node, prev_node)) 4154 node_load[node] = load; 4155 4156 prev_node = node; 4157 load--; 4158 if (order == ZONELIST_ORDER_NODE) 4159 build_zonelists_in_node_order(pgdat, node); 4160 else 4161 node_order[j++] = node; /* remember order */ 4162 } 4163 4164 if (order == ZONELIST_ORDER_ZONE) { 4165 /* calculate node order -- i.e., DMA last! */ 4166 build_zonelists_in_zone_order(pgdat, j); 4167 } 4168 4169 build_thisnode_zonelists(pgdat); 4170 } 4171 4172 /* Construct the zonelist performance cache - see further mmzone.h */ 4173 static void build_zonelist_cache(pg_data_t *pgdat) 4174 { 4175 struct zonelist *zonelist; 4176 struct zonelist_cache *zlc; 4177 struct zoneref *z; 4178 4179 zonelist = &pgdat->node_zonelists[0]; 4180 zonelist->zlcache_ptr = zlc = &zonelist->zlcache; 4181 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST); 4182 for (z = zonelist->_zonerefs; z->zone; z++) 4183 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z); 4184 } 4185 4186 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 4187 /* 4188 * Return node id of node used for "local" allocations. 4189 * I.e., first node id of first zone in arg node's generic zonelist. 4190 * Used for initializing percpu 'numa_mem', which is used primarily 4191 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 4192 */ 4193 int local_memory_node(int node) 4194 { 4195 struct zone *zone; 4196 4197 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 4198 gfp_zone(GFP_KERNEL), 4199 NULL, 4200 &zone); 4201 return zone->node; 4202 } 4203 #endif 4204 4205 #else /* CONFIG_NUMA */ 4206 4207 static void set_zonelist_order(void) 4208 { 4209 current_zonelist_order = ZONELIST_ORDER_ZONE; 4210 } 4211 4212 static void build_zonelists(pg_data_t *pgdat) 4213 { 4214 int node, local_node; 4215 enum zone_type j; 4216 struct zonelist *zonelist; 4217 4218 local_node = pgdat->node_id; 4219 4220 zonelist = &pgdat->node_zonelists[0]; 4221 j = build_zonelists_node(pgdat, zonelist, 0); 4222 4223 /* 4224 * Now we build the zonelist so that it contains the zones 4225 * of all the other nodes. 4226 * We don't want to pressure a particular node, so when 4227 * building the zones for node N, we make sure that the 4228 * zones coming right after the local ones are those from 4229 * node N+1 (modulo N) 4230 */ 4231 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 4232 if (!node_online(node)) 4233 continue; 4234 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4235 } 4236 for (node = 0; node < local_node; node++) { 4237 if (!node_online(node)) 4238 continue; 4239 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4240 } 4241 4242 zonelist->_zonerefs[j].zone = NULL; 4243 zonelist->_zonerefs[j].zone_idx = 0; 4244 } 4245 4246 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */ 4247 static void build_zonelist_cache(pg_data_t *pgdat) 4248 { 4249 pgdat->node_zonelists[0].zlcache_ptr = NULL; 4250 } 4251 4252 #endif /* CONFIG_NUMA */ 4253 4254 /* 4255 * Boot pageset table. One per cpu which is going to be used for all 4256 * zones and all nodes. The parameters will be set in such a way 4257 * that an item put on a list will immediately be handed over to 4258 * the buddy list. This is safe since pageset manipulation is done 4259 * with interrupts disabled. 4260 * 4261 * The boot_pagesets must be kept even after bootup is complete for 4262 * unused processors and/or zones. They do play a role for bootstrapping 4263 * hotplugged processors. 4264 * 4265 * zoneinfo_show() and maybe other functions do 4266 * not check if the processor is online before following the pageset pointer. 4267 * Other parts of the kernel may not check if the zone is available. 4268 */ 4269 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 4270 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 4271 static void setup_zone_pageset(struct zone *zone); 4272 4273 /* 4274 * Global mutex to protect against size modification of zonelists 4275 * as well as to serialize pageset setup for the new populated zone. 4276 */ 4277 DEFINE_MUTEX(zonelists_mutex); 4278 4279 /* return values int ....just for stop_machine() */ 4280 static int __build_all_zonelists(void *data) 4281 { 4282 int nid; 4283 int cpu; 4284 pg_data_t *self = data; 4285 4286 #ifdef CONFIG_NUMA 4287 memset(node_load, 0, sizeof(node_load)); 4288 #endif 4289 4290 if (self && !node_online(self->node_id)) { 4291 build_zonelists(self); 4292 build_zonelist_cache(self); 4293 } 4294 4295 for_each_online_node(nid) { 4296 pg_data_t *pgdat = NODE_DATA(nid); 4297 4298 build_zonelists(pgdat); 4299 build_zonelist_cache(pgdat); 4300 } 4301 4302 /* 4303 * Initialize the boot_pagesets that are going to be used 4304 * for bootstrapping processors. The real pagesets for 4305 * each zone will be allocated later when the per cpu 4306 * allocator is available. 4307 * 4308 * boot_pagesets are used also for bootstrapping offline 4309 * cpus if the system is already booted because the pagesets 4310 * are needed to initialize allocators on a specific cpu too. 4311 * F.e. the percpu allocator needs the page allocator which 4312 * needs the percpu allocator in order to allocate its pagesets 4313 * (a chicken-egg dilemma). 4314 */ 4315 for_each_possible_cpu(cpu) { 4316 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 4317 4318 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 4319 /* 4320 * We now know the "local memory node" for each node-- 4321 * i.e., the node of the first zone in the generic zonelist. 4322 * Set up numa_mem percpu variable for on-line cpus. During 4323 * boot, only the boot cpu should be on-line; we'll init the 4324 * secondary cpus' numa_mem as they come on-line. During 4325 * node/memory hotplug, we'll fixup all on-line cpus. 4326 */ 4327 if (cpu_online(cpu)) 4328 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 4329 #endif 4330 } 4331 4332 return 0; 4333 } 4334 4335 static noinline void __init 4336 build_all_zonelists_init(void) 4337 { 4338 __build_all_zonelists(NULL); 4339 mminit_verify_zonelist(); 4340 cpuset_init_current_mems_allowed(); 4341 } 4342 4343 /* 4344 * Called with zonelists_mutex held always 4345 * unless system_state == SYSTEM_BOOTING. 4346 * 4347 * __ref due to (1) call of __meminit annotated setup_zone_pageset 4348 * [we're only called with non-NULL zone through __meminit paths] and 4349 * (2) call of __init annotated helper build_all_zonelists_init 4350 * [protected by SYSTEM_BOOTING]. 4351 */ 4352 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) 4353 { 4354 set_zonelist_order(); 4355 4356 if (system_state == SYSTEM_BOOTING) { 4357 build_all_zonelists_init(); 4358 } else { 4359 #ifdef CONFIG_MEMORY_HOTPLUG 4360 if (zone) 4361 setup_zone_pageset(zone); 4362 #endif 4363 /* we have to stop all cpus to guarantee there is no user 4364 of zonelist */ 4365 stop_machine(__build_all_zonelists, pgdat, NULL); 4366 /* cpuset refresh routine should be here */ 4367 } 4368 vm_total_pages = nr_free_pagecache_pages(); 4369 /* 4370 * Disable grouping by mobility if the number of pages in the 4371 * system is too low to allow the mechanism to work. It would be 4372 * more accurate, but expensive to check per-zone. This check is 4373 * made on memory-hotadd so a system can start with mobility 4374 * disabled and enable it later 4375 */ 4376 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 4377 page_group_by_mobility_disabled = 1; 4378 else 4379 page_group_by_mobility_disabled = 0; 4380 4381 pr_info("Built %i zonelists in %s order, mobility grouping %s. " 4382 "Total pages: %ld\n", 4383 nr_online_nodes, 4384 zonelist_order_name[current_zonelist_order], 4385 page_group_by_mobility_disabled ? "off" : "on", 4386 vm_total_pages); 4387 #ifdef CONFIG_NUMA 4388 pr_info("Policy zone: %s\n", zone_names[policy_zone]); 4389 #endif 4390 } 4391 4392 /* 4393 * Helper functions to size the waitqueue hash table. 4394 * Essentially these want to choose hash table sizes sufficiently 4395 * large so that collisions trying to wait on pages are rare. 4396 * But in fact, the number of active page waitqueues on typical 4397 * systems is ridiculously low, less than 200. So this is even 4398 * conservative, even though it seems large. 4399 * 4400 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to 4401 * waitqueues, i.e. the size of the waitq table given the number of pages. 4402 */ 4403 #define PAGES_PER_WAITQUEUE 256 4404 4405 #ifndef CONFIG_MEMORY_HOTPLUG 4406 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 4407 { 4408 unsigned long size = 1; 4409 4410 pages /= PAGES_PER_WAITQUEUE; 4411 4412 while (size < pages) 4413 size <<= 1; 4414 4415 /* 4416 * Once we have dozens or even hundreds of threads sleeping 4417 * on IO we've got bigger problems than wait queue collision. 4418 * Limit the size of the wait table to a reasonable size. 4419 */ 4420 size = min(size, 4096UL); 4421 4422 return max(size, 4UL); 4423 } 4424 #else 4425 /* 4426 * A zone's size might be changed by hot-add, so it is not possible to determine 4427 * a suitable size for its wait_table. So we use the maximum size now. 4428 * 4429 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie: 4430 * 4431 * i386 (preemption config) : 4096 x 16 = 64Kbyte. 4432 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte. 4433 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte. 4434 * 4435 * The maximum entries are prepared when a zone's memory is (512K + 256) pages 4436 * or more by the traditional way. (See above). It equals: 4437 * 4438 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte. 4439 * ia64(16K page size) : = ( 8G + 4M)byte. 4440 * powerpc (64K page size) : = (32G +16M)byte. 4441 */ 4442 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages) 4443 { 4444 return 4096UL; 4445 } 4446 #endif 4447 4448 /* 4449 * This is an integer logarithm so that shifts can be used later 4450 * to extract the more random high bits from the multiplicative 4451 * hash function before the remainder is taken. 4452 */ 4453 static inline unsigned long wait_table_bits(unsigned long size) 4454 { 4455 return ffz(~size); 4456 } 4457 4458 /* 4459 * Check if a pageblock contains reserved pages 4460 */ 4461 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn) 4462 { 4463 unsigned long pfn; 4464 4465 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 4466 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn))) 4467 return 1; 4468 } 4469 return 0; 4470 } 4471 4472 /* 4473 * Mark a number of pageblocks as MIGRATE_RESERVE. The number 4474 * of blocks reserved is based on min_wmark_pages(zone). The memory within 4475 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes 4476 * higher will lead to a bigger reserve which will get freed as contiguous 4477 * blocks as reclaim kicks in 4478 */ 4479 static void setup_zone_migrate_reserve(struct zone *zone) 4480 { 4481 unsigned long start_pfn, pfn, end_pfn, block_end_pfn; 4482 struct page *page; 4483 unsigned long block_migratetype; 4484 int reserve; 4485 int old_reserve; 4486 4487 /* 4488 * Get the start pfn, end pfn and the number of blocks to reserve 4489 * We have to be careful to be aligned to pageblock_nr_pages to 4490 * make sure that we always check pfn_valid for the first page in 4491 * the block. 