1 /* 2 * linux/mm/page_alloc.c 3 * 4 * Manages the free list, the system allocates free pages here. 5 * Note that kmalloc() lives in slab.c 6 * 7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds 8 * Swap reorganised 29.12.95, Stephen Tweedie 9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999 10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999 11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999 12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000 13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002 14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton) 15 */ 16 17 #include <linux/stddef.h> 18 #include <linux/mm.h> 19 #include <linux/swap.h> 20 #include <linux/interrupt.h> 21 #include <linux/pagemap.h> 22 #include <linux/jiffies.h> 23 #include <linux/bootmem.h> 24 #include <linux/memblock.h> 25 #include <linux/compiler.h> 26 #include <linux/kernel.h> 27 #include <linux/kmemcheck.h> 28 #include <linux/kasan.h> 29 #include <linux/module.h> 30 #include <linux/suspend.h> 31 #include <linux/pagevec.h> 32 #include <linux/blkdev.h> 33 #include <linux/slab.h> 34 #include <linux/ratelimit.h> 35 #include <linux/oom.h> 36 #include <linux/notifier.h> 37 #include <linux/topology.h> 38 #include <linux/sysctl.h> 39 #include <linux/cpu.h> 40 #include <linux/cpuset.h> 41 #include <linux/memory_hotplug.h> 42 #include <linux/nodemask.h> 43 #include <linux/vmalloc.h> 44 #include <linux/vmstat.h> 45 #include <linux/mempolicy.h> 46 #include <linux/memremap.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 <trace/events/oom.h> 59 #include <linux/prefetch.h> 60 #include <linux/mm_inline.h> 61 #include <linux/migrate.h> 62 #include <linux/hugetlb.h> 63 #include <linux/sched/rt.h> 64 #include <linux/sched/mm.h> 65 #include <linux/page_owner.h> 66 #include <linux/kthread.h> 67 #include <linux/memcontrol.h> 68 69 #include <asm/sections.h> 70 #include <asm/tlbflush.h> 71 #include <asm/div64.h> 72 #include "internal.h" 73 74 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */ 75 static DEFINE_MUTEX(pcp_batch_high_lock); 76 #define MIN_PERCPU_PAGELIST_FRACTION (8) 77 78 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID 79 DEFINE_PER_CPU(int, numa_node); 80 EXPORT_PER_CPU_SYMBOL(numa_node); 81 #endif 82 83 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 84 /* 85 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly. 86 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined. 87 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem() 88 * defined in <linux/topology.h>. 89 */ 90 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */ 91 EXPORT_PER_CPU_SYMBOL(_numa_mem_); 92 int _node_numa_mem_[MAX_NUMNODES]; 93 #endif 94 95 /* work_structs for global per-cpu drains */ 96 DEFINE_MUTEX(pcpu_drain_mutex); 97 DEFINE_PER_CPU(struct work_struct, pcpu_drain); 98 99 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY 100 volatile unsigned long latent_entropy __latent_entropy; 101 EXPORT_SYMBOL(latent_entropy); 102 #endif 103 104 /* 105 * Array of node states. 106 */ 107 nodemask_t node_states[NR_NODE_STATES] __read_mostly = { 108 [N_POSSIBLE] = NODE_MASK_ALL, 109 [N_ONLINE] = { { [0] = 1UL } }, 110 #ifndef CONFIG_NUMA 111 [N_NORMAL_MEMORY] = { { [0] = 1UL } }, 112 #ifdef CONFIG_HIGHMEM 113 [N_HIGH_MEMORY] = { { [0] = 1UL } }, 114 #endif 115 #ifdef CONFIG_MOVABLE_NODE 116 [N_MEMORY] = { { [0] = 1UL } }, 117 #endif 118 [N_CPU] = { { [0] = 1UL } }, 119 #endif /* NUMA */ 120 }; 121 EXPORT_SYMBOL(node_states); 122 123 /* Protect totalram_pages and zone->managed_pages */ 124 static DEFINE_SPINLOCK(managed_page_count_lock); 125 126 unsigned long totalram_pages __read_mostly; 127 unsigned long totalreserve_pages __read_mostly; 128 unsigned long totalcma_pages __read_mostly; 129 130 int percpu_pagelist_fraction; 131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK; 132 133 /* 134 * A cached value of the page's pageblock's migratetype, used when the page is 135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when 136 * freeing from pcplists in most cases, at the cost of possibly becoming stale. 137 * Also the migratetype set in the page does not necessarily match the pcplist 138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any 139 * other index - this ensures that it will be put on the correct CMA freelist. 140 */ 141 static inline int get_pcppage_migratetype(struct page *page) 142 { 143 return page->index; 144 } 145 146 static inline void set_pcppage_migratetype(struct page *page, int migratetype) 147 { 148 page->index = migratetype; 149 } 150 151 #ifdef CONFIG_PM_SLEEP 152 /* 153 * The following functions are used by the suspend/hibernate code to temporarily 154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations 155 * while devices are suspended. To avoid races with the suspend/hibernate code, 156 * they should always be called with pm_mutex held (gfp_allowed_mask also should 157 * only be modified with pm_mutex held, unless the suspend/hibernate code is 158 * guaranteed not to run in parallel with that modification). 159 */ 160 161 static gfp_t saved_gfp_mask; 162 163 void pm_restore_gfp_mask(void) 164 { 165 WARN_ON(!mutex_is_locked(&pm_mutex)); 166 if (saved_gfp_mask) { 167 gfp_allowed_mask = saved_gfp_mask; 168 saved_gfp_mask = 0; 169 } 170 } 171 172 void pm_restrict_gfp_mask(void) 173 { 174 WARN_ON(!mutex_is_locked(&pm_mutex)); 175 WARN_ON(saved_gfp_mask); 176 saved_gfp_mask = gfp_allowed_mask; 177 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS); 178 } 179 180 bool pm_suspended_storage(void) 181 { 182 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) 183 return false; 184 return true; 185 } 186 #endif /* CONFIG_PM_SLEEP */ 187 188 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 189 unsigned int pageblock_order __read_mostly; 190 #endif 191 192 static void __free_pages_ok(struct page *page, unsigned int order); 193 194 /* 195 * results with 256, 32 in the lowmem_reserve sysctl: 196 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high) 197 * 1G machine -> (16M dma, 784M normal, 224M high) 198 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA 199 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL 200 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA 201 * 202 * TBD: should special case ZONE_DMA32 machines here - in those we normally 203 * don't need any ZONE_NORMAL reservation 204 */ 205 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = { 206 #ifdef CONFIG_ZONE_DMA 207 256, 208 #endif 209 #ifdef CONFIG_ZONE_DMA32 210 256, 211 #endif 212 #ifdef CONFIG_HIGHMEM 213 32, 214 #endif 215 32, 216 }; 217 218 EXPORT_SYMBOL(totalram_pages); 219 220 static char * const zone_names[MAX_NR_ZONES] = { 221 #ifdef CONFIG_ZONE_DMA 222 "DMA", 223 #endif 224 #ifdef CONFIG_ZONE_DMA32 225 "DMA32", 226 #endif 227 "Normal", 228 #ifdef CONFIG_HIGHMEM 229 "HighMem", 230 #endif 231 "Movable", 232 #ifdef CONFIG_ZONE_DEVICE 233 "Device", 234 #endif 235 }; 236 237 char * const migratetype_names[MIGRATE_TYPES] = { 238 "Unmovable", 239 "Movable", 240 "Reclaimable", 241 "HighAtomic", 242 #ifdef CONFIG_CMA 243 "CMA", 244 #endif 245 #ifdef CONFIG_MEMORY_ISOLATION 246 "Isolate", 247 #endif 248 }; 249 250 compound_page_dtor * const compound_page_dtors[] = { 251 NULL, 252 free_compound_page, 253 #ifdef CONFIG_HUGETLB_PAGE 254 free_huge_page, 255 #endif 256 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 257 free_transhuge_page, 258 #endif 259 }; 260 261 int min_free_kbytes = 1024; 262 int user_min_free_kbytes = -1; 263 int watermark_scale_factor = 10; 264 265 static unsigned long __meminitdata nr_kernel_pages; 266 static unsigned long __meminitdata nr_all_pages; 267 static unsigned long __meminitdata dma_reserve; 268 269 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 270 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES]; 271 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES]; 272 static unsigned long __initdata required_kernelcore; 273 static unsigned long __initdata required_movablecore; 274 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES]; 275 static bool mirrored_kernelcore; 276 277 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */ 278 int movable_zone; 279 EXPORT_SYMBOL(movable_zone); 280 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 281 282 #if MAX_NUMNODES > 1 283 int nr_node_ids __read_mostly = MAX_NUMNODES; 284 int nr_online_nodes __read_mostly = 1; 285 EXPORT_SYMBOL(nr_node_ids); 286 EXPORT_SYMBOL(nr_online_nodes); 287 #endif 288 289 int page_group_by_mobility_disabled __read_mostly; 290 291 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 292 static inline void reset_deferred_meminit(pg_data_t *pgdat) 293 { 294 pgdat->first_deferred_pfn = ULONG_MAX; 295 } 296 297 /* Returns true if the struct page for the pfn is uninitialised */ 298 static inline bool __meminit early_page_uninitialised(unsigned long pfn) 299 { 300 int nid = early_pfn_to_nid(pfn); 301 302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn) 303 return true; 304 305 return false; 306 } 307 308 /* 309 * Returns false when the remaining initialisation should be deferred until 310 * later in the boot cycle when it can be parallelised. 311 */ 312 static inline bool update_defer_init(pg_data_t *pgdat, 313 unsigned long pfn, unsigned long zone_end, 314 unsigned long *nr_initialised) 315 { 316 unsigned long max_initialise; 317 318 /* Always populate low zones for address-contrained allocations */ 319 if (zone_end < pgdat_end_pfn(pgdat)) 320 return true; 321 /* 322 * Initialise at least 2G of a node but also take into account that 323 * two large system hashes that can take up 1GB for 0.25TB/node. 324 */ 325 max_initialise = max(2UL << (30 - PAGE_SHIFT), 326 (pgdat->node_spanned_pages >> 8)); 327 328 (*nr_initialised)++; 329 if ((*nr_initialised > max_initialise) && 330 (pfn & (PAGES_PER_SECTION - 1)) == 0) { 331 pgdat->first_deferred_pfn = pfn; 332 return false; 333 } 334 335 return true; 336 } 337 #else 338 static inline void reset_deferred_meminit(pg_data_t *pgdat) 339 { 340 } 341 342 static inline bool early_page_uninitialised(unsigned long pfn) 343 { 344 return false; 345 } 346 347 static inline bool update_defer_init(pg_data_t *pgdat, 348 unsigned long pfn, unsigned long zone_end, 349 unsigned long *nr_initialised) 350 { 351 return true; 352 } 353 #endif 354 355 /* Return a pointer to the bitmap storing bits affecting a block of pages */ 356 static inline unsigned long *get_pageblock_bitmap(struct page *page, 357 unsigned long pfn) 358 { 359 #ifdef CONFIG_SPARSEMEM 360 return __pfn_to_section(pfn)->pageblock_flags; 361 #else 362 return page_zone(page)->pageblock_flags; 363 #endif /* CONFIG_SPARSEMEM */ 364 } 365 366 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn) 367 { 368 #ifdef CONFIG_SPARSEMEM 369 pfn &= (PAGES_PER_SECTION-1); 370 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 371 #else 372 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages); 373 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS; 374 #endif /* CONFIG_SPARSEMEM */ 375 } 376 377 /** 378 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages 379 * @page: The page within the block of interest 380 * @pfn: The target page frame number 381 * @end_bitidx: The last bit of interest to retrieve 382 * @mask: mask of bits that the caller is interested in 383 * 384 * Return: pageblock_bits flags 385 */ 386 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page, 387 unsigned long pfn, 388 unsigned long end_bitidx, 389 unsigned long mask) 390 { 391 unsigned long *bitmap; 392 unsigned long bitidx, word_bitidx; 393 unsigned long word; 394 395 bitmap = get_pageblock_bitmap(page, pfn); 396 bitidx = pfn_to_bitidx(page, pfn); 397 word_bitidx = bitidx / BITS_PER_LONG; 398 bitidx &= (BITS_PER_LONG-1); 399 400 word = bitmap[word_bitidx]; 401 bitidx += end_bitidx; 402 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask; 403 } 404 405 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn, 406 unsigned long end_bitidx, 407 unsigned long mask) 408 { 409 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask); 410 } 411 412 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn) 413 { 414 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK); 415 } 416 417 /** 418 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages 419 * @page: The page within the block of interest 420 * @flags: The flags to set 421 * @pfn: The target page frame number 422 * @end_bitidx: The last bit of interest 423 * @mask: mask of bits that the caller is interested in 424 */ 425 void set_pfnblock_flags_mask(struct page *page, unsigned long flags, 426 unsigned long pfn, 427 unsigned long end_bitidx, 428 unsigned long mask) 429 { 430 unsigned long *bitmap; 431 unsigned long bitidx, word_bitidx; 432 unsigned long old_word, word; 433 434 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4); 435 436 bitmap = get_pageblock_bitmap(page, pfn); 437 bitidx = pfn_to_bitidx(page, pfn); 438 word_bitidx = bitidx / BITS_PER_LONG; 439 bitidx &= (BITS_PER_LONG-1); 440 441 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page); 442 443 bitidx += end_bitidx; 444 mask <<= (BITS_PER_LONG - bitidx - 1); 445 flags <<= (BITS_PER_LONG - bitidx - 1); 446 447 word = READ_ONCE(bitmap[word_bitidx]); 448 for (;;) { 449 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags); 450 if (word == old_word) 451 break; 452 word = old_word; 453 } 454 } 455 456 void set_pageblock_migratetype(struct page *page, int migratetype) 457 { 458 if (unlikely(page_group_by_mobility_disabled && 459 migratetype < MIGRATE_PCPTYPES)) 460 migratetype = MIGRATE_UNMOVABLE; 461 462 set_pageblock_flags_group(page, (unsigned long)migratetype, 463 PB_migrate, PB_migrate_end); 464 } 465 466 #ifdef CONFIG_DEBUG_VM 467 static int page_outside_zone_boundaries(struct zone *zone, struct page *page) 468 { 469 int ret = 0; 470 unsigned seq; 471 unsigned long pfn = page_to_pfn(page); 472 unsigned long sp, start_pfn; 473 474 do { 475 seq = zone_span_seqbegin(zone); 476 start_pfn = zone->zone_start_pfn; 477 sp = zone->spanned_pages; 478 if (!zone_spans_pfn(zone, pfn)) 479 ret = 1; 480 } while (zone_span_seqretry(zone, seq)); 481 482 if (ret) 483 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n", 484 pfn, zone_to_nid(zone), zone->name, 485 start_pfn, start_pfn + sp); 486 487 return ret; 488 } 489 490 static int page_is_consistent(struct zone *zone, struct page *page) 491 { 492 if (!pfn_valid_within(page_to_pfn(page))) 493 return 0; 494 if (zone != page_zone(page)) 495 return 0; 496 497 return 1; 498 } 499 /* 500 * Temporary debugging check for pages not lying within a given zone. 501 */ 502 static int bad_range(struct zone *zone, struct page *page) 503 { 504 if (page_outside_zone_boundaries(zone, page)) 505 return 1; 506 if (!page_is_consistent(zone, page)) 507 return 1; 508 509 return 0; 510 } 511 #else 512 static inline int bad_range(struct zone *zone, struct page *page) 513 { 514 return 0; 515 } 516 #endif 517 518 static void bad_page(struct page *page, const char *reason, 519 unsigned long bad_flags) 520 { 521 static unsigned long resume; 522 static unsigned long nr_shown; 523 static unsigned long nr_unshown; 524 525 /* 526 * Allow a burst of 60 reports, then keep quiet for that minute; 527 * or allow a steady drip of one report per second. 528 */ 529 if (nr_shown == 60) { 530 if (time_before(jiffies, resume)) { 531 nr_unshown++; 532 goto out; 533 } 534 if (nr_unshown) { 535 pr_alert( 536 "BUG: Bad page state: %lu messages suppressed\n", 537 nr_unshown); 538 nr_unshown = 0; 539 } 540 nr_shown = 0; 541 } 542 if (nr_shown++ == 0) 543 resume = jiffies + 60 * HZ; 544 545 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n", 546 current->comm, page_to_pfn(page)); 547 __dump_page(page, reason); 548 bad_flags &= page->flags; 549 if (bad_flags) 550 pr_alert("bad because of flags: %#lx(%pGp)\n", 551 bad_flags, &bad_flags); 552 dump_page_owner(page); 553 554 print_modules(); 555 dump_stack(); 556 out: 557 /* Leave bad fields for debug, except PageBuddy could make trouble */ 558 page_mapcount_reset(page); /* remove PageBuddy */ 559 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); 560 } 561 562 /* 563 * Higher-order pages are called "compound pages". They are structured thusly: 564 * 565 * The first PAGE_SIZE page is called the "head page" and have PG_head set. 566 * 567 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded 568 * in bit 0 of page->compound_head. The rest of bits is pointer to head page. 569 * 570 * The first tail page's ->compound_dtor holds the offset in array of compound 571 * page destructors. See compound_page_dtors. 572 * 573 * The first tail page's ->compound_order holds the order of allocation. 574 * This usage means that zero-order pages may not be compound. 575 */ 576 577 void free_compound_page(struct page *page) 578 { 579 __free_pages_ok(page, compound_order(page)); 580 } 581 582 void prep_compound_page(struct page *page, unsigned int order) 583 { 584 int i; 585 int nr_pages = 1 << order; 586 587 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); 588 set_compound_order(page, order); 589 __SetPageHead(page); 590 for (i = 1; i < nr_pages; i++) { 591 struct page *p = page + i; 592 set_page_count(p, 0); 593 p->mapping = TAIL_MAPPING; 594 set_compound_head(p, page); 595 } 596 atomic_set(compound_mapcount_ptr(page), -1); 597 } 598 599 #ifdef CONFIG_DEBUG_PAGEALLOC 600 unsigned int _debug_guardpage_minorder; 601 bool _debug_pagealloc_enabled __read_mostly 602 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT); 603 EXPORT_SYMBOL(_debug_pagealloc_enabled); 604 bool _debug_guardpage_enabled __read_mostly; 605 606 static int __init early_debug_pagealloc(char *buf) 607 { 608 if (!buf) 609 return -EINVAL; 610 return kstrtobool(buf, &_debug_pagealloc_enabled); 611 } 612 early_param("debug_pagealloc", early_debug_pagealloc); 613 614 static bool need_debug_guardpage(void) 615 { 616 /* If we don't use debug_pagealloc, we don't need guard page */ 617 if (!debug_pagealloc_enabled()) 618 return false; 619 620 if (!debug_guardpage_minorder()) 621 return false; 622 623 return true; 624 } 625 626 static void init_debug_guardpage(void) 627 { 628 if (!debug_pagealloc_enabled()) 629 return; 630 631 if (!debug_guardpage_minorder()) 632 return; 633 634 _debug_guardpage_enabled = true; 635 } 636 637 struct page_ext_operations debug_guardpage_ops = { 638 .need = need_debug_guardpage, 639 .init = init_debug_guardpage, 640 }; 641 642 static int __init debug_guardpage_minorder_setup(char *buf) 643 { 644 unsigned long res; 645 646 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) { 647 pr_err("Bad debug_guardpage_minorder value\n"); 648 return 0; 649 } 650 _debug_guardpage_minorder = res; 651 pr_info("Setting debug_guardpage_minorder to %lu\n", res); 652 return 0; 653 } 654 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup); 655 656 static inline bool set_page_guard(struct zone *zone, struct page *page, 657 unsigned int order, int migratetype) 658 { 659 struct page_ext *page_ext; 660 661 if (!debug_guardpage_enabled()) 662 return false; 663 664 if (order >= debug_guardpage_minorder()) 665 return false; 666 667 page_ext = lookup_page_ext(page); 668 if (unlikely(!page_ext)) 669 return false; 670 671 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 672 673 INIT_LIST_HEAD(&page->lru); 674 set_page_private(page, order); 675 /* Guard pages are not available for any usage */ 676 __mod_zone_freepage_state(zone, -(1 << order), migratetype); 677 678 return true; 679 } 680 681 static inline void clear_page_guard(struct zone *zone, struct page *page, 682 unsigned int order, int migratetype) 683 { 684 struct page_ext *page_ext; 685 686 if (!debug_guardpage_enabled()) 687 return; 688 689 page_ext = lookup_page_ext(page); 690 if (unlikely(!page_ext)) 691 return; 692 693 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags); 694 695 set_page_private(page, 0); 696 if (!is_migrate_isolate(migratetype)) 697 __mod_zone_freepage_state(zone, (1 << order), migratetype); 698 } 699 #else 700 struct page_ext_operations debug_guardpage_ops; 701 static inline bool set_page_guard(struct zone *zone, struct page *page, 702 unsigned int order, int migratetype) { return false; } 703 static inline void clear_page_guard(struct zone *zone, struct page *page, 704 unsigned int order, int migratetype) {} 705 #endif 706 707 static inline void set_page_order(struct page *page, unsigned int order) 708 { 709 set_page_private(page, order); 710 __SetPageBuddy(page); 711 } 712 713 static inline void rmv_page_order(struct page *page) 714 { 715 __ClearPageBuddy(page); 716 set_page_private(page, 0); 717 } 718 719 /* 720 * This function checks whether a page is free && is the buddy 721 * we can do coalesce a page and its buddy if 722 * (a) the buddy is not in a hole (check before calling!) && 723 * (b) the buddy is in the buddy system && 724 * (c) a page and its buddy have the same order && 725 * (d) a page and its buddy are in the same zone. 726 * 727 * For recording whether a page is in the buddy system, we set ->_mapcount 728 * PAGE_BUDDY_MAPCOUNT_VALUE. 729 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is 730 * serialized by zone->lock. 731 * 732 * For recording page's order, we use page_private(page). 733 */ 734 static inline int page_is_buddy(struct page *page, struct page *buddy, 735 unsigned int order) 736 { 737 if (page_is_guard(buddy) && page_order(buddy) == order) { 738 if (page_zone_id(page) != page_zone_id(buddy)) 739 return 0; 740 741 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 742 743 return 1; 744 } 745 746 if (PageBuddy(buddy) && page_order(buddy) == order) { 747 /* 748 * zone check is done late to avoid uselessly 749 * calculating zone/node ids for pages that could 750 * never merge. 751 */ 752 if (page_zone_id(page) != page_zone_id(buddy)) 753 return 0; 754 755 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy); 756 757 return 1; 758 } 759 return 0; 760 } 761 762 /* 763 * Freeing function for a buddy system allocator. 764 * 765 * The concept of a buddy system is to maintain direct-mapped table 766 * (containing bit values) for memory blocks of various "orders". 767 * The bottom level table contains the map for the smallest allocatable 768 * units of memory (here, pages), and each level above it describes 769 * pairs of units from the levels below, hence, "buddies". 770 * At a high level, all that happens here is marking the table entry 771 * at the bottom level available, and propagating the changes upward 772 * as necessary, plus some accounting needed to play nicely with other 773 * parts of the VM system. 774 * At each level, we keep a list of pages, which are heads of continuous 775 * free pages of length of (1 << order) and marked with _mapcount 776 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page) 777 * field. 778 * So when we are allocating or freeing one, we can derive the state of the 779 * other. That is, if we allocate a small block, and both were 780 * free, the remainder of the region must be split into blocks. 781 * If a block is freed, and its buddy is also free, then this 782 * triggers coalescing into a block of larger size. 783 * 784 * -- nyc 785 */ 786 787 static inline void __free_one_page(struct page *page, 788 unsigned long pfn, 789 struct zone *zone, unsigned int order, 790 int migratetype) 791 { 792 unsigned long combined_pfn; 793 unsigned long uninitialized_var(buddy_pfn); 794 struct page *buddy; 795 unsigned int max_order; 796 797 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1); 798 799 VM_BUG_ON(!zone_is_initialized(zone)); 800 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page); 801 802 VM_BUG_ON(migratetype == -1); 803 if (likely(!is_migrate_isolate(migratetype))) 804 __mod_zone_freepage_state(zone, 1 << order, migratetype); 805 806 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page); 807 VM_BUG_ON_PAGE(bad_range(zone, page), page); 808 809 continue_merging: 810 while (order < max_order - 1) { 811 buddy_pfn = __find_buddy_pfn(pfn, order); 812 buddy = page + (buddy_pfn - pfn); 813 814 if (!pfn_valid_within(buddy_pfn)) 815 goto done_merging; 816 if (!page_is_buddy(page, buddy, order)) 817 goto done_merging; 818 /* 819 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page, 820 * merge with it and move up one order. 821 */ 822 if (page_is_guard(buddy)) { 823 clear_page_guard(zone, buddy, order, migratetype); 824 } else { 825 list_del(&buddy->lru); 826 zone->free_area[order].nr_free--; 827 rmv_page_order(buddy); 828 } 829 combined_pfn = buddy_pfn & pfn; 830 page = page + (combined_pfn - pfn); 831 pfn = combined_pfn; 832 order++; 833 } 834 if (max_order < MAX_ORDER) { 835 /* If we are here, it means order is >= pageblock_order. 836 * We want to prevent merge between freepages on isolate 837 * pageblock and normal pageblock. Without this, pageblock 838 * isolation could cause incorrect freepage or CMA accounting. 839 * 840 * We don't want to hit this code for the more frequent 841 * low-order merging. 842 */ 843 if (unlikely(has_isolate_pageblock(zone))) { 844 int buddy_mt; 845 846 buddy_pfn = __find_buddy_pfn(pfn, order); 847 buddy = page + (buddy_pfn - pfn); 848 buddy_mt = get_pageblock_migratetype(buddy); 849 850 if (migratetype != buddy_mt 851 && (is_migrate_isolate(migratetype) || 852 is_migrate_isolate(buddy_mt))) 853 goto done_merging; 854 } 855 max_order++; 856 goto continue_merging; 857 } 858 859 done_merging: 860 set_page_order(page, order); 861 862 /* 863 * If this is not the largest possible page, check if the buddy 864 * of the next-highest order is free. If it is, it's possible 865 * that pages are being freed that will coalesce soon. In case, 866 * that is happening, add the free page to the tail of the list 867 * so it's less likely to be used soon and more likely to be merged 868 * as a higher order page 869 */ 870 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) { 871 struct page *higher_page, *higher_buddy; 872 combined_pfn = buddy_pfn & pfn; 873 higher_page = page + (combined_pfn - pfn); 874 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1); 875 higher_buddy = higher_page + (buddy_pfn - combined_pfn); 876 if (pfn_valid_within(buddy_pfn) && 877 page_is_buddy(higher_page, higher_buddy, order + 1)) { 878 list_add_tail(&page->lru, 879 &zone->free_area[order].free_list[migratetype]); 880 goto out; 881 } 882 } 883 884 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]); 885 out: 886 zone->free_area[order].nr_free++; 887 } 888 889 /* 890 * A bad page could be due to a number of fields. Instead of multiple branches, 891 * try and check multiple fields with one check. The caller must do a detailed 892 * check if necessary. 893 */ 894 static inline bool page_expected_state(struct page *page, 895 unsigned long check_flags) 896 { 897 if (unlikely(atomic_read(&page->_mapcount) != -1)) 898 return false; 899 900 if (unlikely((unsigned long)page->mapping | 901 page_ref_count(page) | 902 #ifdef CONFIG_MEMCG 903 (unsigned long)page->mem_cgroup | 904 #endif 905 (page->flags & check_flags))) 906 return false; 907 908 return true; 909 } 910 911 static void free_pages_check_bad(struct page *page) 912 { 913 const char *bad_reason; 914 unsigned long bad_flags; 915 916 bad_reason = NULL; 917 bad_flags = 0; 918 919 if (unlikely(atomic_read(&page->_mapcount) != -1)) 920 bad_reason = "nonzero mapcount"; 921 if (unlikely(page->mapping != NULL)) 922 bad_reason = "non-NULL mapping"; 923 if (unlikely(page_ref_count(page) != 0)) 924 bad_reason = "nonzero _refcount"; 925 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) { 926 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set"; 927 bad_flags = PAGE_FLAGS_CHECK_AT_FREE; 928 } 929 #ifdef CONFIG_MEMCG 930 if (unlikely(page->mem_cgroup)) 931 bad_reason = "page still charged to cgroup"; 932 #endif 933 bad_page(page, bad_reason, bad_flags); 934 } 935 936 static inline int free_pages_check(struct page *page) 937 { 938 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE))) 939 return 0; 940 941 /* Something has gone sideways, find it */ 942 free_pages_check_bad(page); 943 return 1; 944 } 945 946 static int free_tail_pages_check(struct page *head_page, struct page *page) 947 { 948 int ret = 1; 949 950 /* 951 * We rely page->lru.next never has bit 0 set, unless the page 952 * is PageTail(). Let's make sure that's true even for poisoned ->lru. 953 */ 954 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1); 955 956 if (!IS_ENABLED(CONFIG_DEBUG_VM)) { 957 ret = 0; 958 goto out; 959 } 960 switch (page - head_page) { 961 case 1: 962 /* the first tail page: ->mapping is compound_mapcount() */ 963 if (unlikely(compound_mapcount(page))) { 964 bad_page(page, "nonzero compound_mapcount", 0); 965 goto out; 966 } 967 break; 968 case 2: 969 /* 970 * the second tail page: ->mapping is 971 * page_deferred_list().next -- ignore value. 972 */ 973 break; 974 default: 975 if (page->mapping != TAIL_MAPPING) { 976 bad_page(page, "corrupted mapping in tail page", 0); 977 goto out; 978 } 979 break; 980 } 981 if (unlikely(!PageTail(page))) { 982 bad_page(page, "PageTail not set", 0); 983 goto out; 984 } 985 if (unlikely(compound_head(page) != head_page)) { 986 bad_page(page, "compound_head not consistent", 0); 987 goto out; 988 } 989 ret = 0; 990 out: 991 page->mapping = NULL; 992 clear_compound_head(page); 993 return ret; 994 } 995 996 static __always_inline bool free_pages_prepare(struct page *page, 997 unsigned int order, bool check_free) 998 { 999 int bad = 0; 1000 1001 VM_BUG_ON_PAGE(PageTail(page), page); 1002 1003 trace_mm_page_free(page, order); 1004 kmemcheck_free_shadow(page, order); 1005 1006 /* 1007 * Check tail pages before head page information is cleared to 1008 * avoid checking PageCompound for order-0 pages. 1009 */ 1010 if (unlikely(order)) { 1011 bool compound = PageCompound(page); 1012 int i; 1013 1014 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page); 1015 1016 if (compound) 1017 ClearPageDoubleMap(page); 1018 for (i = 1; i < (1 << order); i++) { 1019 if (compound) 1020 bad += free_tail_pages_check(page, page + i); 1021 if (unlikely(free_pages_check(page + i))) { 1022 bad++; 1023 continue; 1024 } 1025 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 1026 } 1027 } 1028 if (PageMappingFlags(page)) 1029 page->mapping = NULL; 1030 if (memcg_kmem_enabled() && PageKmemcg(page)) 1031 memcg_kmem_uncharge(page, order); 1032 if (check_free) 1033 bad += free_pages_check(page); 1034 if (bad) 1035 return false; 1036 1037 page_cpupid_reset_last(page); 1038 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP; 1039 reset_page_owner(page, order); 1040 1041 if (!PageHighMem(page)) { 1042 debug_check_no_locks_freed(page_address(page), 1043 PAGE_SIZE << order); 1044 debug_check_no_obj_freed(page_address(page), 1045 PAGE_SIZE << order); 1046 } 1047 arch_free_page(page, order); 1048 kernel_poison_pages(page, 1 << order, 0); 1049 kernel_map_pages(page, 1 << order, 0); 1050 kasan_free_pages(page, order); 1051 1052 return true; 1053 } 1054 1055 #ifdef CONFIG_DEBUG_VM 1056 static inline bool free_pcp_prepare(struct page *page) 1057 { 1058 return free_pages_prepare(page, 0, true); 1059 } 1060 1061 static inline bool bulkfree_pcp_prepare(struct page *page) 1062 { 1063 return false; 1064 } 1065 #else 1066 static bool free_pcp_prepare(struct page *page) 1067 { 1068 return free_pages_prepare(page, 0, false); 1069 } 1070 1071 static bool bulkfree_pcp_prepare(struct page *page) 1072 { 1073 return free_pages_check(page); 1074 } 1075 #endif /* CONFIG_DEBUG_VM */ 1076 1077 /* 1078 * Frees a number of pages from the PCP lists 1079 * Assumes all pages on list are in same zone, and of same order. 1080 * count is the number of pages to free. 1081 * 1082 * If the zone was previously in an "all pages pinned" state then look to 1083 * see if this freeing clears that state. 1084 * 1085 * And clear the zone's pages_scanned counter, to hold off the "all pages are 1086 * pinned" detection logic. 1087 */ 1088 static void free_pcppages_bulk(struct zone *zone, int count, 1089 struct per_cpu_pages *pcp) 1090 { 1091 int migratetype = 0; 1092 int batch_free = 0; 1093 unsigned long nr_scanned; 1094 bool isolated_pageblocks; 1095 1096 spin_lock(&zone->lock); 1097 isolated_pageblocks = has_isolate_pageblock(zone); 1098 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED); 1099 if (nr_scanned) 1100 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned); 1101 1102 while (count) { 1103 struct page *page; 1104 struct list_head *list; 1105 1106 /* 1107 * Remove pages from lists in a round-robin fashion. A 1108 * batch_free count is maintained that is incremented when an 1109 * empty list is encountered. This is so more pages are freed 1110 * off fuller lists instead of spinning excessively around empty 1111 * lists 1112 */ 1113 do { 1114 batch_free++; 1115 if (++migratetype == MIGRATE_PCPTYPES) 1116 migratetype = 0; 1117 list = &pcp->lists[migratetype]; 1118 } while (list_empty(list)); 1119 1120 /* This is the only non-empty list. Free them all. */ 1121 if (batch_free == MIGRATE_PCPTYPES) 1122 batch_free = count; 1123 1124 do { 1125 int mt; /* migratetype of the to-be-freed page */ 1126 1127 page = list_last_entry(list, struct page, lru); 1128 /* must delete as __free_one_page list manipulates */ 1129 list_del(&page->lru); 1130 1131 mt = get_pcppage_migratetype(page); 1132 /* MIGRATE_ISOLATE page should not go to pcplists */ 1133 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page); 1134 /* Pageblock could have been isolated meanwhile */ 1135 if (unlikely(isolated_pageblocks)) 1136 mt = get_pageblock_migratetype(page); 1137 1138 if (bulkfree_pcp_prepare(page)) 1139 continue; 1140 1141 __free_one_page(page, page_to_pfn(page), zone, 0, mt); 1142 trace_mm_page_pcpu_drain(page, 0, mt); 1143 } while (--count && --batch_free && !list_empty(list)); 1144 } 1145 spin_unlock(&zone->lock); 1146 } 1147 1148 static void free_one_page(struct zone *zone, 1149 struct page *page, unsigned long pfn, 1150 unsigned int order, 1151 int migratetype) 1152 { 1153 unsigned long nr_scanned; 1154 spin_lock(&zone->lock); 1155 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED); 1156 if (nr_scanned) 1157 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned); 1158 1159 if (unlikely(has_isolate_pageblock(zone) || 1160 is_migrate_isolate(migratetype))) { 1161 migratetype = get_pfnblock_migratetype(page, pfn); 1162 } 1163 __free_one_page(page, pfn, zone, order, migratetype); 1164 spin_unlock(&zone->lock); 1165 } 1166 1167 static void __meminit __init_single_page(struct page *page, unsigned long pfn, 1168 unsigned long zone, int nid) 1169 { 1170 set_page_links(page, zone, nid, pfn); 1171 init_page_count(page); 1172 page_mapcount_reset(page); 1173 page_cpupid_reset_last(page); 1174 1175 INIT_LIST_HEAD(&page->lru); 1176 #ifdef WANT_PAGE_VIRTUAL 1177 /* The shift won't overflow because ZONE_NORMAL is below 4G. */ 1178 if (!is_highmem_idx(zone)) 1179 set_page_address(page, __va(pfn << PAGE_SHIFT)); 1180 #endif 1181 } 1182 1183 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone, 1184 int nid) 1185 { 1186 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid); 1187 } 1188 1189 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 1190 static void init_reserved_page(unsigned long pfn) 1191 { 1192 pg_data_t *pgdat; 1193 int nid, zid; 1194 1195 if (!early_page_uninitialised(pfn)) 1196 return; 1197 1198 nid = early_pfn_to_nid(pfn); 1199 pgdat = NODE_DATA(nid); 1200 1201 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1202 struct zone *zone = &pgdat->node_zones[zid]; 1203 1204 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone)) 1205 break; 1206 } 1207 __init_single_pfn(pfn, zid, nid); 1208 } 1209 #else 1210 static inline void init_reserved_page(unsigned long pfn) 1211 { 1212 } 1213 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 1214 1215 /* 1216 * Initialised pages do not have PageReserved set. This function is 1217 * called for each range allocated by the bootmem allocator and 1218 * marks the pages PageReserved. The remaining valid pages are later 1219 * sent to the buddy page allocator. 1220 */ 1221 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end) 1222 { 1223 unsigned long start_pfn = PFN_DOWN(start); 1224 unsigned long end_pfn = PFN_UP(end); 1225 1226 for (; start_pfn < end_pfn; start_pfn++) { 1227 if (pfn_valid(start_pfn)) { 1228 struct page *page = pfn_to_page(start_pfn); 1229 1230 init_reserved_page(start_pfn); 1231 1232 /* Avoid false-positive PageTail() */ 1233 INIT_LIST_HEAD(&page->lru); 1234 1235 SetPageReserved(page); 1236 } 1237 } 1238 } 1239 1240 static void __free_pages_ok(struct page *page, unsigned int order) 1241 { 1242 unsigned long flags; 1243 int migratetype; 1244 unsigned long pfn = page_to_pfn(page); 1245 1246 if (!free_pages_prepare(page, order, true)) 1247 return; 1248 1249 migratetype = get_pfnblock_migratetype(page, pfn); 1250 local_irq_save(flags); 1251 __count_vm_events(PGFREE, 1 << order); 1252 free_one_page(page_zone(page), page, pfn, order, migratetype); 1253 local_irq_restore(flags); 1254 } 1255 1256 static void __init __free_pages_boot_core(struct page *page, unsigned int order) 1257 { 1258 unsigned int nr_pages = 1 << order; 1259 struct page *p = page; 1260 unsigned int loop; 1261 1262 prefetchw(p); 1263 for (loop = 0; loop < (nr_pages - 1); loop++, p++) { 1264 prefetchw(p + 1); 1265 __ClearPageReserved(p); 1266 set_page_count(p, 0); 1267 } 1268 __ClearPageReserved(p); 1269 set_page_count(p, 0); 1270 1271 page_zone(page)->managed_pages += nr_pages; 1272 set_page_refcounted(page); 1273 __free_pages(page, order); 1274 } 1275 1276 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \ 1277 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) 1278 1279 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata; 1280 1281 int __meminit early_pfn_to_nid(unsigned long pfn) 1282 { 1283 static DEFINE_SPINLOCK(early_pfn_lock); 1284 int nid; 1285 1286 spin_lock(&early_pfn_lock); 1287 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache); 1288 if (nid < 0) 1289 nid = first_online_node; 1290 spin_unlock(&early_pfn_lock); 1291 1292 return nid; 1293 } 1294 #endif 1295 1296 #ifdef CONFIG_NODES_SPAN_OTHER_NODES 1297 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1298 struct mminit_pfnnid_cache *state) 1299 { 1300 int nid; 1301 1302 nid = __early_pfn_to_nid(pfn, state); 1303 if (nid >= 0 && nid != node) 1304 return false; 1305 return true; 1306 } 1307 1308 /* Only safe to use early in boot when initialisation is single-threaded */ 1309 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1310 { 1311 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache); 1312 } 1313 1314 #else 1315 1316 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node) 1317 { 1318 return true; 1319 } 1320 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node, 1321 struct mminit_pfnnid_cache *state) 1322 { 1323 return true; 1324 } 1325 #endif 1326 1327 1328 void __init __free_pages_bootmem(struct page *page, unsigned long pfn, 1329 unsigned int order) 1330 { 1331 if (early_page_uninitialised(pfn)) 1332 return; 1333 return __free_pages_boot_core(page, order); 1334 } 1335 1336 /* 1337 * Check that the whole (or subset of) a pageblock given by the interval of 1338 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it 1339 * with the migration of free compaction scanner. The scanners then need to 1340 * use only pfn_valid_within() check for arches that allow holes within 1341 * pageblocks. 1342 * 1343 * Return struct page pointer of start_pfn, or NULL if checks were not passed. 1344 * 1345 * It's possible on some configurations to have a setup like node0 node1 node0 1346 * i.e. it's possible that all pages within a zones range of pages do not 1347 * belong to a single zone. We assume that a border between node0 and node1 1348 * can occur within a single pageblock, but not a node0 node1 node0 1349 * interleaving within a single pageblock. It is therefore sufficient to check 1350 * the first and last page of a pageblock and avoid checking each individual 1351 * page in a pageblock. 1352 */ 1353 struct page *__pageblock_pfn_to_page(unsigned long start_pfn, 1354 unsigned long end_pfn, struct zone *zone) 1355 { 1356 struct page *start_page; 1357 struct page *end_page; 1358 1359 /* end_pfn is one past the range we are checking */ 1360 end_pfn--; 1361 1362 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn)) 1363 return NULL; 1364 1365 start_page = pfn_to_page(start_pfn); 1366 1367 if (page_zone(start_page) != zone) 1368 return NULL; 1369 1370 end_page = pfn_to_page(end_pfn); 1371 1372 /* This gives a shorter code than deriving page_zone(end_page) */ 1373 if (page_zone_id(start_page) != page_zone_id(end_page)) 1374 return NULL; 1375 1376 return start_page; 1377 } 1378 1379 void set_zone_contiguous(struct zone *zone) 1380 { 1381 unsigned long block_start_pfn = zone->zone_start_pfn; 1382 unsigned long block_end_pfn; 1383 1384 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages); 1385 for (; block_start_pfn < zone_end_pfn(zone); 1386 block_start_pfn = block_end_pfn, 1387 block_end_pfn += pageblock_nr_pages) { 1388 1389 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone)); 1390 1391 if (!__pageblock_pfn_to_page(block_start_pfn, 1392 block_end_pfn, zone)) 1393 return; 1394 } 1395 1396 /* We confirm that there is no hole */ 1397 zone->contiguous = true; 1398 } 1399 1400 void clear_zone_contiguous(struct zone *zone) 1401 { 1402 zone->contiguous = false; 1403 } 1404 1405 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 1406 static void __init deferred_free_range(struct page *page, 1407 unsigned long pfn, int nr_pages) 1408 { 1409 int i; 1410 1411 if (!page) 1412 return; 1413 1414 /* Free a large naturally-aligned chunk if possible */ 1415 if (nr_pages == pageblock_nr_pages && 1416 (pfn & (pageblock_nr_pages - 1)) == 0) { 1417 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1418 __free_pages_boot_core(page, pageblock_order); 1419 return; 1420 } 1421 1422 for (i = 0; i < nr_pages; i++, page++, pfn++) { 1423 if ((pfn & (pageblock_nr_pages - 1)) == 0) 1424 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 1425 __free_pages_boot_core(page, 0); 1426 } 1427 } 1428 1429 /* Completion tracking for deferred_init_memmap() threads */ 1430 static atomic_t pgdat_init_n_undone __initdata; 1431 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp); 1432 1433 static inline void __init pgdat_init_report_one_done(void) 1434 { 1435 if (atomic_dec_and_test(&pgdat_init_n_undone)) 1436 complete(&pgdat_init_all_done_comp); 1437 } 1438 1439 /* Initialise remaining memory on a node */ 1440 static int __init deferred_init_memmap(void *data) 1441 { 1442 pg_data_t *pgdat = data; 1443 int nid = pgdat->node_id; 1444 struct mminit_pfnnid_cache nid_init_state = { }; 1445 unsigned long start = jiffies; 1446 unsigned long nr_pages = 0; 1447 unsigned long walk_start, walk_end; 1448 int i, zid; 1449 struct zone *zone; 1450 unsigned long first_init_pfn = pgdat->first_deferred_pfn; 1451 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); 1452 1453 if (first_init_pfn == ULONG_MAX) { 1454 pgdat_init_report_one_done(); 1455 return 0; 1456 } 1457 1458 /* Bind memory initialisation thread to a local node if possible */ 1459 if (!cpumask_empty(cpumask)) 1460 set_cpus_allowed_ptr(current, cpumask); 1461 1462 /* Sanity check boundaries */ 1463 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn); 1464 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat)); 1465 pgdat->first_deferred_pfn = ULONG_MAX; 1466 1467 /* Only the highest zone is deferred so find it */ 1468 for (zid = 0; zid < MAX_NR_ZONES; zid++) { 1469 zone = pgdat->node_zones + zid; 1470 if (first_init_pfn < zone_end_pfn(zone)) 1471 break; 1472 } 1473 1474 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) { 1475 unsigned long pfn, end_pfn; 1476 struct page *page = NULL; 1477 struct page *free_base_page = NULL; 1478 unsigned long free_base_pfn = 0; 1479 int nr_to_free = 0; 1480 1481 end_pfn = min(walk_end, zone_end_pfn(zone)); 1482 pfn = first_init_pfn; 1483 if (pfn < walk_start) 1484 pfn = walk_start; 1485 if (pfn < zone->zone_start_pfn) 1486 pfn = zone->zone_start_pfn; 1487 1488 for (; pfn < end_pfn; pfn++) { 1489 if (!pfn_valid_within(pfn)) 1490 goto free_range; 1491 1492 /* 1493 * Ensure pfn_valid is checked every 1494 * pageblock_nr_pages for memory holes 1495 */ 1496 if ((pfn & (pageblock_nr_pages - 1)) == 0) { 1497 if (!pfn_valid(pfn)) { 1498 page = NULL; 1499 goto free_range; 1500 } 1501 } 1502 1503 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) { 1504 page = NULL; 1505 goto free_range; 1506 } 1507 1508 /* Minimise pfn page lookups and scheduler checks */ 1509 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) { 1510 page++; 1511 } else { 1512 nr_pages += nr_to_free; 1513 deferred_free_range(free_base_page, 1514 free_base_pfn, nr_to_free); 1515 free_base_page = NULL; 1516 free_base_pfn = nr_to_free = 0; 1517 1518 page = pfn_to_page(pfn); 1519 cond_resched(); 1520 } 1521 1522 if (page->flags) { 1523 VM_BUG_ON(page_zone(page) != zone); 1524 goto free_range; 1525 } 1526 1527 __init_single_page(page, pfn, zid, nid); 1528 if (!free_base_page) { 1529 free_base_page = page; 1530 free_base_pfn = pfn; 1531 nr_to_free = 0; 1532 } 1533 nr_to_free++; 1534 1535 /* Where possible, batch up pages for a single free */ 1536 continue; 1537 free_range: 1538 /* Free the current block of pages to allocator */ 1539 nr_pages += nr_to_free; 1540 deferred_free_range(free_base_page, free_base_pfn, 1541 nr_to_free); 1542 free_base_page = NULL; 1543 free_base_pfn = nr_to_free = 0; 1544 } 1545 /* Free the last block of pages to allocator */ 1546 nr_pages += nr_to_free; 1547 deferred_free_range(free_base_page, free_base_pfn, nr_to_free); 1548 1549 first_init_pfn = max(end_pfn, first_init_pfn); 1550 } 1551 1552 /* Sanity check that the next zone really is unpopulated */ 1553 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone)); 1554 1555 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages, 1556 jiffies_to_msecs(jiffies - start)); 1557 1558 pgdat_init_report_one_done(); 1559 return 0; 1560 } 1561 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */ 1562 1563 void __init page_alloc_init_late(void) 1564 { 1565 struct zone *zone; 1566 1567 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT 1568 int nid; 1569 1570 /* There will be num_node_state(N_MEMORY) threads */ 1571 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY)); 1572 for_each_node_state(nid, N_MEMORY) { 1573 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid); 1574 } 1575 1576 /* Block until all are initialised */ 1577 wait_for_completion(&pgdat_init_all_done_comp); 1578 1579 /* Reinit limits that are based on free pages after the kernel is up */ 1580 files_maxfiles_init(); 1581 #endif 1582 1583 for_each_populated_zone(zone) 1584 set_zone_contiguous(zone); 1585 } 1586 1587 #ifdef CONFIG_CMA 1588 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */ 1589 void __init init_cma_reserved_pageblock(struct page *page) 1590 { 1591 unsigned i = pageblock_nr_pages; 1592 struct page *p = page; 1593 1594 do { 1595 __ClearPageReserved(p); 1596 set_page_count(p, 0); 1597 } while (++p, --i); 1598 1599 set_pageblock_migratetype(page, MIGRATE_CMA); 1600 1601 if (pageblock_order >= MAX_ORDER) { 1602 i = pageblock_nr_pages; 1603 p = page; 1604 do { 1605 set_page_refcounted(p); 1606 __free_pages(p, MAX_ORDER - 1); 1607 p += MAX_ORDER_NR_PAGES; 1608 } while (i -= MAX_ORDER_NR_PAGES); 1609 } else { 1610 set_page_refcounted(page); 1611 __free_pages(page, pageblock_order); 1612 } 1613 1614 adjust_managed_page_count(page, pageblock_nr_pages); 1615 } 1616 #endif 1617 1618 /* 1619 * The order of subdivision here is critical for the IO subsystem. 1620 * Please do not alter this order without good reasons and regression 1621 * testing. Specifically, as large blocks of memory are subdivided, 1622 * the order in which smaller blocks are delivered depends on the order 1623 * they're subdivided in this function. This is the primary factor 1624 * influencing the order in which pages are delivered to the IO 1625 * subsystem according to empirical testing, and this is also justified 1626 * by considering the behavior of a buddy system containing a single 1627 * large block of memory acted on by a series of small allocations. 1628 * This behavior is a critical factor in sglist merging's success. 1629 * 1630 * -- nyc 1631 */ 1632 static inline void expand(struct zone *zone, struct page *page, 1633 int low, int high, struct free_area *area, 1634 int migratetype) 1635 { 1636 unsigned long size = 1 << high; 1637 1638 while (high > low) { 1639 area--; 1640 high--; 1641 size >>= 1; 1642 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]); 1643 1644 /* 1645 * Mark as guard pages (or page), that will allow to 1646 * merge back to allocator when buddy will be freed. 1647 * Corresponding page table entries will not be touched, 1648 * pages will stay not present in virtual address space 1649 */ 1650 if (set_page_guard(zone, &page[size], high, migratetype)) 1651 continue; 1652 1653 list_add(&page[size].lru, &area->free_list[migratetype]); 1654 area->nr_free++; 1655 set_page_order(&page[size], high); 1656 } 1657 } 1658 1659 static void check_new_page_bad(struct page *page) 1660 { 1661 const char *bad_reason = NULL; 1662 unsigned long bad_flags = 0; 1663 1664 if (unlikely(atomic_read(&page->_mapcount) != -1)) 1665 bad_reason = "nonzero mapcount"; 1666 if (unlikely(page->mapping != NULL)) 1667 bad_reason = "non-NULL mapping"; 1668 if (unlikely(page_ref_count(page) != 0)) 1669 bad_reason = "nonzero _count"; 1670 if (unlikely(page->flags & __PG_HWPOISON)) { 1671 bad_reason = "HWPoisoned (hardware-corrupted)"; 1672 bad_flags = __PG_HWPOISON; 1673 /* Don't complain about hwpoisoned pages */ 1674 page_mapcount_reset(page); /* remove PageBuddy */ 1675 return; 1676 } 1677 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) { 1678 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set"; 1679 bad_flags = PAGE_FLAGS_CHECK_AT_PREP; 1680 } 1681 #ifdef CONFIG_MEMCG 1682 if (unlikely(page->mem_cgroup)) 1683 bad_reason = "page still charged to cgroup"; 1684 #endif 1685 bad_page(page, bad_reason, bad_flags); 1686 } 1687 1688 /* 1689 * This page is about to be returned from the page allocator 1690 */ 1691 static inline int check_new_page(struct page *page) 1692 { 1693 if (likely(page_expected_state(page, 1694 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON))) 1695 return 0; 1696 1697 check_new_page_bad(page); 1698 return 1; 1699 } 1700 1701 static inline bool free_pages_prezeroed(bool poisoned) 1702 { 1703 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) && 1704 page_poisoning_enabled() && poisoned; 1705 } 1706 1707 #ifdef CONFIG_DEBUG_VM 1708 static bool check_pcp_refill(struct page *page) 1709 { 1710 return false; 1711 } 1712 1713 static bool check_new_pcp(struct page *page) 1714 { 1715 return check_new_page(page); 1716 } 1717 #else 1718 static bool check_pcp_refill(struct page *page) 1719 { 1720 return check_new_page(page); 1721 } 1722 static bool check_new_pcp(struct page *page) 1723 { 1724 return false; 1725 } 1726 #endif /* CONFIG_DEBUG_VM */ 1727 1728 static bool check_new_pages(struct page *page, unsigned int order) 1729 { 1730 int i; 1731 for (i = 0; i < (1 << order); i++) { 1732 struct page *p = page + i; 1733 1734 if (unlikely(check_new_page(p))) 1735 return true; 1736 } 1737 1738 return false; 1739 } 1740 1741 inline void post_alloc_hook(struct page *page, unsigned int order, 1742 gfp_t gfp_flags) 1743 { 1744 set_page_private(page, 0); 1745 set_page_refcounted(page); 1746 1747 arch_alloc_page(page, order); 1748 kernel_map_pages(page, 1 << order, 1); 1749 kernel_poison_pages(page, 1 << order, 1); 1750 kasan_alloc_pages(page, order); 1751 set_page_owner(page, order, gfp_flags); 1752 } 1753 1754 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags, 1755 unsigned int alloc_flags) 1756 { 1757 int i; 1758 bool poisoned = true; 1759 1760 for (i = 0; i < (1 << order); i++) { 1761 struct page *p = page + i; 1762 if (poisoned) 1763 poisoned &= page_is_poisoned(p); 1764 } 1765 1766 post_alloc_hook(page, order, gfp_flags); 1767 1768 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO)) 1769 for (i = 0; i < (1 << order); i++) 1770 clear_highpage(page + i); 1771 1772 if (order && (gfp_flags & __GFP_COMP)) 1773 prep_compound_page(page, order); 1774 1775 /* 1776 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to 1777 * allocate the page. The expectation is that the caller is taking 1778 * steps that will free more memory. The caller should avoid the page 1779 * being used for !PFMEMALLOC purposes. 1780 */ 1781 if (alloc_flags & ALLOC_NO_WATERMARKS) 1782 set_page_pfmemalloc(page); 1783 else 1784 clear_page_pfmemalloc(page); 1785 } 1786 1787 /* 1788 * Go through the free lists for the given migratetype and remove 1789 * the smallest available page from the freelists 1790 */ 1791 static inline 1792 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order, 1793 int migratetype) 1794 { 1795 unsigned int current_order; 1796 struct free_area *area; 1797 struct page *page; 1798 1799 /* Find a page of the appropriate size in the preferred list */ 1800 for (current_order = order; current_order < MAX_ORDER; ++current_order) { 1801 area = &(zone->free_area[current_order]); 1802 page = list_first_entry_or_null(&area->free_list[migratetype], 1803 struct page, lru); 1804 if (!page) 1805 continue; 1806 list_del(&page->lru); 1807 rmv_page_order(page); 1808 area->nr_free--; 1809 expand(zone, page, order, current_order, area, migratetype); 1810 set_pcppage_migratetype(page, migratetype); 1811 return page; 1812 } 1813 1814 return NULL; 1815 } 1816 1817 1818 /* 1819 * This array describes the order lists are fallen back to when 1820 * the free lists for the desirable migrate type are depleted 1821 */ 1822 static int fallbacks[MIGRATE_TYPES][4] = { 1823 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES }, 1824 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES }, 1825 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES }, 1826 #ifdef CONFIG_CMA 1827 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */ 1828 #endif 1829 #ifdef CONFIG_MEMORY_ISOLATION 1830 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */ 1831 #endif 1832 }; 1833 1834 #ifdef CONFIG_CMA 1835 static struct page *__rmqueue_cma_fallback(struct zone *zone, 1836 unsigned int order) 1837 { 1838 return __rmqueue_smallest(zone, order, MIGRATE_CMA); 1839 } 1840 #else 1841 static inline struct page *__rmqueue_cma_fallback(struct zone *zone, 1842 unsigned int order) { return NULL; } 1843 #endif 1844 1845 /* 1846 * Move the free pages in a range to the free lists of the requested type. 1847 * Note that start_page and end_pages are not aligned on a pageblock 1848 * boundary. If alignment is required, use move_freepages_block() 1849 */ 1850 int move_freepages(struct zone *zone, 1851 struct page *start_page, struct page *end_page, 1852 int migratetype) 1853 { 1854 struct page *page; 1855 unsigned int order; 1856 int pages_moved = 0; 1857 1858 #ifndef CONFIG_HOLES_IN_ZONE 1859 /* 1860 * page_zone is not safe to call in this context when 1861 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant 1862 * anyway as we check zone boundaries in move_freepages_block(). 1863 * Remove at a later date when no bug reports exist related to 1864 * grouping pages by mobility 1865 */ 1866 VM_BUG_ON(page_zone(start_page) != page_zone(end_page)); 1867 #endif 1868 1869 for (page = start_page; page <= end_page;) { 1870 if (!pfn_valid_within(page_to_pfn(page))) { 1871 page++; 1872 continue; 1873 } 1874 1875 /* Make sure we are not inadvertently changing nodes */ 1876 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page); 1877 1878 if (!PageBuddy(page)) { 1879 page++; 1880 continue; 1881 } 1882 1883 order = page_order(page); 1884 list_move(&page->lru, 1885 &zone->free_area[order].free_list[migratetype]); 1886 page += 1 << order; 1887 pages_moved += 1 << order; 1888 } 1889 1890 return pages_moved; 1891 } 1892 1893 int move_freepages_block(struct zone *zone, struct page *page, 1894 int migratetype) 1895 { 1896 unsigned long start_pfn, end_pfn; 1897 struct page *start_page, *end_page; 1898 1899 start_pfn = page_to_pfn(page); 1900 start_pfn = start_pfn & ~(pageblock_nr_pages-1); 1901 start_page = pfn_to_page(start_pfn); 1902 end_page = start_page + pageblock_nr_pages - 1; 1903 end_pfn = start_pfn + pageblock_nr_pages - 1; 1904 1905 /* Do not cross zone boundaries */ 1906 if (!zone_spans_pfn(zone, start_pfn)) 1907 start_page = page; 1908 if (!zone_spans_pfn(zone, end_pfn)) 1909 return 0; 1910 1911 return move_freepages(zone, start_page, end_page, migratetype); 1912 } 1913 1914 static void change_pageblock_range(struct page *pageblock_page, 1915 int start_order, int migratetype) 1916 { 1917 int nr_pageblocks = 1 << (start_order - pageblock_order); 1918 1919 while (nr_pageblocks--) { 1920 set_pageblock_migratetype(pageblock_page, migratetype); 1921 pageblock_page += pageblock_nr_pages; 1922 } 1923 } 1924 1925 /* 1926 * When we are falling back to another migratetype during allocation, try to 1927 * steal extra free pages from the same pageblocks to satisfy further 1928 * allocations, instead of polluting multiple pageblocks. 1929 * 1930 * If we are stealing a relatively large buddy page, it is likely there will 1931 * be more free pages in the pageblock, so try to steal them all. For 1932 * reclaimable and unmovable allocations, we steal regardless of page size, 1933 * as fragmentation caused by those allocations polluting movable pageblocks 1934 * is worse than movable allocations stealing from unmovable and reclaimable 1935 * pageblocks. 1936 */ 1937 static bool can_steal_fallback(unsigned int order, int start_mt) 1938 { 1939 /* 1940 * Leaving this order check is intended, although there is 1941 * relaxed order check in next check. The reason is that 1942 * we can actually steal whole pageblock if this condition met, 1943 * but, below check doesn't guarantee it and that is just heuristic 1944 * so could be changed anytime. 1945 */ 1946 if (order >= pageblock_order) 1947 return true; 1948 1949 if (order >= pageblock_order / 2 || 1950 start_mt == MIGRATE_RECLAIMABLE || 1951 start_mt == MIGRATE_UNMOVABLE || 1952 page_group_by_mobility_disabled) 1953 return true; 1954 1955 return false; 1956 } 1957 1958 /* 1959 * This function implements actual steal behaviour. If order is large enough, 1960 * we can steal whole pageblock. If not, we first move freepages in this 1961 * pageblock and check whether half of pages are moved or not. If half of 1962 * pages are moved, we can change migratetype of pageblock and permanently 1963 * use it's pages as requested migratetype in the future. 