4492 */ 4493 start_pfn = zone->zone_start_pfn; 4494 end_pfn = zone_end_pfn(zone); 4495 start_pfn = roundup(start_pfn, pageblock_nr_pages); 4496 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >> 4497 pageblock_order; 4498 4499 /* 4500 * Reserve blocks are generally in place to help high-order atomic 4501 * allocations that are short-lived. A min_free_kbytes value that 4502 * would result in more than 2 reserve blocks for atomic allocations 4503 * is assumed to be in place to help anti-fragmentation for the 4504 * future allocation of hugepages at runtime. 4505 */ 4506 reserve = min(2, reserve); 4507 old_reserve = zone->nr_migrate_reserve_block; 4508 4509 /* When memory hot-add, we almost always need to do nothing */ 4510 if (reserve == old_reserve) 4511 return; 4512 zone->nr_migrate_reserve_block = reserve; 4513 4514 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) { 4515 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone))) 4516 return; 4517 4518 if (!pfn_valid(pfn)) 4519 continue; 4520 page = pfn_to_page(pfn); 4521 4522 /* Watch out for overlapping nodes */ 4523 if (page_to_nid(page) != zone_to_nid(zone)) 4524 continue; 4525 4526 block_migratetype = get_pageblock_migratetype(page); 4527 4528 /* Only test what is necessary when the reserves are not met */ 4529 if (reserve > 0) { 4530 /* 4531 * Blocks with reserved pages will never free, skip 4532 * them. 4533 */ 4534 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn); 4535 if (pageblock_is_reserved(pfn, block_end_pfn)) 4536 continue; 4537 4538 /* If this block is reserved, account for it */ 4539 if (block_migratetype == MIGRATE_RESERVE) { 4540 reserve--; 4541 continue; 4542 } 4543 4544 /* Suitable for reserving if this block is movable */ 4545 if (block_migratetype == MIGRATE_MOVABLE) { 4546 set_pageblock_migratetype(page, 4547 MIGRATE_RESERVE); 4548 move_freepages_block(zone, page, 4549 MIGRATE_RESERVE); 4550 reserve--; 4551 continue; 4552 } 4553 } else if (!old_reserve) { 4554 /* 4555 * At boot time we don't need to scan the whole zone 4556 * for turning off MIGRATE_RESERVE. 4557 */ 4558 break; 4559 } 4560 4561 /* 4562 * If the reserve is met and this is a previous reserved block, 4563 * take it back 4564 */ 4565 if (block_migratetype == MIGRATE_RESERVE) { 4566 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4567 move_freepages_block(zone, page, MIGRATE_MOVABLE); 4568 } 4569 } 4570 } 4571 4572 /* 4573 * Initially all pages are reserved - free ones are freed 4574 * up by free_all_bootmem() once the early boot process is 4575 * done. Non-atomic initialization, single-pass. 4576 */ 4577 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 4578 unsigned long start_pfn, enum memmap_context context) 4579 { 4580 pg_data_t *pgdat = NODE_DATA(nid); 4581 unsigned long end_pfn = start_pfn + size; 4582 unsigned long pfn; 4583 struct zone *z; 4584 unsigned long nr_initialised = 0; 4585 4586 if (highest_memmap_pfn < end_pfn - 1) 4587 highest_memmap_pfn = end_pfn - 1; 4588 4589 z = &pgdat->node_zones[zone]; 4590 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 4591 /* 4592 * There can be holes in boot-time mem_map[]s 4593 * handed to this function. They do not 4594 * exist on hotplugged memory. 4595 */ 4596 if (context == MEMMAP_EARLY) { 4597 if (!early_pfn_valid(pfn)) 4598 continue; 4599 if (!early_pfn_in_nid(pfn, nid)) 4600 continue; 4601 if (!update_defer_init(pgdat, pfn, end_pfn, 4602 &nr_initialised)) 4603 break; 4604 } 4605 4606 /* 4607 * Mark the block movable so that blocks are reserved for 4608 * movable at startup. This will force kernel allocations 4609 * to reserve their blocks rather than leaking throughout 4610 * the address space during boot when many long-lived 4611 * kernel allocations are made. Later some blocks near 4612 * the start are marked MIGRATE_RESERVE by 4613 * setup_zone_migrate_reserve() 4614 * 4615 * bitmap is created for zone's valid pfn range. but memmap 4616 * can be created for invalid pages (for alignment) 4617 * check here not to call set_pageblock_migratetype() against 4618 * pfn out of zone. 4619 */ 4620 if (!(pfn & (pageblock_nr_pages - 1))) { 4621 struct page *page = pfn_to_page(pfn); 4622 4623 __init_single_page(page, pfn, zone, nid); 4624 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 4625 } else { 4626 __init_single_pfn(pfn, zone, nid); 4627 } 4628 } 4629 } 4630 4631 static void __meminit zone_init_free_lists(struct zone *zone) 4632 { 4633 unsigned int order, t; 4634 for_each_migratetype_order(order, t) { 4635 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 4636 zone->free_area[order].nr_free = 0; 4637 } 4638 } 4639 4640 #ifndef __HAVE_ARCH_MEMMAP_INIT 4641 #define memmap_init(size, nid, zone, start_pfn) \ 4642 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 4643 #endif 4644 4645 static int zone_batchsize(struct zone *zone) 4646 { 4647 #ifdef CONFIG_MMU 4648 int batch; 4649 4650 /* 4651 * The per-cpu-pages pools are set to around 1000th of the 4652 * size of the zone. But no more than 1/2 of a meg. 4653 * 4654 * OK, so we don't know how big the cache is. So guess. 4655 */ 4656 batch = zone->managed_pages / 1024; 4657 if (batch * PAGE_SIZE > 512 * 1024) 4658 batch = (512 * 1024) / PAGE_SIZE; 4659 batch /= 4; /* We effectively *= 4 below */ 4660 if (batch < 1) 4661 batch = 1; 4662 4663 /* 4664 * Clamp the batch to a 2^n - 1 value. Having a power 4665 * of 2 value was found to be more likely to have 4666 * suboptimal cache aliasing properties in some cases. 4667 * 4668 * For example if 2 tasks are alternately allocating 4669 * batches of pages, one task can end up with a lot 4670 * of pages of one half of the possible page colors 4671 * and the other with pages of the other colors. 4672 */ 4673 batch = rounddown_pow_of_two(batch + batch/2) - 1; 4674 4675 return batch; 4676 4677 #else 4678 /* The deferral and batching of frees should be suppressed under NOMMU 4679 * conditions. 4680 * 4681 * The problem is that NOMMU needs to be able to allocate large chunks 4682 * of contiguous memory as there's no hardware page translation to 4683 * assemble apparent contiguous memory from discontiguous pages. 4684 * 4685 * Queueing large contiguous runs of pages for batching, however, 4686 * causes the pages to actually be freed in smaller chunks. As there 4687 * can be a significant delay between the individual batches being 4688 * recycled, this leads to the once large chunks of space being 4689 * fragmented and becoming unavailable for high-order allocations. 4690 */ 4691 return 0; 4692 #endif 4693 } 4694 4695 /* 4696 * pcp->high and pcp->batch values are related and dependent on one another: 4697 * ->batch must never be higher then ->high. 4698 * The following function updates them in a safe manner without read side 4699 * locking. 4700 * 4701 * Any new users of pcp->batch and pcp->high should ensure they can cope with 4702 * those fields changing asynchronously (acording the the above rule). 4703 * 4704 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function 4705 * outside of boot time (or some other assurance that no concurrent updaters 4706 * exist). 4707 */ 4708 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, 4709 unsigned long batch) 4710 { 4711 /* start with a fail safe value for batch */ 4712 pcp->batch = 1; 4713 smp_wmb(); 4714 4715 /* Update high, then batch, in order */ 4716 pcp->high = high; 4717 smp_wmb(); 4718 4719 pcp->batch = batch; 4720 } 4721 4722 /* a companion to pageset_set_high() */ 4723 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) 4724 { 4725 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); 4726 } 4727 4728 static void pageset_init(struct per_cpu_pageset *p) 4729 { 4730 struct per_cpu_pages *pcp; 4731 int migratetype; 4732 4733 memset(p, 0, sizeof(*p)); 4734 4735 pcp = &p->pcp; 4736 pcp->count = 0; 4737 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 4738 INIT_LIST_HEAD(&pcp->lists[migratetype]); 4739 } 4740 4741 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 4742 { 4743 pageset_init(p); 4744 pageset_set_batch(p, batch); 4745 } 4746 4747 /* 4748 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist 4749 * to the value high for the pageset p. 4750 */ 4751 static void pageset_set_high(struct per_cpu_pageset *p, 4752 unsigned long high) 4753 { 4754 unsigned long batch = max(1UL, high / 4); 4755 if ((high / 4) > (PAGE_SHIFT * 8)) 4756 batch = PAGE_SHIFT * 8; 4757 4758 pageset_update(&p->pcp, high, batch); 4759 } 4760 4761 static void pageset_set_high_and_batch(struct zone *zone, 4762 struct per_cpu_pageset *pcp) 4763 { 4764 if (percpu_pagelist_fraction) 4765 pageset_set_high(pcp, 4766 (zone->managed_pages / 4767 percpu_pagelist_fraction)); 4768 else 4769 pageset_set_batch(pcp, zone_batchsize(zone)); 4770 } 4771 4772 static void __meminit zone_pageset_init(struct zone *zone, int cpu) 4773 { 4774 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 4775 4776 pageset_init(pcp); 4777 pageset_set_high_and_batch(zone, pcp); 4778 } 4779 4780 static void __meminit setup_zone_pageset(struct zone *zone) 4781 { 4782 int cpu; 4783 zone->pageset = alloc_percpu(struct per_cpu_pageset); 4784 for_each_possible_cpu(cpu) 4785 zone_pageset_init(zone, cpu); 4786 } 4787 4788 /* 4789 * Allocate per cpu pagesets and initialize them. 4790 * Before this call only boot pagesets were available. 4791 */ 4792 void __init setup_per_cpu_pageset(void) 4793 { 4794 struct zone *zone; 4795 4796 for_each_populated_zone(zone) 4797 setup_zone_pageset(zone); 4798 } 4799 4800 static noinline __init_refok 4801 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages) 4802 { 4803 int i; 4804 size_t alloc_size; 4805 4806 /* 4807 * The per-page waitqueue mechanism uses hashed waitqueues 4808 * per zone. 4809 */ 4810 zone->wait_table_hash_nr_entries = 4811 wait_table_hash_nr_entries(zone_size_pages); 4812 zone->wait_table_bits = 4813 wait_table_bits(zone->wait_table_hash_nr_entries); 4814 alloc_size = zone->wait_table_hash_nr_entries 4815 * sizeof(wait_queue_head_t); 4816 4817 if (!slab_is_available()) { 4818 zone->wait_table = (wait_queue_head_t *) 4819 memblock_virt_alloc_node_nopanic( 4820 alloc_size, zone->zone_pgdat->node_id); 4821 } else { 4822 /* 4823 * This case means that a zone whose size was 0 gets new memory 4824 * via memory hot-add. 4825 * But it may be the case that a new node was hot-added. In 4826 * this case vmalloc() will not be able to use this new node's 4827 * memory - this wait_table must be initialized to use this new 4828 * node itself as well. 4829 * To use this new node's memory, further consideration will be 4830 * necessary. 