1964 */ 1965 static void steal_suitable_fallback(struct zone *zone, struct page *page, 1966 int start_type) 1967 { 1968 unsigned int current_order = page_order(page); 1969 int pages; 1970 1971 /* Take ownership for orders >= pageblock_order */ 1972 if (current_order >= pageblock_order) { 1973 change_pageblock_range(page, current_order, start_type); 1974 return; 1975 } 1976 1977 pages = move_freepages_block(zone, page, start_type); 1978 1979 /* Claim the whole block if over half of it is free */ 1980 if (pages >= (1 << (pageblock_order-1)) || 1981 page_group_by_mobility_disabled) 1982 set_pageblock_migratetype(page, start_type); 1983 } 1984 1985 /* 1986 * Check whether there is a suitable fallback freepage with requested order. 1987 * If only_stealable is true, this function returns fallback_mt only if 1988 * we can steal other freepages all together. This would help to reduce 1989 * fragmentation due to mixed migratetype pages in one pageblock. 1990 */ 1991 int find_suitable_fallback(struct free_area *area, unsigned int order, 1992 int migratetype, bool only_stealable, bool *can_steal) 1993 { 1994 int i; 1995 int fallback_mt; 1996 1997 if (area->nr_free == 0) 1998 return -1; 1999 2000 *can_steal = false; 2001 for (i = 0;; i++) { 2002 fallback_mt = fallbacks[migratetype][i]; 2003 if (fallback_mt == MIGRATE_TYPES) 2004 break; 2005 2006 if (list_empty(&area->free_list[fallback_mt])) 2007 continue; 2008 2009 if (can_steal_fallback(order, migratetype)) 2010 *can_steal = true; 2011 2012 if (!only_stealable) 2013 return fallback_mt; 2014 2015 if (*can_steal) 2016 return fallback_mt; 2017 } 2018 2019 return -1; 2020 } 2021 2022 /* 2023 * Reserve a pageblock for exclusive use of high-order atomic allocations if 2024 * there are no empty page blocks that contain a page with a suitable order 2025 */ 2026 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone, 2027 unsigned int alloc_order) 2028 { 2029 int mt; 2030 unsigned long max_managed, flags; 2031 2032 /* 2033 * Limit the number reserved to 1 pageblock or roughly 1% of a zone. 2034 * Check is race-prone but harmless. 2035 */ 2036 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages; 2037 if (zone->nr_reserved_highatomic >= max_managed) 2038 return; 2039 2040 spin_lock_irqsave(&zone->lock, flags); 2041 2042 /* Recheck the nr_reserved_highatomic limit under the lock */ 2043 if (zone->nr_reserved_highatomic >= max_managed) 2044 goto out_unlock; 2045 2046 /* Yoink! */ 2047 mt = get_pageblock_migratetype(page); 2048 if (mt != MIGRATE_HIGHATOMIC && 2049 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) { 2050 zone->nr_reserved_highatomic += pageblock_nr_pages; 2051 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC); 2052 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC); 2053 } 2054 2055 out_unlock: 2056 spin_unlock_irqrestore(&zone->lock, flags); 2057 } 2058 2059 /* 2060 * Used when an allocation is about to fail under memory pressure. This 2061 * potentially hurts the reliability of high-order allocations when under 2062 * intense memory pressure but failed atomic allocations should be easier 2063 * to recover from than an OOM. 2064 * 2065 * If @force is true, try to unreserve a pageblock even though highatomic 2066 * pageblock is exhausted. 2067 */ 2068 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac, 2069 bool force) 2070 { 2071 struct zonelist *zonelist = ac->zonelist; 2072 unsigned long flags; 2073 struct zoneref *z; 2074 struct zone *zone; 2075 struct page *page; 2076 int order; 2077 bool ret; 2078 2079 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx, 2080 ac->nodemask) { 2081 /* 2082 * Preserve at least one pageblock unless memory pressure 2083 * is really high. 2084 */ 2085 if (!force && zone->nr_reserved_highatomic <= 2086 pageblock_nr_pages) 2087 continue; 2088 2089 spin_lock_irqsave(&zone->lock, flags); 2090 for (order = 0; order < MAX_ORDER; order++) { 2091 struct free_area *area = &(zone->free_area[order]); 2092 2093 page = list_first_entry_or_null( 2094 &area->free_list[MIGRATE_HIGHATOMIC], 2095 struct page, lru); 2096 if (!page) 2097 continue; 2098 2099 /* 2100 * In page freeing path, migratetype change is racy so 2101 * we can counter several free pages in a pageblock 2102 * in this loop althoug we changed the pageblock type 2103 * from highatomic to ac->migratetype. So we should 2104 * adjust the count once. 2105 */ 2106 if (get_pageblock_migratetype(page) == 2107 MIGRATE_HIGHATOMIC) { 2108 /* 2109 * It should never happen but changes to 2110 * locking could inadvertently allow a per-cpu 2111 * drain to add pages to MIGRATE_HIGHATOMIC 2112 * while unreserving so be safe and watch for 2113 * underflows. 2114 */ 2115 zone->nr_reserved_highatomic -= min( 2116 pageblock_nr_pages, 2117 zone->nr_reserved_highatomic); 2118 } 2119 2120 /* 2121 * Convert to ac->migratetype and avoid the normal 2122 * pageblock stealing heuristics. Minimally, the caller 2123 * is doing the work and needs the pages. More 2124 * importantly, if the block was always converted to 2125 * MIGRATE_UNMOVABLE or another type then the number 2126 * of pageblocks that cannot be completely freed 2127 * may increase. 2128 */ 2129 set_pageblock_migratetype(page, ac->migratetype); 2130 ret = move_freepages_block(zone, page, ac->migratetype); 2131 if (ret) { 2132 spin_unlock_irqrestore(&zone->lock, flags); 2133 return ret; 2134 } 2135 } 2136 spin_unlock_irqrestore(&zone->lock, flags); 2137 } 2138 2139 return false; 2140 } 2141 2142 /* Remove an element from the buddy allocator from the fallback list */ 2143 static inline struct page * 2144 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype) 2145 { 2146 struct free_area *area; 2147 unsigned int current_order; 2148 struct page *page; 2149 int fallback_mt; 2150 bool can_steal; 2151 2152 /* Find the largest possible block of pages in the other list */ 2153 for (current_order = MAX_ORDER-1; 2154 current_order >= order && current_order <= MAX_ORDER-1; 2155 --current_order) { 2156 area = &(zone->free_area[current_order]); 2157 fallback_mt = find_suitable_fallback(area, current_order, 2158 start_migratetype, false, &can_steal); 2159 if (fallback_mt == -1) 2160 continue; 2161 2162 page = list_first_entry(&area->free_list[fallback_mt], 2163 struct page, lru); 2164 if (can_steal && 2165 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC) 2166 steal_suitable_fallback(zone, page, start_migratetype); 2167 2168 /* Remove the page from the freelists */ 2169 area->nr_free--; 2170 list_del(&page->lru); 2171 rmv_page_order(page); 2172 2173 expand(zone, page, order, current_order, area, 2174 start_migratetype); 2175 /* 2176 * The pcppage_migratetype may differ from pageblock's 2177 * migratetype depending on the decisions in 2178 * find_suitable_fallback(). This is OK as long as it does not 2179 * differ for MIGRATE_CMA pageblocks. Those can be used as 2180 * fallback only via special __rmqueue_cma_fallback() function 2181 */ 2182 set_pcppage_migratetype(page, start_migratetype); 2183 2184 trace_mm_page_alloc_extfrag(page, order, current_order, 2185 start_migratetype, fallback_mt); 2186 2187 return page; 2188 } 2189 2190 return NULL; 2191 } 2192 2193 /* 2194 * Do the hard work of removing an element from the buddy allocator. 2195 * Call me with the zone->lock already held. 2196 */ 2197 static struct page *__rmqueue(struct zone *zone, unsigned int order, 2198 int migratetype) 2199 { 2200 struct page *page; 2201 2202 page = __rmqueue_smallest(zone, order, migratetype); 2203 if (unlikely(!page)) { 2204 if (migratetype == MIGRATE_MOVABLE) 2205 page = __rmqueue_cma_fallback(zone, order); 2206 2207 if (!page) 2208 page = __rmqueue_fallback(zone, order, migratetype); 2209 } 2210 2211 trace_mm_page_alloc_zone_locked(page, order, migratetype); 2212 return page; 2213 } 2214 2215 /* 2216 * Obtain a specified number of elements from the buddy allocator, all under 2217 * a single hold of the lock, for efficiency. Add them to the supplied list. 2218 * Returns the number of new pages which were placed at *list. 2219 */ 2220 static int rmqueue_bulk(struct zone *zone, unsigned int order, 2221 unsigned long count, struct list_head *list, 2222 int migratetype, bool cold) 2223 { 2224 int i, alloced = 0; 2225 2226 spin_lock(&zone->lock); 2227 for (i = 0; i < count; ++i) { 2228 struct page *page = __rmqueue(zone, order, migratetype); 2229 if (unlikely(page == NULL)) 2230 break; 2231 2232 if (unlikely(check_pcp_refill(page))) 2233 continue; 2234 2235 /* 2236 * Split buddy pages returned by expand() are received here 2237 * in physical page order. The page is added to the callers and 2238 * list and the list head then moves forward. From the callers 2239 * perspective, the linked list is ordered by page number in 2240 * some conditions. This is useful for IO devices that can 2241 * merge IO requests if the physical pages are ordered 2242 * properly. 2243 */ 2244 if (likely(!cold)) 2245 list_add(&page->lru, list); 2246 else 2247 list_add_tail(&page->lru, list); 2248 list = &page->lru; 2249 alloced++; 2250 if (is_migrate_cma(get_pcppage_migratetype(page))) 2251 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 2252 -(1 << order)); 2253 } 2254 2255 /* 2256 * i pages were removed from the buddy list even if some leak due 2257 * to check_pcp_refill failing so adjust NR_FREE_PAGES based 2258 * on i. Do not confuse with 'alloced' which is the number of 2259 * pages added to the pcp list. 2260 */ 2261 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order)); 2262 spin_unlock(&zone->lock); 2263 return alloced; 2264 } 2265 2266 #ifdef CONFIG_NUMA 2267 /* 2268 * Called from the vmstat counter updater to drain pagesets of this 2269 * currently executing processor on remote nodes after they have 2270 * expired. 2271 * 2272 * Note that this function must be called with the thread pinned to 2273 * a single processor. 2274 */ 2275 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp) 2276 { 2277 unsigned long flags; 2278 int to_drain, batch; 2279 2280 local_irq_save(flags); 2281 batch = READ_ONCE(pcp->batch); 2282 to_drain = min(pcp->count, batch); 2283 if (to_drain > 0) { 2284 free_pcppages_bulk(zone, to_drain, pcp); 2285 pcp->count -= to_drain; 2286 } 2287 local_irq_restore(flags); 2288 } 2289 #endif 2290 2291 /* 2292 * Drain pcplists of the indicated processor and zone. 2293 * 2294 * The processor must either be the current processor and the 2295 * thread pinned to the current processor or a processor that 2296 * is not online. 2297 */ 2298 static void drain_pages_zone(unsigned int cpu, struct zone *zone) 2299 { 2300 unsigned long flags; 2301 struct per_cpu_pageset *pset; 2302 struct per_cpu_pages *pcp; 2303 2304 local_irq_save(flags); 2305 pset = per_cpu_ptr(zone->pageset, cpu); 2306 2307 pcp = &pset->pcp; 2308 if (pcp->count) { 2309 free_pcppages_bulk(zone, pcp->count, pcp); 2310 pcp->count = 0; 2311 } 2312 local_irq_restore(flags); 2313 } 2314 2315 /* 2316 * Drain pcplists of all zones on the indicated processor. 2317 * 2318 * The processor must either be the current processor and the 2319 * thread pinned to the current processor or a processor that 2320 * is not online. 2321 */ 2322 static void drain_pages(unsigned int cpu) 2323 { 2324 struct zone *zone; 2325 2326 for_each_populated_zone(zone) { 2327 drain_pages_zone(cpu, zone); 2328 } 2329 } 2330 2331 /* 2332 * Spill all of this CPU's per-cpu pages back into the buddy allocator. 2333 * 2334 * The CPU has to be pinned. When zone parameter is non-NULL, spill just 2335 * the single zone's pages. 2336 */ 2337 void drain_local_pages(struct zone *zone) 2338 { 2339 int cpu = smp_processor_id(); 2340 2341 if (zone) 2342 drain_pages_zone(cpu, zone); 2343 else 2344 drain_pages(cpu); 2345 } 2346 2347 static void drain_local_pages_wq(struct work_struct *work) 2348 { 2349 /* 2350 * drain_all_pages doesn't use proper cpu hotplug protection so 2351 * we can race with cpu offline when the WQ can move this from 2352 * a cpu pinned worker to an unbound one. We can operate on a different 2353 * cpu which is allright but we also have to make sure to not move to 2354 * a different one. 2355 */ 2356 preempt_disable(); 2357 drain_local_pages(NULL); 2358 preempt_enable(); 2359 } 2360 2361 /* 2362 * Spill all the per-cpu pages from all CPUs back into the buddy allocator. 2363 * 2364 * When zone parameter is non-NULL, spill just the single zone's pages. 2365 * 2366 * Note that this can be extremely slow as the draining happens in a workqueue. 2367 */ 2368 void drain_all_pages(struct zone *zone) 2369 { 2370 int cpu; 2371 2372 /* 2373 * Allocate in the BSS so we wont require allocation in 2374 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y 2375 */ 2376 static cpumask_t cpus_with_pcps; 2377 2378 /* 2379 * Make sure nobody triggers this path before mm_percpu_wq is fully 2380 * initialized. 2381 */ 2382 if (WARN_ON_ONCE(!mm_percpu_wq)) 2383 return; 2384 2385 /* Workqueues cannot recurse */ 2386 if (current->flags & PF_WQ_WORKER) 2387 return; 2388 2389 /* 2390 * Do not drain if one is already in progress unless it's specific to 2391 * a zone. Such callers are primarily CMA and memory hotplug and need 2392 * the drain to be complete when the call returns. 2393 */ 2394 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) { 2395 if (!zone) 2396 return; 2397 mutex_lock(&pcpu_drain_mutex); 2398 } 2399 2400 /* 2401 * We don't care about racing with CPU hotplug event 2402 * as offline notification will cause the notified 2403 * cpu to drain that CPU pcps and on_each_cpu_mask 2404 * disables preemption as part of its processing 2405 */ 2406 for_each_online_cpu(cpu) { 2407 struct per_cpu_pageset *pcp; 2408 struct zone *z; 2409 bool has_pcps = false; 2410 2411 if (zone) { 2412 pcp = per_cpu_ptr(zone->pageset, cpu); 2413 if (pcp->pcp.count) 2414 has_pcps = true; 2415 } else { 2416 for_each_populated_zone(z) { 2417 pcp = per_cpu_ptr(z->pageset, cpu); 2418 if (pcp->pcp.count) { 2419 has_pcps = true; 2420 break; 2421 } 2422 } 2423 } 2424 2425 if (has_pcps) 2426 cpumask_set_cpu(cpu, &cpus_with_pcps); 2427 else 2428 cpumask_clear_cpu(cpu, &cpus_with_pcps); 2429 } 2430 2431 for_each_cpu(cpu, &cpus_with_pcps) { 2432 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu); 2433 INIT_WORK(work, drain_local_pages_wq); 2434 queue_work_on(cpu, mm_percpu_wq, work); 2435 } 2436 for_each_cpu(cpu, &cpus_with_pcps) 2437 flush_work(per_cpu_ptr(&pcpu_drain, cpu)); 2438 2439 mutex_unlock(&pcpu_drain_mutex); 2440 } 2441 2442 #ifdef CONFIG_HIBERNATION 2443 2444 void mark_free_pages(struct zone *zone) 2445 { 2446 unsigned long pfn, max_zone_pfn; 2447 unsigned long flags; 2448 unsigned int order, t; 2449 struct page *page; 2450 2451 if (zone_is_empty(zone)) 2452 return; 2453 2454 spin_lock_irqsave(&zone->lock, flags); 2455 2456 max_zone_pfn = zone_end_pfn(zone); 2457 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) 2458 if (pfn_valid(pfn)) { 2459 page = pfn_to_page(pfn); 2460 2461 if (page_zone(page) != zone) 2462 continue; 2463 2464 if (!swsusp_page_is_forbidden(page)) 2465 swsusp_unset_page_free(page); 2466 } 2467 2468 for_each_migratetype_order(order, t) { 2469 list_for_each_entry(page, 2470 &zone->free_area[order].free_list[t], lru) { 2471 unsigned long i; 2472 2473 pfn = page_to_pfn(page); 2474 for (i = 0; i < (1UL << order); i++) 2475 swsusp_set_page_free(pfn_to_page(pfn + i)); 2476 } 2477 } 2478 spin_unlock_irqrestore(&zone->lock, flags); 2479 } 2480 #endif /* CONFIG_PM */ 2481 2482 /* 2483 * Free a 0-order page 2484 * cold == true ? free a cold page : free a hot page 2485 */ 2486 void free_hot_cold_page(struct page *page, bool cold) 2487 { 2488 struct zone *zone = page_zone(page); 2489 struct per_cpu_pages *pcp; 2490 unsigned long flags; 2491 unsigned long pfn = page_to_pfn(page); 2492 int migratetype; 2493 2494 if (!free_pcp_prepare(page)) 2495 return; 2496 2497 migratetype = get_pfnblock_migratetype(page, pfn); 2498 set_pcppage_migratetype(page, migratetype); 2499 local_irq_save(flags); 2500 __count_vm_event(PGFREE); 2501 2502 /* 2503 * We only track unmovable, reclaimable and movable on pcp lists. 2504 * Free ISOLATE pages back to the allocator because they are being 2505 * offlined but treat RESERVE as movable pages so we can get those 2506 * areas back if necessary. Otherwise, we may have to free 2507 * excessively into the page allocator 2508 */ 2509 if (migratetype >= MIGRATE_PCPTYPES) { 2510 if (unlikely(is_migrate_isolate(migratetype))) { 2511 free_one_page(zone, page, pfn, 0, migratetype); 2512 goto out; 2513 } 2514 migratetype = MIGRATE_MOVABLE; 2515 } 2516 2517 pcp = &this_cpu_ptr(zone->pageset)->pcp; 2518 if (!cold) 2519 list_add(&page->lru, &pcp->lists[migratetype]); 2520 else 2521 list_add_tail(&page->lru, &pcp->lists[migratetype]); 2522 pcp->count++; 2523 if (pcp->count >= pcp->high) { 2524 unsigned long batch = READ_ONCE(pcp->batch); 2525 free_pcppages_bulk(zone, batch, pcp); 2526 pcp->count -= batch; 2527 } 2528 2529 out: 2530 local_irq_restore(flags); 2531 } 2532 2533 /* 2534 * Free a list of 0-order pages 2535 */ 2536 void free_hot_cold_page_list(struct list_head *list, bool cold) 2537 { 2538 struct page *page, *next; 2539 2540 list_for_each_entry_safe(page, next, list, lru) { 2541 trace_mm_page_free_batched(page, cold); 2542 free_hot_cold_page(page, cold); 2543 } 2544 } 2545 2546 /* 2547 * split_page takes a non-compound higher-order page, and splits it into 2548 * n (1<<order) sub-pages: page[0..n] 2549 * Each sub-page must be freed individually. 2550 * 2551 * Note: this is probably too low level an operation for use in drivers. 2552 * Please consult with lkml before using this in your driver. 2553 */ 2554 void split_page(struct page *page, unsigned int order) 2555 { 2556 int i; 2557 2558 VM_BUG_ON_PAGE(PageCompound(page), page); 2559 VM_BUG_ON_PAGE(!page_count(page), page); 2560 2561 #ifdef CONFIG_KMEMCHECK 2562 /* 2563 * Split shadow pages too, because free(page[0]) would 2564 * otherwise free the whole shadow. 2565 */ 2566 if (kmemcheck_page_is_tracked(page)) 2567 split_page(virt_to_page(page[0].shadow), order); 2568 #endif 2569 2570 for (i = 1; i < (1 << order); i++) 2571 set_page_refcounted(page + i); 2572 split_page_owner(page, order); 2573 } 2574 EXPORT_SYMBOL_GPL(split_page); 2575 2576 int __isolate_free_page(struct page *page, unsigned int order) 2577 { 2578 unsigned long watermark; 2579 struct zone *zone; 2580 int mt; 2581 2582 BUG_ON(!PageBuddy(page)); 2583 2584 zone = page_zone(page); 2585 mt = get_pageblock_migratetype(page); 2586 2587 if (!is_migrate_isolate(mt)) { 2588 /* 2589 * Obey watermarks as if the page was being allocated. We can 2590 * emulate a high-order watermark check with a raised order-0 2591 * watermark, because we already know our high-order page 2592 * exists. 2593 */ 2594 watermark = min_wmark_pages(zone) + (1UL << order); 2595 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA)) 2596 return 0; 2597 2598 __mod_zone_freepage_state(zone, -(1UL << order), mt); 2599 } 2600 2601 /* Remove page from free list */ 2602 list_del(&page->lru); 2603 zone->free_area[order].nr_free--; 2604 rmv_page_order(page); 2605 2606 /* 2607 * Set the pageblock if the isolated page is at least half of a 2608 * pageblock 2609 */ 2610 if (order >= pageblock_order - 1) { 2611 struct page *endpage = page + (1 << order) - 1; 2612 for (; page < endpage; page += pageblock_nr_pages) { 2613 int mt = get_pageblock_migratetype(page); 2614 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt) 2615 && mt != MIGRATE_HIGHATOMIC) 2616 set_pageblock_migratetype(page, 2617 MIGRATE_MOVABLE); 2618 } 2619 } 2620 2621 2622 return 1UL << order; 2623 } 2624 2625 /* 2626 * Update NUMA hit/miss statistics 2627 * 2628 * Must be called with interrupts disabled. 2629 */ 2630 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z) 2631 { 2632 #ifdef CONFIG_NUMA 2633 enum zone_stat_item local_stat = NUMA_LOCAL; 2634 2635 if (z->node != numa_node_id()) 2636 local_stat = NUMA_OTHER; 2637 2638 if (z->node == preferred_zone->node) 2639 __inc_zone_state(z, NUMA_HIT); 2640 else { 2641 __inc_zone_state(z, NUMA_MISS); 2642 __inc_zone_state(preferred_zone, NUMA_FOREIGN); 2643 } 2644 __inc_zone_state(z, local_stat); 2645 #endif 2646 } 2647 2648 /* Remove page from the per-cpu list, caller must protect the list */ 2649 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype, 2650 bool cold, struct per_cpu_pages *pcp, 2651 struct list_head *list) 2652 { 2653 struct page *page; 2654 2655 do { 2656 if (list_empty(list)) { 2657 pcp->count += rmqueue_bulk(zone, 0, 2658 pcp->batch, list, 2659 migratetype, cold); 2660 if (unlikely(list_empty(list))) 2661 return NULL; 2662 } 2663 2664 if (cold) 2665 page = list_last_entry(list, struct page, lru); 2666 else 2667 page = list_first_entry(list, struct page, lru); 2668 2669 list_del(&page->lru); 2670 pcp->count--; 2671 } while (check_new_pcp(page)); 2672 2673 return page; 2674 } 2675 2676 /* Lock and remove page from the per-cpu list */ 2677 static struct page *rmqueue_pcplist(struct zone *preferred_zone, 2678 struct zone *zone, unsigned int order, 2679 gfp_t gfp_flags, int migratetype) 2680 { 2681 struct per_cpu_pages *pcp; 2682 struct list_head *list; 2683 bool cold = ((gfp_flags & __GFP_COLD) != 0); 2684 struct page *page; 2685 unsigned long flags; 2686 2687 local_irq_save(flags); 2688 pcp = &this_cpu_ptr(zone->pageset)->pcp; 2689 list = &pcp->lists[migratetype]; 2690 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list); 2691 if (page) { 2692 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); 2693 zone_statistics(preferred_zone, zone); 2694 } 2695 local_irq_restore(flags); 2696 return page; 2697 } 2698 2699 /* 2700 * Allocate a page from the given zone. Use pcplists for order-0 allocations. 2701 */ 2702 static inline 2703 struct page *rmqueue(struct zone *preferred_zone, 2704 struct zone *zone, unsigned int order, 2705 gfp_t gfp_flags, unsigned int alloc_flags, 2706 int migratetype) 2707 { 2708 unsigned long flags; 2709 struct page *page; 2710 2711 if (likely(order == 0)) { 2712 page = rmqueue_pcplist(preferred_zone, zone, order, 2713 gfp_flags, migratetype); 2714 goto out; 2715 } 2716 2717 /* 2718 * We most definitely don't want callers attempting to 2719 * allocate greater than order-1 page units with __GFP_NOFAIL. 2720 */ 2721 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1)); 2722 spin_lock_irqsave(&zone->lock, flags); 2723 2724 do { 2725 page = NULL; 2726 if (alloc_flags & ALLOC_HARDER) { 2727 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC); 2728 if (page) 2729 trace_mm_page_alloc_zone_locked(page, order, migratetype); 2730 } 2731 if (!page) 2732 page = __rmqueue(zone, order, migratetype); 2733 } while (page && check_new_pages(page, order)); 2734 spin_unlock(&zone->lock); 2735 if (!page) 2736 goto failed; 2737 __mod_zone_freepage_state(zone, -(1 << order), 2738 get_pcppage_migratetype(page)); 2739 2740 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order); 2741 zone_statistics(preferred_zone, zone); 2742 local_irq_restore(flags); 2743 2744 out: 2745 VM_BUG_ON_PAGE(page && bad_range(zone, page), page); 2746 return page; 2747 2748 failed: 2749 local_irq_restore(flags); 2750 return NULL; 2751 } 2752 2753 #ifdef CONFIG_FAIL_PAGE_ALLOC 2754 2755 static struct { 2756 struct fault_attr attr; 2757 2758 bool ignore_gfp_highmem; 2759 bool ignore_gfp_reclaim; 2760 u32 min_order; 2761 } fail_page_alloc = { 2762 .attr = FAULT_ATTR_INITIALIZER, 2763 .ignore_gfp_reclaim = true, 2764 .ignore_gfp_highmem = true, 2765 .min_order = 1, 2766 }; 2767 2768 static int __init setup_fail_page_alloc(char *str) 2769 { 2770 return setup_fault_attr(&fail_page_alloc.attr, str); 2771 } 2772 __setup("fail_page_alloc=", setup_fail_page_alloc); 2773 2774 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2775 { 2776 if (order < fail_page_alloc.min_order) 2777 return false; 2778 if (gfp_mask & __GFP_NOFAIL) 2779 return false; 2780 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM)) 2781 return false; 2782 if (fail_page_alloc.ignore_gfp_reclaim && 2783 (gfp_mask & __GFP_DIRECT_RECLAIM)) 2784 return false; 2785 2786 return should_fail(&fail_page_alloc.attr, 1 << order); 2787 } 2788 2789 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS 2790 2791 static int __init fail_page_alloc_debugfs(void) 2792 { 2793 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR; 2794 struct dentry *dir; 2795 2796 dir = fault_create_debugfs_attr("fail_page_alloc", NULL, 2797 &fail_page_alloc.attr); 2798 if (IS_ERR(dir)) 2799 return PTR_ERR(dir); 2800 2801 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir, 2802 &fail_page_alloc.ignore_gfp_reclaim)) 2803 goto fail; 2804 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir, 2805 &fail_page_alloc.ignore_gfp_highmem)) 2806 goto fail; 2807 if (!debugfs_create_u32("min-order", mode, dir, 2808 &fail_page_alloc.min_order)) 2809 goto fail; 2810 2811 return 0; 2812 fail: 2813 debugfs_remove_recursive(dir); 2814 2815 return -ENOMEM; 2816 } 2817 2818 late_initcall(fail_page_alloc_debugfs); 2819 2820 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ 2821 2822 #else /* CONFIG_FAIL_PAGE_ALLOC */ 2823 2824 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order) 2825 { 2826 return false; 2827 } 2828 2829 #endif /* CONFIG_FAIL_PAGE_ALLOC */ 2830 2831 /* 2832 * Return true if free base pages are above 'mark'. For high-order checks it 2833 * will return true of the order-0 watermark is reached and there is at least 2834 * one free page of a suitable size. Checking now avoids taking the zone lock 2835 * to check in the allocation paths if no pages are free. 2836 */ 2837 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, 2838 int classzone_idx, unsigned int alloc_flags, 2839 long free_pages) 2840 { 2841 long min = mark; 2842 int o; 2843 const bool alloc_harder = (alloc_flags & ALLOC_HARDER); 2844 2845 /* free_pages may go negative - that's OK */ 2846 free_pages -= (1 << order) - 1; 2847 2848 if (alloc_flags & ALLOC_HIGH) 2849 min -= min / 2; 2850 2851 /* 2852 * If the caller does not have rights to ALLOC_HARDER then subtract 2853 * the high-atomic reserves. This will over-estimate the size of the 2854 * atomic reserve but it avoids a search. 2855 */ 2856 if (likely(!alloc_harder)) 2857 free_pages -= z->nr_reserved_highatomic; 2858 else 2859 min -= min / 4; 2860 2861 #ifdef CONFIG_CMA 2862 /* If allocation can't use CMA areas don't use free CMA pages */ 2863 if (!(alloc_flags & ALLOC_CMA)) 2864 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES); 2865 #endif 2866 2867 /* 2868 * Check watermarks for an order-0 allocation request. If these 2869 * are not met, then a high-order request also cannot go ahead 2870 * even if a suitable page happened to be free. 2871 */ 2872 if (free_pages <= min + z->lowmem_reserve[classzone_idx]) 2873 return false; 2874 2875 /* If this is an order-0 request then the watermark is fine */ 2876 if (!order) 2877 return true; 2878 2879 /* For a high-order request, check at least one suitable page is free */ 2880 for (o = order; o < MAX_ORDER; o++) { 2881 struct free_area *area = &z->free_area[o]; 2882 int mt; 2883 2884 if (!area->nr_free) 2885 continue; 2886 2887 if (alloc_harder) 2888 return true; 2889 2890 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) { 2891 if (!list_empty(&area->free_list[mt])) 2892 return true; 2893 } 2894 2895 #ifdef CONFIG_CMA 2896 if ((alloc_flags & ALLOC_CMA) && 2897 !list_empty(&area->free_list[MIGRATE_CMA])) { 2898 return true; 2899 } 2900 #endif 2901 } 2902 return false; 2903 } 2904 2905 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark, 2906 int classzone_idx, unsigned int alloc_flags) 2907 { 2908 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2909 zone_page_state(z, NR_FREE_PAGES)); 2910 } 2911 2912 static inline bool zone_watermark_fast(struct zone *z, unsigned int order, 2913 unsigned long mark, int classzone_idx, unsigned int alloc_flags) 2914 { 2915 long free_pages = zone_page_state(z, NR_FREE_PAGES); 2916 long cma_pages = 0; 2917 2918 #ifdef CONFIG_CMA 2919 /* If allocation can't use CMA areas don't use free CMA pages */ 2920 if (!(alloc_flags & ALLOC_CMA)) 2921 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES); 2922 #endif 2923 2924 /* 2925 * Fast check for order-0 only. If this fails then the reserves 2926 * need to be calculated. There is a corner case where the check 2927 * passes but only the high-order atomic reserve are free. If 2928 * the caller is !atomic then it'll uselessly search the free 2929 * list. That corner case is then slower but it is harmless. 2930 */ 2931 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx]) 2932 return true; 2933 2934 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags, 2935 free_pages); 2936 } 2937 2938 bool zone_watermark_ok_safe(struct zone *z, unsigned int order, 2939 unsigned long mark, int classzone_idx) 2940 { 2941 long free_pages = zone_page_state(z, NR_FREE_PAGES); 2942 2943 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark) 2944 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES); 2945 2946 return __zone_watermark_ok(z, order, mark, classzone_idx, 0, 2947 free_pages); 2948 } 2949 2950 #ifdef CONFIG_NUMA 2951 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2952 { 2953 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <= 2954 RECLAIM_DISTANCE; 2955 } 2956 #else /* CONFIG_NUMA */ 2957 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone) 2958 { 2959 return true; 2960 } 2961 #endif /* CONFIG_NUMA */ 2962 2963 /* 2964 * get_page_from_freelist goes through the zonelist trying to allocate 2965 * a page. 2966 */ 2967 static struct page * 2968 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags, 2969 const struct alloc_context *ac) 2970 { 2971 struct zoneref *z = ac->preferred_zoneref; 2972 struct zone *zone; 2973 struct pglist_data *last_pgdat_dirty_limit = NULL; 2974 2975 /* 2976 * Scan zonelist, looking for a zone with enough free. 2977 * See also __cpuset_node_allowed() comment in kernel/cpuset.c. 2978 */ 2979 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, 2980 ac->nodemask) { 2981 struct page *page; 2982 unsigned long mark; 2983 2984 if (cpusets_enabled() && 2985 (alloc_flags & ALLOC_CPUSET) && 2986 !__cpuset_zone_allowed(zone, gfp_mask)) 2987 continue; 2988 /* 2989 * When allocating a page cache page for writing, we 2990 * want to get it from a node that is within its dirty 2991 * limit, such that no single node holds more than its 2992 * proportional share of globally allowed dirty pages. 2993 * The dirty limits take into account the node's 2994 * lowmem reserves and high watermark so that kswapd 2995 * should be able to balance it without having to 2996 * write pages from its LRU list. 2997 * 2998 * XXX: For now, allow allocations to potentially 2999 * exceed the per-node dirty limit in the slowpath 3000 * (spread_dirty_pages unset) before going into reclaim, 3001 * which is important when on a NUMA setup the allowed 3002 * nodes are together not big enough to reach the 3003 * global limit. The proper fix for these situations 3004 * will require awareness of nodes in the 3005 * dirty-throttling and the flusher threads. 3006 */ 3007 if (ac->spread_dirty_pages) { 3008 if (last_pgdat_dirty_limit == zone->zone_pgdat) 3009 continue; 3010 3011 if (!node_dirty_ok(zone->zone_pgdat)) { 3012 last_pgdat_dirty_limit = zone->zone_pgdat; 3013 continue; 3014 } 3015 } 3016 3017 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK]; 3018 if (!zone_watermark_fast(zone, order, mark, 3019 ac_classzone_idx(ac), alloc_flags)) { 3020 int ret; 3021 3022 /* Checked here to keep the fast path fast */ 3023 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK); 3024 if (alloc_flags & ALLOC_NO_WATERMARKS) 3025 goto try_this_zone; 3026 3027 if (node_reclaim_mode == 0 || 3028 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone)) 3029 continue; 3030 3031 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order); 3032 switch (ret) { 3033 case NODE_RECLAIM_NOSCAN: 3034 /* did not scan */ 3035 continue; 3036 case NODE_RECLAIM_FULL: 3037 /* scanned but unreclaimable */ 3038 continue; 3039 default: 3040 /* did we reclaim enough */ 3041 if (zone_watermark_ok(zone, order, mark, 3042 ac_classzone_idx(ac), alloc_flags)) 3043 goto try_this_zone; 3044 3045 continue; 3046 } 3047 } 3048 3049 try_this_zone: 3050 page = rmqueue(ac->preferred_zoneref->zone, zone, order, 3051 gfp_mask, alloc_flags, ac->migratetype); 3052 if (page) { 3053 prep_new_page(page, order, gfp_mask, alloc_flags); 3054 3055 /* 3056 * If this is a high-order atomic allocation then check 3057 * if the pageblock should be reserved for the future 3058 */ 3059 if (unlikely(order && (alloc_flags & ALLOC_HARDER))) 3060 reserve_highatomic_pageblock(page, zone, order); 3061 3062 return page; 3063 } 3064 } 3065 3066 return NULL; 3067 } 3068 3069 /* 3070 * Large machines with many possible nodes should not always dump per-node 3071 * meminfo in irq context. 3072 */ 3073 static inline bool should_suppress_show_mem(void) 3074 { 3075 bool ret = false; 3076 3077 #if NODES_SHIFT > 8 3078 ret = in_interrupt(); 3079 #endif 3080 return ret; 3081 } 3082 3083 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask) 3084 { 3085 unsigned int filter = SHOW_MEM_FILTER_NODES; 3086 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1); 3087 3088 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs)) 3089 return; 3090 3091 /* 3092 * This documents exceptions given to allocations in certain 3093 * contexts that are allowed to allocate outside current's set 3094 * of allowed nodes. 3095 */ 3096 if (!(gfp_mask & __GFP_NOMEMALLOC)) 3097 if (test_thread_flag(TIF_MEMDIE) || 3098 (current->flags & (PF_MEMALLOC | PF_EXITING))) 3099 filter &= ~SHOW_MEM_FILTER_NODES; 3100 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM)) 3101 filter &= ~SHOW_MEM_FILTER_NODES; 3102 3103 show_mem(filter, nodemask); 3104 } 3105 3106 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...) 3107 { 3108 struct va_format vaf; 3109 va_list args; 3110 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL, 3111 DEFAULT_RATELIMIT_BURST); 3112 3113 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) || 3114 debug_guardpage_minorder() > 0) 3115 return; 3116 3117 pr_warn("%s: ", current->comm); 3118 3119 va_start(args, fmt); 3120 vaf.fmt = fmt; 3121 vaf.va = &args; 3122 pr_cont("%pV", &vaf); 3123 va_end(args); 3124 3125 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask); 3126 if (nodemask) 3127 pr_cont("%*pbl\n", nodemask_pr_args(nodemask)); 3128 else 3129 pr_cont("(null)\n"); 3130 3131 cpuset_print_current_mems_allowed(); 3132 3133 dump_stack(); 3134 warn_alloc_show_mem(gfp_mask, nodemask); 3135 } 3136 3137 static inline struct page * 3138 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order, 3139 unsigned int alloc_flags, 3140 const struct alloc_context *ac) 3141 { 3142 struct page *page; 3143 3144 page = get_page_from_freelist(gfp_mask, order, 3145 alloc_flags|ALLOC_CPUSET, ac); 3146 /* 3147 * fallback to ignore cpuset restriction if our nodes 3148 * are depleted 3149 */ 3150 if (!page) 3151 page = get_page_from_freelist(gfp_mask, order, 3152 alloc_flags, ac); 3153 3154 return page; 3155 } 3156 3157 static inline struct page * 3158 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order, 3159 const struct alloc_context *ac, unsigned long *did_some_progress) 3160 { 3161 struct oom_control oc = { 3162 .zonelist = ac->zonelist, 3163 .nodemask = ac->nodemask, 3164 .memcg = NULL, 3165 .gfp_mask = gfp_mask, 3166 .order = order, 3167 }; 3168 struct page *page; 3169 3170 *did_some_progress = 0; 3171 3172 /* 3173 * Acquire the oom lock. If that fails, somebody else is 3174 * making progress for us. 3175 */ 3176 if (!mutex_trylock(&oom_lock)) { 3177 *did_some_progress = 1; 3178 schedule_timeout_uninterruptible(1); 3179 return NULL; 3180 } 3181 3182 /* 3183 * Go through the zonelist yet one more time, keep very high watermark 3184 * here, this is only to catch a parallel oom killing, we must fail if 3185 * we're still under heavy pressure. 3186 */ 3187 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order, 3188 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac); 3189 if (page) 3190 goto out; 3191 3192 /* Coredumps can quickly deplete all memory reserves */ 3193 if (current->flags & PF_DUMPCORE) 3194 goto out; 3195 /* The OOM killer will not help higher order allocs */ 3196 if (order > PAGE_ALLOC_COSTLY_ORDER) 3197 goto out; 3198 /* The OOM killer does not needlessly kill tasks for lowmem */ 3199 if (ac->high_zoneidx < ZONE_NORMAL) 3200 goto out; 3201 if (pm_suspended_storage()) 3202 goto out; 3203 /* 3204 * XXX: GFP_NOFS allocations should rather fail than rely on 3205 * other request to make a forward progress. 3206 * We are in an unfortunate situation where out_of_memory cannot 3207 * do much for this context but let's try it to at least get 3208 * access to memory reserved if the current task is killed (see 3209 * out_of_memory). Once filesystems are ready to handle allocation 3210 * failures more gracefully we should just bail out here. 3211 */ 3212 3213 /* The OOM killer may not free memory on a specific node */ 3214 if (gfp_mask & __GFP_THISNODE) 3215 goto out; 3216 3217 /* Exhausted what can be done so it's blamo time */ 3218 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) { 3219 *did_some_progress = 1; 3220 3221 /* 3222 * Help non-failing allocations by giving them access to memory 3223 * reserves 3224 */ 3225 if (gfp_mask & __GFP_NOFAIL) 3226 page = __alloc_pages_cpuset_fallback(gfp_mask, order, 3227 ALLOC_NO_WATERMARKS, ac); 3228 } 3229 out: 3230 mutex_unlock(&oom_lock); 3231 return page; 3232 } 3233 3234 /* 3235 * Maximum number of compaction retries wit a progress before OOM 3236 * killer is consider as the only way to move forward. 3237 */ 3238 #define MAX_COMPACT_RETRIES 16 3239 3240 #ifdef CONFIG_COMPACTION 3241 /* Try memory compaction for high-order allocations before reclaim */ 3242 static struct page * 3243 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 3244 unsigned int alloc_flags, const struct alloc_context *ac, 3245 enum compact_priority prio, enum compact_result *compact_result) 3246 { 3247 struct page *page; 3248 3249 if (!order) 3250 return NULL; 3251 3252 current->flags |= PF_MEMALLOC; 3253 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac, 3254 prio); 3255 current->flags &= ~PF_MEMALLOC; 3256 3257 if (*compact_result <= COMPACT_INACTIVE) 3258 return NULL; 3259 3260 /* 3261 * At least in one zone compaction wasn't deferred or skipped, so let's 3262 * count a compaction stall 3263 */ 3264 count_vm_event(COMPACTSTALL); 3265 3266 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); 3267 3268 if (page) { 3269 struct zone *zone = page_zone(page); 3270 3271 zone->compact_blockskip_flush = false; 3272 compaction_defer_reset(zone, order, true); 3273 count_vm_event(COMPACTSUCCESS); 3274 return page; 3275 } 3276 3277 /* 3278 * It's bad if compaction run occurs and fails. The most likely reason 3279 * is that pages exist, but not enough to satisfy watermarks. 3280 */ 3281 count_vm_event(COMPACTFAIL); 3282 3283 cond_resched(); 3284 3285 return NULL; 3286 } 3287 3288 static inline bool 3289 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, 3290 enum compact_result compact_result, 3291 enum compact_priority *compact_priority, 3292 int *compaction_retries) 3293 { 3294 int max_retries = MAX_COMPACT_RETRIES; 3295 int min_priority; 3296 bool ret = false; 3297 int retries = *compaction_retries; 3298 enum compact_priority priority = *compact_priority; 3299 3300 if (!order) 3301 return false; 3302 3303 if (compaction_made_progress(compact_result)) 3304 (*compaction_retries)++; 3305 3306 /* 3307 * compaction considers all the zone as desperately out of memory 3308 * so it doesn't really make much sense to retry except when the 3309 * failure could be caused by insufficient priority 3310 */ 3311 if (compaction_failed(compact_result)) 3312 goto check_priority; 3313 3314 /* 3315 * make sure the compaction wasn't deferred or didn't bail out early 3316 * due to locks contention before we declare that we should give up. 3317 * But do not retry if the given zonelist is not suitable for 3318 * compaction. 3319 */ 3320 if (compaction_withdrawn(compact_result)) { 3321 ret = compaction_zonelist_suitable(ac, order, alloc_flags); 3322 goto out; 3323 } 3324 3325 /* 3326 * !costly requests are much more important than __GFP_REPEAT 3327 * costly ones because they are de facto nofail and invoke OOM 3328 * killer to move on while costly can fail and users are ready 3329 * to cope with that. 1/4 retries is rather arbitrary but we 3330 * would need much more detailed feedback from compaction to 3331 * make a better decision. 3332 */ 3333 if (order > PAGE_ALLOC_COSTLY_ORDER) 3334 max_retries /= 4; 3335 if (*compaction_retries <= max_retries) { 3336 ret = true; 3337 goto out; 3338 } 3339 3340 /* 3341 * Make sure there are attempts at the highest priority if we exhausted 3342 * all retries or failed at the lower priorities. 3343 */ 3344 check_priority: 3345 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ? 3346 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY; 3347 3348 if (*compact_priority > min_priority) { 3349 (*compact_priority)--; 3350 *compaction_retries = 0; 3351 ret = true; 3352 } 3353 out: 3354 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret); 3355 return ret; 3356 } 3357 #else 3358 static inline struct page * 3359 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order, 3360 unsigned int alloc_flags, const struct alloc_context *ac, 3361 enum compact_priority prio, enum compact_result *compact_result) 3362 { 3363 *compact_result = COMPACT_SKIPPED; 3364 return NULL; 3365 } 3366 3367 static inline bool 3368 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags, 3369 enum compact_result compact_result, 3370 enum compact_priority *compact_priority, 3371 int *compaction_retries) 3372 { 3373 struct zone *zone; 3374 struct zoneref *z; 3375 3376 if (!order || order > PAGE_ALLOC_COSTLY_ORDER) 3377 return false; 3378 3379 /* 3380 * There are setups with compaction disabled which would prefer to loop 3381 * inside the allocator rather than hit the oom killer prematurely. 3382 * Let's give them a good hope and keep retrying while the order-0 3383 * watermarks are OK. 3384 */ 3385 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, 3386 ac->nodemask) { 3387 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone), 3388 ac_classzone_idx(ac), alloc_flags)) 3389 return true; 3390 } 3391 return false; 3392 } 3393 #endif /* CONFIG_COMPACTION */ 3394 3395 /* Perform direct synchronous page reclaim */ 3396 static int 3397 __perform_reclaim(gfp_t gfp_mask, unsigned int order, 3398 const struct alloc_context *ac) 3399 { 3400 struct reclaim_state reclaim_state; 3401 int progress; 3402 3403 cond_resched(); 3404 3405 /* We now go into synchronous reclaim */ 3406 cpuset_memory_pressure_bump(); 3407 current->flags |= PF_MEMALLOC; 3408 lockdep_set_current_reclaim_state(gfp_mask); 3409 reclaim_state.reclaimed_slab = 0; 3410 current->reclaim_state = &reclaim_state; 3411 3412 progress = try_to_free_pages(ac->zonelist, order, gfp_mask, 3413 ac->nodemask); 3414 3415 current->reclaim_state = NULL; 3416 lockdep_clear_current_reclaim_state(); 3417 current->flags &= ~PF_MEMALLOC; 3418 3419 cond_resched(); 3420 3421 return progress; 3422 } 3423 3424 /* The really slow allocator path where we enter direct reclaim */ 3425 static inline struct page * 3426 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order, 3427 unsigned int alloc_flags, const struct alloc_context *ac, 3428 unsigned long *did_some_progress) 3429 { 3430 struct page *page = NULL; 3431 bool drained = false; 3432 3433 *did_some_progress = __perform_reclaim(gfp_mask, order, ac); 3434 if (unlikely(!(*did_some_progress))) 3435 return NULL; 3436 3437 retry: 3438 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); 3439 3440 /* 3441 * If an allocation failed after direct reclaim, it could be because 3442 * pages are pinned on the per-cpu lists or in high alloc reserves. 3443 * Shrink them them and try again 3444 */ 3445 if (!page && !drained) { 3446 unreserve_highatomic_pageblock(ac, false); 3447 drain_all_pages(NULL); 3448 drained = true; 3449 goto retry; 3450 } 3451 3452 return page; 3453 } 3454 3455 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac) 3456 { 3457 struct zoneref *z; 3458 struct zone *zone; 3459 pg_data_t *last_pgdat = NULL; 3460 3461 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, 3462 ac->high_zoneidx, ac->nodemask) { 3463 if (last_pgdat != zone->zone_pgdat) 3464 wakeup_kswapd(zone, order, ac->high_zoneidx); 3465 last_pgdat = zone->zone_pgdat; 3466 } 3467 } 3468 3469 static inline unsigned int 3470 gfp_to_alloc_flags(gfp_t gfp_mask) 3471 { 3472 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET; 3473 3474 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */ 3475 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH); 3476 3477 /* 3478 * The caller may dip into page reserves a bit more if the caller 3479 * cannot run direct reclaim, or if the caller has realtime scheduling 3480 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will 3481 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH). 3482 */ 3483 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH); 3484 3485 if (gfp_mask & __GFP_ATOMIC) { 3486 /* 3487 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even 3488 * if it can't schedule. 3489 */ 3490 if (!(gfp_mask & __GFP_NOMEMALLOC)) 3491 alloc_flags |= ALLOC_HARDER; 3492 /* 3493 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the 3494 * comment for __cpuset_node_allowed(). 3495 */ 3496 alloc_flags &= ~ALLOC_CPUSET; 3497 } else if (unlikely(rt_task(current)) && !in_interrupt()) 3498 alloc_flags |= ALLOC_HARDER; 3499 3500 #ifdef CONFIG_CMA 3501 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE) 3502 alloc_flags |= ALLOC_CMA; 3503 #endif 3504 return alloc_flags; 3505 } 3506 3507 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask) 3508 { 3509 if (unlikely(gfp_mask & __GFP_NOMEMALLOC)) 3510 return false; 3511 3512 if (gfp_mask & __GFP_MEMALLOC) 3513 return true; 3514 if (in_serving_softirq() && (current->flags & PF_MEMALLOC)) 3515 return true; 3516 if (!in_interrupt() && 3517 ((current->flags & PF_MEMALLOC) || 3518 unlikely(test_thread_flag(TIF_MEMDIE)))) 3519 return true; 3520 3521 return false; 3522 } 3523 3524 /* 3525 * Maximum number of reclaim retries without any progress before OOM killer 3526 * is consider as the only way to move forward. 3527 */ 3528 #define MAX_RECLAIM_RETRIES 16 3529 3530 /* 3531 * Checks whether it makes sense to retry the reclaim to make a forward progress 3532 * for the given allocation request. 3533 * The reclaim feedback represented by did_some_progress (any progress during 3534 * the last reclaim round) and no_progress_loops (number of reclaim rounds without 3535 * any progress in a row) is considered as well as the reclaimable pages on the 3536 * applicable zone list (with a backoff mechanism which is a function of 3537 * no_progress_loops). 3538 * 3539 * Returns true if a retry is viable or false to enter the oom path. 3540 */ 3541 static inline bool 3542 should_reclaim_retry(gfp_t gfp_mask, unsigned order, 3543 struct alloc_context *ac, int alloc_flags, 3544 bool did_some_progress, int *no_progress_loops) 3545 { 3546 struct zone *zone; 3547 struct zoneref *z; 3548 3549 /* 3550 * Costly allocations might have made a progress but this doesn't mean 3551 * their order will become available due to high fragmentation so 3552 * always increment the no progress counter for them 3553 */ 3554 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) 3555 *no_progress_loops = 0; 3556 else 3557 (*no_progress_loops)++; 3558 3559 /* 3560 * Make sure we converge to OOM if we cannot make any progress 3561 * several times in the row. 3562 */ 3563 if (*no_progress_loops > MAX_RECLAIM_RETRIES) { 3564 /* Before OOM, exhaust highatomic_reserve */ 3565 return unreserve_highatomic_pageblock(ac, true); 3566 } 3567 3568 /* 3569 * Keep reclaiming pages while there is a chance this will lead 3570 * somewhere. If none of the target zones can satisfy our allocation 3571 * request even if all reclaimable pages are considered then we are 3572 * screwed and have to go OOM. 3573 */ 3574 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx, 3575 ac->nodemask) { 3576 unsigned long available; 3577 unsigned long reclaimable; 3578 unsigned long min_wmark = min_wmark_pages(zone); 3579 bool wmark; 3580 3581 available = reclaimable = zone_reclaimable_pages(zone); 3582 available -= DIV_ROUND_UP((*no_progress_loops) * available, 3583 MAX_RECLAIM_RETRIES); 3584 available += zone_page_state_snapshot(zone, NR_FREE_PAGES); 3585 3586 /* 3587 * Would the allocation succeed if we reclaimed the whole 3588 * available? 3589 */ 3590 wmark = __zone_watermark_ok(zone, order, min_wmark, 3591 ac_classzone_idx(ac), alloc_flags, available); 3592 trace_reclaim_retry_zone(z, order, reclaimable, 3593 available, min_wmark, *no_progress_loops, wmark); 3594 if (wmark) { 3595 /* 3596 * If we didn't make any progress and have a lot of 3597 * dirty + writeback pages then we should wait for 3598 * an IO to complete to slow down the reclaim and 3599 * prevent from pre mature OOM 3600 */ 3601 if (!did_some_progress) { 3602 unsigned long write_pending; 3603 3604 write_pending = zone_page_state_snapshot(zone, 3605 NR_ZONE_WRITE_PENDING); 3606 3607 if (2 * write_pending > reclaimable) { 3608 congestion_wait(BLK_RW_ASYNC, HZ/10); 3609 return true; 3610 } 3611 } 3612 3613 /* 3614 * Memory allocation/reclaim might be called from a WQ 3615 * context and the current implementation of the WQ 3616 * concurrency control doesn't recognize that 3617 * a particular WQ is congested if the worker thread is 3618 * looping without ever sleeping. Therefore we have to 3619 * do a short sleep here rather than calling 3620 * cond_resched(). 3621 */ 3622 if (current->flags & PF_WQ_WORKER) 3623 schedule_timeout_uninterruptible(1); 3624 else 3625 cond_resched(); 3626 3627 return true; 3628 } 3629 } 3630 3631 return false; 3632 } 3633 3634 static inline struct page * 3635 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order, 3636 struct alloc_context *ac) 3637 { 3638 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM; 3639 struct page *page = NULL; 3640 unsigned int alloc_flags; 3641 unsigned long did_some_progress; 3642 enum compact_priority compact_priority; 3643 enum compact_result compact_result; 3644 int compaction_retries; 3645 int no_progress_loops; 3646 unsigned long alloc_start = jiffies; 3647 unsigned int stall_timeout = 10 * HZ; 3648 unsigned int cpuset_mems_cookie; 3649 3650 /* 3651 * In the slowpath, we sanity check order to avoid ever trying to 3652 * reclaim >= MAX_ORDER areas which will never succeed. Callers may 3653 * be using allocators in order of preference for an area that is 3654 * too large. 3655 */ 3656 if (order >= MAX_ORDER) { 3657 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN)); 3658 return NULL; 3659 } 3660 3661 /* 3662 * We also sanity check to catch abuse of atomic reserves being used by 3663 * callers that are not in atomic context. 3664 */ 3665 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) == 3666 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM))) 3667 gfp_mask &= ~__GFP_ATOMIC; 3668 3669 retry_cpuset: 3670 compaction_retries = 0; 3671 no_progress_loops = 0; 3672 compact_priority = DEF_COMPACT_PRIORITY; 3673 cpuset_mems_cookie = read_mems_allowed_begin(); 3674 3675 /* 3676 * The fast path uses conservative alloc_flags to succeed only until 3677 * kswapd needs to be woken up, and to avoid the cost of setting up 3678 * alloc_flags precisely. So we do that now. 3679 */ 3680 alloc_flags = gfp_to_alloc_flags(gfp_mask); 3681 3682 /* 3683 * We need to recalculate the starting point for the zonelist iterator 3684 * because we might have used different nodemask in the fast path, or 3685 * there was a cpuset modification and we are retrying - otherwise we 3686 * could end up iterating over non-eligible zones endlessly. 3687 */ 3688 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, 3689 ac->high_zoneidx, ac->nodemask); 3690 if (!ac->preferred_zoneref->zone) 3691 goto nopage; 3692 3693 if (gfp_mask & __GFP_KSWAPD_RECLAIM) 3694 wake_all_kswapds(order, ac); 3695 3696 /* 3697 * The adjusted alloc_flags might result in immediate success, so try 3698 * that first 3699 */ 3700 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); 3701 if (page) 3702 goto got_pg; 3703 3704 /* 3705 * For costly allocations, try direct compaction first, as it's likely 3706 * that we have enough base pages and don't need to reclaim. Don't try 3707 * that for allocations that are allowed to ignore watermarks, as the 3708 * ALLOC_NO_WATERMARKS attempt didn't yet happen. 3709 */ 3710 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER && 3711 !gfp_pfmemalloc_allowed(gfp_mask)) { 3712 page = __alloc_pages_direct_compact(gfp_mask, order, 3713 alloc_flags, ac, 3714 INIT_COMPACT_PRIORITY, 3715 &compact_result); 3716 if (page) 3717 goto got_pg; 3718 3719 /* 3720 * Checks for costly allocations with __GFP_NORETRY, which 3721 * includes THP page fault allocations 3722 */ 3723 if (gfp_mask & __GFP_NORETRY) { 3724 /* 3725 * If compaction is deferred for high-order allocations, 3726 * it is because sync compaction recently failed. If 3727 * this is the case and the caller requested a THP 3728 * allocation, we do not want to heavily disrupt the 3729 * system, so we fail the allocation instead of entering 3730 * direct reclaim. 3731 */ 3732 if (compact_result == COMPACT_DEFERRED) 3733 goto nopage; 3734 3735 /* 3736 * Looks like reclaim/compaction is worth trying, but 3737 * sync compaction could be very expensive, so keep 3738 * using async compaction. 3739 */ 3740 compact_priority = INIT_COMPACT_PRIORITY; 3741 } 3742 } 3743 3744 retry: 3745 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */ 3746 if (gfp_mask & __GFP_KSWAPD_RECLAIM) 3747 wake_all_kswapds(order, ac); 3748 3749 if (gfp_pfmemalloc_allowed(gfp_mask)) 3750 alloc_flags = ALLOC_NO_WATERMARKS; 3751 3752 /* 3753 * Reset the zonelist iterators if memory policies can be ignored. 3754 * These allocations are high priority and system rather than user 3755 * orientated. 3756 */ 3757 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) { 3758 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask); 3759 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, 3760 ac->high_zoneidx, ac->nodemask); 3761 } 3762 3763 /* Attempt with potentially adjusted zonelist and alloc_flags */ 3764 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac); 3765 if (page) 3766 goto got_pg; 3767 3768 /* Caller is not willing to reclaim, we can't balance anything */ 3769 if (!can_direct_reclaim) 3770 goto nopage; 3771 3772 /* Make sure we know about allocations which stall for too long */ 3773 if (time_after(jiffies, alloc_start + stall_timeout)) { 3774 warn_alloc(gfp_mask, ac->nodemask, 3775 "page allocation stalls for %ums, order:%u", 3776 jiffies_to_msecs(jiffies-alloc_start), order); 3777 stall_timeout += 10 * HZ; 3778 } 3779 3780 /* Avoid recursion of direct reclaim */ 3781 if (current->flags & PF_MEMALLOC) 3782 goto nopage; 3783 3784 /* Try direct reclaim and then allocating */ 3785 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac, 3786 &did_some_progress); 3787 if (page) 3788 goto got_pg; 3789 3790 /* Try direct compaction and then allocating */ 3791 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac, 3792 compact_priority, &compact_result); 3793 if (page) 3794 goto got_pg; 3795 3796 /* Do not loop if specifically requested */ 3797 if (gfp_mask & __GFP_NORETRY) 3798 goto nopage; 3799 3800 /* 3801 * Do not retry costly high order allocations unless they are 3802 * __GFP_REPEAT 3803 */ 3804 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT)) 3805 goto nopage; 3806 3807 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags, 3808 did_some_progress > 0, &no_progress_loops)) 3809 goto retry; 3810 3811 /* 3812 * It doesn't make any sense to retry for the compaction if the order-0 3813 * reclaim is not able to make any progress because the current 3814 * implementation of the compaction depends on the sufficient amount 3815 * of free memory (see __compaction_suitable) 3816 */ 3817 if (did_some_progress > 0 && 3818 should_compact_retry(ac, order, alloc_flags, 3819 compact_result, &compact_priority, 3820 &compaction_retries)) 3821 goto retry; 3822 3823 /* 3824 * It's possible we raced with cpuset update so the OOM would be 3825 * premature (see below the nopage: label for full explanation). 3826 */ 3827 if (read_mems_allowed_retry(cpuset_mems_cookie)) 3828 goto retry_cpuset; 3829 3830 /* Reclaim has failed us, start killing things */ 3831 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress); 3832 if (page) 3833 goto got_pg; 3834 3835 /* Avoid allocations with no watermarks from looping endlessly */ 3836 if (test_thread_flag(TIF_MEMDIE)) 3837 goto nopage; 3838 3839 /* Retry as long as the OOM killer is making progress */ 3840 if (did_some_progress) { 3841 no_progress_loops = 0; 3842 goto retry; 3843 } 3844 3845 nopage: 3846 /* 3847 * When updating a task's mems_allowed or mempolicy nodemask, it is 3848 * possible to race with parallel threads in such a way that our 3849 * allocation can fail while the mask is being updated. If we are about 3850 * to fail, check if the cpuset changed during allocation and if so, 3851 * retry. 