4831 */ 4832 zone->wait_table = vmalloc(alloc_size); 4833 } 4834 if (!zone->wait_table) 4835 return -ENOMEM; 4836 4837 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i) 4838 init_waitqueue_head(zone->wait_table + i); 4839 4840 return 0; 4841 } 4842 4843 static __meminit void zone_pcp_init(struct zone *zone) 4844 { 4845 /* 4846 * per cpu subsystem is not up at this point. The following code 4847 * relies on the ability of the linker to provide the 4848 * offset of a (static) per cpu variable into the per cpu area. 4849 */ 4850 zone->pageset = &boot_pageset; 4851 4852 if (populated_zone(zone)) 4853 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 4854 zone->name, zone->present_pages, 4855 zone_batchsize(zone)); 4856 } 4857 4858 int __meminit init_currently_empty_zone(struct zone *zone, 4859 unsigned long zone_start_pfn, 4860 unsigned long size, 4861 enum memmap_context context) 4862 { 4863 struct pglist_data *pgdat = zone->zone_pgdat; 4864 int ret; 4865 ret = zone_wait_table_init(zone, size); 4866 if (ret) 4867 return ret; 4868 pgdat->nr_zones = zone_idx(zone) + 1; 4869 4870 zone->zone_start_pfn = zone_start_pfn; 4871 4872 mminit_dprintk(MMINIT_TRACE, "memmap_init", 4873 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 4874 pgdat->node_id, 4875 (unsigned long)zone_idx(zone), 4876 zone_start_pfn, (zone_start_pfn + size)); 4877 4878 zone_init_free_lists(zone); 4879 4880 return 0; 4881 } 4882 4883 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 4884 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 4885 4886 /* 4887 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 4888 */ 4889 int __meminit __early_pfn_to_nid(unsigned long pfn, 4890 struct mminit_pfnnid_cache *state) 4891 { 4892 unsigned long start_pfn, end_pfn; 4893 int nid; 4894 4895 if (state->last_start <= pfn && pfn < state->last_end) 4896 return state->last_nid; 4897 4898 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); 4899 if (nid != -1) { 4900 state->last_start = start_pfn; 4901 state->last_end = end_pfn; 4902 state->last_nid = nid; 4903 } 4904 4905 return nid; 4906 } 4907 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 4908 4909 /** 4910 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range 4911 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 4912 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid 4913 * 4914 * If an architecture guarantees that all ranges registered contain no holes 4915 * and may be freed, this this function may be used instead of calling 4916 * memblock_free_early_nid() manually. 4917 */ 4918 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 4919 { 4920 unsigned long start_pfn, end_pfn; 4921 int i, this_nid; 4922 4923 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 4924 start_pfn = min(start_pfn, max_low_pfn); 4925 end_pfn = min(end_pfn, max_low_pfn); 4926 4927 if (start_pfn < end_pfn) 4928 memblock_free_early_nid(PFN_PHYS(start_pfn), 4929 (end_pfn - start_pfn) << PAGE_SHIFT, 4930 this_nid); 4931 } 4932 } 4933 4934 /** 4935 * sparse_memory_present_with_active_regions - Call memory_present for each active range 4936 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 4937 * 4938 * If an architecture guarantees that all ranges registered contain no holes and may 4939 * be freed, this function may be used instead of calling memory_present() manually. 4940 */ 4941 void __init sparse_memory_present_with_active_regions(int nid) 4942 { 4943 unsigned long start_pfn, end_pfn; 4944 int i, this_nid; 4945 4946 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 4947 memory_present(this_nid, start_pfn, end_pfn); 4948 } 4949 4950 /** 4951 * get_pfn_range_for_nid - Return the start and end page frames for a node 4952 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 4953 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 4954 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 4955 * 4956 * It returns the start and end page frame of a node based on information 4957 * provided by memblock_set_node(). If called for a node 4958 * with no available memory, a warning is printed and the start and end 4959 * PFNs will be 0. 4960 */ 4961 void __meminit get_pfn_range_for_nid(unsigned int nid, 4962 unsigned long *start_pfn, unsigned long *end_pfn) 4963 { 4964 unsigned long this_start_pfn, this_end_pfn; 4965 int i; 4966 4967 *start_pfn = -1UL; 4968 *end_pfn = 0; 4969 4970 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 4971 *start_pfn = min(*start_pfn, this_start_pfn); 4972 *end_pfn = max(*end_pfn, this_end_pfn); 4973 } 4974 4975 if (*start_pfn == -1UL) 4976 *start_pfn = 0; 4977 } 4978 4979 /* 4980 * This finds a zone that can be used for ZONE_MOVABLE pages. The 4981 * assumption is made that zones within a node are ordered in monotonic 4982 * increasing memory addresses so that the "highest" populated zone is used 4983 */ 4984 static void __init find_usable_zone_for_movable(void) 4985 { 4986 int zone_index; 4987 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 4988 if (zone_index == ZONE_MOVABLE) 4989 continue; 4990 4991 if (arch_zone_highest_possible_pfn[zone_index] > 4992 arch_zone_lowest_possible_pfn[zone_index]) 4993 break; 4994 } 4995 4996 VM_BUG_ON(zone_index == -1); 4997 movable_zone = zone_index; 4998 } 4999 5000 /* 5001 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 5002 * because it is sized independent of architecture. Unlike the other zones, 5003 * the starting point for ZONE_MOVABLE is not fixed. It may be different 5004 * in each node depending on the size of each node and how evenly kernelcore 5005 * is distributed. This helper function adjusts the zone ranges 5006 * provided by the architecture for a given node by using the end of the 5007 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 5008 * zones within a node are in order of monotonic increases memory addresses 5009 */ 5010 static void __meminit adjust_zone_range_for_zone_movable(int nid, 5011 unsigned long zone_type, 5012 unsigned long node_start_pfn, 5013 unsigned long node_end_pfn, 5014 unsigned long *zone_start_pfn, 5015 unsigned long *zone_end_pfn) 5016 { 5017 /* Only adjust if ZONE_MOVABLE is on this node */ 5018 if (zone_movable_pfn[nid]) { 5019 /* Size ZONE_MOVABLE */ 5020 if (zone_type == ZONE_MOVABLE) { 5021 *zone_start_pfn = zone_movable_pfn[nid]; 5022 *zone_end_pfn = min(node_end_pfn, 5023 arch_zone_highest_possible_pfn[movable_zone]); 5024 5025 /* Adjust for ZONE_MOVABLE starting within this range */ 5026 } else if (*zone_start_pfn < zone_movable_pfn[nid] && 5027 *zone_end_pfn > zone_movable_pfn[nid]) { 5028 *zone_end_pfn = zone_movable_pfn[nid]; 5029 5030 /* Check if this whole range is within ZONE_MOVABLE */ 5031 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 5032 *zone_start_pfn = *zone_end_pfn; 5033 } 5034 } 5035 5036 /* 5037 * Return the number of pages a zone spans in a node, including holes 5038 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 5039 */ 5040 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 5041 unsigned long zone_type, 5042 unsigned long node_start_pfn, 5043 unsigned long node_end_pfn, 5044 unsigned long *ignored) 5045 { 5046 unsigned long zone_start_pfn, zone_end_pfn; 5047 5048 /* Get the start and end of the zone */ 5049 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 5050 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 5051 adjust_zone_range_for_zone_movable(nid, zone_type, 5052 node_start_pfn, node_end_pfn, 5053 &zone_start_pfn, &zone_end_pfn); 5054 5055 /* Check that this node has pages within the zone's required range */ 5056 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn) 5057 return 0; 5058 5059 /* Move the zone boundaries inside the node if necessary */ 5060 zone_end_pfn = min(zone_end_pfn, node_end_pfn); 5061 zone_start_pfn = max(zone_start_pfn, node_start_pfn); 5062 5063 /* Return the spanned pages */ 5064 return zone_end_pfn - zone_start_pfn; 5065 } 5066 5067 /* 5068 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 5069 * then all holes in the requested range will be accounted for. 5070 */ 5071 unsigned long __meminit __absent_pages_in_range(int nid, 5072 unsigned long range_start_pfn, 5073 unsigned long range_end_pfn) 5074 { 5075 unsigned long nr_absent = range_end_pfn - range_start_pfn; 5076 unsigned long start_pfn, end_pfn; 5077 int i; 5078 5079 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5080 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 5081 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 5082 nr_absent -= end_pfn - start_pfn; 5083 } 5084 return nr_absent; 5085 } 5086 5087 /** 5088 * absent_pages_in_range - Return number of page frames in holes within a range 5089 * @start_pfn: The start PFN to start searching for holes 5090 * @end_pfn: The end PFN to stop searching for holes 5091 * 5092 * It returns the number of pages frames in memory holes within a range. 5093 */ 5094 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 5095 unsigned long end_pfn) 5096 { 5097 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 5098 } 5099 5100 /* Return the number of page frames in holes in a zone on a node */ 5101 static unsigned long __meminit zone_absent_pages_in_node(int nid, 5102 unsigned long zone_type, 5103 unsigned long node_start_pfn, 5104 unsigned long node_end_pfn, 5105 unsigned long *ignored) 5106 { 5107 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 5108 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 5109 unsigned long zone_start_pfn, zone_end_pfn; 5110 5111 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 5112 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 5113 5114 adjust_zone_range_for_zone_movable(nid, zone_type, 5115 node_start_pfn, node_end_pfn, 5116 &zone_start_pfn, &zone_end_pfn); 5117 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 5118 } 5119 5120 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5121 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 5122 unsigned long zone_type, 5123 unsigned long node_start_pfn, 5124 unsigned long node_end_pfn, 5125 unsigned long *zones_size) 5126 { 5127 return zones_size[zone_type]; 5128 } 5129 5130 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 5131 unsigned long zone_type, 5132 unsigned long node_start_pfn, 5133 unsigned long node_end_pfn, 5134 unsigned long *zholes_size) 5135 { 5136 if (!zholes_size) 5137 return 0; 5138 5139 return zholes_size[zone_type]; 5140 } 5141 5142 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5143 5144 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 5145 unsigned long node_start_pfn, 5146 unsigned long node_end_pfn, 5147 unsigned long *zones_size, 5148 unsigned long *zholes_size) 5149 { 5150 unsigned long realtotalpages = 0, totalpages = 0; 5151 enum zone_type i; 5152 5153 for (i = 0; i < MAX_NR_ZONES; i++) { 5154 struct zone *zone = pgdat->node_zones + i; 5155 unsigned long size, real_size; 5156 5157 size = zone_spanned_pages_in_node(pgdat->node_id, i, 5158 node_start_pfn, 5159 node_end_pfn, 5160 zones_size); 5161 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i, 5162 node_start_pfn, node_end_pfn, 5163 zholes_size); 5164 zone->spanned_pages = size; 5165 zone->present_pages = real_size; 5166 5167 totalpages += size; 5168 realtotalpages += real_size; 5169 } 5170 5171 pgdat->node_spanned_pages = totalpages; 5172 pgdat->node_present_pages = realtotalpages; 5173 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 5174 realtotalpages); 5175 } 5176 5177 #ifndef CONFIG_SPARSEMEM 5178 /* 5179 * Calculate the size of the zone->blockflags rounded to an unsigned long 5180 * Start by making sure zonesize is a multiple of pageblock_order by rounding 5181 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 5182 * round what is now in bits to nearest long in bits, then return it in 5183 * bytes. 5184 */ 5185 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) 5186 { 5187 unsigned long usemapsize; 5188 5189 zonesize += zone_start_pfn & (pageblock_nr_pages-1); 5190 usemapsize = roundup(zonesize, pageblock_nr_pages); 5191 usemapsize = usemapsize >> pageblock_order; 5192 usemapsize *= NR_PAGEBLOCK_BITS; 5193 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 5194 5195 return usemapsize / 8; 5196 } 5197 5198 static void __init setup_usemap(struct pglist_data *pgdat, 5199 struct zone *zone, 5200 unsigned long zone_start_pfn, 5201 unsigned long zonesize) 5202 { 5203 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); 5204 zone->pageblock_flags = NULL; 5205 if (usemapsize) 5206 zone->pageblock_flags = 5207 memblock_virt_alloc_node_nopanic(usemapsize, 5208 pgdat->node_id); 5209 } 5210 #else 5211 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, 5212 unsigned long zone_start_pfn, unsigned long zonesize) {} 5213 #endif /* CONFIG_SPARSEMEM */ 5214 5215 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 5216 5217 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 5218 void __paginginit set_pageblock_order(void) 5219 { 5220 unsigned int order; 5221 5222 /* Check that pageblock_nr_pages has not already been setup */ 5223 if (pageblock_order) 5224 return; 5225 5226 if (HPAGE_SHIFT > PAGE_SHIFT) 5227 order = HUGETLB_PAGE_ORDER; 5228 else 5229 order = MAX_ORDER - 1; 5230 5231 /* 5232 * Assume the largest contiguous order of interest is a huge page. 5233 * This value may be variable depending on boot parameters on IA64 and 5234 * powerpc. 5235 */ 5236 pageblock_order = order; 5237 } 5238 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5239 5240 /* 5241 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 5242 * is unused as pageblock_order is set at compile-time. See 5243 * include/linux/pageblock-flags.h for the values of pageblock_order based on 5244 * the kernel config 5245 */ 5246 void __paginginit set_pageblock_order(void) 5247 { 5248 } 5249 5250 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5251 5252 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, 5253 unsigned long present_pages) 5254 { 5255 unsigned long pages = spanned_pages; 5256 5257 /* 5258 * Provide a more accurate estimation if there are holes within 5259 * the zone and SPARSEMEM is in use. If there are holes within the 5260 * zone, each populated memory region may cost us one or two extra 5261 * memmap pages due to alignment because memmap pages for each 5262 * populated regions may not naturally algined on page boundary. 5263 * So the (present_pages >> 4) heuristic is a tradeoff for that. 5264 */ 5265 if (spanned_pages > present_pages + (present_pages >> 4) && 5266 IS_ENABLED(CONFIG_SPARSEMEM)) 5267 pages = present_pages; 5268 5269 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; 5270 } 5271 5272 /* 5273 * Set up the zone data structures: 5274 * - mark all pages reserved 5275 * - mark all memory queues empty 5276 * - clear the memory bitmaps 5277 * 5278 * NOTE: pgdat should get zeroed by caller. 5279 */ 5280 static void __paginginit free_area_init_core(struct pglist_data *pgdat, 5281 unsigned long node_start_pfn, unsigned long node_end_pfn) 5282 { 5283 enum zone_type j; 5284 int nid = pgdat->node_id; 5285 unsigned long zone_start_pfn = pgdat->node_start_pfn; 5286 int ret; 5287 5288 pgdat_resize_init(pgdat); 5289 #ifdef CONFIG_NUMA_BALANCING 5290 spin_lock_init(&pgdat->numabalancing_migrate_lock); 5291 pgdat->numabalancing_migrate_nr_pages = 0; 5292 pgdat->numabalancing_migrate_next_window = jiffies; 5293 #endif 5294 init_waitqueue_head(&pgdat->kswapd_wait); 5295 init_waitqueue_head(&pgdat->pfmemalloc_wait); 5296 pgdat_page_ext_init(pgdat); 5297 5298 for (j = 0; j < MAX_NR_ZONES; j++) { 5299 struct zone *zone = pgdat->node_zones + j; 5300 unsigned long size, realsize, freesize, memmap_pages; 5301 5302 size = zone->spanned_pages; 5303 realsize = freesize = zone->present_pages; 5304 5305 /* 5306 * Adjust freesize so that it accounts for how much memory 5307 * is used by this zone for memmap. This affects the watermark 5308 * and per-cpu initialisations 5309 */ 5310 memmap_pages = calc_memmap_size(size, realsize); 5311 if (!is_highmem_idx(j)) { 5312 if (freesize >= memmap_pages) { 5313 freesize -= memmap_pages; 5314 if (memmap_pages) 5315 printk(KERN_DEBUG 5316 " %s zone: %lu pages used for memmap\n", 5317 zone_names[j], memmap_pages); 5318 } else 5319 printk(KERN_WARNING 5320 " %s zone: %lu pages exceeds freesize %lu\n", 5321 zone_names[j], memmap_pages, freesize); 5322 } 5323 5324 /* Account for reserved pages */ 5325 if (j == 0 && freesize > dma_reserve) { 5326 freesize -= dma_reserve; 5327 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 5328 zone_names[0], dma_reserve); 5329 } 5330 5331 if (!is_highmem_idx(j)) 5332 nr_kernel_pages += freesize; 5333 /* Charge for highmem memmap if there are enough kernel pages */ 5334 else if (nr_kernel_pages > memmap_pages * 2) 5335 nr_kernel_pages -= memmap_pages; 5336 nr_all_pages += freesize; 5337 5338 /* 5339 * Set an approximate value for lowmem here, it will be adjusted 5340 * when the bootmem allocator frees pages into the buddy system. 5341 * And all highmem pages will be managed by the buddy system. 5342 */ 5343 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; 5344 #ifdef CONFIG_NUMA 5345 zone->node = nid; 5346 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio) 5347 / 100; 5348 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100; 5349 #endif 5350 zone->name = zone_names[j]; 5351 spin_lock_init(&zone->lock); 5352 spin_lock_init(&zone->lru_lock); 5353 zone_seqlock_init(zone); 5354 zone->zone_pgdat = pgdat; 5355 zone_pcp_init(zone); 5356 5357 /* For bootup, initialized properly in watermark setup */ 5358 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages); 5359 5360 lruvec_init(&zone->lruvec); 5361 if (!size) 5362 continue; 5363 5364 set_pageblock_order(); 5365 setup_usemap(pgdat, zone, zone_start_pfn, size); 5366 ret = init_currently_empty_zone(zone, zone_start_pfn, 5367 size, MEMMAP_EARLY); 5368 BUG_ON(ret); 5369 memmap_init(size, nid, j, zone_start_pfn); 5370 zone_start_pfn += size; 5371 } 5372 } 5373 5374 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat) 5375 { 5376 /* Skip empty nodes */ 5377 if (!pgdat->node_spanned_pages) 5378 return; 5379 5380 #ifdef CONFIG_FLAT_NODE_MEM_MAP 5381 /* ia64 gets its own node_mem_map, before this, without bootmem */ 5382 if (!pgdat->node_mem_map) { 5383 unsigned long size, start, end; 5384 struct page *map; 5385 5386 /* 5387 * The zone's endpoints aren't required to be MAX_ORDER 5388 * aligned but the node_mem_map endpoints must be in order 5389 * for the buddy allocator to function correctly. 5390 */ 5391 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 5392 end = pgdat_end_pfn(pgdat); 5393 end = ALIGN(end, MAX_ORDER_NR_PAGES); 5394 size = (end - start) * sizeof(struct page); 5395 map = alloc_remap(pgdat->node_id, size); 5396 if (!map) 5397 map = memblock_virt_alloc_node_nopanic(size, 5398 pgdat->node_id); 5399 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start); 5400 } 5401 #ifndef CONFIG_NEED_MULTIPLE_NODES 5402 /* 5403 * With no DISCONTIG, the global mem_map is just set as node 0's 5404 */ 5405 if (pgdat == NODE_DATA(0)) { 5406 mem_map = NODE_DATA(0)->node_mem_map; 5407 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5408 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 5409 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET); 5410 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5411 } 5412 #endif 5413 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 5414 } 5415 5416 void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 5417 unsigned long node_start_pfn, unsigned long *zholes_size) 5418 { 5419 pg_data_t *pgdat = NODE_DATA(nid); 5420 unsigned long start_pfn = 0; 5421 unsigned long end_pfn = 0; 5422 5423 /* pg_data_t should be reset to zero when it's allocated */ 5424 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx); 5425 5426 reset_deferred_meminit(pgdat); 5427 pgdat->node_id = nid; 5428 pgdat->node_start_pfn = node_start_pfn; 5429 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5430 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 5431 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid, 5432 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1); 5433 #endif 5434 calculate_node_totalpages(pgdat, start_pfn, end_pfn, 5435 zones_size, zholes_size); 5436 5437 alloc_node_mem_map(pgdat); 5438 #ifdef CONFIG_FLAT_NODE_MEM_MAP 5439 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 5440 nid, (unsigned long)pgdat, 5441 (unsigned long)pgdat->node_mem_map); 5442 #endif 5443 5444 free_area_init_core(pgdat, start_pfn, end_pfn); 5445 } 5446 5447 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5448 5449 #if MAX_NUMNODES > 1 5450 /* 5451 * Figure out the number of possible node ids. 5452 */ 5453 void __init setup_nr_node_ids(void) 5454 { 5455 unsigned int node; 5456 unsigned int highest = 0; 5457 5458 for_each_node_mask(node, node_possible_map) 5459 highest = node; 5460 nr_node_ids = highest + 1; 5461 } 5462 #endif 5463 5464 /** 5465 * node_map_pfn_alignment - determine the maximum internode alignment 5466 * 5467 * This function should be called after node map is populated and sorted. 5468 * It calculates the maximum power of two alignment which can distinguish 5469 * all the nodes. 5470 * 5471 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 5472 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 5473 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 5474 * shifted, 1GiB is enough and this function will indicate so. 5475 * 5476 * This is used to test whether pfn -> nid mapping of the chosen memory 5477 * model has fine enough granularity to avoid incorrect mapping for the 5478 * populated node map. 5479 * 5480 * Returns the determined alignment in pfn's. 0 if there is no alignment 5481 * requirement (single node). 5482 */ 5483 unsigned long __init node_map_pfn_alignment(void) 5484 { 5485 unsigned long accl_mask = 0, last_end = 0; 5486 unsigned long start, end, mask; 5487 int last_nid = -1; 5488 int i, nid; 5489 5490 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 5491 if (!start || last_nid < 0 || last_nid == nid) { 5492 last_nid = nid; 5493 last_end = end; 5494 continue; 5495 } 5496 5497 /* 5498 * Start with a mask granular enough to pin-point to the 5499 * start pfn and tick off bits one-by-one until it becomes 5500 * too coarse to separate the current node from the last. 5501 */ 5502 mask = ~((1 << __ffs(start)) - 1); 5503 while (mask && last_end <= (start & (mask << 1))) 5504 mask <<= 1; 5505 5506 /* accumulate all internode masks */ 5507 accl_mask |= mask; 5508 } 5509 5510 /* convert mask to number of pages */ 5511 return ~accl_mask + 1; 5512 } 5513 5514 /* Find the lowest pfn for a node */ 5515 static unsigned long __init find_min_pfn_for_node(int nid) 5516 { 5517 unsigned long min_pfn = ULONG_MAX; 5518 unsigned long start_pfn; 5519 int i; 5520 5521 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 5522 min_pfn = min(min_pfn, start_pfn); 5523 5524 if (min_pfn == ULONG_MAX) { 5525 printk(KERN_WARNING 5526 "Could not find start_pfn for node %d\n", nid); 5527 return 0; 5528 } 5529 5530 return min_pfn; 5531 } 5532 5533 /** 5534 * find_min_pfn_with_active_regions - Find the minimum PFN registered 5535 * 5536 * It returns the minimum PFN based on information provided via 5537 * memblock_set_node(). 