3852 */ 3853 if (read_mems_allowed_retry(cpuset_mems_cookie)) 3854 goto retry_cpuset; 3855 3856 /* 3857 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure 3858 * we always retry 3859 */ 3860 if (gfp_mask & __GFP_NOFAIL) { 3861 /* 3862 * All existing users of the __GFP_NOFAIL are blockable, so warn 3863 * of any new users that actually require GFP_NOWAIT 3864 */ 3865 if (WARN_ON_ONCE(!can_direct_reclaim)) 3866 goto fail; 3867 3868 /* 3869 * PF_MEMALLOC request from this context is rather bizarre 3870 * because we cannot reclaim anything and only can loop waiting 3871 * for somebody to do a work for us 3872 */ 3873 WARN_ON_ONCE(current->flags & PF_MEMALLOC); 3874 3875 /* 3876 * non failing costly orders are a hard requirement which we 3877 * are not prepared for much so let's warn about these users 3878 * so that we can identify them and convert them to something 3879 * else. 3880 */ 3881 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER); 3882 3883 /* 3884 * Help non-failing allocations by giving them access to memory 3885 * reserves but do not use ALLOC_NO_WATERMARKS because this 3886 * could deplete whole memory reserves which would just make 3887 * the situation worse 3888 */ 3889 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac); 3890 if (page) 3891 goto got_pg; 3892 3893 cond_resched(); 3894 goto retry; 3895 } 3896 fail: 3897 warn_alloc(gfp_mask, ac->nodemask, 3898 "page allocation failure: order:%u", order); 3899 got_pg: 3900 return page; 3901 } 3902 3903 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order, 3904 struct zonelist *zonelist, nodemask_t *nodemask, 3905 struct alloc_context *ac, gfp_t *alloc_mask, 3906 unsigned int *alloc_flags) 3907 { 3908 ac->high_zoneidx = gfp_zone(gfp_mask); 3909 ac->zonelist = zonelist; 3910 ac->nodemask = nodemask; 3911 ac->migratetype = gfpflags_to_migratetype(gfp_mask); 3912 3913 if (cpusets_enabled()) { 3914 *alloc_mask |= __GFP_HARDWALL; 3915 if (!ac->nodemask) 3916 ac->nodemask = &cpuset_current_mems_allowed; 3917 else 3918 *alloc_flags |= ALLOC_CPUSET; 3919 } 3920 3921 lockdep_trace_alloc(gfp_mask); 3922 3923 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM); 3924 3925 if (should_fail_alloc_page(gfp_mask, order)) 3926 return false; 3927 3928 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE) 3929 *alloc_flags |= ALLOC_CMA; 3930 3931 return true; 3932 } 3933 3934 /* Determine whether to spread dirty pages and what the first usable zone */ 3935 static inline void finalise_ac(gfp_t gfp_mask, 3936 unsigned int order, struct alloc_context *ac) 3937 { 3938 /* Dirty zone balancing only done in the fast path */ 3939 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE); 3940 3941 /* 3942 * The preferred zone is used for statistics but crucially it is 3943 * also used as the starting point for the zonelist iterator. It 3944 * may get reset for allocations that ignore memory policies. 3945 */ 3946 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist, 3947 ac->high_zoneidx, ac->nodemask); 3948 } 3949 3950 /* 3951 * This is the 'heart' of the zoned buddy allocator. 3952 */ 3953 struct page * 3954 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, 3955 struct zonelist *zonelist, nodemask_t *nodemask) 3956 { 3957 struct page *page; 3958 unsigned int alloc_flags = ALLOC_WMARK_LOW; 3959 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */ 3960 struct alloc_context ac = { }; 3961 3962 gfp_mask &= gfp_allowed_mask; 3963 if (!prepare_alloc_pages(gfp_mask, order, zonelist, nodemask, &ac, &alloc_mask, &alloc_flags)) 3964 return NULL; 3965 3966 finalise_ac(gfp_mask, order, &ac); 3967 3968 /* First allocation attempt */ 3969 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac); 3970 if (likely(page)) 3971 goto out; 3972 3973 /* 3974 * Runtime PM, block IO and its error handling path can deadlock 3975 * because I/O on the device might not complete. 3976 */ 3977 alloc_mask = memalloc_noio_flags(gfp_mask); 3978 ac.spread_dirty_pages = false; 3979 3980 /* 3981 * Restore the original nodemask if it was potentially replaced with 3982 * &cpuset_current_mems_allowed to optimize the fast-path attempt. 3983 */ 3984 if (unlikely(ac.nodemask != nodemask)) 3985 ac.nodemask = nodemask; 3986 3987 page = __alloc_pages_slowpath(alloc_mask, order, &ac); 3988 3989 out: 3990 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page && 3991 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) { 3992 __free_pages(page, order); 3993 page = NULL; 3994 } 3995 3996 if (kmemcheck_enabled && page) 3997 kmemcheck_pagealloc_alloc(page, order, gfp_mask); 3998 3999 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype); 4000 4001 return page; 4002 } 4003 EXPORT_SYMBOL(__alloc_pages_nodemask); 4004 4005 /* 4006 * Common helper functions. 4007 */ 4008 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order) 4009 { 4010 struct page *page; 4011 4012 /* 4013 * __get_free_pages() returns a 32-bit address, which cannot represent 4014 * a highmem page 4015 */ 4016 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0); 4017 4018 page = alloc_pages(gfp_mask, order); 4019 if (!page) 4020 return 0; 4021 return (unsigned long) page_address(page); 4022 } 4023 EXPORT_SYMBOL(__get_free_pages); 4024 4025 unsigned long get_zeroed_page(gfp_t gfp_mask) 4026 { 4027 return __get_free_pages(gfp_mask | __GFP_ZERO, 0); 4028 } 4029 EXPORT_SYMBOL(get_zeroed_page); 4030 4031 void __free_pages(struct page *page, unsigned int order) 4032 { 4033 if (put_page_testzero(page)) { 4034 if (order == 0) 4035 free_hot_cold_page(page, false); 4036 else 4037 __free_pages_ok(page, order); 4038 } 4039 } 4040 4041 EXPORT_SYMBOL(__free_pages); 4042 4043 void free_pages(unsigned long addr, unsigned int order) 4044 { 4045 if (addr != 0) { 4046 VM_BUG_ON(!virt_addr_valid((void *)addr)); 4047 __free_pages(virt_to_page((void *)addr), order); 4048 } 4049 } 4050 4051 EXPORT_SYMBOL(free_pages); 4052 4053 /* 4054 * Page Fragment: 4055 * An arbitrary-length arbitrary-offset area of memory which resides 4056 * within a 0 or higher order page. Multiple fragments within that page 4057 * are individually refcounted, in the page's reference counter. 4058 * 4059 * The page_frag functions below provide a simple allocation framework for 4060 * page fragments. This is used by the network stack and network device 4061 * drivers to provide a backing region of memory for use as either an 4062 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info. 4063 */ 4064 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc, 4065 gfp_t gfp_mask) 4066 { 4067 struct page *page = NULL; 4068 gfp_t gfp = gfp_mask; 4069 4070 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 4071 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY | 4072 __GFP_NOMEMALLOC; 4073 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask, 4074 PAGE_FRAG_CACHE_MAX_ORDER); 4075 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE; 4076 #endif 4077 if (unlikely(!page)) 4078 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0); 4079 4080 nc->va = page ? page_address(page) : NULL; 4081 4082 return page; 4083 } 4084 4085 void __page_frag_cache_drain(struct page *page, unsigned int count) 4086 { 4087 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page); 4088 4089 if (page_ref_sub_and_test(page, count)) { 4090 unsigned int order = compound_order(page); 4091 4092 if (order == 0) 4093 free_hot_cold_page(page, false); 4094 else 4095 __free_pages_ok(page, order); 4096 } 4097 } 4098 EXPORT_SYMBOL(__page_frag_cache_drain); 4099 4100 void *page_frag_alloc(struct page_frag_cache *nc, 4101 unsigned int fragsz, gfp_t gfp_mask) 4102 { 4103 unsigned int size = PAGE_SIZE; 4104 struct page *page; 4105 int offset; 4106 4107 if (unlikely(!nc->va)) { 4108 refill: 4109 page = __page_frag_cache_refill(nc, gfp_mask); 4110 if (!page) 4111 return NULL; 4112 4113 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 4114 /* if size can vary use size else just use PAGE_SIZE */ 4115 size = nc->size; 4116 #endif 4117 /* Even if we own the page, we do not use atomic_set(). 4118 * This would break get_page_unless_zero() users. 4119 */ 4120 page_ref_add(page, size - 1); 4121 4122 /* reset page count bias and offset to start of new frag */ 4123 nc->pfmemalloc = page_is_pfmemalloc(page); 4124 nc->pagecnt_bias = size; 4125 nc->offset = size; 4126 } 4127 4128 offset = nc->offset - fragsz; 4129 if (unlikely(offset < 0)) { 4130 page = virt_to_page(nc->va); 4131 4132 if (!page_ref_sub_and_test(page, nc->pagecnt_bias)) 4133 goto refill; 4134 4135 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE) 4136 /* if size can vary use size else just use PAGE_SIZE */ 4137 size = nc->size; 4138 #endif 4139 /* OK, page count is 0, we can safely set it */ 4140 set_page_count(page, size); 4141 4142 /* reset page count bias and offset to start of new frag */ 4143 nc->pagecnt_bias = size; 4144 offset = size - fragsz; 4145 } 4146 4147 nc->pagecnt_bias--; 4148 nc->offset = offset; 4149 4150 return nc->va + offset; 4151 } 4152 EXPORT_SYMBOL(page_frag_alloc); 4153 4154 /* 4155 * Frees a page fragment allocated out of either a compound or order 0 page. 4156 */ 4157 void page_frag_free(void *addr) 4158 { 4159 struct page *page = virt_to_head_page(addr); 4160 4161 if (unlikely(put_page_testzero(page))) 4162 __free_pages_ok(page, compound_order(page)); 4163 } 4164 EXPORT_SYMBOL(page_frag_free); 4165 4166 static void *make_alloc_exact(unsigned long addr, unsigned int order, 4167 size_t size) 4168 { 4169 if (addr) { 4170 unsigned long alloc_end = addr + (PAGE_SIZE << order); 4171 unsigned long used = addr + PAGE_ALIGN(size); 4172 4173 split_page(virt_to_page((void *)addr), order); 4174 while (used < alloc_end) { 4175 free_page(used); 4176 used += PAGE_SIZE; 4177 } 4178 } 4179 return (void *)addr; 4180 } 4181 4182 /** 4183 * alloc_pages_exact - allocate an exact number physically-contiguous pages. 4184 * @size: the number of bytes to allocate 4185 * @gfp_mask: GFP flags for the allocation 4186 * 4187 * This function is similar to alloc_pages(), except that it allocates the 4188 * minimum number of pages to satisfy the request. alloc_pages() can only 4189 * allocate memory in power-of-two pages. 4190 * 4191 * This function is also limited by MAX_ORDER. 4192 * 4193 * Memory allocated by this function must be released by free_pages_exact(). 4194 */ 4195 void *alloc_pages_exact(size_t size, gfp_t gfp_mask) 4196 { 4197 unsigned int order = get_order(size); 4198 unsigned long addr; 4199 4200 addr = __get_free_pages(gfp_mask, order); 4201 return make_alloc_exact(addr, order, size); 4202 } 4203 EXPORT_SYMBOL(alloc_pages_exact); 4204 4205 /** 4206 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous 4207 * pages on a node. 4208 * @nid: the preferred node ID where memory should be allocated 4209 * @size: the number of bytes to allocate 4210 * @gfp_mask: GFP flags for the allocation 4211 * 4212 * Like alloc_pages_exact(), but try to allocate on node nid first before falling 4213 * back. 4214 */ 4215 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask) 4216 { 4217 unsigned int order = get_order(size); 4218 struct page *p = alloc_pages_node(nid, gfp_mask, order); 4219 if (!p) 4220 return NULL; 4221 return make_alloc_exact((unsigned long)page_address(p), order, size); 4222 } 4223 4224 /** 4225 * free_pages_exact - release memory allocated via alloc_pages_exact() 4226 * @virt: the value returned by alloc_pages_exact. 4227 * @size: size of allocation, same value as passed to alloc_pages_exact(). 4228 * 4229 * Release the memory allocated by a previous call to alloc_pages_exact. 4230 */ 4231 void free_pages_exact(void *virt, size_t size) 4232 { 4233 unsigned long addr = (unsigned long)virt; 4234 unsigned long end = addr + PAGE_ALIGN(size); 4235 4236 while (addr < end) { 4237 free_page(addr); 4238 addr += PAGE_SIZE; 4239 } 4240 } 4241 EXPORT_SYMBOL(free_pages_exact); 4242 4243 /** 4244 * nr_free_zone_pages - count number of pages beyond high watermark 4245 * @offset: The zone index of the highest zone 4246 * 4247 * nr_free_zone_pages() counts the number of counts pages which are beyond the 4248 * high watermark within all zones at or below a given zone index. For each 4249 * zone, the number of pages is calculated as: 4250 * managed_pages - high_pages 4251 */ 4252 static unsigned long nr_free_zone_pages(int offset) 4253 { 4254 struct zoneref *z; 4255 struct zone *zone; 4256 4257 /* Just pick one node, since fallback list is circular */ 4258 unsigned long sum = 0; 4259 4260 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL); 4261 4262 for_each_zone_zonelist(zone, z, zonelist, offset) { 4263 unsigned long size = zone->managed_pages; 4264 unsigned long high = high_wmark_pages(zone); 4265 if (size > high) 4266 sum += size - high; 4267 } 4268 4269 return sum; 4270 } 4271 4272 /** 4273 * nr_free_buffer_pages - count number of pages beyond high watermark 4274 * 4275 * nr_free_buffer_pages() counts the number of pages which are beyond the high 4276 * watermark within ZONE_DMA and ZONE_NORMAL. 4277 */ 4278 unsigned long nr_free_buffer_pages(void) 4279 { 4280 return nr_free_zone_pages(gfp_zone(GFP_USER)); 4281 } 4282 EXPORT_SYMBOL_GPL(nr_free_buffer_pages); 4283 4284 /** 4285 * nr_free_pagecache_pages - count number of pages beyond high watermark 4286 * 4287 * nr_free_pagecache_pages() counts the number of pages which are beyond the 4288 * high watermark within all zones. 4289 */ 4290 unsigned long nr_free_pagecache_pages(void) 4291 { 4292 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE)); 4293 } 4294 4295 static inline void show_node(struct zone *zone) 4296 { 4297 if (IS_ENABLED(CONFIG_NUMA)) 4298 printk("Node %d ", zone_to_nid(zone)); 4299 } 4300 4301 long si_mem_available(void) 4302 { 4303 long available; 4304 unsigned long pagecache; 4305 unsigned long wmark_low = 0; 4306 unsigned long pages[NR_LRU_LISTS]; 4307 struct zone *zone; 4308 int lru; 4309 4310 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++) 4311 pages[lru] = global_node_page_state(NR_LRU_BASE + lru); 4312 4313 for_each_zone(zone) 4314 wmark_low += zone->watermark[WMARK_LOW]; 4315 4316 /* 4317 * Estimate the amount of memory available for userspace allocations, 4318 * without causing swapping. 4319 */ 4320 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages; 4321 4322 /* 4323 * Not all the page cache can be freed, otherwise the system will 4324 * start swapping. Assume at least half of the page cache, or the 4325 * low watermark worth of cache, needs to stay. 4326 */ 4327 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE]; 4328 pagecache -= min(pagecache / 2, wmark_low); 4329 available += pagecache; 4330 4331 /* 4332 * Part of the reclaimable slab consists of items that are in use, 4333 * and cannot be freed. Cap this estimate at the low watermark. 4334 */ 4335 available += global_page_state(NR_SLAB_RECLAIMABLE) - 4336 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low); 4337 4338 if (available < 0) 4339 available = 0; 4340 return available; 4341 } 4342 EXPORT_SYMBOL_GPL(si_mem_available); 4343 4344 void si_meminfo(struct sysinfo *val) 4345 { 4346 val->totalram = totalram_pages; 4347 val->sharedram = global_node_page_state(NR_SHMEM); 4348 val->freeram = global_page_state(NR_FREE_PAGES); 4349 val->bufferram = nr_blockdev_pages(); 4350 val->totalhigh = totalhigh_pages; 4351 val->freehigh = nr_free_highpages(); 4352 val->mem_unit = PAGE_SIZE; 4353 } 4354 4355 EXPORT_SYMBOL(si_meminfo); 4356 4357 #ifdef CONFIG_NUMA 4358 void si_meminfo_node(struct sysinfo *val, int nid) 4359 { 4360 int zone_type; /* needs to be signed */ 4361 unsigned long managed_pages = 0; 4362 unsigned long managed_highpages = 0; 4363 unsigned long free_highpages = 0; 4364 pg_data_t *pgdat = NODE_DATA(nid); 4365 4366 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) 4367 managed_pages += pgdat->node_zones[zone_type].managed_pages; 4368 val->totalram = managed_pages; 4369 val->sharedram = node_page_state(pgdat, NR_SHMEM); 4370 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES); 4371 #ifdef CONFIG_HIGHMEM 4372 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) { 4373 struct zone *zone = &pgdat->node_zones[zone_type]; 4374 4375 if (is_highmem(zone)) { 4376 managed_highpages += zone->managed_pages; 4377 free_highpages += zone_page_state(zone, NR_FREE_PAGES); 4378 } 4379 } 4380 val->totalhigh = managed_highpages; 4381 val->freehigh = free_highpages; 4382 #else 4383 val->totalhigh = managed_highpages; 4384 val->freehigh = free_highpages; 4385 #endif 4386 val->mem_unit = PAGE_SIZE; 4387 } 4388 #endif 4389 4390 /* 4391 * Determine whether the node should be displayed or not, depending on whether 4392 * SHOW_MEM_FILTER_NODES was passed to show_free_areas(). 4393 */ 4394 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask) 4395 { 4396 if (!(flags & SHOW_MEM_FILTER_NODES)) 4397 return false; 4398 4399 /* 4400 * no node mask - aka implicit memory numa policy. Do not bother with 4401 * the synchronization - read_mems_allowed_begin - because we do not 4402 * have to be precise here. 4403 */ 4404 if (!nodemask) 4405 nodemask = &cpuset_current_mems_allowed; 4406 4407 return !node_isset(nid, *nodemask); 4408 } 4409 4410 #define K(x) ((x) << (PAGE_SHIFT-10)) 4411 4412 static void show_migration_types(unsigned char type) 4413 { 4414 static const char types[MIGRATE_TYPES] = { 4415 [MIGRATE_UNMOVABLE] = 'U', 4416 [MIGRATE_MOVABLE] = 'M', 4417 [MIGRATE_RECLAIMABLE] = 'E', 4418 [MIGRATE_HIGHATOMIC] = 'H', 4419 #ifdef CONFIG_CMA 4420 [MIGRATE_CMA] = 'C', 4421 #endif 4422 #ifdef CONFIG_MEMORY_ISOLATION 4423 [MIGRATE_ISOLATE] = 'I', 4424 #endif 4425 }; 4426 char tmp[MIGRATE_TYPES + 1]; 4427 char *p = tmp; 4428 int i; 4429 4430 for (i = 0; i < MIGRATE_TYPES; i++) { 4431 if (type & (1 << i)) 4432 *p++ = types[i]; 4433 } 4434 4435 *p = '\0'; 4436 printk(KERN_CONT "(%s) ", tmp); 4437 } 4438 4439 /* 4440 * Show free area list (used inside shift_scroll-lock stuff) 4441 * We also calculate the percentage fragmentation. We do this by counting the 4442 * memory on each free list with the exception of the first item on the list. 4443 * 4444 * Bits in @filter: 4445 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's 4446 * cpuset. 4447 */ 4448 void show_free_areas(unsigned int filter, nodemask_t *nodemask) 4449 { 4450 unsigned long free_pcp = 0; 4451 int cpu; 4452 struct zone *zone; 4453 pg_data_t *pgdat; 4454 4455 for_each_populated_zone(zone) { 4456 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask)) 4457 continue; 4458 4459 for_each_online_cpu(cpu) 4460 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 4461 } 4462 4463 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n" 4464 " active_file:%lu inactive_file:%lu isolated_file:%lu\n" 4465 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n" 4466 " slab_reclaimable:%lu slab_unreclaimable:%lu\n" 4467 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n" 4468 " free:%lu free_pcp:%lu free_cma:%lu\n", 4469 global_node_page_state(NR_ACTIVE_ANON), 4470 global_node_page_state(NR_INACTIVE_ANON), 4471 global_node_page_state(NR_ISOLATED_ANON), 4472 global_node_page_state(NR_ACTIVE_FILE), 4473 global_node_page_state(NR_INACTIVE_FILE), 4474 global_node_page_state(NR_ISOLATED_FILE), 4475 global_node_page_state(NR_UNEVICTABLE), 4476 global_node_page_state(NR_FILE_DIRTY), 4477 global_node_page_state(NR_WRITEBACK), 4478 global_node_page_state(NR_UNSTABLE_NFS), 4479 global_page_state(NR_SLAB_RECLAIMABLE), 4480 global_page_state(NR_SLAB_UNRECLAIMABLE), 4481 global_node_page_state(NR_FILE_MAPPED), 4482 global_node_page_state(NR_SHMEM), 4483 global_page_state(NR_PAGETABLE), 4484 global_page_state(NR_BOUNCE), 4485 global_page_state(NR_FREE_PAGES), 4486 free_pcp, 4487 global_page_state(NR_FREE_CMA_PAGES)); 4488 4489 for_each_online_pgdat(pgdat) { 4490 if (show_mem_node_skip(filter, pgdat->node_id, nodemask)) 4491 continue; 4492 4493 printk("Node %d" 4494 " active_anon:%lukB" 4495 " inactive_anon:%lukB" 4496 " active_file:%lukB" 4497 " inactive_file:%lukB" 4498 " unevictable:%lukB" 4499 " isolated(anon):%lukB" 4500 " isolated(file):%lukB" 4501 " mapped:%lukB" 4502 " dirty:%lukB" 4503 " writeback:%lukB" 4504 " shmem:%lukB" 4505 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4506 " shmem_thp: %lukB" 4507 " shmem_pmdmapped: %lukB" 4508 " anon_thp: %lukB" 4509 #endif 4510 " writeback_tmp:%lukB" 4511 " unstable:%lukB" 4512 " pages_scanned:%lu" 4513 " all_unreclaimable? %s" 4514 "\n", 4515 pgdat->node_id, 4516 K(node_page_state(pgdat, NR_ACTIVE_ANON)), 4517 K(node_page_state(pgdat, NR_INACTIVE_ANON)), 4518 K(node_page_state(pgdat, NR_ACTIVE_FILE)), 4519 K(node_page_state(pgdat, NR_INACTIVE_FILE)), 4520 K(node_page_state(pgdat, NR_UNEVICTABLE)), 4521 K(node_page_state(pgdat, NR_ISOLATED_ANON)), 4522 K(node_page_state(pgdat, NR_ISOLATED_FILE)), 4523 K(node_page_state(pgdat, NR_FILE_MAPPED)), 4524 K(node_page_state(pgdat, NR_FILE_DIRTY)), 4525 K(node_page_state(pgdat, NR_WRITEBACK)), 4526 K(node_page_state(pgdat, NR_SHMEM)), 4527 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 4528 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR), 4529 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED) 4530 * HPAGE_PMD_NR), 4531 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR), 4532 #endif 4533 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)), 4534 K(node_page_state(pgdat, NR_UNSTABLE_NFS)), 4535 node_page_state(pgdat, NR_PAGES_SCANNED), 4536 !pgdat_reclaimable(pgdat) ? "yes" : "no"); 4537 } 4538 4539 for_each_populated_zone(zone) { 4540 int i; 4541 4542 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask)) 4543 continue; 4544 4545 free_pcp = 0; 4546 for_each_online_cpu(cpu) 4547 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count; 4548 4549 show_node(zone); 4550 printk(KERN_CONT 4551 "%s" 4552 " free:%lukB" 4553 " min:%lukB" 4554 " low:%lukB" 4555 " high:%lukB" 4556 " active_anon:%lukB" 4557 " inactive_anon:%lukB" 4558 " active_file:%lukB" 4559 " inactive_file:%lukB" 4560 " unevictable:%lukB" 4561 " writepending:%lukB" 4562 " present:%lukB" 4563 " managed:%lukB" 4564 " mlocked:%lukB" 4565 " slab_reclaimable:%lukB" 4566 " slab_unreclaimable:%lukB" 4567 " kernel_stack:%lukB" 4568 " pagetables:%lukB" 4569 " bounce:%lukB" 4570 " free_pcp:%lukB" 4571 " local_pcp:%ukB" 4572 " free_cma:%lukB" 4573 "\n", 4574 zone->name, 4575 K(zone_page_state(zone, NR_FREE_PAGES)), 4576 K(min_wmark_pages(zone)), 4577 K(low_wmark_pages(zone)), 4578 K(high_wmark_pages(zone)), 4579 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)), 4580 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)), 4581 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)), 4582 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)), 4583 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)), 4584 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)), 4585 K(zone->present_pages), 4586 K(zone->managed_pages), 4587 K(zone_page_state(zone, NR_MLOCK)), 4588 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)), 4589 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)), 4590 zone_page_state(zone, NR_KERNEL_STACK_KB), 4591 K(zone_page_state(zone, NR_PAGETABLE)), 4592 K(zone_page_state(zone, NR_BOUNCE)), 4593 K(free_pcp), 4594 K(this_cpu_read(zone->pageset->pcp.count)), 4595 K(zone_page_state(zone, NR_FREE_CMA_PAGES))); 4596 printk("lowmem_reserve[]:"); 4597 for (i = 0; i < MAX_NR_ZONES; i++) 4598 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]); 4599 printk(KERN_CONT "\n"); 4600 } 4601 4602 for_each_populated_zone(zone) { 4603 unsigned int order; 4604 unsigned long nr[MAX_ORDER], flags, total = 0; 4605 unsigned char types[MAX_ORDER]; 4606 4607 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask)) 4608 continue; 4609 show_node(zone); 4610 printk(KERN_CONT "%s: ", zone->name); 4611 4612 spin_lock_irqsave(&zone->lock, flags); 4613 for (order = 0; order < MAX_ORDER; order++) { 4614 struct free_area *area = &zone->free_area[order]; 4615 int type; 4616 4617 nr[order] = area->nr_free; 4618 total += nr[order] << order; 4619 4620 types[order] = 0; 4621 for (type = 0; type < MIGRATE_TYPES; type++) { 4622 if (!list_empty(&area->free_list[type])) 4623 types[order] |= 1 << type; 4624 } 4625 } 4626 spin_unlock_irqrestore(&zone->lock, flags); 4627 for (order = 0; order < MAX_ORDER; order++) { 4628 printk(KERN_CONT "%lu*%lukB ", 4629 nr[order], K(1UL) << order); 4630 if (nr[order]) 4631 show_migration_types(types[order]); 4632 } 4633 printk(KERN_CONT "= %lukB\n", K(total)); 4634 } 4635 4636 hugetlb_show_meminfo(); 4637 4638 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES)); 4639 4640 show_swap_cache_info(); 4641 } 4642 4643 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref) 4644 { 4645 zoneref->zone = zone; 4646 zoneref->zone_idx = zone_idx(zone); 4647 } 4648 4649 /* 4650 * Builds allocation fallback zone lists. 4651 * 4652 * Add all populated zones of a node to the zonelist. 4653 */ 4654 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist, 4655 int nr_zones) 4656 { 4657 struct zone *zone; 4658 enum zone_type zone_type = MAX_NR_ZONES; 4659 4660 do { 4661 zone_type--; 4662 zone = pgdat->node_zones + zone_type; 4663 if (managed_zone(zone)) { 4664 zoneref_set_zone(zone, 4665 &zonelist->_zonerefs[nr_zones++]); 4666 check_highest_zone(zone_type); 4667 } 4668 } while (zone_type); 4669 4670 return nr_zones; 4671 } 4672 4673 4674 /* 4675 * zonelist_order: 4676 * 0 = automatic detection of better ordering. 4677 * 1 = order by ([node] distance, -zonetype) 4678 * 2 = order by (-zonetype, [node] distance) 4679 * 4680 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create 4681 * the same zonelist. So only NUMA can configure this param. 4682 */ 4683 #define ZONELIST_ORDER_DEFAULT 0 4684 #define ZONELIST_ORDER_NODE 1 4685 #define ZONELIST_ORDER_ZONE 2 4686 4687 /* zonelist order in the kernel. 4688 * set_zonelist_order() will set this to NODE or ZONE. 4689 */ 4690 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT; 4691 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"}; 4692 4693 4694 #ifdef CONFIG_NUMA 4695 /* The value user specified ....changed by config */ 4696 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT; 4697 /* string for sysctl */ 4698 #define NUMA_ZONELIST_ORDER_LEN 16 4699 char numa_zonelist_order[16] = "default"; 4700 4701 /* 4702 * interface for configure zonelist ordering. 4703 * command line option "numa_zonelist_order" 4704 * = "[dD]efault - default, automatic configuration. 4705 * = "[nN]ode - order by node locality, then by zone within node 4706 * = "[zZ]one - order by zone, then by locality within zone 4707 */ 4708 4709 static int __parse_numa_zonelist_order(char *s) 4710 { 4711 if (*s == 'd' || *s == 'D') { 4712 user_zonelist_order = ZONELIST_ORDER_DEFAULT; 4713 } else if (*s == 'n' || *s == 'N') { 4714 user_zonelist_order = ZONELIST_ORDER_NODE; 4715 } else if (*s == 'z' || *s == 'Z') { 4716 user_zonelist_order = ZONELIST_ORDER_ZONE; 4717 } else { 4718 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s); 4719 return -EINVAL; 4720 } 4721 return 0; 4722 } 4723 4724 static __init int setup_numa_zonelist_order(char *s) 4725 { 4726 int ret; 4727 4728 if (!s) 4729 return 0; 4730 4731 ret = __parse_numa_zonelist_order(s); 4732 if (ret == 0) 4733 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN); 4734 4735 return ret; 4736 } 4737 early_param("numa_zonelist_order", setup_numa_zonelist_order); 4738 4739 /* 4740 * sysctl handler for numa_zonelist_order 4741 */ 4742 int numa_zonelist_order_handler(struct ctl_table *table, int write, 4743 void __user *buffer, size_t *length, 4744 loff_t *ppos) 4745 { 4746 char saved_string[NUMA_ZONELIST_ORDER_LEN]; 4747 int ret; 4748 static DEFINE_MUTEX(zl_order_mutex); 4749 4750 mutex_lock(&zl_order_mutex); 4751 if (write) { 4752 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) { 4753 ret = -EINVAL; 4754 goto out; 4755 } 4756 strcpy(saved_string, (char *)table->data); 4757 } 4758 ret = proc_dostring(table, write, buffer, length, ppos); 4759 if (ret) 4760 goto out; 4761 if (write) { 4762 int oldval = user_zonelist_order; 4763 4764 ret = __parse_numa_zonelist_order((char *)table->data); 4765 if (ret) { 4766 /* 4767 * bogus value. restore saved string 4768 */ 4769 strncpy((char *)table->data, saved_string, 4770 NUMA_ZONELIST_ORDER_LEN); 4771 user_zonelist_order = oldval; 4772 } else if (oldval != user_zonelist_order) { 4773 mutex_lock(&zonelists_mutex); 4774 build_all_zonelists(NULL, NULL); 4775 mutex_unlock(&zonelists_mutex); 4776 } 4777 } 4778 out: 4779 mutex_unlock(&zl_order_mutex); 4780 return ret; 4781 } 4782 4783 4784 #define MAX_NODE_LOAD (nr_online_nodes) 4785 static int node_load[MAX_NUMNODES]; 4786 4787 /** 4788 * find_next_best_node - find the next node that should appear in a given node's fallback list 4789 * @node: node whose fallback list we're appending 4790 * @used_node_mask: nodemask_t of already used nodes 4791 * 4792 * We use a number of factors to determine which is the next node that should 4793 * appear on a given node's fallback list. The node should not have appeared 4794 * already in @node's fallback list, and it should be the next closest node 4795 * according to the distance array (which contains arbitrary distance values 4796 * from each node to each node in the system), and should also prefer nodes 4797 * with no CPUs, since presumably they'll have very little allocation pressure 4798 * on them otherwise. 4799 * It returns -1 if no node is found. 4800 */ 4801 static int find_next_best_node(int node, nodemask_t *used_node_mask) 4802 { 4803 int n, val; 4804 int min_val = INT_MAX; 4805 int best_node = NUMA_NO_NODE; 4806 const struct cpumask *tmp = cpumask_of_node(0); 4807 4808 /* Use the local node if we haven't already */ 4809 if (!node_isset(node, *used_node_mask)) { 4810 node_set(node, *used_node_mask); 4811 return node; 4812 } 4813 4814 for_each_node_state(n, N_MEMORY) { 4815 4816 /* Don't want a node to appear more than once */ 4817 if (node_isset(n, *used_node_mask)) 4818 continue; 4819 4820 /* Use the distance array to find the distance */ 4821 val = node_distance(node, n); 4822 4823 /* Penalize nodes under us ("prefer the next node") */ 4824 val += (n < node); 4825 4826 /* Give preference to headless and unused nodes */ 4827 tmp = cpumask_of_node(n); 4828 if (!cpumask_empty(tmp)) 4829 val += PENALTY_FOR_NODE_WITH_CPUS; 4830 4831 /* Slight preference for less loaded node */ 4832 val *= (MAX_NODE_LOAD*MAX_NUMNODES); 4833 val += node_load[n]; 4834 4835 if (val < min_val) { 4836 min_val = val; 4837 best_node = n; 4838 } 4839 } 4840 4841 if (best_node >= 0) 4842 node_set(best_node, *used_node_mask); 4843 4844 return best_node; 4845 } 4846 4847 4848 /* 4849 * Build zonelists ordered by node and zones within node. 4850 * This results in maximum locality--normal zone overflows into local 4851 * DMA zone, if any--but risks exhausting DMA zone. 4852 */ 4853 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node) 4854 { 4855 int j; 4856 struct zonelist *zonelist; 4857 4858 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK]; 4859 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++) 4860 ; 4861 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 4862 zonelist->_zonerefs[j].zone = NULL; 4863 zonelist->_zonerefs[j].zone_idx = 0; 4864 } 4865 4866 /* 4867 * Build gfp_thisnode zonelists 4868 */ 4869 static void build_thisnode_zonelists(pg_data_t *pgdat) 4870 { 4871 int j; 4872 struct zonelist *zonelist; 4873 4874 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK]; 4875 j = build_zonelists_node(pgdat, zonelist, 0); 4876 zonelist->_zonerefs[j].zone = NULL; 4877 zonelist->_zonerefs[j].zone_idx = 0; 4878 } 4879 4880 /* 4881 * Build zonelists ordered by zone and nodes within zones. 