5538 */ 5539 unsigned long __init find_min_pfn_with_active_regions(void) 5540 { 5541 return find_min_pfn_for_node(MAX_NUMNODES); 5542 } 5543 5544 /* 5545 * early_calculate_totalpages() 5546 * Sum pages in active regions for movable zone. 5547 * Populate N_MEMORY for calculating usable_nodes. 5548 */ 5549 static unsigned long __init early_calculate_totalpages(void) 5550 { 5551 unsigned long totalpages = 0; 5552 unsigned long start_pfn, end_pfn; 5553 int i, nid; 5554 5555 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 5556 unsigned long pages = end_pfn - start_pfn; 5557 5558 totalpages += pages; 5559 if (pages) 5560 node_set_state(nid, N_MEMORY); 5561 } 5562 return totalpages; 5563 } 5564 5565 /* 5566 * Find the PFN the Movable zone begins in each node. Kernel memory 5567 * is spread evenly between nodes as long as the nodes have enough 5568 * memory. When they don't, some nodes will have more kernelcore than 5569 * others 5570 */ 5571 static void __init find_zone_movable_pfns_for_nodes(void) 5572 { 5573 int i, nid; 5574 unsigned long usable_startpfn; 5575 unsigned long kernelcore_node, kernelcore_remaining; 5576 /* save the state before borrow the nodemask */ 5577 nodemask_t saved_node_state = node_states[N_MEMORY]; 5578 unsigned long totalpages = early_calculate_totalpages(); 5579 int usable_nodes = nodes_weight(node_states[N_MEMORY]); 5580 struct memblock_region *r; 5581 5582 /* Need to find movable_zone earlier when movable_node is specified. */ 5583 find_usable_zone_for_movable(); 5584 5585 /* 5586 * If movable_node is specified, ignore kernelcore and movablecore 5587 * options. 5588 */ 5589 if (movable_node_is_enabled()) { 5590 for_each_memblock(memory, r) { 5591 if (!memblock_is_hotpluggable(r)) 5592 continue; 5593 5594 nid = r->nid; 5595 5596 usable_startpfn = PFN_DOWN(r->base); 5597 zone_movable_pfn[nid] = zone_movable_pfn[nid] ? 5598 min(usable_startpfn, zone_movable_pfn[nid]) : 5599 usable_startpfn; 5600 } 5601 5602 goto out2; 5603 } 5604 5605 /* 5606 * If movablecore=nn[KMG] was specified, calculate what size of 5607 * kernelcore that corresponds so that memory usable for 5608 * any allocation type is evenly spread. If both kernelcore 5609 * and movablecore are specified, then the value of kernelcore 5610 * will be used for required_kernelcore if it's greater than 5611 * what movablecore would have allowed. 5612 */ 5613 if (required_movablecore) { 5614 unsigned long corepages; 5615 5616 /* 5617 * Round-up so that ZONE_MOVABLE is at least as large as what 5618 * was requested by the user 5619 */ 5620 required_movablecore = 5621 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 5622 corepages = totalpages - required_movablecore; 5623 5624 required_kernelcore = max(required_kernelcore, corepages); 5625 } 5626 5627 /* If kernelcore was not specified, there is no ZONE_MOVABLE */ 5628 if (!required_kernelcore) 5629 goto out; 5630 5631 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 5632 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 5633 5634 restart: 5635 /* Spread kernelcore memory as evenly as possible throughout nodes */ 5636 kernelcore_node = required_kernelcore / usable_nodes; 5637 for_each_node_state(nid, N_MEMORY) { 5638 unsigned long start_pfn, end_pfn; 5639 5640 /* 5641 * Recalculate kernelcore_node if the division per node 5642 * now exceeds what is necessary to satisfy the requested 5643 * amount of memory for the kernel 5644 */ 5645 if (required_kernelcore < kernelcore_node) 5646 kernelcore_node = required_kernelcore / usable_nodes; 5647 5648 /* 5649 * As the map is walked, we track how much memory is usable 5650 * by the kernel using kernelcore_remaining. When it is 5651 * 0, the rest of the node is usable by ZONE_MOVABLE 5652 */ 5653 kernelcore_remaining = kernelcore_node; 5654 5655 /* Go through each range of PFNs within this node */ 5656 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5657 unsigned long size_pages; 5658 5659 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 5660 if (start_pfn >= end_pfn) 5661 continue; 5662 5663 /* Account for what is only usable for kernelcore */ 5664 if (start_pfn < usable_startpfn) { 5665 unsigned long kernel_pages; 5666 kernel_pages = min(end_pfn, usable_startpfn) 5667 - start_pfn; 5668 5669 kernelcore_remaining -= min(kernel_pages, 5670 kernelcore_remaining); 5671 required_kernelcore -= min(kernel_pages, 5672 required_kernelcore); 5673 5674 /* Continue if range is now fully accounted */ 5675 if (end_pfn <= usable_startpfn) { 5676 5677 /* 5678 * Push zone_movable_pfn to the end so 5679 * that if we have to rebalance 5680 * kernelcore across nodes, we will 5681 * not double account here 5682 */ 5683 zone_movable_pfn[nid] = end_pfn; 5684 continue; 5685 } 5686 start_pfn = usable_startpfn; 5687 } 5688 5689 /* 5690 * The usable PFN range for ZONE_MOVABLE is from 5691 * start_pfn->end_pfn. Calculate size_pages as the 5692 * number of pages used as kernelcore 5693 */ 5694 size_pages = end_pfn - start_pfn; 5695 if (size_pages > kernelcore_remaining) 5696 size_pages = kernelcore_remaining; 5697 zone_movable_pfn[nid] = start_pfn + size_pages; 5698 5699 /* 5700 * Some kernelcore has been met, update counts and 5701 * break if the kernelcore for this node has been 5702 * satisfied 5703 */ 5704 required_kernelcore -= min(required_kernelcore, 5705 size_pages); 5706 kernelcore_remaining -= size_pages; 5707 if (!kernelcore_remaining) 5708 break; 5709 } 5710 } 5711 5712 /* 5713 * If there is still required_kernelcore, we do another pass with one 5714 * less node in the count. This will push zone_movable_pfn[nid] further 5715 * along on the nodes that still have memory until kernelcore is 5716 * satisfied 5717 */ 5718 usable_nodes--; 5719 if (usable_nodes && required_kernelcore > usable_nodes) 5720 goto restart; 5721 5722 out2: 5723 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 5724 for (nid = 0; nid < MAX_NUMNODES; nid++) 5725 zone_movable_pfn[nid] = 5726 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 5727 5728 out: 5729 /* restore the node_state */ 5730 node_states[N_MEMORY] = saved_node_state; 5731 } 5732 5733 /* Any regular or high memory on that node ? */ 5734 static void check_for_memory(pg_data_t *pgdat, int nid) 5735 { 5736 enum zone_type zone_type; 5737 5738 if (N_MEMORY == N_NORMAL_MEMORY) 5739 return; 5740 5741 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { 5742 struct zone *zone = &pgdat->node_zones[zone_type]; 5743 if (populated_zone(zone)) { 5744 node_set_state(nid, N_HIGH_MEMORY); 5745 if (N_NORMAL_MEMORY != N_HIGH_MEMORY && 5746 zone_type <= ZONE_NORMAL) 5747 node_set_state(nid, N_NORMAL_MEMORY); 5748 break; 5749 } 5750 } 5751 } 5752 5753 /** 5754 * free_area_init_nodes - Initialise all pg_data_t and zone data 5755 * @max_zone_pfn: an array of max PFNs for each zone 5756 * 5757 * This will call free_area_init_node() for each active node in the system. 5758 * Using the page ranges provided by memblock_set_node(), the size of each 5759 * zone in each node and their holes is calculated. If the maximum PFN 5760 * between two adjacent zones match, it is assumed that the zone is empty. 5761 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 5762 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 5763 * starts where the previous one ended. For example, ZONE_DMA32 starts 5764 * at arch_max_dma_pfn. 5765 */ 5766 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 5767 { 5768 unsigned long start_pfn, end_pfn; 5769 int i, nid; 5770 5771 /* Record where the zone boundaries are */ 5772 memset(arch_zone_lowest_possible_pfn, 0, 5773 sizeof(arch_zone_lowest_possible_pfn)); 5774 memset(arch_zone_highest_possible_pfn, 0, 5775 sizeof(arch_zone_highest_possible_pfn)); 5776 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions(); 5777 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0]; 5778 for (i = 1; i < MAX_NR_ZONES; i++) { 5779 if (i == ZONE_MOVABLE) 5780 continue; 5781 arch_zone_lowest_possible_pfn[i] = 5782 arch_zone_highest_possible_pfn[i-1]; 5783 arch_zone_highest_possible_pfn[i] = 5784 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]); 5785 } 5786 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0; 5787 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0; 5788 5789 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 5790 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 5791 find_zone_movable_pfns_for_nodes(); 5792 5793 /* Print out the zone ranges */ 5794 pr_info("Zone ranges:\n"); 5795 for (i = 0; i < MAX_NR_ZONES; i++) { 5796 if (i == ZONE_MOVABLE) 5797 continue; 5798 pr_info(" %-8s ", zone_names[i]); 5799 if (arch_zone_lowest_possible_pfn[i] == 5800 arch_zone_highest_possible_pfn[i]) 5801 pr_cont("empty\n"); 5802 else 5803 pr_cont("[mem %#018Lx-%#018Lx]\n", 5804 (u64)arch_zone_lowest_possible_pfn[i] 5805 << PAGE_SHIFT, 5806 ((u64)arch_zone_highest_possible_pfn[i] 5807 << PAGE_SHIFT) - 1); 5808 } 5809 5810 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 5811 pr_info("Movable zone start for each node\n"); 5812 for (i = 0; i < MAX_NUMNODES; i++) { 5813 if (zone_movable_pfn[i]) 5814 pr_info(" Node %d: %#018Lx\n", i, 5815 (u64)zone_movable_pfn[i] << PAGE_SHIFT); 5816 } 5817 5818 /* Print out the early node map */ 5819 pr_info("Early memory node ranges\n"); 5820 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 5821 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid, 5822 (u64)start_pfn << PAGE_SHIFT, 5823 ((u64)end_pfn << PAGE_SHIFT) - 1); 5824 5825 /* Initialise every node */ 5826 mminit_verify_pageflags_layout(); 5827 setup_nr_node_ids(); 5828 for_each_online_node(nid) { 5829 pg_data_t *pgdat = NODE_DATA(nid); 5830 free_area_init_node(nid, NULL, 5831 find_min_pfn_for_node(nid), NULL); 5832 5833 /* Any memory on that node */ 5834 if (pgdat->node_present_pages) 5835 node_set_state(nid, N_MEMORY); 5836 check_for_memory(pgdat, nid); 5837 } 5838 } 5839 5840 static int __init cmdline_parse_core(char *p, unsigned long *core) 5841 { 5842 unsigned long long coremem; 5843 if (!p) 5844 return -EINVAL; 5845 5846 coremem = memparse(p, &p); 5847 *core = coremem >> PAGE_SHIFT; 5848 5849 /* Paranoid check that UL is enough for the coremem value */ 5850 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 5851 5852 return 0; 5853 } 5854 5855 /* 5856 * kernelcore=size sets the amount of memory for use for allocations that 5857 * cannot be reclaimed or migrated. 5858 */ 5859 static int __init cmdline_parse_kernelcore(char *p) 5860 { 5861 return cmdline_parse_core(p, &required_kernelcore); 5862 } 5863 5864 /* 5865 * movablecore=size sets the amount of memory for use for allocations that 5866 * can be reclaimed or migrated. 