4882 * This results in conserving DMA zone[s] until all Normal memory is 4883 * exhausted, but results in overflowing to remote node while memory 4884 * may still exist in local DMA zone. 4885 */ 4886 static int node_order[MAX_NUMNODES]; 4887 4888 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes) 4889 { 4890 int pos, j, node; 4891 int zone_type; /* needs to be signed */ 4892 struct zone *z; 4893 struct zonelist *zonelist; 4894 4895 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK]; 4896 pos = 0; 4897 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) { 4898 for (j = 0; j < nr_nodes; j++) { 4899 node = node_order[j]; 4900 z = &NODE_DATA(node)->node_zones[zone_type]; 4901 if (managed_zone(z)) { 4902 zoneref_set_zone(z, 4903 &zonelist->_zonerefs[pos++]); 4904 check_highest_zone(zone_type); 4905 } 4906 } 4907 } 4908 zonelist->_zonerefs[pos].zone = NULL; 4909 zonelist->_zonerefs[pos].zone_idx = 0; 4910 } 4911 4912 #if defined(CONFIG_64BIT) 4913 /* 4914 * Devices that require DMA32/DMA are relatively rare and do not justify a 4915 * penalty to every machine in case the specialised case applies. Default 4916 * to Node-ordering on 64-bit NUMA machines 4917 */ 4918 static int default_zonelist_order(void) 4919 { 4920 return ZONELIST_ORDER_NODE; 4921 } 4922 #else 4923 /* 4924 * On 32-bit, the Normal zone needs to be preserved for allocations accessible 4925 * by the kernel. If processes running on node 0 deplete the low memory zone 4926 * then reclaim will occur more frequency increasing stalls and potentially 4927 * be easier to OOM if a large percentage of the zone is under writeback or 4928 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set. 4929 * Hence, default to zone ordering on 32-bit. 4930 */ 4931 static int default_zonelist_order(void) 4932 { 4933 return ZONELIST_ORDER_ZONE; 4934 } 4935 #endif /* CONFIG_64BIT */ 4936 4937 static void set_zonelist_order(void) 4938 { 4939 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT) 4940 current_zonelist_order = default_zonelist_order(); 4941 else 4942 current_zonelist_order = user_zonelist_order; 4943 } 4944 4945 static void build_zonelists(pg_data_t *pgdat) 4946 { 4947 int i, node, load; 4948 nodemask_t used_mask; 4949 int local_node, prev_node; 4950 struct zonelist *zonelist; 4951 unsigned int order = current_zonelist_order; 4952 4953 /* initialize zonelists */ 4954 for (i = 0; i < MAX_ZONELISTS; i++) { 4955 zonelist = pgdat->node_zonelists + i; 4956 zonelist->_zonerefs[0].zone = NULL; 4957 zonelist->_zonerefs[0].zone_idx = 0; 4958 } 4959 4960 /* NUMA-aware ordering of nodes */ 4961 local_node = pgdat->node_id; 4962 load = nr_online_nodes; 4963 prev_node = local_node; 4964 nodes_clear(used_mask); 4965 4966 memset(node_order, 0, sizeof(node_order)); 4967 i = 0; 4968 4969 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) { 4970 /* 4971 * We don't want to pressure a particular node. 4972 * So adding penalty to the first node in same 4973 * distance group to make it round-robin. 4974 */ 4975 if (node_distance(local_node, node) != 4976 node_distance(local_node, prev_node)) 4977 node_load[node] = load; 4978 4979 prev_node = node; 4980 load--; 4981 if (order == ZONELIST_ORDER_NODE) 4982 build_zonelists_in_node_order(pgdat, node); 4983 else 4984 node_order[i++] = node; /* remember order */ 4985 } 4986 4987 if (order == ZONELIST_ORDER_ZONE) { 4988 /* calculate node order -- i.e., DMA last! */ 4989 build_zonelists_in_zone_order(pgdat, i); 4990 } 4991 4992 build_thisnode_zonelists(pgdat); 4993 } 4994 4995 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 4996 /* 4997 * Return node id of node used for "local" allocations. 4998 * I.e., first node id of first zone in arg node's generic zonelist. 4999 * Used for initializing percpu 'numa_mem', which is used primarily 5000 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist. 5001 */ 5002 int local_memory_node(int node) 5003 { 5004 struct zoneref *z; 5005 5006 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL), 5007 gfp_zone(GFP_KERNEL), 5008 NULL); 5009 return z->zone->node; 5010 } 5011 #endif 5012 5013 static void setup_min_unmapped_ratio(void); 5014 static void setup_min_slab_ratio(void); 5015 #else /* CONFIG_NUMA */ 5016 5017 static void set_zonelist_order(void) 5018 { 5019 current_zonelist_order = ZONELIST_ORDER_ZONE; 5020 } 5021 5022 static void build_zonelists(pg_data_t *pgdat) 5023 { 5024 int node, local_node; 5025 enum zone_type j; 5026 struct zonelist *zonelist; 5027 5028 local_node = pgdat->node_id; 5029 5030 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK]; 5031 j = build_zonelists_node(pgdat, zonelist, 0); 5032 5033 /* 5034 * Now we build the zonelist so that it contains the zones 5035 * of all the other nodes. 5036 * We don't want to pressure a particular node, so when 5037 * building the zones for node N, we make sure that the 5038 * zones coming right after the local ones are those from 5039 * node N+1 (modulo N) 5040 */ 5041 for (node = local_node + 1; node < MAX_NUMNODES; node++) { 5042 if (!node_online(node)) 5043 continue; 5044 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 5045 } 5046 for (node = 0; node < local_node; node++) { 5047 if (!node_online(node)) 5048 continue; 5049 j = build_zonelists_node(NODE_DATA(node), zonelist, j); 5050 } 5051 5052 zonelist->_zonerefs[j].zone = NULL; 5053 zonelist->_zonerefs[j].zone_idx = 0; 5054 } 5055 5056 #endif /* CONFIG_NUMA */ 5057 5058 /* 5059 * Boot pageset table. One per cpu which is going to be used for all 5060 * zones and all nodes. The parameters will be set in such a way 5061 * that an item put on a list will immediately be handed over to 5062 * the buddy list. This is safe since pageset manipulation is done 5063 * with interrupts disabled. 5064 * 5065 * The boot_pagesets must be kept even after bootup is complete for 5066 * unused processors and/or zones. They do play a role for bootstrapping 5067 * hotplugged processors. 5068 * 5069 * zoneinfo_show() and maybe other functions do 5070 * not check if the processor is online before following the pageset pointer. 5071 * Other parts of the kernel may not check if the zone is available. 5072 */ 5073 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch); 5074 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset); 5075 static void setup_zone_pageset(struct zone *zone); 5076 5077 /* 5078 * Global mutex to protect against size modification of zonelists 5079 * as well as to serialize pageset setup for the new populated zone. 5080 */ 5081 DEFINE_MUTEX(zonelists_mutex); 5082 5083 /* return values int ....just for stop_machine() */ 5084 static int __build_all_zonelists(void *data) 5085 { 5086 int nid; 5087 int cpu; 5088 pg_data_t *self = data; 5089 5090 #ifdef CONFIG_NUMA 5091 memset(node_load, 0, sizeof(node_load)); 5092 #endif 5093 5094 if (self && !node_online(self->node_id)) { 5095 build_zonelists(self); 5096 } 5097 5098 for_each_online_node(nid) { 5099 pg_data_t *pgdat = NODE_DATA(nid); 5100 5101 build_zonelists(pgdat); 5102 } 5103 5104 /* 5105 * Initialize the boot_pagesets that are going to be used 5106 * for bootstrapping processors. The real pagesets for 5107 * each zone will be allocated later when the per cpu 5108 * allocator is available. 5109 * 5110 * boot_pagesets are used also for bootstrapping offline 5111 * cpus if the system is already booted because the pagesets 5112 * are needed to initialize allocators on a specific cpu too. 5113 * F.e. the percpu allocator needs the page allocator which 5114 * needs the percpu allocator in order to allocate its pagesets 5115 * (a chicken-egg dilemma). 5116 */ 5117 for_each_possible_cpu(cpu) { 5118 setup_pageset(&per_cpu(boot_pageset, cpu), 0); 5119 5120 #ifdef CONFIG_HAVE_MEMORYLESS_NODES 5121 /* 5122 * We now know the "local memory node" for each node-- 5123 * i.e., the node of the first zone in the generic zonelist. 5124 * Set up numa_mem percpu variable for on-line cpus. During 5125 * boot, only the boot cpu should be on-line; we'll init the 5126 * secondary cpus' numa_mem as they come on-line. During 5127 * node/memory hotplug, we'll fixup all on-line cpus. 5128 */ 5129 if (cpu_online(cpu)) 5130 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu))); 5131 #endif 5132 } 5133 5134 return 0; 5135 } 5136 5137 static noinline void __init 5138 build_all_zonelists_init(void) 5139 { 5140 __build_all_zonelists(NULL); 5141 mminit_verify_zonelist(); 5142 cpuset_init_current_mems_allowed(); 5143 } 5144 5145 /* 5146 * Called with zonelists_mutex held always 5147 * unless system_state == SYSTEM_BOOTING. 5148 * 5149 * __ref due to (1) call of __meminit annotated setup_zone_pageset 5150 * [we're only called with non-NULL zone through __meminit paths] and 5151 * (2) call of __init annotated helper build_all_zonelists_init 5152 * [protected by SYSTEM_BOOTING]. 5153 */ 5154 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone) 5155 { 5156 set_zonelist_order(); 5157 5158 if (system_state == SYSTEM_BOOTING) { 5159 build_all_zonelists_init(); 5160 } else { 5161 #ifdef CONFIG_MEMORY_HOTPLUG 5162 if (zone) 5163 setup_zone_pageset(zone); 5164 #endif 5165 /* we have to stop all cpus to guarantee there is no user 5166 of zonelist */ 5167 stop_machine(__build_all_zonelists, pgdat, NULL); 5168 /* cpuset refresh routine should be here */ 5169 } 5170 vm_total_pages = nr_free_pagecache_pages(); 5171 /* 5172 * Disable grouping by mobility if the number of pages in the 5173 * system is too low to allow the mechanism to work. It would be 5174 * more accurate, but expensive to check per-zone. This check is 5175 * made on memory-hotadd so a system can start with mobility 5176 * disabled and enable it later 5177 */ 5178 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES)) 5179 page_group_by_mobility_disabled = 1; 5180 else 5181 page_group_by_mobility_disabled = 0; 5182 5183 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n", 5184 nr_online_nodes, 5185 zonelist_order_name[current_zonelist_order], 5186 page_group_by_mobility_disabled ? "off" : "on", 5187 vm_total_pages); 5188 #ifdef CONFIG_NUMA 5189 pr_info("Policy zone: %s\n", zone_names[policy_zone]); 5190 #endif 5191 } 5192 5193 /* 5194 * Initially all pages are reserved - free ones are freed 5195 * up by free_all_bootmem() once the early boot process is 5196 * done. Non-atomic initialization, single-pass. 5197 */ 5198 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone, 5199 unsigned long start_pfn, enum memmap_context context) 5200 { 5201 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn)); 5202 unsigned long end_pfn = start_pfn + size; 5203 pg_data_t *pgdat = NODE_DATA(nid); 5204 unsigned long pfn; 5205 unsigned long nr_initialised = 0; 5206 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5207 struct memblock_region *r = NULL, *tmp; 5208 #endif 5209 5210 if (highest_memmap_pfn < end_pfn - 1) 5211 highest_memmap_pfn = end_pfn - 1; 5212 5213 /* 5214 * Honor reservation requested by the driver for this ZONE_DEVICE 5215 * memory 5216 */ 5217 if (altmap && start_pfn == altmap->base_pfn) 5218 start_pfn += altmap->reserve; 5219 5220 for (pfn = start_pfn; pfn < end_pfn; pfn++) { 5221 /* 5222 * There can be holes in boot-time mem_map[]s handed to this 5223 * function. They do not exist on hotplugged memory. 5224 */ 5225 if (context != MEMMAP_EARLY) 5226 goto not_early; 5227 5228 if (!early_pfn_valid(pfn)) { 5229 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5230 /* 5231 * Skip to the pfn preceding the next valid one (or 5232 * end_pfn), such that we hit a valid pfn (or end_pfn) 5233 * on our next iteration of the loop. 5234 */ 5235 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1; 5236 #endif 5237 continue; 5238 } 5239 if (!early_pfn_in_nid(pfn, nid)) 5240 continue; 5241 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised)) 5242 break; 5243 5244 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5245 /* 5246 * Check given memblock attribute by firmware which can affect 5247 * kernel memory layout. If zone==ZONE_MOVABLE but memory is 5248 * mirrored, it's an overlapped memmap init. skip it. 5249 */ 5250 if (mirrored_kernelcore && zone == ZONE_MOVABLE) { 5251 if (!r || pfn >= memblock_region_memory_end_pfn(r)) { 5252 for_each_memblock(memory, tmp) 5253 if (pfn < memblock_region_memory_end_pfn(tmp)) 5254 break; 5255 r = tmp; 5256 } 5257 if (pfn >= memblock_region_memory_base_pfn(r) && 5258 memblock_is_mirror(r)) { 5259 /* already initialized as NORMAL */ 5260 pfn = memblock_region_memory_end_pfn(r); 5261 continue; 5262 } 5263 } 5264 #endif 5265 5266 not_early: 5267 /* 5268 * Mark the block movable so that blocks are reserved for 5269 * movable at startup. This will force kernel allocations 5270 * to reserve their blocks rather than leaking throughout 5271 * the address space during boot when many long-lived 5272 * kernel allocations are made. 5273 * 5274 * bitmap is created for zone's valid pfn range. but memmap 5275 * can be created for invalid pages (for alignment) 5276 * check here not to call set_pageblock_migratetype() against 5277 * pfn out of zone. 5278 */ 5279 if (!(pfn & (pageblock_nr_pages - 1))) { 5280 struct page *page = pfn_to_page(pfn); 5281 5282 __init_single_page(page, pfn, zone, nid); 5283 set_pageblock_migratetype(page, MIGRATE_MOVABLE); 5284 } else { 5285 __init_single_pfn(pfn, zone, nid); 5286 } 5287 } 5288 } 5289 5290 static void __meminit zone_init_free_lists(struct zone *zone) 5291 { 5292 unsigned int order, t; 5293 for_each_migratetype_order(order, t) { 5294 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]); 5295 zone->free_area[order].nr_free = 0; 5296 } 5297 } 5298 5299 #ifndef __HAVE_ARCH_MEMMAP_INIT 5300 #define memmap_init(size, nid, zone, start_pfn) \ 5301 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY) 5302 #endif 5303 5304 static int zone_batchsize(struct zone *zone) 5305 { 5306 #ifdef CONFIG_MMU 5307 int batch; 5308 5309 /* 5310 * The per-cpu-pages pools are set to around 1000th of the 5311 * size of the zone. But no more than 1/2 of a meg. 5312 * 5313 * OK, so we don't know how big the cache is. So guess. 5314 */ 5315 batch = zone->managed_pages / 1024; 5316 if (batch * PAGE_SIZE > 512 * 1024) 5317 batch = (512 * 1024) / PAGE_SIZE; 5318 batch /= 4; /* We effectively *= 4 below */ 5319 if (batch < 1) 5320 batch = 1; 5321 5322 /* 5323 * Clamp the batch to a 2^n - 1 value. Having a power 5324 * of 2 value was found to be more likely to have 5325 * suboptimal cache aliasing properties in some cases. 5326 * 5327 * For example if 2 tasks are alternately allocating 5328 * batches of pages, one task can end up with a lot 5329 * of pages of one half of the possible page colors 5330 * and the other with pages of the other colors. 5331 */ 5332 batch = rounddown_pow_of_two(batch + batch/2) - 1; 5333 5334 return batch; 5335 5336 #else 5337 /* The deferral and batching of frees should be suppressed under NOMMU 5338 * conditions. 5339 * 5340 * The problem is that NOMMU needs to be able to allocate large chunks 5341 * of contiguous memory as there's no hardware page translation to 5342 * assemble apparent contiguous memory from discontiguous pages. 5343 * 5344 * Queueing large contiguous runs of pages for batching, however, 5345 * causes the pages to actually be freed in smaller chunks. As there 5346 * can be a significant delay between the individual batches being 5347 * recycled, this leads to the once large chunks of space being 5348 * fragmented and becoming unavailable for high-order allocations. 5349 */ 5350 return 0; 5351 #endif 5352 } 5353 5354 /* 5355 * pcp->high and pcp->batch values are related and dependent on one another: 5356 * ->batch must never be higher then ->high. 5357 * The following function updates them in a safe manner without read side 5358 * locking. 5359 * 5360 * Any new users of pcp->batch and pcp->high should ensure they can cope with 5361 * those fields changing asynchronously (acording the the above rule). 5362 * 5363 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function 5364 * outside of boot time (or some other assurance that no concurrent updaters 5365 * exist). 5366 */ 5367 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high, 5368 unsigned long batch) 5369 { 5370 /* start with a fail safe value for batch */ 5371 pcp->batch = 1; 5372 smp_wmb(); 5373 5374 /* Update high, then batch, in order */ 5375 pcp->high = high; 5376 smp_wmb(); 5377 5378 pcp->batch = batch; 5379 } 5380 5381 /* a companion to pageset_set_high() */ 5382 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch) 5383 { 5384 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch)); 5385 } 5386 5387 static void pageset_init(struct per_cpu_pageset *p) 5388 { 5389 struct per_cpu_pages *pcp; 5390 int migratetype; 5391 5392 memset(p, 0, sizeof(*p)); 5393 5394 pcp = &p->pcp; 5395 pcp->count = 0; 5396 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++) 5397 INIT_LIST_HEAD(&pcp->lists[migratetype]); 5398 } 5399 5400 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch) 5401 { 5402 pageset_init(p); 5403 pageset_set_batch(p, batch); 5404 } 5405 5406 /* 5407 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist 5408 * to the value high for the pageset p. 5409 */ 5410 static void pageset_set_high(struct per_cpu_pageset *p, 5411 unsigned long high) 5412 { 5413 unsigned long batch = max(1UL, high / 4); 5414 if ((high / 4) > (PAGE_SHIFT * 8)) 5415 batch = PAGE_SHIFT * 8; 5416 5417 pageset_update(&p->pcp, high, batch); 5418 } 5419 5420 static void pageset_set_high_and_batch(struct zone *zone, 5421 struct per_cpu_pageset *pcp) 5422 { 5423 if (percpu_pagelist_fraction) 5424 pageset_set_high(pcp, 5425 (zone->managed_pages / 5426 percpu_pagelist_fraction)); 5427 else 5428 pageset_set_batch(pcp, zone_batchsize(zone)); 5429 } 5430 5431 static void __meminit zone_pageset_init(struct zone *zone, int cpu) 5432 { 5433 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu); 5434 5435 pageset_init(pcp); 5436 pageset_set_high_and_batch(zone, pcp); 5437 } 5438 5439 static void __meminit setup_zone_pageset(struct zone *zone) 5440 { 5441 int cpu; 5442 zone->pageset = alloc_percpu(struct per_cpu_pageset); 5443 for_each_possible_cpu(cpu) 5444 zone_pageset_init(zone, cpu); 5445 } 5446 5447 /* 5448 * Allocate per cpu pagesets and initialize them. 5449 * Before this call only boot pagesets were available. 5450 */ 5451 void __init setup_per_cpu_pageset(void) 5452 { 5453 struct pglist_data *pgdat; 5454 struct zone *zone; 5455 5456 for_each_populated_zone(zone) 5457 setup_zone_pageset(zone); 5458 5459 for_each_online_pgdat(pgdat) 5460 pgdat->per_cpu_nodestats = 5461 alloc_percpu(struct per_cpu_nodestat); 5462 } 5463 5464 static __meminit void zone_pcp_init(struct zone *zone) 5465 { 5466 /* 5467 * per cpu subsystem is not up at this point. The following code 5468 * relies on the ability of the linker to provide the 5469 * offset of a (static) per cpu variable into the per cpu area. 5470 */ 5471 zone->pageset = &boot_pageset; 5472 5473 if (populated_zone(zone)) 5474 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n", 5475 zone->name, zone->present_pages, 5476 zone_batchsize(zone)); 5477 } 5478 5479 int __meminit init_currently_empty_zone(struct zone *zone, 5480 unsigned long zone_start_pfn, 5481 unsigned long size) 5482 { 5483 struct pglist_data *pgdat = zone->zone_pgdat; 5484 5485 pgdat->nr_zones = zone_idx(zone) + 1; 5486 5487 zone->zone_start_pfn = zone_start_pfn; 5488 5489 mminit_dprintk(MMINIT_TRACE, "memmap_init", 5490 "Initialising map node %d zone %lu pfns %lu -> %lu\n", 5491 pgdat->node_id, 5492 (unsigned long)zone_idx(zone), 5493 zone_start_pfn, (zone_start_pfn + size)); 5494 5495 zone_init_free_lists(zone); 5496 zone->initialized = 1; 5497 5498 return 0; 5499 } 5500 5501 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 5502 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID 5503 5504 /* 5505 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on. 5506 */ 5507 int __meminit __early_pfn_to_nid(unsigned long pfn, 5508 struct mminit_pfnnid_cache *state) 5509 { 5510 unsigned long start_pfn, end_pfn; 5511 int nid; 5512 5513 if (state->last_start <= pfn && pfn < state->last_end) 5514 return state->last_nid; 5515 5516 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn); 5517 if (nid != -1) { 5518 state->last_start = start_pfn; 5519 state->last_end = end_pfn; 5520 state->last_nid = nid; 5521 } 5522 5523 return nid; 5524 } 5525 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */ 5526 5527 /** 5528 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range 5529 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed. 5530 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid 5531 * 5532 * If an architecture guarantees that all ranges registered contain no holes 5533 * and may be freed, this this function may be used instead of calling 5534 * memblock_free_early_nid() manually. 5535 */ 5536 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn) 5537 { 5538 unsigned long start_pfn, end_pfn; 5539 int i, this_nid; 5540 5541 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) { 5542 start_pfn = min(start_pfn, max_low_pfn); 5543 end_pfn = min(end_pfn, max_low_pfn); 5544 5545 if (start_pfn < end_pfn) 5546 memblock_free_early_nid(PFN_PHYS(start_pfn), 5547 (end_pfn - start_pfn) << PAGE_SHIFT, 5548 this_nid); 5549 } 5550 } 5551 5552 /** 5553 * sparse_memory_present_with_active_regions - Call memory_present for each active range 5554 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used. 5555 * 5556 * If an architecture guarantees that all ranges registered contain no holes and may 5557 * be freed, this function may be used instead of calling memory_present() manually. 5558 */ 5559 void __init sparse_memory_present_with_active_regions(int nid) 5560 { 5561 unsigned long start_pfn, end_pfn; 5562 int i, this_nid; 5563 5564 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) 5565 memory_present(this_nid, start_pfn, end_pfn); 5566 } 5567 5568 /** 5569 * get_pfn_range_for_nid - Return the start and end page frames for a node 5570 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned. 5571 * @start_pfn: Passed by reference. On return, it will have the node start_pfn. 5572 * @end_pfn: Passed by reference. On return, it will have the node end_pfn. 5573 * 5574 * It returns the start and end page frame of a node based on information 5575 * provided by memblock_set_node(). If called for a node 5576 * with no available memory, a warning is printed and the start and end 5577 * PFNs will be 0. 5578 */ 5579 void __meminit get_pfn_range_for_nid(unsigned int nid, 5580 unsigned long *start_pfn, unsigned long *end_pfn) 5581 { 5582 unsigned long this_start_pfn, this_end_pfn; 5583 int i; 5584 5585 *start_pfn = -1UL; 5586 *end_pfn = 0; 5587 5588 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) { 5589 *start_pfn = min(*start_pfn, this_start_pfn); 5590 *end_pfn = max(*end_pfn, this_end_pfn); 5591 } 5592 5593 if (*start_pfn == -1UL) 5594 *start_pfn = 0; 5595 } 5596 5597 /* 5598 * This finds a zone that can be used for ZONE_MOVABLE pages. The 5599 * assumption is made that zones within a node are ordered in monotonic 5600 * increasing memory addresses so that the "highest" populated zone is used 5601 */ 5602 static void __init find_usable_zone_for_movable(void) 5603 { 5604 int zone_index; 5605 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) { 5606 if (zone_index == ZONE_MOVABLE) 5607 continue; 5608 5609 if (arch_zone_highest_possible_pfn[zone_index] > 5610 arch_zone_lowest_possible_pfn[zone_index]) 5611 break; 5612 } 5613 5614 VM_BUG_ON(zone_index == -1); 5615 movable_zone = zone_index; 5616 } 5617 5618 /* 5619 * The zone ranges provided by the architecture do not include ZONE_MOVABLE 5620 * because it is sized independent of architecture. Unlike the other zones, 5621 * the starting point for ZONE_MOVABLE is not fixed. It may be different 5622 * in each node depending on the size of each node and how evenly kernelcore 5623 * is distributed. This helper function adjusts the zone ranges 5624 * provided by the architecture for a given node by using the end of the 5625 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that 5626 * zones within a node are in order of monotonic increases memory addresses 5627 */ 5628 static void __meminit adjust_zone_range_for_zone_movable(int nid, 5629 unsigned long zone_type, 5630 unsigned long node_start_pfn, 5631 unsigned long node_end_pfn, 5632 unsigned long *zone_start_pfn, 5633 unsigned long *zone_end_pfn) 5634 { 5635 /* Only adjust if ZONE_MOVABLE is on this node */ 5636 if (zone_movable_pfn[nid]) { 5637 /* Size ZONE_MOVABLE */ 5638 if (zone_type == ZONE_MOVABLE) { 5639 *zone_start_pfn = zone_movable_pfn[nid]; 5640 *zone_end_pfn = min(node_end_pfn, 5641 arch_zone_highest_possible_pfn[movable_zone]); 5642 5643 /* Adjust for ZONE_MOVABLE starting within this range */ 5644 } else if (!mirrored_kernelcore && 5645 *zone_start_pfn < zone_movable_pfn[nid] && 5646 *zone_end_pfn > zone_movable_pfn[nid]) { 5647 *zone_end_pfn = zone_movable_pfn[nid]; 5648 5649 /* Check if this whole range is within ZONE_MOVABLE */ 5650 } else if (*zone_start_pfn >= zone_movable_pfn[nid]) 5651 *zone_start_pfn = *zone_end_pfn; 5652 } 5653 } 5654 5655 /* 5656 * Return the number of pages a zone spans in a node, including holes 5657 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node() 5658 */ 5659 static unsigned long __meminit zone_spanned_pages_in_node(int nid, 5660 unsigned long zone_type, 5661 unsigned long node_start_pfn, 5662 unsigned long node_end_pfn, 5663 unsigned long *zone_start_pfn, 5664 unsigned long *zone_end_pfn, 5665 unsigned long *ignored) 5666 { 5667 /* When hotadd a new node from cpu_up(), the node should be empty */ 5668 if (!node_start_pfn && !node_end_pfn) 5669 return 0; 5670 5671 /* Get the start and end of the zone */ 5672 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type]; 5673 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type]; 5674 adjust_zone_range_for_zone_movable(nid, zone_type, 5675 node_start_pfn, node_end_pfn, 5676 zone_start_pfn, zone_end_pfn); 5677 5678 /* Check that this node has pages within the zone's required range */ 5679 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn) 5680 return 0; 5681 5682 /* Move the zone boundaries inside the node if necessary */ 5683 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn); 5684 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn); 5685 5686 /* Return the spanned pages */ 5687 return *zone_end_pfn - *zone_start_pfn; 5688 } 5689 5690 /* 5691 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES, 5692 * then all holes in the requested range will be accounted for. 5693 */ 5694 unsigned long __meminit __absent_pages_in_range(int nid, 5695 unsigned long range_start_pfn, 5696 unsigned long range_end_pfn) 5697 { 5698 unsigned long nr_absent = range_end_pfn - range_start_pfn; 5699 unsigned long start_pfn, end_pfn; 5700 int i; 5701 5702 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 5703 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn); 5704 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn); 5705 nr_absent -= end_pfn - start_pfn; 5706 } 5707 return nr_absent; 5708 } 5709 5710 /** 5711 * absent_pages_in_range - Return number of page frames in holes within a range 5712 * @start_pfn: The start PFN to start searching for holes 5713 * @end_pfn: The end PFN to stop searching for holes 5714 * 5715 * It returns the number of pages frames in memory holes within a range. 5716 */ 5717 unsigned long __init absent_pages_in_range(unsigned long start_pfn, 5718 unsigned long end_pfn) 5719 { 5720 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn); 5721 } 5722 5723 /* Return the number of page frames in holes in a zone on a node */ 5724 static unsigned long __meminit zone_absent_pages_in_node(int nid, 5725 unsigned long zone_type, 5726 unsigned long node_start_pfn, 5727 unsigned long node_end_pfn, 5728 unsigned long *ignored) 5729 { 5730 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type]; 5731 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type]; 5732 unsigned long zone_start_pfn, zone_end_pfn; 5733 unsigned long nr_absent; 5734 5735 /* When hotadd a new node from cpu_up(), the node should be empty */ 5736 if (!node_start_pfn && !node_end_pfn) 5737 return 0; 5738 5739 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high); 5740 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high); 5741 5742 adjust_zone_range_for_zone_movable(nid, zone_type, 5743 node_start_pfn, node_end_pfn, 5744 &zone_start_pfn, &zone_end_pfn); 5745 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn); 5746 5747 /* 5748 * ZONE_MOVABLE handling. 5749 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages 5750 * and vice versa. 5751 */ 5752 if (mirrored_kernelcore && zone_movable_pfn[nid]) { 5753 unsigned long start_pfn, end_pfn; 5754 struct memblock_region *r; 5755 5756 for_each_memblock(memory, r) { 5757 start_pfn = clamp(memblock_region_memory_base_pfn(r), 5758 zone_start_pfn, zone_end_pfn); 5759 end_pfn = clamp(memblock_region_memory_end_pfn(r), 5760 zone_start_pfn, zone_end_pfn); 5761 5762 if (zone_type == ZONE_MOVABLE && 5763 memblock_is_mirror(r)) 5764 nr_absent += end_pfn - start_pfn; 5765 5766 if (zone_type == ZONE_NORMAL && 5767 !memblock_is_mirror(r)) 5768 nr_absent += end_pfn - start_pfn; 5769 } 5770 } 5771 5772 return nr_absent; 5773 } 5774 5775 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5776 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid, 5777 unsigned long zone_type, 5778 unsigned long node_start_pfn, 5779 unsigned long node_end_pfn, 5780 unsigned long *zone_start_pfn, 5781 unsigned long *zone_end_pfn, 5782 unsigned long *zones_size) 5783 { 5784 unsigned int zone; 5785 5786 *zone_start_pfn = node_start_pfn; 5787 for (zone = 0; zone < zone_type; zone++) 5788 *zone_start_pfn += zones_size[zone]; 5789 5790 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type]; 5791 5792 return zones_size[zone_type]; 5793 } 5794 5795 static inline unsigned long __meminit zone_absent_pages_in_node(int nid, 5796 unsigned long zone_type, 5797 unsigned long node_start_pfn, 5798 unsigned long node_end_pfn, 5799 unsigned long *zholes_size) 5800 { 5801 if (!zholes_size) 5802 return 0; 5803 5804 return zholes_size[zone_type]; 5805 } 5806 5807 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 5808 5809 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat, 5810 unsigned long node_start_pfn, 5811 unsigned long node_end_pfn, 5812 unsigned long *zones_size, 5813 unsigned long *zholes_size) 5814 { 5815 unsigned long realtotalpages = 0, totalpages = 0; 5816 enum zone_type i; 5817 5818 for (i = 0; i < MAX_NR_ZONES; i++) { 5819 struct zone *zone = pgdat->node_zones + i; 5820 unsigned long zone_start_pfn, zone_end_pfn; 5821 unsigned long size, real_size; 5822 5823 size = zone_spanned_pages_in_node(pgdat->node_id, i, 5824 node_start_pfn, 5825 node_end_pfn, 5826 &zone_start_pfn, 5827 &zone_end_pfn, 5828 zones_size); 5829 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i, 5830 node_start_pfn, node_end_pfn, 5831 zholes_size); 5832 if (size) 5833 zone->zone_start_pfn = zone_start_pfn; 5834 else 5835 zone->zone_start_pfn = 0; 5836 zone->spanned_pages = size; 5837 zone->present_pages = real_size; 5838 5839 totalpages += size; 5840 realtotalpages += real_size; 5841 } 5842 5843 pgdat->node_spanned_pages = totalpages; 5844 pgdat->node_present_pages = realtotalpages; 5845 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id, 5846 realtotalpages); 5847 } 5848 5849 #ifndef CONFIG_SPARSEMEM 5850 /* 5851 * Calculate the size of the zone->blockflags rounded to an unsigned long 5852 * Start by making sure zonesize is a multiple of pageblock_order by rounding 5853 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally 5854 * round what is now in bits to nearest long in bits, then return it in 5855 * bytes. 5856 */ 5857 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize) 5858 { 5859 unsigned long usemapsize; 5860 5861 zonesize += zone_start_pfn & (pageblock_nr_pages-1); 5862 usemapsize = roundup(zonesize, pageblock_nr_pages); 5863 usemapsize = usemapsize >> pageblock_order; 5864 usemapsize *= NR_PAGEBLOCK_BITS; 5865 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long)); 5866 5867 return usemapsize / 8; 5868 } 5869 5870 static void __init setup_usemap(struct pglist_data *pgdat, 5871 struct zone *zone, 5872 unsigned long zone_start_pfn, 5873 unsigned long zonesize) 5874 { 5875 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize); 5876 zone->pageblock_flags = NULL; 5877 if (usemapsize) 5878 zone->pageblock_flags = 5879 memblock_virt_alloc_node_nopanic(usemapsize, 5880 pgdat->node_id); 5881 } 5882 #else 5883 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone, 5884 unsigned long zone_start_pfn, unsigned long zonesize) {} 5885 #endif /* CONFIG_SPARSEMEM */ 5886 5887 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE 5888 5889 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */ 5890 void __paginginit set_pageblock_order(void) 5891 { 5892 unsigned int order; 5893 5894 /* Check that pageblock_nr_pages has not already been setup */ 5895 if (pageblock_order) 5896 return; 5897 5898 if (HPAGE_SHIFT > PAGE_SHIFT) 5899 order = HUGETLB_PAGE_ORDER; 5900 else 5901 order = MAX_ORDER - 1; 5902 5903 /* 5904 * Assume the largest contiguous order of interest is a huge page. 5905 * This value may be variable depending on boot parameters on IA64 and 5906 * powerpc. 5907 */ 5908 pageblock_order = order; 5909 } 5910 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5911 5912 /* 5913 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order() 5914 * is unused as pageblock_order is set at compile-time. See 5915 * include/linux/pageblock-flags.h for the values of pageblock_order based on 5916 * the kernel config 5917 */ 5918 void __paginginit set_pageblock_order(void) 5919 { 5920 } 5921 5922 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */ 5923 5924 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages, 5925 unsigned long present_pages) 5926 { 5927 unsigned long pages = spanned_pages; 5928 5929 /* 5930 * Provide a more accurate estimation if there are holes within 5931 * the zone and SPARSEMEM is in use. If there are holes within the 5932 * zone, each populated memory region may cost us one or two extra 5933 * memmap pages due to alignment because memmap pages for each 5934 * populated regions may not be naturally aligned on page boundary. 5935 * So the (present_pages >> 4) heuristic is a tradeoff for that. 5936 */ 5937 if (spanned_pages > present_pages + (present_pages >> 4) && 5938 IS_ENABLED(CONFIG_SPARSEMEM)) 5939 pages = present_pages; 5940 5941 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT; 5942 } 5943 5944 /* 5945 * Set up the zone data structures: 5946 * - mark all pages reserved 5947 * - mark all memory queues empty 5948 * - clear the memory bitmaps 5949 * 5950 * NOTE: pgdat should get zeroed by caller. 5951 */ 5952 static void __paginginit free_area_init_core(struct pglist_data *pgdat) 5953 { 5954 enum zone_type j; 5955 int nid = pgdat->node_id; 5956 int ret; 5957 5958 pgdat_resize_init(pgdat); 5959 #ifdef CONFIG_NUMA_BALANCING 5960 spin_lock_init(&pgdat->numabalancing_migrate_lock); 5961 pgdat->numabalancing_migrate_nr_pages = 0; 5962 pgdat->numabalancing_migrate_next_window = jiffies; 5963 #endif 5964 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 5965 spin_lock_init(&pgdat->split_queue_lock); 5966 INIT_LIST_HEAD(&pgdat->split_queue); 5967 pgdat->split_queue_len = 0; 5968 #endif 5969 init_waitqueue_head(&pgdat->kswapd_wait); 5970 init_waitqueue_head(&pgdat->pfmemalloc_wait); 5971 #ifdef CONFIG_COMPACTION 5972 init_waitqueue_head(&pgdat->kcompactd_wait); 5973 #endif 5974 pgdat_page_ext_init(pgdat); 5975 spin_lock_init(&pgdat->lru_lock); 5976 lruvec_init(node_lruvec(pgdat)); 5977 5978 for (j = 0; j < MAX_NR_ZONES; j++) { 5979 struct zone *zone = pgdat->node_zones + j; 5980 unsigned long size, realsize, freesize, memmap_pages; 5981 unsigned long zone_start_pfn = zone->zone_start_pfn; 5982 5983 size = zone->spanned_pages; 5984 realsize = freesize = zone->present_pages; 5985 5986 /* 5987 * Adjust freesize so that it accounts for how much memory 5988 * is used by this zone for memmap. This affects the watermark 5989 * and per-cpu initialisations 5990 */ 5991 memmap_pages = calc_memmap_size(size, realsize); 5992 if (!is_highmem_idx(j)) { 5993 if (freesize >= memmap_pages) { 5994 freesize -= memmap_pages; 5995 if (memmap_pages) 5996 printk(KERN_DEBUG 5997 " %s zone: %lu pages used for memmap\n", 5998 zone_names[j], memmap_pages); 5999 } else 6000 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n", 6001 zone_names[j], memmap_pages, freesize); 6002 } 6003 6004 /* Account for reserved pages */ 6005 if (j == 0 && freesize > dma_reserve) { 6006 freesize -= dma_reserve; 6007 printk(KERN_DEBUG " %s zone: %lu pages reserved\n", 6008 zone_names[0], dma_reserve); 6009 } 6010 6011 if (!is_highmem_idx(j)) 6012 nr_kernel_pages += freesize; 6013 /* Charge for highmem memmap if there are enough kernel pages */ 6014 else if (nr_kernel_pages > memmap_pages * 2) 6015 nr_kernel_pages -= memmap_pages; 6016 nr_all_pages += freesize; 6017 6018 /* 6019 * Set an approximate value for lowmem here, it will be adjusted 6020 * when the bootmem allocator frees pages into the buddy system. 6021 * And all highmem pages will be managed by the buddy system. 6022 */ 6023 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize; 6024 #ifdef CONFIG_NUMA 6025 zone->node = nid; 6026 #endif 6027 zone->name = zone_names[j]; 6028 zone->zone_pgdat = pgdat; 6029 spin_lock_init(&zone->lock); 6030 zone_seqlock_init(zone); 6031 zone_pcp_init(zone); 6032 6033 if (!size) 6034 continue; 6035 6036 set_pageblock_order(); 6037 setup_usemap(pgdat, zone, zone_start_pfn, size); 6038 ret = init_currently_empty_zone(zone, zone_start_pfn, size); 6039 BUG_ON(ret); 6040 memmap_init(size, nid, j, zone_start_pfn); 6041 } 6042 } 6043 6044 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) 6045 { 6046 unsigned long __maybe_unused start = 0; 6047 unsigned long __maybe_unused offset = 0; 6048 6049 /* Skip empty nodes */ 6050 if (!pgdat->node_spanned_pages) 6051 return; 6052 6053 #ifdef CONFIG_FLAT_NODE_MEM_MAP 6054 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1); 6055 offset = pgdat->node_start_pfn - start; 6056 /* ia64 gets its own node_mem_map, before this, without bootmem */ 6057 if (!pgdat->node_mem_map) { 6058 unsigned long size, end; 6059 struct page *map; 6060 6061 /* 6062 * The zone's endpoints aren't required to be MAX_ORDER 6063 * aligned but the node_mem_map endpoints must be in order 6064 * for the buddy allocator to function correctly. 6065 */ 6066 end = pgdat_end_pfn(pgdat); 6067 end = ALIGN(end, MAX_ORDER_NR_PAGES); 6068 size = (end - start) * sizeof(struct page); 6069 map = alloc_remap(pgdat->node_id, size); 6070 if (!map) 6071 map = memblock_virt_alloc_node_nopanic(size, 6072 pgdat->node_id); 6073 pgdat->node_mem_map = map + offset; 6074 } 6075 #ifndef CONFIG_NEED_MULTIPLE_NODES 6076 /* 6077 * With no DISCONTIG, the global mem_map is just set as node 0's 6078 */ 6079 if (pgdat == NODE_DATA(0)) { 6080 mem_map = NODE_DATA(0)->node_mem_map; 6081 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM) 6082 if (page_to_pfn(mem_map) != pgdat->node_start_pfn) 6083 mem_map -= offset; 6084 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 6085 } 6086 #endif 6087 #endif /* CONFIG_FLAT_NODE_MEM_MAP */ 6088 } 6089 6090 void __paginginit free_area_init_node(int nid, unsigned long *zones_size, 6091 unsigned long node_start_pfn, unsigned long *zholes_size) 6092 { 6093 pg_data_t *pgdat = NODE_DATA(nid); 6094 unsigned long start_pfn = 0; 6095 unsigned long end_pfn = 0; 6096 6097 /* pg_data_t should be reset to zero when it's allocated */ 6098 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx); 6099 6100 reset_deferred_meminit(pgdat); 6101 pgdat->node_id = nid; 6102 pgdat->node_start_pfn = node_start_pfn; 6103 pgdat->per_cpu_nodestats = NULL; 6104 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 6105 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 6106 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid, 6107 (u64)start_pfn << PAGE_SHIFT, 6108 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0); 6109 #else 6110 start_pfn = node_start_pfn; 6111 #endif 6112 calculate_node_totalpages(pgdat, start_pfn, end_pfn, 6113 zones_size, zholes_size); 6114 6115 alloc_node_mem_map(pgdat); 6116 #ifdef CONFIG_FLAT_NODE_MEM_MAP 6117 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n", 6118 nid, (unsigned long)pgdat, 6119 (unsigned long)pgdat->node_mem_map); 6120 #endif 6121 6122 free_area_init_core(pgdat); 6123 } 6124 6125 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP 6126 6127 #if MAX_NUMNODES > 1 6128 /* 6129 * Figure out the number of possible node ids. 6130 */ 6131 void __init setup_nr_node_ids(void) 6132 { 6133 unsigned int highest; 6134 6135 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES); 6136 nr_node_ids = highest + 1; 6137 } 6138 #endif 6139 6140 /** 6141 * node_map_pfn_alignment - determine the maximum internode alignment 6142 * 6143 * This function should be called after node map is populated and sorted. 6144 * It calculates the maximum power of two alignment which can distinguish 6145 * all the nodes. 6146 * 6147 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value 6148 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the 6149 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is 6150 * shifted, 1GiB is enough and this function will indicate so. 6151 * 6152 * This is used to test whether pfn -> nid mapping of the chosen memory 6153 * model has fine enough granularity to avoid incorrect mapping for the 6154 * populated node map. 6155 * 6156 * Returns the determined alignment in pfn's. 0 if there is no alignment 6157 * requirement (single node). 6158 */ 6159 unsigned long __init node_map_pfn_alignment(void) 6160 { 6161 unsigned long accl_mask = 0, last_end = 0; 6162 unsigned long start, end, mask; 6163 int last_nid = -1; 6164 int i, nid; 6165 6166 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) { 6167 if (!start || last_nid < 0 || last_nid == nid) { 6168 last_nid = nid; 6169 last_end = end; 6170 continue; 6171 } 6172 6173 /* 6174 * Start with a mask granular enough to pin-point to the 6175 * start pfn and tick off bits one-by-one until it becomes 6176 * too coarse to separate the current node from the last. 6177 */ 6178 mask = ~((1 << __ffs(start)) - 1); 6179 while (mask && last_end <= (start & (mask << 1))) 6180 mask <<= 1; 6181 6182 /* accumulate all internode masks */ 6183 accl_mask |= mask; 6184 } 6185 6186 /* convert mask to number of pages */ 6187 return ~accl_mask + 1; 6188 } 6189 6190 /* Find the lowest pfn for a node */ 6191 static unsigned long __init find_min_pfn_for_node(int nid) 6192 { 6193 unsigned long min_pfn = ULONG_MAX; 6194 unsigned long start_pfn; 6195 int i; 6196 6197 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL) 6198 min_pfn = min(min_pfn, start_pfn); 6199 6200 if (min_pfn == ULONG_MAX) { 6201 pr_warn("Could not find start_pfn for node %d\n", nid); 6202 return 0; 6203 } 6204 6205 return min_pfn; 6206 } 6207 6208 /** 6209 * find_min_pfn_with_active_regions - Find the minimum PFN registered 6210 * 6211 * It returns the minimum PFN based on information provided via 6212 * memblock_set_node(). 6213 */ 6214 unsigned long __init find_min_pfn_with_active_regions(void) 6215 { 6216 return find_min_pfn_for_node(MAX_NUMNODES); 6217 } 6218 6219 /* 6220 * early_calculate_totalpages() 6221 * Sum pages in active regions for movable zone. 6222 * Populate N_MEMORY for calculating usable_nodes. 6223 */ 6224 static unsigned long __init early_calculate_totalpages(void) 6225 { 6226 unsigned long totalpages = 0; 6227 unsigned long start_pfn, end_pfn; 6228 int i, nid; 6229 6230 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) { 6231 unsigned long pages = end_pfn - start_pfn; 6232 6233 totalpages += pages; 6234 if (pages) 6235 node_set_state(nid, N_MEMORY); 6236 } 6237 return totalpages; 6238 } 6239 6240 /* 6241 * Find the PFN the Movable zone begins in each node. Kernel memory 6242 * is spread evenly between nodes as long as the nodes have enough 6243 * memory. When they don't, some nodes will have more kernelcore than 6244 * others 6245 */ 6246 static void __init find_zone_movable_pfns_for_nodes(void) 6247 { 6248 int i, nid; 6249 unsigned long usable_startpfn; 6250 unsigned long kernelcore_node, kernelcore_remaining; 6251 /* save the state before borrow the nodemask */ 6252 nodemask_t saved_node_state = node_states[N_MEMORY]; 6253 unsigned long totalpages = early_calculate_totalpages(); 6254 int usable_nodes = nodes_weight(node_states[N_MEMORY]); 6255 struct memblock_region *r; 6256 6257 /* Need to find movable_zone earlier when movable_node is specified. */ 6258 find_usable_zone_for_movable(); 6259 6260 /* 6261 * If movable_node is specified, ignore kernelcore and movablecore 6262 * options. 6263 */ 6264 if (movable_node_is_enabled()) { 6265 for_each_memblock(memory, r) { 6266 if (!memblock_is_hotpluggable(r)) 6267 continue; 6268 6269 nid = r->nid; 6270 6271 usable_startpfn = PFN_DOWN(r->base); 6272 zone_movable_pfn[nid] = zone_movable_pfn[nid] ? 6273 min(usable_startpfn, zone_movable_pfn[nid]) : 6274 usable_startpfn; 6275 } 6276 6277 goto out2; 6278 } 6279 6280 /* 6281 * If kernelcore=mirror is specified, ignore movablecore option 6282 */ 6283 if (mirrored_kernelcore) { 6284 bool mem_below_4gb_not_mirrored = false; 6285 6286 for_each_memblock(memory, r) { 6287 if (memblock_is_mirror(r)) 6288 continue; 6289 6290 nid = r->nid; 6291 6292 usable_startpfn = memblock_region_memory_base_pfn(r); 6293 6294 if (usable_startpfn < 0x100000) { 6295 mem_below_4gb_not_mirrored = true; 6296 continue; 6297 } 6298 6299 zone_movable_pfn[nid] = zone_movable_pfn[nid] ? 6300 min(usable_startpfn, zone_movable_pfn[nid]) : 6301 usable_startpfn; 6302 } 6303 6304 if (mem_below_4gb_not_mirrored) 6305 pr_warn("This configuration results in unmirrored kernel memory."); 6306 6307 goto out2; 6308 } 6309 6310 /* 6311 * If movablecore=nn[KMG] was specified, calculate what size of 6312 * kernelcore that corresponds so that memory usable for 6313 * any allocation type is evenly spread. If both kernelcore 6314 * and movablecore are specified, then the value of kernelcore 6315 * will be used for required_kernelcore if it's greater than 6316 * what movablecore would have allowed. 6317 */ 6318 if (required_movablecore) { 6319 unsigned long corepages; 6320 6321 /* 6322 * Round-up so that ZONE_MOVABLE is at least as large as what 6323 * was requested by the user 6324 */ 6325 required_movablecore = 6326 roundup(required_movablecore, MAX_ORDER_NR_PAGES); 6327 required_movablecore = min(totalpages, required_movablecore); 6328 corepages = totalpages - required_movablecore; 6329 6330 required_kernelcore = max(required_kernelcore, corepages); 6331 } 6332 6333 /* 6334 * If kernelcore was not specified or kernelcore size is larger 6335 * than totalpages, there is no ZONE_MOVABLE. 6336 */ 6337 if (!required_kernelcore || required_kernelcore >= totalpages) 6338 goto out; 6339 6340 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */ 6341 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone]; 6342 6343 restart: 6344 /* Spread kernelcore memory as evenly as possible throughout nodes */ 6345 kernelcore_node = required_kernelcore / usable_nodes; 6346 for_each_node_state(nid, N_MEMORY) { 6347 unsigned long start_pfn, end_pfn; 6348 6349 /* 6350 * Recalculate kernelcore_node if the division per node 6351 * now exceeds what is necessary to satisfy the requested 6352 * amount of memory for the kernel 6353 */ 6354 if (required_kernelcore < kernelcore_node) 6355 kernelcore_node = required_kernelcore / usable_nodes; 6356 6357 /* 6358 * As the map is walked, we track how much memory is usable 6359 * by the kernel using kernelcore_remaining. When it is 6360 * 0, the rest of the node is usable by ZONE_MOVABLE 6361 */ 6362 kernelcore_remaining = kernelcore_node; 6363 6364 /* Go through each range of PFNs within this node */ 6365 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) { 6366 unsigned long size_pages; 6367 6368 start_pfn = max(start_pfn, zone_movable_pfn[nid]); 6369 if (start_pfn >= end_pfn) 6370 continue; 6371 6372 /* Account for what is only usable for kernelcore */ 6373 if (start_pfn < usable_startpfn) { 6374 unsigned long kernel_pages; 6375 kernel_pages = min(end_pfn, usable_startpfn) 6376 - start_pfn; 6377 6378 kernelcore_remaining -= min(kernel_pages, 6379 kernelcore_remaining); 6380 required_kernelcore -= min(kernel_pages, 6381 required_kernelcore); 6382 6383 /* Continue if range is now fully accounted */ 6384 if (end_pfn <= usable_startpfn) { 6385 6386 /* 6387 * Push zone_movable_pfn to the end so 6388 * that if we have to rebalance 6389 * kernelcore across nodes, we will 6390 * not double account here 6391 */ 6392 zone_movable_pfn[nid] = end_pfn; 6393 continue; 6394 } 6395 start_pfn = usable_startpfn; 6396 } 6397 6398 /* 6399 * The usable PFN range for ZONE_MOVABLE is from 6400 * start_pfn->end_pfn. Calculate size_pages as the 6401 * number of pages used as kernelcore 6402 */ 6403 size_pages = end_pfn - start_pfn; 6404 if (size_pages > kernelcore_remaining) 6405 size_pages = kernelcore_remaining; 6406 zone_movable_pfn[nid] = start_pfn + size_pages; 6407 6408 /* 6409 * Some kernelcore has been met, update counts and 6410 * break if the kernelcore for this node has been 6411 * satisfied 6412 */ 6413 required_kernelcore -= min(required_kernelcore, 6414 size_pages); 6415 kernelcore_remaining -= size_pages; 6416 if (!kernelcore_remaining) 6417 break; 6418 } 6419 } 6420 6421 /* 6422 * If there is still required_kernelcore, we do another pass with one 6423 * less node in the count. This will push zone_movable_pfn[nid] further 6424 * along on the nodes that still have memory until kernelcore is 6425 * satisfied 6426 */ 6427 usable_nodes--; 6428 if (usable_nodes && required_kernelcore > usable_nodes) 6429 goto restart; 6430 6431 out2: 6432 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */ 6433 for (nid = 0; nid < MAX_NUMNODES; nid++) 6434 zone_movable_pfn[nid] = 6435 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES); 6436 6437 out: 6438 /* restore the node_state */ 6439 node_states[N_MEMORY] = saved_node_state; 6440 } 6441 6442 /* Any regular or high memory on that node ? */ 6443 static void check_for_memory(pg_data_t *pgdat, int nid) 6444 { 6445 enum zone_type zone_type; 6446 6447 if (N_MEMORY == N_NORMAL_MEMORY) 6448 return; 6449 6450 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) { 6451 struct zone *zone = &pgdat->node_zones[zone_type]; 6452 if (populated_zone(zone)) { 6453 node_set_state(nid, N_HIGH_MEMORY); 6454 if (N_NORMAL_MEMORY != N_HIGH_MEMORY && 6455 zone_type <= ZONE_NORMAL) 6456 node_set_state(nid, N_NORMAL_MEMORY); 6457 break; 6458 } 6459 } 6460 } 6461 6462 /** 6463 * free_area_init_nodes - Initialise all pg_data_t and zone data 6464 * @max_zone_pfn: an array of max PFNs for each zone 6465 * 6466 * This will call free_area_init_node() for each active node in the system. 6467 * Using the page ranges provided by memblock_set_node(), the size of each 6468 * zone in each node and their holes is calculated. If the maximum PFN 6469 * between two adjacent zones match, it is assumed that the zone is empty. 6470 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed 6471 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone 6472 * starts where the previous one ended. For example, ZONE_DMA32 starts 6473 * at arch_max_dma_pfn. 6474 */ 6475 void __init free_area_init_nodes(unsigned long *max_zone_pfn) 6476 { 6477 unsigned long start_pfn, end_pfn; 6478 int i, nid; 6479 6480 /* Record where the zone boundaries are */ 6481 memset(arch_zone_lowest_possible_pfn, 0, 6482 sizeof(arch_zone_lowest_possible_pfn)); 6483 memset(arch_zone_highest_possible_pfn, 0, 6484 sizeof(arch_zone_highest_possible_pfn)); 6485 6486 start_pfn = find_min_pfn_with_active_regions(); 6487 6488 for (i = 0; i < MAX_NR_ZONES; i++) { 6489 if (i == ZONE_MOVABLE) 6490 continue; 6491 6492 end_pfn = max(max_zone_pfn[i], start_pfn); 6493 arch_zone_lowest_possible_pfn[i] = start_pfn; 6494 arch_zone_highest_possible_pfn[i] = end_pfn; 6495 6496 start_pfn = end_pfn; 6497 } 6498 6499 /* Find the PFNs that ZONE_MOVABLE begins at in each node */ 6500 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn)); 6501 find_zone_movable_pfns_for_nodes(); 6502 6503 /* Print out the zone ranges */ 6504 pr_info("Zone ranges:\n"); 6505 for (i = 0; i < MAX_NR_ZONES; i++) { 6506 if (i == ZONE_MOVABLE) 6507 continue; 6508 pr_info(" %-8s ", zone_names[i]); 6509 if (arch_zone_lowest_possible_pfn[i] == 6510 arch_zone_highest_possible_pfn[i]) 6511 pr_cont("empty\n"); 6512 else 6513 pr_cont("[mem %#018Lx-%#018Lx]\n", 6514 (u64)arch_zone_lowest_possible_pfn[i] 6515 << PAGE_SHIFT, 6516 ((u64)arch_zone_highest_possible_pfn[i] 6517 << PAGE_SHIFT) - 1); 6518 } 6519 6520 /* Print out the PFNs ZONE_MOVABLE begins at in each node */ 6521 pr_info("Movable zone start for each node\n"); 6522 for (i = 0; i < MAX_NUMNODES; i++) { 6523 if (zone_movable_pfn[i]) 6524 pr_info(" Node %d: %#018Lx\n", i, 6525 (u64)zone_movable_pfn[i] << PAGE_SHIFT); 6526 } 6527 6528 /* Print out the early node map */ 6529 pr_info("Early memory node ranges\n"); 6530 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) 6531 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid, 6532 (u64)start_pfn << PAGE_SHIFT, 6533 ((u64)end_pfn << PAGE_SHIFT) - 1); 6534 6535 /* Initialise every node */ 6536 mminit_verify_pageflags_layout(); 6537 setup_nr_node_ids(); 6538 for_each_online_node(nid) { 6539 pg_data_t *pgdat = NODE_DATA(nid); 6540 free_area_init_node(nid, NULL, 6541 find_min_pfn_for_node(nid), NULL); 6542 6543 /* Any memory on that node */ 6544 if (pgdat->node_present_pages) 6545 node_set_state(nid, N_MEMORY); 6546 check_for_memory(pgdat, nid); 6547 } 6548 } 6549 6550 static int __init cmdline_parse_core(char *p, unsigned long *core) 6551 { 6552 unsigned long long coremem; 6553 if (!p) 6554 return -EINVAL; 6555 6556 coremem = memparse(p, &p); 6557 *core = coremem >> PAGE_SHIFT; 6558 6559 /* Paranoid check that UL is enough for the coremem value */ 6560 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX); 6561 6562 return 0; 6563 } 6564 6565 /* 6566 * kernelcore=size sets the amount of memory for use for allocations that 6567 * cannot be reclaimed or migrated. 6568 */ 6569 static int __init cmdline_parse_kernelcore(char *p) 6570 { 6571 /* parse kernelcore=mirror */ 6572 if (parse_option_str(p, "mirror")) { 6573 mirrored_kernelcore = true; 6574 return 0; 6575 } 6576 6577 return cmdline_parse_core(p, &required_kernelcore); 6578 } 6579 6580 /* 6581 * movablecore=size sets the amount of memory for use for allocations that 6582 * can be reclaimed or migrated. 6583 */ 6584 static int __init cmdline_parse_movablecore(char *p) 6585 { 6586 return cmdline_parse_core(p, &required_movablecore); 6587 } 6588 6589 early_param("kernelcore", cmdline_parse_kernelcore); 6590 early_param("movablecore", cmdline_parse_movablecore); 6591 6592 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */ 6593 6594 void adjust_managed_page_count(struct page *page, long count) 6595 { 6596 spin_lock(&managed_page_count_lock); 6597 page_zone(page)->managed_pages += count; 6598 totalram_pages += count; 6599 #ifdef CONFIG_HIGHMEM 6600 if (PageHighMem(page)) 6601 totalhigh_pages += count; 6602 #endif 6603 spin_unlock(&managed_page_count_lock); 6604 } 6605 EXPORT_SYMBOL(adjust_managed_page_count); 6606 6607 unsigned long free_reserved_area(void *start, void *end, int poison, char *s) 6608 { 6609 void *pos; 6610 unsigned long pages = 0; 6611 6612 start = (void *)PAGE_ALIGN((unsigned long)start); 6613 end = (void *)((unsigned long)end & PAGE_MASK); 6614 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) { 6615 if ((unsigned int)poison <= 0xFF) 6616 memset(pos, poison, PAGE_SIZE); 6617 free_reserved_page(virt_to_page(pos)); 6618 } 6619 6620 if (pages && s) 6621 pr_info("Freeing %s memory: %ldK\n", 6622 s, pages << (PAGE_SHIFT - 10)); 6623 6624 return pages; 6625 } 6626 EXPORT_SYMBOL(free_reserved_area); 6627 6628 #ifdef CONFIG_HIGHMEM 6629 void free_highmem_page(struct page *page) 6630 { 6631 __free_reserved_page(page); 6632 totalram_pages++; 6633 page_zone(page)->managed_pages++; 6634 totalhigh_pages++; 6635 } 6636 #endif 6637 6638 6639 void __init mem_init_print_info(const char *str) 6640 { 6641 unsigned long physpages, codesize, datasize, rosize, bss_size; 6642 unsigned long init_code_size, init_data_size; 6643 6644 physpages = get_num_physpages(); 6645 codesize = _etext - _stext; 6646 datasize = _edata - _sdata; 6647 rosize = __end_rodata - __start_rodata; 6648 bss_size = __bss_stop - __bss_start; 6649 init_data_size = __init_end - __init_begin; 6650 init_code_size = _einittext - _sinittext; 6651 6652 /* 6653 * Detect special cases and adjust section sizes accordingly: 6654 * 1) .init.* may be embedded into .data sections 6655 * 2) .init.text.* may be out of [__init_begin, __init_end], 6656 * please refer to arch/tile/kernel/vmlinux.lds.S. 6657 * 3) .rodata.* may be embedded into .text or .data sections. 