5867 */ 5868 static int __init cmdline_parse_movablecore(char *p) 5869 { 5870 return cmdline_parse_core(p, &required_movablecore); 5871 } 5872 5873 early_param("kernelcore", cmdline_parse_kernelcore); 5874 early_param("movablecore", cmdline_parse_movablecore); 5875 5876 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5877 5878 void adjust_managed_page_count(struct page *page, long count) 5879 { 5880 spin_lock(&managed_page_count_lock); 5881 page_zone(page)->managed_pages += count; 5882 totalram_pages += count; 5883 #ifdef CONFIG_HIGHMEM 5884 if (PageHighMem(page)) 5885 totalhigh_pages += count; 5886 #endif 5887 spin_unlock(&managed_page_count_lock); 5888 } 5889 EXPORT_SYMBOL(adjust_managed_page_count); 5890 5891 unsigned long free_reserved_area(void *start, void *end, int poison, char *s) 5892 { 5893 void *pos; 5894 unsigned long pages = 0; 5895 5896 start = (void *)PAGE_ALIGN((unsigned long)start); 5897 end = (void *)((unsigned long)end & PAGE_MASK); 5898 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { 5899 if ((unsigned int)poison <= 0xFF) 5900 memset(pos, poison, PAGE_SIZE); 5901 free_reserved_page(virt_to_page(pos)); 5902 } 5903 5904 if (pages && s) 5905 pr_info("Freeing %s memory: %ldK (%p - %p)\n", 5906 s, pages << (PAGE_SHIFT - 10), start, end); 5907 5908 return pages; 5909 } 5910 EXPORT_SYMBOL(free_reserved_area); 5911 5912 #ifdef CONFIG_HIGHMEM 5913 void free_highmem_page(struct page *page) 5914 { 5915 __free_reserved_page(page); 5916 totalram_pages++; 5917 page_zone(page)->managed_pages++; 5918 totalhigh_pages++; 5919 } 5920 #endif 5921 5922 5923 void __init mem_init_print_info(const char *str) 5924 { 5925 unsigned long physpages, codesize, datasize, rosize, bss_size; 5926 unsigned long init_code_size, init_data_size; 5927 5928 physpages = get_num_physpages(); 5929 codesize = _etext - _stext; 5930 datasize = _edata - _sdata; 5931 rosize = __end_rodata - __start_rodata; 5932 bss_size = __bss_stop - __bss_start; 5933 init_data_size = __init_end - __init_begin; 5934 init_code_size = _einittext - _sinittext; 5935 5936 /* 5937 * Detect special cases and adjust section sizes accordingly: 5938 * 1) .init.* may be embedded into .data sections 5939 * 2) .init.text.* may be out of [__init_begin, __init_end], 5940 * please refer to arch/tile/kernel/vmlinux.lds.S. 5941 * 3) .rodata.* may be embedded into .text or .data sections. 5942 */ 5943 #define adj_init_size(start, end, size, pos, adj) \ 5944 do { \ 5945 if (start <= pos && pos < end && size > adj) \ 5946 size -= adj; \ 5947 } while (0) 5948 5949 adj_init_size(__init_begin, __init_end, init_data_size, 5950 _sinittext, init_code_size); 5951 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); 5952 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); 5953 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); 5954 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); 5955 5956 #undef adj_init_size 5957 5958 pr_info("Memory: %luK/%luK available " 5959 "(%luK kernel code, %luK rwdata, %luK rodata, " 5960 "%luK init, %luK bss, %luK reserved, %luK cma-reserved" 5961 #ifdef CONFIG_HIGHMEM 5962 ", %luK highmem" 5963 #endif 5964 "%s%s)\n", 5965 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10), 5966 codesize >> 10, datasize >> 10, rosize >> 10, 5967 (init_data_size + init_code_size) >> 10, bss_size >> 10, 5968 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10), 5969 totalcma_pages << (PAGE_SHIFT-10), 5970 #ifdef CONFIG_HIGHMEM 5971 totalhigh_pages << (PAGE_SHIFT-10), 5972 #endif 5973 str ? ", " : "", str ? str : ""); 5974 } 5975 5976 /** 5977 * set_dma_reserve - set the specified number of pages reserved in the first zone 5978 * @new_dma_reserve: The number of pages to mark reserved 5979 * 5980 * The per-cpu batchsize and zone watermarks are determined by present_pages. 5981 * In the DMA zone, a significant percentage may be consumed by kernel image 5982 * and other unfreeable allocations which can skew the watermarks badly. This 5983 * function may optionally be used to account for unfreeable pages in the 5984 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 5985 * smaller per-cpu batchsize. 5986 */ 5987 void __init set_dma_reserve(unsigned long new_dma_reserve) 5988 { 5989 dma_reserve = new_dma_reserve; 5990 } 5991 5992 void __init free_area_init(unsigned long *zones_size) 5993 { 5994 free_area_init_node(0, zones_size, 5995 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 5996 } 5997 5998 static int page_alloc_cpu_notify(struct notifier_block *self, 5999 unsigned long action, void *hcpu) 6000 { 6001 int cpu = (unsigned long)hcpu; 6002 6003 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) { 6004 lru_add_drain_cpu(cpu); 6005 drain_pages(cpu); 6006 6007 /* 6008 * Spill the event counters of the dead processor 6009 * into the current processors event counters. 6010 * This artificially elevates the count of the current 6011 * processor. 6012 */ 6013 vm_events_fold_cpu(cpu); 6014 6015 /* 6016 * Zero the differential counters of the dead processor 6017 * so that the vm statistics are consistent. 6018 * 6019 * This is only okay since the processor is dead and cannot 6020 * race with what we are doing. 6021 */ 6022 cpu_vm_stats_fold(cpu); 6023 } 6024 return NOTIFY_OK; 6025 } 6026 6027 void __init page_alloc_init(void) 6028 { 6029 hotcpu_notifier(page_alloc_cpu_notify, 0); 6030 } 6031 6032 /* 6033 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio 6034 * or min_free_kbytes changes. 6035 */ 6036 static void calculate_totalreserve_pages(void) 6037 { 6038 struct pglist_data *pgdat; 6039 unsigned long reserve_pages = 0; 6040 enum zone_type i, j; 6041 6042 for_each_online_pgdat(pgdat) { 6043 for (i = 0; i < MAX_NR_ZONES; i++) { 6044 struct zone *zone = pgdat->node_zones + i; 6045 long max = 0; 6046 6047 /* Find valid and maximum lowmem_reserve in the zone */ 6048 for (j = i; j < MAX_NR_ZONES; j++) { 6049 if (zone->lowmem_reserve[j] > max) 6050 max = zone->lowmem_reserve[j]; 6051 } 6052 6053 /* we treat the high watermark as reserved pages. */ 6054 max += high_wmark_pages(zone); 6055 6056 if (max > zone->managed_pages) 6057 max = zone->managed_pages; 6058 reserve_pages += max; 6059 /* 6060 * Lowmem reserves are not available to 6061 * GFP_HIGHUSER page cache allocations and 6062 * kswapd tries to balance zones to their high 6063 * watermark. As a result, neither should be 6064 * regarded as dirtyable memory, to prevent a 6065 * situation where reclaim has to clean pages 6066 * in order to balance the zones. 6067 */ 6068 zone->dirty_balance_reserve = max; 6069 } 6070 } 6071 dirty_balance_reserve = reserve_pages; 6072 totalreserve_pages = reserve_pages; 6073 } 6074 6075 /* 6076 * setup_per_zone_lowmem_reserve - called whenever 6077 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone 6078 * has a correct pages reserved value, so an adequate number of 6079 * pages are left in the zone after a successful __alloc_pages(). 6080 */ 6081 static void setup_per_zone_lowmem_reserve(void) 6082 { 6083 struct pglist_data *pgdat; 6084 enum zone_type j, idx; 6085 6086 for_each_online_pgdat(pgdat) { 6087 for (j = 0; j < MAX_NR_ZONES; j++) { 6088 struct zone *zone = pgdat->node_zones + j; 6089 unsigned long managed_pages = zone->managed_pages; 6090 6091 zone->lowmem_reserve[j] = 0; 6092 6093 idx = j; 6094 while (idx) { 6095 struct zone *lower_zone; 6096 6097 idx--; 6098 6099 if (sysctl_lowmem_reserve_ratio[idx] < 1) 6100 sysctl_lowmem_reserve_ratio[idx] = 1; 6101 6102 lower_zone = pgdat->node_zones + idx; 6103 lower_zone->lowmem_reserve[j] = managed_pages / 6104 sysctl_lowmem_reserve_ratio[idx]; 6105 managed_pages += lower_zone->managed_pages; 6106 } 6107 } 6108 } 6109 6110 /* update totalreserve_pages */ 6111 calculate_totalreserve_pages(); 6112 } 6113 6114 static void __setup_per_zone_wmarks(void) 6115 { 6116 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 6117 unsigned long lowmem_pages = 0; 6118 struct zone *zone; 6119 unsigned long flags; 6120 6121 /* Calculate total number of !ZONE_HIGHMEM pages */ 6122 for_each_zone(zone) { 6123 if (!is_highmem(zone)) 6124 lowmem_pages += zone->managed_pages; 6125 } 6126 6127 for_each_zone(zone) { 6128 u64 tmp; 6129 6130 spin_lock_irqsave(&zone->lock, flags); 6131 tmp = (u64)pages_min * zone->managed_pages; 6132 do_div(tmp, lowmem_pages); 6133 if (is_highmem(zone)) { 6134 /* 6135 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 6136 * need highmem pages, so cap pages_min to a small 6137 * value here. 6138 * 6139 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 6140 * deltas control asynch page reclaim, and so should 6141 * not be capped for highmem. 6142 */ 6143 unsigned long min_pages; 6144 6145 min_pages = zone->managed_pages / 1024; 6146 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); 6147 zone->watermark[WMARK_MIN] = min_pages; 6148 } else { 6149 /* 6150 * If it's a lowmem zone, reserve a number of pages 6151 * proportionate to the zone's size. 6152 */ 6153 zone->watermark[WMARK_MIN] = tmp; 6154 } 6155 6156 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2); 6157 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1); 6158 6159 __mod_zone_page_state(zone, NR_ALLOC_BATCH, 6160 high_wmark_pages(zone) - low_wmark_pages(zone) - 6161 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH])); 6162 6163 setup_zone_migrate_reserve(zone); 6164 spin_unlock_irqrestore(&zone->lock, flags); 6165 } 6166 6167 /* update totalreserve_pages */ 6168 calculate_totalreserve_pages(); 6169 } 6170 6171 /** 6172 * setup_per_zone_wmarks - called when min_free_kbytes changes 6173 * or when memory is hot-{added|removed} 6174 * 6175 * Ensures that the watermark[min,low,high] values for each zone are set 6176 * correctly with respect to min_free_kbytes. 6177 */ 6178 void setup_per_zone_wmarks(void) 6179 { 6180 mutex_lock(&zonelists_mutex); 6181 __setup_per_zone_wmarks(); 6182 mutex_unlock(&zonelists_mutex); 6183 } 6184 6185 /* 6186 * The inactive anon list should be small enough that the VM never has to 6187 * do too much work, but large enough that each inactive page has a chance 6188 * to be referenced again before it is swapped out. 6189 * 6190 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to 6191 * INACTIVE_ANON pages on this zone's LRU, maintained by the 6192 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of 6193 * the anonymous pages are kept on the inactive list. 6194 * 6195 * total target max 6196 * memory ratio inactive anon 6197 * ------------------------------------- 6198 * 10MB 1 5MB 6199 * 100MB 1 50MB 6200 * 1GB 3 250MB 6201 * 10GB 10 0.9GB 6202 * 100GB 31 3GB 6203 * 1TB 101 10GB 6204 * 10TB 320 32GB 6205 */ 6206 static void __meminit calculate_zone_inactive_ratio(struct zone *zone) 6207 { 6208 unsigned int gb, ratio; 6209 6210 /* Zone size in gigabytes */ 6211 gb = zone->managed_pages >> (30 - PAGE_SHIFT); 6212 if (gb) 6213 ratio = int_sqrt(10 * gb); 6214 else 6215 ratio = 1; 6216 6217 zone->inactive_ratio = ratio; 6218 } 6219 6220 static void __meminit setup_per_zone_inactive_ratio(void) 6221 { 6222 struct zone *zone; 6223 6224 for_each_zone(zone) 6225 calculate_zone_inactive_ratio(zone); 6226 } 6227 6228 /* 6229 * Initialise min_free_kbytes. 6230 * 6231 * For small machines we want it small (128k min). For large machines 6232 * we want it large (64MB max). But it is not linear, because network 6233 * bandwidth does not increase linearly with machine size. We use 6234 * 6235 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 6236 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 6237 * 6238 * which yields 6239 * 6240 * 16MB: 512k 6241 * 32MB: 724k 6242 * 64MB: 1024k 6243 * 128MB: 1448k 6244 * 256MB: 2048k 6245 * 512MB: 2896k 6246 * 1024MB: 4096k 6247 * 2048MB: 5792k 6248 * 4096MB: 8192k 6249 * 8192MB: 11584k 6250 * 16384MB: 16384k 6251 */ 6252 int __meminit init_per_zone_wmark_min(void) 6253 { 6254 unsigned long lowmem_kbytes; 6255 int new_min_free_kbytes; 6256 6257 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 6258 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 6259 6260 if (new_min_free_kbytes > user_min_free_kbytes) { 6261 min_free_kbytes = new_min_free_kbytes; 6262 if (min_free_kbytes < 128) 6263 min_free_kbytes = 128; 6264 if (min_free_kbytes > 65536) 6265 min_free_kbytes = 65536; 6266 } else { 6267 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", 6268 new_min_free_kbytes, user_min_free_kbytes); 6269 } 6270 setup_per_zone_wmarks(); 6271 refresh_zone_stat_thresholds(); 6272 setup_per_zone_lowmem_reserve(); 6273 setup_per_zone_inactive_ratio(); 6274 return 0; 6275 } 6276 module_init(init_per_zone_wmark_min) 6277 6278 /* 6279 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 6280 * that we can call two helper functions whenever min_free_kbytes 6281 * changes. 6282 */ 6283 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, 6284 void __user *buffer, size_t *length, loff_t *ppos) 6285 { 6286 int rc; 6287 6288 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6289 if (rc) 6290 return rc; 6291 6292 if (write) { 6293 user_min_free_kbytes = min_free_kbytes; 6294 setup_per_zone_wmarks(); 6295 } 6296 return 0; 6297 } 6298 6299 #ifdef CONFIG_NUMA 6300 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, 6301 void __user *buffer, size_t *length, loff_t *ppos) 6302 { 6303 struct zone *zone; 6304 int rc; 6305 6306 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6307 if (rc) 6308 return rc; 6309 6310 for_each_zone(zone) 6311 zone->min_unmapped_pages = (zone->managed_pages * 6312 sysctl_min_unmapped_ratio) / 100; 6313 return 0; 6314 } 6315 6316 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, 6317 void __user *buffer, size_t *length, loff_t *ppos) 6318 { 6319 struct zone *zone; 6320 int rc; 6321 6322 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6323 if (rc) 6324 return rc; 6325 6326 for_each_zone(zone) 6327 zone->min_slab_pages = (zone->managed_pages * 6328 sysctl_min_slab_ratio) / 100; 6329 return 0; 6330 } 6331 #endif 6332 6333 /* 6334 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 6335 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 6336 * whenever sysctl_lowmem_reserve_ratio changes. 6337 * 6338 * The reserve ratio obviously has absolutely no relation with the 6339 * minimum watermarks. The lowmem reserve ratio can only make sense 6340 * if in function of the boot time zone sizes. 6341 */ 6342 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write, 6343 void __user *buffer, size_t *length, loff_t *ppos) 6344 { 6345 proc_dointvec_minmax(table, write, buffer, length, ppos); 6346 setup_per_zone_lowmem_reserve(); 6347 return 0; 6348 } 6349 6350 /* 6351 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 6352 * cpu. It is the fraction of total pages in each zone that a hot per cpu 6353 * pagelist can have before it gets flushed back to buddy allocator. 6354 */ 6355 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, 6356 void __user *buffer, size_t *length, loff_t *ppos) 6357 { 6358 struct zone *zone; 6359 int old_percpu_pagelist_fraction; 6360 int ret; 6361 6362 mutex_lock(&pcp_batch_high_lock); 6363 old_percpu_pagelist_fraction = percpu_pagelist_fraction; 6364 6365 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 6366 if (!write || ret < 0) 6367 goto out; 6368 6369 /* Sanity checking to avoid pcp imbalance */ 6370 if (percpu_pagelist_fraction && 6371 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) { 6372 percpu_pagelist_fraction = old_percpu_pagelist_fraction; 6373 ret = -EINVAL; 6374 goto out; 6375 } 6376 6377 /* No change? */ 6378 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction) 6379 goto out; 6380 6381 for_each_populated_zone(zone) { 6382 unsigned int cpu; 6383 6384 for_each_possible_cpu(cpu) 6385 pageset_set_high_and_batch(zone, 6386 per_cpu_ptr(zone->pageset, cpu)); 6387 } 6388 out: 6389 mutex_unlock(&pcp_batch_high_lock); 6390 return ret; 6391 } 6392 6393 #ifdef CONFIG_NUMA 6394 int hashdist = HASHDIST_DEFAULT; 6395 6396 static int __init set_hashdist(char *str) 6397 { 6398 if (!str) 6399 return 0; 6400 hashdist = simple_strtoul(str, &str, 0); 6401 return 1; 6402 } 6403 __setup("hashdist=", set_hashdist); 6404 #endif 6405 6406 /* 6407 * allocate a large system hash table from bootmem 6408 * - it is assumed that the hash table must contain an exact power-of-2 6409 * quantity of entries 6410 * - limit is the number of hash buckets, not the total allocation size 6411 */ 6412 void *__init alloc_large_system_hash(const char *tablename, 6413 unsigned long bucketsize, 6414 unsigned long numentries, 6415 int scale, 6416 int flags, 6417 unsigned int *_hash_shift, 6418 unsigned int *_hash_mask, 6419 unsigned long low_limit, 6420 unsigned long high_limit) 6421 { 6422 unsigned long long max = high_limit; 6423 unsigned long log2qty, size; 6424 void *table = NULL; 6425 6426 /* allow the kernel cmdline to have a say */ 6427 if (!numentries) { 6428 /* round applicable memory size up to nearest megabyte */ 6429 numentries = nr_kernel_pages; 6430 6431 /* It isn't necessary when PAGE_SIZE >= 1MB */ 6432 if (PAGE_SHIFT < 20) 6433 numentries = round_up(numentries, (1<<20)/PAGE_SIZE); 6434 6435 /* limit to 1 bucket per 2^scale bytes of low memory */ 6436 if (scale > PAGE_SHIFT) 6437 numentries >>= (scale - PAGE_SHIFT); 6438 else 6439 numentries <<= (PAGE_SHIFT - scale); 6440 6441 /* Make sure we've got at least a 0-order allocation.. */ 6442 if (unlikely(flags & HASH_SMALL)) { 6443 /* Makes no sense without HASH_EARLY */ 6444 WARN_ON(!(flags & HASH_EARLY)); 6445 if (!(numentries >> *_hash_shift)) { 6446 numentries = 1UL << *_hash_shift; 6447 BUG_ON(!numentries); 6448 } 6449 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 6450 numentries = PAGE_SIZE / bucketsize; 6451 } 6452 numentries = roundup_pow_of_two(numentries); 6453 6454 /* limit allocation size to 1/16 total memory by default */ 6455 if (max == 0) { 6456 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 6457 do_div(max, bucketsize); 6458 } 6459 max = min(max, 0x80000000ULL); 6460 6461 if (numentries < low_limit) 6462 numentries = low_limit; 6463 if (numentries > max) 6464 numentries = max; 6465 6466 log2qty = ilog2(numentries); 6467 6468 do { 6469 size = bucketsize << log2qty; 6470 if (flags & HASH_EARLY) 6471 table = memblock_virt_alloc_nopanic(size, 0); 6472 else if (hashdist) 6473 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 6474 else { 6475 /* 6476 * If bucketsize is not a power-of-two, we may free 6477 * some pages at the end of hash table which 6478 * alloc_pages_exact() automatically does 6479 */ 6480 if (get_order(size) < MAX_ORDER) { 6481 table = alloc_pages_exact(size, GFP_ATOMIC); 6482 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 6483 } 6484 } 6485 } while (!table && size > PAGE_SIZE && --log2qty); 6486 6487 if (!table) 6488 panic("Failed to allocate %s hash table\n", tablename); 6489 6490 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n", 6491 tablename, 6492 (1UL << log2qty), 6493 ilog2(size) - PAGE_SHIFT, 6494 size); 6495 6496 if (_hash_shift) 6497 *_hash_shift = log2qty; 6498 if (_hash_mask) 6499 *_hash_mask = (1 << log2qty) - 1; 6500 6501 return table; 6502 } 6503 6504 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 6505 static inline unsigned long *get_pageblock_bitmap(struct zone *zone, 6506 unsigned long pfn) 6507 { 6508 #ifdef CONFIG_SPARSEMEM 6509 return __pfn_to_section(pfn)->pageblock_flags; 6510 #else 6511 return zone->pageblock_flags; 6512 #endif /* CONFIG_SPARSEMEM */ 6513 } 6514 6515 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn) 6516 { 6517 #ifdef CONFIG_SPARSEMEM 6518 pfn &= (PAGES_PER_SECTION-1); 6519 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6520 #else 6521 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages); 6522 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 6523 #endif /* CONFIG_SPARSEMEM */ 6524 } 6525 6526 /** 6527 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages 6528 * @page: The page within the block of interest 6529 * @pfn: The target page frame number 6530 * @end_bitidx: The last bit of interest to retrieve 6531 * @mask: mask of bits that the caller is interested in 6532 * 6533 * Return: pageblock_bits flags 6534 */ 6535 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, 6536 unsigned long end_bitidx, 6537 unsigned long mask) 6538 { 6539 struct zone *zone; 6540 unsigned long *bitmap; 6541 unsigned long bitidx, word_bitidx; 6542 unsigned long word; 6543 6544 zone = page_zone(page); 6545 bitmap = get_pageblock_bitmap(zone, pfn); 6546 bitidx = pfn_to_bitidx(zone, pfn); 6547 word_bitidx = bitidx / BITS_PER_LONG; 6548 bitidx &= (BITS_PER_LONG-1); 6549 6550 word = bitmap[word_bitidx]; 6551 bitidx += end_bitidx; 6552 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; 6553 } 6554 6555 /** 6556 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages 6557 * @page: The page within the block of interest 6558 * @flags: The flags to set 6559 * @pfn: The target page frame number 6560 * @end_bitidx: The last bit of interest 6561 * @mask: mask of bits that the caller is interested in 6562 */ 6563 void set_pfnblock_flags_mask(struct page *page, unsigned long flags, 6564 unsigned long pfn, 6565 unsigned long end_bitidx, 6566 unsigned long mask) 6567 { 6568 struct zone *zone; 6569 unsigned long *bitmap; 6570 unsigned long bitidx, word_bitidx; 6571 unsigned long old_word, word; 6572 6573 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); 6574 6575 zone = page_zone(page); 6576 bitmap = get_pageblock_bitmap(zone, pfn); 6577 bitidx = pfn_to_bitidx(zone, pfn); 6578 word_bitidx = bitidx / BITS_PER_LONG; 6579 bitidx &= (BITS_PER_LONG-1); 6580 6581 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page); 6582 6583 bitidx += end_bitidx; 6584 mask <<= (BITS_PER_LONG - bitidx - 1); 6585 flags <<= (BITS_PER_LONG - bitidx - 1); 6586 6587 word = READ_ONCE(bitmap[word_bitidx]); 6588 for (;;) { 6589 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); 6590 if (word == old_word) 6591 break; 6592 word = old_word; 6593 } 6594 } 6595 6596 /* 6597 * This function checks whether pageblock includes unmovable pages or not. 6598 * If @count is not zero, it is okay to include less @count unmovable pages 6599 * 6600 * PageLRU check without isolation or lru_lock could race so that 6601 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't 6602 * expect this function should be exact. 6603 */ 6604 bool has_unmovable_pages(struct zone *zone, struct page *page, int count, 6605 bool skip_hwpoisoned_pages) 6606 { 6607 unsigned long pfn, iter, found; 6608 int mt; 6609 6610 /* 6611 * For avoiding noise data, lru_add_drain_all() should be called 6612 * If ZONE_MOVABLE, the zone never contains unmovable pages 6613 */ 6614 if (zone_idx(zone) == ZONE_MOVABLE) 6615 return false; 6616 mt = get_pageblock_migratetype(page); 6617 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) 6618 return false; 6619 6620 pfn = page_to_pfn(page); 6621 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 6622 unsigned long check = pfn + iter; 6623 6624 if (!pfn_valid_within(check)) 6625 continue; 6626 6627 page = pfn_to_page(check); 6628 6629 /* 6630 * Hugepages are not in LRU lists, but they're movable. 6631 * We need not scan over tail pages bacause we don't 6632 * handle each tail page individually in migration. 