6658 */ 6659 #define adj_init_size(start, end, size, pos, adj) \ 6660 do { \ 6661 if (start <= pos && pos < end && size > adj) \ 6662 size -= adj; \ 6663 } while (0) 6664 6665 adj_init_size(__init_begin, __init_end, init_data_size, 6666 _sinittext, init_code_size); 6667 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size); 6668 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size); 6669 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize); 6670 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize); 6671 6672 #undef adj_init_size 6673 6674 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved" 6675 #ifdef CONFIG_HIGHMEM 6676 ", %luK highmem" 6677 #endif 6678 "%s%s)\n", 6679 nr_free_pages() << (PAGE_SHIFT - 10), 6680 physpages << (PAGE_SHIFT - 10), 6681 codesize >> 10, datasize >> 10, rosize >> 10, 6682 (init_data_size + init_code_size) >> 10, bss_size >> 10, 6683 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10), 6684 totalcma_pages << (PAGE_SHIFT - 10), 6685 #ifdef CONFIG_HIGHMEM 6686 totalhigh_pages << (PAGE_SHIFT - 10), 6687 #endif 6688 str ? ", " : "", str ? str : ""); 6689 } 6690 6691 /** 6692 * set_dma_reserve - set the specified number of pages reserved in the first zone 6693 * @new_dma_reserve: The number of pages to mark reserved 6694 * 6695 * The per-cpu batchsize and zone watermarks are determined by managed_pages. 6696 * In the DMA zone, a significant percentage may be consumed by kernel image 6697 * and other unfreeable allocations which can skew the watermarks badly. This 6698 * function may optionally be used to account for unfreeable pages in the 6699 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and 6700 * smaller per-cpu batchsize. 6701 */ 6702 void __init set_dma_reserve(unsigned long new_dma_reserve) 6703 { 6704 dma_reserve = new_dma_reserve; 6705 } 6706 6707 void __init free_area_init(unsigned long *zones_size) 6708 { 6709 free_area_init_node(0, zones_size, 6710 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL); 6711 } 6712 6713 static int page_alloc_cpu_dead(unsigned int cpu) 6714 { 6715 6716 lru_add_drain_cpu(cpu); 6717 drain_pages(cpu); 6718 6719 /* 6720 * Spill the event counters of the dead processor 6721 * into the current processors event counters. 6722 * This artificially elevates the count of the current 6723 * processor. 6724 */ 6725 vm_events_fold_cpu(cpu); 6726 6727 /* 6728 * Zero the differential counters of the dead processor 6729 * so that the vm statistics are consistent. 6730 * 6731 * This is only okay since the processor is dead and cannot 6732 * race with what we are doing. 6733 */ 6734 cpu_vm_stats_fold(cpu); 6735 return 0; 6736 } 6737 6738 void __init page_alloc_init(void) 6739 { 6740 int ret; 6741 6742 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD, 6743 "mm/page_alloc:dead", NULL, 6744 page_alloc_cpu_dead); 6745 WARN_ON(ret < 0); 6746 } 6747 6748 /* 6749 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio 6750 * or min_free_kbytes changes. 6751 */ 6752 static void calculate_totalreserve_pages(void) 6753 { 6754 struct pglist_data *pgdat; 6755 unsigned long reserve_pages = 0; 6756 enum zone_type i, j; 6757 6758 for_each_online_pgdat(pgdat) { 6759 6760 pgdat->totalreserve_pages = 0; 6761 6762 for (i = 0; i < MAX_NR_ZONES; i++) { 6763 struct zone *zone = pgdat->node_zones + i; 6764 long max = 0; 6765 6766 /* Find valid and maximum lowmem_reserve in the zone */ 6767 for (j = i; j < MAX_NR_ZONES; j++) { 6768 if (zone->lowmem_reserve[j] > max) 6769 max = zone->lowmem_reserve[j]; 6770 } 6771 6772 /* we treat the high watermark as reserved pages. */ 6773 max += high_wmark_pages(zone); 6774 6775 if (max > zone->managed_pages) 6776 max = zone->managed_pages; 6777 6778 pgdat->totalreserve_pages += max; 6779 6780 reserve_pages += max; 6781 } 6782 } 6783 totalreserve_pages = reserve_pages; 6784 } 6785 6786 /* 6787 * setup_per_zone_lowmem_reserve - called whenever 6788 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone 6789 * has a correct pages reserved value, so an adequate number of 6790 * pages are left in the zone after a successful __alloc_pages(). 6791 */ 6792 static void setup_per_zone_lowmem_reserve(void) 6793 { 6794 struct pglist_data *pgdat; 6795 enum zone_type j, idx; 6796 6797 for_each_online_pgdat(pgdat) { 6798 for (j = 0; j < MAX_NR_ZONES; j++) { 6799 struct zone *zone = pgdat->node_zones + j; 6800 unsigned long managed_pages = zone->managed_pages; 6801 6802 zone->lowmem_reserve[j] = 0; 6803 6804 idx = j; 6805 while (idx) { 6806 struct zone *lower_zone; 6807 6808 idx--; 6809 6810 if (sysctl_lowmem_reserve_ratio[idx] < 1) 6811 sysctl_lowmem_reserve_ratio[idx] = 1; 6812 6813 lower_zone = pgdat->node_zones + idx; 6814 lower_zone->lowmem_reserve[j] = managed_pages / 6815 sysctl_lowmem_reserve_ratio[idx]; 6816 managed_pages += lower_zone->managed_pages; 6817 } 6818 } 6819 } 6820 6821 /* update totalreserve_pages */ 6822 calculate_totalreserve_pages(); 6823 } 6824 6825 static void __setup_per_zone_wmarks(void) 6826 { 6827 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10); 6828 unsigned long lowmem_pages = 0; 6829 struct zone *zone; 6830 unsigned long flags; 6831 6832 /* Calculate total number of !ZONE_HIGHMEM pages */ 6833 for_each_zone(zone) { 6834 if (!is_highmem(zone)) 6835 lowmem_pages += zone->managed_pages; 6836 } 6837 6838 for_each_zone(zone) { 6839 u64 tmp; 6840 6841 spin_lock_irqsave(&zone->lock, flags); 6842 tmp = (u64)pages_min * zone->managed_pages; 6843 do_div(tmp, lowmem_pages); 6844 if (is_highmem(zone)) { 6845 /* 6846 * __GFP_HIGH and PF_MEMALLOC allocations usually don't 6847 * need highmem pages, so cap pages_min to a small 6848 * value here. 6849 * 6850 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN) 6851 * deltas control asynch page reclaim, and so should 6852 * not be capped for highmem. 6853 */ 6854 unsigned long min_pages; 6855 6856 min_pages = zone->managed_pages / 1024; 6857 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL); 6858 zone->watermark[WMARK_MIN] = min_pages; 6859 } else { 6860 /* 6861 * If it's a lowmem zone, reserve a number of pages 6862 * proportionate to the zone's size. 6863 */ 6864 zone->watermark[WMARK_MIN] = tmp; 6865 } 6866 6867 /* 6868 * Set the kswapd watermarks distance according to the 6869 * scale factor in proportion to available memory, but 6870 * ensure a minimum size on small systems. 6871 */ 6872 tmp = max_t(u64, tmp >> 2, 6873 mult_frac(zone->managed_pages, 6874 watermark_scale_factor, 10000)); 6875 6876 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp; 6877 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2; 6878 6879 spin_unlock_irqrestore(&zone->lock, flags); 6880 } 6881 6882 /* update totalreserve_pages */ 6883 calculate_totalreserve_pages(); 6884 } 6885 6886 /** 6887 * setup_per_zone_wmarks - called when min_free_kbytes changes 6888 * or when memory is hot-{added|removed} 6889 * 6890 * Ensures that the watermark[min,low,high] values for each zone are set 6891 * correctly with respect to min_free_kbytes. 6892 */ 6893 void setup_per_zone_wmarks(void) 6894 { 6895 mutex_lock(&zonelists_mutex); 6896 __setup_per_zone_wmarks(); 6897 mutex_unlock(&zonelists_mutex); 6898 } 6899 6900 /* 6901 * Initialise min_free_kbytes. 6902 * 6903 * For small machines we want it small (128k min). For large machines 6904 * we want it large (64MB max). But it is not linear, because network 6905 * bandwidth does not increase linearly with machine size. We use 6906 * 6907 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy: 6908 * min_free_kbytes = sqrt(lowmem_kbytes * 16) 6909 * 6910 * which yields 6911 * 6912 * 16MB: 512k 6913 * 32MB: 724k 6914 * 64MB: 1024k 6915 * 128MB: 1448k 6916 * 256MB: 2048k 6917 * 512MB: 2896k 6918 * 1024MB: 4096k 6919 * 2048MB: 5792k 6920 * 4096MB: 8192k 6921 * 8192MB: 11584k 6922 * 16384MB: 16384k 6923 */ 6924 int __meminit init_per_zone_wmark_min(void) 6925 { 6926 unsigned long lowmem_kbytes; 6927 int new_min_free_kbytes; 6928 6929 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10); 6930 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16); 6931 6932 if (new_min_free_kbytes > user_min_free_kbytes) { 6933 min_free_kbytes = new_min_free_kbytes; 6934 if (min_free_kbytes < 128) 6935 min_free_kbytes = 128; 6936 if (min_free_kbytes > 65536) 6937 min_free_kbytes = 65536; 6938 } else { 6939 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n", 6940 new_min_free_kbytes, user_min_free_kbytes); 6941 } 6942 setup_per_zone_wmarks(); 6943 refresh_zone_stat_thresholds(); 6944 setup_per_zone_lowmem_reserve(); 6945 6946 #ifdef CONFIG_NUMA 6947 setup_min_unmapped_ratio(); 6948 setup_min_slab_ratio(); 6949 #endif 6950 6951 return 0; 6952 } 6953 core_initcall(init_per_zone_wmark_min) 6954 6955 /* 6956 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so 6957 * that we can call two helper functions whenever min_free_kbytes 6958 * changes. 6959 */ 6960 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write, 6961 void __user *buffer, size_t *length, loff_t *ppos) 6962 { 6963 int rc; 6964 6965 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6966 if (rc) 6967 return rc; 6968 6969 if (write) { 6970 user_min_free_kbytes = min_free_kbytes; 6971 setup_per_zone_wmarks(); 6972 } 6973 return 0; 6974 } 6975 6976 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write, 6977 void __user *buffer, size_t *length, loff_t *ppos) 6978 { 6979 int rc; 6980 6981 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 6982 if (rc) 6983 return rc; 6984 6985 if (write) 6986 setup_per_zone_wmarks(); 6987 6988 return 0; 6989 } 6990 6991 #ifdef CONFIG_NUMA 6992 static void setup_min_unmapped_ratio(void) 6993 { 6994 pg_data_t *pgdat; 6995 struct zone *zone; 6996 6997 for_each_online_pgdat(pgdat) 6998 pgdat->min_unmapped_pages = 0; 6999 7000 for_each_zone(zone) 7001 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages * 7002 sysctl_min_unmapped_ratio) / 100; 7003 } 7004 7005 7006 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write, 7007 void __user *buffer, size_t *length, loff_t *ppos) 7008 { 7009 int rc; 7010 7011 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 7012 if (rc) 7013 return rc; 7014 7015 setup_min_unmapped_ratio(); 7016 7017 return 0; 7018 } 7019 7020 static void setup_min_slab_ratio(void) 7021 { 7022 pg_data_t *pgdat; 7023 struct zone *zone; 7024 7025 for_each_online_pgdat(pgdat) 7026 pgdat->min_slab_pages = 0; 7027 7028 for_each_zone(zone) 7029 zone->zone_pgdat->min_slab_pages += (zone->managed_pages * 7030 sysctl_min_slab_ratio) / 100; 7031 } 7032 7033 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write, 7034 void __user *buffer, size_t *length, loff_t *ppos) 7035 { 7036 int rc; 7037 7038 rc = proc_dointvec_minmax(table, write, buffer, length, ppos); 7039 if (rc) 7040 return rc; 7041 7042 setup_min_slab_ratio(); 7043 7044 return 0; 7045 } 7046 #endif 7047 7048 /* 7049 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around 7050 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve() 7051 * whenever sysctl_lowmem_reserve_ratio changes. 7052 * 7053 * The reserve ratio obviously has absolutely no relation with the 7054 * minimum watermarks. The lowmem reserve ratio can only make sense 7055 * if in function of the boot time zone sizes. 7056 */ 7057 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write, 7058 void __user *buffer, size_t *length, loff_t *ppos) 7059 { 7060 proc_dointvec_minmax(table, write, buffer, length, ppos); 7061 setup_per_zone_lowmem_reserve(); 7062 return 0; 7063 } 7064 7065 /* 7066 * percpu_pagelist_fraction - changes the pcp->high for each zone on each 7067 * cpu. It is the fraction of total pages in each zone that a hot per cpu 7068 * pagelist can have before it gets flushed back to buddy allocator. 7069 */ 7070 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write, 7071 void __user *buffer, size_t *length, loff_t *ppos) 7072 { 7073 struct zone *zone; 7074 int old_percpu_pagelist_fraction; 7075 int ret; 7076 7077 mutex_lock(&pcp_batch_high_lock); 7078 old_percpu_pagelist_fraction = percpu_pagelist_fraction; 7079 7080 ret = proc_dointvec_minmax(table, write, buffer, length, ppos); 7081 if (!write || ret < 0) 7082 goto out; 7083 7084 /* Sanity checking to avoid pcp imbalance */ 7085 if (percpu_pagelist_fraction && 7086 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) { 7087 percpu_pagelist_fraction = old_percpu_pagelist_fraction; 7088 ret = -EINVAL; 7089 goto out; 7090 } 7091 7092 /* No change? */ 7093 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction) 7094 goto out; 7095 7096 for_each_populated_zone(zone) { 7097 unsigned int cpu; 7098 7099 for_each_possible_cpu(cpu) 7100 pageset_set_high_and_batch(zone, 7101 per_cpu_ptr(zone->pageset, cpu)); 7102 } 7103 out: 7104 mutex_unlock(&pcp_batch_high_lock); 7105 return ret; 7106 } 7107 7108 #ifdef CONFIG_NUMA 7109 int hashdist = HASHDIST_DEFAULT; 7110 7111 static int __init set_hashdist(char *str) 7112 { 7113 if (!str) 7114 return 0; 7115 hashdist = simple_strtoul(str, &str, 0); 7116 return 1; 7117 } 7118 __setup("hashdist=", set_hashdist); 7119 #endif 7120 7121 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES 7122 /* 7123 * Returns the number of pages that arch has reserved but 7124 * is not known to alloc_large_system_hash(). 7125 */ 7126 static unsigned long __init arch_reserved_kernel_pages(void) 7127 { 7128 return 0; 7129 } 7130 #endif 7131 7132 /* 7133 * allocate a large system hash table from bootmem 7134 * - it is assumed that the hash table must contain an exact power-of-2 7135 * quantity of entries 7136 * - limit is the number of hash buckets, not the total allocation size 7137 */ 7138 void *__init alloc_large_system_hash(const char *tablename, 7139 unsigned long bucketsize, 7140 unsigned long numentries, 7141 int scale, 7142 int flags, 7143 unsigned int *_hash_shift, 7144 unsigned int *_hash_mask, 7145 unsigned long low_limit, 7146 unsigned long high_limit) 7147 { 7148 unsigned long long max = high_limit; 7149 unsigned long log2qty, size; 7150 void *table = NULL; 7151 7152 /* allow the kernel cmdline to have a say */ 7153 if (!numentries) { 7154 /* round applicable memory size up to nearest megabyte */ 7155 numentries = nr_kernel_pages; 7156 numentries -= arch_reserved_kernel_pages(); 7157 7158 /* It isn't necessary when PAGE_SIZE >= 1MB */ 7159 if (PAGE_SHIFT < 20) 7160 numentries = round_up(numentries, (1<<20)/PAGE_SIZE); 7161 7162 /* limit to 1 bucket per 2^scale bytes of low memory */ 7163 if (scale > PAGE_SHIFT) 7164 numentries >>= (scale - PAGE_SHIFT); 7165 else 7166 numentries <<= (PAGE_SHIFT - scale); 7167 7168 /* Make sure we've got at least a 0-order allocation.. */ 7169 if (unlikely(flags & HASH_SMALL)) { 7170 /* Makes no sense without HASH_EARLY */ 7171 WARN_ON(!(flags & HASH_EARLY)); 7172 if (!(numentries >> *_hash_shift)) { 7173 numentries = 1UL << *_hash_shift; 7174 BUG_ON(!numentries); 7175 } 7176 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE)) 7177 numentries = PAGE_SIZE / bucketsize; 7178 } 7179 numentries = roundup_pow_of_two(numentries); 7180 7181 /* limit allocation size to 1/16 total memory by default */ 7182 if (max == 0) { 7183 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4; 7184 do_div(max, bucketsize); 7185 } 7186 max = min(max, 0x80000000ULL); 7187 7188 if (numentries < low_limit) 7189 numentries = low_limit; 7190 if (numentries > max) 7191 numentries = max; 7192 7193 log2qty = ilog2(numentries); 7194 7195 do { 7196 size = bucketsize << log2qty; 7197 if (flags & HASH_EARLY) 7198 table = memblock_virt_alloc_nopanic(size, 0); 7199 else if (hashdist) 7200 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL); 7201 else { 7202 /* 7203 * If bucketsize is not a power-of-two, we may free 7204 * some pages at the end of hash table which 7205 * alloc_pages_exact() automatically does 7206 */ 7207 if (get_order(size) < MAX_ORDER) { 7208 table = alloc_pages_exact(size, GFP_ATOMIC); 7209 kmemleak_alloc(table, size, 1, GFP_ATOMIC); 7210 } 7211 } 7212 } while (!table && size > PAGE_SIZE && --log2qty); 7213 7214 if (!table) 7215 panic("Failed to allocate %s hash table\n", tablename); 7216 7217 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n", 7218 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size); 7219 7220 if (_hash_shift) 7221 *_hash_shift = log2qty; 7222 if (_hash_mask) 7223 *_hash_mask = (1 << log2qty) - 1; 7224 7225 return table; 7226 } 7227 7228 /* 7229 * This function checks whether pageblock includes unmovable pages or not. 7230 * If @count is not zero, it is okay to include less @count unmovable pages 7231 * 7232 * PageLRU check without isolation or lru_lock could race so that 7233 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable 7234 * check without lock_page also may miss some movable non-lru pages at 7235 * race condition. So you can't expect this function should be exact. 7236 */ 7237 bool has_unmovable_pages(struct zone *zone, struct page *page, int count, 7238 bool skip_hwpoisoned_pages) 7239 { 7240 unsigned long pfn, iter, found; 7241 int mt; 7242 7243 /* 7244 * For avoiding noise data, lru_add_drain_all() should be called 7245 * If ZONE_MOVABLE, the zone never contains unmovable pages 7246 */ 7247 if (zone_idx(zone) == ZONE_MOVABLE) 7248 return false; 7249 mt = get_pageblock_migratetype(page); 7250 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt)) 7251 return false; 7252 7253 pfn = page_to_pfn(page); 7254 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) { 7255 unsigned long check = pfn + iter; 7256 7257 if (!pfn_valid_within(check)) 7258 continue; 7259 7260 page = pfn_to_page(check); 7261 7262 /* 7263 * Hugepages are not in LRU lists, but they're movable. 7264 * We need not scan over tail pages bacause we don't 7265 * handle each tail page individually in migration. 7266 */ 7267 if (PageHuge(page)) { 7268 iter = round_up(iter + 1, 1<<compound_order(page)) - 1; 7269 continue; 7270 } 7271 7272 /* 7273 * We can't use page_count without pin a page 7274 * because another CPU can free compound page. 7275 * This check already skips compound tails of THP 7276 * because their page->_refcount is zero at all time. 7277 */ 7278 if (!page_ref_count(page)) { 7279 if (PageBuddy(page)) 7280 iter += (1 << page_order(page)) - 1; 7281 continue; 7282 } 7283 7284 /* 7285 * The HWPoisoned page may be not in buddy system, and 7286 * page_count() is not 0. 7287 */ 7288 if (skip_hwpoisoned_pages && PageHWPoison(page)) 7289 continue; 7290 7291 if (__PageMovable(page)) 7292 continue; 7293 7294 if (!PageLRU(page)) 7295 found++; 7296 /* 7297 * If there are RECLAIMABLE pages, we need to check 7298 * it. But now, memory offline itself doesn't call 7299 * shrink_node_slabs() and it still to be fixed. 7300 */ 7301 /* 7302 * If the page is not RAM, page_count()should be 0. 7303 * we don't need more check. This is an _used_ not-movable page. 7304 * 7305 * The problematic thing here is PG_reserved pages. PG_reserved 7306 * is set to both of a memory hole page and a _used_ kernel 7307 * page at boot. 7308 */ 7309 if (found > count) 7310 return true; 7311 } 7312 return false; 7313 } 7314 7315 bool is_pageblock_removable_nolock(struct page *page) 7316 { 7317 struct zone *zone; 7318 unsigned long pfn; 7319 7320 /* 7321 * We have to be careful here because we are iterating over memory 7322 * sections which are not zone aware so we might end up outside of 7323 * the zone but still within the section. 7324 * We have to take care about the node as well. If the node is offline 7325 * its NODE_DATA will be NULL - see page_zone. 7326 */ 7327 if (!node_online(page_to_nid(page))) 7328 return false; 7329 7330 zone = page_zone(page); 7331 pfn = page_to_pfn(page); 7332 if (!zone_spans_pfn(zone, pfn)) 7333 return false; 7334 7335 return !has_unmovable_pages(zone, page, 0, true); 7336 } 7337 7338 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA) 7339 7340 static unsigned long pfn_max_align_down(unsigned long pfn) 7341 { 7342 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES, 7343 pageblock_nr_pages) - 1); 7344 } 7345 7346 static unsigned long pfn_max_align_up(unsigned long pfn) 7347 { 7348 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES, 7349 pageblock_nr_pages)); 7350 } 7351 7352 /* [start, end) must belong to a single zone. */ 7353 static int __alloc_contig_migrate_range(struct compact_control *cc, 7354 unsigned long start, unsigned long end) 7355 { 7356 /* This function is based on compact_zone() from compaction.c. */ 7357 unsigned long nr_reclaimed; 7358 unsigned long pfn = start; 7359 unsigned int tries = 0; 7360 int ret = 0; 7361 7362 migrate_prep(); 7363 7364 while (pfn < end || !list_empty(&cc->migratepages)) { 7365 if (fatal_signal_pending(current)) { 7366 ret = -EINTR; 7367 break; 7368 } 7369 7370 if (list_empty(&cc->migratepages)) { 7371 cc->nr_migratepages = 0; 7372 pfn = isolate_migratepages_range(cc, pfn, end); 7373 if (!pfn) { 7374 ret = -EINTR; 7375 break; 7376 } 7377 tries = 0; 7378 } else if (++tries == 5) { 7379 ret = ret < 0 ? ret : -EBUSY; 7380 break; 7381 } 7382 7383 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone, 7384 &cc->migratepages); 7385 cc->nr_migratepages -= nr_reclaimed; 7386 7387 ret = migrate_pages(&cc->migratepages, alloc_migrate_target, 7388 NULL, 0, cc->mode, MR_CMA); 7389 } 7390 if (ret < 0) { 7391 putback_movable_pages(&cc->migratepages); 7392 return ret; 7393 } 7394 return 0; 7395 } 7396 7397 /** 7398 * alloc_contig_range() -- tries to allocate given range of pages 7399 * @start: start PFN to allocate 7400 * @end: one-past-the-last PFN to allocate 7401 * @migratetype: migratetype of the underlaying pageblocks (either 7402 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks 7403 * in range must have the same migratetype and it must 7404 * be either of the two. 7405 * @gfp_mask: GFP mask to use during compaction 7406 * 7407 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES 7408 * aligned, however it's the caller's responsibility to guarantee that 7409 * we are the only thread that changes migrate type of pageblocks the 7410 * pages fall in. 7411 * 7412 * The PFN range must belong to a single zone. 7413 * 7414 * Returns zero on success or negative error code. On success all 7415 * pages which PFN is in [start, end) are allocated for the caller and 7416 * need to be freed with free_contig_range(). 7417 */ 7418 int alloc_contig_range(unsigned long start, unsigned long end, 7419 unsigned migratetype, gfp_t gfp_mask) 7420 { 7421 unsigned long outer_start, outer_end; 7422 unsigned int order; 7423 int ret = 0; 7424 7425 struct compact_control cc = { 7426 .nr_migratepages = 0, 7427 .order = -1, 7428 .zone = page_zone(pfn_to_page(start)), 7429 .mode = MIGRATE_SYNC, 7430 .ignore_skip_hint = true, 7431 .gfp_mask = memalloc_noio_flags(gfp_mask), 7432 }; 7433 INIT_LIST_HEAD(&cc.migratepages); 7434 7435 /* 7436 * What we do here is we mark all pageblocks in range as 7437 * MIGRATE_ISOLATE. Because pageblock and max order pages may 7438 * have different sizes, and due to the way page allocator 7439 * work, we align the range to biggest of the two pages so 7440 * that page allocator won't try to merge buddies from 7441 * different pageblocks and change MIGRATE_ISOLATE to some 7442 * other migration type. 7443 * 7444 * Once the pageblocks are marked as MIGRATE_ISOLATE, we 7445 * migrate the pages from an unaligned range (ie. pages that 7446 * we are interested in). This will put all the pages in 7447 * range back to page allocator as MIGRATE_ISOLATE. 7448 * 7449 * When this is done, we take the pages in range from page 7450 * allocator removing them from the buddy system. This way 7451 * page allocator will never consider using them. 7452 * 7453 * This lets us mark the pageblocks back as 7454 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the 7455 * aligned range but not in the unaligned, original range are 7456 * put back to page allocator so that buddy can use them. 7457 */ 7458 7459 ret = start_isolate_page_range(pfn_max_align_down(start), 7460 pfn_max_align_up(end), migratetype, 7461 false); 7462 if (ret) 7463 return ret; 7464 7465 /* 7466 * In case of -EBUSY, we'd like to know which page causes problem. 7467 * So, just fall through. We will check it in test_pages_isolated(). 7468 */ 7469 ret = __alloc_contig_migrate_range(&cc, start, end); 7470 if (ret && ret != -EBUSY) 7471 goto done; 7472 7473 /* 7474 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES 7475 * aligned blocks that are marked as MIGRATE_ISOLATE. What's 7476 * more, all pages in [start, end) are free in page allocator. 7477 * What we are going to do is to allocate all pages from 7478 * [start, end) (that is remove them from page allocator). 7479 * 7480 * The only problem is that pages at the beginning and at the 7481 * end of interesting range may be not aligned with pages that 7482 * page allocator holds, ie. they can be part of higher order 7483 * pages. Because of this, we reserve the bigger range and 7484 * once this is done free the pages we are not interested in. 7485 * 7486 * We don't have to hold zone->lock here because the pages are 7487 * isolated thus they won't get removed from buddy. 7488 */ 7489 7490 lru_add_drain_all(); 7491 drain_all_pages(cc.zone); 7492 7493 order = 0; 7494 outer_start = start; 7495 while (!PageBuddy(pfn_to_page(outer_start))) { 7496 if (++order >= MAX_ORDER) { 7497 outer_start = start; 7498 break; 7499 } 7500 outer_start &= ~0UL << order; 7501 } 7502 7503 if (outer_start != start) { 7504 order = page_order(pfn_to_page(outer_start)); 7505 7506 /* 7507 * outer_start page could be small order buddy page and 7508 * it doesn't include start page. Adjust outer_start 7509 * in this case to report failed page properly 7510 * on tracepoint in test_pages_isolated() 7511 */ 7512 if (outer_start + (1UL << order) <= start) 7513 outer_start = start; 7514 } 7515 7516 /* Make sure the range is really isolated. */ 7517 if (test_pages_isolated(outer_start, end, false)) { 7518 pr_info("%s: [%lx, %lx) PFNs busy\n", 7519 __func__, outer_start, end); 7520 ret = -EBUSY; 7521 goto done; 7522 } 7523 7524 /* Grab isolated pages from freelists. */ 7525 outer_end = isolate_freepages_range(&cc, outer_start, end); 7526 if (!outer_end) { 7527 ret = -EBUSY; 7528 goto done; 7529 } 7530 7531 /* Free head and tail (if any) */ 7532 if (start != outer_start) 7533 free_contig_range(outer_start, start - outer_start); 7534 if (end != outer_end) 7535 free_contig_range(end, outer_end - end); 7536 7537 done: 7538 undo_isolate_page_range(pfn_max_align_down(start), 7539 pfn_max_align_up(end), migratetype); 7540 return ret; 7541 } 7542 7543 void free_contig_range(unsigned long pfn, unsigned nr_pages) 7544 { 7545 unsigned int count = 0; 7546 7547 for (; nr_pages--; pfn++) { 7548 struct page *page = pfn_to_page(pfn); 7549 7550 count += page_count(page) != 1; 7551 __free_page(page); 7552 } 7553 WARN(count != 0, "%d pages are still in use!\n", count); 7554 } 7555 #endif 7556 7557 #ifdef CONFIG_MEMORY_HOTPLUG 7558 /* 7559 * The zone indicated has a new number of managed_pages; batch sizes and percpu 7560 * page high values need to be recalulated. 7561 */ 7562 void __meminit zone_pcp_update(struct zone *zone) 7563 { 7564 unsigned cpu; 7565 mutex_lock(&pcp_batch_high_lock); 7566 for_each_possible_cpu(cpu) 7567 pageset_set_high_and_batch(zone, 7568 per_cpu_ptr(zone->pageset, cpu)); 7569 mutex_unlock(&pcp_batch_high_lock); 7570 } 7571 #endif 7572 7573 void zone_pcp_reset(struct zone *zone) 7574 { 7575 unsigned long flags; 7576 int cpu; 7577 struct per_cpu_pageset *pset; 7578 7579 /* avoid races with drain_pages() */ 7580 local_irq_save(flags); 7581 if (zone->pageset != &boot_pageset) { 7582 for_each_online_cpu(cpu) { 7583 pset = per_cpu_ptr(zone->pageset, cpu); 7584 drain_zonestat(zone, pset); 7585 } 7586 free_percpu(zone->pageset); 7587 zone->pageset = &boot_pageset; 7588 } 7589 local_irq_restore(flags); 7590 } 7591 7592 #ifdef CONFIG_MEMORY_HOTREMOVE 7593 /* 7594 * All pages in the range must be in a single zone and isolated 7595 * before calling this. 7596 */ 7597 void 7598 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn) 7599 { 7600 struct page *page; 7601 struct zone *zone; 7602 unsigned int order, i; 7603 unsigned long pfn; 7604 unsigned long flags; 7605 /* find the first valid pfn */ 7606 for (pfn = start_pfn; pfn < end_pfn; pfn++) 7607 if (pfn_valid(pfn)) 7608 break; 7609 if (pfn == end_pfn) 7610 return; 7611 zone = page_zone(pfn_to_page(pfn)); 7612 spin_lock_irqsave(&zone->lock, flags); 7613 pfn = start_pfn; 7614 while (pfn < end_pfn) { 7615 if (!pfn_valid(pfn)) { 7616 pfn++; 7617 continue; 7618 } 7619 page = pfn_to_page(pfn); 7620 /* 7621 * The HWPoisoned page may be not in buddy system, and 7622 * page_count() is not 0. 7623 */ 7624 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) { 7625 pfn++; 7626 SetPageReserved(page); 7627 continue; 7628 } 7629 7630 BUG_ON(page_count(page)); 7631 BUG_ON(!PageBuddy(page)); 7632 order = page_order(page); 7633 #ifdef CONFIG_DEBUG_VM 7634 pr_info("remove from free list %lx %d %lx\n", 7635 pfn, 1 << order, end_pfn); 7636 #endif 7637 list_del(&page->lru); 7638 rmv_page_order(page); 7639 zone->free_area[order].nr_free--; 7640 for (i = 0; i < (1 << order); i++) 7641 SetPageReserved((page+i)); 7642 pfn += (1 << order); 7643 } 7644 spin_unlock_irqrestore(&zone->lock, flags); 7645 } 7646 #endif 7647 7648 bool is_free_buddy_page(struct page *page) 7649 { 7650 struct zone *zone = page_zone(page); 7651 unsigned long pfn = page_to_pfn(page); 7652 unsigned long flags; 7653 unsigned int order; 7654 7655 spin_lock_irqsave(&zone->lock, flags); 7656 for (order = 0; order < MAX_ORDER; order++) { 7657 struct page *page_head = page - (pfn & ((1 << order) - 1)); 7658 7659 if (PageBuddy(page_head) && page_order(page_head) >= order) 7660 break; 7661 } 7662 spin_unlock_irqrestore(&zone->lock, flags); 7663 7664 return order < MAX_ORDER; 7665 } 7666