6633 */ 6634 if (PageHuge(page)) { 6635 iter = round_up(iter + 1, 1<<compound_order(page)) - 1; 6636 continue; 6637 } 6638 6639 /* 6640 * We can't use page_count without pin a page 6641 * because another CPU can free compound page. 6642 * This check already skips compound tails of THP 6643 * because their page->_count is zero at all time. 6644 */ 6645 if (!atomic_read(&page->_count)) { 6646 if (PageBuddy(page)) 6647 iter += (1 << page_order(page)) - 1; 6648 continue; 6649 } 6650 6651 /* 6652 * The HWPoisoned page may be not in buddy system, and 6653 * page_count() is not 0. 6654 */ 6655 if (skip_hwpoisoned_pages && PageHWPoison(page)) 6656 continue; 6657 6658 if (!PageLRU(page)) 6659 found++; 6660 /* 6661 * If there are RECLAIMABLE pages, we need to check 6662 * it. But now, memory offline itself doesn't call 6663 * shrink_node_slabs() and it still to be fixed. 6664 */ 6665 /* 6666 * If the page is not RAM, page_count()should be 0. 6667 * we don't need more check. This is an _used_ not-movable page. 6668 * 6669 * The problematic thing here is PG_reserved pages. PG_reserved 6670 * is set to both of a memory hole page and a _used_ kernel 6671 * page at boot. 6672 */ 6673 if (found > count) 6674 return true; 6675 } 6676 return false; 6677 } 6678 6679 bool is_pageblock_removable_nolock(struct page *page) 6680 { 6681 struct zone *zone; 6682 unsigned long pfn; 6683 6684 /* 6685 * We have to be careful here because we are iterating over memory 6686 * sections which are not zone aware so we might end up outside of 6687 * the zone but still within the section. 6688 * We have to take care about the node as well. If the node is offline 6689 * its NODE_DATA will be NULL - see page_zone. 6690 */ 6691 if (!node_online(page_to_nid(page))) 6692 return false; 6693 6694 zone = page_zone(page); 6695 pfn = page_to_pfn(page); 6696 if (!zone_spans_pfn(zone, pfn)) 6697 return false; 6698 6699 return !has_unmovable_pages(zone, page, 0, true); 6700 } 6701 6702 #ifdef CONFIG_CMA 6703 6704 static unsigned long pfn_max_align_down(unsigned long pfn) 6705 { 6706 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, 6707 pageblock_nr_pages) - 1); 6708 } 6709 6710 static unsigned long pfn_max_align_up(unsigned long pfn) 6711 { 6712 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, 6713 pageblock_nr_pages)); 6714 } 6715 6716 /* [start, end) must belong to a single zone. */ 6717 static int __alloc_contig_migrate_range(struct compact_control *cc, 6718 unsigned long start, unsigned long end) 6719 { 6720 /* This function is based on compact_zone() from compaction.c. */ 6721 unsigned long nr_reclaimed; 6722 unsigned long pfn = start; 6723 unsigned int tries = 0; 6724 int ret = 0; 6725 6726 migrate_prep(); 6727 6728 while (pfn < end || !list_empty(&cc->migratepages)) { 6729 if (fatal_signal_pending(current)) { 6730 ret = -EINTR; 6731 break; 6732 } 6733 6734 if (list_empty(&cc->migratepages)) { 6735 cc->nr_migratepages = 0; 6736 pfn = isolate_migratepages_range(cc, pfn, end); 6737 if (!pfn) { 6738 ret = -EINTR; 6739 break; 6740 } 6741 tries = 0; 6742 } else if (++tries == 5) { 6743 ret = ret < 0 ? ret : -EBUSY; 6744 break; 6745 } 6746 6747 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, 6748 &cc->migratepages); 6749 cc->nr_migratepages -= nr_reclaimed; 6750 6751 ret = migrate_pages(&cc->migratepages, alloc_migrate_target, 6752 NULL, 0, cc->mode, MR_CMA); 6753 } 6754 if (ret < 0) { 6755 putback_movable_pages(&cc->migratepages); 6756 return ret; 6757 } 6758 return 0; 6759 } 6760 6761 /** 6762 * alloc_contig_range() -- tries to allocate given range of pages 6763 * @start: start PFN to allocate 6764 * @end: one-past-the-last PFN to allocate 6765 * @migratetype: migratetype of the underlaying pageblocks (either 6766 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks 6767 * in range must have the same migratetype and it must 6768 * be either of the two. 6769 * 6770 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES 6771 * aligned, however it's the caller's responsibility to guarantee that 6772 * we are the only thread that changes migrate type of pageblocks the 6773 * pages fall in. 6774 * 6775 * The PFN range must belong to a single zone. 6776 * 6777 * Returns zero on success or negative error code. On success all 6778 * pages which PFN is in [start, end) are allocated for the caller and 6779 * need to be freed with free_contig_range(). 6780 */ 6781 int alloc_contig_range(unsigned long start, unsigned long end, 6782 unsigned migratetype) 6783 { 6784 unsigned long outer_start, outer_end; 6785 int ret = 0, order; 6786 6787 struct compact_control cc = { 6788 .nr_migratepages = 0, 6789 .order = -1, 6790 .zone = page_zone(pfn_to_page(start)), 6791 .mode = MIGRATE_SYNC, 6792 .ignore_skip_hint = true, 6793 }; 6794 INIT_LIST_HEAD(&cc.migratepages); 6795 6796 /* 6797 * What we do here is we mark all pageblocks in range as 6798 * MIGRATE_ISOLATE. Because pageblock and max order pages may 6799 * have different sizes, and due to the way page allocator 6800 * work, we align the range to biggest of the two pages so 6801 * that page allocator won't try to merge buddies from 6802 * different pageblocks and change MIGRATE_ISOLATE to some 6803 * other migration type. 6804 * 6805 * Once the pageblocks are marked as MIGRATE_ISOLATE, we 6806 * migrate the pages from an unaligned range (ie. pages that 6807 * we are interested in). This will put all the pages in 6808 * range back to page allocator as MIGRATE_ISOLATE. 6809 * 6810 * When this is done, we take the pages in range from page 6811 * allocator removing them from the buddy system. This way 6812 * page allocator will never consider using them. 6813 * 6814 * This lets us mark the pageblocks back as 6815 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the 6816 * aligned range but not in the unaligned, original range are 6817 * put back to page allocator so that buddy can use them. 6818 */ 6819 6820 ret = start_isolate_page_range(pfn_max_align_down(start), 6821 pfn_max_align_up(end), migratetype, 6822 false); 6823 if (ret) 6824 return ret; 6825 6826 ret = __alloc_contig_migrate_range(&cc, start, end); 6827 if (ret) 6828 goto done; 6829 6830 /* 6831 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES 6832 * aligned blocks that are marked as MIGRATE_ISOLATE. What's 6833 * more, all pages in [start, end) are free in page allocator. 6834 * What we are going to do is to allocate all pages from 6835 * [start, end) (that is remove them from page allocator). 6836 * 6837 * The only problem is that pages at the beginning and at the 6838 * end of interesting range may be not aligned with pages that 6839 * page allocator holds, ie. they can be part of higher order 6840 * pages. Because of this, we reserve the bigger range and 6841 * once this is done free the pages we are not interested in. 6842 * 6843 * We don't have to hold zone->lock here because the pages are 6844 * isolated thus they won't get removed from buddy. 6845 */ 6846 6847 lru_add_drain_all(); 6848 drain_all_pages(cc.zone); 6849 6850 order = 0; 6851 outer_start = start; 6852 while (!PageBuddy(pfn_to_page(outer_start))) { 6853 if (++order >= MAX_ORDER) { 6854 ret = -EBUSY; 6855 goto done; 6856 } 6857 outer_start &= ~0UL << order; 6858 } 6859 6860 /* Make sure the range is really isolated. */ 6861 if (test_pages_isolated(outer_start, end, false)) { 6862 pr_info("%s: [%lx, %lx) PFNs busy\n", 6863 __func__, outer_start, end); 6864 ret = -EBUSY; 6865 goto done; 6866 } 6867 6868 /* Grab isolated pages from freelists. */ 6869 outer_end = isolate_freepages_range(&cc, outer_start, end); 6870 if (!outer_end) { 6871 ret = -EBUSY; 6872 goto done; 6873 } 6874 6875 /* Free head and tail (if any) */ 6876 if (start != outer_start) 6877 free_contig_range(outer_start, start - outer_start); 6878 if (end != outer_end) 6879 free_contig_range(end, outer_end - end); 6880 6881 done: 6882 undo_isolate_page_range(pfn_max_align_down(start), 6883 pfn_max_align_up(end), migratetype); 6884 return ret; 6885 } 6886 6887 void free_contig_range(unsigned long pfn, unsigned nr_pages) 6888 { 6889 unsigned int count = 0; 6890 6891 for (; nr_pages--; pfn++) { 6892 struct page *page = pfn_to_page(pfn); 6893 6894 count += page_count(page) != 1; 6895 __free_page(page); 6896 } 6897 WARN(count != 0, "%d pages are still in use!\n", count); 6898 } 6899 #endif 6900 6901 #ifdef CONFIG_MEMORY_HOTPLUG 6902 /* 6903 * The zone indicated has a new number of managed_pages; batch sizes and percpu 6904 * page high values need to be recalulated. 6905 */ 6906 void __meminit zone_pcp_update(struct zone *zone) 6907 { 6908 unsigned cpu; 6909 mutex_lock(&pcp_batch_high_lock); 6910 for_each_possible_cpu(cpu) 6911 pageset_set_high_and_batch(zone, 6912 per_cpu_ptr(zone->pageset, cpu)); 6913 mutex_unlock(&pcp_batch_high_lock); 6914 } 6915 #endif 6916 6917 void zone_pcp_reset(struct zone *zone) 6918 { 6919 unsigned long flags; 6920 int cpu; 6921 struct per_cpu_pageset *pset; 6922 6923 /* avoid races with drain_pages() */ 6924 local_irq_save(flags); 6925 if (zone->pageset != &boot_pageset) { 6926 for_each_online_cpu(cpu) { 6927 pset = per_cpu_ptr(zone->pageset, cpu); 6928 drain_zonestat(zone, pset); 6929 } 6930 free_percpu(zone->pageset); 6931 zone->pageset = &boot_pageset; 6932 } 6933 local_irq_restore(flags); 6934 } 6935 6936 #ifdef CONFIG_MEMORY_HOTREMOVE 6937 /* 6938 * All pages in the range must be isolated before calling this. 6939 */ 6940 void 6941 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 6942 { 6943 struct page *page; 6944 struct zone *zone; 6945 unsigned int order, i; 6946 unsigned long pfn; 6947 unsigned long flags; 6948 /* find the first valid pfn */ 6949 for (pfn = start_pfn; pfn < end_pfn; pfn++) 6950 if (pfn_valid(pfn)) 6951 break; 6952 if (pfn == end_pfn) 6953 return; 6954 zone = page_zone(pfn_to_page(pfn)); 6955 spin_lock_irqsave(&zone->lock, flags); 6956 pfn = start_pfn; 6957 while (pfn < end_pfn) { 6958 if (!pfn_valid(pfn)) { 6959 pfn++; 6960 continue; 6961 } 6962 page = pfn_to_page(pfn); 6963 /* 6964 * The HWPoisoned page may be not in buddy system, and 6965 * page_count() is not 0. 6966 */ 6967 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { 6968 pfn++; 6969 SetPageReserved(page); 6970 continue; 6971 } 6972 6973 BUG_ON(page_count(page)); 6974 BUG_ON(!PageBuddy(page)); 6975 order = page_order(page); 6976 #ifdef CONFIG_DEBUG_VM 6977 printk(KERN_INFO "remove from free list %lx %d %lx\n", 6978 pfn, 1 << order, end_pfn); 6979 #endif 6980 list_del(&page->lru); 6981 rmv_page_order(page); 6982 zone->free_area[order].nr_free--; 6983 for (i = 0; i < (1 << order); i++) 6984 SetPageReserved((page+i)); 6985 pfn += (1 << order); 6986 } 6987 spin_unlock_irqrestore(&zone->lock, flags); 6988 } 6989 #endif 6990 6991 #ifdef CONFIG_MEMORY_FAILURE 6992 bool is_free_buddy_page(struct page *page) 6993 { 6994 struct zone *zone = page_zone(page); 6995 unsigned long pfn = page_to_pfn(page); 6996 unsigned long flags; 6997 unsigned int order; 6998 6999 spin_lock_irqsave(&zone->lock, flags); 7000 for (order = 0; order < MAX_ORDER; order++) { 7001 struct page *page_head = page - (pfn & ((1 << order) - 1)); 7002 7003 if (PageBuddy(page_head) && page_order(page_head) >= order) 7004 break; 7005 } 7006 spin_unlock_irqrestore(&zone->lock, flags); 7007 7008 return order < MAX_ORDER; 7009 } 7010 #endif 7011